diff --git a/sys/amd64/amd64/pmap.c b/sys/amd64/amd64/pmap.c index 4325acd1255a..0672c957ff75 100644 --- a/sys/amd64/amd64/pmap.c +++ b/sys/amd64/amd64/pmap.c @@ -1,11893 +1,11893 @@ /*- * SPDX-License-Identifier: BSD-4-Clause * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2003 Peter Wemm * All rights reserved. * Copyright (c) 2005-2010 Alan L. Cox * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department and William Jolitz of UUNET Technologies Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 */ /*- * Copyright (c) 2003 Networks Associates Technology, Inc. * Copyright (c) 2014-2020 The FreeBSD Foundation * All rights reserved. * * This software was developed for the FreeBSD Project by Jake Burkholder, * Safeport Network Services, and Network Associates Laboratories, the * Security Research Division of Network Associates, Inc. under * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA * CHATS research program. * * Portions of this software were developed by * Konstantin Belousov under sponsorship from * the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #define AMD64_NPT_AWARE #include __FBSDID("$FreeBSD$"); /* * Manages physical address maps. * * Since the information managed by this module is * also stored by the logical address mapping module, * this module may throw away valid virtual-to-physical * mappings at almost any time. However, invalidations * of virtual-to-physical mappings must be done as * requested. * * In order to cope with hardware architectures which * make virtual-to-physical map invalidates expensive, * this module may delay invalidate or reduced protection * operations until such time as they are actually * necessary. This module is given full information as * to which processors are currently using which maps, * and to when physical maps must be made correct. */ #include "opt_ddb.h" #include "opt_pmap.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DDB #include #include #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef SMP #include #endif #include #include #ifdef NUMA #define PMAP_MEMDOM MAXMEMDOM #else #define PMAP_MEMDOM 1 #endif static __inline boolean_t pmap_type_guest(pmap_t pmap) { return ((pmap->pm_type == PT_EPT) || (pmap->pm_type == PT_RVI)); } static __inline boolean_t pmap_emulate_ad_bits(pmap_t pmap) { return ((pmap->pm_flags & PMAP_EMULATE_AD_BITS) != 0); } static __inline pt_entry_t pmap_valid_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = X86_PG_V; break; case PT_EPT: if (pmap_emulate_ad_bits(pmap)) mask = EPT_PG_EMUL_V; else mask = EPT_PG_READ; break; default: panic("pmap_valid_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline pt_entry_t pmap_rw_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = X86_PG_RW; break; case PT_EPT: if (pmap_emulate_ad_bits(pmap)) mask = EPT_PG_EMUL_RW; else mask = EPT_PG_WRITE; break; default: panic("pmap_rw_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static pt_entry_t pg_g; static __inline pt_entry_t pmap_global_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: mask = pg_g; break; case PT_RVI: case PT_EPT: mask = 0; break; default: panic("pmap_global_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline pt_entry_t pmap_accessed_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = X86_PG_A; break; case PT_EPT: if (pmap_emulate_ad_bits(pmap)) mask = EPT_PG_READ; else mask = EPT_PG_A; break; default: panic("pmap_accessed_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline pt_entry_t pmap_modified_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = X86_PG_M; break; case PT_EPT: if (pmap_emulate_ad_bits(pmap)) mask = EPT_PG_WRITE; else mask = EPT_PG_M; break; default: panic("pmap_modified_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline pt_entry_t pmap_pku_mask_bit(pmap_t pmap) { return (pmap->pm_type == PT_X86 ? X86_PG_PKU_MASK : 0); } #if !defined(DIAGNOSTIC) #ifdef __GNUC_GNU_INLINE__ #define PMAP_INLINE __attribute__((__gnu_inline__)) inline #else #define PMAP_INLINE extern inline #endif #else #define PMAP_INLINE #endif #ifdef PV_STATS #define PV_STAT(x) do { x ; } while (0) #else #define PV_STAT(x) do { } while (0) #endif #undef pa_index #ifdef NUMA #define pa_index(pa) ({ \ KASSERT((pa) <= vm_phys_segs[vm_phys_nsegs - 1].end, \ ("address %lx beyond the last segment", (pa))); \ (pa) >> PDRSHIFT; \ }) #define pa_to_pmdp(pa) (&pv_table[pa_index(pa)]) #define pa_to_pvh(pa) (&(pa_to_pmdp(pa)->pv_page)) #define PHYS_TO_PV_LIST_LOCK(pa) ({ \ struct rwlock *_lock; \ if (__predict_false((pa) > pmap_last_pa)) \ _lock = &pv_dummy_large.pv_lock; \ else \ _lock = &(pa_to_pmdp(pa)->pv_lock); \ _lock; \ }) #else #define pa_index(pa) ((pa) >> PDRSHIFT) #define pa_to_pvh(pa) (&pv_table[pa_index(pa)]) #define NPV_LIST_LOCKS MAXCPU #define PHYS_TO_PV_LIST_LOCK(pa) \ (&pv_list_locks[pa_index(pa) % NPV_LIST_LOCKS]) #endif #define CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa) do { \ struct rwlock **_lockp = (lockp); \ struct rwlock *_new_lock; \ \ _new_lock = PHYS_TO_PV_LIST_LOCK(pa); \ if (_new_lock != *_lockp) { \ if (*_lockp != NULL) \ rw_wunlock(*_lockp); \ *_lockp = _new_lock; \ rw_wlock(*_lockp); \ } \ } while (0) #define CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m) \ CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, VM_PAGE_TO_PHYS(m)) #define RELEASE_PV_LIST_LOCK(lockp) do { \ struct rwlock **_lockp = (lockp); \ \ if (*_lockp != NULL) { \ rw_wunlock(*_lockp); \ *_lockp = NULL; \ } \ } while (0) #define VM_PAGE_TO_PV_LIST_LOCK(m) \ PHYS_TO_PV_LIST_LOCK(VM_PAGE_TO_PHYS(m)) struct pmap kernel_pmap_store; vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ int nkpt; SYSCTL_INT(_machdep, OID_AUTO, nkpt, CTLFLAG_RD, &nkpt, 0, "Number of kernel page table pages allocated on bootup"); static int ndmpdp; vm_paddr_t dmaplimit; vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS; pt_entry_t pg_nx; static SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "VM/pmap parameters"); static int pg_ps_enabled = 1; SYSCTL_INT(_vm_pmap, OID_AUTO, pg_ps_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pg_ps_enabled, 0, "Are large page mappings enabled?"); int __read_frequently la57 = 0; SYSCTL_INT(_vm_pmap, OID_AUTO, la57, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &la57, 0, "5-level paging for host is enabled"); static bool pmap_is_la57(pmap_t pmap) { if (pmap->pm_type == PT_X86) return (la57); return (false); /* XXXKIB handle EPT */ } #define PAT_INDEX_SIZE 8 static int pat_index[PAT_INDEX_SIZE]; /* cache mode to PAT index conversion */ static u_int64_t KPTphys; /* phys addr of kernel level 1 */ static u_int64_t KPDphys; /* phys addr of kernel level 2 */ static u_int64_t KPDPphys; /* phys addr of kernel level 3 */ u_int64_t KPML4phys; /* phys addr of kernel level 4 */ u_int64_t KPML5phys; /* phys addr of kernel level 5, if supported */ #ifdef KASAN static uint64_t KASANPDPphys; #endif static pml4_entry_t *kernel_pml4; static u_int64_t DMPDphys; /* phys addr of direct mapped level 2 */ static u_int64_t DMPDPphys; /* phys addr of direct mapped level 3 */ static int ndmpdpphys; /* number of DMPDPphys pages */ vm_paddr_t kernphys; /* phys addr of start of bootstrap data */ vm_paddr_t KERNend; /* and the end */ /* * pmap_mapdev support pre initialization (i.e. console) */ #define PMAP_PREINIT_MAPPING_COUNT 8 static struct pmap_preinit_mapping { vm_paddr_t pa; vm_offset_t va; vm_size_t sz; int mode; } pmap_preinit_mapping[PMAP_PREINIT_MAPPING_COUNT]; static int pmap_initialized; /* * Data for the pv entry allocation mechanism. * Updates to pv_invl_gen are protected by the pv list lock but reads are not. */ #ifdef NUMA static __inline int pc_to_domain(struct pv_chunk *pc) { return (vm_phys_domain(DMAP_TO_PHYS((vm_offset_t)pc))); } #else static __inline int pc_to_domain(struct pv_chunk *pc __unused) { return (0); } #endif struct pv_chunks_list { struct mtx pvc_lock; TAILQ_HEAD(pch, pv_chunk) pvc_list; int active_reclaims; } __aligned(CACHE_LINE_SIZE); struct pv_chunks_list __exclusive_cache_line pv_chunks[PMAP_MEMDOM]; #ifdef NUMA struct pmap_large_md_page { struct rwlock pv_lock; struct md_page pv_page; u_long pv_invl_gen; }; __exclusive_cache_line static struct pmap_large_md_page pv_dummy_large; #define pv_dummy pv_dummy_large.pv_page __read_mostly static struct pmap_large_md_page *pv_table; __read_mostly vm_paddr_t pmap_last_pa; #else static struct rwlock __exclusive_cache_line pv_list_locks[NPV_LIST_LOCKS]; static u_long pv_invl_gen[NPV_LIST_LOCKS]; static struct md_page *pv_table; static struct md_page pv_dummy; #endif /* * All those kernel PT submaps that BSD is so fond of */ pt_entry_t *CMAP1 = NULL; caddr_t CADDR1 = 0; static vm_offset_t qframe = 0; static struct mtx qframe_mtx; static int pmap_flags = PMAP_PDE_SUPERPAGE; /* flags for x86 pmaps */ static vmem_t *large_vmem; static u_int lm_ents; #define PMAP_ADDRESS_IN_LARGEMAP(va) ((va) >= LARGEMAP_MIN_ADDRESS && \ (va) < LARGEMAP_MIN_ADDRESS + NBPML4 * (u_long)lm_ents) int pmap_pcid_enabled = 1; SYSCTL_INT(_vm_pmap, OID_AUTO, pcid_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pmap_pcid_enabled, 0, "Is TLB Context ID enabled ?"); int invpcid_works = 0; SYSCTL_INT(_vm_pmap, OID_AUTO, invpcid_works, CTLFLAG_RD, &invpcid_works, 0, "Is the invpcid instruction available ?"); int __read_frequently pti = 0; SYSCTL_INT(_vm_pmap, OID_AUTO, pti, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pti, 0, "Page Table Isolation enabled"); static vm_object_t pti_obj; static pml4_entry_t *pti_pml4; static vm_pindex_t pti_pg_idx; static bool pti_finalized; struct pmap_pkru_range { struct rs_el pkru_rs_el; u_int pkru_keyidx; int pkru_flags; }; static uma_zone_t pmap_pkru_ranges_zone; static bool pmap_pkru_same(pmap_t pmap, vm_offset_t sva, vm_offset_t eva); static pt_entry_t pmap_pkru_get(pmap_t pmap, vm_offset_t va); static void pmap_pkru_on_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva); static void *pkru_dup_range(void *ctx, void *data); static void pkru_free_range(void *ctx, void *node); static int pmap_pkru_copy(pmap_t dst_pmap, pmap_t src_pmap); static int pmap_pkru_deassign(pmap_t pmap, vm_offset_t sva, vm_offset_t eva); static void pmap_pkru_deassign_all(pmap_t pmap); static COUNTER_U64_DEFINE_EARLY(pcid_save_cnt); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pcid_save_cnt, CTLFLAG_RD, &pcid_save_cnt, "Count of saved TLB context on switch"); static LIST_HEAD(, pmap_invl_gen) pmap_invl_gen_tracker = LIST_HEAD_INITIALIZER(&pmap_invl_gen_tracker); static struct mtx invl_gen_mtx; /* Fake lock object to satisfy turnstiles interface. */ static struct lock_object invl_gen_ts = { .lo_name = "invlts", }; static struct pmap_invl_gen pmap_invl_gen_head = { .gen = 1, .next = NULL, }; static u_long pmap_invl_gen = 1; static int pmap_invl_waiters; static struct callout pmap_invl_callout; static bool pmap_invl_callout_inited; #define PMAP_ASSERT_NOT_IN_DI() \ KASSERT(pmap_not_in_di(), ("DI already started")) static bool pmap_di_locked(void) { int tun; if ((cpu_feature2 & CPUID2_CX16) == 0) return (true); tun = 0; TUNABLE_INT_FETCH("vm.pmap.di_locked", &tun); return (tun != 0); } static int sysctl_pmap_di_locked(SYSCTL_HANDLER_ARGS) { int locked; locked = pmap_di_locked(); return (sysctl_handle_int(oidp, &locked, 0, req)); } SYSCTL_PROC(_vm_pmap, OID_AUTO, di_locked, CTLTYPE_INT | CTLFLAG_RDTUN | CTLFLAG_MPSAFE, 0, 0, sysctl_pmap_di_locked, "", "Locked delayed invalidation"); static bool pmap_not_in_di_l(void); static bool pmap_not_in_di_u(void); DEFINE_IFUNC(, bool, pmap_not_in_di, (void)) { return (pmap_di_locked() ? pmap_not_in_di_l : pmap_not_in_di_u); } static bool pmap_not_in_di_l(void) { struct pmap_invl_gen *invl_gen; invl_gen = &curthread->td_md.md_invl_gen; return (invl_gen->gen == 0); } static void pmap_thread_init_invl_gen_l(struct thread *td) { struct pmap_invl_gen *invl_gen; invl_gen = &td->td_md.md_invl_gen; invl_gen->gen = 0; } static void pmap_delayed_invl_wait_block(u_long *m_gen, u_long *invl_gen) { struct turnstile *ts; ts = turnstile_trywait(&invl_gen_ts); if (*m_gen > atomic_load_long(invl_gen)) turnstile_wait(ts, NULL, TS_SHARED_QUEUE); else turnstile_cancel(ts); } static void pmap_delayed_invl_finish_unblock(u_long new_gen) { struct turnstile *ts; turnstile_chain_lock(&invl_gen_ts); ts = turnstile_lookup(&invl_gen_ts); if (new_gen != 0) pmap_invl_gen = new_gen; if (ts != NULL) { turnstile_broadcast(ts, TS_SHARED_QUEUE); turnstile_unpend(ts); } turnstile_chain_unlock(&invl_gen_ts); } /* * Start a new Delayed Invalidation (DI) block of code, executed by * the current thread. Within a DI block, the current thread may * destroy both the page table and PV list entries for a mapping and * then release the corresponding PV list lock before ensuring that * the mapping is flushed from the TLBs of any processors with the * pmap active. */ static void pmap_delayed_invl_start_l(void) { struct pmap_invl_gen *invl_gen; u_long currgen; invl_gen = &curthread->td_md.md_invl_gen; PMAP_ASSERT_NOT_IN_DI(); mtx_lock(&invl_gen_mtx); if (LIST_EMPTY(&pmap_invl_gen_tracker)) currgen = pmap_invl_gen; else currgen = LIST_FIRST(&pmap_invl_gen_tracker)->gen; invl_gen->gen = currgen + 1; LIST_INSERT_HEAD(&pmap_invl_gen_tracker, invl_gen, link); mtx_unlock(&invl_gen_mtx); } /* * Finish the DI block, previously started by the current thread. All * required TLB flushes for the pages marked by * pmap_delayed_invl_page() must be finished before this function is * called. * * This function works by bumping the global DI generation number to * the generation number of the current thread's DI, unless there is a * pending DI that started earlier. In the latter case, bumping the * global DI generation number would incorrectly signal that the * earlier DI had finished. Instead, this function bumps the earlier * DI's generation number to match the generation number of the * current thread's DI. */ static void pmap_delayed_invl_finish_l(void) { struct pmap_invl_gen *invl_gen, *next; invl_gen = &curthread->td_md.md_invl_gen; KASSERT(invl_gen->gen != 0, ("missed invl_start")); mtx_lock(&invl_gen_mtx); next = LIST_NEXT(invl_gen, link); if (next == NULL) pmap_delayed_invl_finish_unblock(invl_gen->gen); else next->gen = invl_gen->gen; LIST_REMOVE(invl_gen, link); mtx_unlock(&invl_gen_mtx); invl_gen->gen = 0; } static bool pmap_not_in_di_u(void) { struct pmap_invl_gen *invl_gen; invl_gen = &curthread->td_md.md_invl_gen; return (((uintptr_t)invl_gen->next & PMAP_INVL_GEN_NEXT_INVALID) != 0); } static void pmap_thread_init_invl_gen_u(struct thread *td) { struct pmap_invl_gen *invl_gen; invl_gen = &td->td_md.md_invl_gen; invl_gen->gen = 0; invl_gen->next = (void *)PMAP_INVL_GEN_NEXT_INVALID; } static bool pmap_di_load_invl(struct pmap_invl_gen *ptr, struct pmap_invl_gen *out) { uint64_t new_high, new_low, old_high, old_low; char res; old_low = new_low = 0; old_high = new_high = (uintptr_t)0; __asm volatile("lock;cmpxchg16b\t%1" : "=@cce" (res), "+m" (*ptr), "+a" (old_low), "+d" (old_high) : "b"(new_low), "c" (new_high) : "memory", "cc"); if (res == 0) { if ((old_high & PMAP_INVL_GEN_NEXT_INVALID) != 0) return (false); out->gen = old_low; out->next = (void *)old_high; } else { out->gen = new_low; out->next = (void *)new_high; } return (true); } static bool pmap_di_store_invl(struct pmap_invl_gen *ptr, struct pmap_invl_gen *old_val, struct pmap_invl_gen *new_val) { uint64_t new_high, new_low, old_high, old_low; char res; new_low = new_val->gen; new_high = (uintptr_t)new_val->next; old_low = old_val->gen; old_high = (uintptr_t)old_val->next; __asm volatile("lock;cmpxchg16b\t%1" : "=@cce" (res), "+m" (*ptr), "+a" (old_low), "+d" (old_high) : "b"(new_low), "c" (new_high) : "memory", "cc"); return (res); } static COUNTER_U64_DEFINE_EARLY(pv_page_count); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pv_page_count, CTLFLAG_RD, &pv_page_count, "Current number of allocated pv pages"); static COUNTER_U64_DEFINE_EARLY(user_pt_page_count); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, user_pt_page_count, CTLFLAG_RD, &user_pt_page_count, "Current number of allocated page table pages for userspace"); static COUNTER_U64_DEFINE_EARLY(kernel_pt_page_count); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, kernel_pt_page_count, CTLFLAG_RD, &kernel_pt_page_count, "Current number of allocated page table pages for the kernel"); #ifdef PV_STATS static COUNTER_U64_DEFINE_EARLY(invl_start_restart); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, invl_start_restart, CTLFLAG_RD, &invl_start_restart, "Number of delayed TLB invalidation request restarts"); static COUNTER_U64_DEFINE_EARLY(invl_finish_restart); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, invl_finish_restart, CTLFLAG_RD, &invl_finish_restart, "Number of delayed TLB invalidation completion restarts"); static int invl_max_qlen; SYSCTL_INT(_vm_pmap, OID_AUTO, invl_max_qlen, CTLFLAG_RD, &invl_max_qlen, 0, "Maximum delayed TLB invalidation request queue length"); #endif #define di_delay locks_delay static void pmap_delayed_invl_start_u(void) { struct pmap_invl_gen *invl_gen, *p, prev, new_prev; struct thread *td; struct lock_delay_arg lda; uintptr_t prevl; u_char pri; #ifdef PV_STATS int i, ii; #endif td = curthread; invl_gen = &td->td_md.md_invl_gen; PMAP_ASSERT_NOT_IN_DI(); lock_delay_arg_init(&lda, &di_delay); invl_gen->saved_pri = 0; pri = td->td_base_pri; if (pri > PVM) { thread_lock(td); pri = td->td_base_pri; if (pri > PVM) { invl_gen->saved_pri = pri; sched_prio(td, PVM); } thread_unlock(td); } again: PV_STAT(i = 0); for (p = &pmap_invl_gen_head;; p = prev.next) { PV_STAT(i++); prevl = (uintptr_t)atomic_load_ptr(&p->next); if ((prevl & PMAP_INVL_GEN_NEXT_INVALID) != 0) { PV_STAT(counter_u64_add(invl_start_restart, 1)); lock_delay(&lda); goto again; } if (prevl == 0) break; prev.next = (void *)prevl; } #ifdef PV_STATS if ((ii = invl_max_qlen) < i) atomic_cmpset_int(&invl_max_qlen, ii, i); #endif if (!pmap_di_load_invl(p, &prev) || prev.next != NULL) { PV_STAT(counter_u64_add(invl_start_restart, 1)); lock_delay(&lda); goto again; } new_prev.gen = prev.gen; new_prev.next = invl_gen; invl_gen->gen = prev.gen + 1; /* Formal fence between store to invl->gen and updating *p. */ atomic_thread_fence_rel(); /* * After inserting an invl_gen element with invalid bit set, * this thread blocks any other thread trying to enter the * delayed invalidation block. Do not allow to remove us from * the CPU, because it causes starvation for other threads. */ critical_enter(); /* * ABA for *p is not possible there, since p->gen can only * increase. So if the *p thread finished its di, then * started a new one and got inserted into the list at the * same place, its gen will appear greater than the previously * read gen. */ if (!pmap_di_store_invl(p, &prev, &new_prev)) { critical_exit(); PV_STAT(counter_u64_add(invl_start_restart, 1)); lock_delay(&lda); goto again; } /* * There we clear PMAP_INVL_GEN_NEXT_INVALID in * invl_gen->next, allowing other threads to iterate past us. * pmap_di_store_invl() provides fence between the generation * write and the update of next. */ invl_gen->next = NULL; critical_exit(); } static bool pmap_delayed_invl_finish_u_crit(struct pmap_invl_gen *invl_gen, struct pmap_invl_gen *p) { struct pmap_invl_gen prev, new_prev; u_long mygen; /* * Load invl_gen->gen after setting invl_gen->next * PMAP_INVL_GEN_NEXT_INVALID. This prevents larger * generations to propagate to our invl_gen->gen. Lock prefix * in atomic_set_ptr() worked as seq_cst fence. */ mygen = atomic_load_long(&invl_gen->gen); if (!pmap_di_load_invl(p, &prev) || prev.next != invl_gen) return (false); KASSERT(prev.gen < mygen, ("invalid di gen sequence %lu %lu", prev.gen, mygen)); new_prev.gen = mygen; new_prev.next = (void *)((uintptr_t)invl_gen->next & ~PMAP_INVL_GEN_NEXT_INVALID); /* Formal fence between load of prev and storing update to it. */ atomic_thread_fence_rel(); return (pmap_di_store_invl(p, &prev, &new_prev)); } static void pmap_delayed_invl_finish_u(void) { struct pmap_invl_gen *invl_gen, *p; struct thread *td; struct lock_delay_arg lda; uintptr_t prevl; td = curthread; invl_gen = &td->td_md.md_invl_gen; KASSERT(invl_gen->gen != 0, ("missed invl_start: gen 0")); KASSERT(((uintptr_t)invl_gen->next & PMAP_INVL_GEN_NEXT_INVALID) == 0, ("missed invl_start: INVALID")); lock_delay_arg_init(&lda, &di_delay); again: for (p = &pmap_invl_gen_head; p != NULL; p = (void *)prevl) { prevl = (uintptr_t)atomic_load_ptr(&p->next); if ((prevl & PMAP_INVL_GEN_NEXT_INVALID) != 0) { PV_STAT(counter_u64_add(invl_finish_restart, 1)); lock_delay(&lda); goto again; } if ((void *)prevl == invl_gen) break; } /* * It is legitimate to not find ourself on the list if a * thread before us finished its DI and started it again. */ if (__predict_false(p == NULL)) { PV_STAT(counter_u64_add(invl_finish_restart, 1)); lock_delay(&lda); goto again; } critical_enter(); atomic_set_ptr((uintptr_t *)&invl_gen->next, PMAP_INVL_GEN_NEXT_INVALID); if (!pmap_delayed_invl_finish_u_crit(invl_gen, p)) { atomic_clear_ptr((uintptr_t *)&invl_gen->next, PMAP_INVL_GEN_NEXT_INVALID); critical_exit(); PV_STAT(counter_u64_add(invl_finish_restart, 1)); lock_delay(&lda); goto again; } critical_exit(); if (atomic_load_int(&pmap_invl_waiters) > 0) pmap_delayed_invl_finish_unblock(0); if (invl_gen->saved_pri != 0) { thread_lock(td); sched_prio(td, invl_gen->saved_pri); thread_unlock(td); } } #ifdef DDB DB_SHOW_COMMAND(di_queue, pmap_di_queue) { struct pmap_invl_gen *p, *pn; struct thread *td; uintptr_t nextl; bool first; for (p = &pmap_invl_gen_head, first = true; p != NULL; p = pn, first = false) { nextl = (uintptr_t)atomic_load_ptr(&p->next); pn = (void *)(nextl & ~PMAP_INVL_GEN_NEXT_INVALID); td = first ? NULL : __containerof(p, struct thread, td_md.md_invl_gen); db_printf("gen %lu inv %d td %p tid %d\n", p->gen, (nextl & PMAP_INVL_GEN_NEXT_INVALID) != 0, td, td != NULL ? td->td_tid : -1); } } #endif #ifdef PV_STATS static COUNTER_U64_DEFINE_EARLY(invl_wait); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, invl_wait, CTLFLAG_RD, &invl_wait, "Number of times DI invalidation blocked pmap_remove_all/write"); static COUNTER_U64_DEFINE_EARLY(invl_wait_slow); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, invl_wait_slow, CTLFLAG_RD, &invl_wait_slow, "Number of slow invalidation waits for lockless DI"); #endif #ifdef NUMA static u_long * pmap_delayed_invl_genp(vm_page_t m) { vm_paddr_t pa; u_long *gen; pa = VM_PAGE_TO_PHYS(m); if (__predict_false((pa) > pmap_last_pa)) gen = &pv_dummy_large.pv_invl_gen; else gen = &(pa_to_pmdp(pa)->pv_invl_gen); return (gen); } #else static u_long * pmap_delayed_invl_genp(vm_page_t m) { return (&pv_invl_gen[pa_index(VM_PAGE_TO_PHYS(m)) % NPV_LIST_LOCKS]); } #endif static void pmap_delayed_invl_callout_func(void *arg __unused) { if (atomic_load_int(&pmap_invl_waiters) == 0) return; pmap_delayed_invl_finish_unblock(0); } static void pmap_delayed_invl_callout_init(void *arg __unused) { if (pmap_di_locked()) return; callout_init(&pmap_invl_callout, 1); pmap_invl_callout_inited = true; } SYSINIT(pmap_di_callout, SI_SUB_CPU + 1, SI_ORDER_ANY, pmap_delayed_invl_callout_init, NULL); /* * Ensure that all currently executing DI blocks, that need to flush * TLB for the given page m, actually flushed the TLB at the time the * function returned. If the page m has an empty PV list and we call * pmap_delayed_invl_wait(), upon its return we know that no CPU has a * valid mapping for the page m in either its page table or TLB. * * This function works by blocking until the global DI generation * number catches up with the generation number associated with the * given page m and its PV list. Since this function's callers * typically own an object lock and sometimes own a page lock, it * cannot sleep. Instead, it blocks on a turnstile to relinquish the * processor. */ static void pmap_delayed_invl_wait_l(vm_page_t m) { u_long *m_gen; #ifdef PV_STATS bool accounted = false; #endif m_gen = pmap_delayed_invl_genp(m); while (*m_gen > pmap_invl_gen) { #ifdef PV_STATS if (!accounted) { counter_u64_add(invl_wait, 1); accounted = true; } #endif pmap_delayed_invl_wait_block(m_gen, &pmap_invl_gen); } } static void pmap_delayed_invl_wait_u(vm_page_t m) { u_long *m_gen; struct lock_delay_arg lda; bool fast; fast = true; m_gen = pmap_delayed_invl_genp(m); lock_delay_arg_init(&lda, &di_delay); while (*m_gen > atomic_load_long(&pmap_invl_gen_head.gen)) { if (fast || !pmap_invl_callout_inited) { PV_STAT(counter_u64_add(invl_wait, 1)); lock_delay(&lda); fast = false; } else { /* * The page's invalidation generation number * is still below the current thread's number. * Prepare to block so that we do not waste * CPU cycles or worse, suffer livelock. * * Since it is impossible to block without * racing with pmap_delayed_invl_finish_u(), * prepare for the race by incrementing * pmap_invl_waiters and arming a 1-tick * callout which will unblock us if we lose * the race. */ atomic_add_int(&pmap_invl_waiters, 1); /* * Re-check the current thread's invalidation * generation after incrementing * pmap_invl_waiters, so that there is no race * with pmap_delayed_invl_finish_u() setting * the page generation and checking * pmap_invl_waiters. The only race allowed * is for a missed unblock, which is handled * by the callout. */ if (*m_gen > atomic_load_long(&pmap_invl_gen_head.gen)) { callout_reset(&pmap_invl_callout, 1, pmap_delayed_invl_callout_func, NULL); PV_STAT(counter_u64_add(invl_wait_slow, 1)); pmap_delayed_invl_wait_block(m_gen, &pmap_invl_gen_head.gen); } atomic_add_int(&pmap_invl_waiters, -1); } } } DEFINE_IFUNC(, void, pmap_thread_init_invl_gen, (struct thread *)) { return (pmap_di_locked() ? pmap_thread_init_invl_gen_l : pmap_thread_init_invl_gen_u); } DEFINE_IFUNC(static, void, pmap_delayed_invl_start, (void)) { return (pmap_di_locked() ? pmap_delayed_invl_start_l : pmap_delayed_invl_start_u); } DEFINE_IFUNC(static, void, pmap_delayed_invl_finish, (void)) { return (pmap_di_locked() ? pmap_delayed_invl_finish_l : pmap_delayed_invl_finish_u); } DEFINE_IFUNC(static, void, pmap_delayed_invl_wait, (vm_page_t)) { return (pmap_di_locked() ? pmap_delayed_invl_wait_l : pmap_delayed_invl_wait_u); } /* * Mark the page m's PV list as participating in the current thread's * DI block. Any threads concurrently using m's PV list to remove or * restrict all mappings to m will wait for the current thread's DI * block to complete before proceeding. * * The function works by setting the DI generation number for m's PV * list to at least the DI generation number of the current thread. * This forces a caller of pmap_delayed_invl_wait() to block until * current thread calls pmap_delayed_invl_finish(). */ static void pmap_delayed_invl_page(vm_page_t m) { u_long gen, *m_gen; rw_assert(VM_PAGE_TO_PV_LIST_LOCK(m), RA_WLOCKED); gen = curthread->td_md.md_invl_gen.gen; if (gen == 0) return; m_gen = pmap_delayed_invl_genp(m); if (*m_gen < gen) *m_gen = gen; } /* * Crashdump maps. */ static caddr_t crashdumpmap; /* * Internal flags for pmap_enter()'s helper functions. */ #define PMAP_ENTER_NORECLAIM 0x1000000 /* Don't reclaim PV entries. */ #define PMAP_ENTER_NOREPLACE 0x2000000 /* Don't replace mappings. */ /* * Internal flags for pmap_mapdev_internal() and * pmap_change_props_locked(). */ #define MAPDEV_FLUSHCACHE 0x00000001 /* Flush cache after mapping. */ #define MAPDEV_SETATTR 0x00000002 /* Modify existing attrs. */ #define MAPDEV_ASSERTVALID 0x00000004 /* Assert mapping validity. */ TAILQ_HEAD(pv_chunklist, pv_chunk); static void free_pv_chunk(struct pv_chunk *pc); static void free_pv_chunk_batch(struct pv_chunklist *batch); static void free_pv_entry(pmap_t pmap, pv_entry_t pv); static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp); static int popcnt_pc_map_pq(uint64_t *map); static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp); static void reserve_pv_entries(pmap_t pmap, int needed, struct rwlock **lockp); static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp); static bool pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, pd_entry_t pde, u_int flags, struct rwlock **lockp); #if VM_NRESERVLEVEL > 0 static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp); #endif static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static void pmap_abort_ptp(pmap_t pmap, vm_offset_t va, vm_page_t mpte); static int pmap_change_props_locked(vm_offset_t va, vm_size_t size, vm_prot_t prot, int mode, int flags); static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va); static boolean_t pmap_demote_pde_locked(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, struct rwlock **lockp); static boolean_t pmap_demote_pdpe(pmap_t pmap, pdp_entry_t *pdpe, vm_offset_t va); static bool pmap_enter_2mpage(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, struct rwlock **lockp); static int pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde, u_int flags, vm_page_t m, struct rwlock **lockp); static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp); static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte); static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted); static void pmap_invalidate_cache_range_selfsnoop(vm_offset_t sva, vm_offset_t eva); static void pmap_invalidate_cache_range_all(vm_offset_t sva, vm_offset_t eva); static void pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, pd_entry_t pde); static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode); static vm_page_t pmap_large_map_getptp_unlocked(void); static vm_paddr_t pmap_large_map_kextract(vm_offset_t va); #if VM_NRESERVLEVEL > 0 static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, struct rwlock **lockp); #endif static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot); static void pmap_pte_props(pt_entry_t *pte, u_long bits, u_long mask); static void pmap_pti_add_kva_locked(vm_offset_t sva, vm_offset_t eva, bool exec); static pdp_entry_t *pmap_pti_pdpe(vm_offset_t va); static pd_entry_t *pmap_pti_pde(vm_offset_t va); static void pmap_pti_wire_pte(void *pte); static int pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva, struct spglist *free, struct rwlock **lockp); static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva, pd_entry_t ptepde, struct spglist *free, struct rwlock **lockp); static vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va); static void pmap_remove_page(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, struct spglist *free); static bool pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, pd_entry_t *pde, struct spglist *free, struct rwlock **lockp); static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m, struct rwlock **lockp); static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde); static void pmap_update_pde_invalidate(pmap_t, vm_offset_t va, pd_entry_t pde); static pd_entry_t *pmap_alloc_pde(pmap_t pmap, vm_offset_t va, vm_page_t *pdpgp, struct rwlock **lockp); static vm_page_t pmap_allocpte_alloc(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp, vm_offset_t va); static vm_page_t pmap_allocpte_nosleep(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp, vm_offset_t va); static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, struct rwlock **lockp); static void _pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free); static int pmap_unuse_pt(pmap_t, vm_offset_t, pd_entry_t, struct spglist *); static vm_page_t pmap_alloc_pt_page(pmap_t, vm_pindex_t, int); static void pmap_free_pt_page(pmap_t, vm_page_t, bool); /********************/ /* Inline functions */ /********************/ /* * Return a non-clipped indexes for a given VA, which are page table * pages indexes at the corresponding level. */ static __inline vm_pindex_t pmap_pde_pindex(vm_offset_t va) { return (va >> PDRSHIFT); } static __inline vm_pindex_t pmap_pdpe_pindex(vm_offset_t va) { return (NUPDE + (va >> PDPSHIFT)); } static __inline vm_pindex_t pmap_pml4e_pindex(vm_offset_t va) { return (NUPDE + NUPDPE + (va >> PML4SHIFT)); } static __inline vm_pindex_t pmap_pml5e_pindex(vm_offset_t va) { return (NUPDE + NUPDPE + NUPML4E + (va >> PML5SHIFT)); } static __inline pml4_entry_t * pmap_pml5e(pmap_t pmap, vm_offset_t va) { MPASS(pmap_is_la57(pmap)); return (&pmap->pm_pmltop[pmap_pml5e_index(va)]); } static __inline pml4_entry_t * pmap_pml5e_u(pmap_t pmap, vm_offset_t va) { MPASS(pmap_is_la57(pmap)); return (&pmap->pm_pmltopu[pmap_pml5e_index(va)]); } static __inline pml4_entry_t * pmap_pml5e_to_pml4e(pml5_entry_t *pml5e, vm_offset_t va) { pml4_entry_t *pml4e; /* XXX MPASS(pmap_is_la57(pmap); */ pml4e = (pml4_entry_t *)PHYS_TO_DMAP(*pml5e & PG_FRAME); return (&pml4e[pmap_pml4e_index(va)]); } /* Return a pointer to the PML4 slot that corresponds to a VA */ static __inline pml4_entry_t * pmap_pml4e(pmap_t pmap, vm_offset_t va) { pml5_entry_t *pml5e; pml4_entry_t *pml4e; pt_entry_t PG_V; if (pmap_is_la57(pmap)) { pml5e = pmap_pml5e(pmap, va); PG_V = pmap_valid_bit(pmap); if ((*pml5e & PG_V) == 0) return (NULL); pml4e = (pml4_entry_t *)PHYS_TO_DMAP(*pml5e & PG_FRAME); } else { pml4e = pmap->pm_pmltop; } return (&pml4e[pmap_pml4e_index(va)]); } static __inline pml4_entry_t * pmap_pml4e_u(pmap_t pmap, vm_offset_t va) { MPASS(!pmap_is_la57(pmap)); return (&pmap->pm_pmltopu[pmap_pml4e_index(va)]); } /* Return a pointer to the PDP slot that corresponds to a VA */ static __inline pdp_entry_t * pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va) { pdp_entry_t *pdpe; pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME); return (&pdpe[pmap_pdpe_index(va)]); } /* Return a pointer to the PDP slot that corresponds to a VA */ static __inline pdp_entry_t * pmap_pdpe(pmap_t pmap, vm_offset_t va) { pml4_entry_t *pml4e; pt_entry_t PG_V; PG_V = pmap_valid_bit(pmap); pml4e = pmap_pml4e(pmap, va); if (pml4e == NULL || (*pml4e & PG_V) == 0) return (NULL); return (pmap_pml4e_to_pdpe(pml4e, va)); } /* Return a pointer to the PD slot that corresponds to a VA */ static __inline pd_entry_t * pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va) { pd_entry_t *pde; KASSERT((*pdpe & PG_PS) == 0, ("%s: pdpe %#lx is a leaf", __func__, *pdpe)); pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME); return (&pde[pmap_pde_index(va)]); } /* Return a pointer to the PD slot that corresponds to a VA */ static __inline pd_entry_t * pmap_pde(pmap_t pmap, vm_offset_t va) { pdp_entry_t *pdpe; pt_entry_t PG_V; PG_V = pmap_valid_bit(pmap); pdpe = pmap_pdpe(pmap, va); if (pdpe == NULL || (*pdpe & PG_V) == 0) return (NULL); KASSERT((*pdpe & PG_PS) == 0, ("pmap_pde for 1G page, pmap %p va %#lx", pmap, va)); return (pmap_pdpe_to_pde(pdpe, va)); } /* Return a pointer to the PT slot that corresponds to a VA */ static __inline pt_entry_t * pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va) { pt_entry_t *pte; KASSERT((*pde & PG_PS) == 0, ("%s: pde %#lx is a leaf", __func__, *pde)); pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME); return (&pte[pmap_pte_index(va)]); } /* Return a pointer to the PT slot that corresponds to a VA */ static __inline pt_entry_t * pmap_pte(pmap_t pmap, vm_offset_t va) { pd_entry_t *pde; pt_entry_t PG_V; PG_V = pmap_valid_bit(pmap); pde = pmap_pde(pmap, va); if (pde == NULL || (*pde & PG_V) == 0) return (NULL); if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */ return ((pt_entry_t *)pde); return (pmap_pde_to_pte(pde, va)); } static __inline void pmap_resident_count_adj(pmap_t pmap, int count) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(pmap->pm_stats.resident_count + count >= 0, ("pmap %p resident count underflow %ld %d", pmap, pmap->pm_stats.resident_count, count)); pmap->pm_stats.resident_count += count; } static __inline void pmap_pt_page_count_pinit(pmap_t pmap, int count) { KASSERT(pmap->pm_stats.resident_count + count >= 0, ("pmap %p resident count underflow %ld %d", pmap, pmap->pm_stats.resident_count, count)); pmap->pm_stats.resident_count += count; } static __inline void pmap_pt_page_count_adj(pmap_t pmap, int count) { if (pmap == kernel_pmap) counter_u64_add(kernel_pt_page_count, count); else { if (pmap != NULL) pmap_resident_count_adj(pmap, count); counter_u64_add(user_pt_page_count, count); } } PMAP_INLINE pt_entry_t * vtopte(vm_offset_t va) { u_int64_t mask; KASSERT(va >= VM_MAXUSER_ADDRESS, ("vtopte on a uva/gpa 0x%0lx", va)); if (la57) { mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT + NPML5EPGSHIFT)) - 1); return (P5Tmap + ((va >> PAGE_SHIFT) & mask)); } else { mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1); return (P4Tmap + ((va >> PAGE_SHIFT) & mask)); } } static __inline pd_entry_t * vtopde(vm_offset_t va) { u_int64_t mask; KASSERT(va >= VM_MAXUSER_ADDRESS, ("vtopde on a uva/gpa 0x%0lx", va)); if (la57) { mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT + NPML5EPGSHIFT)) - 1); return (P5Dmap + ((va >> PDRSHIFT) & mask)); } else { mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1); return (P4Dmap + ((va >> PDRSHIFT) & mask)); } } static u_int64_t allocpages(vm_paddr_t *firstaddr, int n) { u_int64_t ret; ret = *firstaddr; bzero((void *)ret, n * PAGE_SIZE); *firstaddr += n * PAGE_SIZE; return (ret); } CTASSERT(powerof2(NDMPML4E)); /* number of kernel PDP slots */ #define NKPDPE(ptpgs) howmany(ptpgs, NPDEPG) static void nkpt_init(vm_paddr_t addr) { int pt_pages; #ifdef NKPT pt_pages = NKPT; #else pt_pages = howmany(addr - kernphys, NBPDR) + 1; /* +1 for 2M hole @0 */ pt_pages += NKPDPE(pt_pages); /* * Add some slop beyond the bare minimum required for bootstrapping * the kernel. * * This is quite important when allocating KVA for kernel modules. * The modules are required to be linked in the negative 2GB of * the address space. If we run out of KVA in this region then * pmap_growkernel() will need to allocate page table pages to map * the entire 512GB of KVA space which is an unnecessary tax on * physical memory. * * Secondly, device memory mapped as part of setting up the low- * level console(s) is taken from KVA, starting at virtual_avail. * This is because cninit() is called after pmap_bootstrap() but * before vm_init() and pmap_init(). 20MB for a frame buffer is * not uncommon. */ pt_pages += 32; /* 64MB additional slop. */ #endif nkpt = pt_pages; } /* * Returns the proper write/execute permission for a physical page that is * part of the initial boot allocations. * * If the page has kernel text, it is marked as read-only. If the page has * kernel read-only data, it is marked as read-only/not-executable. If the * page has only read-write data, it is marked as read-write/not-executable. * If the page is below/above the kernel range, it is marked as read-write. * * This function operates on 2M pages, since we map the kernel space that * way. */ static inline pt_entry_t bootaddr_rwx(vm_paddr_t pa) { /* * The kernel is loaded at a 2MB-aligned address, and memory below that * need not be executable. The .bss section is padded to a 2MB * boundary, so memory following the kernel need not be executable * either. Preloaded kernel modules have their mapping permissions * fixed up by the linker. */ if (pa < trunc_2mpage(kernphys + btext - KERNSTART) || pa >= trunc_2mpage(kernphys + _end - KERNSTART)) return (X86_PG_RW | pg_nx); /* * The linker should ensure that the read-only and read-write * portions don't share the same 2M page, so this shouldn't * impact read-only data. However, in any case, any page with * read-write data needs to be read-write. */ if (pa >= trunc_2mpage(kernphys + brwsection - KERNSTART)) return (X86_PG_RW | pg_nx); /* * Mark any 2M page containing kernel text as read-only. Mark * other pages with read-only data as read-only and not executable. * (It is likely a small portion of the read-only data section will * be marked as read-only, but executable. This should be acceptable * since the read-only protection will keep the data from changing.) * Note that fixups to the .text section will still work until we * set CR0.WP. */ if (pa < round_2mpage(kernphys + etext - KERNSTART)) return (0); return (pg_nx); } static void create_pagetables(vm_paddr_t *firstaddr) { pd_entry_t *pd_p; pdp_entry_t *pdp_p; pml4_entry_t *p4_p; uint64_t DMPDkernphys; vm_paddr_t pax; #ifdef KASAN pt_entry_t *pt_p; uint64_t KASANPDphys, KASANPTphys, KASANphys; vm_offset_t kasankernbase; int kasankpdpi, kasankpdi, nkasanpte; #endif int i, j, ndm1g, nkpdpe, nkdmpde; /* Allocate page table pages for the direct map */ ndmpdp = howmany(ptoa(Maxmem), NBPDP); if (ndmpdp < 4) /* Minimum 4GB of dirmap */ ndmpdp = 4; ndmpdpphys = howmany(ndmpdp, NPDPEPG); if (ndmpdpphys > NDMPML4E) { /* * Each NDMPML4E allows 512 GB, so limit to that, * and then readjust ndmpdp and ndmpdpphys. */ printf("NDMPML4E limits system to %d GB\n", NDMPML4E * 512); Maxmem = atop(NDMPML4E * NBPML4); ndmpdpphys = NDMPML4E; ndmpdp = NDMPML4E * NPDEPG; } DMPDPphys = allocpages(firstaddr, ndmpdpphys); ndm1g = 0; if ((amd_feature & AMDID_PAGE1GB) != 0) { /* * Calculate the number of 1G pages that will fully fit in * Maxmem. */ ndm1g = ptoa(Maxmem) >> PDPSHIFT; /* * Allocate 2M pages for the kernel. These will be used in * place of the one or more 1G pages from ndm1g that maps * kernel memory into DMAP. */ nkdmpde = howmany((vm_offset_t)brwsection - KERNSTART + kernphys - rounddown2(kernphys, NBPDP), NBPDP); DMPDkernphys = allocpages(firstaddr, nkdmpde); } if (ndm1g < ndmpdp) DMPDphys = allocpages(firstaddr, ndmpdp - ndm1g); dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT; /* Allocate pages */ KPML4phys = allocpages(firstaddr, 1); KPDPphys = allocpages(firstaddr, NKPML4E); #ifdef KASAN KASANPDPphys = allocpages(firstaddr, NKASANPML4E); KASANPDphys = allocpages(firstaddr, 1); #endif /* * Allocate the initial number of kernel page table pages required to * bootstrap. We defer this until after all memory-size dependent * allocations are done (e.g. direct map), so that we don't have to * build in too much slop in our estimate. * * Note that when NKPML4E > 1, we have an empty page underneath * all but the KPML4I'th one, so we need NKPML4E-1 extra (zeroed) * pages. (pmap_enter requires a PD page to exist for each KPML4E.) */ nkpt_init(*firstaddr); nkpdpe = NKPDPE(nkpt); KPTphys = allocpages(firstaddr, nkpt); KPDphys = allocpages(firstaddr, nkpdpe); #ifdef KASAN nkasanpte = howmany(nkpt, KASAN_SHADOW_SCALE); KASANPTphys = allocpages(firstaddr, nkasanpte); KASANphys = allocpages(firstaddr, nkasanpte * NPTEPG); #endif /* * Connect the zero-filled PT pages to their PD entries. This * implicitly maps the PT pages at their correct locations within * the PTmap. */ pd_p = (pd_entry_t *)KPDphys; for (i = 0; i < nkpt; i++) pd_p[i] = (KPTphys + ptoa(i)) | X86_PG_RW | X86_PG_V; /* * Map from start of the kernel in physical memory (staging * area) to the end of loader preallocated memory using 2MB * pages. This replaces some of the PD entries created above. * For compatibility, identity map 2M at the start. */ pd_p[0] = X86_PG_V | PG_PS | pg_g | X86_PG_M | X86_PG_A | X86_PG_RW | pg_nx; for (i = 1, pax = kernphys; pax < KERNend; i++, pax += NBPDR) { /* Preset PG_M and PG_A because demotion expects it. */ pd_p[i] = pax | X86_PG_V | PG_PS | pg_g | X86_PG_M | X86_PG_A | bootaddr_rwx(pax); } /* * Because we map the physical blocks in 2M pages, adjust firstaddr * to record the physical blocks we've actually mapped into kernel * virtual address space. */ if (*firstaddr < round_2mpage(KERNend)) *firstaddr = round_2mpage(KERNend); /* And connect up the PD to the PDP (leaving room for L4 pages) */ pdp_p = (pdp_entry_t *)(KPDPphys + ptoa(KPML4I - KPML4BASE)); for (i = 0; i < nkpdpe; i++) pdp_p[i + KPDPI] = (KPDphys + ptoa(i)) | X86_PG_RW | X86_PG_V; #ifdef KASAN kasankernbase = kasan_md_addr_to_shad(KERNBASE); kasankpdpi = pmap_pdpe_index(kasankernbase); kasankpdi = pmap_pde_index(kasankernbase); pdp_p = (pdp_entry_t *)KASANPDPphys; pdp_p[kasankpdpi] = (KASANPDphys | X86_PG_RW | X86_PG_V | pg_nx); pd_p = (pd_entry_t *)KASANPDphys; for (i = 0; i < nkasanpte; i++) pd_p[i + kasankpdi] = (KASANPTphys + ptoa(i)) | X86_PG_RW | X86_PG_V | pg_nx; pt_p = (pt_entry_t *)KASANPTphys; for (i = 0; i < nkasanpte * NPTEPG; i++) pt_p[i] = (KASANphys + ptoa(i)) | X86_PG_RW | X86_PG_V | X86_PG_M | X86_PG_A | pg_nx; #endif /* * Now, set up the direct map region using 2MB and/or 1GB pages. If * the end of physical memory is not aligned to a 1GB page boundary, * then the residual physical memory is mapped with 2MB pages. Later, * if pmap_mapdev{_attr}() uses the direct map for non-write-back * memory, pmap_change_attr() will demote any 2MB or 1GB page mappings * that are partially used. */ pd_p = (pd_entry_t *)DMPDphys; for (i = NPDEPG * ndm1g, j = 0; i < NPDEPG * ndmpdp; i++, j++) { pd_p[j] = (vm_paddr_t)i << PDRSHIFT; /* Preset PG_M and PG_A because demotion expects it. */ pd_p[j] |= X86_PG_RW | X86_PG_V | PG_PS | pg_g | X86_PG_M | X86_PG_A | pg_nx; } pdp_p = (pdp_entry_t *)DMPDPphys; for (i = 0; i < ndm1g; i++) { pdp_p[i] = (vm_paddr_t)i << PDPSHIFT; /* Preset PG_M and PG_A because demotion expects it. */ pdp_p[i] |= X86_PG_RW | X86_PG_V | PG_PS | pg_g | X86_PG_M | X86_PG_A | pg_nx; } for (j = 0; i < ndmpdp; i++, j++) { pdp_p[i] = DMPDphys + ptoa(j); pdp_p[i] |= X86_PG_RW | X86_PG_V | pg_nx; } /* * Instead of using a 1G page for the memory containing the kernel, * use 2M pages with read-only and no-execute permissions. (If using 1G * pages, this will partially overwrite the PDPEs above.) */ if (ndm1g > 0) { pd_p = (pd_entry_t *)DMPDkernphys; for (i = 0, pax = rounddown2(kernphys, NBPDP); i < NPDEPG * nkdmpde; i++, pax += NBPDR) { pd_p[i] = pax | X86_PG_V | PG_PS | pg_g | X86_PG_M | X86_PG_A | pg_nx | bootaddr_rwx(pax); } j = rounddown2(kernphys, NBPDP) >> PDPSHIFT; for (i = 0; i < nkdmpde; i++) { pdp_p[i + j] = (DMPDkernphys + ptoa(i)) | X86_PG_RW | X86_PG_V | pg_nx; } } /* And recursively map PML4 to itself in order to get PTmap */ p4_p = (pml4_entry_t *)KPML4phys; p4_p[PML4PML4I] = KPML4phys; p4_p[PML4PML4I] |= X86_PG_RW | X86_PG_V | pg_nx; #ifdef KASAN /* Connect the KASAN shadow map slots up to the PML4. */ for (i = 0; i < NKASANPML4E; i++) { p4_p[KASANPML4I + i] = KASANPDPphys + ptoa(i); p4_p[KASANPML4I + i] |= X86_PG_RW | X86_PG_V | pg_nx; } #endif /* Connect the Direct Map slots up to the PML4. */ for (i = 0; i < ndmpdpphys; i++) { p4_p[DMPML4I + i] = DMPDPphys + ptoa(i); p4_p[DMPML4I + i] |= X86_PG_RW | X86_PG_V | pg_nx; } /* Connect the KVA slots up to the PML4 */ for (i = 0; i < NKPML4E; i++) { p4_p[KPML4BASE + i] = KPDPphys + ptoa(i); p4_p[KPML4BASE + i] |= X86_PG_RW | X86_PG_V; } kernel_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys); } /* * Bootstrap the system enough to run with virtual memory. * * On amd64 this is called after mapping has already been enabled * and just syncs the pmap module with what has already been done. * [We can't call it easily with mapping off since the kernel is not * mapped with PA == VA, hence we would have to relocate every address * from the linked base (virtual) address "KERNBASE" to the actual * (physical) address starting relative to 0] */ void pmap_bootstrap(vm_paddr_t *firstaddr) { vm_offset_t va; pt_entry_t *pte, *pcpu_pte; struct region_descriptor r_gdt; uint64_t cr4, pcpu_phys; u_long res; int i; KERNend = *firstaddr; res = atop(KERNend - (vm_paddr_t)kernphys); if (!pti) pg_g = X86_PG_G; /* * Create an initial set of page tables to run the kernel in. */ create_pagetables(firstaddr); pcpu_phys = allocpages(firstaddr, MAXCPU); /* * Add a physical memory segment (vm_phys_seg) corresponding to the * preallocated kernel page table pages so that vm_page structures * representing these pages will be created. The vm_page structures * are required for promotion of the corresponding kernel virtual * addresses to superpage mappings. */ vm_phys_early_add_seg(KPTphys, KPTphys + ptoa(nkpt)); /* * Account for the virtual addresses mapped by create_pagetables(). */ virtual_avail = (vm_offset_t)KERNSTART + round_2mpage(KERNend - (vm_paddr_t)kernphys); virtual_end = VM_MAX_KERNEL_ADDRESS; /* * Enable PG_G global pages, then switch to the kernel page * table from the bootstrap page table. After the switch, it * is possible to enable SMEP and SMAP since PG_U bits are * correct now. */ cr4 = rcr4(); cr4 |= CR4_PGE; load_cr4(cr4); load_cr3(KPML4phys); if (cpu_stdext_feature & CPUID_STDEXT_SMEP) cr4 |= CR4_SMEP; if (cpu_stdext_feature & CPUID_STDEXT_SMAP) cr4 |= CR4_SMAP; load_cr4(cr4); /* * Initialize the kernel pmap (which is statically allocated). * Count bootstrap data as being resident in case any of this data is * later unmapped (using pmap_remove()) and freed. */ PMAP_LOCK_INIT(kernel_pmap); kernel_pmap->pm_pmltop = kernel_pml4; kernel_pmap->pm_cr3 = KPML4phys; kernel_pmap->pm_ucr3 = PMAP_NO_CR3; TAILQ_INIT(&kernel_pmap->pm_pvchunk); kernel_pmap->pm_stats.resident_count = res; kernel_pmap->pm_flags = pmap_flags; /* * The kernel pmap is always active on all CPUs. Once CPUs are * enumerated, the mask will be set equal to all_cpus. */ CPU_FILL(&kernel_pmap->pm_active); /* * Initialize the TLB invalidations generation number lock. */ mtx_init(&invl_gen_mtx, "invlgn", NULL, MTX_DEF); /* * Reserve some special page table entries/VA space for temporary * mapping of pages. */ #define SYSMAP(c, p, v, n) \ v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); va = virtual_avail; pte = vtopte(va); /* * Crashdump maps. The first page is reused as CMAP1 for the * memory test. */ SYSMAP(caddr_t, CMAP1, crashdumpmap, MAXDUMPPGS) CADDR1 = crashdumpmap; SYSMAP(struct pcpu *, pcpu_pte, __pcpu, MAXCPU); virtual_avail = va; for (i = 0; i < MAXCPU; i++) { pcpu_pte[i] = (pcpu_phys + ptoa(i)) | X86_PG_V | X86_PG_RW | pg_g | pg_nx | X86_PG_M | X86_PG_A; } /* * Re-initialize PCPU area for BSP after switching. * Make hardware use gdt and common_tss from the new PCPU. */ STAILQ_INIT(&cpuhead); wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[0]); pcpu_init(&__pcpu[0], 0, sizeof(struct pcpu)); amd64_bsp_pcpu_init1(&__pcpu[0]); amd64_bsp_ist_init(&__pcpu[0]); __pcpu[0].pc_common_tss.tss_iobase = sizeof(struct amd64tss) + IOPERM_BITMAP_SIZE; memcpy(__pcpu[0].pc_gdt, temp_bsp_pcpu.pc_gdt, NGDT * sizeof(struct user_segment_descriptor)); gdt_segs[GPROC0_SEL].ssd_base = (uintptr_t)&__pcpu[0].pc_common_tss; ssdtosyssd(&gdt_segs[GPROC0_SEL], (struct system_segment_descriptor *)&__pcpu[0].pc_gdt[GPROC0_SEL]); r_gdt.rd_limit = NGDT * sizeof(struct user_segment_descriptor) - 1; r_gdt.rd_base = (long)__pcpu[0].pc_gdt; lgdt(&r_gdt); wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[0]); ltr(GSEL(GPROC0_SEL, SEL_KPL)); __pcpu[0].pc_dynamic = temp_bsp_pcpu.pc_dynamic; __pcpu[0].pc_acpi_id = temp_bsp_pcpu.pc_acpi_id; /* * Initialize the PAT MSR. * pmap_init_pat() clears and sets CR4_PGE, which, as a * side-effect, invalidates stale PG_G TLB entries that might * have been created in our pre-boot environment. */ pmap_init_pat(); /* Initialize TLB Context Id. */ if (pmap_pcid_enabled) { for (i = 0; i < MAXCPU; i++) { kernel_pmap->pm_pcids[i].pm_pcid = PMAP_PCID_KERN; kernel_pmap->pm_pcids[i].pm_gen = 1; } /* * PMAP_PCID_KERN + 1 is used for initialization of * proc0 pmap. The pmap' pcid state might be used by * EFIRT entry before first context switch, so it * needs to be valid. */ PCPU_SET(pcid_next, PMAP_PCID_KERN + 2); PCPU_SET(pcid_gen, 1); /* * pcpu area for APs is zeroed during AP startup. * pc_pcid_next and pc_pcid_gen are initialized by AP * during pcpu setup. */ load_cr4(rcr4() | CR4_PCIDE); } } /* * Setup the PAT MSR. */ void pmap_init_pat(void) { uint64_t pat_msr; u_long cr0, cr4; int i; /* Bail if this CPU doesn't implement PAT. */ if ((cpu_feature & CPUID_PAT) == 0) panic("no PAT??"); /* Set default PAT index table. */ for (i = 0; i < PAT_INDEX_SIZE; i++) pat_index[i] = -1; pat_index[PAT_WRITE_BACK] = 0; pat_index[PAT_WRITE_THROUGH] = 1; pat_index[PAT_UNCACHEABLE] = 3; pat_index[PAT_WRITE_COMBINING] = 6; pat_index[PAT_WRITE_PROTECTED] = 5; pat_index[PAT_UNCACHED] = 2; /* * Initialize default PAT entries. * Leave the indices 0-3 at the default of WB, WT, UC-, and UC. * Program 5 and 6 as WP and WC. * * Leave 4 and 7 as WB and UC. Note that a recursive page table * mapping for a 2M page uses a PAT value with the bit 3 set due * to its overload with PG_PS. */ pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | PAT_VALUE(1, PAT_WRITE_THROUGH) | PAT_VALUE(2, PAT_UNCACHED) | PAT_VALUE(3, PAT_UNCACHEABLE) | PAT_VALUE(4, PAT_WRITE_BACK) | PAT_VALUE(5, PAT_WRITE_PROTECTED) | PAT_VALUE(6, PAT_WRITE_COMBINING) | PAT_VALUE(7, PAT_UNCACHEABLE); /* Disable PGE. */ cr4 = rcr4(); load_cr4(cr4 & ~CR4_PGE); /* Disable caches (CD = 1, NW = 0). */ cr0 = rcr0(); load_cr0((cr0 & ~CR0_NW) | CR0_CD); /* Flushes caches and TLBs. */ wbinvd(); invltlb(); /* Update PAT and index table. */ wrmsr(MSR_PAT, pat_msr); /* Flush caches and TLBs again. */ wbinvd(); invltlb(); /* Restore caches and PGE. */ load_cr0(cr0); load_cr4(cr4); } vm_page_t pmap_page_alloc_below_4g(bool zeroed) { return (vm_page_alloc_noobj_contig((zeroed ? VM_ALLOC_ZERO : 0), 1, 0, (1ULL << 32), PAGE_SIZE, 0, VM_MEMATTR_DEFAULT)); } extern const char la57_trampoline[], la57_trampoline_gdt_desc[], la57_trampoline_gdt[], la57_trampoline_end[]; static void pmap_bootstrap_la57(void *arg __unused) { char *v_code; pml5_entry_t *v_pml5; pml4_entry_t *v_pml4; pdp_entry_t *v_pdp; pd_entry_t *v_pd; pt_entry_t *v_pt; vm_page_t m_code, m_pml4, m_pdp, m_pd, m_pt, m_pml5; void (*la57_tramp)(uint64_t pml5); struct region_descriptor r_gdt; if ((cpu_stdext_feature2 & CPUID_STDEXT2_LA57) == 0) return; TUNABLE_INT_FETCH("vm.pmap.la57", &la57); if (!la57) return; r_gdt.rd_limit = NGDT * sizeof(struct user_segment_descriptor) - 1; r_gdt.rd_base = (long)__pcpu[0].pc_gdt; m_code = pmap_page_alloc_below_4g(true); v_code = (char *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_code)); m_pml5 = pmap_page_alloc_below_4g(true); KPML5phys = VM_PAGE_TO_PHYS(m_pml5); v_pml5 = (pml5_entry_t *)PHYS_TO_DMAP(KPML5phys); m_pml4 = pmap_page_alloc_below_4g(true); v_pml4 = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pml4)); m_pdp = pmap_page_alloc_below_4g(true); v_pdp = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pdp)); m_pd = pmap_page_alloc_below_4g(true); v_pd = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pd)); m_pt = pmap_page_alloc_below_4g(true); v_pt = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pt)); /* * Map m_code 1:1, it appears below 4G in KVA due to physical * address being below 4G. Since kernel KVA is in upper half, * the pml4e should be zero and free for temporary use. */ kernel_pmap->pm_pmltop[pmap_pml4e_index(VM_PAGE_TO_PHYS(m_code))] = VM_PAGE_TO_PHYS(m_pdp) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; v_pdp[pmap_pdpe_index(VM_PAGE_TO_PHYS(m_code))] = VM_PAGE_TO_PHYS(m_pd) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; v_pd[pmap_pde_index(VM_PAGE_TO_PHYS(m_code))] = VM_PAGE_TO_PHYS(m_pt) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; v_pt[pmap_pte_index(VM_PAGE_TO_PHYS(m_code))] = VM_PAGE_TO_PHYS(m_code) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; /* * Add pml5 entry at top of KVA pointing to existing pml4 table, * entering all existing kernel mappings into level 5 table. */ v_pml5[pmap_pml5e_index(UPT_MAX_ADDRESS)] = KPML4phys | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M | pg_g; /* * Add pml5 entry for 1:1 trampoline mapping after LA57 is turned on. */ v_pml5[pmap_pml5e_index(VM_PAGE_TO_PHYS(m_code))] = VM_PAGE_TO_PHYS(m_pml4) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; v_pml4[pmap_pml4e_index(VM_PAGE_TO_PHYS(m_code))] = VM_PAGE_TO_PHYS(m_pdp) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; /* * Copy and call the 48->57 trampoline, hope we return there, alive. */ bcopy(la57_trampoline, v_code, la57_trampoline_end - la57_trampoline); *(u_long *)(v_code + 2 + (la57_trampoline_gdt_desc - la57_trampoline)) = la57_trampoline_gdt - la57_trampoline + VM_PAGE_TO_PHYS(m_code); la57_tramp = (void (*)(uint64_t))VM_PAGE_TO_PHYS(m_code); invlpg((vm_offset_t)la57_tramp); la57_tramp(KPML5phys); /* * gdt was necessary reset, switch back to our gdt. */ lgdt(&r_gdt); wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[0]); load_ds(_udatasel); load_es(_udatasel); load_fs(_ufssel); ssdtosyssd(&gdt_segs[GPROC0_SEL], (struct system_segment_descriptor *)&__pcpu[0].pc_gdt[GPROC0_SEL]); ltr(GSEL(GPROC0_SEL, SEL_KPL)); /* * Now unmap the trampoline, and free the pages. * Clear pml5 entry used for 1:1 trampoline mapping. */ pte_clear(&v_pml5[pmap_pml5e_index(VM_PAGE_TO_PHYS(m_code))]); invlpg((vm_offset_t)v_code); vm_page_free(m_code); vm_page_free(m_pdp); vm_page_free(m_pd); vm_page_free(m_pt); /* * Recursively map PML5 to itself in order to get PTmap and * PDmap. */ v_pml5[PML5PML5I] = KPML5phys | X86_PG_RW | X86_PG_V | pg_nx; kernel_pmap->pm_cr3 = KPML5phys; kernel_pmap->pm_pmltop = v_pml5; pmap_pt_page_count_adj(kernel_pmap, 1); } SYSINIT(la57, SI_SUB_KMEM, SI_ORDER_ANY, pmap_bootstrap_la57, NULL); /* * Initialize a vm_page's machine-dependent fields. */ void pmap_page_init(vm_page_t m) { TAILQ_INIT(&m->md.pv_list); m->md.pat_mode = PAT_WRITE_BACK; } static int pmap_allow_2m_x_ept; SYSCTL_INT(_vm_pmap, OID_AUTO, allow_2m_x_ept, CTLFLAG_RWTUN | CTLFLAG_NOFETCH, &pmap_allow_2m_x_ept, 0, "Allow executable superpage mappings in EPT"); void pmap_allow_2m_x_ept_recalculate(void) { /* * SKL002, SKL012S. Since the EPT format is only used by * Intel CPUs, the vendor check is merely a formality. */ if (!