Index: head/sys/kern/kern_cpuset.c =================================================================== --- head/sys/kern/kern_cpuset.c (revision 339924) +++ head/sys/kern/kern_cpuset.c (revision 339925) @@ -1,2285 +1,2295 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2008, Jeffrey Roberson * All rights reserved. * * Copyright (c) 2008 Nokia Corporation * 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 unmodified, 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. * */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.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 #ifdef DDB #include #endif /* DDB */ /* * cpusets provide a mechanism for creating and manipulating sets of * processors for the purpose of constraining the scheduling of threads to * specific processors. * * Each process belongs to an identified set, by default this is set 1. Each * thread may further restrict the cpus it may run on to a subset of this * named set. This creates an anonymous set which other threads and processes * may not join by number. * * The named set is referred to herein as the 'base' set to avoid ambiguity. * This set is usually a child of a 'root' set while the anonymous set may * simply be referred to as a mask. In the syscall api these are referred to * as the ROOT, CPUSET, and MASK levels where CPUSET is called 'base' here. * * Threads inherit their set from their creator whether it be anonymous or * not. This means that anonymous sets are immutable because they may be * shared. To modify an anonymous set a new set is created with the desired * mask and the same parent as the existing anonymous set. This gives the * illusion of each thread having a private mask. * * Via the syscall apis a user may ask to retrieve or modify the root, base, * or mask that is discovered via a pid, tid, or setid. Modifying a set * modifies all numbered and anonymous child sets to comply with the new mask. * Modifying a pid or tid's mask applies only to that tid but must still * exist within the assigned parent set. * * A thread may not be assigned to a group separate from other threads in * the process. This is to remove ambiguity when the setid is queried with * a pid argument. There is no other technical limitation. * * This somewhat complex arrangement is intended to make it easy for * applications to query available processors and bind their threads to * specific processors while also allowing administrators to dynamically * reprovision by changing sets which apply to groups of processes. * * A simple application should not concern itself with sets at all and * rather apply masks to its own threads via CPU_WHICH_TID and a -1 id * meaning 'curthread'. It may query available cpus for that tid with a * getaffinity call using (CPU_LEVEL_CPUSET, CPU_WHICH_PID, -1, ...). */ LIST_HEAD(domainlist, domainset); struct domainset __read_mostly domainset_prefer[MAXMEMDOM]; struct domainset __read_mostly domainset_roundrobin; static uma_zone_t cpuset_zone; static uma_zone_t domainset_zone; static struct mtx cpuset_lock; static struct setlist cpuset_ids; static struct domainlist cpuset_domains; static struct unrhdr *cpuset_unr; static struct cpuset *cpuset_zero, *cpuset_default, *cpuset_kernel; static struct domainset domainset0, domainset2; /* Return the size of cpuset_t at the kernel level */ SYSCTL_INT(_kern_sched, OID_AUTO, cpusetsize, CTLFLAG_RD | CTLFLAG_CAPRD, SYSCTL_NULL_INT_PTR, sizeof(cpuset_t), "sizeof(cpuset_t)"); cpuset_t *cpuset_root; cpuset_t cpuset_domain[MAXMEMDOM]; static int domainset_valid(const struct domainset *, const struct domainset *); /* * Find the first non-anonymous set starting from 'set'. */ static struct cpuset * cpuset_getbase(struct cpuset *set) { if (set->cs_id == CPUSET_INVALID) set = set->cs_parent; return (set); } /* * Walks up the tree from 'set' to find the root. */ static struct cpuset * cpuset_getroot(struct cpuset *set) { while ((set->cs_flags & CPU_SET_ROOT) == 0 && set->cs_parent != NULL) set = set->cs_parent; return (set); } /* * Acquire a reference to a cpuset, all pointers must be tracked with refs. */ struct cpuset * cpuset_ref(struct cpuset *set) { refcount_acquire(&set->cs_ref); return (set); } /* * Walks up the tree from 'set' to find the root. Returns the root * referenced. */ static struct cpuset * cpuset_refroot(struct cpuset *set) { return (cpuset_ref(cpuset_getroot(set))); } /* * Find the first non-anonymous set starting from 'set'. Returns this set * referenced. May return the passed in set with an extra ref if it is * not anonymous. */ static struct cpuset * cpuset_refbase(struct cpuset *set) { return (cpuset_ref(cpuset_getbase(set))); } /* * Release a reference in a context where it is safe to allocate. */ void cpuset_rel(struct cpuset *set) { cpusetid_t id; if (refcount_release(&set->cs_ref) == 0) return; mtx_lock_spin(&cpuset_lock); LIST_REMOVE(set, cs_siblings); id = set->cs_id; if (id != CPUSET_INVALID) LIST_REMOVE(set, cs_link); mtx_unlock_spin(&cpuset_lock); cpuset_rel(set->cs_parent); uma_zfree(cpuset_zone, set); if (id != CPUSET_INVALID) free_unr(cpuset_unr, id); } /* * Deferred release must be used when in a context that is not safe to * allocate/free. This places any unreferenced sets on the list 'head'. */ static void cpuset_rel_defer(struct setlist *head, struct cpuset *set) { if (refcount_release(&set->cs_ref) == 0) return; mtx_lock_spin(&cpuset_lock); LIST_REMOVE(set, cs_siblings); if (set->cs_id != CPUSET_INVALID) LIST_REMOVE(set, cs_link); LIST_INSERT_HEAD(head, set, cs_link); mtx_unlock_spin(&cpuset_lock); } /* * Complete a deferred release. Removes the set from the list provided to * cpuset_rel_defer. */ static void cpuset_rel_complete(struct cpuset *set) { LIST_REMOVE(set, cs_link); cpuset_rel(set->cs_parent); uma_zfree(cpuset_zone, set); } /* * Find a set based on an id. Returns it with a ref. */ static struct cpuset * cpuset_lookup(cpusetid_t setid, struct thread *td) { struct cpuset *set; if (setid == CPUSET_INVALID) return (NULL); mtx_lock_spin(&cpuset_lock); LIST_FOREACH(set, &cpuset_ids, cs_link) if (set->cs_id == setid) break; if (set) cpuset_ref(set); mtx_unlock_spin(&cpuset_lock); KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__)); if (set != NULL && jailed(td->td_ucred)) { struct cpuset *jset, *tset; jset = td->td_ucred->cr_prison->pr_cpuset; for (tset = set; tset != NULL; tset = tset->cs_parent) if (tset == jset) break; if (tset == NULL) { cpuset_rel(set); set = NULL; } } return (set); } /* * Create a set in the space provided in 'set' with the provided parameters. * The set is returned with a single ref. May return EDEADLK if the set * will have no valid cpu based on restrictions from the parent. */ static int _cpuset_create(struct cpuset *set, struct cpuset *parent, const cpuset_t *mask, struct domainset *domain, cpusetid_t id) { if (domain == NULL) domain = parent->cs_domain; if (mask == NULL) mask = &parent->cs_mask; if (!CPU_OVERLAP(&parent->cs_mask, mask)) return (EDEADLK); /* The domain must be prepared ahead of time. */ if (!domainset_valid(parent->cs_domain, domain)) return (EDEADLK); CPU_COPY(mask, &set->cs_mask); LIST_INIT(&set->cs_children); refcount_init(&set->cs_ref, 1); set->cs_flags = 0; mtx_lock_spin(&cpuset_lock); set->cs_domain = domain; CPU_AND(&set->cs_mask, &parent->cs_mask); set->cs_id = id; set->cs_parent = cpuset_ref(parent); LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings); if (set->cs_id != CPUSET_INVALID) LIST_INSERT_HEAD(&cpuset_ids, set, cs_link); mtx_unlock_spin(&cpuset_lock); return (0); } /* * Create a new non-anonymous set with the requested parent and mask. May * return failures if the mask is invalid or a new number can not be * allocated. */ static int cpuset_create(struct cpuset **setp, struct cpuset *parent, const cpuset_t *mask) { struct cpuset *set; cpusetid_t id; int error; id = alloc_unr(cpuset_unr); if (id == -1) return (ENFILE); *setp = set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO); error = _cpuset_create(set, parent, mask, NULL, id); if (error == 0) return (0); free_unr(cpuset_unr, id); uma_zfree(cpuset_zone, set); return (error); } static void cpuset_freelist_add(struct setlist *list, int count) { struct cpuset *set; int i; for (i = 0; i < count; i++) { set = uma_zalloc(cpuset_zone, M_ZERO | M_WAITOK); LIST_INSERT_HEAD(list, set, cs_link); } } static void cpuset_freelist_init(struct setlist *list, int count) { LIST_INIT(list); cpuset_freelist_add(list, count); } static void cpuset_freelist_free(struct setlist *list) { struct cpuset *set; while ((set = LIST_FIRST(list)) != NULL) { LIST_REMOVE(set, cs_link); uma_zfree(cpuset_zone, set); } } static void domainset_freelist_add(struct domainlist *list, int count) { struct domainset *set; int i; for (i = 0; i < count; i++) { set = uma_zalloc(domainset_zone, M_ZERO | M_WAITOK); LIST_INSERT_HEAD(list, set, ds_link); } } static void domainset_freelist_init(struct domainlist *list, int count) { LIST_INIT(list); domainset_freelist_add(list, count); } static void domainset_freelist_free(struct domainlist *list) { struct domainset *set; while ((set = LIST_FIRST(list)) != NULL) { LIST_REMOVE(set, ds_link); uma_zfree(domainset_zone, set); } } /* Copy a domainset preserving mask and policy. */ static void domainset_copy(const struct domainset *from, struct domainset *to) { DOMAINSET_COPY(&from->ds_mask, &to->ds_mask); to->ds_policy = from->ds_policy; to->ds_prefer = from->ds_prefer; } /* Return 1 if mask and policy are equal, otherwise 0. */ static int domainset_equal(const struct domainset *one, const struct domainset *two) { return (DOMAINSET_CMP(&one->ds_mask, &two->ds_mask) == 0 && one->ds_policy == two->ds_policy && one->ds_prefer == two->ds_prefer); } /* Return 1 if child is a valid subset of parent. */ static int domainset_valid(const struct domainset *parent, const struct domainset *child) { if (child->ds_policy != DOMAINSET_POLICY_PREFER) return (DOMAINSET_SUBSET(&parent->ds_mask, &child->ds_mask)); return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask)); } static int domainset_restrict(const struct domainset *parent, const struct domainset *child) { if (child->ds_policy != DOMAINSET_POLICY_PREFER) return (DOMAINSET_OVERLAP(&parent->ds_mask, &child->ds_mask)); return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask)); } /* * Lookup or create a domainset. The key is provided in ds_mask and * ds_policy. If the domainset does not yet exist the storage in * 'domain' is used to insert. Otherwise this storage is freed to the * domainset_zone and the existing domainset is returned. */ static struct domainset * _domainset_create(struct domainset *domain, struct domainlist *freelist) { struct domainset *ndomain; int i, j, max; KASSERT(domain->ds_cnt <= vm_ndomains, ("invalid domain count in domainset %p", domain)); KASSERT(domain->ds_policy != DOMAINSET_POLICY_PREFER || domain->ds_prefer < vm_ndomains, ("invalid preferred domain in domains %p", domain)); mtx_lock_spin(&cpuset_lock); LIST_FOREACH(ndomain, &cpuset_domains, ds_link) if (domainset_equal(ndomain, domain)) break; /* * If the domain does not yet exist we insert it and initialize * various iteration helpers which are not part of the key. */ if (ndomain == NULL) { LIST_INSERT_HEAD(&cpuset_domains, domain, ds_link); domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask); max = DOMAINSET_FLS(&domain->ds_mask) + 1; for (i = 0, j = 0; i < max; i++) if (DOMAINSET_ISSET(i, &domain->ds_mask)) domain->ds_order[j++] = i; } mtx_unlock_spin(&cpuset_lock); if (ndomain == NULL) return (domain); if (freelist != NULL) LIST_INSERT_HEAD(freelist, domain, ds_link); else uma_zfree(domainset_zone, domain); return (ndomain); } /* - * Are any of the domains in the mask empty? If so, silently - * remove them. If only empty domains are present, we must - * return failure. + * Are any of the domains in the mask empty? If so, silently + * remove them and update the domainset accordingly. If only empty + * domains are present, we must return failure. */ static bool domainset_empty_vm(struct domainset *domain) { - int i, max; + int i, j, max; max = DOMAINSET_FLS(&domain->ds_mask) + 1; - for (i = 0; i < max; i++) { - if (DOMAINSET_ISSET(i, &domain->ds_mask) && - VM_DOMAIN_EMPTY(i)) + for (i = 0; i < max; i++) + if (DOMAINSET_ISSET(i, &domain->ds_mask) && VM_DOMAIN_EMPTY(i)) DOMAINSET_CLR(i, &domain->ds_mask); + domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask); + max = DOMAINSET_FLS(&domain->ds_mask) + 1; + for (i = j = 0; i < max; i++) { + if (DOMAINSET_ISSET(i, &domain->ds_mask)) + domain->ds_order[j++] = i; + else if (domain->ds_policy == DOMAINSET_POLICY_PREFER && + domain->ds_prefer == i && domain->ds_cnt > 1) { + domain->ds_policy = DOMAINSET_POLICY_ROUNDROBIN; + domain->ds_prefer = -1; + } } return (DOMAINSET_EMPTY(&domain->ds_mask)); } /* * Create or lookup a domainset based on the key held in 'domain'. */ struct domainset * domainset_create(const struct domainset *domain) { struct domainset *ndomain; /* * Validate the policy. It must specify a useable policy number with * only valid domains. Preferred must include the preferred domain * in the mask. */ if (domain->ds_policy <= DOMAINSET_POLICY_INVALID || domain->ds_policy > DOMAINSET_POLICY_MAX) return (NULL); if (domain->ds_policy == DOMAINSET_POLICY_PREFER && !DOMAINSET_ISSET(domain->ds_prefer, &domain->ds_mask)) return (NULL); if (!DOMAINSET_SUBSET(&domainset0.ds_mask, &domain->ds_mask)) return (NULL); ndomain = uma_zalloc(domainset_zone, M_WAITOK | M_ZERO); domainset_copy(domain, ndomain); return _domainset_create(ndomain, NULL); } /* * Update thread domainset pointers. */ static void domainset_notify(void) { struct thread *td; struct proc *p; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); if (p->p_state == PRS_NEW) { PROC_UNLOCK(p); continue; } FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); td->td_domain.dr_policy = td->td_cpuset->cs_domain; thread_unlock(td); } PROC_UNLOCK(p); } sx_sunlock(&allproc_lock); kernel_object->domain.dr_policy = cpuset_kernel->cs_domain; } /* * Create a new set that is a subset of a parent. */ static struct domainset * domainset_shadow(const struct domainset *pdomain, const struct domainset *domain, struct domainlist *freelist) { struct domainset *ndomain; ndomain = LIST_FIRST(freelist); LIST_REMOVE(ndomain, ds_link); /* * Initialize the key from the request. */ domainset_copy(domain, ndomain); /* * Restrict the key by the parent. */ DOMAINSET_AND(&ndomain->ds_mask, &pdomain->ds_mask); return _domainset_create(ndomain, freelist); } /* * Recursively check for errors that would occur from applying mask to * the tree of sets starting at 'set'. Checks for sets that would become * empty as well as RDONLY flags. */ static int cpuset_testupdate(struct cpuset *set, cpuset_t *mask, int check_mask) { struct cpuset *nset; cpuset_t newmask; int error; mtx_assert(&cpuset_lock, MA_OWNED); if (set->cs_flags & CPU_SET_RDONLY) return (EPERM); if (check_mask) { if (!CPU_OVERLAP(&set->cs_mask, mask)) return (EDEADLK); CPU_COPY(&set->cs_mask, &newmask); CPU_AND(&newmask, mask); } else CPU_COPY(mask, &newmask); error = 0; LIST_FOREACH(nset, &set->cs_children, cs_siblings) if ((error = cpuset_testupdate(nset, &newmask, 1)) != 0) break; return (error); } /* * Applies the mask 'mask' without checking for empty sets or permissions. */ static void cpuset_update(struct cpuset *set, cpuset_t *mask) { struct cpuset *nset; mtx_assert(&cpuset_lock, MA_OWNED); CPU_AND(&set->cs_mask, mask); LIST_FOREACH(nset, &set->cs_children, cs_siblings) cpuset_update(nset, &set->cs_mask); return; } /* * Modify the set 'set' to use a copy of the mask provided. Apply this new * mask to restrict all children in the tree. Checks for validity before * applying the changes. */ static int cpuset_modify(struct cpuset *set, cpuset_t *mask) { struct cpuset *root; int error; error = priv_check(curthread, PRIV_SCHED_CPUSET); if (error) return (error); /* * In case we are called from within the jail * we do not allow modifying the dedicated root * cpuset of the jail but may still allow to * change child sets. */ if (jailed(curthread->td_ucred) && set->cs_flags & CPU_SET_ROOT) return (EPERM); /* * Verify that we have access to this set of * cpus. */ root = cpuset_getroot(set); mtx_lock_spin(&cpuset_lock); if (root && !CPU_SUBSET(&root->cs_mask, mask)) { error = EINVAL; goto out; } error = cpuset_testupdate(set, mask, 0); if (error) goto out; CPU_COPY(mask, &set->cs_mask); cpuset_update(set, mask); out: mtx_unlock_spin(&cpuset_lock); return (error); } /* * Recursively check for errors that would occur from applying mask to * the tree of sets starting at 'set'. Checks for sets that would become * empty as well as RDONLY flags. */ static int cpuset_testupdate_domain(struct cpuset *set, struct domainset *dset, struct domainset *orig, int *count, int check_mask) { struct cpuset *nset; struct domainset *domain; struct domainset newset; int error; mtx_assert(&cpuset_lock, MA_OWNED); if (set->cs_flags & CPU_SET_RDONLY) return (EPERM); domain = set->cs_domain; domainset_copy(domain, &newset); if (!domainset_equal(domain, orig)) { if (!domainset_restrict(domain, dset)) return (EDEADLK); DOMAINSET_AND(&newset.ds_mask, &dset->ds_mask); /* Count the number of domains that are changing. */ (*count)++; } error = 0; LIST_FOREACH(nset, &set->cs_children, cs_siblings) if ((error = cpuset_testupdate_domain(nset, &newset, domain, count, 1)) != 0) break; return (error); } /* * Applies the mask 'mask' without checking for empty sets or permissions. */ static void cpuset_update_domain(struct cpuset *set, struct domainset *domain, struct domainset *orig, struct domainlist *domains) { struct cpuset *nset; mtx_assert(&cpuset_lock, MA_OWNED); /* * If this domainset has changed from the parent we must calculate * a new set. Otherwise it simply inherits from the parent. When * we inherit from the parent we get a new mask and policy. If the * set is modified from the parent we keep the policy and only * update the mask. */ if (set->cs_domain != orig) { orig = set->cs_domain; set->cs_domain = domainset_shadow(domain, orig, domains); } else set->cs_domain = domain; LIST_FOREACH(nset, &set->cs_children, cs_siblings) cpuset_update_domain(nset, set->cs_domain, orig, domains); return; } /* * Modify the set 'set' to use a copy the domainset provided. Apply this new * mask to restrict all children in the tree. Checks for validity before * applying the changes. */ static int cpuset_modify_domain(struct cpuset *set, struct domainset *domain) { struct domainlist domains; struct domainset temp; struct domainset *dset; struct cpuset *root; int ndomains, needed; int error; error = priv_check(curthread, PRIV_SCHED_CPUSET); if (error) return (error); /* * In case we are called from within the jail * we do not allow modifying the dedicated root * cpuset of the jail but may still allow to * change child sets. */ if (jailed(curthread->td_ucred) && set->cs_flags & CPU_SET_ROOT) return (EPERM); domainset_freelist_init(&domains, 0); domain = domainset_create(domain); ndomains = needed = 0; do { if (ndomains < needed) { domainset_freelist_add(&domains, needed - ndomains); ndomains = needed; } root = cpuset_getroot(set); mtx_lock_spin(&cpuset_lock); dset = root->cs_domain; /* * Verify that we have access to this set of domains. */ if (root && !domainset_valid(dset, domain)) { error = EINVAL; goto out; } /* * If applying prefer we keep the current set as the fallback. */ if (domain->ds_policy == DOMAINSET_POLICY_PREFER) DOMAINSET_COPY(&set->cs_domain->ds_mask, &domain->ds_mask); /* * Determine whether we can apply this set of domains and * how many new domain structures it will require. */ domainset_copy(domain, &temp); needed = 0; error = cpuset_testupdate_domain(set, &temp, set->cs_domain, &needed, 0); if (error) goto out; } while (ndomains < needed); dset = set->cs_domain; cpuset_update_domain(set, domain, dset, &domains); out: mtx_unlock_spin(&cpuset_lock); domainset_freelist_free(&domains); if (error == 0) domainset_notify(); return (error); } /* * Resolve the 'which' parameter of several cpuset apis. * * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid. Also * checks for permission via p_cansched(). * * For WHICH_SET returns a valid set with a new reference. * * -1 may be supplied for any argument to mean the current proc/thread or * the base set of the current thread. May fail with ESRCH/EPERM. */ int cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp, struct cpuset **setp) { struct cpuset *set; struct thread *td; struct proc *p; int error; *pp = p = NULL; *tdp = td = NULL; *setp = set = NULL; switch (which) { case CPU_WHICH_PID: if (id == -1) { PROC_LOCK(curproc); p = curproc; break; } if ((p = pfind(id)) == NULL) return (ESRCH); break; case CPU_WHICH_TID: if (id == -1) { PROC_LOCK(curproc); p = curproc; td = curthread; break; } td = tdfind(id, -1); if (td == NULL) return (ESRCH); p = td->td_proc; break; case CPU_WHICH_CPUSET: if (id == -1) { thread_lock(curthread); set = cpuset_refbase(curthread->td_cpuset); thread_unlock(curthread); } else set = cpuset_lookup(id, curthread); if (set) { *setp = set; return (0); } return (ESRCH); case CPU_WHICH_JAIL: { /* Find `set' for prison with given id. */ struct prison *pr; sx_slock(&allprison_lock); pr = prison_find_child(curthread->td_ucred->cr_prison, id); sx_sunlock(&allprison_lock); if (pr == NULL) return (ESRCH); cpuset_ref(pr->pr_cpuset); *setp = pr->pr_cpuset; mtx_unlock(&pr->pr_mtx); return (0); } case CPU_WHICH_IRQ: case CPU_WHICH_DOMAIN: return (0); default: return (EINVAL); } error = p_cansched(curthread, p); if (error) { PROC_UNLOCK(p); return (error); } if (td == NULL) td = FIRST_THREAD_IN_PROC(p); *pp = p; *tdp = td; return (0); } static int cpuset_testshadow(struct cpuset *set, const cpuset_t *mask, const struct domainset *domain) { struct cpuset *parent; struct domainset *dset; parent = cpuset_getbase(set); /* * If we are restricting a cpu mask it must be a subset of the * parent or invalid CPUs have been specified. */ if (mask != NULL && !CPU_SUBSET(&parent->cs_mask, mask)) return (EINVAL); /* * If we are restricting a domain mask it must be a subset of the * parent or invalid domains have been specified. */ dset = parent->cs_domain; if (domain != NULL && !domainset_valid(dset, domain)) return (EINVAL); return (0); } /* * Create an anonymous set with the provided mask in the space provided by * 'nset'. If the passed in set is anonymous we use its parent otherwise * the new set is a child of 'set'. */ static int cpuset_shadow(struct cpuset *set, struct cpuset **nsetp, const cpuset_t *mask, const struct domainset *domain, struct setlist *cpusets, struct domainlist *domains) { struct cpuset *parent; struct cpuset *nset; struct domainset *dset; struct domainset *d; int error; error = cpuset_testshadow(set, mask, domain); if (error) return (error); parent = cpuset_getbase(set); dset = parent->cs_domain; if (mask == NULL) mask = &set->cs_mask; if (domain != NULL) d = domainset_shadow(dset, domain, domains); else d = set->cs_domain; nset = LIST_FIRST(cpusets); error = _cpuset_create(nset, parent, mask, d, CPUSET_INVALID); if (error == 0) { LIST_REMOVE(nset, cs_link); *nsetp = nset; } return (error); } static struct cpuset * cpuset_update_thread(struct thread *td, struct cpuset *nset) { struct cpuset *tdset; tdset = td->td_cpuset; td->td_cpuset = nset; td->td_domain.dr_policy = nset->cs_domain; sched_affinity(td); return (tdset); } static int cpuset_setproc_test_maskthread(struct cpuset *tdset, cpuset_t *mask, struct domainset *domain) { struct cpuset *parent; parent = cpuset_getbase(tdset); if (mask == NULL) mask = &tdset->cs_mask; if (domain == NULL) domain = tdset->cs_domain; return cpuset_testshadow(parent, mask, domain); } static int cpuset_setproc_maskthread(struct cpuset *tdset, cpuset_t *mask, struct domainset *domain, struct cpuset **nsetp, struct setlist *freelist, struct domainlist *domainlist) { struct cpuset *parent; parent = cpuset_getbase(tdset); if (mask == NULL) mask = &tdset->cs_mask; if (domain == NULL) domain = tdset->cs_domain; return cpuset_shadow(parent, nsetp, mask, domain, freelist, domainlist); } static int cpuset_setproc_setthread_mask(struct cpuset *tdset, struct cpuset *set, cpuset_t *mask, struct domainset *domain) { struct cpuset *parent; parent = cpuset_getbase(tdset); /* * If the thread restricted its mask then apply that same * restriction to the new set, otherwise take it wholesale. */ if (CPU_CMP(&tdset->cs_mask, &parent->cs_mask) != 0) { CPU_COPY(&tdset->cs_mask, mask); CPU_AND(mask, &set->cs_mask); } else CPU_COPY(&set->cs_mask, mask); /* * If the thread restricted the domain then we apply the * restriction to the new set but retain the policy. */ if (tdset->cs_domain != parent->cs_domain) { domainset_copy(tdset->cs_domain, domain); DOMAINSET_AND(&domain->ds_mask, &set->cs_domain->ds_mask); } else domainset_copy(set->cs_domain, domain); if (CPU_EMPTY(mask) || DOMAINSET_EMPTY(&domain->ds_mask)) return (EDEADLK); return (0); } static int cpuset_setproc_test_setthread(struct cpuset *tdset, struct cpuset *set) { struct domainset domain; cpuset_t mask; if (tdset->cs_id != CPUSET_INVALID) return (0); return cpuset_setproc_setthread_mask(tdset, set, &mask, &domain); } static int cpuset_setproc_setthread(struct cpuset *tdset, struct cpuset *set, struct cpuset **nsetp, struct setlist *freelist, struct domainlist *domainlist) { struct domainset domain; cpuset_t mask; int error; /* * If we're replacing on a thread that has not constrained the * original set we can simply accept the new set. */ if (tdset->cs_id != CPUSET_INVALID) { *nsetp = cpuset_ref(set); return (0); } error = cpuset_setproc_setthread_mask(tdset, set, &mask, &domain); if (error) return (error); return cpuset_shadow(tdset, nsetp, &mask, &domain, freelist, domainlist); } /* * Handle three cases for updating an entire process. * * 1) Set is non-null. This reparents all anonymous sets to the provided * set and replaces all non-anonymous td_cpusets with the provided set. * 2) Mask is non-null. This replaces or creates anonymous sets for every * thread with the existing base as a parent. * 3) domain is non-null. This creates anonymous sets for every thread * and replaces the domain set. * * This is overly complicated because we can't allocate while holding a * spinlock and spinlocks must be held while changing and examining thread * state. */ static int cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask, struct domainset *domain) { struct setlist freelist; struct setlist droplist; struct domainlist domainlist; struct cpuset *nset; struct thread *td; struct proc *p; int threads; int nfree; int error; /* * The algorithm requires two passes due to locking considerations. * * 1) Lookup the process and acquire the locks in the required order. * 2) If enough cpusets have not been allocated release the locks and * allocate them. Loop. */ cpuset_freelist_init(&freelist, 1); domainset_freelist_init(&domainlist, 1); nfree = 1; LIST_INIT(&droplist); nfree = 0; for (;;) { error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset); if (error) goto out; if (nfree >= p->p_numthreads) break; threads = p->p_numthreads; PROC_UNLOCK(p); if (nfree < threads) { cpuset_freelist_add(&freelist, threads - nfree); domainset_freelist_add(&domainlist, threads - nfree); nfree = threads; } } PROC_LOCK_ASSERT(p, MA_OWNED); /* * Now that the appropriate locks are held and we have enough cpusets, * make sure the operation will succeed before applying changes. The * proc lock prevents td_cpuset from changing between calls. */ error = 0; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (set != NULL) error = cpuset_setproc_test_setthread(td->td_cpuset, set); else error = cpuset_setproc_test_maskthread(td->td_cpuset, mask, domain); thread_unlock(td); if (error) goto unlock_out; } /* * Replace each thread's cpuset while using deferred release. We * must do this because the thread lock must be held while operating * on the thread and this limits the type of operations allowed. */ FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (set != NULL) error = cpuset_setproc_setthread(td->td_cpuset, set, &nset, &freelist, &domainlist); else error = cpuset_setproc_maskthread(td->td_cpuset, mask, domain, &nset, &freelist, &domainlist); if (error) { thread_unlock(td); break; } cpuset_rel_defer(&droplist, cpuset_update_thread(td, nset)); thread_unlock(td); } unlock_out: PROC_UNLOCK(p); out: while ((nset = LIST_FIRST(&droplist)) != NULL) cpuset_rel_complete(nset); cpuset_freelist_free(&freelist); domainset_freelist_free(&domainlist); return (error); } static int bitset_strprint(char *buf, size_t bufsiz, const struct bitset *set, int setlen) { size_t bytes; int i, once; char *p; once = 0; p = buf; for (i = 0; i < __bitset_words(setlen); i++) { if (once != 0) { if (bufsiz < 1) return (0); *p = ','; p++; bufsiz--; } else once = 1; if (bufsiz < sizeof(__STRING(ULONG_MAX))) return (0); bytes = snprintf(p, bufsiz, "%lx", set->__bits[i]); p += bytes; bufsiz -= bytes; } return (p - buf); } static int bitset_strscan(struct bitset *set, int setlen, const char *buf) { int i, ret; const char *p; BIT_ZERO(setlen, set); p = buf; for (i = 0; i < __bitset_words(setlen); i++) { if (*p == ',') { p++; continue; } ret = sscanf(p, "%lx", &set->__bits[i]); if (ret == 0 || ret == -1) break; while (isxdigit(*p)) p++; } return (p - buf); } /* * Return a string representing a valid layout for a cpuset_t object. * It expects an incoming buffer at least sized as CPUSETBUFSIZ. */ char * cpusetobj_strprint(char *buf, const cpuset_t *set) { bitset_strprint(buf, CPUSETBUFSIZ, (const struct bitset *)set, CPU_SETSIZE); return (buf); } /* * Build a valid cpuset_t object from a string representation. * It expects an incoming buffer at least sized as CPUSETBUFSIZ. */ int cpusetobj_strscan(cpuset_t *set, const char *buf) { char p; if (strlen(buf) > CPUSETBUFSIZ - 1) return (-1); p = buf[bitset_strscan((struct bitset *)set, CPU_SETSIZE, buf)]; if (p != '\0') return (-1); return (0); } /* * Handle a domainset specifier in the sysctl tree. A poiner to a pointer to * a domainset is in arg1. If the user specifies a valid domainset the * pointer is updated. * * Format is: * hex mask word 0,hex mask word 1,...:decimal policy:decimal preferred */ int sysctl_handle_domainset(SYSCTL_HANDLER_ARGS) { char buf[DOMAINSETBUFSIZ]; struct domainset *dset; struct domainset key; int policy, prefer, error; char *p; dset = *(struct domainset **)arg1; error = 0; if (dset != NULL) { p = buf + bitset_strprint(buf, DOMAINSETBUFSIZ, (const struct bitset *)&dset->ds_mask, DOMAINSET_SETSIZE); sprintf(p, ":%d:%d", dset->ds_policy, dset->ds_prefer); } else sprintf(buf, ""); error = sysctl_handle_string(oidp, buf, sizeof(buf), req); if (error != 0 || req->newptr == NULL) return (error); /* * Read in and validate the string. */ memset(&key, 0, sizeof(key)); p = &buf[bitset_strscan((struct bitset *)&key.ds_mask, DOMAINSET_SETSIZE, buf)]; if (p == buf) return (EINVAL); if (sscanf(p, ":%d:%d", &policy, &prefer) != 2) return (EINVAL); key.ds_policy = policy; key.ds_prefer = prefer; /* Domainset_create() validates the policy.