(cpu_vendor_id != CPU_VENDOR_INTEL || (cpu_ia32_arch_caps & IA32_ARCH_CAP_IF_PSCHANGE_MC_NO) != 0 || (CPUID_TO_FAMILY(cpu_id) == 0x6 && (CPUID_TO_MODEL(cpu_id) == 0x26 || /* Atoms */ CPUID_TO_MODEL(cpu_id) == 0x27 || CPUID_TO_MODEL(cpu_id) == 0x35 || CPUID_TO_MODEL(cpu_id) == 0x36 || CPUID_TO_MODEL(cpu_id) == 0x37 || CPUID_TO_MODEL(cpu_id) == 0x86 || CPUID_TO_MODEL(cpu_id) == 0x1c || CPUID_TO_MODEL(cpu_id) == 0x4a || CPUID_TO_MODEL(cpu_id) == 0x4c || CPUID_TO_MODEL(cpu_id) == 0x4d || CPUID_TO_MODEL(cpu_id) == 0x5a || CPUID_TO_MODEL(cpu_id) == 0x5c || CPUID_TO_MODEL(cpu_id) == 0x5d || CPUID_TO_MODEL(cpu_id) == 0x5f || CPUID_TO_MODEL(cpu_id) == 0x6e || CPUID_TO_MODEL(cpu_id) == 0x7a || CPUID_TO_MODEL(cpu_id) == 0x57 || /* Knights */ CPUID_TO_MODEL(cpu_id) == 0x85)))) pmap_allow_2m_x_ept = 1; TUNABLE_INT_FETCH("hw.allow_2m_x_ept", &pmap_allow_2m_x_ept); } static bool pmap_allow_2m_x_page(pmap_t pmap, bool executable) { return (pmap->pm_type != PT_EPT || !executable || !pmap_allow_2m_x_ept); } #ifdef NUMA static void pmap_init_pv_table(void) { struct pmap_large_md_page *pvd; vm_size_t s; long start, end, highest, pv_npg; int domain, i, j, pages; /* * We strongly depend on the size being a power of two, so the assert * is overzealous. However, should the struct be resized to a * different power of two, the code below needs to be revisited. */ CTASSERT((sizeof(*pvd) == 64)); /* * Calculate the size of the array. */ pmap_last_pa = vm_phys_segs[vm_phys_nsegs - 1].end; pv_npg = howmany(pmap_last_pa, NBPDR); s = (vm_size_t)pv_npg * sizeof(struct pmap_large_md_page); s = round_page(s); pv_table = (struct pmap_large_md_page *)kva_alloc(s); if (pv_table == NULL) panic("%s: kva_alloc failed\n", __func__); /* * Iterate physical segments to allocate space for respective pages. */ highest = -1; s = 0; for (i = 0; i < vm_phys_nsegs; i++) { end = vm_phys_segs[i].end / NBPDR; domain = vm_phys_segs[i].domain; if (highest >= end) continue; start = highest + 1; pvd = &pv_table[start]; pages = end - start + 1; s = round_page(pages * sizeof(*pvd)); highest = start + (s / sizeof(*pvd)) - 1; for (j = 0; j < s; j += PAGE_SIZE) { vm_page_t m = vm_page_alloc_noobj_domain(domain, 0); if (m == NULL) panic("failed to allocate PV table page"); pmap_qenter((vm_offset_t)pvd + j, &m, 1); } for (j = 0; j < s / sizeof(*pvd); j++) { rw_init_flags(&pvd->pv_lock, "pmap pv list", RW_NEW); TAILQ_INIT(&pvd->pv_page.pv_list); pvd->pv_page.pv_gen = 0; pvd->pv_page.pat_mode = 0; pvd->pv_invl_gen = 0; pvd++; } } pvd = &pv_dummy_large; rw_init_flags(&pvd->pv_lock, "pmap pv list dummy", RW_NEW); TAILQ_INIT(&pvd->pv_page.pv_list); pvd->pv_page.pv_gen = 0; pvd->pv_page.pat_mode = 0; pvd->pv_invl_gen = 0; } #else static void pmap_init_pv_table(void) { vm_size_t s; long i, pv_npg; /* * Initialize the pool of pv list locks. */ for (i = 0; i < NPV_LIST_LOCKS; i++) rw_init(&pv_list_locks[i], "pmap pv list"); /* * Calculate the size of the pv head table for superpages. */ pv_npg = howmany(vm_phys_segs[vm_phys_nsegs - 1].end, NBPDR); /* * Allocate memory for the pv head table for superpages. */ s = (vm_size_t)pv_npg * sizeof(struct md_page); s = round_page(s); pv_table = (struct md_page *)kmem_malloc(s, M_WAITOK | M_ZERO); for (i = 0; i < pv_npg; i++) TAILQ_INIT(&pv_table[i].pv_list); TAILQ_INIT(&pv_dummy.pv_list); } #endif /* * Initialize the pmap module. * Called by vm_init, to initialize any structures that the pmap * system needs to map virtual memory. */ void pmap_init(void) { struct pmap_preinit_mapping *ppim; vm_page_t m, mpte; int error, i, ret, skz63; /* L1TF, reserve page @0 unconditionally */ vm_page_blacklist_add(0, bootverbose); /* Detect bare-metal Skylake Server and Skylake-X. */ if (vm_guest == VM_GUEST_NO && cpu_vendor_id == CPU_VENDOR_INTEL && CPUID_TO_FAMILY(cpu_id) == 0x6 && CPUID_TO_MODEL(cpu_id) == 0x55) { /* * Skylake-X errata SKZ63. Processor May Hang When * Executing Code In an HLE Transaction Region between * 40000000H and 403FFFFFH. * * Mark the pages in the range as preallocated. It * seems to be impossible to distinguish between * Skylake Server and Skylake X. */ skz63 = 1; TUNABLE_INT_FETCH("hw.skz63_enable", &skz63); if (skz63 != 0) { if (bootverbose) printf("SKZ63: skipping 4M RAM starting " "at physical 1G\n"); for (i = 0; i < atop(0x400000); i++) { ret = vm_page_blacklist_add(0x40000000 + ptoa(i), FALSE); if (!ret && bootverbose) printf("page at %#lx already used\n", 0x40000000 + ptoa(i)); } } } /* IFU */ pmap_allow_2m_x_ept_recalculate(); /* * Initialize the vm page array entries for the kernel pmap's * page table pages. */ PMAP_LOCK(kernel_pmap); for (i = 0; i < nkpt; i++) { mpte = PHYS_TO_VM_PAGE(KPTphys + (i << PAGE_SHIFT)); KASSERT(mpte >= vm_page_array && mpte < &vm_page_array[vm_page_array_size], ("pmap_init: page table page is out of range")); mpte->pindex = pmap_pde_pindex(KERNBASE) + i; mpte->phys_addr = KPTphys + (i << PAGE_SHIFT); mpte->ref_count = 1; /* * Collect the page table pages that were replaced by a 2MB * page in create_pagetables(). They are zero filled. */ if ((i == 0 || kernphys + ((vm_paddr_t)(i - 1) << PDRSHIFT) < KERNend) && pmap_insert_pt_page(kernel_pmap, mpte, false)) panic("pmap_init: pmap_insert_pt_page failed"); } PMAP_UNLOCK(kernel_pmap); vm_wire_add(nkpt); /* * If the kernel is running on a virtual machine, then it must assume * that MCA is enabled by the hypervisor. Moreover, the kernel must * be prepared for the hypervisor changing the vendor and family that * are reported by CPUID. Consequently, the workaround for AMD Family * 10h Erratum 383 is enabled if the processor's feature set does not * include at least one feature that is only supported by older Intel * or newer AMD processors. */ if (vm_guest != VM_GUEST_NO && (cpu_feature & CPUID_SS) == 0 && (cpu_feature2 & (CPUID2_SSSE3 | CPUID2_SSE41 | CPUID2_AESNI | CPUID2_AVX | CPUID2_XSAVE)) == 0 && (amd_feature2 & (AMDID2_XOP | AMDID2_FMA4)) == 0) workaround_erratum383 = 1; /* * Are large page mappings enabled? */ TUNABLE_INT_FETCH("vm.pmap.pg_ps_enabled", &pg_ps_enabled); if (pg_ps_enabled) { KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0, ("pmap_init: can't assign to pagesizes[1]")); pagesizes[1] = NBPDR; if ((amd_feature & AMDID_PAGE1GB) != 0) { KASSERT(MAXPAGESIZES > 2 && pagesizes[2] == 0, ("pmap_init: can't assign to pagesizes[2]")); pagesizes[2] = NBPDP; } } /* * Initialize pv chunk lists. */ for (i = 0; i < PMAP_MEMDOM; i++) { mtx_init(&pv_chunks[i].pvc_lock, "pmap pv chunk list", NULL, MTX_DEF); TAILQ_INIT(&pv_chunks[i].pvc_list); } pmap_init_pv_table(); pmap_initialized = 1; for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == 0) continue; /* Make the direct map consistent */ if (ppim->pa < dmaplimit && ppim->pa + ppim->sz <= dmaplimit) { (void)pmap_change_attr(PHYS_TO_DMAP(ppim->pa), ppim->sz, ppim->mode); } if (!bootverbose) continue; printf("PPIM %u: PA=%#lx, VA=%#lx, size=%#lx, mode=%#x\n", i, ppim->pa, ppim->va, ppim->sz, ppim->mode); } mtx_init(&qframe_mtx, "qfrmlk", NULL, MTX_SPIN); error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK, (vmem_addr_t *)&qframe); if (error != 0) panic("qframe allocation failed"); lm_ents = 8; TUNABLE_INT_FETCH("vm.pmap.large_map_pml4_entries", &lm_ents); if (lm_ents > LMEPML4I - LMSPML4I + 1) lm_ents = LMEPML4I - LMSPML4I + 1; if (bootverbose) printf("pmap: large map %u PML4 slots (%lu GB)\n", lm_ents, (u_long)lm_ents * (NBPML4 / 1024 / 1024 / 1024)); if (lm_ents != 0) { large_vmem = vmem_create("large", LARGEMAP_MIN_ADDRESS, (vmem_size_t)lm_ents * NBPML4, PAGE_SIZE, 0, M_WAITOK); if (large_vmem == NULL) { printf("pmap: cannot create large map\n"); lm_ents = 0; } for (i = 0; i < lm_ents; i++) { m = pmap_large_map_getptp_unlocked(); /* XXXKIB la57 */ kernel_pml4[LMSPML4I + i] = X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M | pg_nx | VM_PAGE_TO_PHYS(m); } } } SYSCTL_UINT(_vm_pmap, OID_AUTO, large_map_pml4_entries, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &lm_ents, 0, "Maximum number of PML4 entries for use by large map (tunable). " "Each entry corresponds to 512GB of address space."); static SYSCTL_NODE(_vm_pmap, OID_AUTO, pde, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "2MB page mapping counters"); static COUNTER_U64_DEFINE_EARLY(pmap_pde_demotions); SYSCTL_COUNTER_U64(_vm_pmap_pde, OID_AUTO, demotions, CTLFLAG_RD, &pmap_pde_demotions, "2MB page demotions"); static COUNTER_U64_DEFINE_EARLY(pmap_pde_mappings); SYSCTL_COUNTER_U64(_vm_pmap_pde, OID_AUTO, mappings, CTLFLAG_RD, &pmap_pde_mappings, "2MB page mappings"); static COUNTER_U64_DEFINE_EARLY(pmap_pde_p_failures); SYSCTL_COUNTER_U64(_vm_pmap_pde, OID_AUTO, p_failures, CTLFLAG_RD, &pmap_pde_p_failures, "2MB page promotion failures"); static COUNTER_U64_DEFINE_EARLY(pmap_pde_promotions); SYSCTL_COUNTER_U64(_vm_pmap_pde, OID_AUTO, promotions, CTLFLAG_RD, &pmap_pde_promotions, "2MB page promotions"); static SYSCTL_NODE(_vm_pmap, OID_AUTO, pdpe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "1GB page mapping counters"); static COUNTER_U64_DEFINE_EARLY(pmap_pdpe_demotions); SYSCTL_COUNTER_U64(_vm_pmap_pdpe, OID_AUTO, demotions, CTLFLAG_RD, &pmap_pdpe_demotions, "1GB page demotions"); /*************************************************** * Low level helper routines..... ***************************************************/ static pt_entry_t pmap_swap_pat(pmap_t pmap, pt_entry_t entry) { int x86_pat_bits = X86_PG_PTE_PAT | X86_PG_PDE_PAT; switch (pmap->pm_type) { case PT_X86: case PT_RVI: /* Verify that both PAT bits are not set at the same time */ KASSERT((entry & x86_pat_bits) != x86_pat_bits, ("Invalid PAT bits in entry %#lx", entry)); /* Swap the PAT bits if one of them is set */ if ((entry & x86_pat_bits) != 0) entry ^= x86_pat_bits; break; case PT_EPT: /* * Nothing to do - the memory attributes are represented * the same way for regular pages and superpages. */ break; default: panic("pmap_switch_pat_bits: bad pm_type %d", pmap->pm_type); } return (entry); } boolean_t pmap_is_valid_memattr(pmap_t pmap __unused, vm_memattr_t mode) { return (mode >= 0 && mode < PAT_INDEX_SIZE && pat_index[(int)mode] >= 0); } /* * Determine the appropriate bits to set in a PTE or PDE for a specified * caching mode. */ int pmap_cache_bits(pmap_t pmap, int mode, boolean_t is_pde) { int cache_bits, pat_flag, pat_idx; if (!pmap_is_valid_memattr(pmap, mode)) panic("Unknown caching mode %d\n", mode); switch (pmap->pm_type) { case PT_X86: case PT_RVI: /* The PAT bit is different for PTE's and PDE's. */ pat_flag = is_pde ? X86_PG_PDE_PAT : X86_PG_PTE_PAT; /* Map the caching mode to a PAT index. */ pat_idx = pat_index[mode]; /* Map the 3-bit index value into the PAT, PCD, and PWT bits. */ cache_bits = 0; if (pat_idx & 0x4) cache_bits |= pat_flag; if (pat_idx & 0x2) cache_bits |= PG_NC_PCD; if (pat_idx & 0x1) cache_bits |= PG_NC_PWT; break; case PT_EPT: cache_bits = EPT_PG_IGNORE_PAT | EPT_PG_MEMORY_TYPE(mode); break; default: panic("unsupported pmap type %d", pmap->pm_type); } return (cache_bits); } static int pmap_cache_mask(pmap_t pmap, boolean_t is_pde) { int mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = is_pde ? X86_PG_PDE_CACHE : X86_PG_PTE_CACHE; break; case PT_EPT: mask = EPT_PG_IGNORE_PAT | EPT_PG_MEMORY_TYPE(0x7); break; default: panic("pmap_cache_mask: invalid pm_type %d", pmap->pm_type); } return (mask); } static int pmap_pat_index(pmap_t pmap, pt_entry_t pte, bool is_pde) { int pat_flag, pat_idx; pat_idx = 0; switch (pmap->pm_type) { case PT_X86: case PT_RVI: /* The PAT bit is different for PTE's and PDE's. */ pat_flag = is_pde ? X86_PG_PDE_PAT : X86_PG_PTE_PAT; if ((pte & pat_flag) != 0) pat_idx |= 0x4; if ((pte & PG_NC_PCD) != 0) pat_idx |= 0x2; if ((pte & PG_NC_PWT) != 0) pat_idx |= 0x1; break; case PT_EPT: if ((pte & EPT_PG_IGNORE_PAT) != 0) panic("EPT PTE %#lx has no PAT memory type", pte); pat_idx = (pte & EPT_PG_MEMORY_TYPE(0x7)) >> 3; break; } /* See pmap_init_pat(). */ if (pat_idx == 4) pat_idx = 0; if (pat_idx == 7) pat_idx = 3; return (pat_idx); } bool pmap_ps_enabled(pmap_t pmap) { return (pg_ps_enabled && (pmap->pm_flags & PMAP_PDE_SUPERPAGE) != 0); } static void pmap_update_pde_store(pmap_t pmap, pd_entry_t *pde, pd_entry_t newpde) { switch (pmap->pm_type) { case PT_X86: break; case PT_RVI: case PT_EPT: /* * XXX * This is a little bogus since the generation number is * supposed to be bumped up when a region of the address * space is invalidated in the page tables. * * In this case the old PDE entry is valid but yet we want * to make sure that any mappings using the old entry are * invalidated in the TLB. * * The reason this works as expected is because we rendezvous * "all" host cpus and force any vcpu context to exit as a * side-effect. */ atomic_add_long(&pmap->pm_eptgen, 1); break; default: panic("pmap_update_pde_store: bad pm_type %d", pmap->pm_type); } pde_store(pde, newpde); } /* * After changing the page size for the specified virtual address in the page * table, flush the corresponding entries from the processor's TLB. Only the * calling processor's TLB is affected. * * The calling thread must be pinned to a processor. */ static void pmap_update_pde_invalidate(pmap_t pmap, vm_offset_t va, pd_entry_t newpde) { pt_entry_t PG_G; if (pmap_type_guest(pmap)) return; KASSERT(pmap->pm_type == PT_X86, ("pmap_update_pde_invalidate: invalid type %d", pmap->pm_type)); PG_G = pmap_global_bit(pmap); if ((newpde & PG_PS) == 0) /* Demotion: flush a specific 2MB page mapping. */ invlpg(va); else if ((newpde & PG_G) == 0) /* * Promotion: flush every 4KB page mapping from the TLB * because there are too many to flush individually. */ invltlb(); else { /* * Promotion: flush every 4KB page mapping from the TLB, * including any global (PG_G) mappings. */ invltlb_glob(); } } /* * The amd64 pmap uses different approaches to TLB invalidation * depending on the kernel configuration, available hardware features, * and known hardware errata. The kernel configuration option that * has the greatest operational impact on TLB invalidation is PTI, * which is enabled automatically on affected Intel CPUs. The most * impactful hardware features are first PCID, and then INVPCID * instruction presence. PCID usage is quite different for PTI * vs. non-PTI. * * * Kernel Page Table Isolation (PTI or KPTI) is used to mitigate * the Meltdown bug in some Intel CPUs. Under PTI, each user address * space is served by two page tables, user and kernel. The user * page table only maps user space and a kernel trampoline. The * kernel trampoline includes the entirety of the kernel text but * only the kernel data that is needed to switch from user to kernel * mode. The kernel page table maps the user and kernel address * spaces in their entirety. It is identical to the per-process * page table used in non-PTI mode. * * User page tables are only used when the CPU is in user mode. * Consequently, some TLB invalidations can be postponed until the * switch from kernel to user mode. In contrast, the user * space part of the kernel page table is used for copyout(9), so * TLB invalidations on this page table cannot be similarly postponed. * * The existence of a user mode page table for the given pmap is * indicated by a pm_ucr3 value that differs from PMAP_NO_CR3, in * which case pm_ucr3 contains the %cr3 register value for the user * mode page table's root. * * * The pm_active bitmask indicates which CPUs currently have the * pmap active. A CPU's bit is set on context switch to the pmap, and * cleared on switching off this CPU. For the kernel page table, * the pm_active field is immutable and contains all CPUs. The * kernel page table is always logically active on every processor, * but not necessarily in use by the hardware, e.g., in PTI mode. * * When requesting invalidation of virtual addresses with * pmap_invalidate_XXX() functions, the pmap sends shootdown IPIs to * all CPUs recorded as active in pm_active. Updates to and reads * from pm_active are not synchronized, and so they may race with * each other. Shootdown handlers are prepared to handle the race. * * * PCID is an optional feature of the long mode x86 MMU where TLB * entries are tagged with the 'Process ID' of the address space * they belong to. This feature provides a limited namespace for * process identifiers, 12 bits, supporting 4095 simultaneous IDs * total. * * Allocation of a PCID to a pmap is done by an algorithm described * in section 15.12, "Other TLB Consistency Algorithms", of * Vahalia's book "Unix Internals". A PCID cannot be allocated for * the whole lifetime of a pmap in pmap_pinit() due to the limited * namespace. Instead, a per-CPU, per-pmap PCID is assigned when * the CPU is about to start caching TLB entries from a pmap, * i.e., on the context switch that activates the pmap on the CPU. * * The PCID allocator maintains a per-CPU, per-pmap generation * count, pm_gen, which is incremented each time a new PCID is * allocated. On TLB invalidation, the generation counters for the * pmap are zeroed, which signals the context switch code that the * previously allocated PCID is no longer valid. Effectively, * zeroing any of these counters triggers a TLB shootdown for the * given CPU/address space, due to the allocation of a new PCID. * * Zeroing can be performed remotely. Consequently, if a pmap is * inactive on a CPU, then a TLB shootdown for that pmap and CPU can * be initiated by an ordinary memory access to reset the target * CPU's generation count within the pmap. The CPU initiating the * TLB shootdown does not need to send an IPI to the target CPU. * * * PTI + PCID. The available PCIDs are divided into two sets: PCIDs * for complete (kernel) page tables, and PCIDs for user mode page * tables. A user PCID value is obtained from the kernel PCID value * by setting the highest bit, 11, to 1 (0x800 == PMAP_PCID_USER_PT). * * User space page tables are activated on return to user mode, by * loading pm_ucr3 into %cr3. If the PCPU(ucr3_load_mask) requests * clearing bit 63 of the loaded ucr3, this effectively causes * complete invalidation of the user mode TLB entries for the * current pmap. In which case, local invalidations of individual * pages in the user page table are skipped. * * * Local invalidation, all modes. If the requested invalidation is * for a specific address or the total invalidation of a currently * active pmap, then the TLB is flushed using INVLPG for a kernel * page table, and INVPCID(INVPCID_CTXGLOB)/invltlb_glob() for a * user space page table(s). * * If the INVPCID instruction is available, it is used to flush entries * from the kernel page table. * * * mode: PTI disabled, PCID present. The kernel reserves PCID 0 for its * address space, all other 4095 PCIDs are used for user mode spaces * as described above. A context switch allocates a new PCID if * the recorded PCID is zero or the recorded generation does not match * the CPU's generation, effectively flushing the TLB for this address space. * Total remote invalidation is performed by zeroing pm_gen for all CPUs. * local user page: INVLPG * local kernel page: INVLPG * local user total: INVPCID(CTX) * local kernel total: INVPCID(CTXGLOB) or invltlb_glob() * remote user page, inactive pmap: zero pm_gen * remote user page, active pmap: zero pm_gen + IPI:INVLPG * (Both actions are required to handle the aforementioned pm_active races.) * remote kernel page: IPI:INVLPG * remote user total, inactive pmap: zero pm_gen * remote user total, active pmap: zero pm_gen + IPI:(INVPCID(CTX) or * reload %cr3) * (See note above about pm_active races.) * remote kernel total: IPI:(INVPCID(CTXGLOB) or invltlb_glob()) * * PTI enabled, PCID present. * local user page: INVLPG for kpt, INVPCID(ADDR) or (INVLPG for ucr3) * for upt * local kernel page: INVLPG * local user total: INVPCID(CTX) or reload %cr3 for kpt, clear PCID_SAVE * on loading UCR3 into %cr3 for upt * local kernel total: INVPCID(CTXGLOB) or invltlb_glob() * remote user page, inactive pmap: zero pm_gen * remote user page, active pmap: zero pm_gen + IPI:(INVLPG for kpt, * INVPCID(ADDR) for upt) * remote kernel page: IPI:INVLPG * remote user total, inactive pmap: zero pm_gen * remote user total, active pmap: zero pm_gen + IPI:(INVPCID(CTX) for kpt, * clear PCID_SAVE on loading UCR3 into $cr3 for upt) * remote kernel total: IPI:(INVPCID(CTXGLOB) or invltlb_glob()) * * No PCID. * local user page: INVLPG * local kernel page: INVLPG * local user total: reload %cr3 * local kernel total: invltlb_glob() * remote user page, inactive pmap: - * remote user page, active pmap: IPI:INVLPG * remote kernel page: IPI:INVLPG * remote user total, inactive pmap: - * remote user total, active pmap: IPI:(reload %cr3) * remote kernel total: IPI:invltlb_glob() * Since on return to user mode, the reload of %cr3 with ucr3 causes * TLB invalidation, no specific action is required for user page table. * * EPT. EPT pmaps do not map KVA, all mappings are userspace. * XXX TODO */ #ifdef SMP /* * Interrupt the cpus that are executing in the guest context. * This will force the vcpu to exit and the cached EPT mappings * will be invalidated by the host before the next vmresume. */ static __inline void pmap_invalidate_ept(pmap_t pmap) { smr_seq_t goal; int ipinum; sched_pin(); KASSERT(!CPU_ISSET(curcpu, &pmap->pm_active), ("pmap_invalidate_ept: absurd pm_active")); /* * The TLB mappings associated with a vcpu context are not * flushed each time a different vcpu is chosen to execute. * * This is in contrast with a process's vtop mappings that * are flushed from the TLB on each context switch. * * Therefore we need to do more than just a TLB shootdown on * the active cpus in 'pmap->pm_active'. To do this we keep * track of the number of invalidations performed on this pmap. * * Each vcpu keeps a cache of this counter and compares it * just before a vmresume. If the counter is out-of-date an * invept will be done to flush stale mappings from the TLB. * * To ensure that all vCPU threads have observed the new counter * value before returning, we use SMR. Ordering is important here: * the VMM enters an SMR read section before loading the counter * and after updating the pm_active bit set. Thus, pm_active is * a superset of active readers, and any reader that has observed * the goal has observed the new counter value. */ atomic_add_long(&pmap->pm_eptgen, 1); goal = smr_advance(pmap->pm_eptsmr); /* * Force the vcpu to exit and trap back into the hypervisor. */ ipinum = pmap->pm_flags & PMAP_NESTED_IPIMASK; ipi_selected(pmap->pm_active, ipinum); sched_unpin(); /* * Ensure that all active vCPUs will observe the new generation counter * value before executing any more guest instructions. */ smr_wait(pmap->pm_eptsmr, goal); } static inline void pmap_invalidate_preipi_pcid(pmap_t pmap) { u_int cpuid, i; sched_pin(); cpuid = PCPU_GET(cpuid); if (pmap != PCPU_GET(curpmap)) cpuid = 0xffffffff; /* An impossible value */ CPU_FOREACH(i) { if (cpuid != i) pmap->pm_pcids[i].pm_gen = 0; } /* * The fence is between stores to pm_gen and the read of the * pm_active mask. We need to ensure that it is impossible * for us to miss the bit update in pm_active and * simultaneously observe a non-zero pm_gen in * pmap_activate_sw(), otherwise TLB update is missed. * Without the fence, IA32 allows such an outcome. Note that * pm_active is updated by a locked operation, which provides * the reciprocal fence. */ atomic_thread_fence_seq_cst(); } static void pmap_invalidate_preipi_nopcid(pmap_t pmap __unused) { sched_pin(); } DEFINE_IFUNC(static, void, pmap_invalidate_preipi, (pmap_t)) { return (pmap_pcid_enabled ? pmap_invalidate_preipi_pcid : pmap_invalidate_preipi_nopcid); } static inline void pmap_invalidate_page_pcid_cb(pmap_t pmap, vm_offset_t va, const bool invpcid_works1) { struct invpcid_descr d; uint64_t kcr3, ucr3; uint32_t pcid; u_int cpuid; /* * Because pm_pcid is recalculated on a context switch, we * must ensure there is no preemption, not just pinning. * Otherwise, we might use a stale value below. */ CRITICAL_ASSERT(curthread); /* * No need to do anything with user page tables invalidation * if there is no user page table, or invalidation is deferred * until the return to userspace. ucr3_load_mask is stable * because we have preemption disabled. */ if (pmap->pm_ucr3 == PMAP_NO_CR3 || PCPU_GET(ucr3_load_mask) != PMAP_UCR3_NOMASK) return; cpuid = PCPU_GET(cpuid); pcid = pmap->pm_pcids[cpuid].pm_pcid; if (invpcid_works1) { d.pcid = pcid | PMAP_PCID_USER_PT; d.pad = 0; d.addr = va; invpcid(&d, INVPCID_ADDR); } else { kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE; ucr3 = pmap->pm_ucr3 | pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE; pmap_pti_pcid_invlpg(ucr3, kcr3, va); } } static void pmap_invalidate_page_pcid_invpcid_cb(pmap_t pmap, vm_offset_t va) { pmap_invalidate_page_pcid_cb(pmap, va, true); } static void pmap_invalidate_page_pcid_noinvpcid_cb(pmap_t pmap, vm_offset_t va) { pmap_invalidate_page_pcid_cb(pmap, va, false); } static void pmap_invalidate_page_nopcid_cb(pmap_t pmap __unused, vm_offset_t va __unused) { } DEFINE_IFUNC(static, void, pmap_invalidate_page_cb, (pmap_t, vm_offset_t)) { if (pmap_pcid_enabled) return (invpcid_works ? pmap_invalidate_page_pcid_invpcid_cb : pmap_invalidate_page_pcid_noinvpcid_cb); return (pmap_invalidate_page_nopcid_cb); } static void pmap_invalidate_page_curcpu_cb(pmap_t pmap, vm_offset_t va, vm_offset_t addr2 __unused) { if (pmap == kernel_pmap) { invlpg(va); } else if (pmap == PCPU_GET(curpmap)) { invlpg(va); pmap_invalidate_page_cb(pmap, va); } } void pmap_invalidate_page(pmap_t pmap, vm_offset_t va) { if (pmap_type_guest(pmap)) { pmap_invalidate_ept(pmap); return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_page: invalid type %d", pmap->pm_type)); pmap_invalidate_preipi(pmap); smp_masked_invlpg(va, pmap, pmap_invalidate_page_curcpu_cb); } /* 4k PTEs -- Chosen to exceed the total size of Broadwell L2 TLB */ #define PMAP_INVLPG_THRESHOLD (4 * 1024 * PAGE_SIZE) static void pmap_invalidate_range_pcid_cb(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, const bool invpcid_works1) { struct invpcid_descr d; uint64_t kcr3, ucr3; uint32_t pcid; u_int cpuid; CRITICAL_ASSERT(curthread); if (pmap != PCPU_GET(curpmap) || pmap->pm_ucr3 == PMAP_NO_CR3 || PCPU_GET(ucr3_load_mask) != PMAP_UCR3_NOMASK) return; cpuid = PCPU_GET(cpuid); pcid = pmap->pm_pcids[cpuid].pm_pcid; if (invpcid_works1) { d.pcid = pcid | PMAP_PCID_USER_PT; d.pad = 0; for (d.addr = sva; d.addr < eva; d.addr += PAGE_SIZE) invpcid(&d, INVPCID_ADDR); } else { kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE; ucr3 = pmap->pm_ucr3 | pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE; pmap_pti_pcid_invlrng(ucr3, kcr3, sva, eva); } } static void pmap_invalidate_range_pcid_invpcid_cb(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { pmap_invalidate_range_pcid_cb(pmap, sva, eva, true); } static void pmap_invalidate_range_pcid_noinvpcid_cb(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { pmap_invalidate_range_pcid_cb(pmap, sva, eva, false); } static void pmap_invalidate_range_nopcid_cb(pmap_t pmap __unused, vm_offset_t sva __unused, vm_offset_t eva __unused) { } DEFINE_IFUNC(static, void, pmap_invalidate_range_cb, (pmap_t, vm_offset_t, vm_offset_t)) { if (pmap_pcid_enabled) return (invpcid_works ? pmap_invalidate_range_pcid_invpcid_cb : pmap_invalidate_range_pcid_noinvpcid_cb); return (pmap_invalidate_range_nopcid_cb); } static void pmap_invalidate_range_curcpu_cb(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t addr; if (pmap == kernel_pmap) { for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); } else if (pmap == PCPU_GET(curpmap)) { for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); pmap_invalidate_range_cb(pmap, sva, eva); } } void pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { if (eva - sva >= PMAP_INVLPG_THRESHOLD) { pmap_invalidate_all(pmap); return; } if (pmap_type_guest(pmap)) { pmap_invalidate_ept(pmap); return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_range: invalid type %d", pmap->pm_type)); pmap_invalidate_preipi(pmap); smp_masked_invlpg_range(sva, eva, pmap, pmap_invalidate_range_curcpu_cb); } static inline void pmap_invalidate_all_pcid_cb(pmap_t pmap, bool invpcid_works1) { struct invpcid_descr d; uint64_t kcr3; uint32_t pcid; u_int cpuid; if (pmap == kernel_pmap) { if (invpcid_works1) { bzero(&d, sizeof(d)); invpcid(&d, INVPCID_CTXGLOB); } else { invltlb_glob(); } } else if (pmap == PCPU_GET(curpmap)) { CRITICAL_ASSERT(curthread); cpuid = PCPU_GET(cpuid); pcid = pmap->pm_pcids[cpuid].pm_pcid; if (invpcid_works1) { d.pcid = pcid; d.pad = 0; d.addr = 0; invpcid(&d, INVPCID_CTX); } else { kcr3 = pmap->pm_cr3 | pcid; load_cr3(kcr3); } if (pmap->pm_ucr3 != PMAP_NO_CR3) PCPU_SET(ucr3_load_mask, ~CR3_PCID_SAVE); } } static void pmap_invalidate_all_pcid_invpcid_cb(pmap_t pmap) { pmap_invalidate_all_pcid_cb(pmap, true); } static void pmap_invalidate_all_pcid_noinvpcid_cb(pmap_t pmap) { pmap_invalidate_all_pcid_cb(pmap, false); } static void pmap_invalidate_all_nopcid_cb(pmap_t pmap) { if (pmap == kernel_pmap) invltlb_glob(); else if (pmap == PCPU_GET(curpmap)) invltlb(); } DEFINE_IFUNC(static, void, pmap_invalidate_all_cb, (pmap_t)) { if (pmap_pcid_enabled) return (invpcid_works ? pmap_invalidate_all_pcid_invpcid_cb : pmap_invalidate_all_pcid_noinvpcid_cb); return (pmap_invalidate_all_nopcid_cb); } static void pmap_invalidate_all_curcpu_cb(pmap_t pmap, vm_offset_t addr1 __unused, vm_offset_t addr2 __unused) { pmap_invalidate_all_cb(pmap); } void pmap_invalidate_all(pmap_t pmap) { if (pmap_type_guest(pmap)) { pmap_invalidate_ept(pmap); return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_all: invalid type %d", pmap->pm_type)); pmap_invalidate_preipi(pmap); smp_masked_invltlb(pmap, pmap_invalidate_all_curcpu_cb); } static void pmap_invalidate_cache_curcpu_cb(pmap_t pmap __unused, vm_offset_t va __unused, vm_offset_t addr2 __unused) { wbinvd(); } void pmap_invalidate_cache(void) { sched_pin(); smp_cache_flush(pmap_invalidate_cache_curcpu_cb); } struct pde_action { cpuset_t invalidate; /* processors that invalidate their TLB */ pmap_t pmap; vm_offset_t va; pd_entry_t *pde; pd_entry_t newpde; u_int store; /* processor that updates the PDE */ }; static void pmap_update_pde_action(void *arg) { struct pde_action *act = arg; if (act->store == PCPU_GET(cpuid)) pmap_update_pde_store(act->pmap, act->pde, act->newpde); } static void pmap_update_pde_teardown(void *arg) { struct pde_action *act = arg; if (CPU_ISSET(PCPU_GET(cpuid), &act->invalidate)) pmap_update_pde_invalidate(act->pmap, act->va, act->newpde); } /* * Change the page size for the specified virtual address in a way that * prevents any possibility of the TLB ever having two entries that map the * same virtual address using different page sizes. This is the recommended * workaround for Erratum 383 on AMD Family 10h processors. It prevents a * machine check exception for a TLB state that is improperly diagnosed as a * hardware error. */ static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde) { struct pde_action act; cpuset_t active, other_cpus; u_int cpuid; sched_pin(); cpuid = PCPU_GET(cpuid); other_cpus = all_cpus; CPU_CLR(cpuid, &other_cpus); if (pmap == kernel_pmap || pmap_type_guest(pmap)) active = all_cpus; else { active = pmap->pm_active; } if (CPU_OVERLAP(&active, &other_cpus)) { act.store = cpuid; act.invalidate = active; act.va = va; act.pmap = pmap; act.pde = pde; act.newpde = newpde; CPU_SET(cpuid, &active); smp_rendezvous_cpus(active, smp_no_rendezvous_barrier, pmap_update_pde_action, pmap_update_pde_teardown, &act); } else { pmap_update_pde_store(pmap, pde, newpde); if (CPU_ISSET(cpuid, &active)) pmap_update_pde_invalidate(pmap, va, newpde); } sched_unpin(); } #else /* !SMP */ /* * Normal, non-SMP, invalidation functions. */ void pmap_invalidate_page(pmap_t pmap, vm_offset_t va) { struct invpcid_descr d; uint64_t kcr3, ucr3; uint32_t pcid; if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) { pmap->pm_eptgen++; return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_range: unknown type %d", pmap->pm_type)); if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) { invlpg(va); if (pmap == PCPU_GET(curpmap) && pmap_pcid_enabled && pmap->pm_ucr3 != PMAP_NO_CR3) { critical_enter(); pcid = pmap->pm_pcids[0].pm_pcid; if (invpcid_works) { d.pcid = pcid | PMAP_PCID_USER_PT; d.pad = 0; d.addr = va; invpcid(&d, INVPCID_ADDR); } else { kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE; ucr3 = pmap->pm_ucr3 | pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE; pmap_pti_pcid_invlpg(ucr3, kcr3, va); } critical_exit(); } } else if (pmap_pcid_enabled) pmap->pm_pcids[0].pm_gen = 0; } void pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { struct invpcid_descr d; vm_offset_t addr; uint64_t kcr3, ucr3; if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) { pmap->pm_eptgen++; return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_range: unknown type %d", pmap->pm_type)); if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) { for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); if (pmap == PCPU_GET(curpmap) && pmap_pcid_enabled && pmap->pm_ucr3 != PMAP_NO_CR3) { critical_enter(); if (invpcid_works) { d.pcid = pmap->pm_pcids[0].pm_pcid | PMAP_PCID_USER_PT; d.pad = 0; d.addr = sva; for (; d.addr < eva; d.addr += PAGE_SIZE) invpcid(&d, INVPCID_ADDR); } else { kcr3 = pmap->pm_cr3 | pmap->pm_pcids[0]. pm_pcid | CR3_PCID_SAVE; ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[0]. pm_pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE; pmap_pti_pcid_invlrng(ucr3, kcr3, sva, eva); } critical_exit(); } } else if (pmap_pcid_enabled) { pmap->pm_pcids[0].pm_gen = 0; } } void pmap_invalidate_all(pmap_t pmap) { struct invpcid_descr d; uint64_t kcr3, ucr3; if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) { pmap->pm_eptgen++; return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_all: unknown type %d", pmap->pm_type)); if (pmap == kernel_pmap) { if (pmap_pcid_enabled && invpcid_works) { bzero(&d, sizeof(d)); invpcid(&d, INVPCID_CTXGLOB); } else { invltlb_glob(); } } else if (pmap == PCPU_GET(curpmap)) { if (pmap_pcid_enabled) { critical_enter(); if (invpcid_works) { d.pcid = pmap->pm_pcids[0].pm_pcid; d.pad = 0; d.addr = 0; invpcid(&d, INVPCID_CTX); if (pmap->pm_ucr3 != PMAP_NO_CR3) { d.pcid |= PMAP_PCID_USER_PT; invpcid(&d, INVPCID_CTX); } } else { kcr3 = pmap->pm_cr3 | pmap->pm_pcids[0].pm_pcid; if (pmap->pm_ucr3 != PMAP_NO_CR3) { ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[ 0].pm_pcid | PMAP_PCID_USER_PT; pmap_pti_pcid_invalidate(ucr3, kcr3); } else load_cr3(kcr3); } critical_exit(); } else { invltlb(); } } else if (pmap_pcid_enabled) { pmap->pm_pcids[0].pm_gen = 0; } } PMAP_INLINE void pmap_invalidate_cache(void) { wbinvd(); } static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde) { pmap_update_pde_store(pmap, pde, newpde); if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) pmap_update_pde_invalidate(pmap, va, newpde); else pmap->pm_pcids[0].pm_gen = 0; } #endif /* !SMP */ static void pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, pd_entry_t pde) { /* * When the PDE has PG_PROMOTED set, the 2MB page mapping was created * by a promotion that did not invalidate the 512 4KB page mappings * that might exist in the TLB. Consequently, at this point, the TLB * may hold both 4KB and 2MB page mappings for the address range [va, * va + NBPDR). Therefore, the entire range must be invalidated here. * In contrast, when PG_PROMOTED is clear, the TLB will not hold any * 4KB page mappings for the address range [va, va + NBPDR), and so a * single INVLPG suffices to invalidate the 2MB page mapping from the * TLB. */ if ((pde & PG_PROMOTED) != 0) pmap_invalidate_range(pmap, va, va + NBPDR - 1); else pmap_invalidate_page(pmap, va); } DEFINE_IFUNC(, void, pmap_invalidate_cache_range, (vm_offset_t sva, vm_offset_t eva)) { if ((cpu_feature & CPUID_SS) != 0) return (pmap_invalidate_cache_range_selfsnoop); if ((cpu_feature & CPUID_CLFSH) != 0) return (pmap_force_invalidate_cache_range); return (pmap_invalidate_cache_range_all); } #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024) static void pmap_invalidate_cache_range_check_align(vm_offset_t sva, vm_offset_t eva) { KASSERT((sva & PAGE_MASK) == 0, ("pmap_invalidate_cache_range: sva not page-aligned")); KASSERT((eva & PAGE_MASK) == 0, ("pmap_invalidate_cache_range: eva not page-aligned")); } static void pmap_invalidate_cache_range_selfsnoop(vm_offset_t sva, vm_offset_t eva) { pmap_invalidate_cache_range_check_align(sva, eva); } void pmap_force_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva) { sva &= ~(vm_offset_t)(cpu_clflush_line_size - 1); /* * XXX: Some CPUs fault, hang, or trash the local APIC * registers if we use CLFLUSH on the local APIC range. The * local APIC is always uncached, so we don't need to flush * for that range anyway. */ if (pmap_kextract(sva) == lapic_paddr) return; if ((cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0) { /* * Do per-cache line flush. Use a locked * instruction to insure that previous stores are * included in the write-back. The processor * propagates flush to other processors in the cache * coherence domain. */ atomic_thread_fence_seq_cst(); for (; sva < eva; sva += cpu_clflush_line_size) clflushopt(sva); atomic_thread_fence_seq_cst(); } else { /* * Writes are ordered by CLFLUSH on Intel CPUs. */ if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); for (; sva < eva; sva += cpu_clflush_line_size) clflush(sva); if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); } } static void pmap_invalidate_cache_range_all(vm_offset_t sva, vm_offset_t eva) { pmap_invalidate_cache_range_check_align(sva, eva); pmap_invalidate_cache(); } /* * Remove the specified set of pages from the data and instruction caches. * * In contrast to pmap_invalidate_cache_range(), this function does not * rely on the CPU's self-snoop feature, because it is intended for use * when moving pages into a different cache domain. */ void pmap_invalidate_cache_pages(vm_page_t *pages, int count) { vm_offset_t daddr, eva; int i; bool useclflushopt; useclflushopt = (cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0; if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE || ((cpu_feature & CPUID_CLFSH) == 0 && !useclflushopt)) pmap_invalidate_cache(); else { if (useclflushopt) atomic_thread_fence_seq_cst(); else if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); for (i = 0; i < count; i++) { daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i])); eva = daddr + PAGE_SIZE; for (; daddr < eva; daddr += cpu_clflush_line_size) { if (useclflushopt) clflushopt(daddr); else clflush(daddr); } } if (useclflushopt) atomic_thread_fence_seq_cst(); else if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); } } void pmap_flush_cache_range(vm_offset_t sva, vm_offset_t eva) { pmap_invalidate_cache_range_check_align(sva, eva); if ((cpu_stdext_feature & CPUID_STDEXT_CLWB) == 0) { pmap_force_invalidate_cache_range(sva, eva); return; } /* See comment in pmap_force_invalidate_cache_range(). */ if (pmap_kextract(sva) == lapic_paddr) return; atomic_thread_fence_seq_cst(); for (; sva < eva; sva += cpu_clflush_line_size) clwb(sva); atomic_thread_fence_seq_cst(); } void pmap_flush_cache_phys_range(vm_paddr_t spa, vm_paddr_t epa, vm_memattr_t mattr) { pt_entry_t *pte; vm_offset_t vaddr; int error, pte_bits; KASSERT((spa & PAGE_MASK) == 0, ("pmap_flush_cache_phys_range: spa not page-aligned")); KASSERT((epa & PAGE_MASK) == 0, ("pmap_flush_cache_phys_range: epa not page-aligned")); if (spa < dmaplimit) { pmap_flush_cache_range(PHYS_TO_DMAP(spa), PHYS_TO_DMAP(MIN( dmaplimit, epa))); if (dmaplimit >= epa) return; spa = dmaplimit; } pte_bits = pmap_cache_bits(kernel_pmap, mattr, 0) | X86_PG_RW | X86_PG_V; error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK, &vaddr); KASSERT(error == 0, ("vmem_alloc failed: %d", error)); pte = vtopte(vaddr); for (; spa < epa; spa += PAGE_SIZE) { sched_pin(); pte_store(pte, spa | pte_bits); invlpg(vaddr); /* XXXKIB atomic inside flush_cache_range are excessive */ pmap_flush_cache_range(vaddr, vaddr + PAGE_SIZE); sched_unpin(); } vmem_free(kernel_arena, vaddr, PAGE_SIZE); } /* * Routine: pmap_extract * Function: * Extract the physical page address associated * with the given map/virtual_address pair. */ vm_paddr_t pmap_extract(pmap_t pmap, vm_offset_t va) { pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte, PG_V; vm_paddr_t pa; pa = 0; PG_V = pmap_valid_bit(pmap); PMAP_LOCK(pmap); pdpe = pmap_pdpe(pmap, va); if (pdpe != NULL && (*pdpe & PG_V) != 0) { if ((*pdpe & PG_PS) != 0) pa = (*pdpe & PG_PS_FRAME) | (va & PDPMASK); else { pde = pmap_pdpe_to_pde(pdpe, va); if ((*pde & PG_V) != 0) { if ((*pde & PG_PS) != 0) { pa = (*pde & PG_PS_FRAME) | (va & PDRMASK); } else { pte = pmap_pde_to_pte(pde, va); pa = (*pte & PG_FRAME) | (va & PAGE_MASK); } } } } PMAP_UNLOCK(pmap); return (pa); } /* * Routine: pmap_extract_and_hold * Function: * Atomically extract and hold the physical page * with the given pmap and virtual address pair * if that mapping permits the given protection. */ vm_page_t pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) { pdp_entry_t pdpe, *pdpep; pd_entry_t pde, *pdep; pt_entry_t pte, PG_RW, PG_V; vm_page_t m; m = NULL; PG_RW = pmap_rw_bit(pmap); PG_V = pmap_valid_bit(pmap); PMAP_LOCK(pmap); pdpep = pmap_pdpe(pmap, va); if (pdpep == NULL || ((pdpe = *pdpep) & PG_V) == 0) goto out; if ((pdpe & PG_PS) != 0) { if ((pdpe & PG_RW) == 0 && (prot & VM_PROT_WRITE) != 0) goto out; m = PHYS_TO_VM_PAGE((pdpe & PG_PS_FRAME) | (va & PDPMASK)); goto check_page; } pdep = pmap_pdpe_to_pde(pdpep, va); if (pdep == NULL || ((pde = *pdep) & PG_V) == 0) goto out; if ((pde & PG_PS) != 0) { if ((pde & PG_RW) == 0 && (prot & VM_PROT_WRITE) != 0) goto out; m = PHYS_TO_VM_PAGE((pde & PG_PS_FRAME) | (va & PDRMASK)); goto check_page; } pte = *pmap_pde_to_pte(pdep, va); if ((pte & PG_V) == 0 || ((pte & PG_RW) == 0 && (prot & VM_PROT_WRITE) != 0)) goto out; m = PHYS_TO_VM_PAGE(pte & PG_FRAME); check_page: if (m != NULL && !vm_page_wire_mapped(m)) m = NULL; out: PMAP_UNLOCK(pmap); return (m); } vm_paddr_t pmap_kextract(vm_offset_t va) { pd_entry_t pde; vm_paddr_t pa; if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) { pa = DMAP_TO_PHYS(va); } else if (PMAP_ADDRESS_IN_LARGEMAP(va)) { pa = pmap_large_map_kextract(va); } else { pde = *vtopde(va); if (pde & PG_PS) { pa = (pde & PG_PS_FRAME) | (va & PDRMASK); } else { /* * Beware of a concurrent promotion that changes the * PDE at this point! For example, vtopte() must not * be used to access the PTE because it would use the * new PDE. It is, however, safe to use the old PDE * because the page table page is preserved by the * promotion. */ pa = *pmap_pde_to_pte(&pde, va); pa = (pa & PG_FRAME) | (va & PAGE_MASK); } } return (pa); } /*************************************************** * Low level mapping routines..... ***************************************************/ /* * Add a wired page to the kva. * Note: not SMP coherent. */ PMAP_INLINE void pmap_kenter(vm_offset_t va, vm_paddr_t pa) { pt_entry_t *pte; pte = vtopte(va); pte_store(pte, pa | X86_PG_RW | X86_PG_V | pg_g | pg_nx); } static __inline void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode) { pt_entry_t *pte; int cache_bits; pte = vtopte(va); cache_bits = pmap_cache_bits(kernel_pmap, mode, 0); pte_store(pte, pa | X86_PG_RW | X86_PG_V | pg_g | pg_nx | cache_bits); } /* * Remove a page from the kernel pagetables. * Note: not SMP coherent. */ PMAP_INLINE void pmap_kremove(vm_offset_t va) { pt_entry_t *pte; pte = vtopte(va); pte_clear(pte); } /* * Used to map a range of physical addresses into kernel * virtual address space. * * The value passed in '*virt' is a suggested virtual address for * the mapping. Architectures which can support a direct-mapped * physical to virtual region can return the appropriate address * within that region, leaving '*virt' unchanged. Other * architectures should map the pages starting at '*virt' and * update '*virt' with the first usable address after the mapped * region. */ vm_offset_t pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) { return PHYS_TO_DMAP(start); } /* * Add a list of wired pages to the kva * this routine is only used for temporary * kernel mappings that do not need to have * page modification or references recorded. * Note that old mappings are simply written * over. The page *must* be wired. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count) { pt_entry_t *endpte, oldpte, pa, *pte; vm_page_t m; int cache_bits; oldpte = 0; pte = vtopte(sva); endpte = pte + count; while (pte < endpte) { m = *ma++; cache_bits = pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0); pa = VM_PAGE_TO_PHYS(m) | cache_bits; if ((*pte & (PG_FRAME | X86_PG_PTE_CACHE)) != pa) { oldpte |= *pte; pte_store(pte, pa | pg_g | pg_nx | X86_PG_RW | X86_PG_V); } pte++; } if (__predict_false((oldpte & X86_PG_V) != 0)) pmap_invalidate_range(kernel_pmap, sva, sva + count * PAGE_SIZE); } /* * This routine tears out page mappings from the * kernel -- it is meant only for temporary mappings. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qremove(vm_offset_t sva, int count) { vm_offset_t va; va = sva; while (count-- > 0) { KASSERT(va >= VM_MIN_KERNEL_ADDRESS, ("usermode va %lx", va)); pmap_kremove(va); va += PAGE_SIZE; } pmap_invalidate_range(kernel_pmap, sva, va); } /*************************************************** * Page table page management routines..... ***************************************************/ /* * Schedule the specified unused page table page to be freed. Specifically, * add the page to the specified list of pages that will be released to the * physical memory manager after the TLB has been updated. */ static __inline void pmap_add_delayed_free_list(vm_page_t m, struct spglist *free, boolean_t set_PG_ZERO) { if (set_PG_ZERO) m->flags |= PG_ZERO; else m->flags &= ~PG_ZERO; SLIST_INSERT_HEAD(free, m, plinks.s.ss); } /* * Inserts the specified page table page into the specified pmap's collection * of idle page table pages. Each of a pmap's page table pages is responsible * for mapping a distinct range of virtual addresses. The pmap's collection is * ordered by this virtual address range. * * If "promoted" is false, then the page table page "mpte" must be zero filled. */ static __inline int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); mpte->valid = promoted ? VM_PAGE_BITS_ALL : 0; return (vm_radix_insert(&pmap->pm_root, mpte)); } /* * Removes the page table page mapping the specified virtual address from the * specified pmap's collection of idle page table pages, and returns it. * Otherwise, returns NULL if there is no page table page corresponding to the * specified virtual address. */ static __inline vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); return (vm_radix_remove(&pmap->pm_root, pmap_pde_pindex(va))); } /* * Decrements a page table page's reference count, which is used to record the * number of valid page table entries within the page. If the reference count * drops to zero, then the page table page is unmapped. Returns TRUE if the * page table page was unmapped and FALSE otherwise. */ static inline boolean_t pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { --m->ref_count; if (m->ref_count == 0) { _pmap_unwire_ptp(pmap, va, m, free); return (TRUE); } else return (FALSE); } static void _pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { pml5_entry_t *pml5; pml4_entry_t *pml4; pdp_entry_t *pdp; pd_entry_t *pd; vm_page_t pdpg, pdppg, pml4pg; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * unmap the page table page */ if (m->pindex >= NUPDE + NUPDPE + NUPML4E) { /* PML4 page */ MPASS(pmap_is_la57(pmap)); pml5 = pmap_pml5e(pmap, va); *pml5 = 0; if (pmap->pm_pmltopu != NULL && va <= VM_MAXUSER_ADDRESS) { pml5 = pmap_pml5e_u(pmap, va); *pml5 = 0; } } else if (m->pindex >= NUPDE + NUPDPE) { /* PDP page */ pml4 = pmap_pml4e(pmap, va); *pml4 = 0; if (!pmap_is_la57(pmap) && pmap->pm_pmltopu != NULL && va <= VM_MAXUSER_ADDRESS) { pml4 = pmap_pml4e_u(pmap, va); *pml4 = 0; } } else if (m->pindex >= NUPDE) { /* PD page */ pdp = pmap_pdpe(pmap, va); *pdp = 0; } else { /* PTE page */ pd = pmap_pde(pmap, va); *pd = 0; } if (m->pindex < NUPDE) { /* We just released a PT, unhold the matching PD */ pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME); pmap_unwire_ptp(pmap, va, pdpg, free); } else if (m->pindex < NUPDE + NUPDPE) { /* We just released a PD, unhold the matching PDP */ pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME); pmap_unwire_ptp(pmap, va, pdppg, free); } else if (m->pindex < NUPDE + NUPDPE + NUPML4E && pmap_is_la57(pmap)) { /* We just released a PDP, unhold the matching PML4 */ pml4pg = PHYS_TO_VM_PAGE(*pmap_pml5e(pmap, va) & PG_FRAME); pmap_unwire_ptp(pmap, va, pml4pg, free); } pmap_pt_page_count_adj(pmap, -1); /* * Put page on a list so that it is released after * *ALL* TLB shootdown is done */ pmap_add_delayed_free_list(m, free, TRUE); } /* * After removing a page table entry, this routine is used to * conditionally free the page, and manage the reference count. */ static int pmap_unuse_pt(pmap_t pmap, vm_offset_t va, pd_entry_t ptepde, struct spglist *free) { vm_page_t mpte; if (va >= VM_MAXUSER_ADDRESS) return (0); KASSERT(ptepde != 0, ("pmap_unuse_pt: ptepde != 0")); mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); return (pmap_unwire_ptp(pmap, va, mpte, free)); } /* * Release a page table page reference after a failed attempt to create a * mapping. */ static void pmap_abort_ptp(pmap_t pmap, vm_offset_t va, vm_page_t mpte) { struct spglist free; SLIST_INIT(&free); if (pmap_unwire_ptp(pmap, va, mpte, &free)) { /* * Although "va" was never mapped, paging-structure caches * could nonetheless have entries that refer to the freed * page table pages. Invalidate those entries. */ pmap_invalidate_page(pmap, va); vm_page_free_pages_toq(&free, true); } } void pmap_pinit0(pmap_t pmap) { struct proc *p; struct thread *td; int i; PMAP_LOCK_INIT(pmap); pmap->pm_pmltop = kernel_pmap->pm_pmltop; pmap->pm_pmltopu = NULL; pmap->pm_cr3 = kernel_pmap->pm_cr3; /* hack to keep pmap_pti_pcid_invalidate() alive */ pmap->pm_ucr3 = PMAP_NO_CR3; vm_radix_init(&pmap->pm_root); CPU_ZERO(&pmap->pm_active); TAILQ_INIT(&pmap->pm_pvchunk); bzero(&pmap->pm_stats, sizeof pmap->pm_stats); pmap->pm_flags = pmap_flags; CPU_FOREACH(i) { pmap->pm_pcids[i].pm_pcid = PMAP_PCID_KERN + 1; pmap->pm_pcids[i].pm_gen = 1; } pmap_activate_boot(pmap); td = curthread; if (pti) { p = td->td_proc; PROC_LOCK(p); p->p_md.md_flags |= P_MD_KPTI; PROC_UNLOCK(p); } pmap_thread_init_invl_gen(td); if ((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) { pmap_pkru_ranges_zone = uma_zcreate("pkru ranges", sizeof(struct pmap_pkru_range), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); } } void pmap_pinit_pml4(vm_page_t pml4pg) { pml4_entry_t *pm_pml4; int i; pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml4pg)); /* Wire in kernel global address entries. */ for (i = 0; i < NKPML4E; i++) { pm_pml4[KPML4BASE + i] = (KPDPphys + ptoa(i)) | X86_PG_RW | X86_PG_V; } #ifdef KASAN for (i = 0; i < NKASANPML4E; i++) { pm_pml4[KASANPML4I + i] = (KASANPDPphys + ptoa(i)) | X86_PG_RW | X86_PG_V | pg_nx; } #endif for (i = 0; i < ndmpdpphys; i++) { pm_pml4[DMPML4I + i] = (DMPDPphys + ptoa(i)) | X86_PG_RW | X86_PG_V; } /* install self-referential address mapping entry(s) */ pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pml4pg) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; /* install large map entries if configured */ for (i = 0; i < lm_ents; i++) pm_pml4[LMSPML4I + i] = kernel_pmap->pm_pmltop[LMSPML4I + i]; } void pmap_pinit_pml5(vm_page_t pml5pg) { pml5_entry_t *pm_pml5; pm_pml5 = (pml5_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml5pg)); /* * Add pml5 entry at top of KVA pointing to existing pml4 table, * entering all existing kernel mappings into level 5 table. */ pm_pml5[pmap_pml5e_index(UPT_MAX_ADDRESS)] = KPML4phys | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M | pg_g | pmap_cache_bits(kernel_pmap, VM_MEMATTR_DEFAULT, FALSE); /* * Install self-referential address mapping entry. */ pm_pml5[PML5PML5I] = VM_PAGE_TO_PHYS(pml5pg) | X86_PG_RW | X86_PG_V | X86_PG_M | X86_PG_A | pmap_cache_bits(kernel_pmap, VM_MEMATTR_DEFAULT, FALSE); } static void pmap_pinit_pml4_pti(vm_page_t pml4pgu) { pml4_entry_t *pm_pml4u; int i; pm_pml4u = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml4pgu)); for (i = 0; i < NPML4EPG; i++) pm_pml4u[i] = pti_pml4[i]; } static void pmap_pinit_pml5_pti(vm_page_t pml5pgu) { pml5_entry_t *pm_pml5u; pm_pml5u = (pml5_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml5pgu)); pagezero(pm_pml5u); /* * Add pml5 entry at top of KVA pointing to existing pml4 pti * table, entering all kernel mappings needed for usermode * into level 5 table. */ pm_pml5u[pmap_pml5e_index(UPT_MAX_ADDRESS)] = pmap_kextract((vm_offset_t)pti_pml4) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M | pg_g | pmap_cache_bits(kernel_pmap, VM_MEMATTR_DEFAULT, FALSE); } /* Allocate a page table page and do related bookkeeping */ static vm_page_t pmap_alloc_pt_page(pmap_t pmap, vm_pindex_t pindex, int flags) { vm_page_t m; m = vm_page_alloc_noobj(flags); if (__predict_false(m == NULL)) return (NULL); m->pindex = pindex; pmap_pt_page_count_adj(pmap, 1); return (m); } static void pmap_free_pt_page(pmap_t pmap, vm_page_t m, bool zerofilled) { /* * This function assumes the page will need to be unwired, * even though the counterpart allocation in pmap_alloc_pt_page() * doesn't enforce VM_ALLOC_WIRED. However, all current uses * of pmap_free_pt_page() require unwiring. The case in which * a PT page doesn't require unwiring because its ref_count has * naturally reached 0 is handled through _pmap_unwire_ptp(). */ vm_page_unwire_noq(m); if (zerofilled) vm_page_free_zero(m); else vm_page_free(m); pmap_pt_page_count_adj(pmap, -1); } /* * Initialize a preallocated and zeroed pmap structure, * such as one in a vmspace structure. */ int pmap_pinit_type(pmap_t pmap, enum pmap_type pm_type, int flags) { vm_page_t pmltop_pg, pmltop_pgu; vm_paddr_t pmltop_phys; int i; bzero(&pmap->pm_stats, sizeof pmap->pm_stats); /* * Allocate the page directory page. Pass NULL instead of a * pointer to the pmap here to avoid calling * pmap_resident_count_adj() through pmap_pt_page_count_adj(), * since that requires pmap lock. Instead do the accounting * manually. * * Note that final call to pmap_remove() optimization that * checks for zero resident_count is basically disabled by * accounting for top-level page. But the optimization was * not effective since we started using non-managed mapping of * the shared page. */ pmltop_pg = pmap_alloc_pt_page(NULL, 0, VM_ALLOC_WIRED | VM_ALLOC_ZERO | VM_ALLOC_WAITOK); pmap_pt_page_count_pinit(pmap, 1); pmltop_phys = VM_PAGE_TO_PHYS(pmltop_pg); pmap->pm_pmltop = (pml5_entry_t *)PHYS_TO_DMAP(pmltop_phys); CPU_FOREACH(i) { pmap->pm_pcids[i].pm_pcid = PMAP_PCID_NONE; pmap->pm_pcids[i].pm_gen = 0; } pmap->pm_cr3 = PMAP_NO_CR3; /* initialize to an invalid value */ pmap->pm_ucr3 = PMAP_NO_CR3; pmap->pm_pmltopu = NULL; pmap->pm_type = pm_type; /* * Do not install the host kernel mappings in the nested page * tables. These mappings are meaningless in the guest physical * address space. * Install minimal kernel mappings in PTI case. */ switch (pm_type) { case PT_X86: pmap->pm_cr3 = pmltop_phys; if (pmap_is_la57(pmap)) pmap_pinit_pml5(pmltop_pg); else pmap_pinit_pml4(pmltop_pg); if ((curproc->p_md.md_flags & P_MD_KPTI) != 0) { /* * As with pmltop_pg, pass NULL instead of a * pointer to the pmap to ensure that the PTI * page counted explicitly. */ pmltop_pgu = pmap_alloc_pt_page(NULL, 0, VM_ALLOC_WIRED | VM_ALLOC_WAITOK); pmap_pt_page_count_pinit(pmap, 1); pmap->pm_pmltopu = (pml4_entry_t *)PHYS_TO_DMAP( VM_PAGE_TO_PHYS(pmltop_pgu)); if (pmap_is_la57(pmap)) pmap_pinit_pml5_pti(pmltop_pgu); else pmap_pinit_pml4_pti(pmltop_pgu); pmap->pm_ucr3 = VM_PAGE_TO_PHYS(pmltop_pgu); } if ((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) { rangeset_init(&pmap->pm_pkru, pkru_dup_range, pkru_free_range, pmap, M_NOWAIT); } break; case PT_EPT: case PT_RVI: pmap->pm_eptsmr = smr_create("pmap", 0, 0); break; } vm_radix_init(&pmap->pm_root); CPU_ZERO(&pmap->pm_active); TAILQ_INIT(&pmap->pm_pvchunk); pmap->pm_flags = flags; pmap->pm_eptgen = 0; return (1); } int pmap_pinit(pmap_t pmap) { return (pmap_pinit_type(pmap, PT_X86, pmap_flags)); } static void pmap_allocpte_free_unref(pmap_t pmap, vm_offset_t va, pt_entry_t *pte) { vm_page_t mpg; struct spglist free; mpg = PHYS_TO_VM_PAGE(*pte & PG_FRAME); if (mpg->ref_count != 0) return; SLIST_INIT(&free); _pmap_unwire_ptp(pmap, va, mpg, &free); pmap_invalidate_page(pmap, va); vm_page_free_pages_toq(&free, true); } static pml4_entry_t * pmap_allocpte_getpml4(pmap_t pmap, struct rwlock **lockp, vm_offset_t va, bool addref) { vm_pindex_t pml5index; pml5_entry_t *pml5; pml4_entry_t *pml4; vm_page_t pml4pg; pt_entry_t PG_V; bool allocated; if (!pmap_is_la57(pmap)) return (&pmap->pm_pmltop[pmap_pml4e_index(va)]); PG_V = pmap_valid_bit(pmap); pml5index = pmap_pml5e_index(va); pml5 = &pmap->pm_pmltop[pml5index]; if ((*pml5 & PG_V) == 0) { if (pmap_allocpte_nosleep(pmap, pmap_pml5e_pindex(va), lockp, va) == NULL) return (NULL); allocated = true; } else { allocated = false; } pml4 = (pml4_entry_t *)PHYS_TO_DMAP(*pml5 & PG_FRAME); pml4 = &pml4[pmap_pml4e_index(va)]; if ((*pml4 & PG_V) == 0) { pml4pg = PHYS_TO_VM_PAGE(*pml5 & PG_FRAME); if (allocated && !addref) pml4pg->ref_count--; else if (!allocated && addref) pml4pg->ref_count++; } return (pml4); } static pdp_entry_t * pmap_allocpte_getpdp(pmap_t pmap, struct rwlock **lockp, vm_offset_t va, bool addref) { vm_page_t pdppg; pml4_entry_t *pml4; pdp_entry_t *pdp; pt_entry_t PG_V; bool allocated; PG_V = pmap_valid_bit(pmap); pml4 = pmap_allocpte_getpml4(pmap, lockp, va, false); if (pml4 == NULL) return (NULL); if ((*pml4 & PG_V) == 0) { /* Have to allocate a new pdp, recurse */ if (pmap_allocpte_nosleep(pmap, pmap_pml4e_pindex(va), lockp, va) == NULL) { if (pmap_is_la57(pmap)) pmap_allocpte_free_unref(pmap, va, pmap_pml5e(pmap, va)); return (NULL); } allocated = true; } else { allocated = false; } pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME); pdp = &pdp[pmap_pdpe_index(va)]; if ((*pdp & PG_V) == 0) { pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME); if (allocated && !addref) pdppg->ref_count--; else if (!allocated && addref) pdppg->ref_count++; } return (pdp); } /* * The ptepindexes, i.e. page indices, of the page table pages encountered * while translating virtual address va are defined as follows: * - for the page table page (last level), * ptepindex = pmap_pde_pindex(va) = va >> PDRSHIFT, * in other words, it is just the index of the PDE that maps the page * table page. * - for the page directory page, * ptepindex = NUPDE (number of userland PD entries) + * (pmap_pde_index(va) >> NPDEPGSHIFT) * i.e. index of PDPE is put after the last index of PDE, * - for the page directory pointer page, * ptepindex = NUPDE + NUPDPE + (pmap_pde_index(va) >> (NPDEPGSHIFT + * NPML4EPGSHIFT), * i.e. index of pml4e is put after the last index of PDPE, * - for the PML4 page (if LA57 mode is enabled), * ptepindex = NUPDE + NUPDPE + NUPML4E + (pmap_pde_index(va) >> * (NPDEPGSHIFT + NPML4EPGSHIFT + NPML5EPGSHIFT), * i.e. index of pml5e is put after the last index of PML4E. * * Define an order on the paging entries, where all entries of the * same height are put together, then heights are put from deepest to * root. Then ptexpindex is the sequential number of the * corresponding paging entry in this order. * * The values of NUPDE, NUPDPE, and NUPML4E are determined by the size of * LA57 paging structures even in LA48 paging mode. Moreover, the * ptepindexes are calculated as if the paging structures were 5-level * regardless of the actual mode of operation. * * The root page at PML4/PML5 does not participate in this indexing scheme, * since it is statically allocated by pmap_pinit() and not by pmap_allocpte(). */ static vm_page_t pmap_allocpte_nosleep(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp, vm_offset_t va) { vm_pindex_t pml5index, pml4index; pml5_entry_t *pml5, *pml5u; pml4_entry_t *pml4, *pml4u; pdp_entry_t *pdp; pd_entry_t *pd; vm_page_t m, pdpg; pt_entry_t PG_A, PG_M, PG_RW, PG_V; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); /* * Allocate a page table page. */ m = pmap_alloc_pt_page(pmap, ptepindex, VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (m == NULL) return (NULL); /* * Map the pagetable page into the process address space, if * it isn't already there. */ if (ptepindex >= NUPDE + NUPDPE + NUPML4E) { MPASS(pmap_is_la57(pmap)); pml5index = pmap_pml5e_index(va); pml5 = &pmap->pm_pmltop[pml5index]; KASSERT((*pml5 & PG_V) == 0, ("pmap %p va %#lx pml5 %#lx", pmap, va, *pml5)); *pml5 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; if (pmap->pm_pmltopu != NULL && pml5index < NUPML5E) { if (pmap->pm_ucr3 != PMAP_NO_CR3) *pml5 |= pg_nx; pml5u = &pmap->pm_pmltopu[pml5index]; *pml5u = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; } } else if (ptepindex >= NUPDE + NUPDPE) { pml4index = pmap_pml4e_index(va); /* Wire up a new PDPE page */ pml4 = pmap_allocpte_getpml4(pmap, lockp, va, true); if (pml4 == NULL) { pmap_free_pt_page(pmap, m, true); return (NULL); } KASSERT((*pml4 & PG_V) == 0, ("pmap %p va %#lx pml4 %#lx", pmap, va, *pml4)); *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; if (!pmap_is_la57(pmap) && pmap->pm_pmltopu != NULL && pml4index < NUPML4E) { /* * PTI: Make all user-space mappings in the * kernel-mode page table no-execute so that * we detect any programming errors that leave * the kernel-mode page table active on return * to user space. */ if (pmap->pm_ucr3 != PMAP_NO_CR3) *pml4 |= pg_nx; pml4u = &pmap->pm_pmltopu[pml4index]; *pml4u = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; } } else if (ptepindex >= NUPDE) { /* Wire up a new PDE page */ pdp = pmap_allocpte_getpdp(pmap, lockp, va, true); if (pdp == NULL) { pmap_free_pt_page(pmap, m, true); return (NULL); } KASSERT((*pdp & PG_V) == 0, ("pmap %p va %#lx pdp %#lx", pmap, va, *pdp)); *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; } else { /* Wire up a new PTE page */ pdp = pmap_allocpte_getpdp(pmap, lockp, va, false); if (pdp == NULL) { pmap_free_pt_page(pmap, m, true); return (NULL); } if ((*pdp & PG_V) == 0) { /* Have to allocate a new pd, recurse */ if (pmap_allocpte_nosleep(pmap, pmap_pdpe_pindex(va), lockp, va) == NULL) { pmap_allocpte_free_unref(pmap, va, pmap_pml4e(pmap, va)); pmap_free_pt_page(pmap, m, true); return (NULL); } } else { /* Add reference to the pd page */ pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME); pdpg->ref_count++; } pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME); /* Now we know where the page directory page is */ pd = &pd[pmap_pde_index(va)]; KASSERT((*pd & PG_V) == 0, ("pmap %p va %#lx pd %#lx", pmap, va, *pd)); *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; } return (m); } /* * This routine is called if the desired page table page does not exist. * * If page table page allocation fails, this routine may sleep before * returning NULL. It sleeps only if a lock pointer was given. Sleep * occurs right before returning to the caller. This way, we never * drop pmap lock to sleep while a page table page has ref_count == 0, * which prevents the page from being freed under us. */ static vm_page_t pmap_allocpte_alloc(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp, vm_offset_t va) { vm_page_t m; m = pmap_allocpte_nosleep(pmap, ptepindex, lockp, va); if (m == NULL && lockp != NULL) { RELEASE_PV_LIST_LOCK(lockp); PMAP_UNLOCK(pmap); PMAP_ASSERT_NOT_IN_DI(); vm_wait(NULL); PMAP_LOCK(pmap); } return (m); } static pd_entry_t * pmap_alloc_pde(pmap_t pmap, vm_offset_t va, vm_page_t *pdpgp, struct rwlock **lockp) { pdp_entry_t *pdpe, PG_V; pd_entry_t *pde; vm_page_t pdpg; vm_pindex_t pdpindex; PG_V = pmap_valid_bit(pmap); retry: pdpe = pmap_pdpe(pmap, va); if (pdpe != NULL && (*pdpe & PG_V) != 0) { pde = pmap_pdpe_to_pde(pdpe, va); if (va < VM_MAXUSER_ADDRESS) { /* Add a reference to the pd page. */ pdpg = PHYS_TO_VM_PAGE(*pdpe & PG_FRAME); pdpg->ref_count++; } else pdpg = NULL; } else if (va < VM_MAXUSER_ADDRESS) { /* Allocate a pd page. */ pdpindex = pmap_pde_pindex(va) >> NPDPEPGSHIFT; pdpg = pmap_allocpte_alloc(pmap, NUPDE + pdpindex, lockp, va); if (pdpg == NULL) { if (lockp != NULL) goto retry; else return (NULL); } pde = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pdpg)); pde = &pde[pmap_pde_index(va)]; } else panic("pmap_alloc_pde: missing page table page for va %#lx", va); *pdpgp = pdpg; return (pde); } static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, struct rwlock **lockp) { vm_pindex_t ptepindex; pd_entry_t *pd, PG_V; vm_page_t m; PG_V = pmap_valid_bit(pmap); /* * Calculate pagetable page index */ ptepindex = pmap_pde_pindex(va); retry: /* * Get the page directory entry */ pd = pmap_pde(pmap, va); /* * This supports switching from a 2MB page to a * normal 4K page. */ if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) { if (!pmap_demote_pde_locked(pmap, pd, va, lockp)) { /* * Invalidation of the 2MB page mapping may have caused * the deallocation of the underlying PD page. */ pd = NULL; } } /* * If the page table page is mapped, we just increment the * hold count, and activate it. */ if (pd != NULL && (*pd & PG_V) != 0) { m = PHYS_TO_VM_PAGE(*pd & PG_FRAME); m->ref_count++; } else { /* * Here if the pte page isn't mapped, or if it has been * deallocated. */ m = pmap_allocpte_alloc(pmap, ptepindex, lockp, va); if (m == NULL && lockp != NULL) goto retry; } return (m); } /*************************************************** * Pmap allocation/deallocation routines. ***************************************************/ /* * Release any resources held by the given physical map. * Called when a pmap initialized by pmap_pinit is being released. * Should only be called if the map contains no valid mappings. */ void pmap_release(pmap_t pmap) { vm_page_t m; int i; KASSERT(vm_radix_is_empty(&pmap->pm_root), ("pmap_release: pmap %p has reserved page table page(s)", pmap)); KASSERT(CPU_EMPTY(&pmap->pm_active), ("releasing active pmap %p", pmap)); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pmap->pm_pmltop)); if (pmap_is_la57(pmap)) { pmap->pm_pmltop[pmap_pml5e_index(UPT_MAX_ADDRESS)] = 0; pmap->pm_pmltop[PML5PML5I] = 0; } else { for (i = 0; i < NKPML4E; i++) /* KVA */ pmap->pm_pmltop[KPML4BASE + i] = 0; #ifdef KASAN for (i = 0; i < NKASANPML4E; i++) /* KASAN shadow map */ pmap->pm_pmltop[KASANPML4I + i] = 0; #endif for (i = 0; i < ndmpdpphys; i++)/* Direct Map */ pmap->pm_pmltop[DMPML4I + i] = 0; pmap->pm_pmltop[PML4PML4I] = 0; /* Recursive Mapping */ for (i = 0; i < lm_ents; i++) /* Large Map */ pmap->pm_pmltop[LMSPML4I + i] = 0; } pmap_free_pt_page(NULL, m, true); pmap_pt_page_count_pinit(pmap, -1); if (pmap->pm_pmltopu != NULL) { m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pmap-> pm_pmltopu)); pmap_free_pt_page(NULL, m, false); pmap_pt_page_count_pinit(pmap, -1); } if (pmap->pm_type == PT_X86 && (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) rangeset_fini(&pmap->pm_pkru); KASSERT(pmap->pm_stats.resident_count == 0, ("pmap_release: pmap %p resident count %ld != 0", pmap, pmap->pm_stats.resident_count)); } static int kvm_size(SYSCTL_HANDLER_ARGS) { unsigned long ksize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS; return sysctl_handle_long(oidp, &ksize, 0, req); } SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, kvm_size, "LU", "Size of KVM"); static int kvm_free(SYSCTL_HANDLER_ARGS) { unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; return sysctl_handle_long(oidp, &kfree, 0, req); } SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, kvm_free, "LU", "Amount of KVM free"); /* * Allocate physical memory for the vm_page array and map it into KVA, * attempting to back the vm_pages with domain-local memory. */ void pmap_page_array_startup(long pages) { pdp_entry_t *pdpe; pd_entry_t *pde, newpdir; vm_offset_t va, start, end; vm_paddr_t pa; long pfn; int domain, i; vm_page_array_size = pages; start = VM_MIN_KERNEL_ADDRESS; end = start + pages * sizeof(struct vm_page); for (va = start; va < end; va += NBPDR) { pfn = first_page + (va - start) / sizeof(struct vm_page); domain = vm_phys_domain(ptoa(pfn)); pdpe = pmap_pdpe(kernel_pmap, va); if ((*pdpe & X86_PG_V) == 0) { pa = vm_phys_early_alloc(domain, PAGE_SIZE); dump_add_page(pa); pagezero((void *)PHYS_TO_DMAP(pa)); *pdpe = (pdp_entry_t)(pa | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M); } pde = pmap_pdpe_to_pde(pdpe, va); if ((*pde & X86_PG_V) != 0) panic("Unexpected pde"); pa = vm_phys_early_alloc(domain, NBPDR); for (i = 0; i < NPDEPG; i++) dump_add_page(pa + i * PAGE_SIZE); newpdir = (pd_entry_t)(pa | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M | PG_PS | pg_g | pg_nx); pde_store(pde, newpdir); } vm_page_array = (vm_page_t)start; } /* * grow the number of kernel page table entries, if needed */ void pmap_growkernel(vm_offset_t addr) { vm_paddr_t paddr; vm_page_t nkpg; pd_entry_t *pde, newpdir; pdp_entry_t *pdpe; mtx_assert(&kernel_map->system_mtx, MA_OWNED); /* * Return if "addr" is within the range of kernel page table pages * that were preallocated during pmap bootstrap. Moreover, leave * "kernel_vm_end" and the kernel page table as they were. * * The correctness of this action is based on the following * argument: vm_map_insert() allocates contiguous ranges of the * kernel virtual address space. It calls this function if a range * ends after "kernel_vm_end". If the kernel is mapped between * "kernel_vm_end" and "addr", then the range cannot begin at * "kernel_vm_end". In fact, its beginning address cannot be less * than the kernel. Thus, there is no immediate need to allocate * any new kernel page table pages between "kernel_vm_end" and * "KERNBASE". */ if (KERNBASE < addr && addr <= KERNBASE + nkpt * NBPDR) return; addr = roundup2(addr, NBPDR); if (addr - 1 >= vm_map_max(kernel_map)) addr = vm_map_max(kernel_map); if (kernel_vm_end < addr) kasan_shadow_map(kernel_vm_end, addr - kernel_vm_end); while (kernel_vm_end < addr) { pdpe = pmap_pdpe(kernel_pmap, kernel_vm_end); if ((*pdpe & X86_PG_V) == 0) { /* We need a new PDP entry */ nkpg = pmap_alloc_pt_page(kernel_pmap, kernel_vm_end >> PDPSHIFT, VM_ALLOC_WIRED | VM_ALLOC_INTERRUPT | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); paddr = VM_PAGE_TO_PHYS(nkpg); *pdpe = (pdp_entry_t)(paddr | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M); continue; /* try again */ } pde = pmap_pdpe_to_pde(pdpe, kernel_vm_end); if ((*pde & X86_PG_V) != 0) { kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK; if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) { kernel_vm_end = vm_map_max(kernel_map); break; } continue; } nkpg = pmap_alloc_pt_page(kernel_pmap, pmap_pde_pindex(kernel_vm_end), VM_ALLOC_WIRED | VM_ALLOC_INTERRUPT | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); paddr = VM_PAGE_TO_PHYS(nkpg); newpdir = paddr | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; pde_store(pde, newpdir); kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK; if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) { kernel_vm_end = vm_map_max(kernel_map); break; } } } /*************************************************** * page management routines. ***************************************************/ CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE); CTASSERT(_NPCM == 3); CTASSERT(_NPCPV == 168); static __inline struct pv_chunk * pv_to_chunk(pv_entry_t pv) { return ((struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK)); } #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap) #define PC_FREE0 0xfffffffffffffffful #define PC_FREE1 0xfffffffffffffffful #define PC_FREE2 0x000000fffffffffful static const uint64_t pc_freemask[_NPCM] = { PC_FREE0, PC_FREE1, PC_FREE2 }; #ifdef PV_STATS static COUNTER_U64_DEFINE_EARLY(pc_chunk_count); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, "Current number of pv entry cnunks"); static COUNTER_U64_DEFINE_EARLY(pc_chunk_allocs); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, "Total number of pv entry chunks allocated"); static COUNTER_U64_DEFINE_EARLY(pc_chunk_frees); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, "Total number of pv entry chunks freed"); static COUNTER_U64_DEFINE_EARLY(pc_chunk_tryfail); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, "Number of failed attempts to get a pv entry chunk page"); static COUNTER_U64_DEFINE_EARLY(pv_entry_frees); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, "Total number of pv entries freed"); static COUNTER_U64_DEFINE_EARLY(pv_entry_allocs); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, "Total number of pv entries allocated"); static COUNTER_U64_DEFINE_EARLY(pv_entry_count); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, "Current number of pv entries"); static COUNTER_U64_DEFINE_EARLY(pv_entry_spare); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, "Current number of spare pv entries"); #endif static void reclaim_pv_chunk_leave_pmap(pmap_t pmap, pmap_t locked_pmap, bool start_di) { if (pmap == NULL) return; pmap_invalidate_all(pmap); if (pmap != locked_pmap) PMAP_UNLOCK(pmap); if (start_di) pmap_delayed_invl_finish(); } /* * We are in a serious low memory condition. Resort to * drastic measures to free some pages so we can allocate * another pv entry chunk. * * Returns NULL if PV entries were reclaimed from the specified pmap. * * We do not, however, unmap 2mpages because subsequent accesses will * allocate per-page pv entries until repromotion occurs, thereby * exacerbating the shortage of free pv entries. */ static vm_page_t reclaim_pv_chunk_domain(pmap_t locked_pmap, struct rwlock **lockp, int domain) { struct pv_chunks_list *pvc; struct pv_chunk *pc, *pc_marker, *pc_marker_end; struct pv_chunk_header pc_marker_b, pc_marker_end_b; struct md_page *pvh; pd_entry_t *pde; pmap_t next_pmap, pmap; pt_entry_t *pte, tpte; pt_entry_t PG_G, PG_A, PG_M, PG_RW; pv_entry_t pv; vm_offset_t va; vm_page_t m, m_pc; struct spglist free; uint64_t inuse; int bit, field, freed; bool start_di, restart; PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED); KASSERT(lockp != NULL, ("reclaim_pv_chunk: lockp is NULL")); pmap = NULL; m_pc = NULL; PG_G = PG_A = PG_M = PG_RW = 0; SLIST_INIT(&free); bzero(&pc_marker_b, sizeof(pc_marker_b)); bzero(&pc_marker_end_b, sizeof(pc_marker_end_b)); pc_marker = (struct pv_chunk *)&pc_marker_b; pc_marker_end = (struct pv_chunk *)&pc_marker_end_b; /* * A delayed invalidation block should already be active if * pmap_advise() or pmap_remove() called this function by way * of pmap_demote_pde_locked(). */ start_di = pmap_not_in_di(); pvc = &pv_chunks[domain]; mtx_lock(&pvc->pvc_lock); pvc->active_reclaims++; TAILQ_INSERT_HEAD(&pvc->pvc_list, pc_marker, pc_lru); TAILQ_INSERT_TAIL(&pvc->pvc_list, pc_marker_end, pc_lru); while ((pc = TAILQ_NEXT(pc_marker, pc_lru)) != pc_marker_end && SLIST_EMPTY(&free)) { next_pmap = pc->pc_pmap; if (next_pmap == NULL) { /* * The next chunk is a marker. However, it is * not our marker, so active_reclaims must be * > 1. Consequently, the next_chunk code * will not rotate the pv_chunks list. */ goto next_chunk; } mtx_unlock(&pvc->pvc_lock); /* * A pv_chunk can only be removed from the pc_lru list * when both pc_chunks_mutex is owned and the * corresponding pmap is locked. */ if (pmap != next_pmap) { restart = false; reclaim_pv_chunk_leave_pmap(pmap, locked_pmap, start_di); pmap = next_pmap; /* Avoid deadlock and lock recursion. */ if (pmap > locked_pmap) { RELEASE_PV_LIST_LOCK(lockp); PMAP_LOCK(pmap); if (start_di) pmap_delayed_invl_start(); mtx_lock(&pvc->pvc_lock); restart = true; } else if (pmap != locked_pmap) { if (PMAP_TRYLOCK(pmap)) { if (start_di) pmap_delayed_invl_start(); mtx_lock(&pvc->pvc_lock); restart = true; } else { pmap = NULL; /* pmap is not locked */ mtx_lock(&pvc->pvc_lock); pc = TAILQ_NEXT(pc_marker, pc_lru); if (pc == NULL || pc->pc_pmap != next_pmap) continue; goto next_chunk; } } else if (start_di) pmap_delayed_invl_start(); PG_G = pmap_global_bit(pmap); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); if (restart) continue; } /* * Destroy every non-wired, 4 KB page mapping in the chunk. */ freed = 0; for (field = 0; field < _NPCM; field++) { for (inuse = ~pc->pc_map[field] & pc_freemask[field]; inuse != 0; inuse &= ~(1UL << bit)) { bit = bsfq(inuse); pv = &pc->pc_pventry[field * 64 + bit]; va = pv->pv_va; pde = pmap_pde(pmap, va); if ((*pde & PG_PS) != 0) continue; pte = pmap_pde_to_pte(pde, va); if ((*pte & PG_W) != 0) continue; tpte = pte_load_clear(pte); if ((tpte & PG_G) != 0) pmap_invalidate_page(pmap, va); m = PHYS_TO_VM_PAGE(tpte & PG_FRAME); if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if ((tpte & PG_A) != 0) vm_page_aflag_set(m, PGA_REFERENCED); CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) { vm_page_aflag_clear(m, PGA_WRITEABLE); } } pmap_delayed_invl_page(m); pc->pc_map[field] |= 1UL << bit; pmap_unuse_pt(pmap, va, *pde, &free); freed++; } } if (freed == 0) { mtx_lock(&pvc->pvc_lock); goto next_chunk; } /* Every freed mapping is for a 4 KB page. */ pmap_resident_count_adj(pmap, -freed); PV_STAT(counter_u64_add(pv_entry_frees, freed)); PV_STAT(counter_u64_add(pv_entry_spare, freed)); PV_STAT(counter_u64_add(pv_entry_count, -freed)); TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); if (pc->pc_map[0] == PC_FREE0 && pc->pc_map[1] == PC_FREE1 && pc->pc_map[2] == PC_FREE2) { PV_STAT(counter_u64_add(pv_entry_spare, -_NPCPV)); PV_STAT(counter_u64_add(pc_chunk_count, -1)); PV_STAT(counter_u64_add(pc_chunk_frees, 1)); /* Entire chunk is free; return it. */ m_pc = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc)); dump_drop_page(m_pc->phys_addr); mtx_lock(&pvc->pvc_lock); TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru); break; } TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); mtx_lock(&pvc->pvc_lock); /* One freed pv entry in locked_pmap is sufficient. */ if (pmap == locked_pmap) break; next_chunk: TAILQ_REMOVE(&pvc->pvc_list, pc_marker, pc_lru); TAILQ_INSERT_AFTER(&pvc->pvc_list, pc, pc_marker, pc_lru); if (pvc->active_reclaims == 1 && pmap != NULL) { /* * Rotate the pv chunks list so that we do not * scan the same pv chunks that could not be * freed (because they contained a wired * and/or superpage mapping) on every * invocation of reclaim_pv_chunk(). */ while ((pc = TAILQ_FIRST(&pvc->pvc_list)) != pc_marker) { MPASS(pc->pc_pmap != NULL); TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru); TAILQ_INSERT_TAIL(&pvc->pvc_list, pc, pc_lru); } } } TAILQ_REMOVE(&pvc->pvc_list, pc_marker, pc_lru); TAILQ_REMOVE(&pvc->pvc_list, pc_marker_end, pc_lru); pvc->active_reclaims--; mtx_unlock(&pvc->pvc_lock); reclaim_pv_chunk_leave_pmap(pmap, locked_pmap, start_di); if (m_pc == NULL && !SLIST_EMPTY(&free)) { m_pc = SLIST_FIRST(&free); SLIST_REMOVE_HEAD(&free, plinks.s.ss); /* Recycle a freed page table page. */ m_pc->ref_count = 1; } vm_page_free_pages_toq(&free, true); return (m_pc); } static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp) { vm_page_t m; int i, domain; domain = PCPU_GET(domain); for (i = 0; i < vm_ndomains; i++) { m = reclaim_pv_chunk_domain(locked_pmap, lockp, domain); if (m != NULL) break; domain = (domain + 1) % vm_ndomains; } return (m); } /* * free the pv_entry back to the free list */ static void free_pv_entry(pmap_t pmap, pv_entry_t pv) { struct pv_chunk *pc; int idx, field, bit; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(counter_u64_add(pv_entry_frees, 1)); PV_STAT(counter_u64_add(pv_entry_spare, 1)); PV_STAT(counter_u64_add(pv_entry_count, -1)); pc = pv_to_chunk(pv); idx = pv - &pc->pc_pventry[0]; field = idx / 64; bit = idx % 64; pc->pc_map[field] |= 1ul << bit; if (pc->pc_map[0] != PC_FREE0 || pc->pc_map[1] != PC_FREE1 || pc->pc_map[2] != PC_FREE2) { /* 98% of the time, pc is already at the head of the list. */ if (__predict_false(pc != TAILQ_FIRST(&pmap->pm_pvchunk))) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); } return; } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } static void free_pv_chunk_dequeued(struct pv_chunk *pc) { vm_page_t m; PV_STAT(counter_u64_add(pv_entry_spare, -_NPCPV)); PV_STAT(counter_u64_add(pc_chunk_count, -1)); PV_STAT(counter_u64_add(pc_chunk_frees, 1)); counter_u64_add(pv_page_count, -1); /* entire chunk is free, return it */ m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc)); dump_drop_page(m->phys_addr); vm_page_unwire_noq(m); vm_page_free(m); } static void free_pv_chunk(struct pv_chunk *pc) { struct pv_chunks_list *pvc; pvc = &pv_chunks[pc_to_domain(pc)]; mtx_lock(&pvc->pvc_lock); TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru); mtx_unlock(&pvc->pvc_lock); free_pv_chunk_dequeued(pc); } static void free_pv_chunk_batch(struct pv_chunklist *batch) { struct pv_chunks_list *pvc; struct pv_chunk *pc, *npc; int i; for (i = 0; i < vm_ndomains; i++) { if (TAILQ_EMPTY(&batch[i])) continue; pvc = &pv_chunks[i]; mtx_lock(&pvc->pvc_lock); TAILQ_FOREACH(pc, &batch[i], pc_list) { TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru); } mtx_unlock(&pvc->pvc_lock); } for (i = 0; i < vm_ndomains; i++) { TAILQ_FOREACH_SAFE(pc, &batch[i], pc_list, npc) { free_pv_chunk_dequeued(pc); } } } /* * Returns a new PV entry, allocating a new PV chunk from the system when * needed. If this PV chunk allocation fails and a PV list lock pointer was * given, a PV chunk is reclaimed from an arbitrary pmap. Otherwise, NULL is * returned. * * The given PV list lock may be released. */ static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp) { struct pv_chunks_list *pvc; int bit, field; pv_entry_t pv; struct pv_chunk *pc; vm_page_t m; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(counter_u64_add(pv_entry_allocs, 1)); retry: pc = TAILQ_FIRST(&pmap->pm_pvchunk); if (pc != NULL) { for (field = 0; field < _NPCM; field++) { if (pc->pc_map[field]) { bit = bsfq(pc->pc_map[field]); break; } } if (field < _NPCM) { pv = &pc->pc_pventry[field * 64 + bit]; pc->pc_map[field] &= ~(1ul << bit); /* If this was the last item, move it to tail */ if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 && pc->pc_map[2] == 0) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } PV_STAT(counter_u64_add(pv_entry_count, 1)); PV_STAT(counter_u64_add(pv_entry_spare, -1)); return (pv); } } /* No free items, allocate another chunk */ m = vm_page_alloc_noobj(VM_ALLOC_WIRED); if (m == NULL) { if (lockp == NULL) { PV_STAT(counter_u64_add(pc_chunk_tryfail, 1)); return (NULL); } m = reclaim_pv_chunk(pmap, lockp); if (m == NULL) goto retry; } else counter_u64_add(pv_page_count, 1); PV_STAT(counter_u64_add(pc_chunk_count, 1)); PV_STAT(counter_u64_add(pc_chunk_allocs, 1)); dump_add_page(m->phys_addr); pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; pc->pc_map[0] = PC_FREE0 & ~1ul; /* preallocated bit 0 */ pc->pc_map[1] = PC_FREE1; pc->pc_map[2] = PC_FREE2; pvc = &pv_chunks[vm_page_domain(m)]; mtx_lock(&pvc->pvc_lock); TAILQ_INSERT_TAIL(&pvc->pvc_list, pc, pc_lru); mtx_unlock(&pvc->pvc_lock); pv = &pc->pc_pventry[0]; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); PV_STAT(counter_u64_add(pv_entry_count, 1)); PV_STAT(counter_u64_add(pv_entry_spare, _NPCPV - 1)); return (pv); } /* * Returns the number of one bits within the given PV chunk map. * * The erratas for Intel processors state that "POPCNT Instruction May * Take Longer to Execute Than Expected". It is believed that the * issue is the spurious dependency on the destination register. * Provide a hint to the register rename logic that the destination * value is overwritten, by clearing it, as suggested in the * optimization manual. It should be cheap for unaffected processors * as well. * * Reference numbers for erratas are * 4th Gen Core: HSD146 * 5th Gen Core: BDM85 * 6th Gen Core: SKL029 */ static int popcnt_pc_map_pq(uint64_t *map) { u_long result, tmp; __asm __volatile("xorl %k0,%k0;popcntq %2,%0;" "xorl %k1,%k1;popcntq %3,%1;addl %k1,%k0;" "xorl %k1,%k1;popcntq %4,%1;addl %k1,%k0" : "=&r" (result), "=&r" (tmp) : "m" (map[0]), "m" (map[1]), "m" (map[2])); return (result); } /* * Ensure that the number of spare PV entries in the specified pmap meets or * exceeds the given count, "needed". * * The given PV list lock may be released. */ static void reserve_pv_entries(pmap_t pmap, int needed, struct rwlock **lockp) { struct pv_chunks_list *pvc; struct pch new_tail[PMAP_MEMDOM]; struct pv_chunk *pc; vm_page_t m; int avail, free, i; bool reclaimed; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(lockp != NULL, ("reserve_pv_entries: lockp is NULL")); /* * Newly allocated PV chunks must be stored in a private list until * the required number of PV chunks have been allocated. Otherwise, * reclaim_pv_chunk() could recycle one of these chunks. In * contrast, these chunks must be added to the pmap upon allocation. */ for (i = 0; i < PMAP_MEMDOM; i++) TAILQ_INIT(&new_tail[i]); retry: avail = 0; TAILQ_FOREACH(pc, &pmap->pm_pvchunk, pc_list) { #ifndef __POPCNT__ if ((cpu_feature2 & CPUID2_POPCNT) == 0) bit_count((bitstr_t *)pc->pc_map, 0, sizeof(pc->pc_map) * NBBY, &free); else #endif free = popcnt_pc_map_pq(pc->pc_map); if (free == 0) break; avail += free; if (avail >= needed) break; } for (reclaimed = false; avail < needed; avail += _NPCPV) { m = vm_page_alloc_noobj(VM_ALLOC_WIRED); if (m == NULL) { m = reclaim_pv_chunk(pmap, lockp); if (m == NULL) goto retry; reclaimed = true; } else counter_u64_add(pv_page_count, 1); PV_STAT(counter_u64_add(pc_chunk_count, 1)); PV_STAT(counter_u64_add(pc_chunk_allocs, 1)); dump_add_page(m->phys_addr); pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; pc->pc_map[0] = PC_FREE0; pc->pc_map[1] = PC_FREE1; pc->pc_map[2] = PC_FREE2; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&new_tail[vm_page_domain(m)], pc, pc_lru); PV_STAT(counter_u64_add(pv_entry_spare, _NPCPV)); /* * The reclaim might have freed a chunk from the current pmap. * If that chunk contained available entries, we need to * re-count the number of available entries. */ if (reclaimed) goto retry; } for (i = 0; i < vm_ndomains; i++) { if (TAILQ_EMPTY(&new_tail[i])) continue; pvc = &pv_chunks[i]; mtx_lock(&pvc->pvc_lock); TAILQ_CONCAT(&pvc->pvc_list, &new_tail[i], pc_lru); mtx_unlock(&pvc->pvc_lock); } } /* * First find and then remove the pv entry for the specified pmap and virtual * address from the specified pv list. Returns the pv entry if found and NULL * otherwise. This operation can be performed on pv lists for either 4KB or * 2MB page mappings. */ static __inline pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (pmap == PV_PMAP(pv) && va == pv->pv_va) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; break; } } return (pv); } /* * After demotion from a 2MB page mapping to 512 4KB page mappings, * destroy the pv entry for the 2MB page mapping and reinstantiate the pv * entries for each of the 4KB page mappings. */ static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp) { struct md_page *pvh; struct pv_chunk *pc; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; int bit, field; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((pa & PDRMASK) == 0, ("pmap_pv_demote_pde: pa is not 2mpage aligned")); CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); /* * Transfer the 2mpage's pv entry for this mapping to the first * page's pv list. Once this transfer begins, the pv list lock * must not be released until the last pv entry is reinstantiated. */ pvh = pa_to_pvh(pa); va = trunc_2mpage(va); pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pv_demote_pde: pv not found")); m = PHYS_TO_VM_PAGE(pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; /* Instantiate the remaining NPTEPG - 1 pv entries. */ PV_STAT(counter_u64_add(pv_entry_allocs, NPTEPG - 1)); va_last = va + NBPDR - PAGE_SIZE; for (;;) { pc = TAILQ_FIRST(&pmap->pm_pvchunk); KASSERT(pc->pc_map[0] != 0 || pc->pc_map[1] != 0 || pc->pc_map[2] != 0, ("pmap_pv_demote_pde: missing spare")); for (field = 0; field < _NPCM; field++) { while (pc->pc_map[field]) { bit = bsfq(pc->pc_map[field]); pc->pc_map[field] &= ~(1ul << bit); pv = &pc->pc_pventry[field * 64 + bit]; va += PAGE_SIZE; pv->pv_va = va; m++; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_pv_demote_pde: page %p is not managed", m)); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if (va == va_last) goto out; } } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } out: if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 && pc->pc_map[2] == 0) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } PV_STAT(counter_u64_add(pv_entry_count, NPTEPG - 1)); PV_STAT(counter_u64_add(pv_entry_spare, -(NPTEPG - 1))); } #if VM_NRESERVLEVEL > 0 /* * After promotion from 512 4KB page mappings to a single 2MB page mapping, * replace the many pv entries for the 4KB page mappings by a single pv entry * for the 2MB page mapping. */ static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp) { struct md_page *pvh; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; KASSERT((pa & PDRMASK) == 0, ("pmap_pv_promote_pde: pa is not 2mpage aligned")); CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); /* * Transfer the first page's pv entry for this mapping to the 2mpage's * pv list. Aside from avoiding the cost of a call to get_pv_entry(), * a transfer avoids the possibility that get_pv_entry() calls * reclaim_pv_chunk() and that reclaim_pv_chunk() removes one of the * mappings that is being promoted. */ m = PHYS_TO_VM_PAGE(pa); va = trunc_2mpage(va); pv = pmap_pvh_remove(&m->md, pmap, va); KASSERT(pv != NULL, ("pmap_pv_promote_pde: pv not found")); pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; /* Free the remaining NPTEPG - 1 pv entries. */ va_last = va + NBPDR - PAGE_SIZE; do { m++; va += PAGE_SIZE; pmap_pvh_free(&m->md, pmap, va); } while (va < va_last); } #endif /* VM_NRESERVLEVEL > 0 */ /* * First find and then destroy the pv entry for the specified pmap and virtual * address. This operation can be performed on pv lists for either 4KB or 2MB * page mappings. */ static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pvh_free: pv not found")); free_pv_entry(pmap, pv); } /* * Conditionally create the PV entry for a 4KB page mapping if the required * memory can be allocated without resorting to reclamation. */ static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m, struct rwlock **lockp) { pv_entry_t pv; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* Pass NULL instead of the lock pointer to disable reclamation. */ if ((pv = get_pv_entry(pmap, NULL)) != NULL) { pv->pv_va = va; CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; return (TRUE); } else return (FALSE); } /* * Create the PV entry for a 2MB page mapping. Always returns true unless the * flag PMAP_ENTER_NORECLAIM is specified. If that flag is specified, returns * false if the PV entry cannot be allocated without resorting to reclamation. */ static bool pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, pd_entry_t pde, u_int flags, struct rwlock **lockp) { struct md_page *pvh; pv_entry_t pv; vm_paddr_t pa; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* Pass NULL instead of the lock pointer to disable reclamation. */ if ((pv = get_pv_entry(pmap, (flags & PMAP_ENTER_NORECLAIM) != 0 ? NULL : lockp)) == NULL) return (false); pv->pv_va = va; pa = pde & PG_PS_FRAME; CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; return (true); } /* * Fills a page table page with mappings to consecutive physical pages. */ static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte) { pt_entry_t *pte; for (pte = firstpte; pte < firstpte + NPTEPG; pte++) { *pte = newpte; newpte += PAGE_SIZE; } } /* * Tries to demote a 2MB page mapping. If demotion fails, the 2MB page * mapping is invalidated. */ static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) { struct rwlock *lock; boolean_t rv; lock = NULL; rv = pmap_demote_pde_locked(pmap, pde, va, &lock); if (lock != NULL) rw_wunlock(lock); return (rv); } static void pmap_demote_pde_check(pt_entry_t *firstpte __unused, pt_entry_t newpte __unused) { #ifdef INVARIANTS #ifdef DIAGNOSTIC pt_entry_t *xpte, *ypte; for (xpte = firstpte; xpte < firstpte + NPTEPG; xpte++, newpte += PAGE_SIZE) { if ((*xpte & PG_FRAME) != (newpte & PG_FRAME)) { printf("pmap_demote_pde: xpte %zd and newpte map " "different pages: found %#lx, expected %#lx\n", xpte - firstpte, *xpte, newpte); printf("page table dump\n"); for (ypte = firstpte; ypte < firstpte + NPTEPG; ypte++) printf("%zd %#lx\n", ypte - firstpte, *ypte); panic("firstpte"); } } #else KASSERT((*firstpte & PG_FRAME) == (newpte & PG_FRAME), ("pmap_demote_pde: firstpte and newpte map different physical" " addresses")); #endif #endif } static void pmap_demote_pde_abort(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t oldpde, struct rwlock **lockp) { struct spglist free; vm_offset_t sva; SLIST_INIT(&free); sva = trunc_2mpage(va); pmap_remove_pde(pmap, pde, sva, &free, lockp); if ((oldpde & pmap_global_bit(pmap)) == 0) pmap_invalidate_pde_page(pmap, sva, oldpde); vm_page_free_pages_toq(&free, true); CTR2(KTR_PMAP, "pmap_demote_pde: failure for va %#lx in pmap %p", va, pmap); } static boolean_t pmap_demote_pde_locked(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, struct rwlock **lockp) { pd_entry_t newpde, oldpde; pt_entry_t *firstpte, newpte; pt_entry_t PG_A, PG_G, PG_M, PG_PKU_MASK, PG_RW, PG_V; vm_paddr_t mptepa; vm_page_t mpte; int PG_PTE_CACHE; bool in_kernel; PG_A = pmap_accessed_bit(pmap); PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_PTE_CACHE = pmap_cache_mask(pmap, 0); PG_PKU_MASK = pmap_pku_mask_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); in_kernel = va >= VM_MAXUSER_ADDRESS; oldpde = *pde; KASSERT((oldpde & (PG_PS | PG_V)) == (PG_PS | PG_V), ("pmap_demote_pde: oldpde is missing PG_PS and/or PG_V")); /* * Invalidate the 2MB page mapping and return "failure" if the * mapping was never accessed. */ if ((oldpde & PG_A) == 0) { KASSERT((oldpde & PG_W) == 0, ("pmap_demote_pde: a wired mapping is missing PG_A")); pmap_demote_pde_abort(pmap, va, pde, oldpde, lockp); return (FALSE); } mpte = pmap_remove_pt_page(pmap, va); if (mpte == NULL) { KASSERT((oldpde & PG_W) == 0, ("pmap_demote_pde: page table page for a wired mapping" " is missing")); /* * If the page table page is missing and the mapping * is for a kernel address, the mapping must belong to * the direct map. Page table pages are preallocated * for every other part of the kernel address space, * so the direct map region is the only part of the * kernel address space that must be handled here. */ KASSERT(!in_kernel || (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS), ("pmap_demote_pde: No saved mpte for va %#lx", va)); /* * If the 2MB page mapping belongs to the direct map * region of the kernel's address space, then the page * allocation request specifies the highest possible * priority (VM_ALLOC_INTERRUPT). Otherwise, the * priority is normal. */ mpte = pmap_alloc_pt_page(pmap, pmap_pde_pindex(va), (in_kernel ? VM_ALLOC_INTERRUPT : 0) | VM_ALLOC_WIRED); /* * If the allocation of the new page table page fails, * invalidate the 2MB page mapping and return "failure". */ if (mpte == NULL) { pmap_demote_pde_abort(pmap, va, pde, oldpde, lockp); return (FALSE); } if (!in_kernel) mpte->ref_count = NPTEPG; } mptepa = VM_PAGE_TO_PHYS(mpte); firstpte = (pt_entry_t *)PHYS_TO_DMAP(mptepa); newpde = mptepa | PG_M | PG_A | (oldpde & PG_U) | PG_RW | PG_V; KASSERT((oldpde & (PG_M | PG_RW)) != PG_RW, ("pmap_demote_pde: oldpde is missing PG_M")); newpte = oldpde & ~PG_PS; newpte = pmap_swap_pat(pmap, newpte); /* * If the page table page is not leftover from an earlier promotion, * initialize it. */ if (mpte->valid == 0) pmap_fill_ptp(firstpte, newpte); pmap_demote_pde_check(firstpte, newpte); /* * If the mapping has changed attributes, update the page table * entries. */ if ((*firstpte & PG_PTE_PROMOTE) != (newpte & PG_PTE_PROMOTE)) pmap_fill_ptp(firstpte, newpte); /* * The spare PV entries must be reserved prior to demoting the * mapping, that is, prior to changing the PDE. Otherwise, the state * of the PDE and the PV lists will be inconsistent, which can result * in reclaim_pv_chunk() attempting to remove a PV entry from the * wrong PV list and pmap_pv_demote_pde() failing to find the expected * PV entry for the 2MB page mapping that is being demoted. */ if ((oldpde & PG_MANAGED) != 0) reserve_pv_entries(pmap, NPTEPG - 1, lockp); /* * Demote the mapping. This pmap is locked. The old PDE has * PG_A set. If the old PDE has PG_RW set, it also has PG_M * set. Thus, there is no danger of a race with another * processor changing the setting of PG_A and/or PG_M between * the read above and the store below. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, newpde); else pde_store(pde, newpde); /* * Invalidate a stale recursive mapping of the page table page. */ if (in_kernel) pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va)); /* * Demote the PV entry. */ if ((oldpde & PG_MANAGED) != 0) pmap_pv_demote_pde(pmap, va, oldpde & PG_PS_FRAME, lockp); counter_u64_add(pmap_pde_demotions, 1); CTR2(KTR_PMAP, "pmap_demote_pde: success for va %#lx in pmap %p", va, pmap); return (TRUE); } /* * pmap_remove_kernel_pde: Remove a kernel superpage mapping. */ static void pmap_remove_kernel_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) { pd_entry_t newpde; vm_paddr_t mptepa; vm_page_t mpte; KASSERT(pmap == kernel_pmap, ("pmap %p is not kernel_pmap", pmap)); PMAP_LOCK_ASSERT(pmap, MA_OWNED); mpte = pmap_remove_pt_page(pmap, va); if (mpte == NULL) panic("pmap_remove_kernel_pde: Missing pt page."); mptepa = VM_PAGE_TO_PHYS(mpte); newpde = mptepa | X86_PG_M | X86_PG_A | X86_PG_RW | X86_PG_V; /* * If this page table page was unmapped by a promotion, then it * contains valid mappings. Zero it to invalidate those mappings. */ if (mpte->valid != 0) pagezero((void *)PHYS_TO_DMAP(mptepa)); /* * Demote the mapping. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, newpde); else pde_store(pde, newpde); /* * Invalidate a stale recursive mapping of the page table page. */ pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va)); } /* * pmap_remove_pde: do the things to unmap a superpage in a process */ static int pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva, struct spglist *free, struct rwlock **lockp) { struct md_page *pvh; pd_entry_t oldpde; vm_offset_t eva, va; vm_page_t m, mpte; pt_entry_t PG_G, PG_A, PG_M, PG_RW; PG_G = pmap_global_bit(pmap); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & PDRMASK) == 0, ("pmap_remove_pde: sva is not 2mpage aligned")); oldpde = pte_load_clear(pdq); if (oldpde & PG_W) pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE; if ((oldpde & PG_G) != 0) pmap_invalidate_pde_page(kernel_pmap, sva, oldpde); pmap_resident_count_adj(pmap, -NBPDR / PAGE_SIZE); if (oldpde & PG_MANAGED) { CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, oldpde & PG_PS_FRAME); pvh = pa_to_pvh(oldpde & PG_PS_FRAME); pmap_pvh_free(pvh, pmap, sva); eva = sva + NBPDR; for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME); va < eva; va += PAGE_SIZE, m++) { if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if (oldpde & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); if (TAILQ_EMPTY(&m->md.pv_list) && TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); pmap_delayed_invl_page(m); } } if (pmap == kernel_pmap) { pmap_remove_kernel_pde(pmap, pdq, sva); } else { mpte = pmap_remove_pt_page(pmap, sva); if (mpte != NULL) { KASSERT(mpte->valid == VM_PAGE_BITS_ALL, ("pmap_remove_pde: pte page not promoted")); pmap_resident_count_adj(pmap, -1); KASSERT(mpte->ref_count == NPTEPG, ("pmap_remove_pde: pte page ref count error")); mpte->ref_count = 0; pmap_add_delayed_free_list(mpte, free, FALSE); } } return (pmap_unuse_pt(pmap, sva, *pmap_pdpe(pmap, sva), free)); } /* * pmap_remove_pte: do the things to unmap a page in a process */ static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va, pd_entry_t ptepde, struct spglist *free, struct rwlock **lockp) { struct md_page *pvh; pt_entry_t oldpte, PG_A, PG_M, PG_RW; vm_page_t m; PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldpte = pte_load_clear(ptq); if (oldpte & PG_W) pmap->pm_stats.wired_count -= 1; pmap_resident_count_adj(pmap, -1); if (oldpte & PG_MANAGED) { m = PHYS_TO_VM_PAGE(oldpte & PG_FRAME); if ((oldpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if (oldpte & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); pmap_pvh_free(&m->md, pmap, va); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } pmap_delayed_invl_page(m); } return (pmap_unuse_pt(pmap, va, ptepde, free)); } /* * Remove a single page from a process address space */ static void pmap_remove_page(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, struct spglist *free) { struct rwlock *lock; pt_entry_t *pte, PG_V; PG_V = pmap_valid_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if ((*pde & PG_V) == 0) return; pte = pmap_pde_to_pte(pde, va); if ((*pte & PG_V) == 0) return; lock = NULL; pmap_remove_pte(pmap, pte, va, *pde, free, &lock); if (lock != NULL) rw_wunlock(lock); pmap_invalidate_page(pmap, va); } /* * Removes the specified range of addresses from the page table page. */ static bool pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, pd_entry_t *pde, struct spglist *free, struct rwlock **lockp) { pt_entry_t PG_G, *pte; vm_offset_t va; bool anyvalid; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PG_G = pmap_global_bit(pmap); anyvalid = false; va = eva; for (pte = pmap_pde_to_pte(pde, sva); sva != eva; pte++, sva += PAGE_SIZE) { if (*pte == 0) { if (va != eva) { pmap_invalidate_range(pmap, va, sva); va = eva; } continue; } if ((*pte & PG_G) == 0) anyvalid = true; else if (va == eva) va = sva; if (pmap_remove_pte(pmap, pte, sva, *pde, free, lockp)) { sva += PAGE_SIZE; break; } } if (va != eva) pmap_invalidate_range(pmap, va, sva); return (anyvalid); } /* * Remove the given range of addresses from the specified map. * * It is assumed that the start and end are properly * rounded to the page size. */ void pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { struct rwlock *lock; vm_page_t mt; vm_offset_t va_next; pml5_entry_t *pml5e; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t ptpaddr, *pde; pt_entry_t PG_G, PG_V; struct spglist free; int anyvalid; PG_G = pmap_global_bit(pmap); PG_V = pmap_valid_bit(pmap); /* * If there are no resident pages besides the top level page * table page(s), there is nothing to do. Kernel pmap always * accounts whole preloaded area as resident, which makes its * resident count > 2. * Perform an unsynchronized read. This is, however, safe. */ if (pmap->pm_stats.resident_count <= 1 + (pmap->pm_pmltopu != NULL ? 1 : 0)) return; anyvalid = 0; SLIST_INIT(&free); pmap_delayed_invl_start(); PMAP_LOCK(pmap); pmap_pkru_on_remove(pmap, sva, eva); /* * special handling of removing one page. a very * common operation and easy to short circuit some * code. */ if (sva + PAGE_SIZE == eva) { pde = pmap_pde(pmap, sva); if (pde && (*pde & PG_PS) == 0) { pmap_remove_page(pmap, sva, pde, &free); goto out; } } lock = NULL; for (; sva < eva; sva = va_next) { if (pmap->pm_stats.resident_count == 0) break; if (pmap_is_la57(pmap)) { pml5e = pmap_pml5e(pmap, sva); if ((*pml5e & PG_V) == 0) { va_next = (sva + NBPML5) & ~PML5MASK; if (va_next < sva) va_next = eva; continue; } pml4e = pmap_pml5e_to_pml4e(pml5e, sva); } else { pml4e = pmap_pml4e(pmap, sva); } if ((*pml4e & PG_V) == 0) { va_next = (sva + NBPML4) & ~PML4MASK; if (va_next < sva) va_next = eva; continue; } va_next = (sva + NBPDP) & ~PDPMASK; if (va_next < sva) va_next = eva; pdpe = pmap_pml4e_to_pdpe(pml4e, sva); if ((*pdpe & PG_V) == 0) continue; if ((*pdpe & PG_PS) != 0) { KASSERT(va_next <= eva, ("partial update of non-transparent 1G mapping " "pdpe %#lx sva %#lx eva %#lx va_next %#lx", *pdpe, sva, eva, va_next)); MPASS(pmap != kernel_pmap); /* XXXKIB */ MPASS((*pdpe & (PG_MANAGED | PG_G)) == 0); anyvalid = 1; *pdpe = 0; pmap_resident_count_adj(pmap, -NBPDP / PAGE_SIZE); mt = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, sva) & PG_FRAME); pmap_unwire_ptp(pmap, sva, mt, &free); continue; } /* * Calculate index for next page table. */ va_next = (sva + NBPDR) & ~PDRMASK; if (va_next < sva) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, sva); ptpaddr = *pde; /* * Weed out invalid mappings. */ if (ptpaddr == 0) continue; /* * Check for large page. */ if ((ptpaddr & PG_PS) != 0) { /* * Are we removing the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == va_next && eva >= va_next) { /* * The TLB entry for a PG_G mapping is * invalidated by pmap_remove_pde(). */ if ((ptpaddr & PG_G) == 0) anyvalid = 1; pmap_remove_pde(pmap, pde, sva, &free, &lock); continue; } else if (!pmap_demote_pde_locked(pmap, pde, sva, &lock)) { /* The large page mapping was destroyed. */ continue; } else ptpaddr = *pde; } /* * Limit our scan to either the end of the va represented * by the current page table page, or to the end of the * range being removed. */ if (va_next > eva) va_next = eva; if (pmap_remove_ptes(pmap, sva, va_next, pde, &free, &lock)) anyvalid = 1; } if (lock != NULL) rw_wunlock(lock); out: if (anyvalid) pmap_invalidate_all(pmap); PMAP_UNLOCK(pmap); pmap_delayed_invl_finish(); vm_page_free_pages_toq(&free, true); } /* * Routine: pmap_remove_all * Function: * Removes this physical page from * all physical maps in which it resides. * Reflects back modify bits to the pager. * * Notes: * Original versions of this routine were very * inefficient because they iteratively called * pmap_remove (slow...) */ void pmap_remove_all(vm_page_t m) { struct md_page *pvh; pv_entry_t pv; pmap_t pmap; struct rwlock *lock; pt_entry_t *pte, tpte, PG_A, PG_M, PG_RW; pd_entry_t *pde; vm_offset_t va; struct spglist free; int pvh_gen, md_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_all: page %p is not managed", m)); SLIST_INIT(&free); lock = VM_PAGE_TO_PV_LIST_LOCK(m); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(VM_PAGE_TO_PHYS(m)); rw_wlock(lock); retry: while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } va = pv->pv_va; pde = pmap_pde(pmap, va); (void)pmap_demote_pde_locked(pmap, pde, va, &lock); PMAP_UNLOCK(pmap); } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); pmap_resident_count_adj(pmap, -1); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_remove_all: found" " a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); tpte = pte_load_clear(pte); if (tpte & PG_W) pmap->pm_stats.wired_count--; if (tpte & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); /* * Update the vm_page_t clean and reference bits. */ if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); pmap_unuse_pt(pmap, pv->pv_va, *pde, &free); pmap_invalidate_page(pmap, pv->pv_va); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; free_pv_entry(pmap, pv); PMAP_UNLOCK(pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(lock); pmap_delayed_invl_wait(m); vm_page_free_pages_toq(&free, true); } /* * pmap_protect_pde: do the things to protect a 2mpage in a process */ static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot) { pd_entry_t newpde, oldpde; vm_page_t m, mt; boolean_t anychanged; pt_entry_t PG_G, PG_M, PG_RW; PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & PDRMASK) == 0, ("pmap_protect_pde: sva is not 2mpage aligned")); anychanged = FALSE; retry: oldpde = newpde = *pde; if ((prot & VM_PROT_WRITE) == 0) { if ((oldpde & (PG_MANAGED | PG_M | PG_RW)) == (PG_MANAGED | PG_M | PG_RW)) { m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME); for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) vm_page_dirty(mt); } newpde &= ~(PG_RW | PG_M); } if ((prot & VM_PROT_EXECUTE) == 0) newpde |= pg_nx; if (newpde != oldpde) { /* * As an optimization to future operations on this PDE, clear * PG_PROMOTED. The impending invalidation will remove any * lingering 4KB page mappings from the TLB. */ if (!atomic_cmpset_long(pde, oldpde, newpde & ~PG_PROMOTED)) goto retry; if ((oldpde & PG_G) != 0) pmap_invalidate_pde_page(kernel_pmap, sva, oldpde); else anychanged = TRUE; } return (anychanged); } /* * Set the physical protection on the * specified range of this map as requested. */ void pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { vm_page_t m; vm_offset_t va_next; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t ptpaddr, *pde; pt_entry_t *pte, PG_G, PG_M, PG_RW, PG_V; pt_entry_t obits, pbits; boolean_t anychanged; KASSERT((prot & ~VM_PROT_ALL) == 0, ("invalid prot %x", prot)); if (prot == VM_PROT_NONE) { pmap_remove(pmap, sva, eva); return; } if ((prot & (VM_PROT_WRITE|VM_PROT_EXECUTE)) == (VM_PROT_WRITE|VM_PROT_EXECUTE)) return; PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); anychanged = FALSE; /* * Although this function delays and batches the invalidation * of stale TLB entries, it does not need to call * pmap_delayed_invl_start() and * pmap_delayed_invl_finish(), because it does not * ordinarily destroy mappings. Stale TLB entries from * protection-only changes need only be invalidated before the * pmap lock is released, because protection-only changes do * not destroy PV entries. Even operations that iterate over * a physical page's PV list of mappings, like * pmap_remove_write(), acquire the pmap lock for each * mapping. Consequently, for protection-only changes, the * pmap lock suffices to synchronize both page table and TLB * updates. * * This function only destroys a mapping if pmap_demote_pde() * fails. In that case, stale TLB entries are immediately * invalidated. */ PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { pml4e = pmap_pml4e(pmap, sva); if (pml4e == NULL || (*pml4e & PG_V) == 0) { va_next = (sva + NBPML4) & ~PML4MASK; if (va_next < sva) va_next = eva; continue; } va_next = (sva + NBPDP) & ~PDPMASK; if (va_next < sva) va_next = eva; pdpe = pmap_pml4e_to_pdpe(pml4e, sva); if ((*pdpe & PG_V) == 0) continue; if ((*pdpe & PG_PS) != 0) { KASSERT(va_next <= eva, ("partial update of non-transparent 1G mapping " "pdpe %#lx sva %#lx eva %#lx va_next %#lx", *pdpe, sva, eva, va_next)); retry_pdpe: obits = pbits = *pdpe; MPASS((pbits & (PG_MANAGED | PG_G)) == 0); MPASS(pmap != kernel_pmap); /* XXXKIB */ if ((prot & VM_PROT_WRITE) == 0) pbits &= ~(PG_RW | PG_M); if ((prot & VM_PROT_EXECUTE) == 0) pbits |= pg_nx; if (pbits != obits) { if (!atomic_cmpset_long(pdpe, obits, pbits)) /* PG_PS cannot be cleared under us, */ goto retry_pdpe; anychanged = TRUE; } continue; } va_next = (sva + NBPDR) & ~PDRMASK; if (va_next < sva) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, sva); ptpaddr = *pde; /* * Weed out invalid mappings. */ if (ptpaddr == 0) continue; /* * Check for large page. */ if ((ptpaddr & PG_PS) != 0) { /* * Are we protecting the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == va_next && eva >= va_next) { /* * The TLB entry for a PG_G mapping is * invalidated by pmap_protect_pde(). */ if (pmap_protect_pde(pmap, pde, sva, prot)) anychanged = TRUE; continue; } else if (!pmap_demote_pde(pmap, pde, sva)) { /* * The large page mapping was destroyed. */ continue; } } if (va_next > eva) va_next = eva; for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, sva += PAGE_SIZE) { retry: obits = pbits = *pte; if ((pbits & PG_V) == 0) continue; if ((prot & VM_PROT_WRITE) == 0) { if ((pbits & (PG_MANAGED | PG_M | PG_RW)) == (PG_MANAGED | PG_M | PG_RW)) { m = PHYS_TO_VM_PAGE(pbits & PG_FRAME); vm_page_dirty(m); } pbits &= ~(PG_RW | PG_M); } if ((prot & VM_PROT_EXECUTE) == 0) pbits |= pg_nx; if (pbits != obits) { if (!atomic_cmpset_long(pte, obits, pbits)) goto retry; if (obits & PG_G) pmap_invalidate_page(pmap, sva); else anychanged = TRUE; } } } if (anychanged) pmap_invalidate_all(pmap); PMAP_UNLOCK(pmap); } #if VM_NRESERVLEVEL > 0 static bool pmap_pde_ept_executable(pmap_t pmap, pd_entry_t pde) { if (pmap->pm_type != PT_EPT) return (false); return ((pde & EPT_PG_EXECUTE) != 0); } /* * Tries to promote the 512, contiguous 4KB page mappings that are within a * single page table page (PTP) to a single 2MB page mapping. For promotion * to occur, two conditions must be met: (1) the 4KB page mappings must map * aligned, contiguous physical memory and (2) the 4KB page mappings must have * identical characteristics. */ static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, struct rwlock **lockp) { pd_entry_t newpde; pt_entry_t *firstpte, oldpte, pa, *pte; pt_entry_t PG_G, PG_A, PG_M, PG_RW, PG_V, PG_PKU_MASK; vm_page_t mpte; int PG_PTE_CACHE; PG_A = pmap_accessed_bit(pmap); PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); PG_PKU_MASK = pmap_pku_mask_bit(pmap); PG_PTE_CACHE = pmap_cache_mask(pmap, 0); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * Examine the first PTE in the specified PTP. Abort if this PTE is * either invalid, unused, or does not map the first 4KB physical page * within a 2MB page. */ firstpte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME); newpde = *firstpte; if ((newpde & ((PG_FRAME & PDRMASK) | PG_A | PG_V)) != (PG_A | PG_V) || !pmap_allow_2m_x_page(pmap, pmap_pde_ept_executable(pmap, newpde))) { counter_u64_add(pmap_pde_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx" " in pmap %p", va, pmap); return; } setpde: if ((newpde & (PG_M | PG_RW)) == PG_RW) { /* * When PG_M is already clear, PG_RW can be cleared without * a TLB invalidation. */ if (!atomic_fcmpset_long(firstpte, &newpde, newpde & ~PG_RW)) goto setpde; newpde &= ~PG_RW; } /* * Examine each of the other PTEs in the specified PTP. Abort if this * PTE maps an unexpected 4KB physical page or does not have identical * characteristics to the first PTE. */ pa = (newpde & (PG_PS_FRAME | PG_A | PG_V)) + NBPDR - PAGE_SIZE; for (pte = firstpte + NPTEPG - 1; pte > firstpte; pte--) { oldpte = *pte; if ((oldpte & (PG_FRAME | PG_A | PG_V)) != pa) { counter_u64_add(pmap_pde_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx" " in pmap %p", va, pmap); return; } setpte: if ((oldpte & (PG_M | PG_RW)) == PG_RW) { /* * When PG_M is already clear, PG_RW can be cleared * without a TLB invalidation. */ if (!atomic_fcmpset_long(pte, &oldpte, oldpte & ~PG_RW)) goto setpte; oldpte &= ~PG_RW; CTR2(KTR_PMAP, "pmap_promote_pde: protect for va %#lx" " in pmap %p", (oldpte & PG_FRAME & PDRMASK) | (va & ~PDRMASK), pmap); } if ((oldpte & PG_PTE_PROMOTE) != (newpde & PG_PTE_PROMOTE)) { counter_u64_add(pmap_pde_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx" " in pmap %p", va, pmap); return; } pa -= PAGE_SIZE; } /* * Save the page table page in its current state until the PDE * mapping the superpage is demoted by pmap_demote_pde() or * destroyed by pmap_remove_pde(). */ mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME); KASSERT(mpte >= vm_page_array && mpte < &vm_page_array[vm_page_array_size], ("pmap_promote_pde: page table page is out of range")); KASSERT(mpte->pindex == pmap_pde_pindex(va), ("pmap_promote_pde: page table page's pindex is wrong " "mpte %p pidx %#lx va %#lx va pde pidx %#lx", mpte, mpte->pindex, va, pmap_pde_pindex(va))); if (pmap_insert_pt_page(pmap, mpte, true)) { counter_u64_add(pmap_pde_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx in pmap %p", va, pmap); return; } /* * Promote the pv entries. */ if ((newpde & PG_MANAGED) != 0) pmap_pv_promote_pde(pmap, va, newpde & PG_PS_FRAME, lockp); /* * Propagate the PAT index to its proper position. */ newpde = pmap_swap_pat(pmap, newpde); /* * Map the superpage. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, PG_PS | newpde); else pde_store(pde, PG_PROMOTED | PG_PS | newpde); counter_u64_add(pmap_pde_promotions, 1); CTR2(KTR_PMAP, "pmap_promote_pde: success for va %#lx" " in pmap %p", va, pmap); } #endif /* VM_NRESERVLEVEL > 0 */ static int pmap_enter_largepage(pmap_t pmap, vm_offset_t va, pt_entry_t newpte, int flags, int psind) { vm_page_t mp; pt_entry_t origpte, *pml4e, *pdpe, *pde, pten, PG_V; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(psind > 0 && psind < MAXPAGESIZES && pagesizes[psind] != 0, ("psind %d unexpected", psind)); KASSERT(((newpte & PG_FRAME) & (pagesizes[psind] - 1)) == 0, ("unaligned phys address %#lx newpte %#lx psind %d", newpte & PG_FRAME, newpte, psind)); KASSERT((va & (pagesizes[psind] - 1)) == 0, ("unaligned va %#lx psind %d", va, psind)); KASSERT(va < VM_MAXUSER_ADDRESS, ("kernel mode non-transparent superpage")); /* XXXKIB */ KASSERT(va + pagesizes[psind] < VM_MAXUSER_ADDRESS, ("overflowing user map va %#lx psind %d", va, psind)); /* XXXKIB */ PG_V = pmap_valid_bit(pmap); restart: if (!pmap_pkru_same(pmap, va, va + pagesizes[psind])) return (KERN_PROTECTION_FAILURE); pten = newpte; if (va < VM_MAXUSER_ADDRESS && pmap->pm_type == PT_X86) pten |= pmap_pkru_get(pmap, va); if (psind == 2) { /* 1G */ pml4e = pmap_pml4e(pmap, va); if (pml4e == NULL || (*pml4e & PG_V) == 0) { mp = pmap_allocpte_alloc(pmap, pmap_pml4e_pindex(va), NULL, va); if (mp == NULL) goto allocf; pdpe = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mp)); pdpe = &pdpe[pmap_pdpe_index(va)]; origpte = *pdpe; MPASS(origpte == 0); } else { pdpe = pmap_pml4e_to_pdpe(pml4e, va); KASSERT(pdpe != NULL, ("va %#lx lost pdpe", va)); origpte = *pdpe; if ((origpte & PG_V) == 0) { mp = PHYS_TO_VM_PAGE(*pml4e & PG_FRAME); mp->ref_count++; } } *pdpe = pten; } else /* (psind == 1) */ { /* 2M */ pde = pmap_pde(pmap, va); if (pde == NULL) { mp = pmap_allocpte_alloc(pmap, pmap_pdpe_pindex(va), NULL, va); if (mp == NULL) goto allocf; pde = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mp)); pde = &pde[pmap_pde_index(va)]; origpte = *pde; MPASS(origpte == 0); } else { origpte = *pde; if ((origpte & PG_V) == 0) { pdpe = pmap_pdpe(pmap, va); MPASS(pdpe != NULL && (*pdpe & PG_V) != 0); mp = PHYS_TO_VM_PAGE(*pdpe & PG_FRAME); mp->ref_count++; } } *pde = pten; } KASSERT((origpte & PG_V) == 0 || ((origpte & PG_PS) != 0 && (origpte & PG_PS_FRAME) == (pten & PG_PS_FRAME)), ("va %#lx changing %s phys page origpte %#lx pten %#lx", va, psind == 2 ? "1G" : "2M", origpte, pten)); if ((pten & PG_W) != 0 && (origpte & PG_W) == 0) pmap->pm_stats.wired_count += pagesizes[psind] / PAGE_SIZE; else if ((pten & PG_W) == 0 && (origpte & PG_W) != 0) pmap->pm_stats.wired_count -= pagesizes[psind] / PAGE_SIZE; if ((origpte & PG_V) == 0) pmap_resident_count_adj(pmap, pagesizes[psind] / PAGE_SIZE); return (KERN_SUCCESS); allocf: if ((flags & PMAP_ENTER_NOSLEEP) != 0) return (KERN_RESOURCE_SHORTAGE); PMAP_UNLOCK(pmap); vm_wait(NULL); PMAP_LOCK(pmap); goto restart; } /* * Insert the given physical page (p) at * the specified virtual address (v) in the * target physical map with the protection requested. * * If specified, the page will be wired down, meaning * that the related pte can not be reclaimed. * * NB: This is the only routine which MAY NOT lazy-evaluate * or lose information. That is, this routine must actually * insert this page into the given map NOW. * * When destroying both a page table and PV entry, this function * performs the TLB invalidation before releasing the PV list * lock, so we do not need pmap_delayed_invl_page() calls here. */ int pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind) { struct rwlock *lock; pd_entry_t *pde; pt_entry_t *pte, PG_G, PG_A, PG_M, PG_RW, PG_V; pt_entry_t newpte, origpte; pv_entry_t pv; vm_paddr_t opa, pa; vm_page_t mpte, om; int rv; boolean_t nosleep; PG_A = pmap_accessed_bit(pmap); PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); va = trunc_page(va); KASSERT(va <= VM_MAX_KERNEL_ADDRESS, ("pmap_enter: toobig")); KASSERT(va < UPT_MIN_ADDRESS || va >= UPT_MAX_ADDRESS, ("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va)); KASSERT((m->oflags & VPO_UNMANAGED) != 0 || va < kmi.clean_sva || va >= kmi.clean_eva, ("pmap_enter: managed mapping within the clean submap")); if ((m->oflags & VPO_UNMANAGED) == 0) VM_PAGE_OBJECT_BUSY_ASSERT(m); KASSERT((flags & PMAP_ENTER_RESERVED) == 0, ("pmap_enter: flags %u has reserved bits set", flags)); pa = VM_PAGE_TO_PHYS(m); newpte = (pt_entry_t)(pa | PG_A | PG_V); if ((flags & VM_PROT_WRITE) != 0) newpte |= PG_M; if ((prot & VM_PROT_WRITE) != 0) newpte |= PG_RW; KASSERT((newpte & (PG_M | PG_RW)) != PG_M, ("pmap_enter: flags includes VM_PROT_WRITE but prot doesn't")); if ((prot & VM_PROT_EXECUTE) == 0) newpte |= pg_nx; if ((flags & PMAP_ENTER_WIRED) != 0) newpte |= PG_W; if (va < VM_MAXUSER_ADDRESS) newpte |= PG_U; if (pmap == kernel_pmap) newpte |= PG_G; newpte |= pmap_cache_bits(pmap, m->md.pat_mode, psind > 0); /* * Set modified bit gratuitously for writeable mappings if * the page is unmanaged. We do not want to take a fault * to do the dirty bit accounting for these mappings. */ if ((m->oflags & VPO_UNMANAGED) != 0) { if ((newpte & PG_RW) != 0) newpte |= PG_M; } else newpte |= PG_MANAGED; lock = NULL; PMAP_LOCK(pmap); if ((flags & PMAP_ENTER_LARGEPAGE) != 0) { KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed largepage va %#lx flags %#x", va, flags)); rv = pmap_enter_largepage(pmap, va, newpte | PG_PS, flags, psind); goto out; } if (psind == 1) { /* Assert the required virtual and physical alignment. */ KASSERT((va & PDRMASK) == 0, ("pmap_enter: va unaligned")); KASSERT(m->psind > 0, ("pmap_enter: m->psind < psind")); rv = pmap_enter_pde(pmap, va, newpte | PG_PS, flags, m, &lock); goto out; } mpte = NULL; /* * In the case that a page table page is not * resident, we are creating it here. */ retry: pde = pmap_pde(pmap, va); if (pde != NULL && (*pde & PG_V) != 0 && ((*pde & PG_PS) == 0 || pmap_demote_pde_locked(pmap, pde, va, &lock))) { pte = pmap_pde_to_pte(pde, va); if (va < VM_MAXUSER_ADDRESS && mpte == NULL) { mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME); mpte->ref_count++; } } else if (va < VM_MAXUSER_ADDRESS) { /* * Here if the pte page isn't mapped, or if it has been * deallocated. */ nosleep = (flags & PMAP_ENTER_NOSLEEP) != 0; mpte = pmap_allocpte_alloc(pmap, pmap_pde_pindex(va), nosleep ? NULL : &lock, va); if (mpte == NULL && nosleep) { rv = KERN_RESOURCE_SHORTAGE; goto out; } goto retry; } else panic("pmap_enter: invalid page directory va=%#lx", va); origpte = *pte; pv = NULL; if (va < VM_MAXUSER_ADDRESS && pmap->pm_type == PT_X86) newpte |= pmap_pkru_get(pmap, va); /* * Is the specified virtual address already mapped? */ if ((origpte & PG_V) != 0) { /* * Wiring change, just update stats. We don't worry about * wiring PT pages as they remain resident as long as there * are valid mappings in them. Hence, if a user page is wired, * the PT page will be also. */ if ((newpte & PG_W) != 0 && (origpte & PG_W) == 0) pmap->pm_stats.wired_count++; else if ((newpte & PG_W) == 0 && (origpte & PG_W) != 0) pmap->pm_stats.wired_count--; /* * Remove the extra PT page reference. */ if (mpte != NULL) { mpte->ref_count--; KASSERT(mpte->ref_count > 0, ("pmap_enter: missing reference to page table page," " va: 0x%lx", va)); } /* * Has the physical page changed? */ opa = origpte & PG_FRAME; if (opa == pa) { /* * No, might be a protection or wiring change. */ if ((origpte & PG_MANAGED) != 0 && (newpte & PG_RW) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); if (((origpte ^ newpte) & ~(PG_M | PG_A)) == 0) goto unchanged; goto validate; } /* * The physical page has changed. Temporarily invalidate * the mapping. This ensures that all threads sharing the * pmap keep a consistent view of the mapping, which is * necessary for the correct handling of COW faults. It * also permits reuse of the old mapping's PV entry, * avoiding an allocation. * * For consistency, handle unmanaged mappings the same way. */ origpte = pte_load_clear(pte); KASSERT((origpte & PG_FRAME) == opa, ("pmap_enter: unexpected pa update for %#lx", va)); if ((origpte & PG_MANAGED) != 0) { om = PHYS_TO_VM_PAGE(opa); /* * The pmap lock is sufficient to synchronize with * concurrent calls to pmap_page_test_mappings() and * pmap_ts_referenced(). */ if ((origpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(om); if ((origpte & PG_A) != 0) { pmap_invalidate_page(pmap, va); vm_page_aflag_set(om, PGA_REFERENCED); } CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, opa); pv = pmap_pvh_remove(&om->md, pmap, va); KASSERT(pv != NULL, ("pmap_enter: no PV entry for %#lx", va)); if ((newpte & PG_MANAGED) == 0) free_pv_entry(pmap, pv); if ((om->a.flags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&om->md.pv_list) && ((om->flags & PG_FICTITIOUS) != 0 || TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list))) vm_page_aflag_clear(om, PGA_WRITEABLE); } else { /* * Since this mapping is unmanaged, assume that PG_A * is set. */ pmap_invalidate_page(pmap, va); } origpte = 0; } else { /* * Increment the counters. */ if ((newpte & PG_W) != 0) pmap->pm_stats.wired_count++; pmap_resident_count_adj(pmap, 1); } /* * Enter on the PV list if part of our managed memory. */ if ((newpte & PG_MANAGED) != 0) { if (pv == NULL) { pv = get_pv_entry(pmap, &lock); pv->pv_va = va; } CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if ((newpte & PG_RW) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); } /* * Update the PTE. */ if ((origpte & PG_V) != 0) { validate: origpte = pte_load_store(pte, newpte); KASSERT((origpte & PG_FRAME) == pa, ("pmap_enter: unexpected pa update for %#lx", va)); if ((newpte & PG_M) == 0 && (origpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if ((origpte & PG_MANAGED) != 0) vm_page_dirty(m); /* * Although the PTE may still have PG_RW set, TLB * invalidation may nonetheless be required because * the PTE no longer has PG_M set. */ } else if ((origpte & PG_NX) != 0 || (newpte & PG_NX) == 0) { /* * This PTE change does not require TLB invalidation. */ goto unchanged; } if ((origpte & PG_A) != 0) pmap_invalidate_page(pmap, va); } else pte_store(pte, newpte); unchanged: #if VM_NRESERVLEVEL > 0 /* * If both the page table page and the reservation are fully * populated, then attempt promotion. */ if ((mpte == NULL || mpte->ref_count == NPTEPG) && pmap_ps_enabled(pmap) && (m->flags & PG_FICTITIOUS) == 0 && vm_reserv_level_iffullpop(m) == 0) pmap_promote_pde(pmap, pde, va, &lock); #endif rv = KERN_SUCCESS; out: if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); return (rv); } /* * Tries to create a read- and/or execute-only 2MB page mapping. Returns true * if successful. Returns false if (1) a page table page cannot be allocated * without sleeping, (2) a mapping already exists at the specified virtual * address, or (3) a PV entry cannot be allocated without reclaiming another * PV entry. */ static bool pmap_enter_2mpage(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, struct rwlock **lockp) { pd_entry_t newpde; pt_entry_t PG_V; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PG_V = pmap_valid_bit(pmap); newpde = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(pmap, m->md.pat_mode, 1) | PG_PS | PG_V; if ((m->oflags & VPO_UNMANAGED) == 0) newpde |= PG_MANAGED; if ((prot & VM_PROT_EXECUTE) == 0) newpde |= pg_nx; if (va < VM_MAXUSER_ADDRESS) newpde |= PG_U; return (pmap_enter_pde(pmap, va, newpde, PMAP_ENTER_NOSLEEP | PMAP_ENTER_NOREPLACE | PMAP_ENTER_NORECLAIM, NULL, lockp) == KERN_SUCCESS); } /* * Returns true if every page table entry in the specified page table page is * zero. */ static bool pmap_every_pte_zero(vm_paddr_t pa) { pt_entry_t *pt_end, *pte; KASSERT((pa & PAGE_MASK) == 0, ("pa is misaligned")); pte = (pt_entry_t *)PHYS_TO_DMAP(pa); for (pt_end = pte + NPTEPG; pte < pt_end; pte++) { if (*pte != 0) return (false); } return (true); } /* * Tries to create the specified 2MB page mapping. Returns KERN_SUCCESS if * the mapping was created, and either KERN_FAILURE or KERN_RESOURCE_SHORTAGE * otherwise. Returns KERN_FAILURE if PMAP_ENTER_NOREPLACE was specified and * a mapping already exists at the specified virtual address. Returns * KERN_RESOURCE_SHORTAGE if PMAP_ENTER_NOSLEEP was specified and a page table * page allocation failed. Returns KERN_RESOURCE_SHORTAGE if * PMAP_ENTER_NORECLAIM was specified and a PV entry allocation failed. * * The parameter "m" is only used when creating a managed, writeable mapping. */ static int pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde, u_int flags, vm_page_t m, struct rwlock **lockp) { struct spglist free; pd_entry_t oldpde, *pde; pt_entry_t PG_G, PG_RW, PG_V; vm_page_t mt, pdpg; KASSERT(pmap == kernel_pmap || (newpde & PG_W) == 0, ("pmap_enter_pde: cannot create wired user mapping")); PG_G = pmap_global_bit(pmap); PG_RW = pmap_rw_bit(pmap); KASSERT((newpde & (pmap_modified_bit(pmap) | PG_RW)) != PG_RW, ("pmap_enter_pde: newpde is missing PG_M")); PG_V = pmap_valid_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (!pmap_allow_2m_x_page(pmap, pmap_pde_ept_executable(pmap, newpde))) { CTR2(KTR_PMAP, "pmap_enter_pde: 2m x blocked for va %#lx" " in pmap %p", va, pmap); return (KERN_FAILURE); } if ((pde = pmap_alloc_pde(pmap, va, &pdpg, (flags & PMAP_ENTER_NOSLEEP) != 0 ? NULL : lockp)) == NULL) { CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (KERN_RESOURCE_SHORTAGE); } /* * If pkru is not same for the whole pde range, return failure * and let vm_fault() cope. Check after pde allocation, since * it could sleep. */ if (!pmap_pkru_same(pmap, va, va + NBPDR)) { pmap_abort_ptp(pmap, va, pdpg); - return (KERN_FAILURE); + return (KERN_PROTECTION_FAILURE); } if (va < VM_MAXUSER_ADDRESS && pmap->pm_type == PT_X86) { newpde &= ~X86_PG_PKU_MASK; newpde |= pmap_pkru_get(pmap, va); } /* * If there are existing mappings, either abort or remove them. */ oldpde = *pde; if ((oldpde & PG_V) != 0) { KASSERT(pdpg == NULL || pdpg->ref_count > 1, ("pmap_enter_pde: pdpg's reference count is too low")); if ((flags & PMAP_ENTER_NOREPLACE) != 0 && (va < VM_MAXUSER_ADDRESS || (oldpde & PG_PS) != 0 || !pmap_every_pte_zero(oldpde & PG_FRAME))) { if (pdpg != NULL) pdpg->ref_count--; CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (KERN_FAILURE); } /* Break the existing mapping(s). */ SLIST_INIT(&free); if ((oldpde & PG_PS) != 0) { /* * The reference to the PD page that was acquired by * pmap_alloc_pde() ensures that it won't be freed. * However, if the PDE resulted from a promotion, then * a reserved PT page could be freed. */ (void)pmap_remove_pde(pmap, pde, va, &free, lockp); if ((oldpde & PG_G) == 0) pmap_invalidate_pde_page(pmap, va, oldpde); } else { pmap_delayed_invl_start(); if (pmap_remove_ptes(pmap, va, va + NBPDR, pde, &free, lockp)) pmap_invalidate_all(pmap); pmap_delayed_invl_finish(); } if (va < VM_MAXUSER_ADDRESS) { vm_page_free_pages_toq(&free, true); KASSERT(*pde == 0, ("pmap_enter_pde: non-zero pde %p", pde)); } else { KASSERT(SLIST_EMPTY(&free), ("pmap_enter_pde: freed kernel page table page")); /* * Both pmap_remove_pde() and pmap_remove_ptes() will * leave the kernel page table page zero filled. */ mt = PHYS_TO_VM_PAGE(*pde & PG_FRAME); if (pmap_insert_pt_page(pmap, mt, false)) panic("pmap_enter_pde: trie insert failed"); } } if ((newpde & PG_MANAGED) != 0) { /* * Abort this mapping if its PV entry could not be created. */ if (!pmap_pv_insert_pde(pmap, va, newpde, flags, lockp)) { if (pdpg != NULL) pmap_abort_ptp(pmap, va, pdpg); CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (KERN_RESOURCE_SHORTAGE); } if ((newpde & PG_RW) != 0) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) vm_page_aflag_set(mt, PGA_WRITEABLE); } } /* * Increment counters. */ if ((newpde & PG_W) != 0) pmap->pm_stats.wired_count += NBPDR / PAGE_SIZE; pmap_resident_count_adj(pmap, NBPDR / PAGE_SIZE); /* * Map the superpage. (This is not a promoted mapping; there will not * be any lingering 4KB page mappings in the TLB.) */ pde_store(pde, newpde); counter_u64_add(pmap_pde_mappings, 1); CTR2(KTR_PMAP, "pmap_enter_pde: success for va %#lx in pmap %p", va, pmap); return (KERN_SUCCESS); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ void pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { struct rwlock *lock; vm_offset_t va; vm_page_t m, mpte; vm_pindex_t diff, psize; VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); mpte = NULL; m = m_start; lock = NULL; PMAP_LOCK(pmap); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { va = start + ptoa(diff); if ((va & PDRMASK) == 0 && va + NBPDR <= end && m->psind == 1 && pmap_ps_enabled(pmap) && pmap_enter_2mpage(pmap, va, m, prot, &lock)) m = &m[NBPDR / PAGE_SIZE - 1]; else mpte = pmap_enter_quick_locked(pmap, va, m, prot, mpte, &lock); m = TAILQ_NEXT(m, listq); } if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); } /* * this code makes some *MAJOR* assumptions: * 1. Current pmap & pmap exists. * 2. Not wired. * 3. Read access. * 4. No page table pages. * but is *MUCH* faster than pmap_enter... */ void pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { struct rwlock *lock; lock = NULL; PMAP_LOCK(pmap); (void)pmap_enter_quick_locked(pmap, va, m, prot, NULL, &lock); if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); } static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp) { pt_entry_t newpte, *pte, PG_V; KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva || (m->oflags & VPO_UNMANAGED) != 0, ("pmap_enter_quick_locked: managed mapping within the clean submap")); PG_V = pmap_valid_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * In the case that a page table page is not * resident, we are creating it here. */ if (va < VM_MAXUSER_ADDRESS) { vm_pindex_t ptepindex; pd_entry_t *ptepa; /* * Calculate pagetable page index */ ptepindex = pmap_pde_pindex(va); if (mpte && (mpte->pindex == ptepindex)) { mpte->ref_count++; } else { /* * Get the page directory entry */ ptepa = pmap_pde(pmap, va); /* * If the page table page is mapped, we just increment * the hold count, and activate it. Otherwise, we * attempt to allocate a page table page. If this * attempt fails, we don't retry. Instead, we give up. */ if (ptepa && (*ptepa & PG_V) != 0) { if (*ptepa & PG_PS) return (NULL); mpte = PHYS_TO_VM_PAGE(*ptepa & PG_FRAME); mpte->ref_count++; } else { /* * Pass NULL instead of the PV list lock * pointer, because we don't intend to sleep. */ mpte = pmap_allocpte_alloc(pmap, ptepindex, NULL, va); if (mpte == NULL) return (mpte); } } pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mpte)); pte = &pte[pmap_pte_index(va)]; } else { mpte = NULL; pte = vtopte(va); } if (*pte) { if (mpte != NULL) mpte->ref_count--; return (NULL); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0 && !pmap_try_insert_pv_entry(pmap, va, m, lockp)) { if (mpte != NULL) pmap_abort_ptp(pmap, va, mpte); return (NULL); } /* * Increment counters */ pmap_resident_count_adj(pmap, 1); newpte = VM_PAGE_TO_PHYS(m) | PG_V | pmap_cache_bits(pmap, m->md.pat_mode, 0); if ((m->oflags & VPO_UNMANAGED) == 0) newpte |= PG_MANAGED; if ((prot & VM_PROT_EXECUTE) == 0) newpte |= pg_nx; if (va < VM_MAXUSER_ADDRESS) newpte |= PG_U | pmap_pkru_get(pmap, va); pte_store(pte, newpte); return (mpte); } /* * Make a temporary mapping for a physical address. This is only intended * to be used for panic dumps. */ void * pmap_kenter_temporary(vm_paddr_t pa, int i) { vm_offset_t va; va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); pmap_kenter(va, pa); invlpg(va); return ((void *)crashdumpmap); } /* * This code maps large physical mmap regions into the * processor address space. Note that some shortcuts * are taken, but the code works. */ void pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size) { pd_entry_t *pde; pt_entry_t PG_A, PG_M, PG_RW, PG_V; vm_paddr_t pa, ptepa; vm_page_t p, pdpg; int pat_mode; PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, ("pmap_object_init_pt: non-device object")); if ((addr & (NBPDR - 1)) == 0 && (size & (NBPDR - 1)) == 0) { if (!pmap_ps_enabled(pmap)) return; if (!vm_object_populate(object, pindex, pindex + atop(size))) return; p = vm_page_lookup(object, pindex); KASSERT(p->valid == VM_PAGE_BITS_ALL, ("pmap_object_init_pt: invalid page %p", p)); pat_mode = p->md.pat_mode; /* * Abort the mapping if the first page is not physically * aligned to a 2MB page boundary. */ ptepa = VM_PAGE_TO_PHYS(p); if (ptepa & (NBPDR - 1)) return; /* * Skip the first page. Abort the mapping if the rest of * the pages are not physically contiguous or have differing * memory attributes. */ p = TAILQ_NEXT(p, listq); for (pa = ptepa + PAGE_SIZE; pa < ptepa + size; pa += PAGE_SIZE) { KASSERT(p->valid == VM_PAGE_BITS_ALL, ("pmap_object_init_pt: invalid page %p", p)); if (pa != VM_PAGE_TO_PHYS(p) || pat_mode != p->md.pat_mode) return; p = TAILQ_NEXT(p, listq); } /* * Map using 2MB pages. Since "ptepa" is 2M aligned and * "size" is a multiple of 2M, adding the PAT setting to "pa" * will not affect the termination of this loop. */ PMAP_LOCK(pmap); for (pa = ptepa | pmap_cache_bits(pmap, pat_mode, 1); pa < ptepa + size; pa += NBPDR) { pde = pmap_alloc_pde(pmap, addr, &pdpg, NULL); if (pde == NULL) { /* * The creation of mappings below is only an * optimization. If a page directory page * cannot be allocated without blocking, * continue on to the next mapping rather than * blocking. */ addr += NBPDR; continue; } if ((*pde & PG_V) == 0) { pde_store(pde, pa | PG_PS | PG_M | PG_A | PG_U | PG_RW | PG_V); pmap_resident_count_adj(pmap, NBPDR / PAGE_SIZE); counter_u64_add(pmap_pde_mappings, 1); } else { /* Continue on if the PDE is already valid. */ pdpg->ref_count--; KASSERT(pdpg->ref_count > 0, ("pmap_object_init_pt: missing reference " "to page directory page, va: 0x%lx", addr)); } addr += NBPDR; } PMAP_UNLOCK(pmap); } } /* * Clear the wired attribute from the mappings for the specified range of * addresses in the given pmap. Every valid mapping within that range * must have the wired attribute set. In contrast, invalid mappings * cannot have the wired attribute set, so they are ignored. * * The wired attribute of the page table entry is not a hardware * feature, so there is no need to invalidate any TLB entries. * Since pmap_demote_pde() for the wired entry must never fail, * pmap_delayed_invl_start()/finish() calls around the * function are not needed. */ void pmap_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t va_next; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte, PG_V, PG_G; PG_V = pmap_valid_bit(pmap); PG_G = pmap_global_bit(pmap); PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { pml4e = pmap_pml4e(pmap, sva); if (pml4e == NULL || (*pml4e & PG_V) == 0) { va_next = (sva + NBPML4) & ~PML4MASK; if (va_next < sva) va_next = eva; continue; } va_next = (sva + NBPDP) & ~PDPMASK; if (va_next < sva) va_next = eva; pdpe = pmap_pml4e_to_pdpe(pml4e, sva); if ((*pdpe & PG_V) == 0) continue; if ((*pdpe & PG_PS) != 0) { KASSERT(va_next <= eva, ("partial update of non-transparent 1G mapping " "pdpe %#lx sva %#lx eva %#lx va_next %#lx", *pdpe, sva, eva, va_next)); MPASS(pmap != kernel_pmap); /* XXXKIB */ MPASS((*pdpe & (PG_MANAGED | PG_G)) == 0); atomic_clear_long(pdpe, PG_W); pmap->pm_stats.wired_count -= NBPDP / PAGE_SIZE; continue; } va_next = (sva + NBPDR) & ~PDRMASK; if (va_next < sva) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, sva); if ((*pde & PG_V) == 0) continue; if ((*pde & PG_PS) != 0) { if ((*pde & PG_W) == 0) panic("pmap_unwire: pde %#jx is missing PG_W", (uintmax_t)*pde); /* * Are we unwiring the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == va_next && eva >= va_next) { atomic_clear_long(pde, PG_W); pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE; continue; } else if (!pmap_demote_pde(pmap, pde, sva)) panic("pmap_unwire: demotion failed"); } if (va_next > eva) va_next = eva; for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, sva += PAGE_SIZE) { if ((*pte & PG_V) == 0) continue; if ((*pte & PG_W) == 0) panic("pmap_unwire: pte %#jx is missing PG_W", (uintmax_t)*pte); /* * PG_W must be cleared atomically. Although the pmap * lock synchronizes access to PG_W, another processor * could be setting PG_M and/or PG_A concurrently. */ atomic_clear_long(pte, PG_W); pmap->pm_stats.wired_count--; } } PMAP_UNLOCK(pmap); } /* * Copy the range specified by src_addr/len * from the source map to the range dst_addr/len * in the destination map. * * This routine is only advisory and need not do anything. */ void pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, vm_offset_t src_addr) { struct rwlock *lock; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t *pde, srcptepaddr; pt_entry_t *dst_pte, PG_A, PG_M, PG_V, ptetemp, *src_pte; vm_offset_t addr, end_addr, va_next; vm_page_t dst_pdpg, dstmpte, srcmpte; if (dst_addr != src_addr) return; if (dst_pmap->pm_type != src_pmap->pm_type) return; /* * EPT page table entries that require emulation of A/D bits are * sensitive to clearing the PG_A bit (aka EPT_PG_READ). Although * we clear PG_M (aka EPT_PG_WRITE) concomitantly, the PG_U bit * (aka EPT_PG_EXECUTE) could still be set. Since some EPT * implementations flag an EPT misconfiguration for exec-only * mappings we skip this function entirely for emulated pmaps. */ if (pmap_emulate_ad_bits(dst_pmap)) return; end_addr = src_addr + len; lock = NULL; if (dst_pmap < src_pmap) { PMAP_LOCK(dst_pmap); PMAP_LOCK(src_pmap); } else { PMAP_LOCK(src_pmap); PMAP_LOCK(dst_pmap); } PG_A = pmap_accessed_bit(dst_pmap); PG_M = pmap_modified_bit(dst_pmap); PG_V = pmap_valid_bit(dst_pmap); for (addr = src_addr; addr < end_addr; addr = va_next) { KASSERT(addr < UPT_MIN_ADDRESS, ("pmap_copy: invalid to pmap_copy page tables")); pml4e = pmap_pml4e(src_pmap, addr); if (pml4e == NULL || (*pml4e & PG_V) == 0) { va_next = (addr + NBPML4) & ~PML4MASK; if (va_next < addr) va_next = end_addr; continue; } va_next = (addr + NBPDP) & ~PDPMASK; if (va_next < addr) va_next = end_addr; pdpe = pmap_pml4e_to_pdpe(pml4e, addr); if ((*pdpe & PG_V) == 0) continue; if ((*pdpe & PG_PS) != 0) { KASSERT(va_next <= end_addr, ("partial update of non-transparent 1G mapping " "pdpe %#lx sva %#lx eva %#lx va_next %#lx", *pdpe, addr, end_addr, va_next)); MPASS((addr & PDPMASK) == 0); MPASS((*pdpe & PG_MANAGED) == 0); srcptepaddr = *pdpe; pdpe = pmap_pdpe(dst_pmap, addr); if (pdpe == NULL) { if (pmap_allocpte_alloc(dst_pmap, pmap_pml4e_pindex(addr), NULL, addr) == NULL) break; pdpe = pmap_pdpe(dst_pmap, addr); } else { pml4e = pmap_pml4e(dst_pmap, addr); dst_pdpg = PHYS_TO_VM_PAGE(*pml4e & PG_FRAME); dst_pdpg->ref_count++; } KASSERT(*pdpe == 0, ("1G mapping present in dst pmap " "pdpe %#lx sva %#lx eva %#lx va_next %#lx", *pdpe, addr, end_addr, va_next)); *pdpe = srcptepaddr & ~PG_W; pmap_resident_count_adj(dst_pmap, NBPDP / PAGE_SIZE); continue; } va_next = (addr + NBPDR) & ~PDRMASK; if (va_next < addr) va_next = end_addr; pde = pmap_pdpe_to_pde(pdpe, addr); srcptepaddr = *pde; if (srcptepaddr == 0) continue; if (srcptepaddr & PG_PS) { /* * We can only virtual copy whole superpages. */ if ((addr & PDRMASK) != 0 || addr + NBPDR > end_addr) continue; pde = pmap_alloc_pde(dst_pmap, addr, &dst_pdpg, NULL); if (pde == NULL) break; if (*pde == 0 && ((srcptepaddr & PG_MANAGED) == 0 || pmap_pv_insert_pde(dst_pmap, addr, srcptepaddr, PMAP_ENTER_NORECLAIM, &lock))) { /* * We leave the dirty bit unchanged because * managed read/write superpage mappings are * required to be dirty. However, managed * superpage mappings are not required to * have their accessed bit set, so we clear * it because we don't know if this mapping * will be used. */ srcptepaddr &= ~PG_W; if ((srcptepaddr & PG_MANAGED) != 0) srcptepaddr &= ~PG_A; *pde = srcptepaddr; pmap_resident_count_adj(dst_pmap, NBPDR / PAGE_SIZE); counter_u64_add(pmap_pde_mappings, 1); } else pmap_abort_ptp(dst_pmap, addr, dst_pdpg); continue; } srcptepaddr &= PG_FRAME; srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); KASSERT(srcmpte->ref_count > 0, ("pmap_copy: source page table page is unused")); if (va_next > end_addr) va_next = end_addr; src_pte = (pt_entry_t *)PHYS_TO_DMAP(srcptepaddr); src_pte = &src_pte[pmap_pte_index(addr)]; dstmpte = NULL; for (; addr < va_next; addr += PAGE_SIZE, src_pte++) { ptetemp = *src_pte; /* * We only virtual copy managed pages. */ if ((ptetemp & PG_MANAGED) == 0) continue; if (dstmpte != NULL) { KASSERT(dstmpte->pindex == pmap_pde_pindex(addr), ("dstmpte pindex/addr mismatch")); dstmpte->ref_count++; } else if ((dstmpte = pmap_allocpte(dst_pmap, addr, NULL)) == NULL) goto out; dst_pte = (pt_entry_t *) PHYS_TO_DMAP(VM_PAGE_TO_PHYS(dstmpte)); dst_pte = &dst_pte[pmap_pte_index(addr)]; if (*dst_pte == 0 && pmap_try_insert_pv_entry(dst_pmap, addr, PHYS_TO_VM_PAGE(ptetemp & PG_FRAME), &lock)) { /* * Clear the wired, modified, and accessed * (referenced) bits during the copy. */ *dst_pte = ptetemp & ~(PG_W | PG_M | PG_A); pmap_resident_count_adj(dst_pmap, 1); } else { pmap_abort_ptp(dst_pmap, addr, dstmpte); goto out; } /* Have we copied all of the valid mappings? */ if (dstmpte->ref_count >= srcmpte->ref_count) break; } } out: if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(src_pmap); PMAP_UNLOCK(dst_pmap); } int pmap_vmspace_copy(pmap_t dst_pmap, pmap_t src_pmap) { int error; if (dst_pmap->pm_type != src_pmap->pm_type || dst_pmap->pm_type != PT_X86 || (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) == 0) return (0); for (;;) { if (dst_pmap < src_pmap) { PMAP_LOCK(dst_pmap); PMAP_LOCK(src_pmap); } else { PMAP_LOCK(src_pmap); PMAP_LOCK(dst_pmap); } error = pmap_pkru_copy(dst_pmap, src_pmap); /* Clean up partial copy on failure due to no memory. */ if (error == ENOMEM) pmap_pkru_deassign_all(dst_pmap); PMAP_UNLOCK(src_pmap); PMAP_UNLOCK(dst_pmap); if (error != ENOMEM) break; vm_wait(NULL); } return (error); } /* * Zero the specified hardware page. */ void pmap_zero_page(vm_page_t m) { vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); pagezero((void *)va); } /* * Zero an an area within a single hardware page. off and size must not * cover an area beyond a single hardware page. */ void pmap_zero_page_area(vm_page_t m, int off, int size) { vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); if (off == 0 && size == PAGE_SIZE) pagezero((void *)va); else bzero((char *)va + off, size); } /* * Copy 1 specified hardware page to another. */ void pmap_copy_page(vm_page_t msrc, vm_page_t mdst) { vm_offset_t src = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(msrc)); vm_offset_t dst = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mdst)); pagecopy((void *)src, (void *)dst); } int unmapped_buf_allowed = 1; void pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[], vm_offset_t b_offset, int xfersize) { void *a_cp, *b_cp; vm_page_t pages[2]; vm_offset_t vaddr[2], a_pg_offset, b_pg_offset; int cnt; boolean_t mapped; while (xfersize > 0) { a_pg_offset = a_offset & PAGE_MASK; pages[0] = ma[a_offset >> PAGE_SHIFT]; b_pg_offset = b_offset & PAGE_MASK; pages[1] = mb[b_offset >> PAGE_SHIFT]; cnt = min(xfersize, PAGE_SIZE - a_pg_offset); cnt = min(cnt, PAGE_SIZE - b_pg_offset); mapped = pmap_map_io_transient(pages, vaddr, 2, FALSE); a_cp = (char *)vaddr[0] + a_pg_offset; b_cp = (char *)vaddr[1] + b_pg_offset; bcopy(a_cp, b_cp, cnt); if (__predict_false(mapped)) pmap_unmap_io_transient(pages, vaddr, 2, FALSE); a_offset += cnt; b_offset += cnt; xfersize -= cnt; } } /* * Returns true if the pmap's pv is one of the first * 16 pvs linked to from this page. This count may * be changed upwards or downwards in the future; it * is only necessary that true be returned for a small * subset of pmaps for proper page aging. */ boolean_t pmap_page_exists_quick(pmap_t pmap, vm_page_t m) { struct md_page *pvh; struct rwlock *lock; pv_entry_t pv; int loops = 0; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_page_exists_quick: page %p is not managed", m)); rv = FALSE; lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } } rw_runlock(lock); return (rv); } /* * pmap_page_wired_mappings: * * Return the number of managed mappings to the given physical page * that are wired. */ int pmap_page_wired_mappings(vm_page_t m) { struct rwlock *lock; struct md_page *pvh; pmap_t pmap; pt_entry_t *pte; pv_entry_t pv; int count, md_gen, pvh_gen; if ((m->oflags & VPO_UNMANAGED) != 0) return (0); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); restart: count = 0; TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte(pmap, pv->pv_va); if ((*pte & PG_W) != 0) count++; PMAP_UNLOCK(pmap); } if ((m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen || pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pde(pmap, pv->pv_va); if ((*pte & PG_W) != 0) count++; PMAP_UNLOCK(pmap); } } rw_runlock(lock); return (count); } /* * Returns TRUE if the given page is mapped individually or as part of * a 2mpage. Otherwise, returns FALSE. */ boolean_t pmap_page_is_mapped(vm_page_t m) { struct rwlock *lock; boolean_t rv; if ((m->oflags & VPO_UNMANAGED) != 0) return (FALSE); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); rv = !TAILQ_EMPTY(&m->md.pv_list) || ((m->flags & PG_FICTITIOUS) == 0 && !TAILQ_EMPTY(&pa_to_pvh(VM_PAGE_TO_PHYS(m))->pv_list)); rw_runlock(lock); return (rv); } /* * Destroy all managed, non-wired mappings in the given user-space * pmap. This pmap cannot be active on any processor besides the * caller. * * This function cannot be applied to the kernel pmap. Moreover, it * is not intended for general use. It is only to be used during * process termination. Consequently, it can be implemented in ways * that make it faster than pmap_remove(). First, it can more quickly * destroy mappings by iterating over the pmap's collection of PV * entries, rather than searching the page table. Second, it doesn't * have to test and clear the page table entries atomically, because * no processor is currently accessing the user address space. In * particular, a page table entry's dirty bit won't change state once * this function starts. * * Although this function destroys all of the pmap's managed, * non-wired mappings, it can delay and batch the invalidation of TLB * entries without calling pmap_delayed_invl_start() and * pmap_delayed_invl_finish(). Because the pmap is not active on * any other processor, none of these TLB entries will ever be used * before their eventual invalidation. Consequently, there is no need * for either pmap_remove_all() or pmap_remove_write() to wait for * that eventual TLB invalidation. */ void pmap_remove_pages(pmap_t pmap) { pd_entry_t ptepde; pt_entry_t *pte, tpte; pt_entry_t PG_M, PG_RW, PG_V; struct spglist free; struct pv_chunklist free_chunks[PMAP_MEMDOM]; vm_page_t m, mpte, mt; pv_entry_t pv; struct md_page *pvh; struct pv_chunk *pc, *npc; struct rwlock *lock; int64_t bit; uint64_t inuse, bitmask; int allfree, field, freed, i, idx; boolean_t superpage; vm_paddr_t pa; /* * Assert that the given pmap is only active on the current * CPU. Unfortunately, we cannot block another CPU from * activating the pmap while this function is executing. */ KASSERT(pmap == PCPU_GET(curpmap), ("non-current pmap %p", pmap)); #ifdef INVARIANTS { cpuset_t other_cpus; other_cpus = all_cpus; critical_enter(); CPU_CLR(PCPU_GET(cpuid), &other_cpus); CPU_AND(&other_cpus, &pmap->pm_active); critical_exit(); KASSERT(CPU_EMPTY(&other_cpus), ("pmap active %p", pmap)); } #endif lock = NULL; PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); for (i = 0; i < PMAP_MEMDOM; i++) TAILQ_INIT(&free_chunks[i]); SLIST_INIT(&free); PMAP_LOCK(pmap); TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) { allfree = 1; freed = 0; for (field = 0; field < _NPCM; field++) { inuse = ~pc->pc_map[field] & pc_freemask[field]; while (inuse != 0) { bit = bsfq(inuse); bitmask = 1UL << bit; idx = field * 64 + bit; pv = &pc->pc_pventry[idx]; inuse &= ~bitmask; pte = pmap_pdpe(pmap, pv->pv_va); ptepde = *pte; pte = pmap_pdpe_to_pde(pte, pv->pv_va); tpte = *pte; if ((tpte & (PG_PS | PG_V)) == PG_V) { superpage = FALSE; ptepde = tpte; pte = (pt_entry_t *)PHYS_TO_DMAP(tpte & PG_FRAME); pte = &pte[pmap_pte_index(pv->pv_va)]; tpte = *pte; } else { /* * Keep track whether 'tpte' is a * superpage explicitly instead of * relying on PG_PS being set. * * This is because PG_PS is numerically * identical to PG_PTE_PAT and thus a * regular page could be mistaken for * a superpage. */ superpage = TRUE; } if ((tpte & PG_V) == 0) { panic("bad pte va %lx pte %lx", pv->pv_va, tpte); } /* * We cannot remove wired pages from a process' mapping at this time */ if (tpte & PG_W) { allfree = 0; continue; } /* Mark free */ pc->pc_map[field] |= bitmask; /* * Because this pmap is not active on other * processors, the dirty bit cannot have * changed state since we last loaded pte. */ pte_clear(pte); if (superpage) pa = tpte & PG_PS_FRAME; else pa = tpte & PG_FRAME; m = PHYS_TO_VM_PAGE(pa); KASSERT(m->phys_addr == pa, ("vm_page_t %p phys_addr mismatch %016jx %016jx", m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); KASSERT((m->flags & PG_FICTITIOUS) != 0 || m < &vm_page_array[vm_page_array_size], ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); /* * Update the vm_page_t clean/reference bits. */ if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if (superpage) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) vm_page_dirty(mt); } else vm_page_dirty(m); } CHANGE_PV_LIST_LOCK_TO_VM_PAGE(&lock, m); if (superpage) { pmap_resident_count_adj(pmap, -NBPDR / PAGE_SIZE); pvh = pa_to_pvh(tpte & PG_PS_FRAME); TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; if (TAILQ_EMPTY(&pvh->pv_list)) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) if ((mt->a.flags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&mt->md.pv_list)) vm_page_aflag_clear(mt, PGA_WRITEABLE); } mpte = pmap_remove_pt_page(pmap, pv->pv_va); if (mpte != NULL) { KASSERT(mpte->valid == VM_PAGE_BITS_ALL, ("pmap_remove_pages: pte page not promoted")); pmap_resident_count_adj(pmap, -1); KASSERT(mpte->ref_count == NPTEPG, ("pmap_remove_pages: pte page reference count error")); mpte->ref_count = 0; pmap_add_delayed_free_list(mpte, &free, FALSE); } } else { pmap_resident_count_adj(pmap, -1); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if ((m->a.flags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } } pmap_unuse_pt(pmap, pv->pv_va, ptepde, &free); freed++; } } PV_STAT(counter_u64_add(pv_entry_frees, freed)); PV_STAT(counter_u64_add(pv_entry_spare, freed)); PV_STAT(counter_u64_add(pv_entry_count, -freed)); if (allfree) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&free_chunks[pc_to_domain(pc)], pc, pc_list); } } if (lock != NULL) rw_wunlock(lock); pmap_invalidate_all(pmap); pmap_pkru_deassign_all(pmap); free_pv_chunk_batch((struct pv_chunklist *)&free_chunks); PMAP_UNLOCK(pmap); vm_page_free_pages_toq(&free, true); } static boolean_t pmap_page_test_mappings(vm_page_t m, boolean_t accessed, boolean_t modified) { struct rwlock *lock; pv_entry_t pv; struct md_page *pvh; pt_entry_t *pte, mask; pt_entry_t PG_A, PG_M, PG_RW, PG_V; pmap_t pmap; int md_gen, pvh_gen; boolean_t rv; rv = FALSE; lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); restart: TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte(pmap, pv->pv_va); mask = 0; if (modified) { PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); mask |= PG_RW | PG_M; } if (accessed) { PG_A = pmap_accessed_bit(pmap); PG_V = pmap_valid_bit(pmap); mask |= PG_V | PG_A; } rv = (*pte & mask) == mask; PMAP_UNLOCK(pmap); if (rv) goto out; } if ((m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen || pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pde(pmap, pv->pv_va); mask = 0; if (modified) { PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); mask |= PG_RW | PG_M; } if (accessed) { PG_A = pmap_accessed_bit(pmap); PG_V = pmap_valid_bit(pmap); mask |= PG_V | PG_A; } rv = (*pte & mask) == mask; PMAP_UNLOCK(pmap); if (rv) goto out; } } out: rw_runlock(lock); return (rv); } /* * pmap_is_modified: * * Return whether or not the specified physical page was modified * in any physical maps. */ boolean_t pmap_is_modified(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_modified: page %p is not managed", m)); /* * If the page is not busied then this check is racy. */ if (!pmap_page_is_write_mapped(m)) return (FALSE); return (pmap_page_test_mappings(m, FALSE, TRUE)); } /* * pmap_is_prefaultable: * * Return whether or not the specified virtual address is eligible * for prefault. */ boolean_t pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) { pd_entry_t *pde; pt_entry_t *pte, PG_V; boolean_t rv; PG_V = pmap_valid_bit(pmap); rv = FALSE; PMAP_LOCK(pmap); pde = pmap_pde(pmap, addr); if (pde != NULL && (*pde & (PG_PS | PG_V)) == PG_V) { pte = pmap_pde_to_pte(pde, addr); rv = (*pte & PG_V) == 0; } PMAP_UNLOCK(pmap); return (rv); } /* * pmap_is_referenced: * * Return whether or not the specified physical page was referenced * in any physical maps. */ boolean_t pmap_is_referenced(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_referenced: page %p is not managed", m)); return (pmap_page_test_mappings(m, TRUE, FALSE)); } /* * Clear the write and modified bits in each of the given page's mappings. */ void pmap_remove_write(vm_page_t m) { struct md_page *pvh; pmap_t pmap; struct rwlock *lock; pv_entry_t next_pv, pv; pd_entry_t *pde; pt_entry_t oldpte, *pte, PG_M, PG_RW; vm_offset_t va; int pvh_gen, md_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_write: page %p is not managed", m)); vm_page_assert_busied(m); if (!pmap_page_is_write_mapped(m)) return; lock = VM_PAGE_TO_PV_LIST_LOCK(m); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(VM_PAGE_TO_PHYS(m)); rw_wlock(lock); retry: TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } PG_RW = pmap_rw_bit(pmap); va = pv->pv_va; pde = pmap_pde(pmap, va); if ((*pde & PG_RW) != 0) (void)pmap_demote_pde_locked(pmap, pde, va, &lock); KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); PMAP_UNLOCK(pmap); } TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_remove_write: found a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); oldpte = *pte; if (oldpte & PG_RW) { while (!atomic_fcmpset_long(pte, &oldpte, oldpte & ~(PG_RW | PG_M))) cpu_spinwait(); if ((oldpte & PG_M) != 0) vm_page_dirty(m); pmap_invalidate_page(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } rw_wunlock(lock); vm_page_aflag_clear(m, PGA_WRITEABLE); pmap_delayed_invl_wait(m); } static __inline boolean_t safe_to_clear_referenced(pmap_t pmap, pt_entry_t pte) { if (!pmap_emulate_ad_bits(pmap)) return (TRUE); KASSERT(pmap->pm_type == PT_EPT, ("invalid pm_type %d", pmap->pm_type)); /* * XWR = 010 or 110 will cause an unconditional EPT misconfiguration * so we don't let the referenced (aka EPT_PG_READ) bit to be cleared * if the EPT_PG_WRITE bit is set. */ if ((pte & EPT_PG_WRITE) != 0) return (FALSE); /* * XWR = 100 is allowed only if the PMAP_SUPPORTS_EXEC_ONLY is set. */ if ((pte & EPT_PG_EXECUTE) == 0 || ((pmap->pm_flags & PMAP_SUPPORTS_EXEC_ONLY) != 0)) return (TRUE); else return (FALSE); } /* * pmap_ts_referenced: * * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. * * As an optimization, update the page's dirty field if a modified bit is * found while counting reference bits. This opportunistic update can be * performed at low cost and can eliminate the need for some future calls * to pmap_is_modified(). However, since this function stops after * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some * dirty pages. Those dirty pages will only be detected by a future call * to pmap_is_modified(). * * A DI block is not needed within this function, because * invalidations are performed before the PV list lock is * released. */ int pmap_ts_referenced(vm_page_t m) { struct md_page *pvh; pv_entry_t pv, pvf; pmap_t pmap; struct rwlock *lock; pd_entry_t oldpde, *pde; pt_entry_t *pte, PG_A, PG_M, PG_RW; vm_offset_t va; vm_paddr_t pa; int cleared, md_gen, not_cleared, pvh_gen; struct spglist free; boolean_t demoted; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_ts_referenced: page %p is not managed", m)); SLIST_INIT(&free); cleared = 0; pa = VM_PAGE_TO_PHYS(m); lock = PHYS_TO_PV_LIST_LOCK(pa); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(pa); rw_wlock(lock); retry: not_cleared = 0; if ((pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL) goto small_mappings; pv = pvf; do { if (pvf == NULL) pvf = pv; pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); va = pv->pv_va; pde = pmap_pde(pmap, pv->pv_va); oldpde = *pde; if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) { /* * Although "oldpde" is mapping a 2MB page, because * this function is called at a 4KB page granularity, * we only update the 4KB page under test. */ vm_page_dirty(m); } if ((oldpde & PG_A) != 0) { /* * Since this reference bit is shared by 512 4KB * pages, it should not be cleared every time it is * tested. Apply a simple "hash" function on the * physical page number, the virtual superpage number, * and the pmap address to select one 4KB page out of * the 512 on which testing the reference bit will * result in clearing that reference bit. This * function is designed to avoid the selection of the * same 4KB page for every 2MB page mapping. * * On demotion, a mapping that hasn't been referenced * is simply destroyed. To avoid the possibility of a * subsequent page fault on a demoted wired mapping, * always leave its reference bit set. Moreover, * since the superpage is wired, the current state of * its reference bit won't affect page replacement. */ if ((((pa >> PAGE_SHIFT) ^ (pv->pv_va >> PDRSHIFT) ^ (uintptr_t)pmap) & (NPTEPG - 1)) == 0 && (oldpde & PG_W) == 0) { if (safe_to_clear_referenced(pmap, oldpde)) { atomic_clear_long(pde, PG_A); pmap_invalidate_page(pmap, pv->pv_va); demoted = FALSE; } else if (pmap_demote_pde_locked(pmap, pde, pv->pv_va, &lock)) { /* * Remove the mapping to a single page * so that a subsequent access may * repromote. Since the underlying * page table page is fully populated, * this removal never frees a page * table page. */ demoted = TRUE; va += VM_PAGE_TO_PHYS(m) - (oldpde & PG_PS_FRAME); pte = pmap_pde_to_pte(pde, va); pmap_remove_pte(pmap, pte, va, *pde, NULL, &lock); pmap_invalidate_page(pmap, va); } else demoted = TRUE; if (demoted) { /* * The superpage mapping was removed * entirely and therefore 'pv' is no * longer valid. */ if (pvf == pv) pvf = NULL; pv = NULL; } cleared++; KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); } else not_cleared++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; } if (cleared + not_cleared >= PMAP_TS_REFERENCED_MAX) goto out; } while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf); small_mappings: if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL) goto out; pv = pvf; do { if (pvf == NULL) pvf = pv; pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_ts_referenced: found a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if ((*pte & PG_A) != 0) { if (safe_to_clear_referenced(pmap, *pte)) { atomic_clear_long(pte, PG_A); pmap_invalidate_page(pmap, pv->pv_va); cleared++; } else if ((*pte & PG_W) == 0) { /* * Wired pages cannot be paged out so * doing accessed bit emulation for * them is wasted effort. We do the * hard work for unwired pages only. */ pmap_remove_pte(pmap, pte, pv->pv_va, *pde, &free, &lock); pmap_invalidate_page(pmap, pv->pv_va); cleared++; if (pvf == pv) pvf = NULL; pv = NULL; KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); } else not_cleared++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; } } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && cleared + not_cleared < PMAP_TS_REFERENCED_MAX); out: rw_wunlock(lock); vm_page_free_pages_toq(&free, true); return (cleared + not_cleared); } /* * Apply the given advice to the specified range of addresses within the * given pmap. Depending on the advice, clear the referenced and/or * modified flags in each mapping and set the mapped page's dirty field. */ void pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice) { struct rwlock *lock; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t oldpde, *pde; pt_entry_t *pte, PG_A, PG_G, PG_M, PG_RW, PG_V; vm_offset_t va, va_next; vm_page_t m; bool anychanged; if (advice != MADV_DONTNEED && advice != MADV_FREE) return; /* * A/D bit emulation requires an alternate code path when clearing * the modified and accessed bits below. Since this function is * advisory in nature we skip it entirely for pmaps that require * A/D bit emulation. */ if (pmap_emulate_ad_bits(pmap)) return; PG_A = pmap_accessed_bit(pmap); PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); anychanged = false; pmap_delayed_invl_start(); PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { pml4e = pmap_pml4e(pmap, sva); if (pml4e == NULL || (*pml4e & PG_V) == 0) { va_next = (sva + NBPML4) & ~PML4MASK; if (va_next < sva) va_next = eva; continue; } va_next = (sva + NBPDP) & ~PDPMASK; if (va_next < sva) va_next = eva; pdpe = pmap_pml4e_to_pdpe(pml4e, sva); if ((*pdpe & PG_V) == 0) continue; if ((*pdpe & PG_PS) != 0) { KASSERT(va_next <= eva, ("partial update of non-transparent 1G mapping " "pdpe %#lx sva %#lx eva %#lx va_next %#lx", *pdpe, sva, eva, va_next)); continue; } va_next = (sva + NBPDR) & ~PDRMASK; if (va_next < sva) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, sva); oldpde = *pde; if ((oldpde & PG_V) == 0) continue; else if ((oldpde & PG_PS) != 0) { if ((oldpde & PG_MANAGED) == 0) continue; lock = NULL; if (!pmap_demote_pde_locked(pmap, pde, sva, &lock)) { if (lock != NULL) rw_wunlock(lock); /* * The large page mapping was destroyed. */ continue; } /* * Unless the page mappings are wired, remove the * mapping to a single page so that a subsequent * access may repromote. Choosing the last page * within the address range [sva, min(va_next, eva)) * generally results in more repromotions. Since the * underlying page table page is fully populated, this * removal never frees a page table page. */ if ((oldpde & PG_W) == 0) { va = eva; if (va > va_next) va = va_next; va -= PAGE_SIZE; KASSERT(va >= sva, ("pmap_advise: no address gap")); pte = pmap_pde_to_pte(pde, va); KASSERT((*pte & PG_V) != 0, ("pmap_advise: invalid PTE")); pmap_remove_pte(pmap, pte, va, *pde, NULL, &lock); anychanged = true; } if (lock != NULL) rw_wunlock(lock); } if (va_next > eva) va_next = eva; va = va_next; for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, sva += PAGE_SIZE) { if ((*pte & (PG_MANAGED | PG_V)) != (PG_MANAGED | PG_V)) goto maybe_invlrng; else if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if (advice == MADV_DONTNEED) { /* * Future calls to pmap_is_modified() * can be avoided by making the page * dirty now. */ m = PHYS_TO_VM_PAGE(*pte & PG_FRAME); vm_page_dirty(m); } atomic_clear_long(pte, PG_M | PG_A); } else if ((*pte & PG_A) != 0) atomic_clear_long(pte, PG_A); else goto maybe_invlrng; if ((*pte & PG_G) != 0) { if (va == va_next) va = sva; } else anychanged = true; continue; maybe_invlrng: if (va != va_next) { pmap_invalidate_range(pmap, va, sva); va = va_next; } } if (va != va_next) pmap_invalidate_range(pmap, va, sva); } if (anychanged) pmap_invalidate_all(pmap); PMAP_UNLOCK(pmap); pmap_delayed_invl_finish(); } /* * Clear the modify bits on the specified physical page. */ void pmap_clear_modify(vm_page_t m) { struct md_page *pvh; pmap_t pmap; pv_entry_t next_pv, pv; pd_entry_t oldpde, *pde; pt_entry_t *pte, PG_M, PG_RW; struct rwlock *lock; vm_offset_t va; int md_gen, pvh_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_clear_modify: page %p is not managed", m)); vm_page_assert_busied(m); if (!pmap_page_is_write_mapped(m)) return; pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(VM_PAGE_TO_PHYS(m)); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_wlock(lock); restart: TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); va = pv->pv_va; pde = pmap_pde(pmap, va); oldpde = *pde; /* If oldpde has PG_RW set, then it also has PG_M set. */ if ((oldpde & PG_RW) != 0 && pmap_demote_pde_locked(pmap, pde, va, &lock) && (oldpde & PG_W) == 0) { /* * Write protect the mapping to a single page so that * a subsequent write access may repromote. */ va += VM_PAGE_TO_PHYS(m) - (oldpde & PG_PS_FRAME); pte = pmap_pde_to_pte(pde, va); atomic_clear_long(pte, PG_M | PG_RW); vm_page_dirty(m); pmap_invalidate_page(pmap, va); } PMAP_UNLOCK(pmap); } TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_clear_modify: found" " a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { atomic_clear_long(pte, PG_M); pmap_invalidate_page(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } rw_wunlock(lock); } /* * Miscellaneous support routines follow */ /* Adjust the properties for a leaf page table entry. */ static __inline void pmap_pte_props(pt_entry_t *pte, u_long bits, u_long mask) { u_long opte, npte; opte = *(u_long *)pte; do { npte = opte & ~mask; npte |= bits; } while (npte != opte && !atomic_fcmpset_long((u_long *)pte, &opte, npte)); } /* * Map a set of physical memory pages into the kernel virtual * address space. Return a pointer to where it is mapped. This * routine is intended to be used for mapping device memory, * NOT real memory. */ static void * pmap_mapdev_internal(vm_paddr_t pa, vm_size_t size, int mode, int flags) { struct pmap_preinit_mapping *ppim; vm_offset_t va, offset; vm_size_t tmpsize; int i; offset = pa & PAGE_MASK; size = round_page(offset + size); pa = trunc_page(pa); if (!pmap_initialized) { va = 0; for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == 0) { ppim->pa = pa; ppim->sz = size; ppim->mode = mode; ppim->va = virtual_avail; virtual_avail += size; va = ppim->va; break; } } if (va == 0) panic("%s: too many preinit mappings", __func__); } else { /* * If we have a preinit mapping, re-use it. */ for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->pa == pa && ppim->sz == size && (ppim->mode == mode || (flags & MAPDEV_SETATTR) == 0)) return ((void *)(ppim->va + offset)); } /* * If the specified range of physical addresses fits within * the direct map window, use the direct map. */ if (pa < dmaplimit && pa + size <= dmaplimit) { va = PHYS_TO_DMAP(pa); if ((flags & MAPDEV_SETATTR) != 0) { PMAP_LOCK(kernel_pmap); i = pmap_change_props_locked(va, size, PROT_NONE, mode, flags); PMAP_UNLOCK(kernel_pmap); } else i = 0; if (!i) return ((void *)(va + offset)); } va = kva_alloc(size); if (va == 0) panic("%s: Couldn't allocate KVA", __func__); } for (tmpsize = 0; tmpsize < size; tmpsize += PAGE_SIZE) pmap_kenter_attr(va + tmpsize, pa + tmpsize, mode); pmap_invalidate_range(kernel_pmap, va, va + tmpsize); if ((flags & MAPDEV_FLUSHCACHE) != 0) pmap_invalidate_cache_range(va, va + tmpsize); return ((void *)(va + offset)); } void * pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode) { return (pmap_mapdev_internal(pa, size, mode, MAPDEV_FLUSHCACHE | MAPDEV_SETATTR)); } void * pmap_mapdev(vm_paddr_t pa, vm_size_t size) { return (pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE)); } void * pmap_mapdev_pciecfg(vm_paddr_t pa, vm_size_t size) { return (pmap_mapdev_internal(pa, size, PAT_UNCACHEABLE, MAPDEV_SETATTR)); } void * pmap_mapbios(vm_paddr_t pa, vm_size_t size) { return (pmap_mapdev_internal(pa, size, PAT_WRITE_BACK, MAPDEV_FLUSHCACHE)); } void pmap_unmapdev(vm_offset_t va, vm_size_t size) { struct pmap_preinit_mapping *ppim; vm_offset_t offset; int i; /* If we gave a direct map region in pmap_mapdev, do nothing */ if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) return; offset = va & PAGE_MASK; size = round_page(offset + size); va = trunc_page(va); for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == va && ppim->sz == size) { if (pmap_initialized) return; ppim->pa = 0; ppim->va = 0; ppim->sz = 0; ppim->mode = 0; if (va + size == virtual_avail) virtual_avail = va; return; } } if (pmap_initialized) { pmap_qremove(va, atop(size)); kva_free(va, size); } } /* * Tries to demote a 1GB page mapping. */ static boolean_t pmap_demote_pdpe(pmap_t pmap, pdp_entry_t *pdpe, vm_offset_t va) { pdp_entry_t newpdpe, oldpdpe; pd_entry_t *firstpde, newpde, *pde; pt_entry_t PG_A, PG_M, PG_RW, PG_V; vm_paddr_t pdpgpa; vm_page_t pdpg; PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldpdpe = *pdpe; KASSERT((oldpdpe & (PG_PS | PG_V)) == (PG_PS | PG_V), ("pmap_demote_pdpe: oldpdpe is missing PG_PS and/or PG_V")); pdpg = pmap_alloc_pt_page(pmap, va >> PDPSHIFT, VM_ALLOC_WIRED | VM_ALLOC_INTERRUPT); if (pdpg == NULL) { CTR2(KTR_PMAP, "pmap_demote_pdpe: failure for va %#lx" " in pmap %p", va, pmap); return (FALSE); } pdpgpa = VM_PAGE_TO_PHYS(pdpg); firstpde = (pd_entry_t *)PHYS_TO_DMAP(pdpgpa); newpdpe = pdpgpa | PG_M | PG_A | (oldpdpe & PG_U) | PG_RW | PG_V; KASSERT((oldpdpe & PG_A) != 0, ("pmap_demote_pdpe: oldpdpe is missing PG_A")); KASSERT((oldpdpe & (PG_M | PG_RW)) != PG_RW, ("pmap_demote_pdpe: oldpdpe is missing PG_M")); newpde = oldpdpe; /* * Initialize the page directory page. */ for (pde = firstpde; pde < firstpde + NPDEPG; pde++) { *pde = newpde; newpde += NBPDR; } /* * Demote the mapping. */ *pdpe = newpdpe; /* * Invalidate a stale recursive mapping of the page directory page. */ pmap_invalidate_page(pmap, (vm_offset_t)vtopde(va)); counter_u64_add(pmap_pdpe_demotions, 1); CTR2(KTR_PMAP, "pmap_demote_pdpe: success for va %#lx" " in pmap %p", va, pmap); return (TRUE); } /* * Sets the memory attribute for the specified page. */ void pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma) { m->md.pat_mode = ma; /* * If "m" is a normal page, update its direct mapping. This update * can be relied upon to perform any cache operations that are * required for data coherence. */ if ((m->flags & PG_FICTITIOUS) == 0 && pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE, m->md.pat_mode)) panic("memory attribute change on the direct map failed"); } void pmap_page_set_memattr_noflush(vm_page_t m, vm_memattr_t ma) { int error; m->md.pat_mode = ma; if ((m->flags & PG_FICTITIOUS) != 0) return; PMAP_LOCK(kernel_pmap); error = pmap_change_props_locked(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE, PROT_NONE, m->md.pat_mode, 0); PMAP_UNLOCK(kernel_pmap); if (error != 0) panic("memory attribute change on the direct map failed"); } /* * Changes the specified virtual address range's memory type to that given by * the parameter "mode". The specified virtual address range must be * completely contained within either the direct map or the kernel map. If * the virtual address range is contained within the kernel map, then the * memory type for each of the corresponding ranges of the direct map is also * changed. (The corresponding ranges of the direct map are those ranges that * map the same physical pages as the specified virtual address range.) These * changes to the direct map are necessary because Intel describes the * behavior of their processors as "undefined" if two or more mappings to the * same physical page have different memory types. * * Returns zero if the change completed successfully, and either EINVAL or * ENOMEM if the change failed. Specifically, EINVAL is returned if some part * of the virtual address range was not mapped, and ENOMEM is returned if * there was insufficient memory available to complete the change. In the * latter case, the memory type may have been changed on some part of the * virtual address range or the direct map. */ int pmap_change_attr(vm_offset_t va, vm_size_t size, int mode) { int error; PMAP_LOCK(kernel_pmap); error = pmap_change_props_locked(va, size, PROT_NONE, mode, MAPDEV_FLUSHCACHE); PMAP_UNLOCK(kernel_pmap); return (error); } /* * Changes the specified virtual address range's protections to those * specified by "prot". Like pmap_change_attr(), protections for aliases * in the direct map are updated as well. Protections on aliasing mappings may * be a subset of the requested protections; for example, mappings in the direct * map are never executable. */ int pmap_change_prot(vm_offset_t va, vm_size_t size, vm_prot_t prot) { int error; /* Only supported within the kernel map. */ if (va < VM_MIN_KERNEL_ADDRESS) return (EINVAL); PMAP_LOCK(kernel_pmap); error = pmap_change_props_locked(va, size, prot, -1, MAPDEV_ASSERTVALID); PMAP_UNLOCK(kernel_pmap); return (error); } static int pmap_change_props_locked(vm_offset_t va, vm_size_t size, vm_prot_t prot, int mode, int flags) { vm_offset_t base, offset, tmpva; vm_paddr_t pa_start, pa_end, pa_end1; pdp_entry_t *pdpe; pd_entry_t *pde, pde_bits, pde_mask; pt_entry_t *pte, pte_bits, pte_mask; int error; bool changed; PMAP_LOCK_ASSERT(kernel_pmap, MA_OWNED); base = trunc_page(va); offset = va & PAGE_MASK; size = round_page(offset + size); /* * Only supported on kernel virtual addresses, including the direct * map but excluding the recursive map. */ if (base < DMAP_MIN_ADDRESS) return (EINVAL); /* * Construct our flag sets and masks. "bits" is the subset of * "mask" that will be set in each modified PTE. * * Mappings in the direct map are never allowed to be executable. */ pde_bits = pte_bits = 0; pde_mask = pte_mask = 0; if (mode != -1) { pde_bits |= pmap_cache_bits(kernel_pmap, mode, true); pde_mask |= X86_PG_PDE_CACHE; pte_bits |= pmap_cache_bits(kernel_pmap, mode, false); pte_mask |= X86_PG_PTE_CACHE; } if (prot != VM_PROT_NONE) { if ((prot & VM_PROT_WRITE) != 0) { pde_bits |= X86_PG_RW; pte_bits |= X86_PG_RW; } if ((prot & VM_PROT_EXECUTE) == 0 || va < VM_MIN_KERNEL_ADDRESS) { pde_bits |= pg_nx; pte_bits |= pg_nx; } pde_mask |= X86_PG_RW | pg_nx; pte_mask |= X86_PG_RW | pg_nx; } /* * Pages that aren't mapped aren't supported. Also break down 2MB pages * into 4KB pages if required. */ for (tmpva = base; tmpva < base + size; ) { pdpe = pmap_pdpe(kernel_pmap, tmpva); if (pdpe == NULL || *pdpe == 0) { KASSERT((flags & MAPDEV_ASSERTVALID) == 0, ("%s: addr %#lx is not mapped", __func__, tmpva)); return (EINVAL); } if (*pdpe & PG_PS) { /* * If the current 1GB page already has the required * properties, then we need not demote this page. Just * increment tmpva to the next 1GB page frame. */ if ((*pdpe & pde_mask) == pde_bits) { tmpva = trunc_1gpage(tmpva) + NBPDP; continue; } /* * If the current offset aligns with a 1GB page frame * and there is at least 1GB left within the range, then * we need not break down this page into 2MB pages. */ if ((tmpva & PDPMASK) == 0 && tmpva + PDPMASK < base + size) { tmpva += NBPDP; continue; } if (!pmap_demote_pdpe(kernel_pmap, pdpe, tmpva)) return (ENOMEM); } pde = pmap_pdpe_to_pde(pdpe, tmpva); if (*pde == 0) { KASSERT((flags & MAPDEV_ASSERTVALID) == 0, ("%s: addr %#lx is not mapped", __func__, tmpva)); return (EINVAL); } if (*pde & PG_PS) { /* * If the current 2MB page already has the required * properties, then we need not demote this page. Just * increment tmpva to the next 2MB page frame. */ if ((*pde & pde_mask) == pde_bits) { tmpva = trunc_2mpage(tmpva) + NBPDR; continue; } /* * If the current offset aligns with a 2MB page frame * and there is at least 2MB left within the range, then * we need not break down this page into 4KB pages. */ if ((tmpva & PDRMASK) == 0 && tmpva + PDRMASK < base + size) { tmpva += NBPDR; continue; } if (!pmap_demote_pde(kernel_pmap, pde, tmpva)) return (ENOMEM); } pte = pmap_pde_to_pte(pde, tmpva); if (*pte == 0) { KASSERT((flags & MAPDEV_ASSERTVALID) == 0, ("%s: addr %#lx is not mapped", __func__, tmpva)); return (EINVAL); } tmpva += PAGE_SIZE; } error = 0; /* * Ok, all the pages exist, so run through them updating their * properties if required. */ changed = false; pa_start = pa_end = 0; for (tmpva = base; tmpva < base + size; ) { pdpe = pmap_pdpe(kernel_pmap, tmpva); if (*pdpe & PG_PS) { if ((*pdpe & pde_mask) != pde_bits) { pmap_pte_props(pdpe, pde_bits, pde_mask); changed = true; } if (tmpva >= VM_MIN_KERNEL_ADDRESS && (*pdpe & PG_PS_FRAME) < dmaplimit) { if (pa_start == pa_end) { /* Start physical address run. */ pa_start = *pdpe & PG_PS_FRAME; pa_end = pa_start + NBPDP; } else if (pa_end == (*pdpe & PG_PS_FRAME)) pa_end += NBPDP; else { /* Run ended, update direct map. */ error = pmap_change_props_locked( PHYS_TO_DMAP(pa_start), pa_end - pa_start, prot, mode, flags); if (error != 0) break; /* Start physical address run. */ pa_start = *pdpe & PG_PS_FRAME; pa_end = pa_start + NBPDP; } } tmpva = trunc_1gpage(tmpva) + NBPDP; continue; } pde = pmap_pdpe_to_pde(pdpe, tmpva); if (*pde & PG_PS) { if ((*pde & pde_mask) != pde_bits) { pmap_pte_props(pde, pde_bits, pde_mask); changed = true; } if (tmpva >= VM_MIN_KERNEL_ADDRESS && (*pde & PG_PS_FRAME) < dmaplimit) { if (pa_start == pa_end) { /* Start physical address run. */ pa_start = *pde & PG_PS_FRAME; pa_end = pa_start + NBPDR; } else if (pa_end == (*pde & PG_PS_FRAME)) pa_end += NBPDR; else { /* Run ended, update direct map. */ error = pmap_change_props_locked( PHYS_TO_DMAP(pa_start), pa_end - pa_start, prot, mode, flags); if (error != 0) break; /* Start physical address run. */ pa_start = *pde & PG_PS_FRAME; pa_end = pa_start + NBPDR; } } tmpva = trunc_2mpage(tmpva) + NBPDR; } else { pte = pmap_pde_to_pte(pde, tmpva); if ((*pte & pte_mask) != pte_bits) { pmap_pte_props(pte, pte_bits, pte_mask); changed = true; } if (tmpva >= VM_MIN_KERNEL_ADDRESS && (*pte & PG_FRAME) < dmaplimit) { if (pa_start == pa_end) { /* Start physical address run. */ pa_start = *pte & PG_FRAME; pa_end = pa_start + PAGE_SIZE; } else if (pa_end == (*pte & PG_FRAME)) pa_end += PAGE_SIZE; else { /* Run ended, update direct map. */ error = pmap_change_props_locked( PHYS_TO_DMAP(pa_start), pa_end - pa_start, prot, mode, flags); if (error != 0) break; /* Start physical address run. */ pa_start = *pte & PG_FRAME; pa_end = pa_start + PAGE_SIZE; } } tmpva += PAGE_SIZE; } } if (error == 0 && pa_start != pa_end && pa_start < dmaplimit) { pa_end1 = MIN(pa_end, dmaplimit); if (pa_start != pa_end1) error = pmap_change_props_locked(PHYS_TO_DMAP(pa_start), pa_end1 - pa_start, prot, mode, flags); } /* * Flush CPU caches if required to make sure any data isn't cached that * shouldn't be, etc. */ if (changed) { pmap_invalidate_range(kernel_pmap, base, tmpva); if ((flags & MAPDEV_FLUSHCACHE) != 0) pmap_invalidate_cache_range(base, tmpva); } return (error); } /* * Demotes any mapping within the direct map region that covers more than the * specified range of physical addresses. This range's size must be a power * of two and its starting address must be a multiple of its size. Since the * demotion does not change any attributes of the mapping, a TLB invalidation * is not mandatory. The caller may, however, request a TLB invalidation. */ void pmap_demote_DMAP(vm_paddr_t base, vm_size_t len, boolean_t invalidate) { pdp_entry_t *pdpe; pd_entry_t *pde; vm_offset_t va; boolean_t changed; if (len == 0) return; KASSERT(powerof2(len), ("pmap_demote_DMAP: len is not a power of 2")); KASSERT((base & (len - 1)) == 0, ("pmap_demote_DMAP: base is not a multiple of len")); if (len < NBPDP && base < dmaplimit) { va = PHYS_TO_DMAP(base); changed = FALSE; PMAP_LOCK(kernel_pmap); pdpe = pmap_pdpe(kernel_pmap, va); if ((*pdpe & X86_PG_V) == 0) panic("pmap_demote_DMAP: invalid PDPE"); if ((*pdpe & PG_PS) != 0) { if (!pmap_demote_pdpe(kernel_pmap, pdpe, va)) panic("pmap_demote_DMAP: PDPE failed"); changed = TRUE; } if (len < NBPDR) { pde = pmap_pdpe_to_pde(pdpe, va); if ((*pde & X86_PG_V) == 0) panic("pmap_demote_DMAP: invalid PDE"); if ((*pde & PG_PS) != 0) { if (!pmap_demote_pde(kernel_pmap, pde, va)) panic("pmap_demote_DMAP: PDE failed"); changed = TRUE; } } if (changed && invalidate) pmap_invalidate_page(kernel_pmap, va); PMAP_UNLOCK(kernel_pmap); } } /* * Perform the pmap work for mincore(2). If the page is not both referenced and * modified by this pmap, returns its physical address so that the caller can * find other mappings. */ int pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *pap) { pdp_entry_t *pdpe; pd_entry_t *pdep; pt_entry_t pte, PG_A, PG_M, PG_RW, PG_V; vm_paddr_t pa; int val; PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK(pmap); pte = 0; pa = 0; val = 0; pdpe = pmap_pdpe(pmap, addr); if (pdpe == NULL) goto out; if ((*pdpe & PG_V) != 0) { if ((*pdpe & PG_PS) != 0) { pte = *pdpe; pa = ((pte & PG_PS_PDP_FRAME) | (addr & PDPMASK)) & PG_FRAME; val = MINCORE_PSIND(2); } else { pdep = pmap_pde(pmap, addr); if (pdep != NULL && (*pdep & PG_V) != 0) { if ((*pdep & PG_PS) != 0) { pte = *pdep; /* Compute the physical address of the 4KB page. */ pa = ((pte & PG_PS_FRAME) | (addr & PDRMASK)) & PG_FRAME; val = MINCORE_PSIND(1); } else { pte = *pmap_pde_to_pte(pdep, addr); pa = pte & PG_FRAME; val = 0; } } } } if ((pte & PG_V) != 0) { val |= MINCORE_INCORE; if ((pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER; if ((pte & PG_A) != 0) val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER; } if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) != (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) && (pte & (PG_MANAGED | PG_V)) == (PG_MANAGED | PG_V)) { *pap = pa; } out: PMAP_UNLOCK(pmap); return (val); } static uint64_t pmap_pcid_alloc(pmap_t pmap, u_int cpuid) { uint32_t gen, new_gen, pcid_next; CRITICAL_ASSERT(curthread); gen = PCPU_GET(pcid_gen); if (pmap->pm_pcids[cpuid].pm_pcid == PMAP_PCID_KERN) return (pti ? 