*/ dset = domainset_create(&key); if (dset == NULL) return (EINVAL); *(struct domainset **)arg1 = dset; return (error); } /* * Apply an anonymous mask or a domain to a single thread. */ static int _cpuset_setthread(lwpid_t id, cpuset_t *mask, struct domainset *domain) { struct setlist cpusets; struct domainlist domainlist; struct cpuset *nset; struct cpuset *set; struct thread *td; struct proc *p; int error; cpuset_freelist_init(&cpusets, 1); domainset_freelist_init(&domainlist, domain != NULL); error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set); if (error) goto out; set = NULL; thread_lock(td); error = cpuset_shadow(td->td_cpuset, &nset, mask, domain, &cpusets, &domainlist); if (error == 0) set = cpuset_update_thread(td, nset); thread_unlock(td); PROC_UNLOCK(p); if (set) cpuset_rel(set); out: cpuset_freelist_free(&cpusets); domainset_freelist_free(&domainlist); return (error); } /* * Apply an anonymous mask to a single thread. */ int cpuset_setthread(lwpid_t id, cpuset_t *mask) { return _cpuset_setthread(id, mask, NULL); } /* * Apply new cpumask to the ithread. */ int cpuset_setithread(lwpid_t id, int cpu) { cpuset_t mask; CPU_ZERO(&mask); if (cpu == NOCPU) CPU_COPY(cpuset_root, &mask); else CPU_SET(cpu, &mask); return _cpuset_setthread(id, &mask, NULL); } /* * Initialize static domainsets after NUMA information is available. This is - * called very early during boot. + * called before memory allocators are initialized. */ void domainset_init(void) { struct domainset *dset; int i; dset = &domainset_roundrobin; DOMAINSET_COPY(&all_domains, &dset->ds_mask); dset->ds_policy = DOMAINSET_POLICY_ROUNDROBIN; dset->ds_prefer = -1; _domainset_create(dset, NULL); for (i = 0; i < vm_ndomains; i++) { dset = &domainset_prefer[i]; DOMAINSET_COPY(&all_domains, &dset->ds_mask); dset->ds_policy = DOMAINSET_POLICY_PREFER; dset->ds_prefer = i; _domainset_create(dset, NULL); } } /* * Create the domainset for cpuset 0, 1 and cpuset 2. */ void domainset_zero(void) { - struct domainset *dset; + struct domainset *dset, *tmp; mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE); dset = &domainset0; DOMAINSET_COPY(&all_domains, &dset->ds_mask); dset->ds_policy = DOMAINSET_POLICY_FIRSTTOUCH; dset->ds_prefer = -1; curthread->td_domain.dr_policy = _domainset_create(dset, NULL); domainset_copy(dset, &domainset2); domainset2.ds_policy = DOMAINSET_POLICY_INTERLEAVE; kernel_object->domain.dr_policy = _domainset_create(&domainset2, NULL); /* Remove empty domains from the global policies. */ - LIST_FOREACH(dset, &cpuset_domains, ds_link) - (void)domainset_empty_vm(dset); + LIST_FOREACH_SAFE(dset, &cpuset_domains, ds_link, tmp) + if (domainset_empty_vm(dset)) + LIST_REMOVE(dset, ds_link); } /* * Creates system-wide cpusets and the cpuset for thread0 including three * sets: * * 0 - The root set which should represent all valid processors in the * system. It is initially created with a mask of all processors * because we don't know what processors are valid until cpuset_init() * runs. This set is immutable. * 1 - The default set which all processes are a member of until changed. * This allows an administrator to move all threads off of given cpus to * dedicate them to high priority tasks or save power etc. * 2 - The kernel set which allows restriction and policy to be applied only * to kernel threads and the kernel_object. */ struct cpuset * cpuset_thread0(void) { struct cpuset *set; int i; int error __unused; cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); domainset_zone = uma_zcreate("domainset", sizeof(struct domainset), NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); /* * Create the root system set (0) for the whole machine. Doesn't use * cpuset_create() due to NULL parent. */ set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO); CPU_COPY(&all_cpus, &set->cs_mask); LIST_INIT(&set->cs_children); LIST_INSERT_HEAD(&cpuset_ids, set, cs_link); set->cs_ref = 1; set->cs_flags = CPU_SET_ROOT | CPU_SET_RDONLY; set->cs_domain = &domainset0; cpuset_zero = set; cpuset_root = &set->cs_mask; /* * Now derive a default (1), modifiable set from that to give out. */ set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO); error = _cpuset_create(set, cpuset_zero, NULL, NULL, 1); KASSERT(error == 0, ("Error creating default set: %d\n", error)); cpuset_default = set; /* * Create the kernel set (2). */ set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO); error = _cpuset_create(set, cpuset_zero, NULL, NULL, 2); KASSERT(error == 0, ("Error creating kernel set: %d\n", error)); set->cs_domain = &domainset2; cpuset_kernel = set; /* * Initialize the unit allocator. 0 and 1 are allocated above. */ cpuset_unr = new_unrhdr(2, INT_MAX, NULL); /* * If MD code has not initialized per-domain cpusets, place all * CPUs in domain 0. */ for (i = 0; i < MAXMEMDOM; i++) if (!CPU_EMPTY(&cpuset_domain[i])) goto domains_set; CPU_COPY(&all_cpus, &cpuset_domain[0]); domains_set: return (cpuset_default); } void cpuset_kernthread(struct thread *td) { struct cpuset *set; thread_lock(td); set = td->td_cpuset; td->td_cpuset = cpuset_ref(cpuset_kernel); thread_unlock(td); cpuset_rel(set); } /* * Create a cpuset, which would be cpuset_create() but * mark the new 'set' as root. * * We are not going to reparent the td to it. Use cpuset_setproc_update_set() * for that. * * In case of no error, returns the set in *setp locked with a reference. */ int cpuset_create_root(struct prison *pr, struct cpuset **setp) { struct cpuset *set; int error; KASSERT(pr != NULL, ("[%s:%d] invalid pr", __func__, __LINE__)); KASSERT(setp != NULL, ("[%s:%d] invalid setp", __func__, __LINE__)); error = cpuset_create(setp, pr->pr_cpuset, &pr->pr_cpuset->cs_mask); if (error) return (error); KASSERT(*setp != NULL, ("[%s:%d] cpuset_create returned invalid data", __func__, __LINE__)); /* Mark the set as root. */ set = *setp; set->cs_flags |= CPU_SET_ROOT; return (0); } int cpuset_setproc_update_set(struct proc *p, struct cpuset *set) { int error; KASSERT(p != NULL, ("[%s:%d] invalid proc", __func__, __LINE__)); KASSERT(set != NULL, ("[%s:%d] invalid set", __func__, __LINE__)); cpuset_ref(set); error = cpuset_setproc(p->p_pid, set, NULL, NULL); if (error) return (error); cpuset_rel(set); return (0); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_args { cpusetid_t *setid; }; #endif int sys_cpuset(struct thread *td, struct cpuset_args *uap) { struct cpuset *root; struct cpuset *set; int error; thread_lock(td); root = cpuset_refroot(td->td_cpuset); thread_unlock(td); error = cpuset_create(&set, root, &root->cs_mask); cpuset_rel(root); if (error) return (error); error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id)); if (error == 0) error = cpuset_setproc(-1, set, NULL, NULL); cpuset_rel(set); return (error); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_setid_args { cpuwhich_t which; id_t id; cpusetid_t setid; }; #endif int sys_cpuset_setid(struct thread *td, struct cpuset_setid_args *uap) { return (kern_cpuset_setid(td, uap->which, uap->id, uap->setid)); } int kern_cpuset_setid(struct thread *td, cpuwhich_t which, id_t id, cpusetid_t setid) { struct cpuset *set; int error; /* * Presently we only support per-process sets. */ if (which != CPU_WHICH_PID) return (EINVAL); set = cpuset_lookup(setid, td); if (set == NULL) return (ESRCH); error = cpuset_setproc(id, set, NULL, NULL); cpuset_rel(set); return (error); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_getid_args { cpulevel_t level; cpuwhich_t which; id_t id; cpusetid_t *setid; }; #endif int sys_cpuset_getid(struct thread *td, struct cpuset_getid_args *uap) { return (kern_cpuset_getid(td, uap->level, uap->which, uap->id, uap->setid)); } int kern_cpuset_getid(struct thread *td, cpulevel_t level, cpuwhich_t which, id_t id, cpusetid_t *setid) { struct cpuset *nset; struct cpuset *set; struct thread *ttd; struct proc *p; cpusetid_t tmpid; int error; if (level == CPU_LEVEL_WHICH && which != CPU_WHICH_CPUSET) return (EINVAL); error = cpuset_which(which, id, &p, &ttd, &set); if (error) return (error); switch (which) { case CPU_WHICH_TID: case CPU_WHICH_PID: thread_lock(ttd); set = cpuset_refbase(ttd->td_cpuset); thread_unlock(ttd); PROC_UNLOCK(p); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: break; case CPU_WHICH_IRQ: case CPU_WHICH_DOMAIN: return (EINVAL); } switch (level) { case CPU_LEVEL_ROOT: nset = cpuset_refroot(set); cpuset_rel(set); set = nset; break; case CPU_LEVEL_CPUSET: break; case CPU_LEVEL_WHICH: break; } tmpid = set->cs_id; cpuset_rel(set); if (error == 0) error = copyout(&tmpid, setid, sizeof(tmpid)); return (error); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_getaffinity_args { cpulevel_t level; cpuwhich_t which; id_t id; size_t cpusetsize; cpuset_t *mask; }; #endif int sys_cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap) { return (kern_cpuset_getaffinity(td, uap->level, uap->which, uap->id, uap->cpusetsize, uap->mask)); } int kern_cpuset_getaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which, id_t id, size_t cpusetsize, cpuset_t *maskp) { struct thread *ttd; struct cpuset *nset; struct cpuset *set; struct proc *p; cpuset_t *mask; int error; size_t size; if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY) return (ERANGE); /* In Capability mode, you can only get your own CPU set. */ if (IN_CAPABILITY_MODE(td)) { if (level != CPU_LEVEL_WHICH) return (ECAPMODE); if (which != CPU_WHICH_TID && which != CPU_WHICH_PID) return (ECAPMODE); if (id != -1) return (ECAPMODE); } size = cpusetsize; mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO); error = cpuset_which(which, id, &p, &ttd, &set); if (error) goto out; switch (level) { case CPU_LEVEL_ROOT: case CPU_LEVEL_CPUSET: switch (which) { case CPU_WHICH_TID: case CPU_WHICH_PID: thread_lock(ttd); set = cpuset_ref(ttd->td_cpuset); thread_unlock(ttd); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: case CPU_WHICH_DOMAIN: error = EINVAL; goto out; } if (level == CPU_LEVEL_ROOT) nset = cpuset_refroot(set); else nset = cpuset_refbase(set); CPU_COPY(&nset->cs_mask, mask); cpuset_rel(nset); break; case CPU_LEVEL_WHICH: switch (which) { case CPU_WHICH_TID: thread_lock(ttd); CPU_COPY(&ttd->td_cpuset->cs_mask, mask); thread_unlock(ttd); break; case CPU_WHICH_PID: FOREACH_THREAD_IN_PROC(p, ttd) { thread_lock(ttd); CPU_OR(mask, &ttd->td_cpuset->cs_mask); thread_unlock(ttd); } break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: CPU_COPY(&set->cs_mask, mask); break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: error = intr_getaffinity(id, which, mask); break; case CPU_WHICH_DOMAIN: if (id < 0 || id >= MAXMEMDOM) error = ESRCH; else CPU_COPY(&cpuset_domain[id], mask); break; } break; default: error = EINVAL; break; } if (set) cpuset_rel(set); if (p) PROC_UNLOCK(p); if (error == 0) error = copyout(mask, maskp, size); out: free(mask, M_TEMP); return (error); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_setaffinity_args { cpulevel_t level; cpuwhich_t which; id_t id; size_t cpusetsize; const cpuset_t *mask; }; #endif int sys_cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap) { return (kern_cpuset_setaffinity(td, uap->level, uap->which, uap->id, uap->cpusetsize, uap->mask)); } int kern_cpuset_setaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which, id_t id, size_t cpusetsize, const cpuset_t *maskp) { struct cpuset *nset; struct cpuset *set; struct thread *ttd; struct proc *p; cpuset_t *mask; int error; if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY) return (ERANGE); /* In Capability mode, you can only set your own CPU set. */ if (IN_CAPABILITY_MODE(td)) { if (level != CPU_LEVEL_WHICH) return (ECAPMODE); if (which != CPU_WHICH_TID && which != CPU_WHICH_PID) return (ECAPMODE); if (id != -1) return (ECAPMODE); } mask = malloc(cpusetsize, M_TEMP, M_WAITOK | M_ZERO); error = copyin(maskp, mask, cpusetsize); if (error) goto out; /* * Verify that no high bits are set. */ if (cpusetsize > sizeof(cpuset_t)) { char *end; char *cp; end = cp = (char *)&mask->__bits; end += cpusetsize; cp += sizeof(cpuset_t); while (cp != end) if (*cp++ != 0) { error = EINVAL; goto out; } } switch (level) { case CPU_LEVEL_ROOT: case CPU_LEVEL_CPUSET: error = cpuset_which(which, id, &p, &ttd, &set); if (error) break; switch (which) { case CPU_WHICH_TID: case CPU_WHICH_PID: thread_lock(ttd); set = cpuset_ref(ttd->td_cpuset); thread_unlock(ttd); PROC_UNLOCK(p); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: case CPU_WHICH_DOMAIN: error = EINVAL; goto out; } if (level == CPU_LEVEL_ROOT) nset = cpuset_refroot(set); else nset = cpuset_refbase(set); error = cpuset_modify(nset, mask); cpuset_rel(nset); cpuset_rel(set); break; case CPU_LEVEL_WHICH: switch (which) { case CPU_WHICH_TID: error = cpuset_setthread(id, mask); break; case CPU_WHICH_PID: error = cpuset_setproc(id, NULL, mask, NULL); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: error = cpuset_which(which, id, &p, &ttd, &set); if (error == 0) { error = cpuset_modify(set, mask); cpuset_rel(set); } break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: error = intr_setaffinity(id, which, mask); break; default: error = EINVAL; break; } break; default: error = EINVAL; break; } out: free(mask, M_TEMP); return (error); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_getdomain_args { cpulevel_t level; cpuwhich_t which; id_t id; size_t domainsetsize; domainset_t *mask; int *policy; }; #endif int sys_cpuset_getdomain(struct thread *td, struct cpuset_getdomain_args *uap) { return (kern_cpuset_getdomain(td, uap->level, uap->which, uap->id, uap->domainsetsize, uap->mask, uap->policy)); } int kern_cpuset_getdomain(struct thread *td, cpulevel_t level, cpuwhich_t which, id_t id, size_t domainsetsize, domainset_t *maskp, int *policyp) { struct domainset outset; struct thread *ttd; struct cpuset *nset; struct cpuset *set; struct domainset *dset; struct proc *p; domainset_t *mask; int error; if (domainsetsize < sizeof(domainset_t) || domainsetsize > DOMAINSET_MAXSIZE / NBBY) return (ERANGE); /* In Capability mode, you can only get your own domain set. */ if (IN_CAPABILITY_MODE(td)) { if (level != CPU_LEVEL_WHICH) return (ECAPMODE); if (which != CPU_WHICH_TID && which != CPU_WHICH_PID) return (ECAPMODE); if (id != -1) return (ECAPMODE); } mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO); bzero(&outset, sizeof(outset)); error = cpuset_which(which, id, &p, &ttd, &set); if (error) goto out; switch (level) { case CPU_LEVEL_ROOT: case CPU_LEVEL_CPUSET: switch (which) { case CPU_WHICH_TID: case CPU_WHICH_PID: thread_lock(ttd); set = cpuset_ref(ttd->td_cpuset); thread_unlock(ttd); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: case CPU_WHICH_DOMAIN: error = EINVAL; goto out; } if (level == CPU_LEVEL_ROOT) nset = cpuset_refroot(set); else nset = cpuset_refbase(set); domainset_copy(nset->cs_domain, &outset); cpuset_rel(nset); break; case CPU_LEVEL_WHICH: switch (which) { case CPU_WHICH_TID: thread_lock(ttd); domainset_copy(ttd->td_cpuset->cs_domain, &outset); thread_unlock(ttd); break; case CPU_WHICH_PID: FOREACH_THREAD_IN_PROC(p, ttd) { thread_lock(ttd); dset = ttd->td_cpuset->cs_domain; /* Show all domains in the proc. */ DOMAINSET_OR(&outset.ds_mask, &dset->ds_mask); /* Last policy wins. */ outset.ds_policy = dset->ds_policy; outset.ds_prefer = dset->ds_prefer; thread_unlock(ttd); } break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: domainset_copy(set->cs_domain, &outset); break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: case CPU_WHICH_DOMAIN: error = EINVAL; break; } break; default: error = EINVAL; break; } if (set) cpuset_rel(set); if (p) PROC_UNLOCK(p); /* * Translate prefer into a set containing only the preferred domain, * not the entire fallback set. */ if (outset.ds_policy == DOMAINSET_POLICY_PREFER) { DOMAINSET_ZERO(&outset.ds_mask); DOMAINSET_SET(outset.ds_prefer, &outset.ds_mask); } DOMAINSET_COPY(&outset.ds_mask, mask); if (error == 0) error = copyout(mask, maskp, domainsetsize); if (error == 0) if (suword32(policyp, outset.ds_policy) != 0) error = EFAULT; out: free(mask, M_TEMP); return (error); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_setdomain_args { cpulevel_t level; cpuwhich_t which; id_t id; size_t domainsetsize; domainset_t *mask; int policy; }; #endif int sys_cpuset_setdomain(struct thread *td, struct cpuset_setdomain_args *uap) { return (kern_cpuset_setdomain(td, uap->level, uap->which, uap->id, uap->domainsetsize, uap->mask, uap->policy)); } int kern_cpuset_setdomain(struct thread *td, cpulevel_t level, cpuwhich_t which, id_t id, size_t domainsetsize, const domainset_t *maskp, int policy) { struct cpuset *nset; struct cpuset *set; struct thread *ttd; struct proc *p; struct domainset domain; domainset_t *mask; int error; if (domainsetsize < sizeof(domainset_t) || domainsetsize > DOMAINSET_MAXSIZE / NBBY) return (ERANGE); if (policy <= DOMAINSET_POLICY_INVALID || policy > DOMAINSET_POLICY_MAX) return (EINVAL); /* In Capability mode, you can only set your own CPU set. */ if (IN_CAPABILITY_MODE(td)) { if (level != CPU_LEVEL_WHICH) return (ECAPMODE); if (which != CPU_WHICH_TID && which != CPU_WHICH_PID) return (ECAPMODE); if (id != -1) return (ECAPMODE); } memset(&domain, 0, sizeof(domain)); mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO); error = copyin(maskp, mask, domainsetsize); if (error) goto out; /* * Verify that no high bits are set. */ if (domainsetsize > sizeof(domainset_t)) { char *end; char *cp; end = cp = (char *)&mask->__bits; end += domainsetsize; cp += sizeof(domainset_t); while (cp != end) if (*cp++ != 0) { error = EINVAL; goto out; } } DOMAINSET_COPY(mask, &domain.ds_mask); domain.ds_policy = policy; /* Translate preferred policy