0 : CR3_PCID_SAVE); if (pmap->pm_pcids[cpuid].pm_gen == gen) return (CR3_PCID_SAVE); pcid_next = PCPU_GET(pcid_next); KASSERT((!pti && pcid_next <= PMAP_PCID_OVERMAX) || (pti && pcid_next <= PMAP_PCID_OVERMAX_KERN), ("cpu %d pcid_next %#x", cpuid, pcid_next)); if ((!pti && pcid_next == PMAP_PCID_OVERMAX) || (pti && pcid_next == PMAP_PCID_OVERMAX_KERN)) { new_gen = gen + 1; if (new_gen == 0) new_gen = 1; PCPU_SET(pcid_gen, new_gen); pcid_next = PMAP_PCID_KERN + 1; } else { new_gen = gen; } pmap->pm_pcids[cpuid].pm_pcid = pcid_next; pmap->pm_pcids[cpuid].pm_gen = new_gen; PCPU_SET(pcid_next, pcid_next + 1); return (0); } static uint64_t pmap_pcid_alloc_checked(pmap_t pmap, u_int cpuid) { uint64_t cached; cached = pmap_pcid_alloc(pmap, cpuid); KASSERT(pmap->pm_pcids[cpuid].pm_pcid < PMAP_PCID_OVERMAX, ("pmap %p cpu %d pcid %#x", pmap, cpuid, pmap->pm_pcids[cpuid].pm_pcid)); KASSERT(pmap->pm_pcids[cpuid].pm_pcid != PMAP_PCID_KERN || pmap == kernel_pmap, ("non-kernel pmap pmap %p cpu %d pcid %#x", pmap, cpuid, pmap->pm_pcids[cpuid].pm_pcid)); return (cached); } static void pmap_activate_sw_pti_post(struct thread *td, pmap_t pmap) { PCPU_GET(tssp)->tss_rsp0 = pmap->pm_ucr3 != PMAP_NO_CR3 ? PCPU_GET(pti_rsp0) : (uintptr_t)td->td_md.md_stack_base; } static void pmap_activate_sw_pcid_pti(struct thread *td, pmap_t pmap, u_int cpuid) { pmap_t old_pmap; uint64_t cached, cr3, kcr3, ucr3; KASSERT((read_rflags() & PSL_I) == 0, ("PCID needs interrupts disabled in pmap_activate_sw()")); /* See the comment in pmap_invalidate_page_pcid(). */ if (PCPU_GET(ucr3_load_mask) != PMAP_UCR3_NOMASK) { PCPU_SET(ucr3_load_mask, PMAP_UCR3_NOMASK); old_pmap = PCPU_GET(curpmap); MPASS(old_pmap->pm_ucr3 != PMAP_NO_CR3); old_pmap->pm_pcids[cpuid].pm_gen = 0; } cached = pmap_pcid_alloc_checked(pmap, cpuid); cr3 = rcr3(); if ((cr3 & ~CR3_PCID_MASK) != pmap->pm_cr3) load_cr3(pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid); PCPU_SET(curpmap, pmap); kcr3 = pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid; ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[cpuid].pm_pcid | PMAP_PCID_USER_PT; if (!cached && pmap->pm_ucr3 != PMAP_NO_CR3) PCPU_SET(ucr3_load_mask, ~CR3_PCID_SAVE); PCPU_SET(kcr3, kcr3 | CR3_PCID_SAVE); PCPU_SET(ucr3, ucr3 | CR3_PCID_SAVE); if (cached) counter_u64_add(pcid_save_cnt, 1); pmap_activate_sw_pti_post(td, pmap); } static void pmap_activate_sw_pcid_nopti(struct thread *td __unused, pmap_t pmap, u_int cpuid) { uint64_t cached, cr3; KASSERT((read_rflags() & PSL_I) == 0, ("PCID needs interrupts disabled in pmap_activate_sw()")); cached = pmap_pcid_alloc_checked(pmap, cpuid); cr3 = rcr3(); if (!cached || (cr3 & ~CR3_PCID_MASK) != pmap->pm_cr3) load_cr3(pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid | cached); PCPU_SET(curpmap, pmap); if (cached) counter_u64_add(pcid_save_cnt, 1); } static void pmap_activate_sw_nopcid_nopti(struct thread *td __unused, pmap_t pmap, u_int cpuid __unused) { load_cr3(pmap->pm_cr3); PCPU_SET(curpmap, pmap); } static void pmap_activate_sw_nopcid_pti(struct thread *td, pmap_t pmap, u_int cpuid __unused) { pmap_activate_sw_nopcid_nopti(td, pmap, cpuid); PCPU_SET(kcr3, pmap->pm_cr3); PCPU_SET(ucr3, pmap->pm_ucr3); pmap_activate_sw_pti_post(td, pmap); } DEFINE_IFUNC(static, void, pmap_activate_sw_mode, (struct thread *, pmap_t, u_int)) { if (pmap_pcid_enabled && pti) return (pmap_activate_sw_pcid_pti); else if (pmap_pcid_enabled && !pti) return (pmap_activate_sw_pcid_nopti); else if (!pmap_pcid_enabled && pti) return (pmap_activate_sw_nopcid_pti); else /* if (!pmap_pcid_enabled && !pti) */ return (pmap_activate_sw_nopcid_nopti); } void pmap_activate_sw(struct thread *td) { pmap_t oldpmap, pmap; u_int cpuid; oldpmap = PCPU_GET(curpmap); pmap = vmspace_pmap(td->td_proc->p_vmspace); if (oldpmap == pmap) { if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); return; } cpuid = PCPU_GET(cpuid); #ifdef SMP CPU_SET_ATOMIC(cpuid, &pmap->pm_active); #else CPU_SET(cpuid, &pmap->pm_active); #endif pmap_activate_sw_mode(td, pmap, cpuid); #ifdef SMP CPU_CLR_ATOMIC(cpuid, &oldpmap->pm_active); #else CPU_CLR(cpuid, &oldpmap->pm_active); #endif } void pmap_activate(struct thread *td) { /* * invltlb_{invpcid,}_pcid_handler() is used to handle an * invalidate_all IPI, which checks for curpmap == * smp_tlb_pmap. The below sequence of operations has a * window where %CR3 is loaded with the new pmap's PML4 * address, but the curpmap value has not yet been updated. * This causes the invltlb IPI handler, which is called * between the updates, to execute as a NOP, which leaves * stale TLB entries. * * Note that the most common use of pmap_activate_sw(), from * a context switch, is immune to this race, because * interrupts are disabled (while the thread lock is owned), * so the IPI is delayed until after curpmap is updated. Protect * other callers in a similar way, by disabling interrupts * around the %cr3 register reload and curpmap assignment. */ spinlock_enter(); pmap_activate_sw(td); spinlock_exit(); } void pmap_activate_boot(pmap_t pmap) { uint64_t kcr3; u_int cpuid; /* * kernel_pmap must be never deactivated, and we ensure that * by never activating it at all. */ MPASS(pmap != kernel_pmap); cpuid = PCPU_GET(cpuid); #ifdef SMP CPU_SET_ATOMIC(cpuid, &pmap->pm_active); #else CPU_SET(cpuid, &pmap->pm_active); #endif PCPU_SET(curpmap, pmap); if (pti) { kcr3 = pmap->pm_cr3; if (pmap_pcid_enabled) kcr3 |= pmap->pm_pcids[cpuid].pm_pcid | CR3_PCID_SAVE; } else { kcr3 = PMAP_NO_CR3; } PCPU_SET(kcr3, kcr3); PCPU_SET(ucr3, PMAP_NO_CR3); } void pmap_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz) { } /* * Increase the starting virtual address of the given mapping if a * different alignment might result in more superpage mappings. */ void pmap_align_superpage(vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, vm_size_t size) { vm_offset_t superpage_offset; if (size < NBPDR) return; if (object != NULL && (object->flags & OBJ_COLORED) != 0) offset += ptoa(object->pg_color); superpage_offset = offset & PDRMASK; if (size - ((NBPDR - superpage_offset) & PDRMASK) < NBPDR || (*addr & PDRMASK) == superpage_offset) return; if ((*addr & PDRMASK) < superpage_offset) *addr = (*addr & ~PDRMASK) + superpage_offset; else *addr = ((*addr + PDRMASK) & ~PDRMASK) + superpage_offset; } #ifdef INVARIANTS static unsigned long num_dirty_emulations; SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_dirty_emulations, CTLFLAG_RW, &num_dirty_emulations, 0, NULL); static unsigned long num_accessed_emulations; SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_accessed_emulations, CTLFLAG_RW, &num_accessed_emulations, 0, NULL); static unsigned long num_superpage_accessed_emulations; SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_superpage_accessed_emulations, CTLFLAG_RW, &num_superpage_accessed_emulations, 0, NULL); static unsigned long ad_emulation_superpage_promotions; SYSCTL_ULONG(_vm_pmap, OID_AUTO, ad_emulation_superpage_promotions, CTLFLAG_RW, &ad_emulation_superpage_promotions, 0, NULL); #endif /* INVARIANTS */ int pmap_emulate_accessed_dirty(pmap_t pmap, vm_offset_t va, int ftype) { int rv; struct rwlock *lock; #if VM_NRESERVLEVEL > 0 vm_page_t m, mpte; #endif pd_entry_t *pde; pt_entry_t *pte, PG_A, PG_M, PG_RW, PG_V; KASSERT(ftype == VM_PROT_READ || ftype == VM_PROT_WRITE, ("pmap_emulate_accessed_dirty: invalid fault type %d", ftype)); if (!pmap_emulate_ad_bits(pmap)) return (-1); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); rv = -1; lock = NULL; PMAP_LOCK(pmap); pde = pmap_pde(pmap, va); if (pde == NULL || (*pde & PG_V) == 0) goto done; if ((*pde & PG_PS) != 0) { if (ftype == VM_PROT_READ) { #ifdef INVARIANTS atomic_add_long(&num_superpage_accessed_emulations, 1); #endif *pde |= PG_A; rv = 0; } goto done; } pte = pmap_pde_to_pte(pde, va); if ((*pte & PG_V) == 0) goto done; if (ftype == VM_PROT_WRITE) { if ((*pte & PG_RW) == 0) goto done; /* * Set the modified and accessed bits simultaneously. * * Intel EPT PTEs that do software emulation of A/D bits map * PG_A and PG_M to EPT_PG_READ and EPT_PG_WRITE respectively. * An EPT misconfiguration is triggered if the PTE is writable * but not readable (WR=10). This is avoided by setting PG_A * and PG_M simultaneously. */ *pte |= PG_M | PG_A; } else { *pte |= PG_A; } #if VM_NRESERVLEVEL > 0 /* try to promote the mapping */ if (va < VM_MAXUSER_ADDRESS) mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME); else mpte = NULL; m = PHYS_TO_VM_PAGE(*pte & PG_FRAME); if ((mpte == NULL || mpte->ref_count == NPTEPG) && pmap_ps_enabled(pmap) && (m->flags & PG_FICTITIOUS) == 0 && vm_reserv_level_iffullpop(m) == 0) { pmap_promote_pde(pmap, pde, va, &lock); #ifdef INVARIANTS atomic_add_long(&ad_emulation_superpage_promotions, 1); #endif } #endif #ifdef INVARIANTS if (ftype == VM_PROT_WRITE) atomic_add_long(&num_dirty_emulations, 1); else atomic_add_long(&num_accessed_emulations, 1); #endif rv = 0; /* success */ done: if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); return (rv); } void pmap_get_mapping(pmap_t pmap, vm_offset_t va, uint64_t *ptr, int *num) { pml4_entry_t *pml4; pdp_entry_t *pdp; pd_entry_t *pde; pt_entry_t *pte, PG_V; int idx; idx = 0; PG_V = pmap_valid_bit(pmap); PMAP_LOCK(pmap); pml4 = pmap_pml4e(pmap, va); if (pml4 == NULL) goto done; ptr[idx++] = *pml4; if ((*pml4 & PG_V) == 0) goto done; pdp = pmap_pml4e_to_pdpe(pml4, va); ptr[idx++] = *pdp; if ((*pdp & PG_V) == 0 || (*pdp & PG_PS) != 0) goto done; pde = pmap_pdpe_to_pde(pdp, va); ptr[idx++] = *pde; if ((*pde & PG_V) == 0 || (*pde & PG_PS) != 0) goto done; pte = pmap_pde_to_pte(pde, va); ptr[idx++] = *pte; done: PMAP_UNLOCK(pmap); *num = idx; } /** * Get the kernel virtual address of a set of physical pages. If there are * physical addresses not covered by the DMAP perform a transient mapping * that will be removed when calling pmap_unmap_io_transient. * * \param page The pages the caller wishes to obtain the virtual * address on the kernel memory map. * \param vaddr On return contains the kernel virtual memory address * of the pages passed in the page parameter. * \param count Number of pages passed in. * \param can_fault TRUE if the thread using the mapped pages can take * page faults, FALSE otherwise. * * \returns TRUE if the caller must call pmap_unmap_io_transient when * finished or FALSE otherwise. * */ boolean_t pmap_map_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count, boolean_t can_fault) { vm_paddr_t paddr; boolean_t needs_mapping; pt_entry_t *pte; int cache_bits, error __unused, i; /* * Allocate any KVA space that we need, this is done in a separate * loop to prevent calling vmem_alloc while pinned. */ needs_mapping = FALSE; for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (__predict_false(paddr >= dmaplimit)) { error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK, &vaddr[i]); KASSERT(error == 0, ("vmem_alloc failed: %d", error)); needs_mapping = TRUE; } else { vaddr[i] = PHYS_TO_DMAP(paddr); } } /* Exit early if everything is covered by the DMAP */ if (!needs_mapping) return (FALSE); /* * NB: The sequence of updating a page table followed by accesses * to the corresponding pages used in the !DMAP case is subject to * the situation described in the "AMD64 Architecture Programmer's * Manual Volume 2: System Programming" rev. 3.23, "7.3.1 Special * Coherency Considerations". Therefore, issuing the INVLPG right * after modifying the PTE bits is crucial. */ if (!can_fault) sched_pin(); for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (paddr >= dmaplimit) { if (can_fault) { /* * Slow path, since we can get page faults * while mappings are active don't pin the * thread to the CPU and instead add a global * mapping visible to all CPUs. */ pmap_qenter(vaddr[i], &page[i], 1); } else { pte = vtopte(vaddr[i]); cache_bits = pmap_cache_bits(kernel_pmap, page[i]->md.pat_mode, 0); pte_store(pte, paddr | X86_PG_RW | X86_PG_V | cache_bits); invlpg(vaddr[i]); } } } return (needs_mapping); } void pmap_unmap_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count, boolean_t can_fault) { vm_paddr_t paddr; int i; if (!can_fault) sched_unpin(); for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (paddr >= dmaplimit) { if (can_fault) pmap_qremove(vaddr[i], 1); vmem_free(kernel_arena, vaddr[i], PAGE_SIZE); } } } vm_offset_t pmap_quick_enter_page(vm_page_t m) { vm_paddr_t paddr; paddr = VM_PAGE_TO_PHYS(m); if (paddr < dmaplimit) return (PHYS_TO_DMAP(paddr)); mtx_lock_spin(&qframe_mtx); KASSERT(*vtopte(qframe) == 0, ("qframe busy")); pte_store(vtopte(qframe), paddr | X86_PG_RW | X86_PG_V | X86_PG_A | X86_PG_M | pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0)); return (qframe); } void pmap_quick_remove_page(vm_offset_t addr) { if (addr != qframe) return; pte_store(vtopte(qframe), 0); invlpg(qframe); mtx_unlock_spin(&qframe_mtx); } /* * Pdp pages from the large map are managed differently from either * kernel or user page table pages. They are permanently allocated at * initialization time, and their reference count is permanently set to * zero. The pml4 entries pointing to those pages are copied into * each allocated pmap. * * In contrast, pd and pt pages are managed like user page table * pages. They are dynamically allocated, and their reference count * represents the number of valid entries within the page. */ static vm_page_t pmap_large_map_getptp_unlocked(void) { return (pmap_alloc_pt_page(kernel_pmap, 0, VM_ALLOC_ZERO)); } static vm_page_t pmap_large_map_getptp(void) { vm_page_t m; PMAP_LOCK_ASSERT(kernel_pmap, MA_OWNED); m = pmap_large_map_getptp_unlocked(); if (m == NULL) { PMAP_UNLOCK(kernel_pmap); vm_wait(NULL); PMAP_LOCK(kernel_pmap); /* Callers retry. */ } return (m); } static pdp_entry_t * pmap_large_map_pdpe(vm_offset_t va) { vm_pindex_t pml4_idx; vm_paddr_t mphys; pml4_idx = pmap_pml4e_index(va); KASSERT(LMSPML4I <= pml4_idx && pml4_idx < LMSPML4I + lm_ents, ("pmap_large_map_pdpe: va %#jx out of range idx %#jx LMSPML4I " "%#jx lm_ents %d", (uintmax_t)va, (uintmax_t)pml4_idx, LMSPML4I, lm_ents)); KASSERT((kernel_pml4[pml4_idx] & X86_PG_V) != 0, ("pmap_large_map_pdpe: invalid pml4 for va %#jx idx %#jx " "LMSPML4I %#jx lm_ents %d", (uintmax_t)va, (uintmax_t)pml4_idx, LMSPML4I, lm_ents)); mphys = kernel_pml4[pml4_idx] & PG_FRAME; return ((pdp_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pdpe_index(va)); } static pd_entry_t * pmap_large_map_pde(vm_offset_t va) { pdp_entry_t *pdpe; vm_page_t m; vm_paddr_t mphys; retry: pdpe = pmap_large_map_pdpe(va); if (*pdpe == 0) { m = pmap_large_map_getptp(); if (m == NULL) goto retry; mphys = VM_PAGE_TO_PHYS(m); *pdpe = mphys | X86_PG_A | X86_PG_RW | X86_PG_V | pg_nx; } else { MPASS((*pdpe & X86_PG_PS) == 0); mphys = *pdpe & PG_FRAME; } return ((pd_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pde_index(va)); } static pt_entry_t * pmap_large_map_pte(vm_offset_t va) { pd_entry_t *pde; vm_page_t m; vm_paddr_t mphys; retry: pde = pmap_large_map_pde(va); if (*pde == 0) { m = pmap_large_map_getptp(); if (m == NULL) goto retry; mphys = VM_PAGE_TO_PHYS(m); *pde = mphys | X86_PG_A | X86_PG_RW | X86_PG_V | pg_nx; PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pde))->ref_count++; } else { MPASS((*pde & X86_PG_PS) == 0); mphys = *pde & PG_FRAME; } return ((pt_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pte_index(va)); } static vm_paddr_t pmap_large_map_kextract(vm_offset_t va) { pdp_entry_t *pdpe, pdp; pd_entry_t *pde, pd; pt_entry_t *pte, pt; KASSERT(PMAP_ADDRESS_IN_LARGEMAP(va), ("not largemap range %#lx", (u_long)va)); pdpe = pmap_large_map_pdpe(va); pdp = *pdpe; KASSERT((pdp & X86_PG_V) != 0, ("invalid pdp va %#lx pdpe %#lx pdp %#lx", va, (u_long)pdpe, pdp)); if ((pdp & X86_PG_PS) != 0) { KASSERT((amd_feature & AMDID_PAGE1GB) != 0, ("no 1G pages, va %#lx pdpe %#lx pdp %#lx", va, (u_long)pdpe, pdp)); return ((pdp & PG_PS_PDP_FRAME) | (va & PDPMASK)); } pde = pmap_pdpe_to_pde(pdpe, va); pd = *pde; KASSERT((pd & X86_PG_V) != 0, ("invalid pd va %#lx pde %#lx pd %#lx", va, (u_long)pde, pd)); if ((pd & X86_PG_PS) != 0) return ((pd & PG_PS_FRAME) | (va & PDRMASK)); pte = pmap_pde_to_pte(pde, va); pt = *pte; KASSERT((pt & X86_PG_V) != 0, ("invalid pte va %#lx pte %#lx pt %#lx", va, (u_long)pte, pt)); return ((pt & PG_FRAME) | (va & PAGE_MASK)); } static int pmap_large_map_getva(vm_size_t len, vm_offset_t align, vm_offset_t phase, vmem_addr_t *vmem_res) { /* * Large mappings are all but static. Consequently, there * is no point in waiting for an earlier allocation to be * freed. */ return (vmem_xalloc(large_vmem, len, align, phase, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, M_NOWAIT | M_BESTFIT, vmem_res)); } int pmap_large_map(vm_paddr_t spa, vm_size_t len, void **addr, vm_memattr_t mattr) { pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte; vm_offset_t va, inc; vmem_addr_t vmem_res; vm_paddr_t pa; int error; if (len == 0 || spa + len < spa) return (EINVAL); /* See if DMAP can serve. */ if (spa + len <= dmaplimit) { va = PHYS_TO_DMAP(spa); *addr = (void *)va; return (pmap_change_attr(va, len, mattr)); } /* * No, allocate KVA. Fit the address with best possible * alignment for superpages. Fall back to worse align if * failed. */ error = ENOMEM; if ((amd_feature & AMDID_PAGE1GB) != 0 && rounddown2(spa + len, NBPDP) >= roundup2(spa, NBPDP) + NBPDP) error = pmap_large_map_getva(len, NBPDP, spa & PDPMASK, &vmem_res); if (error != 0 && rounddown2(spa + len, NBPDR) >= roundup2(spa, NBPDR) + NBPDR) error = pmap_large_map_getva(len, NBPDR, spa & PDRMASK, &vmem_res); if (error != 0) error = pmap_large_map_getva(len, PAGE_SIZE, 0, &vmem_res); if (error != 0) return (error); /* * Fill pagetable. PG_M is not pre-set, we scan modified bits * in the pagetable to minimize flushing. No need to * invalidate TLB, since we only update invalid entries. */ PMAP_LOCK(kernel_pmap); for (pa = spa, va = vmem_res; len > 0; pa += inc, va += inc, len -= inc) { if ((amd_feature & AMDID_PAGE1GB) != 0 && len >= NBPDP && (pa & PDPMASK) == 0 && (va & PDPMASK) == 0) { pdpe = pmap_large_map_pdpe(va); MPASS(*pdpe == 0); *pdpe = pa | pg_g | X86_PG_PS | X86_PG_RW | X86_PG_V | X86_PG_A | pg_nx | pmap_cache_bits(kernel_pmap, mattr, TRUE); inc = NBPDP; } else if (len >= NBPDR && (pa & PDRMASK) == 0 && (va & PDRMASK) == 0) { pde = pmap_large_map_pde(va); MPASS(*pde == 0); *pde = pa | pg_g | X86_PG_PS | X86_PG_RW | X86_PG_V | X86_PG_A | pg_nx | pmap_cache_bits(kernel_pmap, mattr, TRUE); PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pde))-> ref_count++; inc = NBPDR; } else { pte = pmap_large_map_pte(va); MPASS(*pte == 0); *pte = pa | pg_g | X86_PG_RW | X86_PG_V | X86_PG_A | pg_nx | pmap_cache_bits(kernel_pmap, mattr, FALSE); PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pte))-> ref_count++; inc = PAGE_SIZE; } } PMAP_UNLOCK(kernel_pmap); MPASS(len == 0); *addr = (void *)vmem_res; return (0); } void pmap_large_unmap(void *svaa, vm_size_t len) { vm_offset_t sva, va; vm_size_t inc; pdp_entry_t *pdpe, pdp; pd_entry_t *pde, pd; pt_entry_t *pte; vm_page_t m; struct spglist spgf; sva = (vm_offset_t)svaa; if (len == 0 || sva + len < sva || (sva >= DMAP_MIN_ADDRESS && sva + len <= DMAP_MIN_ADDRESS + dmaplimit)) return; SLIST_INIT(&spgf); KASSERT(PMAP_ADDRESS_IN_LARGEMAP(sva) && PMAP_ADDRESS_IN_LARGEMAP(sva + len - 1), ("not largemap range %#lx %#lx", (u_long)svaa, (u_long)svaa + len)); PMAP_LOCK(kernel_pmap); for (va = sva; va < sva + len; va += inc) { pdpe = pmap_large_map_pdpe(va); pdp = *pdpe; KASSERT((pdp & X86_PG_V) != 0, ("invalid pdp va %#lx pdpe %#lx pdp %#lx", va, (u_long)pdpe, pdp)); if ((pdp & X86_PG_PS) != 0) { KASSERT((amd_feature & AMDID_PAGE1GB) != 0, ("no 1G pages, va %#lx pdpe %#lx pdp %#lx", va, (u_long)pdpe, pdp)); KASSERT((va & PDPMASK) == 0, ("PDPMASK bit set, va %#lx pdpe %#lx pdp %#lx", va, (u_long)pdpe, pdp)); KASSERT(va + NBPDP <= sva + len, ("unmap covers partial 1GB page, sva %#lx va %#lx " "pdpe %#lx pdp %#lx len %#lx", sva, va, (u_long)pdpe, pdp, len)); *pdpe = 0; inc = NBPDP; continue; } pde = pmap_pdpe_to_pde(pdpe, va); pd = *pde; KASSERT((pd & X86_PG_V) != 0, ("invalid pd va %#lx pde %#lx pd %#lx", va, (u_long)pde, pd)); if ((pd & X86_PG_PS) != 0) { KASSERT((va & PDRMASK) == 0, ("PDRMASK bit set, va %#lx pde %#lx pd %#lx", va, (u_long)pde, pd)); KASSERT(va + NBPDR <= sva + len, ("unmap covers partial 2MB page, sva %#lx va %#lx " "pde %#lx pd %#lx len %#lx", sva, va, (u_long)pde, pd, len)); pde_store(pde, 0); inc = NBPDR; m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pde)); m->ref_count--; if (m->ref_count == 0) { *pdpe = 0; SLIST_INSERT_HEAD(&spgf, m, plinks.s.ss); } continue; } pte = pmap_pde_to_pte(pde, va); KASSERT((*pte & X86_PG_V) != 0, ("invalid pte va %#lx pte %#lx pt %#lx", va, (u_long)pte, *pte)); pte_clear(pte); inc = PAGE_SIZE; m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pte)); m->ref_count--; if (m->ref_count == 0) { *pde = 0; SLIST_INSERT_HEAD(&spgf, m, plinks.s.ss); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pde)); m->ref_count--; if (m->ref_count == 0) { *pdpe = 0; SLIST_INSERT_HEAD(&spgf, m, plinks.s.ss); } } } pmap_invalidate_range(kernel_pmap, sva, sva + len); PMAP_UNLOCK(kernel_pmap); vm_page_free_pages_toq(&spgf, false); vmem_free(large_vmem, sva, len); } static void pmap_large_map_wb_fence_mfence(void) { mfence(); } static void pmap_large_map_wb_fence_atomic(void) { atomic_thread_fence_seq_cst(); } static void pmap_large_map_wb_fence_nop(void) { } DEFINE_IFUNC(static, void, pmap_large_map_wb_fence, (void)) { if (cpu_vendor_id != CPU_VENDOR_INTEL) return (pmap_large_map_wb_fence_mfence); else if ((cpu_stdext_feature & (CPUID_STDEXT_CLWB | CPUID_STDEXT_CLFLUSHOPT)) == 0) return (pmap_large_map_wb_fence_atomic); else /* clflush is strongly enough ordered */ return (pmap_large_map_wb_fence_nop); } static void pmap_large_map_flush_range_clwb(vm_offset_t va, vm_size_t len) { for (; len > 0; len -= cpu_clflush_line_size, va += cpu_clflush_line_size) clwb(va); } static void pmap_large_map_flush_range_clflushopt(vm_offset_t va, vm_size_t len) { for (; len > 0; len -= cpu_clflush_line_size, va += cpu_clflush_line_size) clflushopt(va); } static void pmap_large_map_flush_range_clflush(vm_offset_t va, vm_size_t len) { for (; len > 0; len -= cpu_clflush_line_size, va += cpu_clflush_line_size) clflush(va); } static void pmap_large_map_flush_range_nop(vm_offset_t sva __unused, vm_size_t len __unused) { } DEFINE_IFUNC(static, void, pmap_large_map_flush_range, (vm_offset_t, vm_size_t)) { if ((cpu_stdext_feature & CPUID_STDEXT_CLWB) != 0) return (pmap_large_map_flush_range_clwb); else if ((cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0) return (pmap_large_map_flush_range_clflushopt); else if ((cpu_feature & CPUID_CLFSH) != 0) return (pmap_large_map_flush_range_clflush); else return (pmap_large_map_flush_range_nop); } static void pmap_large_map_wb_large(vm_offset_t sva, vm_offset_t eva) { volatile u_long *pe; u_long p; vm_offset_t va; vm_size_t inc; bool seen_other; for (va = sva; va < eva; va += inc) { inc = 0; if ((amd_feature & AMDID_PAGE1GB) != 0) { pe = (volatile u_long *)pmap_large_map_pdpe(va); p = *pe; if ((p & X86_PG_PS) != 0) inc = NBPDP; } if (inc == 0) { pe = (volatile u_long *)pmap_large_map_pde(va); p = *pe; if ((p & X86_PG_PS) != 0) inc = NBPDR; } if (inc == 0) { pe = (volatile u_long *)pmap_large_map_pte(va); p = *pe; inc = PAGE_SIZE; } seen_other = false; for (;;) { if ((p & X86_PG_AVAIL1) != 0) { /* * Spin-wait for the end of a parallel * write-back. */ cpu_spinwait(); p = *pe; /* * If we saw other write-back * occuring, we cannot rely on PG_M to * indicate state of the cache. The * PG_M bit is cleared before the * flush to avoid ignoring new writes, * and writes which are relevant for * us might happen after. */ seen_other = true; continue; } if ((p & X86_PG_M) != 0 || seen_other) { if (!atomic_fcmpset_long(pe, &p, (p & ~X86_PG_M) | X86_PG_AVAIL1)) /* * If we saw PG_M without * PG_AVAIL1, and then on the * next attempt we do not * observe either PG_M or * PG_AVAIL1, the other * write-back started after us * and finished before us. We * can rely on it doing our * work. */ continue; pmap_large_map_flush_range(va, inc); atomic_clear_long(pe, X86_PG_AVAIL1); } break; } maybe_yield(); } } /* * Write-back cache lines for the given address range. * * Must be called only on the range or sub-range returned from * pmap_large_map(). Must not be called on the coalesced ranges. * * Does nothing on CPUs without CLWB, CLFLUSHOPT, or CLFLUSH * instructions support. */ void pmap_large_map_wb(void *svap, vm_size_t len) { vm_offset_t eva, sva; sva = (vm_offset_t)svap; eva = sva + len; pmap_large_map_wb_fence(); if (sva >= DMAP_MIN_ADDRESS && eva <= DMAP_MIN_ADDRESS + dmaplimit) { pmap_large_map_flush_range(sva, len); } else { KASSERT(sva >= LARGEMAP_MIN_ADDRESS && eva <= LARGEMAP_MIN_ADDRESS + lm_ents * NBPML4, ("pmap_large_map_wb: not largemap %#lx %#lx", sva, len)); pmap_large_map_wb_large(sva, eva); } pmap_large_map_wb_fence(); } static vm_page_t pmap_pti_alloc_page(void) { vm_page_t m; VM_OBJECT_ASSERT_WLOCKED(pti_obj); m = vm_page_grab(pti_obj, pti_pg_idx++, VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | VM_ALLOC_ZERO); return (m); } static bool pmap_pti_free_page(vm_page_t m) { KASSERT(m->ref_count > 0, ("page %p not referenced", m)); if (!vm_page_unwire_noq(m)) return (false); vm_page_free_zero(m); return (true); } static void pmap_pti_init(void) { vm_page_t pml4_pg; pdp_entry_t *pdpe; vm_offset_t va; int i; if (!pti) return; pti_obj = vm_pager_allocate(OBJT_PHYS, NULL, 0, VM_PROT_ALL, 0, NULL); VM_OBJECT_WLOCK(pti_obj); pml4_pg = pmap_pti_alloc_page(); pti_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml4_pg)); for (va = VM_MIN_KERNEL_ADDRESS; va <= VM_MAX_KERNEL_ADDRESS && va >= VM_MIN_KERNEL_ADDRESS && va > NBPML4; va += NBPML4) { pdpe = pmap_pti_pdpe(va); pmap_pti_wire_pte(pdpe); } pmap_pti_add_kva_locked((vm_offset_t)&__pcpu[0], (vm_offset_t)&__pcpu[0] + sizeof(__pcpu[0]) * MAXCPU, false); pmap_pti_add_kva_locked((vm_offset_t)idt, (vm_offset_t)idt + sizeof(struct gate_descriptor) * NIDT, false); CPU_FOREACH(i) { /* Doublefault stack IST 1 */ va = __pcpu[i].pc_common_tss.tss_ist1 + sizeof(struct nmi_pcpu); pmap_pti_add_kva_locked(va - DBLFAULT_STACK_SIZE, va, false); /* NMI stack IST 2 */ va = __pcpu[i].pc_common_tss.tss_ist2 + sizeof(struct nmi_pcpu); pmap_pti_add_kva_locked(va - NMI_STACK_SIZE, va, false); /* MC# stack IST 3 */ va = __pcpu[i].pc_common_tss.tss_ist3 + sizeof(struct nmi_pcpu); pmap_pti_add_kva_locked(va - MCE_STACK_SIZE, va, false); /* DB# stack IST 4 */ va = __pcpu[i].pc_common_tss.tss_ist4 + sizeof(struct nmi_pcpu); pmap_pti_add_kva_locked(va - DBG_STACK_SIZE, va, false); } pmap_pti_add_kva_locked((vm_offset_t)KERNSTART, (vm_offset_t)etext, true); pti_finalized = true; VM_OBJECT_WUNLOCK(pti_obj); } static void pmap_cpu_init(void *arg __unused) { CPU_COPY(&all_cpus, &kernel_pmap->pm_active); pmap_pti_init(); } SYSINIT(pmap_cpu, SI_SUB_CPU + 1, SI_ORDER_ANY, pmap_cpu_init, NULL); static pdp_entry_t * pmap_pti_pdpe(vm_offset_t va) { pml4_entry_t *pml4e; pdp_entry_t *pdpe; vm_page_t m; vm_pindex_t pml4_idx; vm_paddr_t mphys; VM_OBJECT_ASSERT_WLOCKED(pti_obj); pml4_idx = pmap_pml4e_index(va); pml4e = &pti_pml4[pml4_idx]; m = NULL; if (*pml4e == 0) { if (pti_finalized) panic("pml4 alloc after finalization\n"); m = pmap_pti_alloc_page(); if (*pml4e != 0) { pmap_pti_free_page(m); mphys = *pml4e & ~PAGE_MASK; } else { mphys = VM_PAGE_TO_PHYS(m); *pml4e = mphys | X86_PG_RW | X86_PG_V; } } else { mphys = *pml4e & ~PAGE_MASK; } pdpe = (pdp_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pdpe_index(va); return (pdpe); } static void pmap_pti_wire_pte(void *pte) { vm_page_t m; VM_OBJECT_ASSERT_WLOCKED(pti_obj); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pte)); m->ref_count++; } static void pmap_pti_unwire_pde(void *pde, bool only_ref) { vm_page_t m; VM_OBJECT_ASSERT_WLOCKED(pti_obj); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pde)); MPASS(m->ref_count > 0); MPASS(only_ref || m->ref_count > 1); pmap_pti_free_page(m); } static void pmap_pti_unwire_pte(void *pte, vm_offset_t va) { vm_page_t m; pd_entry_t *pde; VM_OBJECT_ASSERT_WLOCKED(pti_obj); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pte)); MPASS(m->ref_count > 0); if (pmap_pti_free_page(m)) { pde = pmap_pti_pde(va); MPASS((*pde & (X86_PG_PS | X86_PG_V)) == X86_PG_V); *pde = 0; pmap_pti_unwire_pde(pde, false); } } static pd_entry_t * pmap_pti_pde(vm_offset_t va) { pdp_entry_t *pdpe; pd_entry_t *pde; vm_page_t m; vm_pindex_t pd_idx; vm_paddr_t mphys; VM_OBJECT_ASSERT_WLOCKED(pti_obj); pdpe = pmap_pti_pdpe(va); if (*pdpe == 0) { m = pmap_pti_alloc_page(); if (*pdpe != 0) { pmap_pti_free_page(m); MPASS((*pdpe & X86_PG_PS) == 0); mphys = *pdpe & ~PAGE_MASK; } else { mphys = VM_PAGE_TO_PHYS(m); *pdpe = mphys | X86_PG_RW | X86_PG_V; } } else { MPASS((*pdpe & X86_PG_PS) == 0); mphys = *pdpe & ~PAGE_MASK; } pde = (pd_entry_t *)PHYS_TO_DMAP(mphys); pd_idx = pmap_pde_index(va); pde += pd_idx; return (pde); } static pt_entry_t * pmap_pti_pte(vm_offset_t va, bool *unwire_pde) { pd_entry_t *pde; pt_entry_t *pte; vm_page_t m; vm_paddr_t mphys; VM_OBJECT_ASSERT_WLOCKED(pti_obj); pde = pmap_pti_pde(va); if (unwire_pde != NULL) { *unwire_pde = true; pmap_pti_wire_pte(pde); } if (*pde == 0) { m = pmap_pti_alloc_page(); if (*pde != 0) { pmap_pti_free_page(m); MPASS((*pde & X86_PG_PS) == 0); mphys = *pde & ~(PAGE_MASK | pg_nx); } else { mphys = VM_PAGE_TO_PHYS(m); *pde = mphys | X86_PG_RW | X86_PG_V; if (unwire_pde != NULL) *unwire_pde = false; } } else { MPASS((*pde & X86_PG_PS) == 0); mphys = *pde & ~(PAGE_MASK | pg_nx); } pte = (pt_entry_t *)PHYS_TO_DMAP(mphys); pte += pmap_pte_index(va); return (pte); } static void pmap_pti_add_kva_locked(vm_offset_t sva, vm_offset_t eva, bool exec) { vm_paddr_t pa; pd_entry_t *pde; pt_entry_t *pte, ptev; bool unwire_pde; VM_OBJECT_ASSERT_WLOCKED(pti_obj); sva = trunc_page(sva); MPASS(sva > VM_MAXUSER_ADDRESS); eva = round_page(eva); MPASS(sva < eva); for (; sva < eva; sva += PAGE_SIZE) { pte = pmap_pti_pte(sva, &unwire_pde); pa = pmap_kextract(sva); ptev = pa | X86_PG_RW | X86_PG_V | X86_PG_A | X86_PG_G | (exec ? 0 : pg_nx) | pmap_cache_bits(kernel_pmap, VM_MEMATTR_DEFAULT, FALSE); if (*pte == 0) { pte_store(pte, ptev); pmap_pti_wire_pte(pte); } else { KASSERT(!pti_finalized, ("pti overlap after fin %#lx %#lx %#lx", sva, *pte, ptev)); KASSERT(*pte == ptev, ("pti non-identical pte after fin %#lx %#lx %#lx", sva, *pte, ptev)); } if (unwire_pde) { pde = pmap_pti_pde(sva); pmap_pti_unwire_pde(pde, true); } } } void pmap_pti_add_kva(vm_offset_t sva, vm_offset_t eva, bool exec) { if (!pti) return; VM_OBJECT_WLOCK(pti_obj); pmap_pti_add_kva_locked(sva, eva, exec); VM_OBJECT_WUNLOCK(pti_obj); } void pmap_pti_remove_kva(vm_offset_t sva, vm_offset_t eva) { pt_entry_t *pte; vm_offset_t va; if (!pti) return; sva = rounddown2(sva, PAGE_SIZE); MPASS(sva > VM_MAXUSER_ADDRESS); eva = roundup2(eva, PAGE_SIZE); MPASS(sva < eva); VM_OBJECT_WLOCK(pti_obj); for (va = sva; va < eva; va += PAGE_SIZE) { pte = pmap_pti_pte(va, NULL); KASSERT((*pte & X86_PG_V) != 0, ("invalid pte va %#lx pte %#lx pt %#lx", va, (u_long)pte, *pte)); pte_clear(pte); pmap_pti_unwire_pte(pte, va); } pmap_invalidate_range(kernel_pmap, sva, eva); VM_OBJECT_WUNLOCK(pti_obj); } static void * pkru_dup_range(void *ctx __unused, void *data) { struct pmap_pkru_range *node, *new_node; new_node = uma_zalloc(pmap_pkru_ranges_zone, M_NOWAIT); if (new_node == NULL) return (NULL); node = data; memcpy(new_node, node, sizeof(*node)); return (new_node); } static void pkru_free_range(void *ctx __unused, void *node) { uma_zfree(pmap_pkru_ranges_zone, node); } static int pmap_pkru_assign(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, u_int keyidx, int flags) { struct pmap_pkru_range *ppr; int error; PMAP_LOCK_ASSERT(pmap, MA_OWNED); MPASS(pmap->pm_type == PT_X86); MPASS((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0); if ((flags & AMD64_PKRU_EXCL) != 0 && !rangeset_check_empty(&pmap->pm_pkru, sva, eva)) return (EBUSY); ppr = uma_zalloc(pmap_pkru_ranges_zone, M_NOWAIT); if (ppr == NULL) return (ENOMEM); ppr->pkru_keyidx = keyidx; ppr->pkru_flags = flags & AMD64_PKRU_PERSIST; error = rangeset_insert(&pmap->pm_pkru, sva, eva, ppr); if (error != 0) uma_zfree(pmap_pkru_ranges_zone, ppr); return (error); } static int pmap_pkru_deassign(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); MPASS(pmap->pm_type == PT_X86); MPASS((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0); return (rangeset_remove(&pmap->pm_pkru, sva, eva)); } static void pmap_pkru_deassign_all(pmap_t pmap) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pmap->pm_type == PT_X86 && (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) rangeset_remove_all(&pmap->pm_pkru); } static bool pmap_pkru_same(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { struct pmap_pkru_range *ppr, *prev_ppr; vm_offset_t va; PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pmap->pm_type != PT_X86 || (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) == 0 || sva >= VM_MAXUSER_ADDRESS) return (true); MPASS(eva <= VM_MAXUSER_ADDRESS); for (va = sva; va < eva; prev_ppr = ppr) { ppr = rangeset_lookup(&pmap->pm_pkru, va); if (va == sva) prev_ppr = ppr; else if ((ppr == NULL) ^ (prev_ppr == NULL)) return (false); if (ppr == NULL) { va += PAGE_SIZE; continue; } if (prev_ppr->pkru_keyidx != ppr->pkru_keyidx) return (false); va = ppr->pkru_rs_el.re_end; } return (true); } static pt_entry_t pmap_pkru_get(pmap_t pmap, vm_offset_t va) { struct pmap_pkru_range *ppr; PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pmap->pm_type != PT_X86 || (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) == 0 || va >= VM_MAXUSER_ADDRESS) return (0); ppr = rangeset_lookup(&pmap->pm_pkru, va); if (ppr != NULL) return (X86_PG_PKU(ppr->pkru_keyidx)); return (0); } static bool pred_pkru_on_remove(void *ctx __unused, void *r) { struct pmap_pkru_range *ppr; ppr = r; return ((ppr->pkru_flags & AMD64_PKRU_PERSIST) == 0); } static void pmap_pkru_on_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pmap->pm_type == PT_X86 && (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) { rangeset_remove_pred(&pmap->pm_pkru, sva, eva, pred_pkru_on_remove); } } static int pmap_pkru_copy(pmap_t dst_pmap, pmap_t src_pmap) { PMAP_LOCK_ASSERT(dst_pmap, MA_OWNED); PMAP_LOCK_ASSERT(src_pmap, MA_OWNED); MPASS(dst_pmap->pm_type == PT_X86); MPASS(src_pmap->pm_type == PT_X86); MPASS((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0); if (src_pmap->pm_pkru.rs_data_ctx == NULL) return (0); return (rangeset_copy(&dst_pmap->pm_pkru, &src_pmap->pm_pkru)); } static void pmap_pkru_update_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, u_int keyidx) { pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t newpde, ptpaddr, *pde; pt_entry_t newpte, *ptep, pte; vm_offset_t va, va_next; bool changed; PMAP_LOCK_ASSERT(pmap, MA_OWNED); MPASS(pmap->pm_type == PT_X86); MPASS(keyidx <= PMAP_MAX_PKRU_IDX); for (changed = false, va = sva; va < eva; va = va_next) { pml4e = pmap_pml4e(pmap, va); if (pml4e == NULL || (*pml4e & X86_PG_V) == 0) { va_next = (va + NBPML4) & ~PML4MASK; if (va_next < va) va_next = eva; continue; } pdpe = pmap_pml4e_to_pdpe(pml4e, va); if ((*pdpe & X86_PG_V) == 0) { va_next = (va + NBPDP) & ~PDPMASK; if (va_next < va) va_next = eva; continue; } va_next = (va + NBPDR) & ~PDRMASK; if (va_next < va) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, va); ptpaddr = *pde; if (ptpaddr == 0) continue; MPASS((ptpaddr & X86_PG_V) != 0); if ((ptpaddr & PG_PS) != 0) { if (va + NBPDR == va_next && eva >= va_next) { newpde = (ptpaddr & ~X86_PG_PKU_MASK) | X86_PG_PKU(keyidx); if (newpde != ptpaddr) { *pde = newpde; changed = true; } continue; } else if (!pmap_demote_pde(pmap, pde, va)) { continue; } } if (va_next > eva) va_next = eva; for (ptep = pmap_pde_to_pte(pde, va); va != va_next; ptep++, va += PAGE_SIZE) { pte = *ptep; if ((pte & X86_PG_V) == 0) continue; newpte = (pte & ~X86_PG_PKU_MASK) | X86_PG_PKU(keyidx); if (newpte != pte) { *ptep = newpte; changed = true; } } } if (changed) pmap_invalidate_range(pmap, sva, eva); } static int pmap_pkru_check_uargs(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, u_int keyidx, int flags) { if (pmap->pm_type != PT_X86 || keyidx > PMAP_MAX_PKRU_IDX || (flags & ~(AMD64_PKRU_PERSIST | AMD64_PKRU_EXCL)) != 0) return (EINVAL); if (eva <= sva || eva > VM_MAXUSER_ADDRESS) return (EFAULT); if ((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) == 0) return (ENOTSUP); return (0); } int pmap_pkru_set(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, u_int keyidx, int flags) { int error; sva = trunc_page(sva); eva = round_page(eva); error = pmap_pkru_check_uargs(pmap, sva, eva, keyidx, flags); if (error != 0) return (error); for (;;) { PMAP_LOCK(pmap); error = pmap_pkru_assign(pmap, sva, eva, keyidx, flags); if (error == 0) pmap_pkru_update_range(pmap, sva, eva, keyidx); PMAP_UNLOCK(pmap); if (error != ENOMEM) break; vm_wait(NULL); } return (error); } int pmap_pkru_clear(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { int error; sva = trunc_page(sva); eva = round_page(eva); error = pmap_pkru_check_uargs(pmap, sva, eva, 0, 0); if (error != 0) return (error); for (;;) { PMAP_LOCK(pmap); error = pmap_pkru_deassign(pmap, sva, eva); if (error == 0) pmap_pkru_update_range(pmap, sva, eva, 0); PMAP_UNLOCK(pmap); if (error != ENOMEM) break; vm_wait(NULL); } return (error); } #ifdef KASAN static vm_page_t pmap_kasan_enter_alloc_4k(void) { vm_page_t m; m = vm_page_alloc_noobj(VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (m == NULL) panic("%s: no memory to grow shadow map", __func__); return (m); } static vm_page_t pmap_kasan_enter_alloc_2m(void) { return (vm_page_alloc_noobj_contig(VM_ALLOC_WIRED | VM_ALLOC_ZERO, NPTEPG, 0, ~0ul, NBPDR, 0, VM_MEMATTR_DEFAULT)); } /* * Grow the shadow map by at least one 4KB page at the specified address. Use * 2MB pages when possible. */ void pmap_kasan_enter(vm_offset_t va) { pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte; vm_page_t m; mtx_assert(&kernel_map->system_mtx, MA_OWNED); pdpe = pmap_pdpe(kernel_pmap, va); if ((*pdpe & X86_PG_V) == 0) { m = pmap_kasan_enter_alloc_4k(); *pdpe = (pdp_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_V | pg_nx); } pde = pmap_pdpe_to_pde(pdpe, va); if ((*pde & X86_PG_V) == 0) { m = pmap_kasan_enter_alloc_2m(); if (m != NULL) { *pde = (pd_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_PS | X86_PG_V | X86_PG_A | X86_PG_M | pg_nx); } else { m = pmap_kasan_enter_alloc_4k(); *pde = (pd_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_V | pg_nx); } } if ((*pde & X86_PG_PS) != 0) return; pte = pmap_pde_to_pte(pde, va); if ((*pte & X86_PG_V) != 0) return; m = pmap_kasan_enter_alloc_4k(); *pte = (pt_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_V | X86_PG_M | X86_PG_A | pg_nx); } #endif #ifdef KMSAN static vm_page_t pmap_kmsan_enter_alloc_4k(void) { vm_page_t m; m = vm_page_alloc_noobj(VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (m == NULL) panic("%s: no memory to grow shadow map", __func__); return (m); } static vm_page_t pmap_kmsan_enter_alloc_2m(void) { return (vm_page_alloc_noobj_contig(VM_ALLOC_ZERO | VM_ALLOC_WIRED, NPTEPG, 0, ~0ul, NBPDR, 0, VM_MEMATTR_DEFAULT)); } /* * Grow the shadow or origin maps by at least one 4KB page at the specified * address. Use 2MB pages when possible. */ void pmap_kmsan_enter(vm_offset_t va) { pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte; vm_page_t m; mtx_assert(&kernel_map->system_mtx, MA_OWNED); pdpe = pmap_pdpe(kernel_pmap, va); if ((*pdpe & X86_PG_V) == 0) { m = pmap_kmsan_enter_alloc_4k(); *pdpe = (pdp_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_V | pg_nx); } pde = pmap_pdpe_to_pde(pdpe, va); if ((*pde & X86_PG_V) == 0) { m = pmap_kmsan_enter_alloc_2m(); if (m != NULL) { *pde = (pd_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_PS | X86_PG_V | X86_PG_A | X86_PG_M | pg_nx); } else { m = pmap_kmsan_enter_alloc_4k(); *pde = (pd_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_V | pg_nx); } } if ((*pde & X86_PG_PS) != 0) return; pte = pmap_pde_to_pte(pde, va); if ((*pte & X86_PG_V) != 0) return; m = pmap_kmsan_enter_alloc_4k(); *pte = (pt_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_V | X86_PG_M | X86_PG_A | pg_nx); } #endif /* * Track a range of the kernel's virtual address space that is contiguous * in various mapping attributes. */ struct pmap_kernel_map_range { vm_offset_t sva; pt_entry_t attrs; int ptes; int pdes; int pdpes; }; static void sysctl_kmaps_dump(struct sbuf *sb, struct pmap_kernel_map_range *range, vm_offset_t eva) { const char *mode; int i, pat_idx; if (eva <= range->sva) return; pat_idx = pmap_pat_index(kernel_pmap, range->attrs, true); for (i = 0; i < PAT_INDEX_SIZE; i++) if (pat_index[i] == pat_idx) break; switch (i) { case PAT_WRITE_BACK: mode = "WB"; break; case PAT_WRITE_THROUGH: mode = "WT"; break; case PAT_UNCACHEABLE: mode = "UC"; break; case PAT_UNCACHED: mode = "U-"; break; case PAT_WRITE_PROTECTED: mode = "WP"; break; case PAT_WRITE_COMBINING: mode = "WC"; break; default: printf("%s: unknown PAT mode %#x for range 0x%016lx-0x%016lx\n", __func__, pat_idx, range->sva, eva); mode = "??"; break; } sbuf_printf(sb, "0x%016lx-0x%016lx r%c%c%c%c %s %d %d %d\n", range->sva, eva, (range->attrs & X86_PG_RW) != 0 ? 'w' : '-', (range->attrs & pg_nx) != 0 ? '-' : 'x', (range->attrs & X86_PG_U) != 0 ? 'u' : 's', (range->attrs & X86_PG_G) != 0 ? 'g' : '-', mode, range->pdpes, range->pdes, range->ptes); /* Reset to sentinel value. */ range->sva = la57 ? KV5ADDR(NPML5EPG - 1, NPML4EPG - 1, NPDPEPG - 1, NPDEPG - 1, NPTEPG - 1) : KV4ADDR(NPML4EPG - 1, NPDPEPG - 1, NPDEPG - 1, NPTEPG - 1); } /* * Determine whether the attributes specified by a page table entry match those * being tracked by the current range. This is not quite as simple as a direct * flag comparison since some PAT modes have multiple representations. */ static bool sysctl_kmaps_match(struct pmap_kernel_map_range *range, pt_entry_t attrs) { pt_entry_t diff, mask; mask = X86_PG_G | X86_PG_RW | X86_PG_U | X86_PG_PDE_CACHE | pg_nx; diff = (range->attrs ^ attrs) & mask; if (diff == 0) return (true); if ((diff & ~X86_PG_PDE_PAT) == 0 && pmap_pat_index(kernel_pmap, range->attrs, true) == pmap_pat_index(kernel_pmap, attrs, true)) return (true); return (false); } static void sysctl_kmaps_reinit(struct pmap_kernel_map_range *range, vm_offset_t va, pt_entry_t attrs) { memset(range, 0, sizeof(*range)); range->sva = va; range->attrs = attrs; } /* * Given a leaf PTE, derive the mapping's attributes. If they do not match * those of the current run, dump the address range and its attributes, and * begin a new run. */ static void sysctl_kmaps_check(struct sbuf *sb, struct pmap_kernel_map_range *range, vm_offset_t va, pml4_entry_t pml4e, pdp_entry_t pdpe, pd_entry_t pde, pt_entry_t pte) { pt_entry_t attrs; attrs = pml4e & (X86_PG_RW | X86_PG_U | pg_nx); attrs |= pdpe & pg_nx; attrs &= pg_nx | (pdpe & (X86_PG_RW | X86_PG_U)); if ((pdpe & PG_PS) != 0) { attrs |= pdpe & (X86_PG_G | X86_PG_PDE_CACHE); } else if (pde != 0) { attrs |= pde & pg_nx; attrs &= pg_nx | (pde & (X86_PG_RW | X86_PG_U)); } if ((pde & PG_PS) != 0) { attrs |= pde & (X86_PG_G | X86_PG_PDE_CACHE); } else if (pte != 0) { attrs |= pte & pg_nx; attrs &= pg_nx | (pte & (X86_PG_RW | X86_PG_U)); attrs |= pte & (X86_PG_G | X86_PG_PTE_CACHE); /* Canonicalize by always using the PDE PAT bit. */ if ((attrs & X86_PG_PTE_PAT) != 0) attrs ^= X86_PG_PDE_PAT | X86_PG_PTE_PAT; } if (range->sva > va || !sysctl_kmaps_match(range, attrs)) { sysctl_kmaps_dump(sb, range, va); sysctl_kmaps_reinit(range, va, attrs); } } static int sysctl_kmaps(SYSCTL_HANDLER_ARGS) { struct pmap_kernel_map_range range; struct sbuf sbuf, *sb; pml4_entry_t pml4e; pdp_entry_t *pdp, pdpe; pd_entry_t *pd, pde; pt_entry_t *pt, pte; vm_offset_t sva; vm_paddr_t pa; int error, i, j, k, l; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sb = &sbuf; sbuf_new_for_sysctl(sb, NULL, PAGE_SIZE, req); /* Sentinel value. */ range.sva = la57 ? KV5ADDR(NPML5EPG - 1, NPML4EPG - 1, NPDPEPG - 1, NPDEPG - 1, NPTEPG - 1) : KV4ADDR(NPML4EPG - 1, NPDPEPG - 1, NPDEPG - 1, NPTEPG - 1); /* * Iterate over the kernel page tables without holding the kernel pmap * lock. Outside of the large map, kernel page table pages are never * freed, so at worst we will observe inconsistencies in the output. * Within the large map, ensure that PDP and PD page addresses are * valid before descending. */ for (sva = 0, i = pmap_pml4e_index(sva); i < NPML4EPG; i++) { switch (i) { case PML4PML4I: sbuf_printf(sb, "\nRecursive map:\n"); break; case DMPML4I: sbuf_printf(sb, "\nDirect map:\n"); break; #ifdef KASAN case KASANPML4I: sbuf_printf(sb, "\nKASAN shadow map:\n"); break; #endif #ifdef KMSAN case KMSANSHADPML4I: sbuf_printf(sb, "\nKMSAN shadow map:\n"); break; case KMSANORIGPML4I: sbuf_printf(sb, "\nKMSAN origin map:\n"); break; #endif case KPML4BASE: sbuf_printf(sb, "\nKernel map:\n"); break; case LMSPML4I: sbuf_printf(sb, "\nLarge map:\n"); break; } /* Convert to canonical form. */ if (sva == 1ul << 47) sva |= -1ul << 48; restart: pml4e = kernel_pml4[i]; if ((pml4e & X86_PG_V) == 0) { sva = rounddown2(sva, NBPML4); sysctl_kmaps_dump(sb, &range, sva); sva += NBPML4; continue; } pa = pml4e & PG_FRAME; pdp = (pdp_entry_t *)PHYS_TO_DMAP(pa); for (j = pmap_pdpe_index(sva); j < NPDPEPG; j++) { pdpe = pdp[j]; if ((pdpe & X86_PG_V) == 0) { sva = rounddown2(sva, NBPDP); sysctl_kmaps_dump(sb, &range, sva); sva += NBPDP; continue; } pa = pdpe & PG_FRAME; if ((pdpe & PG_PS) != 0) { sva = rounddown2(sva, NBPDP); sysctl_kmaps_check(sb, &range, sva, pml4e, pdpe, 0, 0); range.pdpes++; sva += NBPDP; continue; } if (PMAP_ADDRESS_IN_LARGEMAP(sva) && vm_phys_paddr_to_vm_page(pa) == NULL) { /* * Page table pages for the large map may be * freed. Validate the next-level address * before descending. */ goto restart; } pd = (pd_entry_t *)PHYS_TO_DMAP(pa); for (k = pmap_pde_index(sva); k < NPDEPG; k++) { pde = pd[k]; if ((pde & X86_PG_V) == 0) { sva = rounddown2(sva, NBPDR); sysctl_kmaps_dump(sb, &range, sva); sva += NBPDR; continue; } pa = pde & PG_FRAME; if ((pde & PG_PS) != 0) { sva = rounddown2(sva, NBPDR); sysctl_kmaps_check(sb, &range, sva, pml4e, pdpe, pde, 0); range.pdes++; sva += NBPDR; continue; } if (PMAP_ADDRESS_IN_LARGEMAP(sva) && vm_phys_paddr_to_vm_page(pa) == NULL) { /* * Page table pages for the large map * may be freed. Validate the * next-level address before descending. */ goto restart; } pt = (pt_entry_t *)PHYS_TO_DMAP(pa); for (l = pmap_pte_index(sva); l < NPTEPG; l++, sva += PAGE_SIZE) { pte = pt[l]; if ((pte & X86_PG_V) == 0) { sysctl_kmaps_dump(sb, &range, sva); continue; } sysctl_kmaps_check(sb, &range, sva, pml4e, pdpe, pde, pte); range.ptes++; } } } } error = sbuf_finish(sb); sbuf_delete(sb); return (error); } SYSCTL_OID(_vm_pmap, OID_AUTO, kernel_maps, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE | CTLFLAG_SKIP, NULL, 0, sysctl_kmaps, "A", "Dump kernel address layout"); #ifdef DDB DB_SHOW_COMMAND(pte, pmap_print_pte) { pmap_t pmap; pml5_entry_t *pml5; pml4_entry_t *pml4; pdp_entry_t *pdp; pd_entry_t *pde; pt_entry_t *pte, PG_V; vm_offset_t va; if (!have_addr) { db_printf("show pte addr\n"); return; } va = (vm_offset_t)addr; if (kdb_thread != NULL) pmap = vmspace_pmap(kdb_thread->td_proc->p_vmspace); else pmap = PCPU_GET(curpmap); PG_V = pmap_valid_bit(pmap); db_printf("VA 0x%016lx", va); if (pmap_is_la57(pmap)) { pml5 = pmap_pml5e(pmap, va); db_printf(" pml5e 0x%016lx", *pml5); if ((*pml5 & PG_V) == 0) { db_printf("\n"); return; } pml4 = pmap_pml5e_to_pml4e(pml5, va); } else { pml4 = pmap_pml4e(pmap, va); } db_printf(" pml4e 0x%016lx", *pml4); if ((*pml4 & PG_V) == 0) { db_printf("\n"); return; } pdp = pmap_pml4e_to_pdpe(pml4, va); db_printf(" pdpe 0x%016lx", *pdp); if ((*pdp & PG_V) == 0 || (*pdp & PG_PS) != 0) { db_printf("\n"); return; } pde = pmap_pdpe_to_pde(pdp, va); db_printf(" pde 0x%016lx", *pde); if ((*pde & PG_V) == 0 || (*pde & PG_PS) != 0) { db_printf("\n"); return; } pte = pmap_pde_to_pte(pde, va); db_printf(" pte 0x%016lx\n", *pte); } DB_SHOW_COMMAND(phys2dmap, pmap_phys2dmap) { vm_paddr_t a; if (have_addr) { a = (vm_paddr_t)addr; db_printf("0x%jx\n", (uintmax_t)PHYS_TO_DMAP(a)); } else { db_printf("show phys2dmap addr\n"); } } static void ptpages_show_page(int level, int idx, vm_page_t pg) { db_printf("l %d i %d pg %p phys %#lx ref %x\n", level, idx, pg, VM_PAGE_TO_PHYS(pg), pg->ref_count); } static void ptpages_show_complain(int level, int idx, uint64_t pte) { db_printf("l %d i %d pte %#lx\n", level, idx, pte); } static void ptpages_show_pml4(vm_page_t pg4, int num_entries, uint64_t PG_V) { vm_page_t pg3, pg2, pg1; pml4_entry_t *pml4; pdp_entry_t *pdp; pd_entry_t *pd; int i4, i3, i2; pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pg4)); for (i4 = 0; i4 < num_entries; i4++) { if ((pml4[i4] & PG_V) == 0) continue; pg3 = PHYS_TO_VM_PAGE(pml4[i4] & PG_FRAME); if (pg3 == NULL) { ptpages_show_complain(3, i4, pml4[i4]); continue; } ptpages_show_page(3, i4, pg3); pdp = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pg3)); for (i3 = 0; i3 < NPDPEPG; i3++) { if ((pdp[i3] & PG_V) == 0) continue; pg2 = PHYS_TO_VM_PAGE(pdp[i3] & PG_FRAME); if (pg3 == NULL) { ptpages_show_complain(2, i3, pdp[i3]); continue; } ptpages_show_page(2, i3, pg2); pd = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pg2)); for (i2 = 0; i2 < NPDEPG; i2++) { if ((pd[i2] & PG_V) == 0) continue; pg1 = PHYS_TO_VM_PAGE(pd[i2] & PG_FRAME); if (pg1 == NULL) { ptpages_show_complain(1, i2, pd[i2]); continue; } ptpages_show_page(1, i2, pg1); } } } } DB_SHOW_COMMAND(ptpages, pmap_ptpages) { pmap_t pmap; vm_page_t pg; pml5_entry_t *pml5; uint64_t PG_V; int i5; if (have_addr) pmap = (pmap_t)addr; else pmap = PCPU_GET(curpmap); PG_V = pmap_valid_bit(pmap); if (pmap_is_la57(pmap)) { pml5 = pmap->pm_pmltop; for (i5 = 0; i5 < NUPML5E; i5++) { if ((pml5[i5] & PG_V) == 0) continue; pg = PHYS_TO_VM_PAGE(pml5[i5] & PG_FRAME); if (pg == NULL) { ptpages_show_complain(4, i5, pml5[i5]); continue; } ptpages_show_page(4, i5, pg); ptpages_show_pml4(pg, NPML4EPG, PG_V); } } else { ptpages_show_pml4(PHYS_TO_VM_PAGE(DMAP_TO_PHYS( (vm_offset_t)pmap->pm_pmltop)), NUP4ML4E, PG_V); } } #endif diff --git a/sys/mips/include/pmap.h b/sys/mips/include/pmap.h index 7161a02481c7..fb186011bd85 100644 --- a/sys/mips/include/pmap.h +++ b/sys/mips/include/pmap.h @@ -1,199 +1,205 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department and William Jolitz of UUNET Technologies Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * Derived from hp300 version by Mike Hibler, this version by William * Jolitz uses a recursive map [a pde points to the page directory] to * map the page tables using the pagetables themselves. This is done to * reduce the impact on kernel virtual memory for lots of sparse address * space, and to reduce the cost of memory to each process. * * from: hp300: @(#)pmap.h 7.2 (Berkeley) 12/16/90 * from: @(#)pmap.h 7.4 (Berkeley) 5/12/91 * from: src/sys/i386/include/pmap.h,v 1.65.2.2 2000/11/30 01:54:42 peter * JNPR: pmap.h,v 1.7.2.1 2007/09/10 07:44:12 girish * $FreeBSD$ */ #ifndef _MACHINE_PMAP_H_ #define _MACHINE_PMAP_H_ #include #include #if defined(__mips_n32) || defined(__mips_n64) /* PHYSADDR_64BIT */ #define NKPT 256 /* mem > 4G, vm_page_startup needs more KPTs */ #else #define NKPT 120 /* actual number of kernel page tables */ #endif #ifndef LOCORE #include #include #include #include /* * Pmap stuff */ struct pv_entry; struct pv_chunk; struct md_page { int pv_flags; TAILQ_HEAD(, pv_entry) pv_list; }; #define PV_TABLE_REF 0x02 /* referenced */ #define PV_MEMATTR_MASK 0xf0 /* store vm_memattr_t here */ #define PV_MEMATTR_SHIFT 0x04 #define ASID_BITS 8 #define ASIDGEN_BITS (32 - ASID_BITS) #define ASIDGEN_MASK ((1 << ASIDGEN_BITS) - 1) struct pmap { pd_entry_t *pm_segtab; /* KVA of segment table */ TAILQ_HEAD(, pv_chunk) pm_pvchunk; /* list of mappings in pmap */ cpuset_t pm_active; /* active on cpus */ struct { u_int32_t asid:ASID_BITS; /* TLB address space tag */ u_int32_t gen:ASIDGEN_BITS; /* its generation number */ } pm_asid[MAXSMPCPU]; struct pmap_statistics pm_stats; /* pmap statistics */ struct mtx pm_mtx; }; typedef struct pmap *pmap_t; #ifdef _KERNEL pt_entry_t *pmap_pte(pmap_t, vm_offset_t); vm_paddr_t pmap_kextract(vm_offset_t va); #define vtophys(va) pmap_kextract(((vm_offset_t) (va))) #define pmap_asid(pmap) (pmap)->pm_asid[PCPU_GET(cpuid)].asid extern struct pmap kernel_pmap_store; #define kernel_pmap (&kernel_pmap_store) #define PMAP_LOCK(pmap) mtx_lock(&(pmap)->pm_mtx) #define PMAP_LOCK_ASSERT(pmap, type) mtx_assert(&(pmap)->pm_mtx, (type)) #define PMAP_LOCK_DESTROY(pmap) mtx_destroy(&(pmap)->pm_mtx) #define PMAP_LOCK_INIT(pmap) mtx_init(&(pmap)->pm_mtx, "pmap", \ NULL, MTX_DEF) #define PMAP_LOCKED(pmap) mtx_owned(&(pmap)->pm_mtx) #define PMAP_MTX(pmap) (&(pmap)->pm_mtx) #define PMAP_TRYLOCK(pmap) mtx_trylock(&(pmap)->pm_mtx) #define PMAP_UNLOCK(pmap) mtx_unlock(&(pmap)->pm_mtx) /* * For each vm_page_t, there is a list of all currently valid virtual * mappings of that page. An entry is a pv_entry_t, the list is pv_table. */ typedef struct pv_entry { vm_offset_t pv_va; /* virtual address for mapping */ TAILQ_ENTRY(pv_entry) pv_list; } *pv_entry_t; /* * pv_entries are allocated in chunks per-process. This avoids the * need to track per-pmap assignments. */ #ifdef __mips_n64 #define _NPCM 3 #define _NPCPV 168 #else #define _NPCM 11 #define _NPCPV 336 #endif struct pv_chunk { pmap_t pc_pmap; TAILQ_ENTRY(pv_chunk) pc_list; u_long pc_map[_NPCM]; /* bitmap; 1 = free */ TAILQ_ENTRY(pv_chunk) pc_lru; struct pv_entry pc_pventry[_NPCPV]; }; /* * physmem_desc[] is a superset of phys_avail[] and describes all the * memory present in the system. * * phys_avail[] is similar but does not include the memory stolen by * pmap_steal_memory(). * * Each memory region is described by a pair of elements in the array * so we can describe up to (PHYS_AVAIL_ENTRIES / 2) distinct memory * regions. */ extern vm_paddr_t physmem_desc[PHYS_AVAIL_COUNT]; extern vm_offset_t virtual_avail; extern vm_offset_t virtual_end; #define pmap_page_get_memattr(m) (((m)->md.pv_flags & PV_MEMATTR_MASK) >> PV_MEMATTR_SHIFT) #define pmap_page_is_mapped(m) (!TAILQ_EMPTY(&(m)->md.pv_list)) #define pmap_page_is_write_mapped(m) (((m)->a.flags & PGA_WRITEABLE) != 0) #define pmap_vm_page_alloc_check(m) void pmap_bootstrap(void); void *pmap_mapdev(vm_paddr_t, vm_size_t); void *pmap_mapdev_attr(vm_paddr_t, vm_size_t, vm_memattr_t); void pmap_unmapdev(vm_offset_t, vm_size_t); vm_offset_t pmap_steal_memory(vm_size_t size); void pmap_kenter(vm_offset_t va, vm_paddr_t pa); void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, vm_memattr_t attr); void pmap_kenter_device(vm_offset_t, vm_size_t, vm_paddr_t); void pmap_kremove(vm_offset_t va); void pmap_kremove_device(vm_offset_t, vm_size_t); void *pmap_kenter_temporary(vm_paddr_t pa, int i); void pmap_kenter_temporary_free(vm_paddr_t pa); void pmap_flush_pvcache(vm_page_t m); int pmap_emulate_modified(pmap_t pmap, vm_offset_t va); void pmap_page_set_memattr(vm_page_t, vm_memattr_t); int pmap_change_attr(vm_offset_t, vm_size_t, vm_memattr_t); static inline int pmap_vmspace_copy(pmap_t dst_pmap __unused, pmap_t src_pmap __unused) { return (0); } +static inline bool +pmap_ps_enabled(pmap_t pmap __unused) +{ + return (false); +} + #endif /* _KERNEL */ #endif /* !LOCORE */ #endif /* !_MACHINE_PMAP_H_ */ diff --git a/sys/powerpc/aim/mmu_oea.c b/sys/powerpc/aim/mmu_oea.c index 44d8b9518b8e..cb0c905a53f3 100644 --- a/sys/powerpc/aim/mmu_oea.c +++ b/sys/powerpc/aim/mmu_oea.c @@ -1,2829 +1,2837 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD AND BSD-4-Clause * * Copyright (c) 2001 The NetBSD Foundation, Inc. * All rights reserved. * * This code is derived from software contributed to The NetBSD Foundation * by Matt Thomas of Allegro Networks, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE * POSSIBILITY OF SUCH DAMAGE. */ /*- * Copyright (C) 1995, 1996 Wolfgang Solfrank. * Copyright (C) 1995, 1996 TooLs GmbH. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by TooLs GmbH. * 4. The name of TooLs GmbH may not be used to endorse or promote products * derived from this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $ */ /*- * Copyright (C) 2001 Benno Rice. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); /* * Manages physical address maps. * * Since the information managed by this module is also stored by the * logical address mapping module, this module may throw away valid virtual * to physical mappings at almost any time. However, invalidations of * mappings must be done as requested. * * In order to cope with hardware architectures which make virtual to * physical map invalidates expensive, this module may delay invalidate * reduced protection operations until such time as they are actually * necessary. This module is given full information as to which processors * are currently using which maps, and to when physical maps must be made * correct. */ #include "opt_kstack_pages.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define MOEA_DEBUG #define TODO panic("%s: not implemented", __func__); #define VSID_MAKE(sr, hash) ((sr) | (((hash) & 0xfffff) << 4)) #define VSID_TO_SR(vsid) ((vsid) & 0xf) #define VSID_TO_HASH(vsid) (((vsid) >> 4) & 0xfffff) /* Get physical address from PVO. */ #define PVO_PADDR(pvo) ((pvo)->pvo_pte.pte.pte_lo & PTE_RPGN) struct ofw_map { vm_offset_t om_va; vm_size_t om_len; vm_offset_t om_pa; u_int om_mode; }; extern unsigned char _etext[]; extern unsigned char _end[]; /* * Map of physical memory regions. */ static struct mem_region *regions; static struct mem_region *pregions; static u_int phys_avail_count; static int regions_sz, pregions_sz; static struct ofw_map *translations; /* * Lock for the pteg and pvo tables. */ struct mtx moea_table_mutex; struct mtx moea_vsid_mutex; /* tlbie instruction synchronization */ static struct mtx tlbie_mtx; /* * PTEG data. */ static struct pteg *moea_pteg_table; u_int moea_pteg_count; u_int moea_pteg_mask; /* * PVO data. */ struct pvo_head *moea_pvo_table; /* pvo entries by pteg index */ struct pvo_head moea_pvo_kunmanaged = LIST_HEAD_INITIALIZER(moea_pvo_kunmanaged); /* list of unmanaged pages */ static struct rwlock_padalign pvh_global_lock; uma_zone_t moea_upvo_zone; /* zone for pvo entries for unmanaged pages */ uma_zone_t moea_mpvo_zone; /* zone for pvo entries for managed pages */ #define BPVO_POOL_SIZE 32768 static struct pvo_entry *moea_bpvo_pool; static int moea_bpvo_pool_index = 0; #define VSID_NBPW (sizeof(u_int32_t) * 8) static u_int moea_vsid_bitmap[NPMAPS / VSID_NBPW]; static boolean_t moea_initialized = FALSE; /* * Statistics. */ u_int moea_pte_valid = 0; u_int moea_pte_overflow = 0; u_int moea_pte_replacements = 0; u_int moea_pvo_entries = 0; u_int moea_pvo_enter_calls = 0; u_int moea_pvo_remove_calls = 0; u_int moea_pte_spills = 0; SYSCTL_INT(_machdep, OID_AUTO, moea_pte_valid, CTLFLAG_RD, &moea_pte_valid, 0, ""); SYSCTL_INT(_machdep, OID_AUTO, moea_pte_overflow, CTLFLAG_RD, &moea_pte_overflow, 0, ""); SYSCTL_INT(_machdep, OID_AUTO, moea_pte_replacements, CTLFLAG_RD, &moea_pte_replacements, 0, ""); SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_entries, CTLFLAG_RD, &moea_pvo_entries, 0, ""); SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_enter_calls, CTLFLAG_RD, &moea_pvo_enter_calls, 0, ""); SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_remove_calls, CTLFLAG_RD, &moea_pvo_remove_calls, 0, ""); SYSCTL_INT(_machdep, OID_AUTO, moea_pte_spills, CTLFLAG_RD, &moea_pte_spills, 0, ""); /* * Allocate physical memory for use in moea_bootstrap. */ static vm_offset_t moea_bootstrap_alloc(vm_size_t, u_int); /* * PTE calls. */ static int moea_pte_insert(u_int, struct pte *); /* * PVO calls. */ static int moea_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *, vm_offset_t, vm_paddr_t, u_int, int); static void moea_pvo_remove(struct pvo_entry *, int); static struct pvo_entry *moea_pvo_find_va(pmap_t, vm_offset_t, int *); static struct pte *moea_pvo_to_pte(const struct pvo_entry *, int); /* * Utility routines. */ static int moea_enter_locked(pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int, int8_t); static void moea_syncicache(vm_paddr_t, vm_size_t); static boolean_t moea_query_bit(vm_page_t, int); static u_int moea_clear_bit(vm_page_t, int); static void moea_kremove(vm_offset_t); int moea_pte_spill(vm_offset_t); /* * Kernel MMU interface */ void moea_clear_modify(vm_page_t); void moea_copy_page(vm_page_t, vm_page_t); void moea_copy_pages(vm_page_t *ma, vm_offset_t a_offset, vm_page_t *mb, vm_offset_t b_offset, int xfersize); int moea_enter(pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int, int8_t); void moea_enter_object(pmap_t, vm_offset_t, vm_offset_t, vm_page_t, vm_prot_t); void moea_enter_quick(pmap_t, vm_offset_t, vm_page_t, vm_prot_t); vm_paddr_t moea_extract(pmap_t, vm_offset_t); vm_page_t moea_extract_and_hold(pmap_t, vm_offset_t, vm_prot_t); void moea_init(void); boolean_t moea_is_modified(vm_page_t); boolean_t moea_is_prefaultable(pmap_t, vm_offset_t); boolean_t moea_is_referenced(vm_page_t); int moea_ts_referenced(vm_page_t); vm_offset_t moea_map(vm_offset_t *, vm_paddr_t, vm_paddr_t, int); static int moea_mincore(pmap_t, vm_offset_t, vm_paddr_t *); boolean_t moea_page_exists_quick(pmap_t, vm_page_t); void moea_page_init(vm_page_t); int moea_page_wired_mappings(vm_page_t); int moea_pinit(pmap_t); void moea_pinit0(pmap_t); void moea_protect(pmap_t, vm_offset_t, vm_offset_t, vm_prot_t); void moea_qenter(vm_offset_t, vm_page_t *, int); void moea_qremove(vm_offset_t, int); void moea_release(pmap_t); void moea_remove(pmap_t, vm_offset_t, vm_offset_t); void moea_remove_all(vm_page_t); void moea_remove_write(vm_page_t); void moea_unwire(pmap_t, vm_offset_t, vm_offset_t); void moea_zero_page(vm_page_t); void moea_zero_page_area(vm_page_t, int, int); void moea_activate(struct thread *); void moea_deactivate(struct thread *); void moea_cpu_bootstrap(int); void moea_bootstrap(vm_offset_t, vm_offset_t); void *moea_mapdev(vm_paddr_t, vm_size_t); void *moea_mapdev_attr(vm_paddr_t, vm_size_t, vm_memattr_t); void moea_unmapdev(vm_offset_t, vm_size_t); vm_paddr_t moea_kextract(vm_offset_t); void moea_kenter_attr(vm_offset_t, vm_paddr_t, vm_memattr_t); void moea_kenter(vm_offset_t, vm_paddr_t); void moea_page_set_memattr(vm_page_t m, vm_memattr_t ma); boolean_t moea_dev_direct_mapped(vm_paddr_t, vm_size_t); static void moea_sync_icache(pmap_t, vm_offset_t, vm_size_t); void moea_dumpsys_map(vm_paddr_t pa, size_t sz, void **va); void moea_scan_init(void); vm_offset_t moea_quick_enter_page(vm_page_t m); void moea_quick_remove_page(vm_offset_t addr); boolean_t moea_page_is_mapped(vm_page_t m); +bool moea_ps_enabled(pmap_t pmap); static int moea_map_user_ptr(pmap_t pm, volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen); static int moea_decode_kernel_ptr(vm_offset_t addr, int *is_user, vm_offset_t *decoded_addr); static struct pmap_funcs moea_methods = { .clear_modify = moea_clear_modify, .copy_page = moea_copy_page, .copy_pages = moea_copy_pages, .enter = moea_enter, .enter_object = moea_enter_object, .enter_quick = moea_enter_quick, .extract = moea_extract, .extract_and_hold = moea_extract_and_hold, .init = moea_init, .is_modified = moea_is_modified, .is_prefaultable = moea_is_prefaultable, .is_referenced = moea_is_referenced, .ts_referenced = moea_ts_referenced, .map = moea_map, .page_exists_quick = moea_page_exists_quick, .page_init = moea_page_init, .page_wired_mappings = moea_page_wired_mappings, .pinit = moea_pinit, .pinit0 = moea_pinit0, .protect = moea_protect, .qenter = moea_qenter, .qremove = moea_qremove, .release = moea_release, .remove = moea_remove, .remove_all = moea_remove_all, .mincore = moea_mincore, .remove_write = moea_remove_write, .sync_icache = moea_sync_icache, .unwire = moea_unwire, .zero_page = moea_zero_page, .zero_page_area = moea_zero_page_area, .activate = moea_activate, .deactivate = moea_deactivate, .page_set_memattr = moea_page_set_memattr, .quick_enter_page = moea_quick_enter_page, .quick_remove_page = moea_quick_remove_page, .page_is_mapped = moea_page_is_mapped, + .ps_enabled = moea_ps_enabled, /* Internal interfaces */ .bootstrap = moea_bootstrap, .cpu_bootstrap = moea_cpu_bootstrap, .mapdev_attr = moea_mapdev_attr, .mapdev = moea_mapdev, .unmapdev = moea_unmapdev, .kextract = moea_kextract, .kenter = moea_kenter, .kenter_attr = moea_kenter_attr, .dev_direct_mapped = moea_dev_direct_mapped, .dumpsys_pa_init = moea_scan_init, .dumpsys_map_chunk = moea_dumpsys_map, .map_user_ptr = moea_map_user_ptr, .decode_kernel_ptr = moea_decode_kernel_ptr, }; MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods); static __inline uint32_t moea_calc_wimg(vm_paddr_t pa, vm_memattr_t ma) { uint32_t pte_lo; int i; if (ma != VM_MEMATTR_DEFAULT) { switch (ma) { case VM_MEMATTR_UNCACHEABLE: return (PTE_I | PTE_G); case VM_MEMATTR_CACHEABLE: return (PTE_M); case VM_MEMATTR_WRITE_COMBINING: case VM_MEMATTR_WRITE_BACK: case VM_MEMATTR_PREFETCHABLE: return (PTE_I); case VM_MEMATTR_WRITE_THROUGH: return (PTE_W | PTE_M); } } /* * Assume the page is cache inhibited and access is guarded unless * it's in our available memory array. */ pte_lo = PTE_I | PTE_G; for (i = 0; i < pregions_sz; i++) { if ((pa >= pregions[i].mr_start) && (pa < (pregions[i].mr_start + pregions[i].mr_size))) { pte_lo = PTE_M; break; } } return pte_lo; } /* * Translate OFW translations into VM attributes. */ static __inline vm_memattr_t moea_bootstrap_convert_wimg(uint32_t mode) { switch (mode) { case (PTE_I | PTE_G): /* PCI device memory */ return VM_MEMATTR_UNCACHEABLE; case (PTE_M): /* Explicitly coherent */ return VM_MEMATTR_CACHEABLE; case 0: /* Default claim */ case 2: /* Alternate PP bits set by OF for the original payload */ /* "Normal" memory. */ return VM_MEMATTR_DEFAULT; default: /* Err on the side of caution for unknowns */ /* XXX should we panic instead? */ return VM_MEMATTR_UNCACHEABLE; } } static void tlbie(vm_offset_t va) { mtx_lock_spin(&tlbie_mtx); __asm __volatile("ptesync"); __asm __volatile("tlbie %0" :: "r"(va)); __asm __volatile("eieio; tlbsync; ptesync"); mtx_unlock_spin(&tlbie_mtx); } static void tlbia(void) { vm_offset_t va; for (va = 0; va < 0x00040000; va += 0x00001000) { __asm __volatile("tlbie %0" :: "r"(va)); powerpc_sync(); } __asm __volatile("tlbsync"); powerpc_sync(); } static __inline int va_to_sr(u_int *sr, vm_offset_t va) { return (sr[(uintptr_t)va >> ADDR_SR_SHFT]); } static __inline u_int va_to_pteg(u_int sr, vm_offset_t addr) { u_int hash; hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >> ADDR_PIDX_SHFT); return (hash & moea_pteg_mask); } static __inline struct pvo_head * vm_page_to_pvoh(vm_page_t m) { return (&m->md.mdpg_pvoh); } static __inline void moea_attr_clear(vm_page_t m, int ptebit) { rw_assert(&pvh_global_lock, RA_WLOCKED); m->md.mdpg_attrs &= ~ptebit; } static __inline int moea_attr_fetch(vm_page_t m) { return (m->md.mdpg_attrs); } static __inline void moea_attr_save(vm_page_t m, int ptebit) { rw_assert(&pvh_global_lock, RA_WLOCKED); m->md.mdpg_attrs |= ptebit; } static __inline int moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt) { if (pt->pte_hi == pvo_pt->pte_hi) return (1); return (0); } static __inline int moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which) { return (pt->pte_hi & ~PTE_VALID) == (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) | ((va >> ADDR_API_SHFT) & PTE_API) | which); } static __inline void moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo) { mtx_assert(&moea_table_mutex, MA_OWNED); /* * Construct a PTE. Default to IMB initially. Valid bit only gets * set when the real pte is set in memory. * * Note: Don't set the valid bit for correct operation of tlb update. */ pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) | (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API); pt->pte_lo = pte_lo; } static __inline void moea_pte_synch(struct pte *pt, struct pte *pvo_pt) { mtx_assert(&moea_table_mutex, MA_OWNED); pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG); } static __inline void moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit) { mtx_assert(&moea_table_mutex, MA_OWNED); /* * As shown in Section 7.6.3.2.3 */ pt->pte_lo &= ~ptebit; tlbie(va); } static __inline void moea_pte_set(struct pte *pt, struct pte *pvo_pt) { mtx_assert(&moea_table_mutex, MA_OWNED); pvo_pt->pte_hi |= PTE_VALID; /* * Update the PTE as defined in section 7.6.3.1. * Note that the REF/CHG bits are from pvo_pt and thus should have * been saved so this routine can restore them (if desired). */ pt->pte_lo = pvo_pt->pte_lo; powerpc_sync(); pt->pte_hi = pvo_pt->pte_hi; powerpc_sync(); moea_pte_valid++; } static __inline void moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va) { mtx_assert(&moea_table_mutex, MA_OWNED); pvo_pt->pte_hi &= ~PTE_VALID; /* * Force the reg & chg bits back into the PTEs. */ powerpc_sync(); /* * Invalidate the pte. */ pt->pte_hi &= ~PTE_VALID; tlbie(va); /* * Save the reg & chg bits. */ moea_pte_synch(pt, pvo_pt); moea_pte_valid--; } static __inline void moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va) { /* * Invalidate the PTE */ moea_pte_unset(pt, pvo_pt, va); moea_pte_set(pt, pvo_pt); } /* * Quick sort callout for comparing memory regions. */ static int om_cmp(const void *a, const void *b); static int om_cmp(const void *a, const void *b) { const struct ofw_map *mapa; const struct ofw_map *mapb; mapa = a; mapb = b; if (mapa->om_pa < mapb->om_pa) return (-1); else if (mapa->om_pa > mapb->om_pa) return (1); else return (0); } void moea_cpu_bootstrap(int ap) { u_int sdr; int i; if (ap) { powerpc_sync(); __asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu)); __asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl)); isync(); __asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu)); __asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl)); isync(); } __asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu)); __asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl)); isync(); __asm __volatile("mtibatu 1,%0" :: "r"(0)); __asm __volatile("mtdbatu 2,%0" :: "r"(0)); __asm __volatile("mtibatu 2,%0" :: "r"(0)); __asm __volatile("mtdbatu 3,%0" :: "r"(0)); __asm __volatile("mtibatu 3,%0" :: "r"(0)); isync(); for (i = 0; i < 16; i++) mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]); powerpc_sync(); sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10); __asm __volatile("mtsdr1 %0" :: "r"(sdr)); isync(); tlbia(); } void moea_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend) { ihandle_t mmui; phandle_t chosen, mmu; int sz; int i, j; vm_size_t size, physsz, hwphyssz; vm_offset_t pa, va, off; void *dpcpu; /* * Map PCI memory space. */ battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW); battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs); battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW); battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs); battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW); battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs); battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW); battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs); powerpc_sync(); /* map pci space */ __asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu)); __asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl)); isync(); /* set global direct map flag */ hw_direct_map = 1; mem_regions(&pregions, &pregions_sz, ®ions, ®ions_sz); CTR0(KTR_PMAP, "moea_bootstrap: physical memory"); for (i = 0; i < pregions_sz; i++) { vm_offset_t pa; vm_offset_t end; CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)", pregions[i].mr_start, pregions[i].mr_start + pregions[i].mr_size, pregions[i].mr_size); /* * Install entries into the BAT table to allow all * of physmem to be convered by on-demand BAT entries. * The loop will sometimes set the same battable element * twice, but that's fine since they won't be used for * a while yet. */ pa = pregions[i].mr_start & 0xf0000000; end = pregions[i].mr_start + pregions[i].mr_size; do { u_int n = pa >> ADDR_SR_SHFT; battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW); battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs); pa += SEGMENT_LENGTH; } while (pa < end); } if (PHYS_AVAIL_ENTRIES < regions_sz) panic("moea_bootstrap: phys_avail too small"); phys_avail_count = 0; physsz = 0; hwphyssz = 0; TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz); for (i = 0, j = 0; i < regions_sz; i++, j += 2) { CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start, regions[i].mr_start + regions[i].mr_size, regions[i].mr_size); if (hwphyssz != 0 && (physsz + regions[i].mr_size) >= hwphyssz) { if (physsz < hwphyssz) { phys_avail[j] = regions[i].mr_start; phys_avail[j + 1] = regions[i].mr_start + hwphyssz - physsz; physsz = hwphyssz; phys_avail_count++; } break; } phys_avail[j] = regions[i].mr_start; phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size; phys_avail_count++; physsz += regions[i].mr_size; } /* Check for overlap with the kernel and exception vectors */ for (j = 0; j < 2*phys_avail_count; j+=2) { if (phys_avail[j] < EXC_LAST) phys_avail[j] += EXC_LAST; if (kernelstart >= phys_avail[j] && kernelstart < phys_avail[j+1]) { if (kernelend < phys_avail[j+1]) { phys_avail[2*phys_avail_count] = (kernelend & ~PAGE_MASK) + PAGE_SIZE; phys_avail[2*phys_avail_count + 1] = phys_avail[j+1]; phys_avail_count++; } phys_avail[j+1] = kernelstart & ~PAGE_MASK; } if (kernelend >= phys_avail[j] && kernelend < phys_avail[j+1]) { if (kernelstart > phys_avail[j]) { phys_avail[2*phys_avail_count] = phys_avail[j]; phys_avail[2*phys_avail_count + 1] = kernelstart & ~PAGE_MASK; phys_avail_count++; } phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE; } } physmem = btoc(physsz); /* * Allocate PTEG table. */ #ifdef PTEGCOUNT moea_pteg_count = PTEGCOUNT; #else moea_pteg_count = 0x1000; while (moea_pteg_count < physmem) moea_pteg_count <<= 1; moea_pteg_count >>= 1; #endif /* PTEGCOUNT */ size = moea_pteg_count * sizeof(struct pteg); CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count, size); moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size); CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table); bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg)); moea_pteg_mask = moea_pteg_count - 1; /* * Allocate pv/overflow lists. */ size = sizeof(struct pvo_head) * moea_pteg_count; moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size, PAGE_SIZE); CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table); for (i = 0; i < moea_pteg_count; i++) LIST_INIT(&moea_pvo_table[i]); /* * Initialize the lock that synchronizes access to the pteg and pvo * tables. */ mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF | MTX_RECURSE); mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF); mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN); /* * Initialise the unmanaged pvo pool. */ moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc( BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0); moea_bpvo_pool_index = 0; /* * Make sure kernel vsid is allocated as well as VSID 0. */ moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW] |= 1 << (KERNEL_VSIDBITS % VSID_NBPW); moea_vsid_bitmap[0] |= 1; /* * Initialize the kernel pmap (which is statically allocated). */ PMAP_LOCK_INIT(kernel_pmap); for (i = 0; i < 16; i++) kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i; CPU_FILL(&kernel_pmap->pm_active); RB_INIT(&kernel_pmap->pmap_pvo); /* * Initialize the global pv list lock. */ rw_init(&pvh_global_lock, "pmap pv global"); /* * Set up the Open Firmware mappings */ chosen = OF_finddevice("/chosen"); if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1 && (mmu = OF_instance_to_package(mmui)) != -1 && (sz = OF_getproplen(mmu, "translations")) != -1) { translations = NULL; for (i = 0; phys_avail[i] != 0; i += 2) { if (phys_avail[i + 1] >= sz) { translations = (struct ofw_map *)phys_avail[i]; break; } } if (translations == NULL) panic("moea_bootstrap: no space to copy translations"); bzero(translations, sz); if (OF_getprop(mmu, "translations", translations, sz) == -1) panic("moea_bootstrap: can't get ofw translations"); CTR0(KTR_PMAP, "moea_bootstrap: translations"); sz /= sizeof(*translations); qsort(translations, sz, sizeof (*translations), om_cmp); for (i = 0; i < sz; i++) { CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x", translations[i].om_pa, translations[i].om_va, translations[i].om_len); /* * If the mapping is 1:1, let the RAM and device * on-demand BAT tables take care of the translation. * * However, always enter mappings for segment 16, * which is mixed-protection and therefore not * compatible with a BAT entry. */ if ((translations[i].om_va >> ADDR_SR_SHFT) != 0xf && translations[i].om_va == translations[i].om_pa) continue; /* Enter the pages */ for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) moea_kenter_attr(translations[i].om_va + off, translations[i].om_pa + off, moea_bootstrap_convert_wimg(translations[i].om_mode)); } } /* * Calculate the last available physical address. */ for (i = 0; phys_avail[i + 2] != 0; i += 2) ; Maxmem = powerpc_btop(phys_avail[i + 1]); moea_cpu_bootstrap(0); mtmsr(mfmsr() | PSL_DR | PSL_IR); pmap_bootstrapped++; /* * Set the start and end of kva. */ virtual_avail = VM_MIN_KERNEL_ADDRESS; virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS; /* * Allocate a kernel stack with a guard page for thread0 and map it * into the kernel page map. */ pa = moea_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE); va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE; virtual_avail = va + kstack_pages * PAGE_SIZE; CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va); thread0.td_kstack = va; thread0.td_kstack_pages = kstack_pages; for (i = 0; i < kstack_pages; i++) { moea_kenter(va, pa); pa += PAGE_SIZE; va += PAGE_SIZE; } /* * Allocate virtual address space for the message buffer. */ pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE); msgbufp = (struct msgbuf *)virtual_avail; va = virtual_avail; virtual_avail += round_page(msgbufsize); while (va < virtual_avail) { moea_kenter(va, pa); pa += PAGE_SIZE; va += PAGE_SIZE; } /* * Allocate virtual address space for the dynamic percpu area. */ pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE); dpcpu = (void *)virtual_avail; va = virtual_avail; virtual_avail += DPCPU_SIZE; while (va < virtual_avail) { moea_kenter(va, pa); pa += PAGE_SIZE; va += PAGE_SIZE; } dpcpu_init(dpcpu, 0); } /* * Activate a user pmap. The pmap must be activated before it's address * space can be accessed in any way. */ void moea_activate(struct thread *td) { pmap_t pm, pmr; /* * Load all the data we need up front to encourage the compiler to * not issue any loads while we have interrupts disabled below. */ pm = &td->td_proc->p_vmspace->vm_pmap; pmr = pm->pmap_phys; CPU_SET(PCPU_GET(cpuid), &pm->pm_active); PCPU_SET(curpmap, pmr); mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid); } void moea_deactivate(struct thread *td) { pmap_t pm; pm = &td->td_proc->p_vmspace->vm_pmap; CPU_CLR(PCPU_GET(cpuid), &pm->pm_active); PCPU_SET(curpmap, NULL); } void moea_unwire(pmap_t pm, vm_offset_t sva, vm_offset_t eva) { struct pvo_entry key, *pvo; PMAP_LOCK(pm); key.pvo_vaddr = sva; for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key); pvo != NULL && PVO_VADDR(pvo) < eva; pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) { if ((pvo->pvo_vaddr & PVO_WIRED) == 0) panic("moea_unwire: pvo %p is missing PVO_WIRED", pvo); pvo->pvo_vaddr &= ~PVO_WIRED; pm->pm_stats.wired_count--; } PMAP_UNLOCK(pm); } void moea_copy_page(vm_page_t msrc, vm_page_t mdst) { vm_offset_t dst; vm_offset_t src; dst = VM_PAGE_TO_PHYS(mdst); src = VM_PAGE_TO_PHYS(msrc); bcopy((void *)src, (void *)dst, PAGE_SIZE); } void moea_copy_pages(vm_page_t *ma, vm_offset_t a_offset, vm_page_t *mb, vm_offset_t b_offset, int xfersize) { void *a_cp, *b_cp; vm_offset_t a_pg_offset, b_pg_offset; int cnt; while (xfersize > 0) { a_pg_offset = a_offset & PAGE_MASK; cnt = min(xfersize, PAGE_SIZE - a_pg_offset); a_cp = (char *)VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]) + a_pg_offset; b_pg_offset = b_offset & PAGE_MASK; cnt = min(cnt, PAGE_SIZE - b_pg_offset); b_cp = (char *)VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]) + b_pg_offset; bcopy(a_cp, b_cp, cnt); a_offset += cnt; b_offset += cnt; xfersize -= cnt; } } /* * Zero a page of physical memory by temporarily mapping it into the tlb. */ void moea_zero_page(vm_page_t m) { vm_offset_t off, pa = VM_PAGE_TO_PHYS(m); for (off = 0; off < PAGE_SIZE; off += cacheline_size) __asm __volatile("dcbz 0,%0" :: "r"(pa + off)); } void moea_zero_page_area(vm_page_t m, int off, int size) { vm_offset_t pa = VM_PAGE_TO_PHYS(m); void *va = (void *)(pa + off); bzero(va, size); } vm_offset_t moea_quick_enter_page(vm_page_t m) { return (VM_PAGE_TO_PHYS(m)); } void moea_quick_remove_page(vm_offset_t addr) { } boolean_t moea_page_is_mapped(vm_page_t m) { return (!LIST_EMPTY(&(m)->md.mdpg_pvoh)); } +bool +moea_ps_enabled(pmap_t pmap __unused) +{ + return (false); +} + /* * Map the given physical page at the specified virtual address in the * target pmap with the protection requested. If specified the page * will be wired down. */ int moea_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind) { int error; for (;;) { rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); error = moea_enter_locked(pmap, va, m, prot, flags, psind); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); if (error != ENOMEM) return (KERN_SUCCESS); if ((flags & PMAP_ENTER_NOSLEEP) != 0) return (KERN_RESOURCE_SHORTAGE); VM_OBJECT_ASSERT_UNLOCKED(m->object); vm_wait(NULL); } } /* * Map the given physical page at the specified virtual address in the * target pmap with the protection requested. If specified the page * will be wired down. * * The global pvh and pmap must be locked. */ static int moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind __unused) { struct pvo_head *pvo_head; uma_zone_t zone; u_int pte_lo, pvo_flags; int error; if (pmap_bootstrapped) rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if ((m->oflags & VPO_UNMANAGED) == 0) { if ((flags & PMAP_ENTER_QUICK_LOCKED) == 0) VM_PAGE_OBJECT_BUSY_ASSERT(m); else VM_OBJECT_ASSERT_LOCKED(m->object); } if ((m->oflags & VPO_UNMANAGED) != 0 || !moea_initialized) { pvo_head = &moea_pvo_kunmanaged; zone = moea_upvo_zone; pvo_flags = 0; } else { pvo_head = vm_page_to_pvoh(m); zone = moea_mpvo_zone; pvo_flags = PVO_MANAGED; } pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m)); if (prot & VM_PROT_WRITE) { pte_lo |= PTE_BW; if (pmap_bootstrapped && (m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_set(m, PGA_WRITEABLE); } else pte_lo |= PTE_BR; if ((flags & PMAP_ENTER_WIRED) != 0) pvo_flags |= PVO_WIRED; error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m), pte_lo, pvo_flags); /* * Flush the real page from the instruction cache. This has be done * for all user mappings to prevent information leakage via the * instruction cache. moea_pvo_enter() returns ENOENT for the first * mapping for a page. */ if (pmap != kernel_pmap && error == ENOENT && (pte_lo & (PTE_I | PTE_G)) == 0) moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE); return (error); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ void moea_enter_object(pmap_t pm, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { vm_page_t m; vm_pindex_t diff, psize; VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); m = m_start; rw_wlock(&pvh_global_lock); PMAP_LOCK(pm); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { moea_enter_locked(pm, start + ptoa(diff), m, prot & (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_QUICK_LOCKED, 0); m = TAILQ_NEXT(m, listq); } rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pm); } void moea_enter_quick(pmap_t pm, vm_offset_t va, vm_page_t m, vm_prot_t prot) { rw_wlock(&pvh_global_lock); PMAP_LOCK(pm); moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_QUICK_LOCKED, 0); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pm); } vm_paddr_t moea_extract(pmap_t pm, vm_offset_t va) { struct pvo_entry *pvo; vm_paddr_t pa; PMAP_LOCK(pm); pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL); if (pvo == NULL) pa = 0; else pa = PVO_PADDR(pvo) | (va & ADDR_POFF); PMAP_UNLOCK(pm); return (pa); } /* * Atomically extract and hold the physical page with the given * pmap and virtual address pair if that mapping permits the given * protection. */ vm_page_t moea_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) { struct pvo_entry *pvo; vm_page_t m; m = NULL; PMAP_LOCK(pmap); pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL); if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) && ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW || (prot & VM_PROT_WRITE) == 0)) { m = PHYS_TO_VM_PAGE(PVO_PADDR(pvo)); if (!vm_page_wire_mapped(m)) m = NULL; } PMAP_UNLOCK(pmap); return (m); } void moea_init() { moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); moea_initialized = TRUE; } boolean_t moea_is_referenced(vm_page_t m) { boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("moea_is_referenced: page %p is not managed", m)); rw_wlock(&pvh_global_lock); rv = moea_query_bit(m, PTE_REF); rw_wunlock(&pvh_global_lock); return (rv); } boolean_t moea_is_modified(vm_page_t m) { boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("moea_is_modified: page %p is not managed", m)); /* * If the page is not busied then this check is racy. */ if (!pmap_page_is_write_mapped(m)) return (FALSE); rw_wlock(&pvh_global_lock); rv = moea_query_bit(m, PTE_CHG); rw_wunlock(&pvh_global_lock); return (rv); } boolean_t moea_is_prefaultable(pmap_t pmap, vm_offset_t va) { struct pvo_entry *pvo; boolean_t rv; PMAP_LOCK(pmap); pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL); rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0; PMAP_UNLOCK(pmap); return (rv); } void moea_clear_modify(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("moea_clear_modify: page %p is not managed", m)); vm_page_assert_busied(m); if (!pmap_page_is_write_mapped(m)) return; rw_wlock(&pvh_global_lock); moea_clear_bit(m, PTE_CHG); rw_wunlock(&pvh_global_lock); } /* * Clear the write and modified bits in each of the given page's mappings. */ void moea_remove_write(vm_page_t m) { struct pvo_entry *pvo; struct pte *pt; pmap_t pmap; u_int lo; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("moea_remove_write: page %p is not managed", m)); vm_page_assert_busied(m); if (!pmap_page_is_write_mapped(m)) return; rw_wlock(&pvh_global_lock); lo = moea_attr_fetch(m); powerpc_sync(); LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) { pmap = pvo->pvo_pmap; PMAP_LOCK(pmap); if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) { pt = moea_pvo_to_pte(pvo, -1); pvo->pvo_pte.pte.pte_lo &= ~PTE_PP; pvo->pvo_pte.pte.pte_lo |= PTE_BR; if (pt != NULL) { moea_pte_synch(pt, &pvo->pvo_pte.pte); lo |= pvo->pvo_pte.pte.pte_lo; pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG; moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr); mtx_unlock(&moea_table_mutex); } } PMAP_UNLOCK(pmap); } if ((lo & PTE_CHG) != 0) { moea_attr_clear(m, PTE_CHG); vm_page_dirty(m); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(&pvh_global_lock); } /* * moea_ts_referenced: * * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. * * XXX: The exact number of bits to check and clear is a matter that * should be tested and standardized at some point in the future for * optimal aging of shared pages. */ int moea_ts_referenced(vm_page_t m) { int count; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("moea_ts_referenced: page %p is not managed", m)); rw_wlock(&pvh_global_lock); count = moea_clear_bit(m, PTE_REF); rw_wunlock(&pvh_global_lock); return (count); } /* * Modify the WIMG settings of all mappings for a page. */ void moea_page_set_memattr(vm_page_t m, vm_memattr_t ma) { struct pvo_entry *pvo; struct pvo_head *pvo_head; struct pte *pt; pmap_t pmap; u_int lo; if ((m->oflags & VPO_UNMANAGED) != 0) { m->md.mdpg_cache_attrs = ma; return; } rw_wlock(&pvh_global_lock); pvo_head = vm_page_to_pvoh(m); lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma); LIST_FOREACH(pvo, pvo_head, pvo_vlink) { pmap = pvo->pvo_pmap; PMAP_LOCK(pmap); pt = moea_pvo_to_pte(pvo, -1); pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG; pvo->pvo_pte.pte.pte_lo |= lo; if (pt != NULL) { moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr); if (pvo->pvo_pmap == kernel_pmap) isync(); } mtx_unlock(&moea_table_mutex); PMAP_UNLOCK(pmap); } m->md.mdpg_cache_attrs = ma; rw_wunlock(&pvh_global_lock); } /* * Map a wired page into kernel virtual address space. */ void moea_kenter(vm_offset_t va, vm_paddr_t pa) { moea_kenter_attr(va, pa, VM_MEMATTR_DEFAULT); } void moea_kenter_attr(vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma) { u_int pte_lo; int error; #if 0 if (va < VM_MIN_KERNEL_ADDRESS) panic("moea_kenter: attempt to enter non-kernel address %#x", va); #endif pte_lo = moea_calc_wimg(pa, ma); PMAP_LOCK(kernel_pmap); error = moea_pvo_enter(kernel_pmap, moea_upvo_zone, &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED); if (error != 0 && error != ENOENT) panic("moea_kenter: failed to enter va %#x pa %#x: %d", va, pa, error); PMAP_UNLOCK(kernel_pmap); } /* * Extract the physical page address associated with the given kernel virtual * address. */ vm_paddr_t moea_kextract(vm_offset_t va) { struct pvo_entry *pvo; vm_paddr_t pa; /* * Allow direct mappings on 32-bit OEA */ if (va < VM_MIN_KERNEL_ADDRESS) { return (va); } PMAP_LOCK(kernel_pmap); pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL); KASSERT(pvo != NULL, ("moea_kextract: no addr found")); pa = PVO_PADDR(pvo) | (va & ADDR_POFF); PMAP_UNLOCK(kernel_pmap); return (pa); } /* * Remove a wired page from kernel virtual address space. */ void moea_kremove(vm_offset_t va) { moea_remove(kernel_pmap, va, va + PAGE_SIZE); } /* * Provide a kernel pointer corresponding to a given userland pointer. * The returned pointer is valid until the next time this function is * called in this thread. This is used internally in copyin/copyout. */ int moea_map_user_ptr(pmap_t pm, volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen) { size_t l; register_t vsid; *kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK); l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr); if (l > ulen) l = ulen; if (klen) *klen = l; else if (l != ulen) return (EFAULT); vsid = va_to_vsid(pm, (vm_offset_t)uaddr); /* Mark segment no-execute */ vsid |= SR_N; /* If we have already set this VSID, we can just return */ if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == vsid) return (0); __asm __volatile("isync"); curthread->td_pcb->pcb_cpu.aim.usr_segm = (uintptr_t)uaddr >> ADDR_SR_SHFT; curthread->td_pcb->pcb_cpu.aim.usr_vsid = vsid; __asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(vsid)); return (0); } /* * Figure out where a given kernel pointer (usually in a fault) points * to from the VM's perspective, potentially remapping into userland's * address space. */ static int moea_decode_kernel_ptr(vm_offset_t addr, int *is_user, vm_offset_t *decoded_addr) { vm_offset_t user_sr; if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) { user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm; addr &= ADDR_PIDX | ADDR_POFF; addr |= user_sr << ADDR_SR_SHFT; *decoded_addr = addr; *is_user = 1; } else { *decoded_addr = addr; *is_user = 0; } return (0); } /* * Map a range of physical addresses into kernel virtual address space. * * The value passed in *virt is a suggested virtual address for the mapping. * Architectures which can support a direct-mapped physical to virtual region * can return the appropriate address within that region, leaving '*virt' * unchanged. We cannot and therefore do not; *virt is updated with the * first usable address after the mapped region. */ vm_offset_t moea_map(vm_offset_t *virt, vm_paddr_t pa_start, vm_paddr_t pa_end, int prot) { vm_offset_t sva, va; sva = *virt; va = sva; for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE) moea_kenter(va, pa_start); *virt = va; return (sva); } /* * Returns true if the pmap's pv is one of the first * 16 pvs linked to from this page. This count may * be changed upwards or downwards in the future; it * is only necessary that true be returned for a small * subset of pmaps for proper page aging. */ boolean_t moea_page_exists_quick(pmap_t pmap, vm_page_t m) { int loops; struct pvo_entry *pvo; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("moea_page_exists_quick: page %p is not managed", m)); loops = 0; rv = FALSE; rw_wlock(&pvh_global_lock); LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) { if (pvo->pvo_pmap == pmap) { rv = TRUE; break; } if (++loops >= 16) break; } rw_wunlock(&pvh_global_lock); return (rv); } void moea_page_init(vm_page_t m) { m->md.mdpg_attrs = 0; m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT; LIST_INIT(&m->md.mdpg_pvoh); } /* * Return the number of managed mappings to the given physical page * that are wired. */ int moea_page_wired_mappings(vm_page_t m) { struct pvo_entry *pvo; int count; count = 0; if ((m->oflags & VPO_UNMANAGED) != 0) return (count); rw_wlock(&pvh_global_lock); LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) if ((pvo->pvo_vaddr & PVO_WIRED) != 0) count++; rw_wunlock(&pvh_global_lock); return (count); } static u_int moea_vsidcontext; int moea_pinit(pmap_t pmap) { int i, mask; u_int entropy; RB_INIT(&pmap->pmap_pvo); entropy = 0; __asm __volatile("mftb %0" : "=r"(entropy)); if ((pmap->pmap_phys = (pmap_t)moea_kextract((vm_offset_t)pmap)) == NULL) { pmap->pmap_phys = pmap; } mtx_lock(&moea_vsid_mutex); /* * Allocate some segment registers for this pmap. */ for (i = 0; i < NPMAPS; i += VSID_NBPW) { u_int hash, n; /* * Create a new value by mutiplying by a prime and adding in * entropy from the timebase register. This is to make the * VSID more random so that the PT hash function collides * less often. (Note that the prime casues gcc to do shifts * instead of a multiply.) */ moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy; hash = moea_vsidcontext & (NPMAPS - 1); if (hash == 0) /* 0 is special, avoid it */ continue; n = hash >> 5; mask = 1 << (hash & (VSID_NBPW - 1)); hash = (moea_vsidcontext & 0xfffff); if (moea_vsid_bitmap[n] & mask) { /* collision? */ /* anything free in this bucket? */ if (moea_vsid_bitmap[n] == 0xffffffff) { entropy = (moea_vsidcontext >> 20); continue; } i = ffs(~moea_vsid_bitmap[n]) - 1; mask = 1 << i; hash &= rounddown2(0xfffff, VSID_NBPW); hash |= i; } KASSERT(!(moea_vsid_bitmap[n] & mask), ("Allocating in-use VSID group %#x\n", hash)); moea_vsid_bitmap[n] |= mask; for (i = 0; i < 16; i++) pmap->pm_sr[i] = VSID_MAKE(i, hash); mtx_unlock(&moea_vsid_mutex); return (1); } mtx_unlock(&moea_vsid_mutex); panic("moea_pinit: out of segments"); } /* * Initialize the pmap associated with process 0. */ void moea_pinit0(pmap_t pm) { PMAP_LOCK_INIT(pm); moea_pinit(pm); bzero(&pm->pm_stats, sizeof(pm->pm_stats)); } /* * Set the physical protection on the specified range of this map as requested. */ void moea_protect(pmap_t pm, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { struct pvo_entry *pvo, *tpvo, key; struct pte *pt; KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap, ("moea_protect: non current pmap")); if ((prot & VM_PROT_READ) == VM_PROT_NONE) { moea_remove(pm, sva, eva); return; } rw_wlock(&pvh_global_lock); PMAP_LOCK(pm); key.pvo_vaddr = sva; for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key); pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) { tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo); /* * Grab the PTE pointer before we diddle with the cached PTE * copy. */ pt = moea_pvo_to_pte(pvo, -1); /* * Change the protection of the page. */ pvo->pvo_pte.pte.pte_lo &= ~PTE_PP; pvo->pvo_pte.pte.pte_lo |= PTE_BR; /* * If the PVO is in the page table, update that pte as well. */ if (pt != NULL) { moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr); mtx_unlock(&moea_table_mutex); } } rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pm); } /* * Map a list of wired pages into kernel virtual address space. This is * intended for temporary mappings which do not need page modification or * references recorded. Existing mappings in the region are overwritten. */ void moea_qenter(vm_offset_t sva, vm_page_t *m, int count) { vm_offset_t va; va = sva; while (count-- > 0) { moea_kenter(va, VM_PAGE_TO_PHYS(*m)); va += PAGE_SIZE; m++; } } /* * Remove page mappings from kernel virtual address space. Intended for * temporary mappings entered by moea_qenter. */ void moea_qremove(vm_offset_t sva, int count) { vm_offset_t va; va = sva; while (count-- > 0) { moea_kremove(va); va += PAGE_SIZE; } } void moea_release(pmap_t pmap) { int idx, mask; /* * Free segment register's VSID */ if (pmap->pm_sr[0] == 0) panic("moea_release"); mtx_lock(&moea_vsid_mutex); idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1); mask = 1 << (idx % VSID_NBPW); idx /= VSID_NBPW; moea_vsid_bitmap[idx] &= ~mask; mtx_unlock(&moea_vsid_mutex); } /* * Remove the given range of addresses from the specified map. */ void moea_remove(pmap_t pm, vm_offset_t sva, vm_offset_t eva) { struct pvo_entry *pvo, *tpvo, key; rw_wlock(&pvh_global_lock); PMAP_LOCK(pm); key.pvo_vaddr = sva; for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key); pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) { tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo); moea_pvo_remove(pvo, -1); } PMAP_UNLOCK(pm); rw_wunlock(&pvh_global_lock); } /* * Remove physical page from all pmaps in which it resides. moea_pvo_remove() * will reflect changes in pte's back to the vm_page. */ void moea_remove_all(vm_page_t m) { struct pvo_head *pvo_head; struct pvo_entry *pvo, *next_pvo; pmap_t pmap; rw_wlock(&pvh_global_lock); pvo_head = vm_page_to_pvoh(m); for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) { next_pvo = LIST_NEXT(pvo, pvo_vlink); pmap = pvo->pvo_pmap; PMAP_LOCK(pmap); moea_pvo_remove(pvo, -1); PMAP_UNLOCK(pmap); } if ((m->a.flags & PGA_WRITEABLE) && moea_query_bit(m, PTE_CHG)) { moea_attr_clear(m, PTE_CHG); vm_page_dirty(m); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(&pvh_global_lock); } static int moea_mincore(pmap_t pm, vm_offset_t va, vm_paddr_t *pap) { struct pvo_entry *pvo; vm_paddr_t pa; vm_page_t m; int val; bool managed; PMAP_LOCK(pm); pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL); if (pvo != NULL) { pa = PVO_PADDR(pvo); m = PHYS_TO_VM_PAGE(pa); managed = (pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED; val = MINCORE_INCORE; } else { PMAP_UNLOCK(pm); return (0); } PMAP_UNLOCK(pm); if (m == NULL) return (0); if (managed) { if (moea_is_modified(m)) val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER; if (moea_is_referenced(m)) val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER; } if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) != (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) && managed) { *pap = pa; } return (val); } /* * Allocate a physical page of memory directly from the phys_avail map. * Can only be called from moea_bootstrap before avail start and end are * calculated. */ static vm_offset_t moea_bootstrap_alloc(vm_size_t size, u_int align) { vm_offset_t s, e; int i, j; size = round_page(size); for (i = 0; phys_avail[i + 1] != 0; i += 2) { if (align != 0) s = roundup2(phys_avail[i], align); else s = phys_avail[i]; e = s + size; if (s < phys_avail[i] || e > phys_avail[i + 1]) continue; if (s == phys_avail[i]) { phys_avail[i] += size; } else if (e == phys_avail[i + 1]) { phys_avail[i + 1] -= size; } else { for (j = phys_avail_count * 2; j > i; j -= 2) { phys_avail[j] = phys_avail[j - 2]; phys_avail[j + 1] = phys_avail[j - 1]; } phys_avail[i + 3] = phys_avail[i + 1]; phys_avail[i + 1] = s; phys_avail[i + 2] = e; phys_avail_count++; } return (s); } panic("moea_bootstrap_alloc: could not allocate memory"); } static void moea_syncicache(vm_paddr_t pa, vm_size_t len) { __syncicache((void *)pa, len); } static int moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head, vm_offset_t va, vm_paddr_t pa, u_int pte_lo, int flags) { struct pvo_entry *pvo; u_int sr; int first; u_int ptegidx; int i; int bootstrap; moea_pvo_enter_calls++; first = 0; bootstrap = 0; /* * Compute the PTE Group index. */ va &= ~ADDR_POFF; sr = va_to_sr(pm->pm_sr, va); ptegidx = va_to_pteg(sr, va); /* * Remove any existing mapping for this page. Reuse the pvo entry if * there is a mapping. */ mtx_lock(&moea_table_mutex); LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) { if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) { if (PVO_PADDR(pvo) == pa && (pvo->pvo_pte.pte.pte_lo & PTE_PP) == (pte_lo & PTE_PP)) { /* * The PTE is not changing. Instead, this may * be a request to change the mapping's wired * attribute. */ mtx_unlock(&moea_table_mutex); if ((flags & PVO_WIRED) != 0 && (pvo->pvo_vaddr & PVO_WIRED) == 0) { pvo->pvo_vaddr |= PVO_WIRED; pm->pm_stats.wired_count++; } else if ((flags & PVO_WIRED) == 0 && (pvo->pvo_vaddr & PVO_WIRED) != 0) { pvo->pvo_vaddr &= ~PVO_WIRED; pm->pm_stats.wired_count--; } return (0); } moea_pvo_remove(pvo, -1); break; } } /* * If we aren't overwriting a mapping, try to allocate. */ if (moea_initialized) { pvo = uma_zalloc(zone, M_NOWAIT); } else { if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) { panic("moea_enter: bpvo pool exhausted, %d, %d, %d", moea_bpvo_pool_index, BPVO_POOL_SIZE, BPVO_POOL_SIZE * sizeof(struct pvo_entry)); } pvo = &moea_bpvo_pool[moea_bpvo_pool_index]; moea_bpvo_pool_index++; bootstrap = 1; } if (pvo == NULL) { mtx_unlock(&moea_table_mutex); return (ENOMEM); } moea_pvo_entries++; pvo->pvo_vaddr = va; pvo->pvo_pmap = pm; LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink); pvo->pvo_vaddr &= ~ADDR_POFF; if (flags & PVO_WIRED) pvo->pvo_vaddr |= PVO_WIRED; if (pvo_head != &moea_pvo_kunmanaged) pvo->pvo_vaddr |= PVO_MANAGED; if (bootstrap) pvo->pvo_vaddr |= PVO_BOOTSTRAP; moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo); /* * Add to pmap list */ RB_INSERT(pvo_tree, &pm->pmap_pvo, pvo); /* * Remember if the list was empty and therefore will be the first * item. */ if (LIST_FIRST(pvo_head) == NULL) first = 1; LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink); if (pvo->pvo_vaddr & PVO_WIRED) pm->pm_stats.wired_count++; pm->pm_stats.resident_count++; i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte); KASSERT(i < 8, ("Invalid PTE index")); if (i >= 0) { PVO_PTEGIDX_SET(pvo, i); } else { panic("moea_pvo_enter: overflow"); moea_pte_overflow++; } mtx_unlock(&moea_table_mutex); return (first ? ENOENT : 0); } static void moea_pvo_remove(struct pvo_entry *pvo, int pteidx) { struct pte *pt; /* * If there is an active pte entry, we need to deactivate it (and * save the ref & cfg bits). */ pt = moea_pvo_to_pte(pvo, pteidx); if (pt != NULL) { moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr); mtx_unlock(&moea_table_mutex); PVO_PTEGIDX_CLR(pvo); } else { moea_pte_overflow--; } /* * Update our statistics. */ pvo->pvo_pmap->pm_stats.resident_count--; if (pvo->pvo_vaddr & PVO_WIRED) pvo->pvo_pmap->pm_stats.wired_count--; /* * Remove this PVO from the PV and pmap lists. */ LIST_REMOVE(pvo, pvo_vlink); RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo); /* * Save the REF/CHG bits into their cache if the page is managed. * Clear PGA_WRITEABLE if all mappings of the page have been removed. */ if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED) { struct vm_page *pg; pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo)); if (pg != NULL) { moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo & (PTE_REF | PTE_CHG)); if (LIST_EMPTY(&pg->md.mdpg_pvoh)) vm_page_aflag_clear(pg, PGA_WRITEABLE); } } /* * Remove this from the overflow list and return it to the pool * if we aren't going to reuse it. */ LIST_REMOVE(pvo, pvo_olink); if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP)) uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone : moea_upvo_zone, pvo); moea_pvo_entries--; moea_pvo_remove_calls++; } static __inline int moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx) { int pteidx; /* * We can find the actual pte entry without searching by grabbing * the PTEG index from 3 unused bits in pte_lo[11:9] and by * noticing the HID bit. */ pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo); if (pvo->pvo_pte.pte.pte_hi & PTE_HID) pteidx ^= moea_pteg_mask * 8; return (pteidx); } static struct pvo_entry * moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p) { struct pvo_entry *pvo; int ptegidx; u_int sr; va &= ~ADDR_POFF; sr = va_to_sr(pm->pm_sr, va); ptegidx = va_to_pteg(sr, va); mtx_lock(&moea_table_mutex); LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) { if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) { if (pteidx_p) *pteidx_p = moea_pvo_pte_index(pvo, ptegidx); break; } } mtx_unlock(&moea_table_mutex); return (pvo); } static struct pte * moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx) { struct pte *pt; /* * If we haven't been supplied the ptegidx, calculate it. */ if (pteidx == -1) { int ptegidx; u_int sr; sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr); ptegidx = va_to_pteg(sr, pvo->pvo_vaddr); pteidx = moea_pvo_pte_index(pvo, ptegidx); } pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7]; mtx_lock(&moea_table_mutex); if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) { panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no " "valid pte index", pvo); } if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) { panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo " "pvo but no valid pte", pvo); } if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) { if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) { panic("moea_pvo_to_pte: pvo %p has valid pte in " "moea_pteg_table %p but invalid in pvo", pvo, pt); } if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF)) != 0) { panic("moea_pvo_to_pte: pvo %p pte does not match " "pte %p in moea_pteg_table", pvo, pt); } mtx_assert(&moea_table_mutex, MA_OWNED); return (pt); } if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) { panic("moea_pvo_to_pte: pvo %p has invalid pte %p in " "moea_pteg_table but valid in pvo: %8x, %8x", pvo, pt, pvo->pvo_pte.pte.pte_hi, pt->pte_hi); } mtx_unlock(&moea_table_mutex); return (NULL); } /* * XXX: THIS STUFF SHOULD BE IN pte.c? */ int moea_pte_spill(vm_offset_t addr) { struct pvo_entry *source_pvo, *victim_pvo; struct pvo_entry *pvo; int ptegidx, i, j; u_int sr; struct pteg *pteg; struct pte *pt; moea_pte_spills++; sr = mfsrin(addr); ptegidx = va_to_pteg(sr, addr); /* * Have to substitute some entry. Use the primary hash for this. * Use low bits of timebase as random generator. */ pteg = &moea_pteg_table[ptegidx]; mtx_lock(&moea_table_mutex); __asm __volatile("mftb %0" : "=r"(i)); i &= 7; pt = &pteg->pt[i]; source_pvo = NULL; victim_pvo = NULL; LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) { /* * We need to find a pvo entry for this address. */ if (source_pvo == NULL && moea_pte_match(&pvo->pvo_pte.pte, sr, addr, pvo->pvo_pte.pte.pte_hi & PTE_HID)) { /* * Now found an entry to be spilled into the pteg. * The PTE is now valid, so we know it's active. */ j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte); if (j >= 0) { PVO_PTEGIDX_SET(pvo, j); moea_pte_overflow--; mtx_unlock(&moea_table_mutex); return (1); } source_pvo = pvo; if (victim_pvo != NULL) break; } /* * We also need the pvo entry of the victim we are replacing * so save the R & C bits of the PTE. */ if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL && moea_pte_compare(pt, &pvo->pvo_pte.pte)) { victim_pvo = pvo; if (source_pvo != NULL) break; } } if (source_pvo == NULL) { mtx_unlock(&moea_table_mutex); return (0); } if (victim_pvo == NULL) { if ((pt->pte_hi & PTE_HID) == 0) panic("moea_pte_spill: victim p-pte (%p) has no pvo" "entry", pt); /* * If this is a secondary PTE, we need to search it's primary * pvo bucket for the matching PVO. */ LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask], pvo_olink) { /* * We also need the pvo entry of the victim we are * replacing so save the R & C bits of the PTE. */ if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) { victim_pvo = pvo; break; } } if (victim_pvo == NULL) panic("moea_pte_spill: victim s-pte (%p) has no pvo" "entry", pt); } /* * We are invalidating the TLB entry for the EA we are replacing even * though it's valid. If we don't, we lose any ref/chg bit changes * contained in the TLB entry. */ source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID; moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr); moea_pte_set(pt, &source_pvo->pvo_pte.pte); PVO_PTEGIDX_CLR(victim_pvo); PVO_PTEGIDX_SET(source_pvo, i); moea_pte_replacements++; mtx_unlock(&moea_table_mutex); return (1); } static __inline struct pvo_entry * moea_pte_spillable_ident(u_int ptegidx) { struct pte *pt; struct pvo_entry *pvo_walk, *pvo = NULL; LIST_FOREACH(pvo_walk, &moea_pvo_table[ptegidx], pvo_olink) { if (pvo_walk->pvo_vaddr & PVO_WIRED) continue; if (!(pvo_walk->pvo_pte.pte.pte_hi & PTE_VALID)) continue; pt = moea_pvo_to_pte(pvo_walk, -1); if (pt == NULL) continue; pvo = pvo_walk; mtx_unlock(&moea_table_mutex); if (!(pt->pte_lo & PTE_REF)) return (pvo_walk); } return (pvo); } static int moea_pte_insert(u_int ptegidx, struct pte *pvo_pt) { struct pte *pt; struct pvo_entry *victim_pvo; int i; int victim_idx; u_int pteg_bkpidx = ptegidx; mtx_assert(&moea_table_mutex, MA_OWNED); /* * First try primary hash. */ for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) { if ((pt->pte_hi & PTE_VALID) == 0) { pvo_pt->pte_hi &= ~PTE_HID; moea_pte_set(pt, pvo_pt); return (i); } } /* * Now try secondary hash. */ ptegidx ^= moea_pteg_mask; for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) { if ((pt->pte_hi & PTE_VALID) == 0) { pvo_pt->pte_hi |= PTE_HID; moea_pte_set(pt, pvo_pt); return (i); } } /* Try again, but this time try to force a PTE out. */ ptegidx = pteg_bkpidx; victim_pvo = moea_pte_spillable_ident(ptegidx); if (victim_pvo == NULL) { ptegidx ^= moea_pteg_mask; victim_pvo = moea_pte_spillable_ident(ptegidx); } if (victim_pvo == NULL) { panic("moea_pte_insert: overflow"); return (-1); } victim_idx = moea_pvo_pte_index(victim_pvo, ptegidx); if (pteg_bkpidx == ptegidx) pvo_pt->pte_hi &= ~PTE_HID; else pvo_pt->pte_hi |= PTE_HID; /* * Synchronize the sacrifice PTE with its PVO, then mark both * invalid. The PVO will be reused when/if the VM system comes * here after a fault. */ pt = &moea_pteg_table[victim_idx >> 3].pt[victim_idx & 7]; if (pt->pte_hi != victim_pvo->pvo_pte.pte.pte_hi) panic("Victim PVO doesn't match PTE! PVO: %8x, PTE: %8x", victim_pvo->pvo_pte.pte.pte_hi, pt->pte_hi); /* * Set the new PTE. */ moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr); PVO_PTEGIDX_CLR(victim_pvo); moea_pte_overflow++; moea_pte_set(pt, pvo_pt); return (victim_idx & 7); } static boolean_t moea_query_bit(vm_page_t m, int ptebit) { struct pvo_entry *pvo; struct pte *pt; rw_assert(&pvh_global_lock, RA_WLOCKED); if (moea_attr_fetch(m) & ptebit) return (TRUE); LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) { /* * See if we saved the bit off. If so, cache it and return * success. */ if (pvo->pvo_pte.pte.pte_lo & ptebit) { moea_attr_save(m, ptebit); return (TRUE); } } /* * No luck, now go through the hard part of looking at the PTEs * themselves. Sync so that any pending REF/CHG bits are flushed to * the PTEs. */ powerpc_sync(); LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) { /* * See if this pvo has a valid PTE. if so, fetch the * REF/CHG bits from the valid PTE. If the appropriate * ptebit is set, cache it and return success. */ pt = moea_pvo_to_pte(pvo, -1); if (pt != NULL) { moea_pte_synch(pt, &pvo->pvo_pte.pte); mtx_unlock(&moea_table_mutex); if (pvo->pvo_pte.pte.pte_lo & ptebit) { moea_attr_save(m, ptebit); return (TRUE); } } } return (FALSE); } static u_int moea_clear_bit(vm_page_t m, int ptebit) { u_int count; struct pvo_entry *pvo; struct pte *pt; rw_assert(&pvh_global_lock, RA_WLOCKED); /* * Clear the cached value. */ moea_attr_clear(m, ptebit); /* * Sync so that any pending REF/CHG bits are flushed to the PTEs (so * we can reset the right ones). note that since the pvo entries and * list heads are accessed via BAT0 and are never placed in the page * table, we don't have to worry about further accesses setting the * REF/CHG bits. */ powerpc_sync(); /* * For each pvo entry, clear the pvo's ptebit. If this pvo has a * valid pte clear the ptebit from the valid pte. */ count = 0; LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) { pt = moea_pvo_to_pte(pvo, -1); if (pt != NULL) { moea_pte_synch(pt, &pvo->pvo_pte.pte); if (pvo->pvo_pte.pte.pte_lo & ptebit) { count++; moea_pte_clear(pt, PVO_VADDR(pvo), ptebit); } mtx_unlock(&moea_table_mutex); } pvo->pvo_pte.pte.pte_lo &= ~ptebit; } return (count); } /* * Return true if the physical range is encompassed by the battable[idx] */ static int moea_bat_mapped(int idx, vm_paddr_t pa, vm_size_t size) { u_int prot; u_int32_t start; u_int32_t end; u_int32_t bat_ble; /* * Return immediately if not a valid mapping */ if (!(battable[idx].batu & BAT_Vs)) return (EINVAL); /* * The BAT entry must be cache-inhibited, guarded, and r/w * so it can function as an i/o page */ prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW); if (prot != (BAT_I|BAT_G|BAT_PP_RW)) return (EPERM); /* * The address should be within the BAT range. Assume that the * start address in the BAT has the correct alignment (thus * not requiring masking) */ start = battable[idx].batl & BAT_PBS; bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03; end = start | (bat_ble << 15) | 0x7fff; if ((pa < start) || ((pa + size) > end)) return (ERANGE); return (0); } boolean_t moea_dev_direct_mapped(vm_paddr_t pa, vm_size_t size) { int i; /* * This currently does not work for entries that * overlap 256M BAT segments. */ for(i = 0; i < 16; i++) if (moea_bat_mapped(i, pa, size) == 0) return (0); return (EFAULT); } /* * Map a set of physical memory pages into the kernel virtual * address space. Return a pointer to where it is mapped. This * routine is intended to be used for mapping device memory, * NOT real memory. */ void * moea_mapdev(vm_paddr_t pa, vm_size_t size) { return (moea_mapdev_attr(pa, size, VM_MEMATTR_DEFAULT)); } void * moea_mapdev_attr(vm_paddr_t pa, vm_size_t size, vm_memattr_t ma) { vm_offset_t va, tmpva, ppa, offset; int i; ppa = trunc_page(pa); offset = pa & PAGE_MASK; size = roundup(offset + size, PAGE_SIZE); /* * If the physical address lies within a valid BAT table entry, * return the 1:1 mapping. This currently doesn't work * for regions that overlap 256M BAT segments. */ for (i = 0; i < 16; i++) { if (moea_bat_mapped(i, pa, size) == 0) return ((void *) pa); } va = kva_alloc(size); if (!va) panic("moea_mapdev: Couldn't alloc kernel virtual memory"); for (tmpva = va; size > 0;) { moea_kenter_attr(tmpva, ppa, ma); tlbie(tmpva); size -= PAGE_SIZE; tmpva += PAGE_SIZE; ppa += PAGE_SIZE; } return ((void *)(va + offset)); } void moea_unmapdev(vm_offset_t va, vm_size_t size) { vm_offset_t base, offset; /* * If this is outside kernel virtual space, then it's a * battable entry and doesn't require unmapping */ if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) { base = trunc_page(va); offset = va & PAGE_MASK; size = roundup(offset + size, PAGE_SIZE); moea_qremove(base, atop(size)); kva_free(base, size); } } static void moea_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz) { struct pvo_entry *pvo; vm_offset_t lim; vm_paddr_t pa; vm_size_t len; PMAP_LOCK(pm); while (sz > 0) { lim = round_page(va + 1); len = MIN(lim - va, sz); pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL); if (pvo != NULL) { pa = PVO_PADDR(pvo) | (va & ADDR_POFF); moea_syncicache(pa, len); } va += len; sz -= len; } PMAP_UNLOCK(pm); } void moea_dumpsys_map(vm_paddr_t pa, size_t sz, void **va) { *va = (void *)pa; } extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1]; void moea_scan_init() { struct pvo_entry *pvo; vm_offset_t va; int i; if (!do_minidump) { /* Initialize phys. segments for dumpsys(). */ memset(&dump_map, 0, sizeof(dump_map)); mem_regions(&pregions, &pregions_sz, ®ions, ®ions_sz); for (i = 0; i < pregions_sz; i++) { dump_map[i].pa_start = pregions[i].mr_start; dump_map[i].pa_size = pregions[i].mr_size; } return; } /* Virtual segments for minidumps: */ memset(&dump_map, 0, sizeof(dump_map)); /* 1st: kernel .data and .bss. */ dump_map[0].pa_start = trunc_page((uintptr_t)_etext); dump_map[0].pa_size = round_page((uintptr_t)_end) - dump_map[0].pa_start; /* 2nd: msgbuf and tables (see pmap_bootstrap()). */ dump_map[1].pa_start = (vm_paddr_t)msgbufp->msg_ptr; dump_map[1].pa_size = round_page(msgbufp->msg_size); /* 3rd: kernel VM. */ va = dump_map[1].pa_start + dump_map[1].pa_size; /* Find start of next chunk (from va). */ while (va < virtual_end) { /* Don't dump the buffer cache. */ if (va >= kmi.buffer_sva && va < kmi.buffer_eva) { va = kmi.buffer_eva; continue; } pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL); if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID)) break; va += PAGE_SIZE; } if (va < virtual_end) { dump_map[2].pa_start = va; va += PAGE_SIZE; /* Find last page in chunk. */ while (va < virtual_end) { /* Don't run into the buffer cache. */ if (va == kmi.buffer_sva) break; pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL); if (pvo == NULL || !(pvo->pvo_pte.pte.pte_hi & PTE_VALID)) break; va += PAGE_SIZE; } dump_map[2].pa_size = va - dump_map[2].pa_start; } } diff --git a/sys/powerpc/booke/pmap.c b/sys/powerpc/booke/pmap.c index 6bc96b222db8..6c48584096c5 100644 --- a/sys/powerpc/booke/pmap.c +++ b/sys/powerpc/booke/pmap.c @@ -1,3108 +1,3116 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (C) 2007-2009 Semihalf, Rafal Jaworowski * Copyright (C) 2006 Semihalf, Marian Balakowicz * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN * NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * Some hw specific parts of this pmap were derived or influenced * by NetBSD's ibm4xx pmap module. More generic code is shared with * a few other pmap modules from the FreeBSD tree. */ /* * VM layout notes: * * Kernel and user threads run within one common virtual address space * defined by AS=0. * * 32-bit pmap: * Virtual address space layout: * ----------------------------- * 0x0000_0000 - 0x7fff_ffff : user process * 0x8000_0000 - 0xbfff_ffff : pmap_mapdev()-ed area (PCI/PCIE etc.) * 0xc000_0000 - 0xc0ff_ffff : kernel reserved * 0xc000_0000 - data_end : kernel code+data, env, metadata etc. * 0xc100_0000 - 0xffff_ffff : KVA * 0xc100_0000 - 0xc100_3fff : reserved for page zero/copy * 0xc100_4000 - 0xc200_3fff : reserved for ptbl bufs * 0xc200_4000 - 0xc200_8fff : guard page + kstack0 * 0xc200_9000 - 0xfeef_ffff : actual free KVA space * * 64-bit pmap: * Virtual address space layout: * ----------------------------- * 0x0000_0000_0000_0000 - 0xbfff_ffff_ffff_ffff : user process * 0x0000_0000_0000_0000 - 0x8fff_ffff_ffff_ffff : text, data, heap, maps, libraries * 0x9000_0000_0000_0000 - 0xafff_ffff_ffff_ffff : mmio region * 0xb000_0000_0000_0000 - 0xbfff_ffff_ffff_ffff : stack * 0xc000_0000_0000_0000 - 0xcfff_ffff_ffff_ffff : kernel reserved * 0xc000_0000_0000_0000 - endkernel-1 : kernel code & data * endkernel - msgbufp-1 : flat device tree * msgbufp - kernel_pdir-1 : message buffer * kernel_pdir - kernel_pp2d-1 : kernel page directory * kernel_pp2d - . : kernel pointers to page directory * pmap_zero_copy_min - crashdumpmap-1 : reserved for page zero/copy * crashdumpmap - ptbl_buf_pool_vabase-1 : reserved for ptbl bufs * ptbl_buf_pool_vabase - virtual_avail-1 : user page directories and page tables * virtual_avail - 0xcfff_ffff_ffff_ffff : actual free KVA space * 0xd000_0000_0000_0000 - 0xdfff_ffff_ffff_ffff : coprocessor region * 0xe000_0000_0000_0000 - 0xefff_ffff_ffff_ffff : mmio region * 0xf000_0000_0000_0000 - 0xffff_ffff_ffff_ffff : direct map * 0xf000_0000_0000_0000 - +Maxmem : physmem map * - 0xffff_ffff_ffff_ffff : device direct map */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_kstack_pages.