into a mask and fallback. */ if (policy == DOMAINSET_POLICY_PREFER) { /* Only support a single preferred domain. */ if (DOMAINSET_COUNT(&domain.ds_mask) != 1) { error = EINVAL; goto out; } domain.ds_prefer = DOMAINSET_FFS(&domain.ds_mask) - 1; /* This will be constrained by domainset_shadow(). */ DOMAINSET_FILL(&domain.ds_mask); } /* * When given an impossible policy, fall back to interleaving * across all domains */ if (domainset_empty_vm(&domain)) domainset_copy(&domainset2, &domain); switch (level) { case CPU_LEVEL_ROOT: case CPU_LEVEL_CPUSET: error = cpuset_which(which, id, &p, &ttd, &set); if (error) break; switch (which) { case CPU_WHICH_TID: case CPU_WHICH_PID: thread_lock(ttd); set = cpuset_ref(ttd->td_cpuset); thread_unlock(ttd); PROC_UNLOCK(p); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: case CPU_WHICH_DOMAIN: error = EINVAL; goto out; } if (level == CPU_LEVEL_ROOT) nset = cpuset_refroot(set); else nset = cpuset_refbase(set); error = cpuset_modify_domain(nset, &domain); cpuset_rel(nset); cpuset_rel(set); break; case CPU_LEVEL_WHICH: switch (which) { case CPU_WHICH_TID: error = _cpuset_setthread(id, NULL, &domain); break; case CPU_WHICH_PID: error = cpuset_setproc(id, NULL, NULL, &domain); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: error = cpuset_which(which, id, &p, &ttd, &set); if (error == 0) { error = cpuset_modify_domain(set, &domain); cpuset_rel(set); } break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: default: error = EINVAL; break; } break; default: error = EINVAL; break; } out: free(mask, M_TEMP); return (error); } #ifdef DDB static void ddb_display_bitset(const struct bitset *set, int size) { int bit, once; for (once = 0, bit = 0; bit < size; bit++) { if (CPU_ISSET(bit, set)) { if (once == 0) { db_printf("%d", bit); once = 1; } else db_printf(",%d", bit); } } if (once == 0) db_printf(""); } void ddb_display_cpuset(const cpuset_t *set) { ddb_display_bitset((const struct bitset *)set, CPU_SETSIZE); } static void ddb_display_domainset(const domainset_t *set) { ddb_display_bitset((const struct bitset *)set, DOMAINSET_SETSIZE); } DB_SHOW_COMMAND(cpusets, db_show_cpusets) { struct cpuset *set; LIST_FOREACH(set, &cpuset_ids, cs_link) { db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n", set, set->cs_id, set->cs_ref, set->cs_flags, (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0); db_printf(" cpu mask="); ddb_display_cpuset(&set->cs_mask); db_printf("\n"); db_printf(" domain policy %d prefer %d mask=", set->cs_domain->ds_policy, set->cs_domain->ds_prefer); ddb_display_domainset(&set->cs_domain->ds_mask); db_printf("\n"); if (db_pager_quit) break; } } DB_SHOW_COMMAND(domainsets, db_show_domainsets) { struct domainset *set; LIST_FOREACH(set, &cpuset_domains, ds_link) { db_printf("set=%p policy %d prefer %d cnt %d\n", set, set->ds_policy, set->ds_prefer, set->ds_cnt); db_printf(" mask ="); ddb_display_domainset(&set->ds_mask); db_printf("\n"); } } #endif /* DDB */ Index: head/sys/vm/uma_core.c =================================================================== --- head/sys/vm/uma_core.c (revision 339924) +++ head/sys/vm/uma_core.c (revision 339925) @@ -1,4264 +1,4265 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson * Copyright (c) 2004, 2005 Bosko Milekic * Copyright (c) 2004-2006 Robert N. M. Watson * 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 unmodified, 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. */ /* * uma_core.c Implementation of the Universal Memory allocator * * This allocator is intended to replace the multitude of similar object caches * in the standard FreeBSD kernel. The intent is to be flexible as well as * efficient. A primary design goal is to return unused memory to the rest of * the system. This will make the system as a whole more flexible due to the * ability to move memory to subsystems which most need it instead of leaving * pools of reserved memory unused. * * The basic ideas stem from similar slab/zone based allocators whose algorithms * are well known. * */ /* * TODO: * - Improve memory usage for large allocations * - Investigate cache size adjustments */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_param.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 #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DEBUG_MEMGUARD #include #endif /* * This is the zone and keg from which all zones are spawned. */ static uma_zone_t kegs; static uma_zone_t zones; /* This is the zone from which all offpage uma_slab_ts are allocated. */ static uma_zone_t slabzone; /* * The initial hash tables come out of this zone so they can be allocated * prior to malloc coming up. */ static uma_zone_t hashzone; /* The boot-time adjusted value for cache line alignment. */ int uma_align_cache = 64 - 1; static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); /* * Are we allowed to allocate buckets? */ static int bucketdisable = 1; /* Linked list of all kegs in the system */ static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); /* Linked list of all cache-only zones in the system */ static LIST_HEAD(,uma_zone) uma_cachezones = LIST_HEAD_INITIALIZER(uma_cachezones); /* This RW lock protects the keg list */ static struct rwlock_padalign __exclusive_cache_line uma_rwlock; /* * Pointer and counter to pool of pages, that is preallocated at * startup to bootstrap UMA. */ static char *bootmem; static int boot_pages; static struct sx uma_drain_lock; /* kmem soft limit. */ static unsigned long uma_kmem_limit = LONG_MAX; static volatile unsigned long uma_kmem_total; /* Is the VM done starting up? */ static enum { BOOT_COLD = 0, BOOT_STRAPPED, BOOT_PAGEALLOC, BOOT_BUCKETS, BOOT_RUNNING } booted = BOOT_COLD; /* * This is the handle used to schedule events that need to happen * outside of the allocation fast path. */ static struct callout uma_callout; #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ /* * This structure is passed as the zone ctor arg so that I don't have to create * a special allocation function just for zones. */ struct uma_zctor_args { const char *name; size_t size; uma_ctor ctor; uma_dtor dtor; uma_init uminit; uma_fini fini; uma_import import; uma_release release; void *arg; uma_keg_t keg; int align; uint32_t flags; }; struct uma_kctor_args { uma_zone_t zone; size_t size; uma_init uminit; uma_fini fini; int align; uint32_t flags; }; struct uma_bucket_zone { uma_zone_t ubz_zone; char *ubz_name; int ubz_entries; /* Number of items it can hold. */ int ubz_maxsize; /* Maximum allocation size per-item. */ }; /* * Compute the actual number of bucket entries to pack them in power * of two sizes for more efficient space utilization. */ #define BUCKET_SIZE(n) \ (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) #define BUCKET_MAX BUCKET_SIZE(256) struct uma_bucket_zone bucket_zones[] = { { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, { NULL, NULL, 0} }; /* * Flags and enumerations to be passed to internal functions. */ enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; #define UMA_ANYDOMAIN -1 /* Special value for domain search. */ /* Prototypes.. */ int uma_startup_count(int); void uma_startup(void *, int); void uma_startup1(void); void uma_startup2(void); static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int); static void page_free(void *, vm_size_t, uint8_t); static void pcpu_page_free(void *, vm_size_t, uint8_t); static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int); static void cache_drain(uma_zone_t); static void bucket_drain(uma_zone_t, uma_bucket_t); static void bucket_cache_drain(uma_zone_t zone); static int keg_ctor(void *, int, void *, int); static void keg_dtor(void *, int, void *); static int zone_ctor(void *, int, void *, int); static void zone_dtor(void *, int, void *); static int zero_init(void *, int, int); static void keg_small_init(uma_keg_t keg); static void keg_large_init(uma_keg_t keg); static void zone_foreach(void (*zfunc)(uma_zone_t)); static void zone_timeout(uma_zone_t zone); static int hash_alloc(struct uma_hash *); static int hash_expand(struct uma_hash *, struct uma_hash *); static void hash_free(struct uma_hash *hash); static void uma_timeout(void *); static void uma_startup3(void); static void *zone_alloc_item(uma_zone_t, void *, int, int); static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); static void bucket_enable(void); static void bucket_init(void); static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); static void bucket_zone_drain(void); static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int); static uma_slab_t zone_fetch_slab(uma_zone_t, uma_keg_t, int, int); static uma_slab_t zone_fetch_slab_multi(uma_zone_t, uma_keg_t, int, int); static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, int align, uint32_t flags); static int zone_import(uma_zone_t, void **, int, int, int); static void zone_release(uma_zone_t, void **, int); static void uma_zero_item(void *, uma_zone_t); void uma_print_zone(uma_zone_t); void uma_print_stats(void); static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); #ifdef INVARIANTS static bool uma_dbg_kskip(uma_keg_t keg, void *mem); static bool uma_dbg_zskip(uma_zone_t zone, void *mem); static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item); static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item); static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD, 0, "Memory allocation debugging"); static u_int dbg_divisor = 1; SYSCTL_UINT(_vm_debug, OID_AUTO, divisor, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0, "Debug & thrash every this item in memory allocator"); static counter_u64_t uma_dbg_cnt = EARLY_COUNTER; static counter_u64_t uma_skip_cnt = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD, &uma_dbg_cnt, "memory items debugged"); SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD, &uma_skip_cnt, "memory items skipped, not debugged"); #endif SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); static int zone_warnings = 1; SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, "Warn when UMA zones becomes full"); /* Adjust bytes under management by UMA. */ static inline void uma_total_dec(unsigned long size) { atomic_subtract_long(&uma_kmem_total, size); } static inline void uma_total_inc(unsigned long size) { if (atomic_fetchadd_long(&uma_kmem_total, size) > uma_kmem_limit) uma_reclaim_wakeup(); } /* * This routine checks to see whether or not it's safe to enable buckets. */ static void bucket_enable(void) { bucketdisable = vm_page_count_min(); } /* * Initialize bucket_zones, the array of zones of buckets of various sizes. * * For each zone, calculate the memory required for each bucket, consisting * of the header and an array of pointers. */ static void bucket_init(void) { struct uma_bucket_zone *ubz; int size; for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { size = roundup(sizeof(struct uma_bucket), sizeof(void *)); size += sizeof(void *) * ubz->ubz_entries; ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET | UMA_ZONE_NUMA); } } /* * Given a desired number of entries for a bucket, return the zone from which * to allocate the bucket. */ static struct uma_bucket_zone * bucket_zone_lookup(int entries) { struct uma_bucket_zone *ubz; for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) if (ubz->ubz_entries >= entries) return (ubz); ubz--; return (ubz); } static int bucket_select(int size) { struct uma_bucket_zone *ubz; ubz = &bucket_zones[0]; if (size > ubz->ubz_maxsize) return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); for (; ubz->ubz_entries != 0; ubz++) if (ubz->ubz_maxsize < size) break; ubz--; return (ubz->ubz_entries); } static uma_bucket_t bucket_alloc(uma_zone_t zone, void *udata, int flags) { struct uma_bucket_zone *ubz; uma_bucket_t bucket; /* * This is to stop us from allocating per cpu buckets while we're * running out of vm.boot_pages. Otherwise, we would exhaust the * boot pages. This also prevents us from allocating buckets in * low memory situations. */ if (bucketdisable) return (NULL); /* * To limit bucket recursion we store the original zone flags * in a cookie passed via zalloc_arg/zfree_arg. This allows the * NOVM flag to persist even through deep recursions. We also * store ZFLAG_BUCKET once we have recursed attempting to allocate * a bucket for a bucket zone so we do not allow infinite bucket * recursion. This cookie will even persist to frees of unused * buckets via the allocation path or bucket allocations in the * free path. */ if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) udata = (void *)(uintptr_t)zone->uz_flags; else { if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) return (NULL); udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); } if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) flags |= M_NOVM; ubz = bucket_zone_lookup(zone->uz_count); if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) ubz++; bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); if (bucket) { #ifdef INVARIANTS bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); #endif bucket->ub_cnt = 0; bucket->ub_entries = ubz->ubz_entries; } return (bucket); } static void bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) { struct uma_bucket_zone *ubz; KASSERT(bucket->ub_cnt == 0, ("bucket_free: Freeing a non free bucket.")); if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) udata = (void *)(uintptr_t)zone->uz_flags; ubz = bucket_zone_lookup(bucket->ub_entries); uma_zfree_arg(ubz->ubz_zone, bucket, udata); } static void bucket_zone_drain(void) { struct uma_bucket_zone *ubz; for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) zone_drain(ubz->ubz_zone); } static void zone_log_warning(uma_zone_t zone) { static const struct timeval warninterval = { 300, 0 }; if (!zone_warnings || zone->uz_warning == NULL) return; if (ratecheck(&zone->uz_ratecheck, &warninterval)) printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); } static inline void zone_maxaction(uma_zone_t zone) { if (zone->uz_maxaction.ta_func != NULL) taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); } static void zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) { uma_klink_t klink; LIST_FOREACH(klink, &zone->uz_kegs, kl_link) kegfn(klink->kl_keg); } /* * Routine called by timeout which is used to fire off some time interval * based calculations. (stats, hash size, etc.) * * Arguments: * arg Unused * * Returns: * Nothing */ static void uma_timeout(void *unused) { bucket_enable(); zone_foreach(zone_timeout); /* Reschedule this event */ callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); } /* * Routine to perform timeout driven calculations. This expands the * hashes and does per cpu statistics aggregation. * * Returns nothing. */ static void keg_timeout(uma_keg_t keg) { KEG_LOCK(keg); /* * Expand the keg hash table. * * This is done if the number of slabs is larger than the hash size. * What I'm trying to do here is completely reduce collisions. This * may be a little aggressive. Should I allow for two collisions max? */ if (keg->uk_flags & UMA_ZONE_HASH && keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { struct uma_hash newhash; struct uma_hash oldhash; int ret; /* * This is so involved because allocating and freeing * while the keg lock is held will lead to deadlock. * I have to do everything in stages and check for * races. */ newhash = keg->uk_hash; KEG_UNLOCK(keg); ret = hash_alloc(&newhash); KEG_LOCK(keg); if (ret) { if (hash_expand(&keg->uk_hash, &newhash)) { oldhash = keg->uk_hash; keg->uk_hash = newhash; } else oldhash = newhash; KEG_UNLOCK(keg); hash_free(&oldhash); return; } } KEG_UNLOCK(keg); } static void zone_timeout(uma_zone_t zone) { zone_foreach_keg(zone, &keg_timeout); } /* * Allocate and zero fill the next sized hash table from the appropriate * backing store. * * Arguments: * hash A new hash structure with the old hash size in uh_hashsize * * Returns: * 1 on success and 0 on failure. */ static int hash_alloc(struct uma_hash *hash) { int oldsize; int alloc; oldsize = hash->uh_hashsize; /* We're just going to go to a power of two greater */ if (oldsize) { hash->uh_hashsize = oldsize * 2; alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; hash->uh_slab_hash = (struct slabhead *)malloc(alloc, M_UMAHASH, M_NOWAIT); } else { alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, UMA_ANYDOMAIN, M_WAITOK); hash->uh_hashsize = UMA_HASH_SIZE_INIT; } if (hash->uh_slab_hash) { bzero(hash->uh_slab_hash, alloc); hash->uh_hashmask = hash->uh_hashsize - 1; return (1); } return (0); } /* * Expands the hash table for HASH zones. This is done from zone_timeout * to reduce collisions. This must not be done in the regular allocation * path, otherwise, we can recurse on the vm while allocating pages. * * Arguments: * oldhash The hash you want to expand * newhash The hash structure for the new table * * Returns: * Nothing * * Discussion: */ static int hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) { uma_slab_t slab; int hval; int i; if (!newhash->uh_slab_hash) return (0); if (oldhash->uh_hashsize >= newhash->uh_hashsize) return (0); /* * I need to investigate hash algorithms for resizing without a * full rehash. */ for (i = 0; i < oldhash->uh_hashsize; i++) while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); hval = UMA_HASH(newhash, slab->us_data); SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], slab, us_hlink); } return (1); } /* * Free the hash bucket to the appropriate backing store. * * Arguments: * slab_hash The hash bucket we're freeing * hashsize The number of entries in that hash bucket * * Returns: * Nothing */ static void hash_free(struct uma_hash *hash) { if (hash->uh_slab_hash == NULL) return; if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); else free(hash->uh_slab_hash, M_UMAHASH); } /* * Frees all outstanding items in a bucket * * Arguments: * zone The zone to free to, must be unlocked. * bucket The free/alloc bucket with items, cpu queue must be locked. * * Returns: * Nothing */ static void bucket_drain(uma_zone_t zone, uma_bucket_t bucket) { int i; if (bucket == NULL) return; if (zone->uz_fini) for (i = 0; i < bucket->ub_cnt; i++) zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); bucket->ub_cnt = 0; } /* * Drains the per cpu caches for a zone. * * NOTE: This may only be called while the zone is being turn down, and not * during normal operation. This is necessary in order that we do not have * to migrate CPUs to drain the per-CPU caches. * * Arguments: * zone The zone to drain, must be unlocked. * * Returns: * Nothing */ static void cache_drain(uma_zone_t zone) { uma_cache_t cache; int cpu; /* * XXX: It is safe to not lock the per-CPU caches, because we're * tearing down the zone anyway. I.e., there will be no further use * of the caches at this point. * * XXX: It would good to be able to assert that the zone is being * torn down to prevent improper use of cache_drain(). * * XXX: We lock the zone before passing into bucket_cache_drain() as * it is used elsewhere. Should the tear-down path be made special * there in some form? */ CPU_FOREACH(cpu) { cache = &zone->uz_cpu[cpu]; bucket_drain(zone, cache->uc_allocbucket); bucket_drain(zone, cache->uc_freebucket); if (cache->uc_allocbucket != NULL) bucket_free(zone, cache->uc_allocbucket, NULL); if (cache->uc_freebucket != NULL) bucket_free(zone, cache->uc_freebucket, NULL); cache->uc_allocbucket = cache->uc_freebucket = NULL; } ZONE_LOCK(zone); bucket_cache_drain(zone); ZONE_UNLOCK(zone); } static void cache_shrink(uma_zone_t zone) { if (zone->uz_flags & UMA_ZFLAG_INTERNAL) return; ZONE_LOCK(zone); zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; ZONE_UNLOCK(zone); } static void cache_drain_safe_cpu(uma_zone_t zone) { uma_cache_t cache; uma_bucket_t b1, b2; int domain; if (zone->uz_flags & UMA_ZFLAG_INTERNAL) return; b1 = b2 = NULL; ZONE_LOCK(zone); critical_enter(); if (zone->uz_flags & UMA_ZONE_NUMA) domain = PCPU_GET(domain); else domain = 0; cache = &zone->uz_cpu[curcpu]; if (cache->uc_allocbucket) { if (cache->uc_allocbucket->ub_cnt != 0) LIST_INSERT_HEAD(&zone->uz_domain[domain].uzd_buckets, cache->uc_allocbucket, ub_link); else b1 = cache->uc_allocbucket; cache->uc_allocbucket = NULL; } if (cache->uc_freebucket) { if (cache->uc_freebucket->ub_cnt != 0) LIST_INSERT_HEAD(&zone->uz_domain[domain].uzd_buckets, cache->uc_freebucket, ub_link); else b2 = cache->uc_freebucket; cache->uc_freebucket = NULL; } critical_exit(); ZONE_UNLOCK(zone); if (b1) bucket_free(zone, b1, NULL); if (b2) bucket_free(zone, b2, NULL); } /* * Safely drain per-CPU caches of a zone(s) to alloc bucket. * This is an expensive call because it needs to bind to all CPUs * one by one and enter a critical section on each of them in order * to safely access their cache buckets. * Zone lock must not be held on call this function. */ static void cache_drain_safe(uma_zone_t zone) { int cpu; /* * Polite bucket sizes shrinking was not enouth, shrink aggressively. */ if (zone) cache_shrink(zone); else zone_foreach(cache_shrink); CPU_FOREACH(cpu) { thread_lock(curthread); sched_bind(curthread, cpu); thread_unlock(curthread); if (zone) cache_drain_safe_cpu(zone); else zone_foreach(cache_drain_safe_cpu); } thread_lock(curthread); sched_unbind(curthread); thread_unlock(curthread); } /* * Drain the cached buckets from a zone. Expects a locked zone on entry. */ static void bucket_cache_drain(uma_zone_t zone) { uma_zone_domain_t zdom; uma_bucket_t bucket; int i; /* * Drain the bucket queues and free the buckets. */ for (i = 0; i < vm_ndomains; i++) { zdom = &zone->uz_domain[i]; while ((bucket = LIST_FIRST(&zdom->uzd_buckets)) != NULL) { LIST_REMOVE(bucket, ub_link); ZONE_UNLOCK(zone); bucket_drain(zone, bucket); bucket_free(zone, bucket, NULL); ZONE_LOCK(zone); } } /* * Shrink further bucket sizes. Price of single zone lock collision * is probably lower then price of global cache drain. */ if (zone->uz_count > zone->uz_count_min) zone->uz_count--; } static void keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) { uint8_t *mem; int i; uint8_t flags; CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes", keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera); mem = slab->us_data; flags = slab->us_flags; i = start; if (keg->uk_fini != NULL) { for (i--; i > -1; i--) #ifdef INVARIANTS /* * trash_fini implies that dtor was trash_dtor. trash_fini * would check that memory hasn't been modified since free, * which executed trash_dtor. * That's why we need to run uma_dbg_kskip() check here, * albeit we don't make skip check for other init/fini * invocations. */ if (!uma_dbg_kskip(keg, slab->us_data + (keg->uk_rsize * i)) || keg->uk_fini != trash_fini) #endif keg->uk_fini(slab->us_data + (keg->uk_rsize * i), keg->uk_size); } if (keg->uk_flags & UMA_ZONE_OFFPAGE) zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); uma_total_dec(PAGE_SIZE * keg->uk_ppera); } /* * Frees pages from a keg back to the system. This is done on demand from * the pageout daemon. * * Returns nothing. */ static void keg_drain(uma_keg_t keg) { struct slabhead freeslabs = { 0 }; uma_domain_t dom; uma_slab_t slab, tmp; int i; /* * We don't want to take pages from statically allocated kegs at this * time */ if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) return; CTR3(KTR_UMA, "keg_drain %s(%p) free items: %u", keg->uk_name, keg, keg->uk_free); KEG_LOCK(keg); if (keg->uk_free == 0) goto finished; for (i = 0; i < vm_ndomains; i++) { dom = &keg->uk_domain[i]; LIST_FOREACH_SAFE(slab, &dom->ud_free_slab, us_link, tmp) { /* We have nowhere to free these to. */ if (slab->us_flags & UMA_SLAB_BOOT) continue; LIST_REMOVE(slab, us_link); keg->uk_pages -= keg->uk_ppera; keg->uk_free -= keg->uk_ipers; if (keg->uk_flags & UMA_ZONE_HASH) UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); } } finished: KEG_UNLOCK(keg); while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); keg_free_slab(keg, slab, keg->uk_ipers); } } static void zone_drain_wait(uma_zone_t zone, int waitok) { /* * Set draining to interlock with zone_dtor() so we can release our * locks as we go. Only dtor() should do a WAITOK call since it * is the only call that knows the structure will still be available * when it wakes up. */ ZONE_LOCK(zone); while (zone->uz_flags & UMA_ZFLAG_DRAINING) { if (waitok == M_NOWAIT) goto out; msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); } zone->uz_flags |= UMA_ZFLAG_DRAINING; bucket_cache_drain(zone); ZONE_UNLOCK(zone); /* * The DRAINING flag protects us from being freed while * we're running. Normally the uma_rwlock would protect us but we * must be able to release and acquire the right lock for each keg. */ zone_foreach_keg(zone, &keg_drain); ZONE_LOCK(zone); zone->uz_flags &= ~UMA_ZFLAG_DRAINING; wakeup(zone); out: ZONE_UNLOCK(zone); } void zone_drain(uma_zone_t zone) { zone_drain_wait(zone, M_NOWAIT); } /* * Allocate a new slab for a keg. This does not insert the slab onto a list. * If the allocation was successful, the keg lock will be held upon return, * otherwise the keg will be left unlocked. * * Arguments: * wait Shall we wait? * * Returns: * The slab that was allocated or NULL if there is no memory and the * caller specified M_NOWAIT. */ static uma_slab_t keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int wait) { uma_alloc allocf; uma_slab_t slab; unsigned long size; uint8_t *mem; uint8_t flags; int i; KASSERT(domain >= 0 && domain < vm_ndomains, ("keg_alloc_slab: domain %d out of range", domain)); mtx_assert(&keg->uk_lock, MA_OWNED); allocf = keg->uk_allocf; KEG_UNLOCK(keg); slab = NULL; mem = NULL; if (keg->uk_flags & UMA_ZONE_OFFPAGE) { slab = zone_alloc_item(keg->uk_slabzone, NULL, domain, wait); if (slab == NULL) goto out; } /* * This reproduces the old vm_zone behavior of zero filling pages the * first time they are added to a zone. * * Malloced items are zeroed in uma_zalloc. */ if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) wait |= M_ZERO; else wait &= ~M_ZERO; if (keg->uk_flags & UMA_ZONE_NODUMP) wait |= M_NODUMP; /* zone is passed for legacy reasons. */ size = keg->uk_ppera * PAGE_SIZE; mem = allocf(zone, size, domain, &flags, wait); if (mem == NULL) { if (keg->uk_flags & UMA_ZONE_OFFPAGE) zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); slab = NULL; goto out; } uma_total_inc(size); /* Point the slab into the allocated memory */ if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) slab = (uma_slab_t )(mem + keg->uk_pgoff); if (keg->uk_flags & UMA_ZONE_VTOSLAB) for (i = 0; i < keg->uk_ppera; i++) vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); slab->us_keg = keg; slab->us_data = mem; slab->us_freecount = keg->uk_ipers; slab->us_flags = flags; slab->us_domain = domain; BIT_FILL(SLAB_SETSIZE, &slab->us_free); #ifdef INVARIANTS BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); #endif if (keg->uk_init != NULL) { for (i = 0; i < keg->uk_ipers; i++) if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), keg->uk_size, wait) != 0) break; if (i != keg->uk_ipers) { keg_free_slab(keg, slab, i); slab = NULL; goto out; } } KEG_LOCK(keg); CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)", slab, keg->uk_name, keg); if (keg->uk_flags & UMA_ZONE_HASH) UMA_HASH_INSERT(&keg->uk_hash, slab, mem); keg->uk_pages += keg->uk_ppera; keg->uk_free += keg->uk_ipers; out: return (slab); } /* * This function is intended to be used early on in place of page_alloc() so * that we may use the boot time page cache to satisfy allocations before * the VM is ready. */ static void * startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, int wait) { uma_keg_t keg; void *mem; int pages; keg = zone_first_keg(zone); /* * If we are in BOOT_BUCKETS or higher, than switch to real * allocator. Zones with page sized slabs switch at BOOT_PAGEALLOC. */ switch (booted) { case BOOT_COLD: case BOOT_STRAPPED: break; case BOOT_PAGEALLOC: if (keg->uk_ppera > 1) break; case BOOT_BUCKETS: case BOOT_RUNNING: #ifdef UMA_MD_SMALL_ALLOC keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; #else keg->uk_allocf = page_alloc; #endif return keg->uk_allocf(zone, bytes, domain, pflag, wait); } /* * Check our small startup cache to see if it has pages remaining. */ pages = howmany(bytes, PAGE_SIZE); KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__)); if (pages > boot_pages) panic("UMA zone \"%s\": Increase vm.boot_pages", zone->uz_name); #ifdef DIAGNOSTIC printf("%s from \"%s\", %d boot pages left\n", __func__, zone->uz_name, boot_pages); #endif mem = bootmem; boot_pages -= pages; bootmem += pages * PAGE_SIZE; *pflag = UMA_SLAB_BOOT; return (mem); } /* * Allocates a number of pages from the system * * Arguments: * bytes The number of bytes requested * wait Shall we wait? * * Returns: * A pointer to the alloced memory or possibly * NULL if M_NOWAIT is set. */ static void * page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, int wait) { void *p; /* Returned page */ *pflag = UMA_SLAB_KERNEL; p = (void *) kmem_malloc_domain(domain, bytes, wait); return (p); } static void * pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, int wait) { struct pglist alloctail; vm_offset_t addr, zkva; int cpu, flags; vm_page_t p, p_next; #ifdef NUMA struct pcpu *pc; #endif MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE); TAILQ_INIT(&alloctail); flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | malloc2vm_flags(wait); *pflag = UMA_SLAB_KERNEL; for (cpu = 0; cpu <= mp_maxid; cpu++) { if (CPU_ABSENT(cpu)) { p = vm_page_alloc(NULL, 0, flags); } else { #ifndef NUMA p = vm_page_alloc(NULL, 0, flags); #else pc = pcpu_find(cpu); p = vm_page_alloc_domain(NULL, 0, pc->pc_domain, flags); if (__predict_false(p == NULL)) p = vm_page_alloc(NULL, 0, flags); #endif } if (__predict_false(p == NULL)) goto fail; TAILQ_INSERT_TAIL(&alloctail, p, listq); } if ((addr = kva_alloc(bytes)) == 0) goto fail; zkva = addr; TAILQ_FOREACH(p, &alloctail, listq) { pmap_qenter(zkva, &p, 1); zkva += PAGE_SIZE; } return ((void*)addr); fail: TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { vm_page_unwire(p, PQ_NONE); vm_page_free(p); } return (NULL); } /* * Allocates a number of pages from within an object * * Arguments: * bytes The number of bytes requested * wait Shall we wait? * * Returns: * A pointer to the alloced memory or possibly * NULL if M_NOWAIT is set. */ static void * noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, int wait) { TAILQ_HEAD(, vm_page) alloctail; u_long npages; vm_offset_t retkva, zkva; vm_page_t p, p_next; uma_keg_t keg; TAILQ_INIT(&alloctail); keg = zone_first_keg(zone); npages = howmany(bytes, PAGE_SIZE); while (npages > 0) { p = vm_page_alloc_domain(NULL, 0, domain, VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ | ((wait & M_WAITOK) != 0 ? VM_ALLOC_WAITOK : VM_ALLOC_NOWAIT)); if (p != NULL) { /* * Since the page does not belong to an object, its * listq is unused. */ TAILQ_INSERT_TAIL(&alloctail, p, listq); npages--; continue; } /* * Page allocation failed, free intermediate pages and * exit. */ TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { vm_page_unwire(p, PQ_NONE); vm_page_free(p); } return (NULL); } *flags = UMA_SLAB_PRIV; zkva = keg->uk_kva + atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); retkva = zkva; TAILQ_FOREACH(p, &alloctail, listq) { pmap_qenter(zkva, &p, 1); zkva += PAGE_SIZE; } return ((void *)retkva); } /* * Frees a number of pages to the system * * Arguments: * mem A pointer to the memory to be freed * size The size of the memory being freed * flags The original p->us_flags field * * Returns: * Nothing */ static void page_free(void *mem, vm_size_t size, uint8_t flags) { if ((flags & UMA_SLAB_KERNEL) == 0) panic("UMA: page_free used with invalid flags %x", flags); kmem_free((vm_offset_t)mem, size); } /* * Frees pcpu zone allocations * * Arguments: * mem A pointer to the memory to be freed * size The size of the memory being freed * flags The original p->us_flags field * * Returns: * Nothing */ static void pcpu_page_free(void *mem, vm_size_t size, uint8_t flags) { vm_offset_t sva, curva; vm_paddr_t paddr; vm_page_t m; MPASS(size == (mp_maxid+1)*PAGE_SIZE); sva = (vm_offset_t)mem; for (curva = sva; curva < sva + size; curva += PAGE_SIZE) { paddr = pmap_kextract(curva); m = PHYS_TO_VM_PAGE(paddr); vm_page_unwire(m, PQ_NONE); vm_page_free(m); } pmap_qremove(sva, size >> PAGE_SHIFT); kva_free(sva, size); } /* * Zero fill initializer * * Arguments/Returns follow uma_init specifications */ static int zero_init(void *mem, int size, int flags) { bzero(mem, size); return (0); } /* * Finish creating a small