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define SPARSE_MAPDEV /* Use power-of-two mappings in mmu_booke_mapdev(), to save entries. */ #define POW2_MAPPINGS #ifdef DEBUG #define debugf(fmt, args...) printf(fmt, ##args) #else #define debugf(fmt, args...) #endif #ifdef __powerpc64__ #define PRI0ptrX "016lx" #else #define PRI0ptrX "08x" #endif #define TODO panic("%s: not implemented", __func__); extern unsigned char _etext[]; extern unsigned char _end[]; extern uint32_t *bootinfo; vm_paddr_t kernload; vm_offset_t kernstart; vm_size_t kernsize; /* Message buffer and tables. */ static vm_offset_t data_start; static vm_size_t data_end; /* Phys/avail memory regions. */ static struct mem_region *availmem_regions; static int availmem_regions_sz; static struct mem_region *physmem_regions; static int physmem_regions_sz; #ifndef __powerpc64__ /* Reserved KVA space and mutex for mmu_booke_zero_page. */ static vm_offset_t zero_page_va; static struct mtx zero_page_mutex; /* Reserved KVA space and mutex for mmu_booke_copy_page. */ static vm_offset_t copy_page_src_va; static vm_offset_t copy_page_dst_va; static struct mtx copy_page_mutex; #endif static struct mtx tlbivax_mutex; /**************************************************************************/ /* PMAP */ /**************************************************************************/ static int mmu_booke_enter_locked(pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int flags, int8_t psind); unsigned int kptbl_min; /* Index of the first kernel ptbl. */ static uma_zone_t ptbl_root_zone; /* * If user pmap is processed with mmu_booke_remove and the resident count * drops to 0, there are no more pages to remove, so we need not continue. */ #define PMAP_REMOVE_DONE(pmap) \ ((pmap) != kernel_pmap && (pmap)->pm_stats.resident_count == 0) #if defined(COMPAT_FREEBSD32) || !defined(__powerpc64__) extern int elf32_nxstack; #endif /**************************************************************************/ /* TLB and TID handling */ /**************************************************************************/ /* Translation ID busy table */ static volatile pmap_t tidbusy[MAXCPU][TID_MAX + 1]; /* * TLB0 capabilities (entry, way numbers etc.). These can vary between e500 * core revisions and should be read from h/w registers during early config. */ uint32_t tlb0_entries; uint32_t tlb0_ways; uint32_t tlb0_entries_per_way; uint32_t tlb1_entries; #define TLB0_ENTRIES (tlb0_entries) #define TLB0_WAYS (tlb0_ways) #define TLB0_ENTRIES_PER_WAY (tlb0_entries_per_way) #define TLB1_ENTRIES (tlb1_entries) static tlbtid_t tid_alloc(struct pmap *); #ifdef DDB #ifdef __powerpc64__ static void tlb_print_entry(int, uint32_t, uint64_t, uint32_t, uint32_t); #else static void tlb_print_entry(int, uint32_t, uint32_t, uint32_t, uint32_t); #endif #endif static void tlb1_read_entry(tlb_entry_t *, unsigned int); static void tlb1_write_entry(tlb_entry_t *, unsigned int); static int tlb1_iomapped(int, vm_paddr_t, vm_size_t, vm_offset_t *); static vm_size_t tlb1_mapin_region(vm_offset_t, vm_paddr_t, vm_size_t, int); static __inline uint32_t tlb_calc_wimg(vm_paddr_t pa, vm_memattr_t ma); static vm_size_t tsize2size(unsigned int); static unsigned int size2tsize(vm_size_t); static unsigned long ilog2(unsigned long); static void set_mas4_defaults(void); static inline void tlb0_flush_entry(vm_offset_t); static inline unsigned int tlb0_tableidx(vm_offset_t, unsigned int); /**************************************************************************/ /* Page table management */ /**************************************************************************/ static struct rwlock_padalign pvh_global_lock; /* Data for the pv entry allocation mechanism */ static uma_zone_t pvzone; static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; #define PV_ENTRY_ZONE_MIN 2048 /* min pv entries in uma zone */ #ifndef PMAP_SHPGPERPROC #define PMAP_SHPGPERPROC 200 #endif static vm_paddr_t pte_vatopa(pmap_t, vm_offset_t); static int pte_enter(pmap_t, vm_page_t, vm_offset_t, uint32_t, boolean_t); static int pte_remove(pmap_t, vm_offset_t, uint8_t); static pte_t *pte_find(pmap_t, vm_offset_t); static void kernel_pte_alloc(vm_offset_t, vm_offset_t); static pv_entry_t pv_alloc(void); static void pv_free(pv_entry_t); static void pv_insert(pmap_t, vm_offset_t, vm_page_t); static void pv_remove(pmap_t, vm_offset_t, vm_page_t); static void booke_pmap_init_qpages(void); static inline void tlb_miss_lock(void); static inline void tlb_miss_unlock(void); #ifdef SMP extern tlb_entry_t __boot_tlb1[]; void pmap_bootstrap_ap(volatile uint32_t *); #endif /* * Kernel MMU interface */ static void mmu_booke_clear_modify(vm_page_t); static void mmu_booke_copy(pmap_t, pmap_t, vm_offset_t, vm_size_t, vm_offset_t); static void mmu_booke_copy_page(vm_page_t, vm_page_t); static void mmu_booke_copy_pages(vm_page_t *, vm_offset_t, vm_page_t *, vm_offset_t, int); static int mmu_booke_enter(pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int flags, int8_t psind); static void mmu_booke_enter_object(pmap_t, vm_offset_t, vm_offset_t, vm_page_t, vm_prot_t); static void mmu_booke_enter_quick(pmap_t, vm_offset_t, vm_page_t, vm_prot_t); static vm_paddr_t mmu_booke_extract(pmap_t, vm_offset_t); static vm_page_t mmu_booke_extract_and_hold(pmap_t, vm_offset_t, vm_prot_t); static void mmu_booke_init(void); static boolean_t mmu_booke_is_modified(vm_page_t); static boolean_t mmu_booke_is_prefaultable(pmap_t, vm_offset_t); static boolean_t mmu_booke_is_referenced(vm_page_t); static int mmu_booke_ts_referenced(vm_page_t); static vm_offset_t mmu_booke_map(vm_offset_t *, vm_paddr_t, vm_paddr_t, int); static int mmu_booke_mincore(pmap_t, vm_offset_t, vm_paddr_t *); static void mmu_booke_object_init_pt(pmap_t, vm_offset_t, vm_object_t, vm_pindex_t, vm_size_t); static boolean_t mmu_booke_page_exists_quick(pmap_t, vm_page_t); static void mmu_booke_page_init(vm_page_t); static int mmu_booke_page_wired_mappings(vm_page_t); static int mmu_booke_pinit(pmap_t); static void mmu_booke_pinit0(pmap_t); static void mmu_booke_protect(pmap_t, vm_offset_t, vm_offset_t, vm_prot_t); static void mmu_booke_qenter(vm_offset_t, vm_page_t *, int); static void mmu_booke_qremove(vm_offset_t, int); static void mmu_booke_release(pmap_t); static void mmu_booke_remove(pmap_t, vm_offset_t, vm_offset_t); static void mmu_booke_remove_all(vm_page_t); static void mmu_booke_remove_write(vm_page_t); static void mmu_booke_unwire(pmap_t, vm_offset_t, vm_offset_t); static void mmu_booke_zero_page(vm_page_t); static void mmu_booke_zero_page_area(vm_page_t, int, int); static void mmu_booke_activate(struct thread *); static void mmu_booke_deactivate(struct thread *); static void mmu_booke_bootstrap(vm_offset_t, vm_offset_t); static void *mmu_booke_mapdev(vm_paddr_t, vm_size_t); static void *mmu_booke_mapdev_attr(vm_paddr_t, vm_size_t, vm_memattr_t); static void mmu_booke_unmapdev(vm_offset_t, vm_size_t); static vm_paddr_t mmu_booke_kextract(vm_offset_t); static void mmu_booke_kenter(vm_offset_t, vm_paddr_t); static void mmu_booke_kenter_attr(vm_offset_t, vm_paddr_t, vm_memattr_t); static void mmu_booke_kremove(vm_offset_t); static boolean_t mmu_booke_dev_direct_mapped(vm_paddr_t, vm_size_t); static void mmu_booke_sync_icache(pmap_t, vm_offset_t, vm_size_t); static void mmu_booke_dumpsys_map(vm_paddr_t pa, size_t, void **); static void mmu_booke_dumpsys_unmap(vm_paddr_t pa, size_t, void *); static void mmu_booke_scan_init(void); static vm_offset_t mmu_booke_quick_enter_page(vm_page_t m); static void mmu_booke_quick_remove_page(vm_offset_t addr); static int mmu_booke_change_attr(vm_offset_t addr, vm_size_t sz, vm_memattr_t mode); static int mmu_booke_decode_kernel_ptr(vm_offset_t addr, int *is_user, vm_offset_t *decoded_addr); static void mmu_booke_page_array_startup(long); static boolean_t mmu_booke_page_is_mapped(vm_page_t m); +static bool mmu_booke_ps_enabled(pmap_t pmap); static struct pmap_funcs mmu_booke_methods = { /* pmap dispatcher interface */ .clear_modify = mmu_booke_clear_modify, .copy = mmu_booke_copy, .copy_page = mmu_booke_copy_page, .copy_pages = mmu_booke_copy_pages, .enter = mmu_booke_enter, .enter_object = mmu_booke_enter_object, .enter_quick = mmu_booke_enter_quick, .extract = mmu_booke_extract, .extract_and_hold = mmu_booke_extract_and_hold, .init = mmu_booke_init, .is_modified = mmu_booke_is_modified, .is_prefaultable = mmu_booke_is_prefaultable, .is_referenced = mmu_booke_is_referenced, .ts_referenced = mmu_booke_ts_referenced, .map = mmu_booke_map, .mincore = mmu_booke_mincore, .object_init_pt = mmu_booke_object_init_pt, .page_exists_quick = mmu_booke_page_exists_quick, .page_init = mmu_booke_page_init, .page_wired_mappings = mmu_booke_page_wired_mappings, .pinit = mmu_booke_pinit, .pinit0 = mmu_booke_pinit0, .protect = mmu_booke_protect, .qenter = mmu_booke_qenter, .qremove = mmu_booke_qremove, .release = mmu_booke_release, .remove = mmu_booke_remove, .remove_all = mmu_booke_remove_all, .remove_write = mmu_booke_remove_write, .sync_icache = mmu_booke_sync_icache, .unwire = mmu_booke_unwire, .zero_page = mmu_booke_zero_page, .zero_page_area = mmu_booke_zero_page_area, .activate = mmu_booke_activate, .deactivate = mmu_booke_deactivate, .quick_enter_page = mmu_booke_quick_enter_page, .quick_remove_page = mmu_booke_quick_remove_page, .page_array_startup = mmu_booke_page_array_startup, .page_is_mapped = mmu_booke_page_is_mapped, + .ps_enabled = mmu_booke_ps_enabled, /* Internal interfaces */ .bootstrap = mmu_booke_bootstrap, .dev_direct_mapped = mmu_booke_dev_direct_mapped, .mapdev = mmu_booke_mapdev, .mapdev_attr = mmu_booke_mapdev_attr, .kenter = mmu_booke_kenter, .kenter_attr = mmu_booke_kenter_attr, .kextract = mmu_booke_kextract, .kremove = mmu_booke_kremove, .unmapdev = mmu_booke_unmapdev, .change_attr = mmu_booke_change_attr, .decode_kernel_ptr = mmu_booke_decode_kernel_ptr, /* dumpsys() support */ .dumpsys_map_chunk = mmu_booke_dumpsys_map, .dumpsys_unmap_chunk = mmu_booke_dumpsys_unmap, .dumpsys_pa_init = mmu_booke_scan_init, }; MMU_DEF(booke_mmu, MMU_TYPE_BOOKE, mmu_booke_methods); #ifdef __powerpc64__ #include "pmap_64.c" #else #include "pmap_32.c" #endif static vm_offset_t tlb1_map_base = VM_MAPDEV_BASE; static __inline uint32_t tlb_calc_wimg(vm_paddr_t pa, vm_memattr_t ma) { uint32_t attrib; int i; if (ma != VM_MEMATTR_DEFAULT) { switch (ma) { case VM_MEMATTR_UNCACHEABLE: return (MAS2_I | MAS2_G); case VM_MEMATTR_WRITE_COMBINING: case VM_MEMATTR_WRITE_BACK: case VM_MEMATTR_PREFETCHABLE: return (MAS2_I); case VM_MEMATTR_WRITE_THROUGH: return (MAS2_W | MAS2_M); case VM_MEMATTR_CACHEABLE: return (MAS2_M); } } /* * Assume the page is cache inhibited and access is guarded unless * it's in our available memory array. */ attrib = _TLB_ENTRY_IO; for (i = 0; i < physmem_regions_sz; i++) { if ((pa >= physmem_regions[i].mr_start) && (pa < (physmem_regions[i].mr_start + physmem_regions[i].mr_size))) { attrib = _TLB_ENTRY_MEM; break; } } return (attrib); } static inline void tlb_miss_lock(void) { #ifdef SMP struct pcpu *pc; if (!smp_started) return; STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { if (pc != pcpup) { CTR3(KTR_PMAP, "%s: tlb miss LOCK of CPU=%d, " "tlb_lock=%p", __func__, pc->pc_cpuid, pc->pc_booke.tlb_lock); KASSERT((pc->pc_cpuid != PCPU_GET(cpuid)), ("tlb_miss_lock: tried to lock self")); tlb_lock(pc->pc_booke.tlb_lock); CTR1(KTR_PMAP, "%s: locked", __func__); } } #endif } static inline void tlb_miss_unlock(void) { #ifdef SMP struct pcpu *pc; if (!smp_started) return; STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { if (pc != pcpup) { CTR2(KTR_PMAP, "%s: tlb miss UNLOCK of CPU=%d", __func__, pc->pc_cpuid); tlb_unlock(pc->pc_booke.tlb_lock); CTR1(KTR_PMAP, "%s: unlocked", __func__); } } #endif } /* Return number of entries in TLB0. */ static __inline void tlb0_get_tlbconf(void) { uint32_t tlb0_cfg; tlb0_cfg = mfspr(SPR_TLB0CFG); tlb0_entries = tlb0_cfg & TLBCFG_NENTRY_MASK; tlb0_ways = (tlb0_cfg & TLBCFG_ASSOC_MASK) >> TLBCFG_ASSOC_SHIFT; tlb0_entries_per_way = tlb0_entries / tlb0_ways; } /* Return number of entries in TLB1. */ static __inline void tlb1_get_tlbconf(void) { uint32_t tlb1_cfg; tlb1_cfg = mfspr(SPR_TLB1CFG); tlb1_entries = tlb1_cfg & TLBCFG_NENTRY_MASK; } /**************************************************************************/ /* Page table related */ /**************************************************************************/ /* Allocate pv_entry structure. */ pv_entry_t pv_alloc(void) { pv_entry_t pv; pv_entry_count++; if (pv_entry_count > pv_entry_high_water) pagedaemon_wakeup(0); /* XXX powerpc NUMA */ pv = uma_zalloc(pvzone, M_NOWAIT); return (pv); } /* Free pv_entry structure. */ static __inline void pv_free(pv_entry_t pve) { pv_entry_count--; uma_zfree(pvzone, pve); } /* Allocate and initialize pv_entry structure. */ static void pv_insert(pmap_t pmap, vm_offset_t va, vm_page_t m) { pv_entry_t pve; //int su = (pmap == kernel_pmap); //debugf("pv_insert: s (su = %d pmap = 0x%08x va = 0x%08x m = 0x%08x)\n", su, // (u_int32_t)pmap, va, (u_int32_t)m); pve = pv_alloc(); if (pve == NULL) panic("pv_insert: no pv entries!"); pve->pv_pmap = pmap; pve->pv_va = va; /* add to pv_list */ PMAP_LOCK_ASSERT(pmap, MA_OWNED); rw_assert(&pvh_global_lock, RA_WLOCKED); TAILQ_INSERT_TAIL(&m->md.pv_list, pve, pv_link); //debugf("pv_insert: e\n"); } /* Destroy pv entry. */ static void pv_remove(pmap_t pmap, vm_offset_t va, vm_page_t m) { pv_entry_t pve; //int su = (pmap == kernel_pmap); //debugf("pv_remove: s (su = %d pmap = 0x%08x va = 0x%08x)\n", su, (u_int32_t)pmap, va); PMAP_LOCK_ASSERT(pmap, MA_OWNED); rw_assert(&pvh_global_lock, RA_WLOCKED); /* find pv entry */ TAILQ_FOREACH(pve, &m->md.pv_list, pv_link) { if ((pmap == pve->pv_pmap) && (va == pve->pv_va)) { /* remove from pv_list */ TAILQ_REMOVE(&m->md.pv_list, pve, pv_link); if (TAILQ_EMPTY(&m->md.pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); /* free pv entry struct */ pv_free(pve); break; } } //debugf("pv_remove: e\n"); } /**************************************************************************/ /* PMAP related */ /**************************************************************************/ /* * This is called during booke_init, before the system is really initialized. */ static void mmu_booke_bootstrap(vm_offset_t start, vm_offset_t kernelend) { vm_paddr_t phys_kernelend; struct mem_region *mp, *mp1; int cnt, i, j; vm_paddr_t s, e, sz; vm_paddr_t physsz, hwphyssz; u_int phys_avail_count; vm_size_t kstack0_sz; vm_paddr_t kstack0_phys; vm_offset_t kstack0; void *dpcpu; debugf("mmu_booke_bootstrap: entered\n"); /* Set interesting system properties */ #ifdef __powerpc64__ hw_direct_map = 1; #else hw_direct_map = 0; #endif #if defined(COMPAT_FREEBSD32) || !defined(__powerpc64__) elf32_nxstack = 1; #endif /* Initialize invalidation mutex */ mtx_init(&tlbivax_mutex, "tlbivax", NULL, MTX_SPIN); /* Read TLB0 size and associativity. */ tlb0_get_tlbconf(); /* * Align kernel start and end address (kernel image). * Note that kernel end does not necessarily relate to kernsize. * kernsize is the size of the kernel that is actually mapped. */ data_start = round_page(kernelend); data_end = data_start; /* Allocate the dynamic per-cpu area. */ dpcpu = (void *)data_end; data_end += DPCPU_SIZE; /* Allocate space for the message buffer. */ msgbufp = (struct msgbuf *)data_end; data_end += msgbufsize; debugf(" msgbufp at 0x%"PRI0ptrX" end = 0x%"PRI0ptrX"\n", (uintptr_t)msgbufp, data_end); data_end = round_page(data_end); data_end = round_page(mmu_booke_alloc_kernel_pgtables(data_end)); /* Retrieve phys/avail mem regions */ mem_regions(&physmem_regions, &physmem_regions_sz, &availmem_regions, &availmem_regions_sz); if (PHYS_AVAIL_ENTRIES < availmem_regions_sz) panic("mmu_booke_bootstrap: phys_avail too small"); data_end = round_page(data_end); vm_page_array = (vm_page_t)data_end; /* * Get a rough idea (upper bound) on the size of the page array. The * vm_page_array will not handle any more pages than we have in the * avail_regions array, and most likely much less. */ sz = 0; for (mp = availmem_regions; mp->mr_size; mp++) { sz += mp->mr_size; } sz = (round_page(sz) / (PAGE_SIZE + sizeof(struct vm_page))); data_end += round_page(sz * sizeof(struct vm_page)); /* Pre-round up to 1MB. This wastes some space, but saves TLB entries */ data_end = roundup2(data_end, 1 << 20); debugf(" data_end: 0x%"PRI0ptrX"\n", data_end); debugf(" kernstart: %#zx\n", kernstart); debugf(" kernsize: %#zx\n", kernsize); if (data_end - kernstart > kernsize) { kernsize += tlb1_mapin_region(kernstart + kernsize, kernload + kernsize, (data_end - kernstart) - kernsize, _TLB_ENTRY_MEM); } data_end = kernstart + kernsize; debugf(" updated data_end: 0x%"PRI0ptrX"\n", data_end); /* * Clear the structures - note we can only do it safely after the * possible additional TLB1 translations are in place (above) so that * all range up to the currently calculated 'data_end' is covered. */ bzero((void *)data_start, data_end - data_start); dpcpu_init(dpcpu, 0); /*******************************************************/ /* Set the start and end of kva. */ /*******************************************************/ virtual_avail = round_page(data_end); virtual_end = VM_MAX_KERNEL_ADDRESS; #ifndef __powerpc64__ /* Allocate KVA space for page zero/copy operations. */ zero_page_va = virtual_avail; virtual_avail += PAGE_SIZE; copy_page_src_va = virtual_avail; virtual_avail += PAGE_SIZE; copy_page_dst_va = virtual_avail; virtual_avail += PAGE_SIZE; debugf("zero_page_va = 0x%"PRI0ptrX"\n", zero_page_va); debugf("copy_page_src_va = 0x%"PRI0ptrX"\n", copy_page_src_va); debugf("copy_page_dst_va = 0x%"PRI0ptrX"\n", copy_page_dst_va); /* Initialize page zero/copy mutexes. */ mtx_init(&zero_page_mutex, "mmu_booke_zero_page", NULL, MTX_DEF); mtx_init(©_page_mutex, "mmu_booke_copy_page", NULL, MTX_DEF); /* Allocate KVA space for ptbl bufs. */ ptbl_buf_pool_vabase = virtual_avail; virtual_avail += PTBL_BUFS * PTBL_PAGES * PAGE_SIZE; debugf("ptbl_buf_pool_vabase = 0x%"PRI0ptrX" end = 0x%"PRI0ptrX"\n", ptbl_buf_pool_vabase, virtual_avail); #endif /* Calculate corresponding physical addresses for the kernel region. */ phys_kernelend = kernload + kernsize; debugf("kernel image and allocated data:\n"); debugf(" kernload = 0x%09jx\n", (uintmax_t)kernload); debugf(" kernstart = 0x%"PRI0ptrX"\n", kernstart); debugf(" kernsize = 0x%"PRI0ptrX"\n", kernsize); /* * Remove kernel physical address range from avail regions list. Page * align all regions. Non-page aligned memory isn't very interesting * to us. Also, sort the entries for ascending addresses. */ sz = 0; cnt = availmem_regions_sz; debugf("processing avail regions:\n"); for (mp = availmem_regions; mp->mr_size; mp++) { s = mp->mr_start; e = mp->mr_start + mp->mr_size; debugf(" %09jx-%09jx -> ", (uintmax_t)s, (uintmax_t)e); /* Check whether this region holds all of the kernel. */ if (s < kernload && e > phys_kernelend) { availmem_regions[cnt].mr_start = phys_kernelend; availmem_regions[cnt++].mr_size = e - phys_kernelend; e = kernload; } /* Look whether this regions starts within the kernel. */ if (s >= kernload && s < phys_kernelend) { if (e <= phys_kernelend) goto empty; s = phys_kernelend; } /* Now look whether this region ends within the kernel. */ if (e > kernload && e <= phys_kernelend) { if (s >= kernload) goto empty; e = kernload; } /* Now page align the start and size of the region. */ s = round_page(s); e = trunc_page(e); if (e < s) e = s; sz = e - s; debugf("%09jx-%09jx = %jx\n", (uintmax_t)s, (uintmax_t)e, (uintmax_t)sz); /* Check whether some memory is left here. */ if (sz == 0) { empty: memmove(mp, mp + 1, (cnt - (mp - availmem_regions)) * sizeof(*mp)); cnt--; mp--; continue; } /* Do an insertion sort. */ for (mp1 = availmem_regions; mp1 < mp; mp1++) if (s < mp1->mr_start) break; if (mp1 < mp) { memmove(mp1 + 1, mp1, (char *)mp - (char *)mp1); mp1->mr_start = s; mp1->mr_size = sz; } else { mp->mr_start = s; mp->mr_size = sz; } } availmem_regions_sz = cnt; /*******************************************************/ /* Steal physical memory for kernel stack from the end */ /* of the first avail region */ /*******************************************************/ kstack0_sz = kstack_pages * PAGE_SIZE; kstack0_phys = availmem_regions[0].mr_start + availmem_regions[0].mr_size; kstack0_phys -= kstack0_sz; availmem_regions[0].mr_size -= kstack0_sz; /*******************************************************/ /* Fill in phys_avail table, based on availmem_regions */ /*******************************************************/ phys_avail_count = 0; physsz = 0; hwphyssz = 0; TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz); debugf("fill in phys_avail:\n"); for (i = 0, j = 0; i < availmem_regions_sz; i++, j += 2) { debugf(" region: 0x%jx - 0x%jx (0x%jx)\n", (uintmax_t)availmem_regions[i].mr_start, (uintmax_t)availmem_regions[i].mr_start + availmem_regions[i].mr_size, (uintmax_t)availmem_regions[i].mr_size); if (hwphyssz != 0 && (physsz + availmem_regions[i].mr_size) >= hwphyssz) { debugf(" hw.physmem adjust\n"); if (physsz < hwphyssz) { phys_avail[j] = availmem_regions[i].mr_start; phys_avail[j + 1] = availmem_regions[i].mr_start + hwphyssz - physsz; physsz = hwphyssz; phys_avail_count++; dump_avail[j] = phys_avail[j]; dump_avail[j + 1] = phys_avail[j + 1]; } break; } phys_avail[j] = availmem_regions[i].mr_start; phys_avail[j + 1] = availmem_regions[i].mr_start + availmem_regions[i].mr_size; phys_avail_count++; physsz += availmem_regions[i].mr_size; dump_avail[j] = phys_avail[j]; dump_avail[j + 1] = phys_avail[j + 1]; } physmem = btoc(physsz); /* Calculate the last available physical address. */ for (i = 0; phys_avail[i + 2] != 0; i += 2) ; Maxmem = powerpc_btop(phys_avail[i + 1]); debugf("Maxmem = 0x%08lx\n", Maxmem); debugf("phys_avail_count = %d\n", phys_avail_count); debugf("physsz = 0x%09jx physmem = %jd (0x%09jx)\n", (uintmax_t)physsz, (uintmax_t)physmem, (uintmax_t)physmem); #ifdef __powerpc64__ /* * Map the physical memory contiguously in TLB1. * Round so it fits into a single mapping. */ tlb1_mapin_region(DMAP_BASE_ADDRESS, 0, phys_avail[i + 1], _TLB_ENTRY_MEM); #endif /*******************************************************/ /* Initialize (statically allocated) kernel pmap. */ /*******************************************************/ PMAP_LOCK_INIT(kernel_pmap); debugf("kernel_pmap = 0x%"PRI0ptrX"\n", (uintptr_t)kernel_pmap); kernel_pte_alloc(virtual_avail, kernstart); for (i = 0; i < MAXCPU; i++) { kernel_pmap->pm_tid[i] = TID_KERNEL; /* Initialize each CPU's tidbusy entry 0 with kernel_pmap */ tidbusy[i][TID_KERNEL] = kernel_pmap; } /* Mark kernel_pmap active on all CPUs */ CPU_FILL(&kernel_pmap->pm_active); /* * Initialize the global pv list lock. */ rw_init(&pvh_global_lock, "pmap pv global"); /*******************************************************/ /* Final setup */ /*******************************************************/ /* Enter kstack0 into kernel map, provide guard page */ kstack0 = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE; thread0.td_kstack = kstack0; thread0.td_kstack_pages = kstack_pages; debugf("kstack_sz = 0x%08jx\n", (uintmax_t)kstack0_sz); debugf("kstack0_phys at 0x%09jx - 0x%09jx\n", (uintmax_t)kstack0_phys, (uintmax_t)kstack0_phys + kstack0_sz); debugf("kstack0 at 0x%"PRI0ptrX" - 0x%"PRI0ptrX"\n", kstack0, kstack0 + kstack0_sz); virtual_avail += KSTACK_GUARD_PAGES * PAGE_SIZE + kstack0_sz; for (i = 0; i < kstack_pages; i++) { mmu_booke_kenter(kstack0, kstack0_phys); kstack0 += PAGE_SIZE; kstack0_phys += PAGE_SIZE; } pmap_bootstrapped = 1; debugf("virtual_avail = %"PRI0ptrX"\n", virtual_avail); debugf("virtual_end = %"PRI0ptrX"\n", virtual_end); debugf("mmu_booke_bootstrap: exit\n"); } #ifdef SMP void tlb1_ap_prep(void) { tlb_entry_t *e, tmp; unsigned int i; /* Prepare TLB1 image for AP processors */ e = __boot_tlb1; for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&tmp, i); if ((tmp.mas1 & MAS1_VALID) && (tmp.mas2 & _TLB_ENTRY_SHARED)) memcpy(e++, &tmp, sizeof(tmp)); } } void pmap_bootstrap_ap(volatile uint32_t *trcp __unused) { int i; /* * Finish TLB1 configuration: the BSP already set up its TLB1 and we * have the snapshot of its contents in the s/w __boot_tlb1[] table * created by tlb1_ap_prep(), so use these values directly to * (re)program AP's TLB1 hardware. * * Start at index 1 because index 0 has the kernel map. */ for (i = 1; i < TLB1_ENTRIES; i++) { if (__boot_tlb1[i].mas1 & MAS1_VALID) tlb1_write_entry(&__boot_tlb1[i], i); } set_mas4_defaults(); } #endif static void booke_pmap_init_qpages(void) { struct pcpu *pc; int i; CPU_FOREACH(i) { pc = pcpu_find(i); pc->pc_qmap_addr = kva_alloc(PAGE_SIZE); if (pc->pc_qmap_addr == 0) panic("pmap_init_qpages: unable to allocate KVA"); } } SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, booke_pmap_init_qpages, NULL); /* * Get the physical page address for the given pmap/virtual address. */ static vm_paddr_t mmu_booke_extract(pmap_t pmap, vm_offset_t va) { vm_paddr_t pa; PMAP_LOCK(pmap); pa = pte_vatopa(pmap, va); PMAP_UNLOCK(pmap); return (pa); } /* * Extract the physical page address associated with the given * kernel virtual address. */ static vm_paddr_t mmu_booke_kextract(vm_offset_t va) { tlb_entry_t e; vm_paddr_t p = 0; int i; #ifdef __powerpc64__ if (va >= DMAP_BASE_ADDRESS && va <= DMAP_MAX_ADDRESS) return (DMAP_TO_PHYS(va)); #endif if (va >= VM_MIN_KERNEL_ADDRESS && va <= VM_MAX_KERNEL_ADDRESS) p = pte_vatopa(kernel_pmap, va); if (p == 0) { /* Check TLB1 mappings */ for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&e, i); if (!(e.mas1 & MAS1_VALID)) continue; if (va >= e.virt && va < e.virt + e.size) return (e.phys + (va - e.virt)); } } return (p); } /* * Initialize the pmap module. * Called by vm_init, to initialize any structures that the pmap * system needs to map virtual memory. */ static void mmu_booke_init() { int shpgperproc = PMAP_SHPGPERPROC; /* * Initialize the address space (zone) for the pv entries. Set a * high water mark so that the system can recover from excessive * numbers of pv entries. */ pvzone = uma_zcreate("PV ENTRY", sizeof(struct pv_entry), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); pv_entry_max = shpgperproc * maxproc + vm_cnt.v_page_count; TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); pv_entry_high_water = 9 * (pv_entry_max / 10); uma_zone_reserve_kva(pvzone, pv_entry_max); /* Pre-fill pvzone with initial number of pv entries. */ uma_prealloc(pvzone, PV_ENTRY_ZONE_MIN); /* Create a UMA zone for page table roots. */ ptbl_root_zone = uma_zcreate("pmap root", PMAP_ROOT_SIZE, NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, UMA_ZONE_VM); /* Initialize ptbl allocation. */ ptbl_init(); } /* * Map a list of wired pages into kernel virtual address space. This is * intended for temporary mappings which do not need page modification or * references recorded. Existing mappings in the region are overwritten. */ static void mmu_booke_qenter(vm_offset_t sva, vm_page_t *m, int count) { vm_offset_t va; va = sva; while (count-- > 0) { mmu_booke_kenter(va, VM_PAGE_TO_PHYS(*m)); va += PAGE_SIZE; m++; } } /* * Remove page mappings from kernel virtual address space. Intended for * temporary mappings entered by mmu_booke_qenter. */ static void mmu_booke_qremove(vm_offset_t sva, int count) { vm_offset_t va; va = sva; while (count-- > 0) { mmu_booke_kremove(va); va += PAGE_SIZE; } } /* * Map a wired page into kernel virtual address space. */ static void mmu_booke_kenter(vm_offset_t va, vm_paddr_t pa) { mmu_booke_kenter_attr(va, pa, VM_MEMATTR_DEFAULT); } static void mmu_booke_kenter_attr(vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma) { uint32_t flags; pte_t *pte; KASSERT(((va >= VM_MIN_KERNEL_ADDRESS) && (va <= VM_MAX_KERNEL_ADDRESS)), ("mmu_booke_kenter: invalid va")); flags = PTE_SR | PTE_SW | PTE_SX | PTE_WIRED | PTE_VALID; flags |= tlb_calc_wimg(pa, ma) << PTE_MAS2_SHIFT; flags |= PTE_PS_4KB; pte = pte_find(kernel_pmap, va); KASSERT((pte != NULL), ("mmu_booke_kenter: invalid va. NULL PTE")); mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); if (PTE_ISVALID(pte)) { CTR1(KTR_PMAP, "%s: replacing entry!", __func__); /* Flush entry from TLB0 */ tlb0_flush_entry(va); } *pte = PTE_RPN_FROM_PA(pa) | flags; //debugf("mmu_booke_kenter: pdir_idx = %d ptbl_idx = %d va=0x%08x " // "pa=0x%08x rpn=0x%08x flags=0x%08x\n", // pdir_idx, ptbl_idx, va, pa, pte->rpn, pte->flags); /* Flush the real memory from the instruction cache. */ if ((flags & (PTE_I | PTE_G)) == 0) __syncicache((void *)va, PAGE_SIZE); tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); } /* * Remove a page from kernel page table. */ static void mmu_booke_kremove(vm_offset_t va) { pte_t *pte; CTR2(KTR_PMAP,"%s: s (va = 0x%"PRI0ptrX")\n", __func__, va); KASSERT(((va >= VM_MIN_KERNEL_ADDRESS) && (va <= VM_MAX_KERNEL_ADDRESS)), ("mmu_booke_kremove: invalid va")); pte = pte_find(kernel_pmap, va); if (!PTE_ISVALID(pte)) { CTR1(KTR_PMAP, "%s: invalid pte", __func__); return; } mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); /* Invalidate entry in TLB0, update PTE. */ tlb0_flush_entry(va); *pte = 0; tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); } /* * Figure out where a given kernel pointer (usually in a fault) points * to from the VM's perspective, potentially remapping into userland's * address space. */ static int mmu_booke_decode_kernel_ptr(vm_offset_t addr, int *is_user, vm_offset_t *decoded_addr) { if (trunc_page(addr) <= VM_MAXUSER_ADDRESS) *is_user = 1; else *is_user = 0; *decoded_addr = addr; return (0); } static boolean_t mmu_booke_page_is_mapped(vm_page_t m) { return (!TAILQ_EMPTY(&(m)->md.pv_list)); } +static bool +mmu_booke_ps_enabled(pmap_t pmap __unused) +{ + return (false); +} + /* * Initialize pmap associated with process 0. */ static void mmu_booke_pinit0(pmap_t pmap) { PMAP_LOCK_INIT(pmap); mmu_booke_pinit(pmap); PCPU_SET(curpmap, pmap); } /* * Insert the given physical page at the specified virtual address in the * target physical map with the protection requested. If specified the page * will be wired down. */ static int mmu_booke_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind) { int error; rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); error = mmu_booke_enter_locked(pmap, va, m, prot, flags, psind); PMAP_UNLOCK(pmap); rw_wunlock(&pvh_global_lock); return (error); } static int mmu_booke_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int pmap_flags, int8_t psind __unused) { pte_t *pte; vm_paddr_t pa; pte_t flags; int error, su, sync; pa = VM_PAGE_TO_PHYS(m); su = (pmap == kernel_pmap); sync = 0; //debugf("mmu_booke_enter_locked: s (pmap=0x%08x su=%d tid=%d m=0x%08x va=0x%08x " // "pa=0x%08x prot=0x%08x flags=%#x)\n", // (u_int32_t)pmap, su, pmap->pm_tid, // (u_int32_t)m, va, pa, prot, flags); if (su) { KASSERT(((va >= virtual_avail) && (va <= VM_MAX_KERNEL_ADDRESS)), ("mmu_booke_enter_locked: kernel pmap, non kernel va")); } else { KASSERT((va <= VM_MAXUSER_ADDRESS), ("mmu_booke_enter_locked: user pmap, non user va")); } if ((m->oflags & VPO_UNMANAGED) == 0) { if ((pmap_flags & PMAP_ENTER_QUICK_LOCKED) == 0) VM_PAGE_OBJECT_BUSY_ASSERT(m); else VM_OBJECT_ASSERT_LOCKED(m->object); } PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * If there is an existing mapping, and the physical address has not * changed, must be protection or wiring change. */ if (((pte = pte_find(pmap, va)) != NULL) && (PTE_ISVALID(pte)) && (PTE_PA(pte) == pa)) { /* * Before actually updating pte->flags we calculate and * prepare its new value in a helper var. */ flags = *pte; flags &= ~(PTE_UW | PTE_UX | PTE_SW | PTE_SX | PTE_MODIFIED); /* Wiring change, just update stats. */ if ((pmap_flags & PMAP_ENTER_WIRED) != 0) { if (!PTE_ISWIRED(pte)) { flags |= PTE_WIRED; pmap->pm_stats.wired_count++; } } else { if (PTE_ISWIRED(pte)) { flags &= ~PTE_WIRED; pmap->pm_stats.wired_count--; } } if (prot & VM_PROT_WRITE) { /* Add write permissions. */ flags |= PTE_SW; if (!su) flags |= PTE_UW; if ((flags & PTE_MANAGED) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); } else { /* Handle modified pages, sense modify status. */ /* * The PTE_MODIFIED flag could be set by underlying * TLB misses since we last read it (above), possibly * other CPUs could update it so we check in the PTE * directly rather than rely on that saved local flags * copy. */ if (PTE_ISMODIFIED(pte)) vm_page_dirty(m); } if (prot & VM_PROT_EXECUTE) { flags |= PTE_SX; if (!su) flags |= PTE_UX; /* * Check existing flags for execute permissions: if we * are turning execute permissions on, icache should * be flushed. */ if ((*pte & (PTE_UX | PTE_SX)) == 0) sync++; } flags &= ~PTE_REFERENCED; /* * The new flags value is all calculated -- only now actually * update the PTE. */ mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); tlb0_flush_entry(va); *pte &= ~PTE_FLAGS_MASK; *pte |= flags; tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); } else { /* * If there is an existing mapping, but it's for a different * physical address, pte_enter() will delete the old mapping. */ //if ((pte != NULL) && PTE_ISVALID(pte)) // debugf("mmu_booke_enter_locked: replace\n"); //else // debugf("mmu_booke_enter_locked: new\n"); /* Now set up the flags and install the new mapping. */ flags = (PTE_SR | PTE_VALID); flags |= PTE_M; if (!su) flags |= PTE_UR; if (prot & VM_PROT_WRITE) { flags |= PTE_SW; if (!su) flags |= PTE_UW; if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_set(m, PGA_WRITEABLE); } if (prot & VM_PROT_EXECUTE) { flags |= PTE_SX; if (!su) flags |= PTE_UX; } /* If its wired update stats. */ if ((pmap_flags & PMAP_ENTER_WIRED) != 0) flags |= PTE_WIRED; error = pte_enter(pmap, m, va, flags, (pmap_flags & PMAP_ENTER_NOSLEEP) != 0); if (error != 0) return (KERN_RESOURCE_SHORTAGE); if ((flags & PMAP_ENTER_WIRED) != 0) pmap->pm_stats.wired_count++; /* Flush the real memory from the instruction cache. */ if (prot & VM_PROT_EXECUTE) sync++; } if (sync && (su || pmap == PCPU_GET(curpmap))) { __syncicache((void *)va, PAGE_SIZE); sync = 0; } return (KERN_SUCCESS); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ static void mmu_booke_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { vm_page_t m; vm_pindex_t diff, psize; VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); m = m_start; rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { mmu_booke_enter_locked(pmap, start + ptoa(diff), m, prot & (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_NOSLEEP | PMAP_ENTER_QUICK_LOCKED, 0); m = TAILQ_NEXT(m, listq); } PMAP_UNLOCK(pmap); rw_wunlock(&pvh_global_lock); } static void mmu_booke_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); mmu_booke_enter_locked(pmap, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_NOSLEEP | PMAP_ENTER_QUICK_LOCKED, 0); PMAP_UNLOCK(pmap); rw_wunlock(&pvh_global_lock); } /* * Remove the given range of addresses from the specified map. * * It is assumed that the start and end are properly rounded to the page size. */ static void mmu_booke_remove(pmap_t pmap, vm_offset_t va, vm_offset_t endva) { pte_t *pte; uint8_t hold_flag; int su = (pmap == kernel_pmap); //debugf("mmu_booke_remove: s (su = %d pmap=0x%08x tid=%d va=0x%08x endva=0x%08x)\n", // su, (u_int32_t)pmap, pmap->pm_tid, va, endva); if (su) { KASSERT(((va >= virtual_avail) && (va <= VM_MAX_KERNEL_ADDRESS)), ("mmu_booke_remove: kernel pmap, non kernel va")); } else { KASSERT((va <= VM_MAXUSER_ADDRESS), ("mmu_booke_remove: user pmap, non user va")); } if (PMAP_REMOVE_DONE(pmap)) { //debugf("mmu_booke_remove: e (empty)\n"); return; } hold_flag = PTBL_HOLD_FLAG(pmap); //debugf("mmu_booke_remove: hold_flag = %d\n", hold_flag); rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); for (; va < endva; va += PAGE_SIZE) { pte = pte_find_next(pmap, &va); if ((pte == NULL) || !PTE_ISVALID(pte)) break; if (va >= endva) break; pte_remove(pmap, va, hold_flag); } PMAP_UNLOCK(pmap); rw_wunlock(&pvh_global_lock); //debugf("mmu_booke_remove: e\n"); } /* * Remove physical page from all pmaps in which it resides. */ static void mmu_booke_remove_all(vm_page_t m) { pv_entry_t pv, pvn; uint8_t hold_flag; rw_wlock(&pvh_global_lock); TAILQ_FOREACH_SAFE(pv, &m->md.pv_list, pv_link, pvn) { PMAP_LOCK(pv->pv_pmap); hold_flag = PTBL_HOLD_FLAG(pv->pv_pmap); pte_remove(pv->pv_pmap, pv->pv_va, hold_flag); PMAP_UNLOCK(pv->pv_pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(&pvh_global_lock); } /* * Map a range of physical addresses into kernel virtual address space. */ static vm_offset_t mmu_booke_map(vm_offset_t *virt, vm_paddr_t pa_start, vm_paddr_t pa_end, int prot) { vm_offset_t sva = *virt; vm_offset_t va = sva; #ifdef __powerpc64__ /* XXX: Handle memory not starting at 0x0. */ if (pa_end < ctob(Maxmem)) return (PHYS_TO_DMAP(pa_start)); #endif while (pa_start < pa_end) { mmu_booke_kenter(va, pa_start); va += PAGE_SIZE; pa_start += PAGE_SIZE; } *virt = va; return (sva); } /* * The pmap must be activated before it's address space can be accessed in any * way. */ static void mmu_booke_activate(struct thread *td) { pmap_t pmap; u_int cpuid; pmap = &td->td_proc->p_vmspace->vm_pmap; CTR5(KTR_PMAP, "%s: s (td = %p, proc = '%s', id = %d, pmap = 0x%"PRI0ptrX")", __func__, td, td->td_proc->p_comm, td->td_proc->p_pid, pmap); KASSERT((pmap != kernel_pmap), ("mmu_booke_activate: kernel_pmap!")); sched_pin(); cpuid = PCPU_GET(cpuid); CPU_SET_ATOMIC(cpuid, &pmap->pm_active); PCPU_SET(curpmap, pmap); if (pmap->pm_tid[cpuid] == TID_NONE) tid_alloc(pmap); /* Load PID0 register with pmap tid value. */ mtspr(SPR_PID0, pmap->pm_tid[cpuid]); __asm __volatile("isync"); mtspr(SPR_DBCR0, td->td_pcb->pcb_cpu.booke.dbcr0); sched_unpin(); CTR3(KTR_PMAP, "%s: e (tid = %d for '%s')", __func__, pmap->pm_tid[PCPU_GET(cpuid)], td->td_proc->p_comm); } /* * Deactivate the specified process's address space. */ static void mmu_booke_deactivate(struct thread *td) { pmap_t pmap; pmap = &td->td_proc->p_vmspace->vm_pmap; CTR5(KTR_PMAP, "%s: td=%p, proc = '%s', id = %d, pmap = 0x%"PRI0ptrX, __func__, td, td->td_proc->p_comm, td->td_proc->p_pid, pmap); td->td_pcb->pcb_cpu.booke.dbcr0 = mfspr(SPR_DBCR0); CPU_CLR_ATOMIC(PCPU_GET(cpuid), &pmap->pm_active); PCPU_SET(curpmap, NULL); } /* * Copy the range specified by src_addr/len * from the source map to the range dst_addr/len * in the destination map. * * This routine is only advisory and need not do anything. */ static void mmu_booke_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, vm_offset_t src_addr) { } /* * Set the physical protection on the specified range of this map as requested. */ static void mmu_booke_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { vm_offset_t va; vm_page_t m; pte_t *pte; if ((prot & VM_PROT_READ) == VM_PROT_NONE) { mmu_booke_remove(pmap, sva, eva); return; } if (prot & VM_PROT_WRITE) return; PMAP_LOCK(pmap); for (va = sva; va < eva; va += PAGE_SIZE) { if ((pte = pte_find(pmap, va)) != NULL) { if (PTE_ISVALID(pte)) { m = PHYS_TO_VM_PAGE(PTE_PA(pte)); mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); /* Handle modified pages. */ if (PTE_ISMODIFIED(pte) && PTE_ISMANAGED(pte)) vm_page_dirty(m); tlb0_flush_entry(va); *pte &= ~(PTE_UW | PTE_SW | PTE_MODIFIED); tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); } } } PMAP_UNLOCK(pmap); } /* * Clear the write and modified bits in each of the given page's mappings. */ static void mmu_booke_remove_write(vm_page_t m) { pv_entry_t pv; pte_t *pte; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("mmu_booke_remove_write: page %p is not managed", m)); vm_page_assert_busied(m); if (!pmap_page_is_write_mapped(m)) return; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { PMAP_LOCK(pv->pv_pmap); if ((pte = pte_find(pv->pv_pmap, pv->pv_va)) != NULL) { if (PTE_ISVALID(pte)) { m = PHYS_TO_VM_PAGE(PTE_PA(pte)); mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); /* Handle modified pages. */ if (PTE_ISMODIFIED(pte)) vm_page_dirty(m); /* Flush mapping from TLB0. */ *pte &= ~(PTE_UW | PTE_SW | PTE_MODIFIED); tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); } } PMAP_UNLOCK(pv->pv_pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(&pvh_global_lock); } /* * Atomically extract and hold the physical page with the given * pmap and virtual address pair if that mapping permits the given * protection. */ static vm_page_t mmu_booke_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) { pte_t *pte; vm_page_t m; uint32_t pte_wbit; m = NULL; PMAP_LOCK(pmap); pte = pte_find(pmap, va); if ((pte != NULL) && PTE_ISVALID(pte)) { if (pmap == kernel_pmap) pte_wbit = PTE_SW; else pte_wbit = PTE_UW; if ((*pte & pte_wbit) != 0 || (prot & VM_PROT_WRITE) == 0) { m = PHYS_TO_VM_PAGE(PTE_PA(pte)); if (!vm_page_wire_mapped(m)) m = NULL; } } PMAP_UNLOCK(pmap); return (m); } /* * Initialize a vm_page's machine-dependent fields. */ static void mmu_booke_page_init(vm_page_t m) { m->md.pv_tracked = 0; TAILQ_INIT(&m->md.pv_list); } /* * Return whether or not the specified physical page was modified * in any of physical maps. */ static boolean_t mmu_booke_is_modified(vm_page_t m) { pte_t *pte; pv_entry_t pv; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("mmu_booke_is_modified: page %p is not managed", m)); rv = FALSE; /* * If the page is not busied then this check is racy. */ if (!pmap_page_is_write_mapped(m)) return (FALSE); rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { PMAP_LOCK(pv->pv_pmap); if ((pte = pte_find(pv->pv_pmap, pv->pv_va)) != NULL && PTE_ISVALID(pte)) { if (PTE_ISMODIFIED(pte)) rv = TRUE; } PMAP_UNLOCK(pv->pv_pmap); if (rv) break; } rw_wunlock(&pvh_global_lock); return (rv); } /* * Return whether or not the specified virtual address is eligible * for prefault. */ static boolean_t mmu_booke_is_prefaultable(pmap_t pmap, vm_offset_t addr) { return (FALSE); } /* * Return whether or not the specified physical page was referenced * in any physical maps. */ static boolean_t mmu_booke_is_referenced(vm_page_t m) { pte_t *pte; pv_entry_t pv; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("mmu_booke_is_referenced: page %p is not managed", m)); rv = FALSE; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { PMAP_LOCK(pv->pv_pmap); if ((pte = pte_find(pv->pv_pmap, pv->pv_va)) != NULL && PTE_ISVALID(pte)) { if (PTE_ISREFERENCED(pte)) rv = TRUE; } PMAP_UNLOCK(pv->pv_pmap); if (rv) break; } rw_wunlock(&pvh_global_lock); return (rv); } /* * Clear the modify bits on the specified physical page. */ static void mmu_booke_clear_modify(vm_page_t m) { pte_t *pte; pv_entry_t pv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("mmu_booke_clear_modify: page %p is not managed", m)); vm_page_assert_busied(m); if (!pmap_page_is_write_mapped(m)) return; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { PMAP_LOCK(pv->pv_pmap); if ((pte = pte_find(pv->pv_pmap, pv->pv_va)) != NULL && PTE_ISVALID(pte)) { mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); if (*pte & (PTE_SW | PTE_UW | PTE_MODIFIED)) { tlb0_flush_entry(pv->pv_va); *pte &= ~(PTE_SW | PTE_UW | PTE_MODIFIED | PTE_REFERENCED); } tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); } PMAP_UNLOCK(pv->pv_pmap); } rw_wunlock(&pvh_global_lock); } /* * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. * * As an optimization, update the page's dirty field if a modified bit is * found while counting reference bits. This opportunistic update can be * performed at low cost and can eliminate the need for some future calls * to pmap_is_modified(). However, since this function stops after * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some * dirty pages. Those dirty pages will only be detected by a future call * to pmap_is_modified(). */ static int mmu_booke_ts_referenced(vm_page_t m) { pte_t *pte; pv_entry_t pv; int count; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("mmu_booke_ts_referenced: page %p is not managed", m)); count = 0; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { PMAP_LOCK(pv->pv_pmap); if ((pte = pte_find(pv->pv_pmap, pv->pv_va)) != NULL && PTE_ISVALID(pte)) { if (PTE_ISMODIFIED(pte)) vm_page_dirty(m); if (PTE_ISREFERENCED(pte)) { mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); tlb0_flush_entry(pv->pv_va); *pte &= ~PTE_REFERENCED; tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); if (++count >= PMAP_TS_REFERENCED_MAX) { PMAP_UNLOCK(pv->pv_pmap); break; } } } PMAP_UNLOCK(pv->pv_pmap); } rw_wunlock(&pvh_global_lock); return (count); } /* * Clear the wired attribute from the mappings for the specified range of * addresses in the given pmap. Every valid mapping within that range must * have the wired attribute set. In contrast, invalid mappings cannot have * the wired attribute set, so they are ignored. * * The wired attribute of the page table entry is not a hardware feature, so * there is no need to invalidate any TLB entries. */ static void mmu_booke_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t va; pte_t *pte; PMAP_LOCK(pmap); for (va = sva; va < eva; va += PAGE_SIZE) { if ((pte = pte_find(pmap, va)) != NULL && PTE_ISVALID(pte)) { if (!PTE_ISWIRED(pte)) panic("mmu_booke_unwire: pte %p isn't wired", pte); *pte &= ~PTE_WIRED; pmap->pm_stats.wired_count--; } } PMAP_UNLOCK(pmap); } /* * Return true if the pmap's pv is one of the first 16 pvs linked to from this * page. This count may be changed upwards or downwards in the future; it is * only necessary that true be returned for a small subset of pmaps for proper * page aging. */ static boolean_t mmu_booke_page_exists_quick(pmap_t pmap, vm_page_t m) { pv_entry_t pv; int loops; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("mmu_booke_page_exists_quick: page %p is not managed", m)); loops = 0; rv = FALSE; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { if (pv->pv_pmap == pmap) { rv = TRUE; break; } if (++loops >= 16) break; } rw_wunlock(&pvh_global_lock); return (rv); } /* * Return the number of managed mappings to the given physical page that are * wired. */ static int mmu_booke_page_wired_mappings(vm_page_t m) { pv_entry_t pv; pte_t *pte; int count = 0; if ((m->oflags & VPO_UNMANAGED) != 0) return (count); rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { PMAP_LOCK(pv->pv_pmap); if ((pte = pte_find(pv->pv_pmap, pv->pv_va)) != NULL) if (PTE_ISVALID(pte) && PTE_ISWIRED(pte)) count++; PMAP_UNLOCK(pv->pv_pmap); } rw_wunlock(&pvh_global_lock); return (count); } static int mmu_booke_dev_direct_mapped(vm_paddr_t pa, vm_size_t size) { int i; vm_offset_t va; /* * This currently does not work for entries that * overlap TLB1 entries. */ for (i = 0; i < TLB1_ENTRIES; i ++) { if (tlb1_iomapped(i, pa, size, &va) == 0) return (0); } return (EFAULT); } void mmu_booke_dumpsys_map(vm_paddr_t pa, size_t sz, void **va) { vm_paddr_t ppa; vm_offset_t ofs; vm_size_t gran; /* Minidumps are based on virtual memory addresses. */ if (do_minidump) { *va = (void *)(vm_offset_t)pa; return; } /* Raw physical memory dumps don't have a virtual address. */ /* We always map a 256MB page at 256M. */ gran = 256 * 1024 * 1024; ppa = rounddown2(pa, gran); ofs = pa - ppa; *va = (void *)gran; tlb1_set_entry((vm_offset_t)va, ppa, gran, _TLB_ENTRY_IO); if (sz > (gran - ofs)) tlb1_set_entry((vm_offset_t)(va + gran), ppa + gran, gran, _TLB_ENTRY_IO); } void mmu_booke_dumpsys_unmap(vm_paddr_t pa, size_t sz, void *va) { vm_paddr_t ppa; vm_offset_t ofs; vm_size_t gran; tlb_entry_t e; int i; /* Minidumps are based on virtual memory addresses. */ /* Nothing to do... */ if (do_minidump) return; for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&e, i); if (!(e.mas1 & MAS1_VALID)) break; } /* Raw physical memory dumps don't have a virtual address. */ i--; e.mas1 = 0; e.mas2 = 0; e.mas3 = 0; tlb1_write_entry(&e, i); gran = 256 * 1024 * 1024; ppa = rounddown2(pa, gran); ofs = pa - ppa; if (sz > (gran - ofs)) { i--; e.mas1 = 0; e.mas2 = 0; e.mas3 = 0; tlb1_write_entry(&e, i); } } extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1]; void mmu_booke_scan_init() { vm_offset_t va; pte_t *pte; int i; if (!do_minidump) { /* Initialize phys. segments for dumpsys(). */ memset(&dump_map, 0, sizeof(dump_map)); mem_regions(&physmem_regions, &physmem_regions_sz, &availmem_regions, &availmem_regions_sz); for (i = 0; i < physmem_regions_sz; i++) { dump_map[i].pa_start = physmem_regions[i].mr_start; dump_map[i].pa_size = physmem_regions[i].mr_size; } return; } /* Virtual segments for minidumps: */ memset(&dump_map, 0, sizeof(dump_map)); /* 1st: kernel .data and .bss. */ dump_map[0].pa_start = trunc_page((uintptr_t)_etext); dump_map[0].pa_size = round_page((uintptr_t)_end) - dump_map[0].pa_start; /* 2nd: msgbuf and tables (see pmap_bootstrap()). */ dump_map[1].pa_start = data_start; dump_map[1].pa_size = data_end - data_start; /* 3rd: kernel VM. */ va = dump_map[1].pa_start + dump_map[1].pa_size; /* Find start of next chunk (from va). */ while (va < virtual_end) { /* Don't dump the buffer cache. */ if (va >= kmi.buffer_sva && va < kmi.buffer_eva) { va = kmi.buffer_eva; continue; } pte = pte_find(kernel_pmap, va); if (pte != NULL && PTE_ISVALID(pte)) break; va += PAGE_SIZE; } if (va < virtual_end) { dump_map[2].pa_start = va; va += PAGE_SIZE; /* Find last page in chunk. */ while (va < virtual_end) { /* Don't run into the buffer cache. */ if (va == kmi.buffer_sva) break; pte = pte_find(kernel_pmap, va); if (pte == NULL || !PTE_ISVALID(pte)) break; va += PAGE_SIZE; } dump_map[2].pa_size = va - dump_map[2].pa_start; } } /* * Map a set of physical memory pages into the kernel virtual address space. * Return a pointer to where it is mapped. This routine is intended to be used * for mapping device memory, NOT real memory. */ static void * mmu_booke_mapdev(vm_paddr_t pa, vm_size_t size) { return (mmu_booke_mapdev_attr(pa, size, VM_MEMATTR_DEFAULT)); } static int tlb1_find_pa(vm_paddr_t pa, tlb_entry_t *e) { int i; for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(e, i); if ((e->mas1 & MAS1_VALID) == 0) continue; if (e->phys == pa) return (i); } return (-1); } static void * mmu_booke_mapdev_attr(vm_paddr_t pa, vm_size_t size, vm_memattr_t ma) { tlb_entry_t e; vm_paddr_t tmppa; #ifndef __powerpc64__ uintptr_t tmpva; #endif uintptr_t va, retva; vm_size_t sz; int i; int wimge; /* * Check if this is premapped in TLB1. */ sz = size; tmppa = pa; va = ~0; wimge = tlb_calc_wimg(pa, ma); for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&e, i); if (!(e.mas1 & MAS1_VALID)) continue; if (wimge != (e.mas2 & (MAS2_WIMGE_MASK & ~_TLB_ENTRY_SHARED))) continue; if (tmppa >= e.phys && tmppa < e.phys + e.size) { va = e.virt + (pa - e.phys); tmppa = e.phys + e.size; sz -= MIN(sz, e.size - (pa - e.phys)); while (sz > 0 && (i = tlb1_find_pa(tmppa, &e)) != -1) { if (wimge != (e.mas2 & (MAS2_WIMGE_MASK & ~_TLB_ENTRY_SHARED))) break; sz -= MIN(sz, e.size); tmppa = e.phys + e.size; } if (sz != 0) break; return ((void *)va); } } size = roundup(size, PAGE_SIZE); #ifdef __powerpc64__ KASSERT(pa < VM_MAPDEV_PA_MAX, ("Unsupported physical address! %lx", pa)); va = VM_MAPDEV_BASE + pa; retva = va; #ifdef POW2_MAPPINGS /* * Align the mapping to a power of 2 size, taking into account that we * may need to increase the size multiple times to satisfy the size and * alignment requirements. * * This works in the general case because it's very rare (near never?) * to have different access properties (WIMG) within a single * power-of-two region. If a design does call for that, POW2_MAPPINGS * can be undefined, and exact mappings will be used instead. */ sz = size; size = roundup2(size, 1 << ilog2(size)); while (rounddown2(va, size) + size < va + sz) size <<= 1; va = rounddown2(va, size); pa = rounddown2(pa, size); #endif #else /* * The device mapping area is between VM_MAXUSER_ADDRESS and * VM_MIN_KERNEL_ADDRESS. This gives 1GB of device addressing. */ #ifdef SPARSE_MAPDEV /* * With a sparse mapdev, align to the largest starting region. This * could feasibly be optimized for a 'best-fit' alignment, but that * calculation could be very costly. * Align to the smaller of: * - first set bit in overlap of (pa & size mask) * - largest size envelope * * It's possible the device mapping may start at a PA that's not larger * than the size mask, so we need to offset in to maximize the TLB entry * range and minimize the number of used TLB entries. */ do { tmpva = tlb1_map_base; sz = ffsl((~((1 << flsl(size-1)) - 1)) & pa); sz = sz ? min(roundup(sz + 3, 4), flsl(size) - 1) : flsl(size) - 1; va = roundup(tlb1_map_base, 1 << sz) | (((1 << sz) - 1) & pa); } while (!atomic_cmpset_int(&tlb1_map_base, tmpva, va + size)); #endif va = atomic_fetchadd_int(&tlb1_map_base, size); retva = va; #endif if (tlb1_mapin_region(va, pa, size, tlb_calc_wimg(pa, ma)) != size) return (NULL); return ((void *)retva); } /* * 'Unmap' a range mapped by mmu_booke_mapdev(). */ static void mmu_booke_unmapdev(vm_offset_t va, vm_size_t size) { #ifdef SUPPORTS_SHRINKING_TLB1 vm_offset_t base, offset; /* * Unmap only if this is inside kernel virtual space. */ if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= VM_MAX_KERNEL_ADDRESS)) { base = trunc_page(va); offset = va & PAGE_MASK; size = roundup(offset + size, PAGE_SIZE); mmu_booke_qremove(base, atop(size)); kva_free(base, size); } #endif } /* * mmu_booke_object_init_pt preloads the ptes for a given object into the * specified pmap. This eliminates the blast of soft faults on process startup * and immediately after an mmap. */ static void mmu_booke_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, ("mmu_booke_object_init_pt: non-device object")); } /* * Perform the pmap work for mincore. */ static int mmu_booke_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *pap) { /* XXX: this should be implemented at some point */ return (0); } static int mmu_booke_change_attr(vm_offset_t addr, vm_size_t sz, vm_memattr_t mode) { vm_offset_t va; pte_t *pte; int i, j; tlb_entry_t e; addr = trunc_page(addr); /* Only allow changes to mapped kernel addresses. This includes: * - KVA * - DMAP (powerpc64) * - Device mappings */ if (addr <= VM_MAXUSER_ADDRESS || #ifdef __powerpc64__ (addr >= tlb1_map_base && addr < DMAP_BASE_ADDRESS) || (addr > DMAP_MAX_ADDRESS && addr < VM_MIN_KERNEL_ADDRESS) || #else (addr >= tlb1_map_base && addr < VM_MIN_KERNEL_ADDRESS) || #endif (addr > VM_MAX_KERNEL_ADDRESS)) return (EINVAL); /* Check TLB1 mappings */ for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&e, i); if (!(e.mas1 & MAS1_VALID)) continue; if (addr >= e.virt && addr < e.virt + e.size) break; } if (i < TLB1_ENTRIES) { /* Only allow full mappings to be modified for now. */ /* Validate the range. */ for (j = i, va = addr; va < addr + sz; va += e.size, j++) { tlb1_read_entry(&e, j); if (va != e.virt || (sz - (va - addr) < e.size)) return (EINVAL); } for (va = addr; va < addr + sz; va += e.size, i++) { tlb1_read_entry(&e, i); e.mas2 &= ~MAS2_WIMGE_MASK; e.mas2 |= tlb_calc_wimg(e.phys, mode); /* * Write it out to the TLB. Should really re-sync with other * cores. */ tlb1_write_entry(&e, i); } return (0); } /* Not in TLB1, try through pmap */ /* First validate the range. */ for (va = addr; va < addr + sz; va += PAGE_SIZE) { pte = pte_find(kernel_pmap, va); if (pte == NULL || !PTE_ISVALID(pte)) return (EINVAL); } mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); for (va = addr; va < addr + sz; va += PAGE_SIZE) { pte = pte_find(kernel_pmap, va); *pte &= ~(PTE_MAS2_MASK << PTE_MAS2_SHIFT); *pte |= tlb_calc_wimg(PTE_PA(pte), mode) << PTE_MAS2_SHIFT; tlb0_flush_entry(va); } tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); return (0); } static void mmu_booke_page_array_startup(long pages) { vm_page_array_size = pages; } /**************************************************************************/ /* TID handling */ /**************************************************************************/ /* * Allocate a TID. If necessary, steal one from someone else. * The new TID is flushed from the TLB before returning. */ static tlbtid_t tid_alloc(pmap_t pmap) { tlbtid_t tid; int thiscpu; KASSERT((pmap != kernel_pmap), ("tid_alloc: kernel pmap")); CTR2(KTR_PMAP, "%s: s (pmap = %p)", __func__, pmap); thiscpu = PCPU_GET(cpuid); tid = PCPU_GET(booke.tid_next); if (tid > TID_MAX) tid = TID_MIN; PCPU_SET(booke.tid_next, tid + 1); /* If we are stealing TID then clear the relevant pmap's field */ if (tidbusy[thiscpu][tid] != NULL) { CTR2(KTR_PMAP, "%s: warning: stealing tid %d", __func__, tid); tidbusy[thiscpu][tid]->pm_tid[thiscpu] = TID_NONE; /* Flush all entries from TLB0 matching this TID. */ tid_flush(tid); } tidbusy[thiscpu][tid] = pmap; pmap->pm_tid[thiscpu] = tid; __asm __volatile("msync; isync"); CTR3(KTR_PMAP, "%s: e (%02d next = %02d)", __func__, tid, PCPU_GET(booke.tid_next)); return (tid); } /**************************************************************************/ /* TLB0 handling */ /**************************************************************************/ /* Convert TLB0 va and way number to tlb0[] table index. */ static inline unsigned int tlb0_tableidx(vm_offset_t va, unsigned int way) { unsigned int idx; idx = (way * TLB0_ENTRIES_PER_WAY); idx += (va & MAS2_TLB0_ENTRY_IDX_MASK) >> MAS2_TLB0_ENTRY_IDX_SHIFT; return (idx); } /* * Invalidate TLB0 entry. */ static inline void tlb0_flush_entry(vm_offset_t va) { CTR2(KTR_PMAP, "%s: s va=0x%08x", __func__, va); mtx_assert(&tlbivax_mutex, MA_OWNED); __asm __volatile("tlbivax 0, %0" :: "r"(va & MAS2_EPN_MASK)); __asm __volatile("isync; msync"); __asm __volatile("tlbsync; msync"); CTR1(KTR_PMAP, "%s: e", __func__); } /**************************************************************************/ /* TLB1 handling */ /**************************************************************************/ /* * TLB1 mapping notes: * * TLB1[0] Kernel text and data. * TLB1[1-15] Additional kernel text and data mappings (if required), PCI * windows, other devices mappings. */ /* * Read an entry from given TLB1 slot. */ void tlb1_read_entry(tlb_entry_t *entry, unsigned int slot) { register_t msr; uint32_t mas0; KASSERT((entry != NULL), ("%s(): Entry is NULL!", __func__)); msr = mfmsr(); __asm __volatile("wrteei 0"); mas0 = MAS0_TLBSEL(1) | MAS0_ESEL(slot); mtspr(SPR_MAS0, mas0); __asm __volatile("isync; tlbre"); entry->mas1 = mfspr(SPR_MAS1); entry->mas2 = mfspr(SPR_MAS2); entry->mas3 = mfspr(SPR_MAS3); switch ((mfpvr() >> 16) & 0xFFFF) { case FSL_E500v2: case FSL_E500mc: case FSL_E5500: case FSL_E6500: entry->mas7 = mfspr(SPR_MAS7); break; default: entry->mas7 = 0; break; } __asm __volatile("wrtee %0" :: "r"(msr)); entry->virt = entry->mas2 & MAS2_EPN_MASK; entry->phys = ((vm_paddr_t)(entry->mas7 & MAS7_RPN) << 32) | (entry->mas3 & MAS3_RPN); entry->size = tsize2size((entry->mas1 & MAS1_TSIZE_MASK) >> MAS1_TSIZE_SHIFT); } struct tlbwrite_args { tlb_entry_t *e; unsigned int idx; }; static uint32_t tlb1_find_free(void) { tlb_entry_t e; int i; for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&e, i); if ((e.mas1 & MAS1_VALID) == 0) return (i); } return (-1); } static void tlb1_purge_va_range(vm_offset_t va, vm_size_t size) { tlb_entry_t e; int i; for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&e, i); if ((e.mas1 & MAS1_VALID) == 0) continue; if ((e.mas2 & MAS2_EPN_MASK) >= va && (e.mas2 & MAS2_EPN_MASK) < va + size) { mtspr(SPR_MAS1, e.mas1 & ~MAS1_VALID); __asm __volatile("isync; tlbwe; isync; msync"); } } } static void tlb1_write_entry_int(void *arg) { struct tlbwrite_args *args = arg; uint32_t idx, mas0; idx = args->idx; if (idx == -1) { tlb1_purge_va_range(args->e->virt, args->e->size); idx = tlb1_find_free(); if (idx == -1) panic("No free TLB1 entries!\n"); } /* Select entry */ mas0 = MAS0_TLBSEL(1) | MAS0_ESEL(idx); mtspr(SPR_MAS0, mas0); mtspr(SPR_MAS1, args->e->mas1); mtspr(SPR_MAS2, args->e->mas2); mtspr(SPR_MAS3, args->e->mas3); switch ((mfpvr() >> 16) & 0xFFFF) { case FSL_E500mc: case FSL_E5500: case FSL_E6500: mtspr(SPR_MAS8, 0); /* FALLTHROUGH */ case FSL_E500v2: mtspr(SPR_MAS7, args->e->mas7); break; default: break; } __asm __volatile("isync; tlbwe; isync; msync"); } static void tlb1_write_entry_sync(void *arg) { /* Empty synchronization point for smp_rendezvous(). */ } /* * Write given entry to TLB1 hardware. */ static void tlb1_write_entry(tlb_entry_t *e, unsigned int idx) { struct tlbwrite_args args; args.e = e; args.idx = idx; #ifdef SMP if ((e->mas2 & _TLB_ENTRY_SHARED) && smp_started) { mb(); smp_rendezvous(tlb1_write_entry_sync, tlb1_write_entry_int, tlb1_write_entry_sync, &args); } else #endif { register_t msr; msr = mfmsr(); __asm __volatile("wrteei 0"); tlb1_write_entry_int(&args); __asm __volatile("wrtee %0" :: "r"(msr)); } } /* * Convert TLB TSIZE value to mapped region size. */ static vm_size_t tsize2size(unsigned int tsize) { /* * size = 4^tsize KB * size = 4^tsize * 2^10 = 2^(2 * tsize - 10) */ return ((1 << (2 * tsize)) * 1024); } /* * Convert region size (must be power of 4) to TLB TSIZE value. */ static unsigned int size2tsize(vm_size_t size) { return (ilog2(size) / 2 - 5); } /* * Register permanent kernel mapping in TLB1. * * Entries are created starting from index 0 (current free entry is * kept in tlb1_idx) and are not supposed to be invalidated. */ int tlb1_set_entry(vm_offset_t va, vm_paddr_t pa, vm_size_t size, uint32_t flags) { tlb_entry_t e; uint32_t ts, tid; int tsize, index; /* First try to update an existing entry. */ for (index = 0; index < TLB1_ENTRIES; index++) { tlb1_read_entry(&e, index); /* Check if we're just updating the flags, and update them. */ if (e.phys == pa && e.virt == va && e.size == size) { e.mas2 = (va & MAS2_EPN_MASK) | flags; tlb1_write_entry(&e, index); return (0); } } /* Convert size to TSIZE */ tsize = size2tsize(size); tid = (TID_KERNEL << MAS1_TID_SHIFT) & MAS1_TID_MASK; /* XXX TS is hard coded to 0 for now as we only use single address space */ ts = (0 << MAS1_TS_SHIFT) & MAS1_TS_MASK; e.phys = pa; e.virt = va; e.size = size; e.mas1 = MAS1_VALID | MAS1_IPROT | ts | tid; e.mas1 |= ((tsize << MAS1_TSIZE_SHIFT) & MAS1_TSIZE_MASK); e.mas2 = (va & MAS2_EPN_MASK) | flags; /* Set supervisor RWX permission bits */ e.mas3 = (pa & MAS3_RPN) | MAS3_SR | MAS3_SW | MAS3_SX; e.mas7 = (pa >> 32) & MAS7_RPN; tlb1_write_entry(&e, -1); return (0); } /* * Map in contiguous RAM region into the TLB1. */ static vm_size_t tlb1_mapin_region(vm_offset_t va, vm_paddr_t pa, vm_size_t size, int wimge) { vm_offset_t base; vm_size_t mapped, sz, ssize; mapped = 0; base = va; ssize = size; while (size > 0) { sz = 1UL << (ilog2(size) & ~1); /* Align size to PA */ if (pa % sz != 0) { do { sz >>= 2; } while (pa % sz != 0); } /* Now align from there to VA */ if (va % sz != 0) { do { sz >>= 2; } while (va % sz != 0); } #ifdef __powerpc64__ /* * Clamp TLB1 entries to 4G. * * While the e6500 supports up to 1TB mappings, the e5500 * only supports up to 4G mappings. (0b1011) * * If any e6500 machines capable of supporting a very * large amount of memory appear in the future, we can * revisit this. * * For now, though, since we have plenty of space in TLB1, * always avoid creating entries larger than 4GB. */ sz = MIN(sz, 1UL << 32); #endif if (bootverbose) printf("Wiring VA=%p to PA=%jx (size=%lx)\n", (void *)va, (uintmax_t)pa, (long)sz); if (tlb1_set_entry(va, pa, sz, _TLB_ENTRY_SHARED | wimge) < 0) return (mapped); size -= sz; pa += sz; va += sz; } mapped = (va - base); if (bootverbose) printf("mapped size 0x%"PRIxPTR" (wasted space 0x%"PRIxPTR")\n", mapped, mapped - ssize); return (mapped); } /* * TLB1 initialization routine, to be called after the very first * assembler level setup done in locore.S. */ void tlb1_init() { vm_offset_t mas2; uint32_t mas0, mas1, mas3, mas7; uint32_t tsz; tlb1_get_tlbconf(); mas0 = MAS0_TLBSEL(1) | MAS0_ESEL(0); mtspr(SPR_MAS0, mas0); __asm __volatile("isync; tlbre"); mas1 = mfspr(SPR_MAS1); mas2 = mfspr(SPR_MAS2); mas3 = mfspr(SPR_MAS3); mas7 = mfspr(SPR_MAS7); kernload = ((vm_paddr_t)(mas7 & MAS7_RPN) << 32) | (mas3 & MAS3_RPN); tsz = (mas1 & MAS1_TSIZE_MASK) >> MAS1_TSIZE_SHIFT; kernsize += (tsz > 0) ? tsize2size(tsz) : 0; kernstart = trunc_page(mas2); /* Setup TLB miss defaults */ set_mas4_defaults(); } /* * pmap_early_io_unmap() should be used in short conjunction with * pmap_early_io_map(), as in the following snippet: * * x = pmap_early_io_map(...); * * pmap_early_io_unmap(x, size); * * And avoiding more allocations between. */ void pmap_early_io_unmap(vm_offset_t va, vm_size_t size) { int i; tlb_entry_t e; vm_size_t isize; size = roundup(size, PAGE_SIZE); isize = size; for (i = 0; i < TLB1_ENTRIES && size > 0; i++) { tlb1_read_entry(&e, i); if (!(e.mas1 & MAS1_VALID)) continue; if (va <= e.virt && (va + isize) >= (e.virt + e.size)) { size -= e.size; e.mas1 &= ~MAS1_VALID; tlb1_write_entry(&e, i); } } if (tlb1_map_base == va + isize) tlb1_map_base -= isize; } vm_offset_t pmap_early_io_map(vm_paddr_t pa, vm_size_t size) { vm_paddr_t pa_base; vm_offset_t va, sz; int i; tlb_entry_t e; KASSERT(!pmap_bootstrapped, ("Do not use after PMAP is up!")); for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&e, i); if (!(e.mas1 & MAS1_VALID)) continue; if (pa >= e.phys && (pa + size) <= (e.phys + e.size)) return (e.virt + (pa - e.phys)); } pa_base = rounddown(pa, PAGE_SIZE); size = roundup(size + (pa - pa_base), PAGE_SIZE); tlb1_map_base = roundup2(tlb1_map_base, 1 << (ilog2(size) & ~1)); va = tlb1_map_base + (pa - pa_base); do { sz = 1 << (ilog2(size) & ~1); tlb1_set_entry(tlb1_map_base, pa_base, sz, _TLB_ENTRY_SHARED | _TLB_ENTRY_IO); size -= sz; pa_base += sz; tlb1_map_base += sz; } while (size > 0); return (va); } void pmap_track_page(pmap_t pmap, vm_offset_t va) { vm_paddr_t pa; vm_page_t page; struct pv_entry *pve; va = trunc_page(va); pa = pmap_kextract(va); page = PHYS_TO_VM_PAGE(pa); rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); TAILQ_FOREACH(pve, &page->md.pv_list, pv_link) { if ((pmap == pve->pv_pmap) && (va == pve->pv_va)) { goto out; } } page->md.pv_tracked = true; pv_insert(pmap, va, page); out: PMAP_UNLOCK(pmap); rw_wunlock(&pvh_global_lock); } /* * Setup MAS4 defaults. * These values are loaded to MAS0-2 on a TLB miss. */ static void set_mas4_defaults(void) { uint32_t mas4; /* Defaults: TLB0, PID0, TSIZED=4K */ mas4 = MAS4_TLBSELD0; mas4 |= (TLB_SIZE_4K << MAS4_TSIZED_SHIFT) & MAS4_TSIZED_MASK; #ifdef SMP mas4 |= MAS4_MD; #endif mtspr(SPR_MAS4, mas4); __asm __volatile("isync"); } /* * Return 0 if the physical IO range is encompassed by one of the * the TLB1 entries, otherwise return related error code. */ static int tlb1_iomapped(int i, vm_paddr_t pa, vm_size_t size, vm_offset_t *va) { uint32_t prot; vm_paddr_t pa_start; vm_paddr_t pa_end; unsigned int entry_tsize; vm_size_t entry_size; tlb_entry_t e; *va = (vm_offset_t)NULL; tlb1_read_entry(&e, i); /* Skip invalid entries */ if (!(e.mas1 & MAS1_VALID)) return (EINVAL); /* * The entry must be cache-inhibited, guarded, and r/w * so it can function as an i/o page */ prot = e.mas2 & (MAS2_I | MAS2_G); if (prot != (MAS2_I | MAS2_G)) return (EPERM); prot = e.mas3 & (MAS3_SR | MAS3_SW); if (prot != (MAS3_SR | MAS3_SW)) return (EPERM); /* The address should be within the entry range. */ entry_tsize = (e.mas1 & MAS1_TSIZE_MASK) >> MAS1_TSIZE_SHIFT; KASSERT((entry_tsize), ("tlb1_iomapped: invalid entry tsize")); entry_size = tsize2size(entry_tsize); pa_start = (((vm_paddr_t)e.mas7 & MAS7_RPN) << 32) | (e.mas3 & MAS3_RPN); pa_end = pa_start + entry_size; if ((pa < pa_start) || ((pa + size) > pa_end)) return (ERANGE); /* Return virtual address of this mapping. */ *va = (e.mas2 & MAS2_EPN_MASK) + (pa - pa_start); return (0); } #ifdef DDB /* Print out contents of the MAS registers for each TLB0 entry */ static void #ifdef __powerpc64__ tlb_print_entry(int i, uint32_t mas1, uint64_t mas2, uint32_t mas3, #else tlb_print_entry(int i, uint32_t mas1, uint32_t mas2, uint32_t mas3, #endif uint32_t mas7) { int as; char desc[3]; tlbtid_t tid; vm_size_t size; unsigned int tsize; desc[2] = '\0'; if (mas1 & MAS1_VALID) desc[0] = 'V'; else desc[0] = ' '; if (mas1 & MAS1_IPROT) desc[1] = 'P'; else desc[1] = ' '; as = (mas1 & MAS1_TS_MASK) ? 1 : 0; tid = MAS1_GETTID(mas1); tsize = (mas1 & MAS1_TSIZE_MASK) >> MAS1_TSIZE_SHIFT; size = 0; if (tsize) size = tsize2size(tsize); printf("%3d: (%s) [AS=%d] " "sz = 0x%jx tsz = %d tid = %d mas1 = 0x%08x " "mas2(va) = 0x%"PRI0ptrX" mas3(pa) = 0x%08x mas7 = 0x%08x\n", i, desc, as, (uintmax_t)size, tsize, tid, mas1, mas2, mas3, mas7); } DB_SHOW_COMMAND(tlb0, tlb0_print_tlbentries) { uint32_t mas0, mas1, mas3, mas7; #ifdef __powerpc64__ uint64_t mas2; #else uint32_t mas2; #endif int entryidx, way, idx; printf("TLB0 entries:\n"); for (way = 0; way < TLB0_WAYS; way ++) for (entryidx = 0; entryidx < TLB0_ENTRIES_PER_WAY; entryidx++) { mas0 = MAS0_TLBSEL(0) | MAS0_ESEL(way); mtspr(SPR_MAS0, mas0); mas2 = entryidx << MAS2_TLB0_ENTRY_IDX_SHIFT; mtspr(SPR_MAS2, mas2); __asm __volatile("isync; tlbre"); mas1 = mfspr(SPR_MAS1); mas2 = mfspr(SPR_MAS2); mas3 = mfspr(SPR_MAS3); mas7 = mfspr(SPR_MAS7); idx = tlb0_tableidx(mas2, way); tlb_print_entry(idx, mas1, mas2, mas3, mas7); } } /* * Print out contents of the MAS registers for each TLB1 entry */ DB_SHOW_COMMAND(tlb1, tlb1_print_tlbentries) { uint32_t mas0, mas1, mas3, mas7; #ifdef __powerpc64__ uint64_t mas2; #else uint32_t mas2; #endif int i; printf("TLB1 entries:\n"); for (i = 0; i < TLB1_ENTRIES; i++) { mas0 = MAS0_TLBSEL(1) | MAS0_ESEL(i); mtspr(SPR_MAS0, mas0); __asm __volatile("isync; tlbre"); mas1 = mfspr(SPR_MAS1); mas2 = mfspr(SPR_MAS2); mas3 = mfspr(SPR_MAS3); mas7 = mfspr(SPR_MAS7); tlb_print_entry(i, mas1, mas2, mas3, mas7); } } #endif diff --git a/sys/vm/vm_fault.c b/sys/vm/vm_fault.c index 9b80b4188c40..6445f7af59a1 100644 --- a/sys/vm/vm_fault.c +++ b/sys/vm/vm_fault.c @@ -1,2175 +1,2179 @@ /*- * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Page fault handling module. */ #include __FBSDID("$FreeBSD$"); #include "opt_ktrace.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef KTRACE #include #endif #include #include #include #include #include #include #include #include #include #include #include #define PFBAK 4 #define PFFOR 4 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT) #define VM_FAULT_DONTNEED_MIN 1048576 struct faultstate { /* Fault parameters. */ vm_offset_t vaddr; vm_page_t *m_hold; vm_prot_t fault_type; vm_prot_t prot; int fault_flags; boolean_t wired; /* Control state. */ struct timeval oom_start_time; bool oom_started; int nera; /* Page reference for cow. */ vm_page_t m_cow; /* Current object. */ vm_object_t object; vm_pindex_t pindex; vm_page_t m; /* Top-level map object. */ vm_object_t first_object; vm_pindex_t first_pindex; vm_page_t first_m; /* Map state. */ vm_map_t map; vm_map_entry_t entry; int map_generation; bool lookup_still_valid; /* Vnode if locked. */ struct vnode *vp; }; /* * Return codes for internal fault routines. */ enum fault_status { FAULT_SUCCESS = 1, /* Return success to user. */ FAULT_FAILURE, /* Return failure to user. */ FAULT_CONTINUE, /* Continue faulting. */ FAULT_RESTART, /* Restart fault. */ FAULT_OUT_OF_BOUNDS, /* Invalid address for pager. */ FAULT_HARD, /* Performed I/O. */ FAULT_SOFT, /* Found valid page. */ FAULT_PROTECTION_FAILURE, /* Invalid access. */ }; static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead); static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, int backward, int forward, bool obj_locked); static int vm_pfault_oom_attempts = 3; SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN, &vm_pfault_oom_attempts, 0, "Number of page allocation attempts in page fault handler before it " "triggers OOM handling"); static int vm_pfault_oom_wait = 10; SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN, &vm_pfault_oom_wait, 0, "Number of seconds to wait for free pages before retrying " "the page fault handler"); static inline void fault_page_release(vm_page_t *mp) { vm_page_t m; m = *mp; if (m != NULL) { /* * We are likely to loop around again and attempt to busy * this page. Deactivating it leaves it available for * pageout while optimizing fault restarts. */ vm_page_deactivate(m); vm_page_xunbusy(m); *mp = NULL; } } static inline void fault_page_free(vm_page_t *mp) { vm_page_t m; m = *mp; if (m != NULL) { VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_wired(m)) vm_page_free(m); else vm_page_xunbusy(m); *mp = NULL; } } static inline void unlock_map(struct faultstate *fs) { if (fs->lookup_still_valid) { vm_map_lookup_done(fs->map, fs->entry); fs->lookup_still_valid = false; } } static void unlock_vp(struct faultstate *fs) { if (fs->vp != NULL) { vput(fs->vp); fs->vp = NULL; } } static void fault_deallocate(struct faultstate *fs) { fault_page_release(&fs->m_cow); fault_page_release(&fs->m); vm_object_pip_wakeup(fs->object); if (fs->object != fs->first_object) { VM_OBJECT_WLOCK(fs->first_object); fault_page_free(&fs->first_m); VM_OBJECT_WUNLOCK(fs->first_object); vm_object_pip_wakeup(fs->first_object); } vm_object_deallocate(fs->first_object); unlock_map(fs); unlock_vp(fs); } static void unlock_and_deallocate(struct faultstate *fs) { VM_OBJECT_WUNLOCK(fs->object); fault_deallocate(fs); } static void vm_fault_dirty(struct faultstate *fs, vm_page_t m) { bool need_dirty; if (((fs->prot & VM_PROT_WRITE) == 0 && (fs->fault_flags & VM_FAULT_DIRTY) == 0) || (m->oflags & VPO_UNMANAGED) != 0) return; VM_PAGE_OBJECT_BUSY_ASSERT(m); need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 && (fs->fault_flags & VM_FAULT_WIRE) == 0) || (fs->fault_flags & VM_FAULT_DIRTY) != 0; vm_object_set_writeable_dirty(m->object); /* * If the fault is a write, we know that this page is being * written NOW so dirty it explicitly to save on * pmap_is_modified() calls later. * * Also, since the page is now dirty, we can possibly tell * the pager to release any swap backing the page. */ if (need_dirty && vm_page_set_dirty(m) == 0) { /* * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC * if the page is already dirty to prevent data written with * the expectation of being synced from not being synced. * Likewise if this entry does not request NOSYNC then make * sure the page isn't marked NOSYNC. Applications sharing * data should use the same flags to avoid ping ponging. */ if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0) vm_page_aflag_set(m, PGA_NOSYNC); else vm_page_aflag_clear(m, PGA_NOSYNC); } } /* * Unlocks fs.first_object and fs.map on success. */ static enum fault_status vm_fault_soft_fast(struct faultstate *fs) { vm_page_t m, m_map; #if VM_NRESERVLEVEL > 0 vm_page_t m_super; int flags; #endif int psind; vm_offset_t vaddr; enum fault_status res; MPASS(fs->vp == NULL); res = FAULT_SUCCESS; vaddr = fs->vaddr; vm_object_busy(fs->first_object); m = vm_page_lookup(fs->first_object, fs->first_pindex); /* A busy page can be mapped for read|execute access. */ if (m == NULL || ((fs->prot & VM_PROT_WRITE) != 0 && vm_page_busied(m)) || !vm_page_all_valid(m)) { res = FAULT_FAILURE; goto out; } m_map = m; psind = 0; #if VM_NRESERVLEVEL > 0 if ((m->flags & PG_FICTITIOUS) == 0 && (m_super = vm_reserv_to_superpage(m)) != NULL && rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start && roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end && (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) & (pagesizes[m_super->psind] - 1)) && !fs->wired && pmap_ps_enabled(fs->map->pmap)) { flags = PS_ALL_VALID; if ((fs->prot & VM_PROT_WRITE) != 0) { /* * Create a superpage mapping allowing write access * only if none of the constituent pages are busy and * all of them are already dirty (except possibly for * the page that was faulted on). */ flags |= PS_NONE_BUSY; if ((fs->first_object->flags & OBJ_UNMANAGED) == 0) flags |= PS_ALL_DIRTY; } if (vm_page_ps_test(m_super, flags, m)) { m_map = m_super; psind = m_super->psind; vaddr = rounddown2(vaddr, pagesizes[psind]); /* Preset the modified bit for dirty superpages. */ if ((flags & PS_ALL_DIRTY) != 0) fs->fault_type |= VM_PROT_WRITE; } } #endif if (pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type | PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind) != KERN_SUCCESS) { res = FAULT_FAILURE; goto out; } if (fs->m_hold != NULL) { (*fs->m_hold) = m; vm_page_wire(m); } if (psind == 0 && !fs->wired) vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true); VM_OBJECT_RUNLOCK(fs->first_object); vm_fault_dirty(fs, m); vm_map_lookup_done(fs->map, fs->entry); curthread->td_ru.ru_minflt++; out: vm_object_unbusy(fs->first_object); return (res); } static void vm_fault_restore_map_lock(struct faultstate *fs) { VM_OBJECT_ASSERT_WLOCKED(fs->first_object); MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0); if (!vm_map_trylock_read(fs->map)) { VM_OBJECT_WUNLOCK(fs->first_object); vm_map_lock_read(fs->map); VM_OBJECT_WLOCK(fs->first_object); } fs->lookup_still_valid = true; } static void vm_fault_populate_check_page(vm_page_t m) { /* * Check each page to ensure that the pager is obeying the * interface: the page must be installed in the object, fully * valid, and exclusively busied. */ MPASS(m != NULL); MPASS(vm_page_all_valid(m)); MPASS(vm_page_xbusied(m)); } static void vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first, vm_pindex_t last) { vm_page_t m; vm_pindex_t pidx; VM_OBJECT_ASSERT_WLOCKED(object); MPASS(first <= last); for (pidx = first, m = vm_page_lookup(object, pidx); pidx <= last; pidx++, m = vm_page_next(m)) { vm_fault_populate_check_page(m); vm_page_deactivate(m); vm_page_xunbusy(m); } } static enum fault_status vm_fault_populate(struct faultstate *fs) { vm_offset_t vaddr; vm_page_t m; vm_pindex_t map_first, map_last, pager_first, pager_last, pidx; int bdry_idx, i, npages, psind, rv; enum fault_status res; MPASS(fs->object == fs->first_object); VM_OBJECT_ASSERT_WLOCKED(fs->first_object); MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0); MPASS(fs->first_object->backing_object == NULL); MPASS(fs->lookup_still_valid); pager_first = OFF_TO_IDX(fs->entry->offset); pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1; unlock_map(fs); unlock_vp(fs); res = FAULT_SUCCESS; /* * Call the pager (driver) populate() method. * * There is no guarantee that the method will be called again * if the current fault is for read, and a future fault is * for write. Report the entry's maximum allowed protection * to the driver. */ rv = vm_pager_populate(fs->first_object, fs->first_pindex, fs->fault_type, fs->entry->max_protection, &pager_first, &pager_last); VM_OBJECT_ASSERT_WLOCKED(fs->first_object); if (rv == VM_PAGER_BAD) { /* * VM_PAGER_BAD is the backdoor for a pager to request * normal fault handling. */ vm_fault_restore_map_lock(fs); if (fs->map->timestamp != fs->map_generation) return (FAULT_RESTART); return (FAULT_CONTINUE); } if (rv != VM_PAGER_OK) return (FAULT_FAILURE); /* AKA SIGSEGV */ /* Ensure that the driver is obeying the interface. */ MPASS(pager_first <= pager_last); MPASS(fs->first_pindex <= pager_last); MPASS(fs->first_pindex >= pager_first); MPASS(pager_last < fs->first_object->size); vm_fault_restore_map_lock(fs); bdry_idx = (fs->entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; if (fs->map->timestamp != fs->map_generation) { if (bdry_idx == 0) { vm_fault_populate_cleanup(fs->first_object, pager_first, pager_last); } else { m = vm_page_lookup(fs->first_object, pager_first); if (m != fs->m) vm_page_xunbusy(m); } return (FAULT_RESTART); } /* * The map is unchanged after our last unlock. Process the fault. * * First, the special case of largepage mappings, where * populate only busies the first page in superpage run. */ if (bdry_idx != 0) { KASSERT(PMAP_HAS_LARGEPAGES, ("missing pmap support for large pages")); m = vm_page_lookup(fs->first_object, pager_first); vm_fault_populate_check_page(m); VM_OBJECT_WUNLOCK(fs->first_object); vaddr = fs->entry->start + IDX_TO_OFF(pager_first) - fs->entry->offset; /* assert alignment for entry */ KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0, ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx", (uintmax_t)fs->entry->start, (uintmax_t)pager_first, (uintmax_t)fs->entry->offset, (uintmax_t)vaddr)); KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0, ("unaligned superpage m %p %#jx", m, (uintmax_t)VM_PAGE_TO_PHYS(m))); rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) | PMAP_ENTER_LARGEPAGE, bdry_idx); VM_OBJECT_WLOCK(fs->first_object); vm_page_xunbusy(m); if (rv != KERN_SUCCESS) { res = FAULT_FAILURE; goto out; } if ((fs->fault_flags & VM_FAULT_WIRE) != 0) { for (i = 0; i < atop(pagesizes[bdry_idx]); i++) vm_page_wire(m + i); } if (fs->m_hold != NULL) { *fs->m_hold = m + (fs->first_pindex - pager_first); vm_page_wire(*fs->m_hold); } goto out; } /* * The range [pager_first, pager_last] that is given to the * pager is only a hint. The pager may populate any range * within the object that includes the requested page index. * In case the pager expanded the range, clip it to fit into * the map entry. */ map_first = OFF_TO_IDX(fs->entry->offset); if (map_first > pager_first) { vm_fault_populate_cleanup(fs->first_object, pager_first, map_first - 1); pager_first = map_first; } map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1; if (map_last < pager_last) { vm_fault_populate_cleanup(fs->first_object, map_last + 1, pager_last); pager_last = map_last; } for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx); pidx <= pager_last; pidx += npages, m = vm_page_next(&m[npages - 1])) { vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset; -#if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ - __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv) || \ - defined(__powerpc64__) + psind = m->psind; if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 || pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last || !pmap_ps_enabled(fs->map->pmap) || fs->wired)) psind = 0; -#else - psind = 0; -#endif + npages = atop(pagesizes[psind]); for (i = 0; i < npages; i++) { vm_fault_populate_check_page(&m[i]); vm_fault_dirty(fs, &m[i]); } VM_OBJECT_WUNLOCK(fs->first_object); rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0), psind); -#if defined(__amd64__) - if (psind > 0 && rv == KERN_FAILURE) { + + /* + * pmap_enter() may fail for a superpage mapping if additional + * protection policies prevent the full mapping. + * For example, this will happen on amd64 if the entire + * address range does not share the same userspace protection + * key. Revert to single-page mappings if this happens. + */ + MPASS(rv == KERN_SUCCESS || + (psind > 0 && rv == KERN_PROTECTION_FAILURE)); + if (__predict_false(psind > 0 && + rv == KERN_PROTECTION_FAILURE)) { for (i = 0; i < npages; i++) { rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i), &m[i], fs->prot, fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0), 0); MPASS(rv == KERN_SUCCESS); } } -#else - MPASS(rv == KERN_SUCCESS); -#endif + VM_OBJECT_WLOCK(fs->first_object); for (i = 0; i < npages; i++) { if ((fs->fault_flags & VM_FAULT_WIRE) != 0) vm_page_wire(&m[i]); else vm_page_activate(&m[i]); if (fs->m_hold != NULL && m[i].pindex == fs->first_pindex) { (*fs->m_hold) = &m[i]; vm_page_wire(&m[i]); } vm_page_xunbusy(&m[i]); } } out: curthread->td_ru.ru_majflt++; return (res); } static int prot_fault_translation; SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN, &prot_fault_translation, 0, "Control signal to deliver on protection fault"); /* compat definition to keep common code for signal translation */ #define UCODE_PAGEFLT 12 #ifdef T_PAGEFLT _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT"); #endif /* * vm_fault_trap: * * Handle a page fault occurring at the given address, * requiring the given permissions, in the map specified. * If successful, the page is inserted into the * associated physical map. * * NOTE: the given address should be truncated to the * proper page address. * * KERN_SUCCESS is returned if the page fault is handled; otherwise, * a standard error specifying why the fault is fatal is returned. * * The map in question must be referenced, and remains so. * Caller may hold no locks. */ int vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags, int *signo, int *ucode) { int result; MPASS(signo == NULL || ucode != NULL); #ifdef KTRACE if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT)) ktrfault(vaddr, fault_type); #endif result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags, NULL); KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE || result == KERN_INVALID_ADDRESS || result == KERN_RESOURCE_SHORTAGE || result == KERN_PROTECTION_FAILURE || result == KERN_OUT_OF_BOUNDS, ("Unexpected Mach error %d from vm_fault()", result)); #ifdef KTRACE if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND)) ktrfaultend(result); #endif if (result != KERN_SUCCESS && signo != NULL) { switch (result) { case KERN_FAILURE: case KERN_INVALID_ADDRESS: *signo = SIGSEGV; *ucode = SEGV_MAPERR; break; case KERN_RESOURCE_SHORTAGE: *signo = SIGBUS; *ucode = BUS_OOMERR; break; case KERN_OUT_OF_BOUNDS: *signo = SIGBUS; *ucode = BUS_OBJERR; break; case KERN_PROTECTION_FAILURE: if (prot_fault_translation == 0) { /* * Autodetect. This check also covers * the images without the ABI-tag ELF * note. */ if (SV_CURPROC_ABI() == SV_ABI_FREEBSD && curproc->p_osrel >= P_OSREL_SIGSEGV) { *signo = SIGSEGV; *ucode = SEGV_ACCERR; } else { *signo = SIGBUS; *ucode = UCODE_PAGEFLT; } } else if (prot_fault_translation == 1) { /* Always compat mode. */ *signo = SIGBUS; *ucode = UCODE_PAGEFLT; } else { /* Always SIGSEGV mode. */ *signo = SIGSEGV; *ucode = SEGV_ACCERR; } break; default: KASSERT(0, ("Unexpected Mach error %d from vm_fault()", result)); break; } } return (result); } static enum fault_status vm_fault_lock_vnode(struct faultstate *fs, bool objlocked) { struct vnode *vp; int error, locked; if (fs->object->type != OBJT_VNODE) return (FAULT_CONTINUE); vp = fs->object->handle; if (vp == fs->vp) { ASSERT_VOP_LOCKED(vp, "saved vnode is not locked"); return (FAULT_CONTINUE); } /* * Perform an unlock in case the desired vnode changed while * the map was unlocked during a retry. */ unlock_vp(fs); locked = VOP_ISLOCKED(vp); if (locked != LK_EXCLUSIVE) locked = LK_SHARED; /* * We must not sleep acquiring the vnode lock while we have * the page exclusive busied or the object's * paging-in-progress count incremented. Otherwise, we could * deadlock. */ error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT); if (error == 0) { fs->vp = vp; return (FAULT_CONTINUE); } vhold(vp); if (objlocked) unlock_and_deallocate(fs); else fault_deallocate(fs); error = vget(vp, locked | LK_RETRY | LK_CANRECURSE); vdrop(vp); fs->vp = vp; KASSERT(error == 0, ("vm_fault: vget failed %d", error)); return (FAULT_RESTART); } /* * Calculate the desired readahead. Handle drop-behind. * * Returns the number of readahead blocks to pass to the pager. */ static int vm_fault_readahead(struct faultstate *fs) { int era, nera; u_char behavior; KASSERT(fs->lookup_still_valid, ("map unlocked")); era = fs->entry->read_ahead; behavior = vm_map_entry_behavior(fs->entry); if (behavior == MAP_ENTRY_BEHAV_RANDOM) { nera = 0; } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { nera = VM_FAULT_READ_AHEAD_MAX; if (fs->vaddr == fs->entry->next_read) vm_fault_dontneed(fs, fs->vaddr, nera); } else if (fs->vaddr == fs->entry->next_read) { /* * This is a sequential fault. Arithmetically * increase the requested number of pages in * the read-ahead window. The requested * number of pages is "# of sequential faults * x (read ahead min + 1) + read ahead min" */ nera = VM_FAULT_READ_AHEAD_MIN; if (era > 0) { nera += era + 1; if (nera > VM_FAULT_READ_AHEAD_MAX) nera = VM_FAULT_READ_AHEAD_MAX; } if (era == VM_FAULT_READ_AHEAD_MAX) vm_fault_dontneed(fs, fs->vaddr, nera); } else { /* * This is a non-sequential fault. */ nera = 0; } if (era != nera) { /* * A read lock on the map suffices to update * the read ahead count safely. */ fs->entry->read_ahead = nera; } return (nera); } static int vm_fault_lookup(struct faultstate *fs) { int result; KASSERT(!fs->lookup_still_valid, ("vm_fault_lookup: Map already locked.")); result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type | VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object, &fs->first_pindex, &fs->prot, &fs->wired); if (result != KERN_SUCCESS) { unlock_vp(fs); return (result); } fs->map_generation = fs->map->timestamp; if (fs->entry->eflags & MAP_ENTRY_NOFAULT) { panic("%s: fault on nofault entry, addr: %#lx", __func__, (u_long)fs->vaddr); } if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION && fs->entry->wiring_thread != curthread) { vm_map_unlock_read(fs->map); vm_map_lock(fs->map); if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) && (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) { unlock_vp(fs); fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; vm_map_unlock_and_wait(fs->map, 0); } else vm_map_unlock(fs->map); return (KERN_RESOURCE_SHORTAGE); } MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0); if (fs->wired) fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY); else KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0, ("!fs->wired && VM_FAULT_WIRE")); fs->lookup_still_valid = true; return (KERN_SUCCESS); } static int vm_fault_relookup(struct faultstate *fs) { vm_object_t retry_object; vm_pindex_t retry_pindex; vm_prot_t retry_prot; int result; if (!vm_map_trylock_read(fs->map)) return (KERN_RESTART); fs->lookup_still_valid = true; if (fs->map->timestamp == fs->map_generation) return (KERN_SUCCESS); result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type, &fs->entry, &retry_object, &retry_pindex, &retry_prot, &fs->wired); if (result != KERN_SUCCESS) { /* * If retry of map lookup would have blocked then * retry fault from start. */ if (result == KERN_FAILURE) return (KERN_RESTART); return (result); } if (retry_object != fs->first_object || retry_pindex != fs->first_pindex) return (KERN_RESTART); /* * Check whether the protection has changed or the object has * been copied while we left the map unlocked. Changing from * read to write permission is OK - we leave the page * write-protected, and catch the write fault. Changing from * write to read permission means that we can't mark the page * write-enabled after all. */ fs->prot &= retry_prot; fs->fault_type &= retry_prot; if (fs->prot == 0) return (KERN_RESTART); /* Reassert because wired may have changed. */ KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0, ("!wired && VM_FAULT_WIRE")); return (KERN_SUCCESS); } static void vm_fault_cow(struct faultstate *fs) { bool is_first_object_locked; KASSERT(fs->object != fs->first_object, ("source and target COW objects are identical")); /* * This allows pages to be virtually copied from a backing_object * into the first_object, where the backing object has no other * refs to it, and cannot gain any more refs. Instead of a bcopy, * we just move the page from the backing object to the first * object. Note that we must mark the page dirty in the first * object so that it will go out to swap when needed. */ is_first_object_locked = false; if ( /* * Only one shadow object and no other refs. */ fs->object->shadow_count == 1 && fs->object->ref_count == 1 && /* * No other ways to look the object up */ fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 && /* * We don't chase down the shadow chain and we can acquire locks. */ (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) && fs->object == fs->first_object->backing_object && VM_OBJECT_TRYWLOCK(fs->object)) { /* * Remove but keep xbusy for replace. fs->m is moved into * fs->first_object and left busy while fs->first_m is * conditionally freed. */ vm_page_remove_xbusy(fs->m); vm_page_replace(fs->m, fs->first_object, fs->first_pindex, fs->first_m); vm_page_dirty(fs->m); #if VM_NRESERVLEVEL > 0 /* * Rename the reservation. */ vm_reserv_rename(fs->m, fs->first_object, fs->object, OFF_TO_IDX(fs->first_object->backing_object_offset)); #endif VM_OBJECT_WUNLOCK(fs->object); VM_OBJECT_WUNLOCK(fs->first_object); fs->first_m = fs->m; fs->m = NULL; VM_CNT_INC(v_cow_optim); } else { if (is_first_object_locked) VM_OBJECT_WUNLOCK(fs->first_object); /* * Oh, well, lets copy it. */ pmap_copy_page(fs->m, fs->first_m); vm_page_valid(fs->first_m); if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) { vm_page_wire(fs->first_m); vm_page_unwire(fs->m, PQ_INACTIVE); } /* * Save the cow page to be released after * pmap_enter is complete. */ fs->m_cow = fs->m; fs->m = NULL; /* * Typically, the shadow object is either private to this * address space (OBJ_ONEMAPPING) or its pages are read only. * In the highly unusual case where the pages of a shadow object * are read/write shared between this and other address spaces, * we need to ensure that any pmap-level mappings to the * original, copy-on-write page from the backing object are * removed from those other address spaces. * * The flag check is racy, but this is tolerable: if * OBJ_ONEMAPPING is cleared after the check, the busy state * ensures that new mappings of m_cow can't be created. * pmap_enter() will replace an existing mapping in the current * address space. If OBJ_ONEMAPPING is set after the check, * removing mappings will at worse trigger some unnecessary page * faults. */ vm_page_assert_xbusied(fs->m_cow); if ((fs->first_object->flags & OBJ_ONEMAPPING) == 0) pmap_remove_all(fs->m_cow); } vm_object_pip_wakeup(fs->object); /* * Only use the new page below... */ fs->object = fs->first_object; fs->pindex = fs->first_pindex; fs->m = fs->first_m; VM_CNT_INC(v_cow_faults); curthread->td_cow++; } static bool vm_fault_next(struct faultstate *fs) { vm_object_t next_object; /* * The requested page does not exist at this object/ * offset. Remove the invalid page from the object, * waking up anyone waiting for it, and continue on to * the next object. However, if this is the top-level * object, we must leave the busy page in place to * prevent another process from rushing past us, and * inserting the page in that object at the same time * that we are. */ if (fs->object == fs->first_object) { fs->first_m = fs->m; fs->m = NULL; } else fault_page_free(&fs->m); /* * Move on to the next object. Lock the next object before * unlocking the current one. */ VM_OBJECT_ASSERT_WLOCKED(fs->object); next_object = fs->object->backing_object; if (next_object == NULL) return (false); MPASS(fs->first_m != NULL); KASSERT(fs->object != next_object, ("object loop %p", next_object)); VM_OBJECT_WLOCK(next_object); vm_object_pip_add(next_object, 1); if (fs->object != fs->first_object) vm_object_pip_wakeup(fs->object); fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset); VM_OBJECT_WUNLOCK(fs->object); fs->object = next_object; return (true); } static void vm_fault_zerofill(struct faultstate *fs) { /* * If there's no object left, fill the page in the top * object with zeros. */ if (fs->object != fs->first_object) { vm_object_pip_wakeup(fs->object); fs->object = fs->first_object; fs->pindex = fs->first_pindex; } MPASS(fs->first_m != NULL); MPASS(fs->m == NULL); fs->m = fs->first_m; fs->first_m = NULL; /* * Zero the page if necessary and mark it valid. */ if ((fs->m->flags & PG_ZERO) == 0) { pmap_zero_page(fs->m); } else { VM_CNT_INC(v_ozfod); } VM_CNT_INC(v_zfod); vm_page_valid(fs->m); } /* * Initiate page fault after timeout. Returns true if caller should * do vm_waitpfault() after the call. */ static bool vm_fault_allocate_oom(struct faultstate *fs) { struct timeval now; unlock_and_deallocate(fs); if (vm_pfault_oom_attempts < 0) return (true); if (!fs->oom_started) { fs->oom_started = true; getmicrotime(&fs->oom_start_time); return (true); } getmicrotime(&now); timevalsub(&now, &fs->oom_start_time); if (now.tv_sec < vm_pfault_oom_attempts * vm_pfault_oom_wait) return (true); if (bootverbose) printf( "proc %d (%s) failed to alloc page on fault, starting OOM\n", curproc->p_pid, curproc->p_comm); vm_pageout_oom(VM_OOM_MEM_PF); fs->oom_started = false; return (false); } /* * Allocate a page directly or via the object populate method. */ static enum fault_status vm_fault_allocate(struct faultstate *fs) { struct domainset *dset; int alloc_req; enum fault_status res; if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) { res = vm_fault_lock_vnode(fs, true); MPASS(res == FAULT_CONTINUE || res == FAULT_RESTART); if (res == FAULT_RESTART) return (res); } if (fs->pindex >= fs->object->size) { unlock_and_deallocate(fs); return (FAULT_OUT_OF_BOUNDS); } if (fs->object == fs->first_object && (fs->first_object->flags & OBJ_POPULATE) != 0 && fs->first_object->shadow_count == 0) { res = vm_fault_populate(fs); switch (res) { case FAULT_SUCCESS: case FAULT_FAILURE: case FAULT_RESTART: unlock_and_deallocate(fs); return (res); case FAULT_CONTINUE: /* * Pager's populate() method * returned VM_PAGER_BAD. */ break; default: panic("inconsistent return codes"); } } /* * Allocate a new page for this object/offset pair. * * Unlocked read of the p_flag is harmless. At worst, the P_KILLED * might be not observed there, and allocation can fail, causing * restart and new reading of the p_flag. */ dset = fs->object->domain.dr_policy; if (dset == NULL) dset = curthread->td_domain.dr_policy; if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) { #if VM_NRESERVLEVEL > 0 vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex); #endif alloc_req = P_KILLED(curproc) ? VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; if (fs->object->type != OBJT_VNODE && fs->object->backing_object == NULL) alloc_req |= VM_ALLOC_ZERO; fs->m = vm_page_alloc(fs->object, fs->pindex, alloc_req); } if (fs->m == NULL) { if (vm_fault_allocate_oom(fs)) vm_waitpfault(dset, vm_pfault_oom_wait * hz); return (FAULT_RESTART); } fs->oom_started = false; return (FAULT_CONTINUE); } /* * Call the pager to retrieve the page if there is a chance * that the pager has it, and potentially retrieve additional * pages at the same time. */ static enum fault_status vm_fault_getpages(struct faultstate *fs, int *behindp, int *aheadp) { vm_offset_t e_end, e_start; int ahead, behind, cluster_offset, rv; enum fault_status status; u_char behavior; /* * Prepare for unlocking the map. Save the map * entry's start and end addresses, which are used to * optimize the size of the pager operation below. * Even if the map entry's addresses change after * unlocking the map, using the saved addresses is * safe. */ e_start = fs->entry->start; e_end = fs->entry->end; behavior = vm_map_entry_behavior(fs->entry); /* * If the pager for the current object might have * the page, then determine the number of additional * pages to read and potentially reprioritize * previously read pages for earlier reclamation. * These operations should only be performed once per * page fault. Even if the current pager doesn't * have the page, the number of additional pages to * read will apply to subsequent objects in the * shadow chain. */ if (fs->nera == -1 && !P_KILLED(curproc)) fs->nera = vm_fault_readahead(fs); /* * Release the map lock before locking the vnode or * sleeping in the pager. (If the current object has * a shadow, then an earlier iteration of this loop * may have already unlocked the map.) */ unlock_map(fs); status = vm_fault_lock_vnode(fs, false); MPASS(status == FAULT_CONTINUE || status == FAULT_RESTART); if (status == FAULT_RESTART) return (status); KASSERT(fs->vp == NULL || !fs->map->system_map, ("vm_fault: vnode-backed object mapped by system map")); /* * Page in the requested page and hint the pager, * that it may bring up surrounding pages. */ if (fs->nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM || P_KILLED(curproc)) { behind = 0; ahead = 0; } else { /* Is this a sequential fault? */ if (fs->nera > 0) { behind = 0; ahead = fs->nera; } else { /* * Request a cluster of pages that is * aligned to a VM_FAULT_READ_DEFAULT * page offset boundary within the * object. Alignment to a page offset * boundary is more likely to coincide * with the underlying file system * block than alignment to a virtual * address boundary. */ cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT; behind = ulmin(cluster_offset, atop(fs->vaddr - e_start)); ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset; } ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1); } *behindp = behind; *aheadp = ahead; rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp); if (rv == VM_PAGER_OK) return (FAULT_HARD); if (rv == VM_PAGER_ERROR) printf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm); /* * If an I/O error occurred or the requested page was * outside the range of the pager, clean up and return * an error. */ if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) { VM_OBJECT_WLOCK(fs->object); fault_page_free(&fs->m); unlock_and_deallocate(fs); return (FAULT_OUT_OF_BOUNDS); } KASSERT(rv == VM_PAGER_FAIL, ("%s: unepxected pager error %d", __func__, rv)); return (FAULT_CONTINUE); } /* * Wait/Retry if the page is busy. We have to do this if the page is * either exclusive or shared busy because the vm_pager may be using * read busy for pageouts (and even pageins if it is the vnode pager), * and we could end up trying to pagein and pageout the same page * simultaneously. * * We can theoretically allow the busy case on a read fault if the page * is marked valid, but since such pages are typically already pmap'd, * putting that special case in might be more effort then it is worth. * We cannot under any circumstances mess around with a shared busied * page except, perhaps, to pmap it. */ static void vm_fault_busy_sleep(struct faultstate *fs) { /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ vm_page_aflag_set(fs->m, PGA_REFERENCED); if (fs->object != fs->first_object) { fault_page_release(&fs->first_m); vm_object_pip_wakeup(fs->first_object); } vm_object_pip_wakeup(fs->object); unlock_map(fs); if (fs->m != vm_page_lookup(fs->object, fs->pindex) || !vm_page_busy_sleep(fs->m, "vmpfw", 0)) VM_OBJECT_WUNLOCK(fs->object); VM_CNT_INC(v_intrans); vm_object_deallocate(fs->first_object); } /* * Handle page lookup, populate, allocate, page-in for the current * object. * * The object is locked on entry and will remain locked with a return * code of FAULT_CONTINUE so that fault may follow the shadow chain. * Otherwise, the object will be unlocked upon return. */ static enum fault_status vm_fault_object(struct faultstate *fs, int *behindp, int *aheadp) { enum fault_status res; bool dead; /* * If the object is marked for imminent termination, we retry * here, since the collapse pass has raced with us. Otherwise, * if we see terminally dead object, return fail. */ if ((fs->object->flags & OBJ_DEAD) != 0) { dead = fs->object->type == OBJT_DEAD; unlock_and_deallocate(fs); if (dead) return (FAULT_PROTECTION_FAILURE); pause("vmf_de", 1); return (FAULT_RESTART); } /* * See if the page is resident. */ fs->m = vm_page_lookup(fs->object, fs->pindex); if (fs->m != NULL) { if (!vm_page_tryxbusy(fs->m)) { vm_fault_busy_sleep(fs); return (FAULT_RESTART); } /* * The page is marked busy for other processes and the * pagedaemon. If it is still completely valid we are * done. */ if (vm_page_all_valid(fs->m)) { VM_OBJECT_WUNLOCK(fs->object); return (FAULT_SOFT); } } VM_OBJECT_ASSERT_WLOCKED(fs->object); /* * Page is not resident. If the pager might contain the page * or this is the beginning of the search, allocate a new * page. (Default objects are zero-fill, so there is no real * pager for them.) */ if (fs->m == NULL && (fs->object->type != OBJT_DEFAULT || fs->object == fs->first_object)) { res = vm_fault_allocate(fs); if (res != FAULT_CONTINUE) return (res); } /* * Default objects have no pager so no exclusive busy exists * to protect this page in the chain. Skip to the next * object without dropping the lock to preserve atomicity of * shadow faults. */ if (fs->object->type != OBJT_DEFAULT) { /* * At this point, we have either allocated a new page * or found an existing page that is only partially * valid. * * We hold a reference on the current object and the * page is exclusive busied. The exclusive busy * prevents simultaneous faults and collapses while * the object lock is dropped. */ VM_OBJECT_WUNLOCK(fs->object); res = vm_fault_getpages(fs, behindp, aheadp); if (res == FAULT_CONTINUE) VM_OBJECT_WLOCK(fs->object); } else { res = FAULT_CONTINUE; } return (res); } int vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags, vm_page_t *m_hold) { struct faultstate fs; int ahead, behind, faultcount, rv; enum fault_status res; bool hardfault; VM_CNT_INC(v_vm_faults); if ((curthread->td_pflags & TDP_NOFAULTING) != 0) return (KERN_PROTECTION_FAILURE); fs.vp = NULL; fs.vaddr = vaddr; fs.m_hold = m_hold; fs.fault_flags = fault_flags; fs.map = map; fs.lookup_still_valid = false; fs.oom_started = false; fs.nera = -1; faultcount = 0; hardfault = false; RetryFault: fs.fault_type = fault_type; /* * Find the backing store object and offset into it to begin the * search. */ rv = vm_fault_lookup(&fs); if (rv != KERN_SUCCESS) { if (rv == KERN_RESOURCE_SHORTAGE) goto RetryFault; return (rv); } /* * Try to avoid lock contention on the top-level object through * special-case handling of some types of page faults, specifically, * those that are mapping an existing page from the top-level object. * Under this condition, a read lock on the object suffices, allowing * multiple page faults of a similar type to run in parallel. */ if (fs.vp == NULL /* avoid locked vnode leak */ && (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 && (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) { VM_OBJECT_RLOCK(fs.first_object); res = vm_fault_soft_fast(&fs); if (res == FAULT_SUCCESS) return (KERN_SUCCESS); if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) { VM_OBJECT_RUNLOCK(fs.first_object); VM_OBJECT_WLOCK(fs.first_object); } } else { VM_OBJECT_WLOCK(fs.first_object); } /* * Make a reference to this object to prevent its disposal while we * are messing with it. Once we have the reference, the map is free * to be diddled. Since objects reference their shadows (and copies), * they will stay around as well. * * Bump the paging-in-progress count to prevent size changes (e.g. * truncation operations) during I/O. */ vm_object_reference_locked(fs.first_object); vm_object_pip_add(fs.first_object, 1); fs.m_cow = fs.m = fs.first_m = NULL; /* * Search for the page at object/offset. */ fs.object = fs.first_object; fs.pindex = fs.first_pindex; if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) { res = vm_fault_allocate(&fs); switch (res) { case FAULT_RESTART: goto RetryFault; case FAULT_SUCCESS: return (KERN_SUCCESS); case FAULT_FAILURE: return (KERN_FAILURE); case FAULT_OUT_OF_BOUNDS: return (KERN_OUT_OF_BOUNDS); case FAULT_CONTINUE: break; default: panic("vm_fault: Unhandled status %d", res); } } while (TRUE) { KASSERT(fs.m == NULL, ("page still set %p at loop start", fs.m)); res = vm_fault_object(&fs, &behind, &ahead); switch (res) { case FAULT_SOFT: goto found; case FAULT_HARD: faultcount = behind + 1 + ahead; hardfault = true; goto found; case FAULT_RESTART: goto RetryFault; case FAULT_SUCCESS: return (KERN_SUCCESS); case FAULT_FAILURE: return (KERN_FAILURE); case FAULT_OUT_OF_BOUNDS: return (KERN_OUT_OF_BOUNDS); case FAULT_PROTECTION_FAILURE: return (KERN_PROTECTION_FAILURE); case FAULT_CONTINUE: break; default: panic("vm_fault: Unhandled status %d", res); } /* * The page was not found in the current object. Try to * traverse into a backing object or zero fill if none is * found. */ if (vm_fault_next(&fs)) continue; if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) { if (fs.first_object == fs.object) fault_page_free(&fs.first_m); unlock_and_deallocate(&fs); return (KERN_OUT_OF_BOUNDS); } VM_OBJECT_WUNLOCK(fs.object); vm_fault_zerofill(&fs); /* Don't try to prefault neighboring pages. */ faultcount = 1; break; } found: /* * A valid page has been found and exclusively busied. The * object lock must no longer be held. */ vm_page_assert_xbusied(fs.m); VM_OBJECT_ASSERT_UNLOCKED(fs.object); /* * If the page is being written, but isn't already owned by the * top-level object, we have to copy it into a new page owned by the * top-level object. */ if (fs.object != fs.first_object) { /* * We only really need to copy if we want to write it. */ if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { vm_fault_cow(&fs); /* * We only try to prefault read-only mappings to the * neighboring pages when this copy-on-write fault is * a hard fault. In other cases, trying to prefault * is typically wasted effort. */ if (faultcount == 0) faultcount = 1; } else { fs.prot &= ~VM_PROT_WRITE; } } /* * We must verify that the maps have not changed since our last * lookup. */ if (!fs.lookup_still_valid) { rv = vm_fault_relookup(&fs); if (rv != KERN_SUCCESS) { fault_deallocate(&fs); if (rv == KERN_RESTART) goto RetryFault; return (rv); } } VM_OBJECT_ASSERT_UNLOCKED(fs.object); /* * If the page was filled by a pager, save the virtual address that * should be faulted on next under a sequential access pattern to the * map entry. A read lock on the map suffices to update this address * safely. */ if (hardfault) fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE; /* * Page must be completely valid or it is not fit to * map into user space. vm_pager_get_pages() ensures this. */ vm_page_assert_xbusied(fs.m); KASSERT(vm_page_all_valid(fs.m), ("vm_fault: page %p partially invalid", fs.m)); vm_fault_dirty(&fs, fs.m); /* * Put this page into the physical map. We had to do the unlock above * because pmap_enter() may sleep. We don't put the page * back on the active queue until later so that the pageout daemon * won't find it (yet). */ pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0); if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 && fs.wired == 0) vm_fault_prefault(&fs, vaddr, faultcount > 0 ? behind : PFBAK, faultcount > 0 ? ahead : PFFOR, false); /* * If the page is not wired down, then put it where the pageout daemon * can find it. */ if ((fs.fault_flags & VM_FAULT_WIRE) != 0) vm_page_wire(fs.m); else vm_page_activate(fs.m); if (fs.m_hold != NULL) { (*fs.m_hold) = fs.m; vm_page_wire(fs.m); } vm_page_xunbusy(fs.m); fs.m = NULL; /* * Unlock everything, and return */ fault_deallocate(&fs); if (hardfault) { VM_CNT_INC(v_io_faults); curthread->td_ru.ru_majflt++; #ifdef RACCT if (racct_enable && fs.object->type == OBJT_VNODE) { PROC_LOCK(curproc); if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { racct_add_force(curproc, RACCT_WRITEBPS, PAGE_SIZE + behind * PAGE_SIZE); racct_add_force(curproc, RACCT_WRITEIOPS, 1); } else { racct_add_force(curproc, RACCT_READBPS, PAGE_SIZE + ahead * PAGE_SIZE); racct_add_force(curproc, RACCT_READIOPS, 1); } PROC_UNLOCK(curproc); } #endif } else curthread->td_ru.ru_minflt++; return (KERN_SUCCESS); } /* * Speed up the reclamation of pages that precede the faulting pindex within * the first object of the shadow chain. Essentially, perform the equivalent * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes * the faulting pindex by the cluster size when the pages read by vm_fault() * cross a cluster-size boundary. The cluster size is the greater of the * smallest superpage size and VM_FAULT_DONTNEED_MIN. * * When "fs->first_object" is a shadow object, the pages in the backing object * that precede the faulting pindex are deactivated by vm_fault(). So, this * function must only be concerned with pages in the first object. */ static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead) { vm_map_entry_t entry; vm_object_t first_object, object; vm_offset_t end, start; vm_page_t m, m_next; vm_pindex_t pend, pstart; vm_size_t size; object = fs->object; VM_OBJECT_ASSERT_UNLOCKED(object); first_object = fs->first_object; /* Neither fictitious nor unmanaged pages can be reclaimed. */ if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { VM_OBJECT_RLOCK(first_object); size = VM_FAULT_DONTNEED_MIN; if (MAXPAGESIZES > 1 && size < pagesizes[1]) size = pagesizes[1]; end = rounddown2(vaddr, size); if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) && (entry = fs->entry)->start < end) { if (end - entry->start < size) start = entry->start; else start = end - size; pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED); pstart = OFF_TO_IDX(entry->offset) + atop(start - entry->start); m_next = vm_page_find_least(first_object, pstart); pend = OFF_TO_IDX(entry->offset) + atop(end - entry->start); while ((m = m_next) != NULL && m->pindex < pend) { m_next = TAILQ_NEXT(m, listq); if (!vm_page_all_valid(m) || vm_page_busied(m)) continue; /* * Don't clear PGA_REFERENCED, since it would * likely represent a reference by a different * process. * * Typically, at this point, prefetched pages * are still in the inactive queue. Only * pages that triggered page faults are in the * active queue. The test for whether the page * is in the inactive queue is racy; in the * worst case we will requeue the page * unnecessarily. */ if (!vm_page_inactive(m)) vm_page_deactivate(m); } } VM_OBJECT_RUNLOCK(first_object); } } /* * vm_fault_prefault provides a quick way of clustering * pagefaults into a processes address space. It is a "cousin" * of vm_map_pmap_enter, except it runs at page fault time instead * of mmap time. */ static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, int backward, int forward, bool obj_locked) { pmap_t pmap; vm_map_entry_t entry; vm_object_t backing_object, lobject; vm_offset_t addr, starta; vm_pindex_t pindex; vm_page_t m; int i; pmap = fs->map->pmap; if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) return; entry = fs->entry; if (addra < backward * PAGE_SIZE) { starta = entry->start; } else { starta = addra - backward * PAGE_SIZE; if (starta < entry->start) starta = entry->start; } /* * Generate the sequence of virtual addresses that are candidates for * prefaulting in an outward spiral from the faulting virtual address, * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ... * If the candidate address doesn't have a backing physical page, then * the loop immediately terminates. */ for (i = 0; i < 2 * imax(backward, forward); i++) { addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE : PAGE_SIZE); if (addr > addra + forward * PAGE_SIZE) addr = 0; if (addr < starta || addr >= entry->end) continue; if (!pmap_is_prefaultable(pmap, addr)) continue; pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; lobject = entry->object.vm_object; if (!obj_locked) VM_OBJECT_RLOCK(lobject); while ((m = vm_page_lookup(lobject, pindex)) == NULL && lobject->type == OBJT_DEFAULT && (backing_object = lobject->backing_object) != NULL) { KASSERT((lobject->backing_object_offset & PAGE_MASK) == 0, ("vm_fault_prefault: unaligned object offset")); pindex += lobject->backing_object_offset >> PAGE_SHIFT; VM_OBJECT_RLOCK(backing_object); if (!obj_locked || lobject != entry->object.vm_object) VM_OBJECT_RUNLOCK(lobject); lobject = backing_object; } if (m == NULL) { if (!obj_locked || lobject != entry->object.vm_object) VM_OBJECT_RUNLOCK(lobject); break; } if (vm_page_all_valid(m) && (m->flags & PG_FICTITIOUS) == 0) pmap_enter_quick(pmap, addr, m, entry->protection); if (!obj_locked || lobject != entry->object.vm_object) VM_OBJECT_RUNLOCK(lobject); } } /* * Hold each of the physical pages that are mapped by the specified range of * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid * and allow the specified types of access, "prot". If all of the implied * pages are successfully held, then the number of held pages is returned * together with pointers to those pages in the array "ma". However, if any * of the pages cannot be held, -1 is returned. */ int vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, vm_prot_t prot, vm_page_t *ma, int max_count) { vm_offset_t end, va; vm_page_t *mp; int count; boolean_t pmap_failed; if (len == 0) return (0); end = round_page(addr + len); addr = trunc_page(addr); if (!vm_map_range_valid(map, addr, end)) return (-1); if (atop(end - addr) > max_count) panic("vm_fault_quick_hold_pages: count > max_count"); count = atop(end - addr); /* * Most likely, the physical pages are resident in the pmap, so it is * faster to try pmap_extract_and_hold() first. */ pmap_failed = FALSE; for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { *mp = pmap_extract_and_hold(map->pmap, va, prot); if (*mp == NULL) pmap_failed = TRUE; else if ((prot & VM_PROT_WRITE) != 0 && (*mp)->dirty != VM_PAGE_BITS_ALL) { /* * Explicitly dirty the physical page. Otherwise, the * caller's changes may go unnoticed because they are * performed through an unmanaged mapping or by a DMA * operation. * * The object lock is not held here. * See vm_page_clear_dirty_mask(). */ vm_page_dirty(*mp); } } if (pmap_failed) { /* * One or more pages could not be held by the pmap. Either no * page was mapped at the specified virtual address or that * mapping had insufficient permissions. Attempt to fault in * and hold these pages. * * If vm_fault_disable_pagefaults() was called, * i.e., TDP_NOFAULTING is set, we must not sleep nor * acquire MD VM locks, which means we must not call * vm_fault(). Some (out of tree) callers mark * too wide a code area with vm_fault_disable_pagefaults() * already, use the VM_PROT_QUICK_NOFAULT flag to request * the proper behaviour explicitly. */ if ((prot & VM_PROT_QUICK_NOFAULT) != 0 && (curthread->td_pflags & TDP_NOFAULTING) != 0) goto error; for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) if (*mp == NULL && vm_fault(map, va, prot, VM_FAULT_NORMAL, mp) != KERN_SUCCESS) goto error; } return (count); error: for (mp = ma; mp < ma + count; mp++) if (*mp != NULL) vm_page_unwire(*mp, PQ_INACTIVE); return (-1); } /* * Routine: * vm_fault_copy_entry * Function: * Create new shadow object backing dst_entry with private copy of * all underlying pages. When src_entry is equal to dst_entry, * function implements COW for wired-down map entry. Otherwise, * it forks wired entry into dst_map. * * In/out conditions: * The source and destination maps must be locked for write. * The source map entry must be wired down (or be a sharing map * entry corresponding to a main map entry that is wired down). */ void vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, vm_map_entry_t dst_entry, vm_map_entry_t src_entry, vm_ooffset_t *fork_charge) { vm_object_t backing_object, dst_object, object, src_object; vm_pindex_t dst_pindex, pindex, src_pindex; vm_prot_t access, prot; vm_offset_t vaddr; vm_page_t dst_m; vm_page_t src_m; boolean_t upgrade; #ifdef lint src_map++; #endif /* lint */ upgrade = src_entry == dst_entry; access = prot = dst_entry->protection; src_object = src_entry->object.vm_object; src_pindex = OFF_TO_IDX(src_entry->offset); if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { dst_object = src_object; vm_object_reference(dst_object); } else { /* * Create the top-level object for the destination entry. * Doesn't actually shadow anything - we copy the pages * directly. */ dst_object = vm_object_allocate_anon(atop(dst_entry->end - dst_entry->start), NULL, NULL, 0); #if VM_NRESERVLEVEL > 0 dst_object->flags |= OBJ_COLORED; dst_object->pg_color = atop(dst_entry->start); #endif dst_object->domain = src_object->domain; dst_object->charge = dst_entry->end - dst_entry->start; } VM_OBJECT_WLOCK(dst_object); KASSERT(upgrade || dst_entry->object.vm_object == NULL, ("vm_fault_copy_entry: vm_object not NULL")); if (src_object != dst_object) { dst_entry->object.vm_object = dst_object; dst_entry->offset = 0; dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC; } if (fork_charge != NULL) { KASSERT(dst_entry->cred == NULL, ("vm_fault_copy_entry: leaked swp charge")); dst_object->cred = curthread->td_ucred; crhold(dst_object->cred); *fork_charge += dst_object->charge; } else if ((dst_object->type == OBJT_DEFAULT || (dst_object->flags & OBJ_SWAP) != 0) && dst_object->cred == NULL) { KASSERT(dst_entry->cred != NULL, ("no cred for entry %p", dst_entry)); dst_object->cred = dst_entry->cred; dst_entry->cred = NULL; } /* * If not an upgrade, then enter the mappings in the pmap as * read and/or execute accesses. Otherwise, enter them as * write accesses. * * A writeable large page mapping is only created if all of * the constituent small page mappings are modified. Marking * PTEs as modified on inception allows promotion to happen * without taking potentially large number of soft faults. */ if (!upgrade) access &= ~VM_PROT_WRITE; /* * Loop through all of the virtual pages within the entry's * range, copying each page from the source object to the * destination object. Since the source is wired, those pages * must exist. In contrast, the destination is pageable. * Since the destination object doesn't share any backing storage * with the source object, all of its pages must be dirtied, * regardless of whether they can be written. */ for (vaddr = dst_entry->start, dst_pindex = 0; vaddr < dst_entry->end; vaddr += PAGE_SIZE, dst_pindex++) { again: /* * Find the page in the source object, and copy it in. * Because the source is wired down, the page will be * in memory. */ if (src_object != dst_object) VM_OBJECT_RLOCK(src_object); object = src_object; pindex = src_pindex + dst_pindex; while ((src_m = vm_page_lookup(object, pindex)) == NULL && (backing_object = object->backing_object) != NULL) { /* * Unless the source mapping is read-only or * it is presently being upgraded from * read-only, the first object in the shadow * chain should provide all of the pages. In * other words, this loop body should never be * executed when the source mapping is already * read/write. */ KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 || upgrade, ("vm_fault_copy_entry: main object missing page")); VM_OBJECT_RLOCK(backing_object); pindex += OFF_TO_IDX(object->backing_object_offset); if (object != dst_object) VM_OBJECT_RUNLOCK(object); object = backing_object; } KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing")); if (object != dst_object) { /* * Allocate a page in the destination object. */ dst_m = vm_page_alloc(dst_object, (src_object == dst_object ? src_pindex : 0) + dst_pindex, VM_ALLOC_NORMAL); if (dst_m == NULL) { VM_OBJECT_WUNLOCK(dst_object); VM_OBJECT_RUNLOCK(object); vm_wait(dst_object); VM_OBJECT_WLOCK(dst_object); goto again; } pmap_copy_page(src_m, dst_m); VM_OBJECT_RUNLOCK(object); dst_m->dirty = dst_m->valid = src_m->valid; } else { dst_m = src_m; if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0) goto again; if (dst_m->pindex >= dst_object->size) { /* * We are upgrading. Index can occur * out of bounds if the object type is * vnode and the file was truncated. */ vm_page_xunbusy(dst_m); break; } } VM_OBJECT_WUNLOCK(dst_object); /* * Enter it in the pmap. If a wired, copy-on-write * mapping is being replaced by a write-enabled * mapping, then wire that new mapping. * * The page can be invalid if the user called * msync(MS_INVALIDATE) or truncated the backing vnode * or shared memory object. In this case, do not * insert it into pmap, but still do the copy so that * all copies of the wired map entry have similar * backing pages. */ if (vm_page_all_valid(dst_m)) { pmap_enter(dst_map->pmap, vaddr, dst_m, prot, access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); } /* * Mark it no longer busy, and put it on the active list. */ VM_OBJECT_WLOCK(dst_object); if (upgrade) { if (src_m != dst_m) { vm_page_unwire(src_m, PQ_INACTIVE); vm_page_wire(dst_m); } else { KASSERT(vm_page_wired(dst_m), ("dst_m %p is not wired", dst_m)); } } else { vm_page_activate(dst_m); } vm_page_xunbusy(dst_m); } VM_OBJECT_WUNLOCK(dst_object); if (upgrade) { dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); vm_object_deallocate(src_object); } } /* * Block entry into the machine-independent layer's page fault handler by * the calling thread. Subsequent calls to vm_fault() by that thread will * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of * spurious page faults. */ int vm_fault_disable_pagefaults(void) { return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); } void vm_fault_enable_pagefaults(int save) { curthread_pflags_restore(save); }