uma keg. This calculates ipers, and the keg size. * * Arguments * keg The zone we should initialize * * Returns * Nothing */ static void keg_small_init(uma_keg_t keg) { u_int rsize; u_int memused; u_int wastedspace; u_int shsize; u_int slabsize; if (keg->uk_flags & UMA_ZONE_PCPU) { u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU; slabsize = UMA_PCPU_ALLOC_SIZE; keg->uk_ppera = ncpus; } else { slabsize = UMA_SLAB_SIZE; keg->uk_ppera = 1; } /* * Calculate the size of each allocation (rsize) according to * alignment. If the requested size is smaller than we have * allocation bits for we round it up. */ rsize = keg->uk_size; if (rsize < slabsize / SLAB_SETSIZE) rsize = slabsize / SLAB_SETSIZE; if (rsize & keg->uk_align) rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); keg->uk_rsize = rsize; KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || keg->uk_rsize < UMA_PCPU_ALLOC_SIZE, ("%s: size %u too large", __func__, keg->uk_rsize)); if (keg->uk_flags & UMA_ZONE_OFFPAGE) shsize = 0; else shsize = sizeof(struct uma_slab); if (rsize <= slabsize - shsize) keg->uk_ipers = (slabsize - shsize) / rsize; else { /* Handle special case when we have 1 item per slab, so * alignment requirement can be relaxed. */ KASSERT(keg->uk_size <= slabsize - shsize, ("%s: size %u greater than slab", __func__, keg->uk_size)); keg->uk_ipers = 1; } KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); memused = keg->uk_ipers * rsize + shsize; wastedspace = slabsize - memused; /* * We can't do OFFPAGE if we're internal or if we've been * asked to not go to the VM for buckets. If we do this we * may end up going to the VM for slabs which we do not * want to do if we're UMA_ZFLAG_CACHEONLY as a result * of UMA_ZONE_VM, which clearly forbids it. */ if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) return; /* * See if using an OFFPAGE slab will limit our waste. Only do * this if it permits more items per-slab. * * XXX We could try growing slabsize to limit max waste as well. * Historically this was not done because the VM could not * efficiently handle contiguous allocations. */ if ((wastedspace >= slabsize / UMA_MAX_WASTE) && (keg->uk_ipers < (slabsize / keg->uk_rsize))) { keg->uk_ipers = slabsize / keg->uk_rsize; KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); CTR6(KTR_UMA, "UMA decided we need offpage slab headers for " "keg: %s(%p), calculated wastedspace = %d, " "maximum wasted space allowed = %d, " "calculated ipers = %d, " "new wasted space = %d\n", keg->uk_name, keg, wastedspace, slabsize / UMA_MAX_WASTE, keg->uk_ipers, slabsize - keg->uk_ipers * keg->uk_rsize); keg->uk_flags |= UMA_ZONE_OFFPAGE; } if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) keg->uk_flags |= UMA_ZONE_HASH; } /* * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be * more complicated. * * Arguments * keg The keg we should initialize * * Returns * Nothing */ static void keg_large_init(uma_keg_t keg) { u_int shsize; KASSERT(keg != NULL, ("Keg is null in keg_large_init")); KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); keg->uk_ipers = 1; keg->uk_rsize = keg->uk_size; /* Check whether we have enough space to not do OFFPAGE. */ if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) { shsize = sizeof(struct uma_slab); if (shsize & UMA_ALIGN_PTR) shsize = (shsize & ~UMA_ALIGN_PTR) + (UMA_ALIGN_PTR + 1); if (PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < shsize) { /* * We can't do OFFPAGE if we're internal, in which case * we need an extra page per allocation to contain the * slab header. */ if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) == 0) keg->uk_flags |= UMA_ZONE_OFFPAGE; else keg->uk_ppera++; } } if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) keg->uk_flags |= UMA_ZONE_HASH; } static void keg_cachespread_init(uma_keg_t keg) { int alignsize; int trailer; int pages; int rsize; KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); alignsize = keg->uk_align + 1; rsize = keg->uk_size; /* * We want one item to start on every align boundary in a page. To * do this we will span pages. We will also extend the item by the * size of align if it is an even multiple of align. Otherwise, it * would fall on the same boundary every time. */ if (rsize & keg->uk_align) rsize = (rsize & ~keg->uk_align) + alignsize; if ((rsize & alignsize) == 0) rsize += alignsize; trailer = rsize - keg->uk_size; pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; pages = MIN(pages, (128 * 1024) / PAGE_SIZE); keg->uk_rsize = rsize; keg->uk_ppera = pages; keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; KASSERT(keg->uk_ipers <= SLAB_SETSIZE, ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, keg->uk_ipers)); } /* * Keg header ctor. This initializes all fields, locks, etc. And inserts * the keg onto the global keg list. * * Arguments/Returns follow uma_ctor specifications * udata Actually uma_kctor_args */ static int keg_ctor(void *mem, int size, void *udata, int flags) { struct uma_kctor_args *arg = udata; uma_keg_t keg = mem; uma_zone_t zone; bzero(keg, size); keg->uk_size = arg->size; keg->uk_init = arg->uminit; keg->uk_fini = arg->fini; keg->uk_align = arg->align; keg->uk_free = 0; keg->uk_reserve = 0; keg->uk_pages = 0; keg->uk_flags = arg->flags; keg->uk_slabzone = NULL; /* * We use a global round-robin policy by default. Zones with * UMA_ZONE_NUMA set will use first-touch instead, in which case the * iterator is never run. */ keg->uk_dr.dr_policy = DOMAINSET_RR(); keg->uk_dr.dr_iter = 0; /* * The master zone is passed to us at keg-creation time. */ zone = arg->zone; keg->uk_name = zone->uz_name; if (arg->flags & UMA_ZONE_VM) keg->uk_flags |= UMA_ZFLAG_CACHEONLY; if (arg->flags & UMA_ZONE_ZINIT) keg->uk_init = zero_init; if (arg->flags & UMA_ZONE_MALLOC) keg->uk_flags |= UMA_ZONE_VTOSLAB; if (arg->flags & UMA_ZONE_PCPU) #ifdef SMP keg->uk_flags |= UMA_ZONE_OFFPAGE; #else keg->uk_flags &= ~UMA_ZONE_PCPU; #endif if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { keg_cachespread_init(keg); } else { if (keg->uk_size > UMA_SLAB_SPACE) keg_large_init(keg); else keg_small_init(keg); } if (keg->uk_flags & UMA_ZONE_OFFPAGE) keg->uk_slabzone = slabzone; /* * If we haven't booted yet we need allocations to go through the * startup cache until the vm is ready. */ if (booted < BOOT_PAGEALLOC) keg->uk_allocf = startup_alloc; #ifdef UMA_MD_SMALL_ALLOC else if (keg->uk_ppera == 1) keg->uk_allocf = uma_small_alloc; #endif else if (keg->uk_flags & UMA_ZONE_PCPU) keg->uk_allocf = pcpu_page_alloc; else keg->uk_allocf = page_alloc; #ifdef UMA_MD_SMALL_ALLOC if (keg->uk_ppera == 1) keg->uk_freef = uma_small_free; else #endif if (keg->uk_flags & UMA_ZONE_PCPU) keg->uk_freef = pcpu_page_free; else keg->uk_freef = page_free; /* * Initialize keg's lock */ KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); /* * If we're putting the slab header in the actual page we need to * figure out where in each page it goes. This calculates a right * justified offset into the memory on an ALIGN_PTR boundary. */ if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { u_int totsize; /* Size of the slab struct and free list */ totsize = sizeof(struct uma_slab); if (totsize & UMA_ALIGN_PTR) totsize = (totsize & ~UMA_ALIGN_PTR) + (UMA_ALIGN_PTR + 1); keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; /* * The only way the following is possible is if with our * UMA_ALIGN_PTR adjustments we are now bigger than * UMA_SLAB_SIZE. I haven't checked whether this is * mathematically possible for all cases, so we make * sure here anyway. */ totsize = keg->uk_pgoff + sizeof(struct uma_slab); if (totsize > PAGE_SIZE * keg->uk_ppera) { printf("zone %s ipers %d rsize %d size %d\n", zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size); panic("UMA slab won't fit."); } } if (keg->uk_flags & UMA_ZONE_HASH) hash_alloc(&keg->uk_hash); CTR5(KTR_UMA, "keg_ctor %p zone %s(%p) out %d free %d\n", keg, zone->uz_name, zone, (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, keg->uk_free); LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); rw_wlock(&uma_rwlock); LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); rw_wunlock(&uma_rwlock); return (0); } /* * Zone header ctor. This initializes all fields, locks, etc. * * Arguments/Returns follow uma_ctor specifications * udata Actually uma_zctor_args */ static int zone_ctor(void *mem, int size, void *udata, int flags) { struct uma_zctor_args *arg = udata; uma_zone_t zone = mem; uma_zone_t z; uma_keg_t keg; bzero(zone, size); zone->uz_name = arg->name; zone->uz_ctor = arg->ctor; zone->uz_dtor = arg->dtor; zone->uz_slab = zone_fetch_slab; zone->uz_init = NULL; zone->uz_fini = NULL; zone->uz_allocs = 0; zone->uz_frees = 0; zone->uz_fails = 0; zone->uz_sleeps = 0; zone->uz_count = 0; zone->uz_count_min = 0; zone->uz_flags = 0; zone->uz_warning = NULL; /* The domain structures follow the cpu structures. */ zone->uz_domain = (struct uma_zone_domain *)&zone->uz_cpu[mp_ncpus]; timevalclear(&zone->uz_ratecheck); keg = arg->keg; ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); /* * This is a pure cache zone, no kegs. */ if (arg->import) { if (arg->flags & UMA_ZONE_VM) arg->flags |= UMA_ZFLAG_CACHEONLY; zone->uz_flags = arg->flags; zone->uz_size = arg->size; zone->uz_import = arg->import; zone->uz_release = arg->release; zone->uz_arg = arg->arg; zone->uz_lockptr = &zone->uz_lock; rw_wlock(&uma_rwlock); LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); rw_wunlock(&uma_rwlock); goto out; } /* * Use the regular zone/keg/slab allocator. */ zone->uz_import = (uma_import)zone_import; zone->uz_release = (uma_release)zone_release; zone->uz_arg = zone; if (arg->flags & UMA_ZONE_SECONDARY) { KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); zone->uz_init = arg->uminit; zone->uz_fini = arg->fini; zone->uz_lockptr = &keg->uk_lock; zone->uz_flags |= UMA_ZONE_SECONDARY; rw_wlock(&uma_rwlock); ZONE_LOCK(zone); LIST_FOREACH(z, &keg->uk_zones, uz_link) { if (LIST_NEXT(z, uz_link) == NULL) { LIST_INSERT_AFTER(z, zone, uz_link); break; } } ZONE_UNLOCK(zone); rw_wunlock(&uma_rwlock); } else if (keg == NULL) { if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, arg->align, arg->flags)) == NULL) return (ENOMEM); } else { struct uma_kctor_args karg; int error; /* We should only be here from uma_startup() */ karg.size = arg->size; karg.uminit = arg->uminit; karg.fini = arg->fini; karg.align = arg->align; karg.flags = arg->flags; karg.zone = zone; error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, flags); if (error) return (error); } /* * Link in the first keg. */ zone->uz_klink.kl_keg = keg; LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); zone->uz_lockptr = &keg->uk_lock; zone->uz_size = keg->uk_size; zone->uz_flags |= (keg->uk_flags & (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); /* * Some internal zones don't have room allocated for the per cpu * caches. If we're internal, bail out here. */ if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, ("Secondary zone requested UMA_ZFLAG_INTERNAL")); return (0); } out: KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) != (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET), ("Invalid zone flag combination")); if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0) zone->uz_count = BUCKET_MAX; else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0) zone->uz_count = 0; else zone->uz_count = bucket_select(zone->uz_size); zone->uz_count_min = zone->uz_count; return (0); } /* * Keg header dtor. This frees all data, destroys locks, frees the hash * table and removes the keg from the global list. * * Arguments/Returns follow uma_dtor specifications * udata unused */ static void keg_dtor(void *arg, int size, void *udata) { uma_keg_t keg; keg = (uma_keg_t)arg; KEG_LOCK(keg); if (keg->uk_free != 0) { printf("Freed UMA keg (%s) was not empty (%d items). " " Lost %d pages of memory.\n", keg->uk_name ? keg->uk_name : "", keg->uk_free, keg->uk_pages); } KEG_UNLOCK(keg); hash_free(&keg->uk_hash); KEG_LOCK_FINI(keg); } /* * Zone header dtor. * * Arguments/Returns follow uma_dtor specifications * udata unused */ static void zone_dtor(void *arg, int size, void *udata) { uma_klink_t klink; uma_zone_t zone; uma_keg_t keg; zone = (uma_zone_t)arg; keg = zone_first_keg(zone); if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) cache_drain(zone); rw_wlock(&uma_rwlock); LIST_REMOVE(zone, uz_link); rw_wunlock(&uma_rwlock); /* * XXX there are some races here where * the zone can be drained but zone lock * released and then refilled before we * remove it... we dont care for now */ zone_drain_wait(zone, M_WAITOK); /* * Unlink all of our kegs. */ while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { klink->kl_keg = NULL; LIST_REMOVE(klink, kl_link); if (klink == &zone->uz_klink) continue; free(klink, M_TEMP); } /* * We only destroy kegs from non secondary zones. */ if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { rw_wlock(&uma_rwlock); LIST_REMOVE(keg, uk_link); rw_wunlock(&uma_rwlock); zone_free_item(kegs, keg, NULL, SKIP_NONE); } ZONE_LOCK_FINI(zone); } /* * Traverses every zone in the system and calls a callback * * Arguments: * zfunc A pointer to a function which accepts a zone * as an argument. * * Returns: * Nothing */ static void zone_foreach(void (*zfunc)(uma_zone_t)) { uma_keg_t keg; uma_zone_t zone; rw_rlock(&uma_rwlock); LIST_FOREACH(keg, &uma_kegs, uk_link) { LIST_FOREACH(zone, &keg->uk_zones, uz_link) zfunc(zone); } rw_runlock(&uma_rwlock); } /* * Count how many pages do we need to bootstrap. VM supplies * its need in early zones in the argument, we add up our zones, * which consist of: UMA Slabs, UMA Hash and 9 Bucket zones. The * zone of zones and zone of kegs are accounted separately. */ #define UMA_BOOT_ZONES 11 /* Zone of zones and zone of kegs have arbitrary alignment. */ #define UMA_BOOT_ALIGN 32 static int zsize, ksize; int uma_startup_count(int vm_zones) { int zones, pages; ksize = sizeof(struct uma_keg) + (sizeof(struct uma_domain) * vm_ndomains); zsize = sizeof(struct uma_zone) + (sizeof(struct uma_cache) * (mp_maxid + 1)) + (sizeof(struct uma_zone_domain) * vm_ndomains); /* * Memory for the zone of kegs and its keg, * and for zone of zones. */ pages = howmany(roundup(zsize, CACHE_LINE_SIZE) * 2 + roundup(ksize, CACHE_LINE_SIZE), PAGE_SIZE); #ifdef UMA_MD_SMALL_ALLOC zones = UMA_BOOT_ZONES; #else zones = UMA_BOOT_ZONES + vm_zones; vm_zones = 0; #endif /* Memory for the rest of startup zones, UMA and VM, ... */ if (zsize > UMA_SLAB_SPACE) pages += (zones + vm_zones) * howmany(roundup2(zsize, UMA_BOOT_ALIGN), UMA_SLAB_SIZE); else if (roundup2(zsize, UMA_BOOT_ALIGN) > UMA_SLAB_SPACE) pages += zones; else pages += howmany(zones, UMA_SLAB_SPACE / roundup2(zsize, UMA_BOOT_ALIGN)); /* ... and their kegs. Note that zone of zones allocates a keg! */ pages += howmany(zones + 1, UMA_SLAB_SPACE / roundup2(ksize, UMA_BOOT_ALIGN)); /* * Most of startup zones are not going to be offpages, that's * why we use UMA_SLAB_SPACE instead of UMA_SLAB_SIZE in all * calculations. Some large bucket zones will be offpage, and * thus will allocate hashes. We take conservative approach * and assume that all zones may allocate hash. This may give * us some positive inaccuracy, usually an extra single page. */ pages += howmany(zones, UMA_SLAB_SPACE / (sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT)); return (pages); } void uma_startup(void *mem, int npages) { struct uma_zctor_args args; uma_keg_t masterkeg; uintptr_t m; #ifdef DIAGNOSTIC printf("Entering %s with %d boot pages configured\n", __func__, npages); #endif rw_init(&uma_rwlock, "UMA lock"); /* Use bootpages memory for the zone of zones and zone of kegs. */ m = (uintptr_t)mem; zones = (uma_zone_t)m; m += roundup(zsize, CACHE_LINE_SIZE); kegs = (uma_zone_t)m; m += roundup(zsize, CACHE_LINE_SIZE); masterkeg = (uma_keg_t)m; m += roundup(ksize, CACHE_LINE_SIZE); m = roundup(m, PAGE_SIZE); npages -= (m - (uintptr_t)mem) / PAGE_SIZE; mem = (void *)m; /* "manually" create the initial zone */ memset(&args, 0, sizeof(args)); args.name = "UMA Kegs"; args.size = ksize; args.ctor = keg_ctor; args.dtor = keg_dtor; args.uminit = zero_init; args.fini = NULL; args.keg = masterkeg; args.align = UMA_BOOT_ALIGN - 1; args.flags = UMA_ZFLAG_INTERNAL; zone_ctor(kegs, zsize, &args, M_WAITOK); bootmem = mem; boot_pages = npages; args.name = "UMA Zones"; args.size = zsize; args.ctor = zone_ctor; args.dtor = zone_dtor; args.uminit = zero_init; args.fini = NULL; args.keg = NULL; args.align = UMA_BOOT_ALIGN - 1; args.flags = UMA_ZFLAG_INTERNAL; zone_ctor(zones, zsize, &args, M_WAITOK); /* Now make a zone for slab headers */ slabzone = uma_zcreate("UMA Slabs", sizeof(struct uma_slab), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); hashzone = uma_zcreate("UMA Hash", sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); bucket_init(); booted = BOOT_STRAPPED; } void uma_startup1(void) { #ifdef DIAGNOSTIC printf("Entering %s with %d boot pages left\n", __func__, boot_pages); #endif booted = BOOT_PAGEALLOC; } void uma_startup2(void) { #ifdef DIAGNOSTIC printf("Entering %s with %d boot pages left\n", __func__, boot_pages); #endif booted = BOOT_BUCKETS; sx_init(&uma_drain_lock, "umadrain"); bucket_enable(); } /* * Initialize our callout handle * */ static void uma_startup3(void) { #ifdef INVARIANTS TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor); uma_dbg_cnt = counter_u64_alloc(M_WAITOK); uma_skip_cnt = counter_u64_alloc(M_WAITOK); #endif callout_init(&uma_callout, 1); callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); booted = BOOT_RUNNING; } static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, int align, uint32_t flags) { struct uma_kctor_args args; args.size = size; args.uminit = uminit; args.fini = fini; args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; args.flags = flags; args.zone = zone; return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK)); } /* Public functions */ /* See uma.h */ void uma_set_align(int align) { if (align != UMA_ALIGN_CACHE) uma_align_cache = align; } /* See uma.h */ uma_zone_t uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, uma_init uminit, uma_fini fini, int align, uint32_t flags) { struct uma_zctor_args args; uma_zone_t res; bool locked; KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"", align, name)); /* This stuff is essential for the zone ctor */ memset(&args, 0, sizeof(args)); args.name = name; args.size = size; args.ctor = ctor; args.dtor = dtor; args.uminit = uminit; args.fini = fini; #ifdef INVARIANTS /* * If a zone is being created with an empty constructor and * destructor, pass UMA constructor/destructor which checks for * memory use after free. */ if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) && ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) { args.ctor = trash_ctor; args.dtor = trash_dtor; args.uminit = trash_init; args.fini = trash_fini; } #endif args.align = align; args.flags = flags; args.keg = NULL; if (booted < BOOT_BUCKETS) { locked = false; } else { sx_slock(&uma_drain_lock); locked = true; } res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); if (locked) sx_sunlock(&uma_drain_lock); return (res); } /* See uma.h */ uma_zone_t uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, uma_init zinit, uma_fini zfini, uma_zone_t master) { struct uma_zctor_args args; uma_keg_t keg; uma_zone_t res; bool locked; keg = zone_first_keg(master); memset(&args, 0, sizeof(args)); args.name = name; args.size = keg->uk_size; args.ctor = ctor; args.dtor = dtor; args.uminit = zinit; args.fini = zfini; args.align = keg->uk_align; args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; args.keg = keg; if (booted < BOOT_BUCKETS) { locked = false; } else { sx_slock(&uma_drain_lock); locked = true; } /* XXX Attaches only one keg of potentially many. */ res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK); if (locked) sx_sunlock(&uma_drain_lock); return (res); } /* See uma.h */ uma_zone_t uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease, void *arg, int flags) { struct uma_zctor_args args; memset(&args, 0, sizeof(args)); args.name = name; args.size = size; args.ctor = ctor; args.dtor = dtor; args.uminit = zinit; args.fini = zfini; args.import = zimport; args.release = zrelease; args.arg = arg; args.align = 0; args.flags = flags; return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK)); } static void zone_lock_pair(uma_zone_t a, uma_zone_t b) { if (a < b) { ZONE_LOCK(a); mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); } else { ZONE_LOCK(b); mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); } } static void zone_unlock_pair(uma_zone_t a, uma_zone_t b) { ZONE_UNLOCK(a); ZONE_UNLOCK(b); } int uma_zsecond_add(uma_zone_t zone, uma_zone_t master) { uma_klink_t klink; uma_klink_t kl; int error; error = 0; klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); zone_lock_pair(zone, master); /* * zone must use vtoslab() to resolve objects and must already be * a secondary. */ if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { error = EINVAL; goto out; } /* * The new master must also use vtoslab(). */ if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { error = EINVAL; goto out; } /* * The underlying object must be the same size. rsize * may be different. */ if (master->uz_size != zone->uz_size) { error = E2BIG; goto out; } /* * Put it at the end of the list. */ klink->kl_keg = zone_first_keg(master); LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { if (LIST_NEXT(kl, kl_link) == NULL) { LIST_INSERT_AFTER(kl, klink, kl_link); break; } } klink = NULL; zone->uz_flags |= UMA_ZFLAG_MULTI; zone->uz_slab = zone_fetch_slab_multi; out: zone_unlock_pair(zone, master); if (klink != NULL) free(klink, M_TEMP); return (error); } /* See uma.h */ void uma_zdestroy(uma_zone_t zone) { sx_slock(&uma_drain_lock); zone_free_item(zones, zone, NULL, SKIP_NONE); sx_sunlock(&uma_drain_lock); } void uma_zwait(uma_zone_t zone) { void *item; item = uma_zalloc_arg(zone, NULL, M_WAITOK); uma_zfree(zone, item); } void * uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags) { void *item; #ifdef SMP int i; MPASS(zone->uz_flags & UMA_ZONE_PCPU); #endif item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO); if (item != NULL && (flags & M_ZERO)) { #ifdef SMP for (i = 0; i <= mp_maxid; i++) bzero(zpcpu_get_cpu(item, i), zone->uz_size); #else bzero(item, zone->uz_size); #endif } return (item); } /* * A stub while both regular and pcpu cases are identical. */ void uma_zfree_pcpu_arg(uma_zone_t zone, void *item, void *udata) { #ifdef SMP MPASS(zone->uz_flags & UMA_ZONE_PCPU); #endif uma_zfree_arg(zone, item, udata); } /* See uma.h */ void * uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) { uma_zone_domain_t zdom; uma_bucket_t bucket; uma_cache_t cache; void *item; int cpu, domain, lockfail; #ifdef INVARIANTS bool skipdbg; #endif /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); /* This is the fast path allocation */ CTR4(KTR_UMA, "uma_zalloc_arg thread %x zone %s(%p) flags %d", curthread, zone->uz_name, zone, flags); if (flags & M_WAITOK) { WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "uma_zalloc_arg: zone \"%s\"", zone->uz_name); } KASSERT((flags & M_EXEC) == 0, ("uma_zalloc_arg: called with M_EXEC")); KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), ("uma_zalloc_arg: called with spinlock or critical section held")); if (zone->uz_flags & UMA_ZONE_PCPU) KASSERT((flags & M_ZERO) == 0, ("allocating from a pcpu zone " "with M_ZERO passed")); #ifdef DEBUG_MEMGUARD if (memguard_cmp_zone(zone)) { item = memguard_alloc(zone->uz_size, flags); if (item != NULL) { if (zone->uz_init != NULL && zone->uz_init(item, zone->uz_size, flags) != 0) return (NULL); if (zone->uz_ctor != NULL && zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { zone->uz_fini(item, zone->uz_size); return (NULL); } return (item); } /* This is unfortunate but should not be fatal. */ } #endif /* * If possible, allocate from the per-CPU cache. There are two * requirements for safe access to the per-CPU cache: (1) the thread * accessing the cache must not be preempted or yield during access, * and (2) the thread must not migrate CPUs without switching which * cache it accesses. We rely on a critical section to prevent * preemption and migration. We release the critical section in * order to acquire the zone mutex if we are unable to allocate from * the current cache; when we re-acquire the critical section, we * must detect and handle migration if it has occurred. */ zalloc_restart: critical_enter(); cpu = curcpu; cache = &zone->uz_cpu[cpu]; zalloc_start: bucket = cache->uc_allocbucket; if (bucket != NULL && bucket->ub_cnt > 0) { bucket->ub_cnt--; item = bucket->ub_bucket[bucket->ub_cnt]; #ifdef INVARIANTS bucket->ub_bucket[bucket->ub_cnt] = NULL; #endif KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); cache->uc_allocs++; critical_exit(); #ifdef INVARIANTS skipdbg = uma_dbg_zskip(zone, item); #endif if (zone->uz_ctor != NULL && #ifdef INVARIANTS (!skipdbg || zone->uz_ctor != trash_ctor || zone->uz_dtor != trash_dtor) && #endif zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { atomic_add_long(&zone->uz_fails, 1); zone_free_item(zone, item, udata, SKIP_DTOR); return (NULL); } #ifdef INVARIANTS if (!skipdbg) uma_dbg_alloc(zone, NULL, item); #endif if (flags & M_ZERO) uma_zero_item(item, zone); return (item); } /* * We have run out of items in our alloc bucket. * See if we can switch with our free bucket. */ bucket = cache->uc_freebucket; if (bucket != NULL && bucket->ub_cnt > 0) { CTR2(KTR_UMA, "uma_zalloc: zone %s(%p) swapping empty with alloc", zone->uz_name, zone); cache->uc_freebucket = cache->uc_allocbucket; cache->uc_allocbucket = bucket; goto zalloc_start; } /* * Discard any empty allocation bucket while we hold no locks. */ bucket = cache->uc_allocbucket; cache->uc_allocbucket = NULL; critical_exit(); if (bucket != NULL) bucket_free(zone, bucket, udata); if (zone->uz_flags & UMA_ZONE_NUMA) { domain = PCPU_GET(domain); if (VM_DOMAIN_EMPTY(domain)) domain = UMA_ANYDOMAIN; } else domain = UMA_ANYDOMAIN; /* Short-circuit for zones without buckets and low memory. */ if (zone->uz_count == 0 || bucketdisable) goto zalloc_item; /* * Attempt to retrieve the item from the per-CPU cache has failed, so * we must go back to the zone. This requires the zone lock, so we * must drop the critical section, then re-acquire it when we go back * to the cache. Since the critical section is released, we may be * preempted or migrate. As such, make sure not to maintain any * thread-local state specific to the cache from prior to releasing * the critical section. */ lockfail = 0; if (ZONE_TRYLOCK(zone) == 0) { /* Record contention to size the buckets. */ ZONE_LOCK(zone); lockfail = 1; } critical_enter(); cpu = curcpu; cache = &zone->uz_cpu[cpu]; /* See if we lost the race to fill the cache. */ if (cache->uc_allocbucket != NULL) { ZONE_UNLOCK(zone); goto zalloc_start; } /* * Check the zone's cache of buckets. */ if (domain == UMA_ANYDOMAIN) zdom = &zone->uz_domain[0]; else zdom = &zone->uz_domain[domain]; if ((bucket = LIST_FIRST(&zdom->uzd_buckets)) != NULL) { KASSERT(bucket->ub_cnt != 0, ("uma_zalloc_arg: Returning an empty bucket.")); LIST_REMOVE(bucket, ub_link); cache->uc_allocbucket = bucket; ZONE_UNLOCK(zone); goto zalloc_start; } /* We are no longer associated with this CPU. */ critical_exit(); /* * We bump the uz count when the cache size is insufficient to * handle the working set. */ if (lockfail && zone->uz_count < BUCKET_MAX) zone->uz_count++; ZONE_UNLOCK(zone); /* * Now lets just fill a bucket and put it on the free list. If that * works we'll restart the allocation from the beginning and it * will use the just filled bucket. */ bucket = zone_alloc_bucket(zone, udata, domain, flags); CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p", zone->uz_name, zone, bucket); if (bucket != NULL) { ZONE_LOCK(zone); critical_enter(); cpu = curcpu; cache = &zone->uz_cpu[cpu]; /* * See if we lost the race or were migrated. Cache the * initialized bucket to make this less likely or claim * the memory directly. */ if (cache->uc_allocbucket == NULL && ((zone->uz_flags & UMA_ZONE_NUMA) == 0 || domain == PCPU_GET(domain))) { cache->uc_allocbucket = bucket; } else if ((zone->uz_flags & UMA_ZONE_NOBUCKETCACHE) != 0) { critical_exit(); ZONE_UNLOCK(zone); bucket_drain(zone, bucket); bucket_free(zone, bucket, udata); goto zalloc_restart; } else LIST_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link); ZONE_UNLOCK(zone); goto zalloc_start; } /* * We may not be able to get a bucket so return an actual item. */ zalloc_item: item = zone_alloc_item(zone, udata, domain, flags); return (item); } void * uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags) { /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); /* This is the fast path allocation */ CTR5(KTR_UMA, "uma_zalloc_domain thread %x zone %s(%p) domain %d flags %d", curthread, zone->uz_name, zone, domain, flags); if (flags & M_WAITOK) { WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "uma_zalloc_domain: zone \"%s\"", zone->uz_name); } KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), ("uma_zalloc_domain: called with spinlock or critical section held")); return (zone_alloc_item(zone, udata, domain, flags)); } /* * Find a slab with some space. Prefer slabs that are partially used over those * that are totally full. This helps to reduce fragmentation. * * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check * only 'domain'. */ static uma_slab_t keg_first_slab(uma_keg_t keg, int domain, bool rr) { uma_domain_t dom; uma_slab_t slab; int start; KASSERT(domain >= 0 && domain < vm_ndomains, ("keg_first_slab: domain %d out of range", domain)); slab = NULL; start = domain; do { dom = &keg->uk_domain[domain]; if (!LIST_EMPTY(&dom->ud_part_slab)) return (LIST_FIRST(&dom->ud_part_slab)); if (!LIST_EMPTY(&dom->ud_free_slab)) { slab = LIST_FIRST(&dom->ud_free_slab); LIST_REMOVE(slab, us_link); LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); return (slab); } if (rr) domain = (domain + 1) % vm_ndomains; } while (domain != start); return (NULL); } static uma_slab_t keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags) { uint32_t reserve; mtx_assert(&keg->uk_lock, MA_OWNED); reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve; if (keg->uk_free <= reserve) return (NULL); return (keg_first_slab(keg, domain, rr)); } static uma_slab_t keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags) { struct vm_domainset_iter di; uma_domain_t dom; uma_slab_t slab; int aflags, domain; bool rr; restart: mtx_assert(&keg->uk_lock, MA_OWNED); /* * Use the keg's policy if upper layers haven't already specified a * domain (as happens with first-touch zones). * * To avoid races we run the iterator with the keg lock held, but that * means that we cannot allow the vm_domainset layer to sleep. Thus, * clear M_WAITOK and handle low memory conditions locally. */ rr = rdomain == UMA_ANYDOMAIN; if (rr) { aflags = (flags & ~M_WAITOK) | M_NOWAIT; vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, &aflags); } else { aflags = flags; domain = rdomain; } for (;;) { slab = keg_fetch_free_slab(keg, domain, rr, flags); if (slab != NULL) { MPASS(slab->us_keg == keg); return (slab); } /* * M_NOVM means don't ask at all! */ if (flags & M_NOVM) break; if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { keg->uk_flags |= UMA_ZFLAG_FULL; /* * If this is not a multi-zone, set the FULL bit. * Otherwise slab_multi() takes care of it. */ if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { zone->uz_flags |= UMA_ZFLAG_FULL; zone_log_warning(zone); zone_maxaction(zone); } if (flags & M_NOWAIT) return (NULL); zone->uz_sleeps++; msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); continue; } slab = keg_alloc_slab(keg, zone, domain, aflags); /* * If we got a slab here it's safe to mark it partially used * and return. We assume that the caller is going to remove * at least one item. */ if (slab) { MPASS(slab->us_keg == keg); dom = &keg->uk_domain[slab->us_domain]; LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); return (slab); } KEG_LOCK(keg); if (rr && vm_domainset_iter_policy(&di, &domain) != 0) { if ((flags & M_WAITOK) != 0) { KEG_UNLOCK(keg); vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask); KEG_LOCK(keg); goto restart; } break; } } /* * We might not have been able to get a slab but another cpu * could have while we were unlocked. Check again before we * fail. */ if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL) { MPASS(slab->us_keg == keg); return (slab); } return (NULL); } static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int domain, int flags) { uma_slab_t slab; if (keg == NULL) { keg = zone_first_keg(zone); KEG_LOCK(keg); } for (;;) { slab = keg_fetch_slab(keg, zone, domain, flags); if (slab) return (slab); if (flags & (M_NOWAIT | M_NOVM)) break; } KEG_UNLOCK(keg); return (NULL); } /* * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns * with the keg locked. On NULL no lock is held. * * The last pointer is used to seed the search. It is not required. */ static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int domain, int rflags) { uma_klink_t klink; uma_slab_t slab; uma_keg_t keg; int flags; int empty; int full; /* * Don't wait on the first pass. This will skip limit tests * as well. We don't want to block if we can find a provider * without blocking. */ flags = (rflags & ~M_WAITOK) | M_NOWAIT; /* * Use the last slab allocated as a hint for where to start * the search. */ if (last != NULL) { slab = keg_fetch_slab(last, zone, domain, flags); if (slab) return (slab); KEG_UNLOCK(last); } /* * Loop until we have a slab incase of transient failures * while M_WAITOK is specified. I'm not sure this is 100% * required but we've done it for so long now. */ for (;;) { empty = 0; full = 0; /* * Search the available kegs for slabs. Be careful to hold the * correct lock while calling into the keg layer. */ LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { keg = klink->kl_keg; KEG_LOCK(keg); if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { slab = keg_fetch_slab(keg, zone, domain, flags); if (slab) return (slab); } if (keg->uk_flags & UMA_ZFLAG_FULL) full++; else empty++; KEG_UNLOCK(keg); } if (rflags & (M_NOWAIT | M_NOVM)) break; flags = rflags; /* * All kegs are full. XXX We can't atomically check all kegs * and sleep so just sleep for a short period and retry. */ if (full && !empty) { ZONE_LOCK(zone); zone->uz_flags |= UMA_ZFLAG_FULL; zone->uz_sleeps++; zone_log_warning(zone); zone_maxaction(zone); msleep(zone, zone->uz_lockptr, PVM, "zonelimit", hz/100); zone->uz_flags &= ~UMA_ZFLAG_FULL; ZONE_UNLOCK(zone); continue; } } return (NULL); } static void * slab_alloc_item(uma_keg_t keg, uma_slab_t slab) { uma_domain_t dom; void *item; uint8_t freei; MPASS(keg == slab->us_keg); mtx_assert(&keg->uk_lock, MA_OWNED); freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); item = slab->us_data + (keg->uk_rsize * freei); slab->us_freecount--; keg->uk_free--; /* Move this slab to the full list */ if (slab->us_freecount == 0) { LIST_REMOVE(slab, us_link); dom = &keg->uk_domain[slab->us_domain]; LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link); } return (item); } static int zone_import(uma_zone_t zone, void **bucket, int max, int domain, int flags) { uma_slab_t slab; uma_keg_t keg; #ifdef NUMA int stripe; #endif int i; slab = NULL; keg = NULL; /* Try to keep the buckets totally full */ for (i = 0; i < max; ) { if ((slab = zone->uz_slab(zone, keg, domain, flags)) == NULL) break; keg = slab->us_keg; #ifdef NUMA stripe = howmany(max, vm_ndomains); #endif while (slab->us_freecount && i < max) { bucket[i++] = slab_alloc_item(keg, slab); if (keg->uk_free <= keg->uk_reserve) break; #ifdef NUMA /* * If the zone is striped we pick a new slab for every * N allocations. Eliminating this conditional will * instead pick a new domain for each bucket rather * than stripe within each bucket. The current option * produces more fragmentation and requires more cpu * time but yields better distribution. */ if ((zone->uz_flags & UMA_ZONE_NUMA) == 0 && vm_ndomains > 1 && --stripe == 0) break; #endif } /* Don't block if we allocated any successfully. */ flags &= ~M_WAITOK; flags |= M_NOWAIT; } if (slab != NULL) KEG_UNLOCK(keg); return i; } static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags) { uma_bucket_t bucket; int max; CTR1(KTR_UMA, "zone_alloc:_bucket domain %d)", domain); /* Don't wait for buckets, preserve caller's NOVM setting. */ bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); if (bucket == NULL) return (NULL); max = MIN(bucket->ub_entries, zone->uz_count); bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, max, domain, flags); /* * Initialize the memory if necessary. */ if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { int i; for (i = 0; i < bucket->ub_cnt; i++) if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, flags) != 0) break; /* * If we couldn't initialize the whole bucket, put the * rest back onto the freelist. */ if (i != bucket->ub_cnt) { zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], bucket->ub_cnt - i); #ifdef INVARIANTS bzero(&bucket->ub_bucket[i], sizeof(void *) * (bucket->ub_cnt - i)); #endif bucket->ub_cnt = i; } } if (bucket->ub_cnt == 0) { bucket_free(zone, bucket, udata); atomic_add_long(&zone->uz_fails, 1); return (NULL); } return (bucket); } /* * Allocates a single item from a zone. * * Arguments * zone The zone to alloc for. * udata The data to be passed to the constructor. * domain The domain to allocate from or UMA_ANYDOMAIN. * flags M_WAITOK, M_NOWAIT, M_ZERO. * * Returns * NULL if there is no memory and M_NOWAIT is set * An item if successful */ static void * zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags) { void *item; #ifdef INVARIANTS bool skipdbg; #endif item = NULL; if (domain != UMA_ANYDOMAIN) { /* avoid allocs targeting empty domains */ if (VM_DOMAIN_EMPTY(domain)) domain = UMA_ANYDOMAIN; } if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1) goto fail; atomic_add_long(&zone->uz_allocs, 1); #ifdef INVARIANTS skipdbg = uma_dbg_zskip(zone, item); #endif /* * We have to call both the zone's init (not the keg's init) * and the zone's ctor. This is because the item is going from * a keg slab directly to the user, and the user is expecting it * to be both zone-init'd as well as zone-ctor'd. */ if (zone->uz_init != NULL) { if (zone->uz_init(item, zone->uz_size, flags) != 0) { zone_free_item(zone, item, udata, SKIP_FINI); goto fail; } } if (zone->uz_ctor != NULL && #ifdef INVARIANTS (!skipdbg || zone->uz_ctor != trash_ctor || zone->uz_dtor != trash_dtor) && #endif zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { zone_free_item(zone, item, udata, SKIP_DTOR); goto fail; } #ifdef INVARIANTS if (!skipdbg) uma_dbg_alloc(zone, NULL, item); #endif if (flags & M_ZERO) uma_zero_item(item, zone); CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item, zone->uz_name, zone); return (item); fail: CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)", zone->uz_name, zone); atomic_add_long(&zone->uz_fails, 1); return (NULL); } /* See uma.h */ void uma_zfree_arg(uma_zone_t zone, void *item, void *udata) { uma_cache_t cache; uma_bucket_t bucket; uma_zone_domain_t zdom; int cpu, domain, lockfail; #ifdef INVARIANTS bool skipdbg; #endif /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, zone->uz_name); KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), ("uma_zfree_arg: called with spinlock or critical section held")); /* uma_zfree(..., NULL) does nothing, to match free(9). */ if (item == NULL) return; #ifdef DEBUG_MEMGUARD if (is_memguard_addr(item)) { if (zone->uz_dtor != NULL) zone->uz_dtor(item, zone->uz_size, udata); if (zone->uz_fini != NULL) zone->uz_fini(item, zone->uz_size); memguard_free(item); return; } #endif #ifdef INVARIANTS skipdbg = uma_dbg_zskip(zone, item); if (skipdbg == false) { if (zone->uz_flags & UMA_ZONE_MALLOC) uma_dbg_free(zone, udata, item); else uma_dbg_free(zone, NULL, item); } if (zone->uz_dtor != NULL && (!skipdbg || zone->uz_dtor != trash_dtor || zone->uz_ctor != trash_ctor)) #else if (zone->uz_dtor != NULL) #endif zone->uz_dtor(item, zone->uz_size, udata); /* * The race here is acceptable. If we miss it we'll just have to wait * a little longer for the limits to be reset. */ if (zone->uz_flags & UMA_ZFLAG_FULL) goto zfree_item; /* * If possible, free to the per-CPU cache. There are two * requirements for safe access to the per-CPU cache: (1) the thread * accessing the cache must not be preempted or yield during access, * and (2) the thread must not migrate CPUs without switching which * cache it accesses. We rely on a critical section to prevent * preemption and migration. We release the critical section in * order to acquire the zone mutex if we are unable to free to the * current cache; when we re-acquire the critical section, we must * detect and handle migration if it has occurred. */ zfree_restart: critical_enter(); cpu = curcpu; cache = &zone->uz_cpu[cpu]; zfree_start: /* * Try to free into the allocbucket first to give LIFO ordering * for cache-hot datastructures. Spill over into the freebucket * if necessary. Alloc will swap them if one runs dry. */ bucket = cache->uc_allocbucket; if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) bucket = cache->uc_freebucket; if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, ("uma_zfree: Freeing to non free bucket index.")); bucket->ub_bucket[bucket->ub_cnt] = item; bucket->ub_cnt++; cache->uc_frees++; critical_exit(); return; } /* * We must go back the zone, which requires acquiring the zone lock, * which in turn means we must release and re-acquire the critical * section. Since the critical section is released, we may be * preempted or migrate. As such, make sure not to maintain any * thread-local state specific to the cache from prior to releasing * the critical section. */ critical_exit(); if (zone->uz_count == 0 || bucketdisable) goto zfree_item; lockfail = 0; if (ZONE_TRYLOCK(zone) == 0) { /* Record contention to size the buckets. */ ZONE_LOCK(zone); lockfail = 1; } critical_enter(); cpu = curcpu; cache = &zone->uz_cpu[cpu]; bucket = cache->uc_freebucket; if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { ZONE_UNLOCK(zone); goto zfree_start; } cache->uc_freebucket = NULL; /* We are no longer associated with this CPU. */ critical_exit(); if ((zone->uz_flags & UMA_ZONE_NUMA) != 0) { domain = PCPU_GET(domain); if (VM_DOMAIN_EMPTY(domain)) domain = UMA_ANYDOMAIN; } else domain = 0; zdom = &zone->uz_domain[0]; /* Can we throw this on the zone full list? */ if (bucket != NULL) { CTR3(KTR_UMA, "uma_zfree: zone %s(%p) putting bucket %p on free list", zone->uz_name, zone, bucket); /* ub_cnt is pointing to the last free item */ KASSERT(bucket->ub_cnt != 0, ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); if ((zone->uz_flags & UMA_ZONE_NOBUCKETCACHE) != 0) { ZONE_UNLOCK(zone); bucket_drain(zone, bucket); bucket_free(zone, bucket, udata); goto zfree_restart; } else LIST_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link); } /* * We bump the uz count when the cache size is insufficient to * handle the working set. */ if (lockfail && zone->uz_count < BUCKET_MAX) zone->uz_count++; ZONE_UNLOCK(zone); bucket = bucket_alloc(zone, udata, M_NOWAIT); CTR3(KTR_UMA, "uma_zfree: zone %s(%p) allocated bucket %p", zone->uz_name, zone, bucket); if (bucket) { critical_enter(); cpu = curcpu; cache = &zone->uz_cpu[cpu]; if (cache->uc_freebucket == NULL && ((zone->uz_flags & UMA_ZONE_NUMA) == 0 || domain == PCPU_GET(domain))) { cache->uc_freebucket = bucket; goto zfree_start; } /* * We lost the race, start over. We have to drop our * critical section to free the bucket. */ critical_exit(); bucket_free(zone, bucket, udata); goto zfree_restart; } /* * If nothing else caught this, we'll just do an internal free. */ zfree_item: zone_free_item(zone, item, udata, SKIP_DTOR); return; } void uma_zfree_domain(uma_zone_t zone, void *item, void *udata) { /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA); CTR2(KTR_UMA, "uma_zfree_domain thread %x zone %s", curthread, zone->uz_name); KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), ("uma_zfree_domain: called with spinlock or critical section held")); /* uma_zfree(..., NULL) does nothing, to match free(9). */ if (item == NULL) return; zone_free_item(zone, item, udata, SKIP_NONE); } static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) { uma_domain_t dom; uint8_t freei; mtx_assert(&keg->uk_lock, MA_OWNED); MPASS(keg == slab->us_keg); dom = &keg->uk_domain[slab->us_domain]; /* Do we need to remove from any lists? */ if (slab->us_freecount+1 == keg->uk_ipers) { LIST_REMOVE(slab, us_link); LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); } else if (slab->us_freecount == 0) { LIST_REMOVE(slab, us_link); LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link); } /* Slab management. */ freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); slab->us_freecount++; /* Keg statistics. */ keg->uk_free++; } static void zone_release(uma_zone_t zone, void **bucket, int cnt) { void *item; uma_slab_t slab; uma_keg_t keg; uint8_t *mem; int clearfull; int i; clearfull = 0; keg = zone_first_keg(zone); KEG_LOCK(keg); for (i = 0; i < cnt; i++) { item = bucket[i]; if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); if (zone->uz_flags & UMA_ZONE_HASH) { slab = hash_sfind(&keg->uk_hash, mem); } else { mem += keg->uk_pgoff; slab = (uma_slab_t)mem; } } else { slab = vtoslab((vm_offset_t)item); if (slab->us_keg != keg) { KEG_UNLOCK(keg); keg = slab->us_keg; KEG_LOCK(keg); } } slab_free_item(keg, slab, item); if (keg->uk_flags & UMA_ZFLAG_FULL) { if (keg->uk_pages < keg->uk_maxpages) { keg->uk_flags &= ~UMA_ZFLAG_FULL; clearfull = 1; } /* * We can handle one more allocation. Since we're * clearing ZFLAG_FULL, wake up all procs blocked * on pages. This should be uncommon, so keeping this * simple for now (rather than adding count of blocked * threads etc). */ wakeup(keg); } } KEG_UNLOCK(keg); if (clearfull) { ZONE_LOCK(zone); zone->uz_flags &= ~UMA_ZFLAG_FULL; wakeup(zone); ZONE_UNLOCK(zone); } } /* * Frees a single item to any zone. * * Arguments: * zone The zone to free to * item The item we're freeing * udata User supplied data for the dtor * skip Skip dtors and finis */ static void zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) { #ifdef INVARIANTS bool skipdbg; skipdbg = uma_dbg_zskip(zone, item); if (skip == SKIP_NONE && !skipdbg) { if (zone->uz_flags & UMA_ZONE_MALLOC) uma_dbg_free(zone, udata, item); else uma_dbg_free(zone, NULL, item); } if (skip < SKIP_DTOR && zone->uz_dtor != NULL && (!skipdbg || zone->uz_dtor != trash_dtor || zone->uz_ctor != trash_ctor)) #else if (skip < SKIP_DTOR && zone->uz_dtor != NULL) #endif zone->uz_dtor(item, zone->uz_size, udata); if (skip < SKIP_FINI && zone->uz_fini) zone->uz_fini(item, zone->uz_size); atomic_add_long(&zone->uz_frees, 1); zone->uz_release(zone->uz_arg, &item, 1); } /* See uma.h */ int uma_zone_set_max(uma_zone_t zone, int nitems) { uma_keg_t keg; keg = zone_first_keg(zone); if (keg == NULL) return (0); KEG_LOCK(keg); keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; if (keg->uk_maxpages * keg->uk_ipers < nitems) keg->uk_maxpages += keg->uk_ppera; nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers; KEG_UNLOCK(keg); return (nitems); } /* See uma.h */ int uma_zone_get_max(uma_zone_t zone) { int nitems; uma_keg_t keg; keg = zone_first_keg(zone); if (keg == NULL) return (0); KEG_LOCK(keg); nitems = (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers; KEG_UNLOCK(keg); return (nitems); } /* See uma.h */ void uma_zone_set_warning(uma_zone_t zone, const char *warning) { ZONE_LOCK(zone); zone->uz_warning = warning; ZONE_UNLOCK(zone); } /* See uma.h */ void uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) { ZONE_LOCK(zone); TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); ZONE_UNLOCK(zone); } /* See uma.h */ int uma_zone_get_cur(uma_zone_t zone) { int64_t nitems; u_int i; ZONE_LOCK(zone); nitems = zone->uz_allocs - zone->uz_frees; CPU_FOREACH(i) { /* * See the comment in sysctl_vm_zone_stats() regarding the * safety of accessing the per-cpu caches. With the zone lock * held, it is safe, but can potentially result in stale data. */ nitems += zone->uz_cpu[i].uc_allocs - zone->uz_cpu[i].uc_frees; } ZONE_UNLOCK(zone); return (nitems < 0 ? 0 : nitems); } /* See uma.h */ void uma_zone_set_init(uma_zone_t zone, uma_init uminit) { uma_keg_t keg; keg = zone_first_keg(zone); KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); KEG_LOCK(keg); KASSERT(keg->uk_pages == 0, ("uma_zone_set_init on non-empty keg")); keg->uk_init = uminit; KEG_UNLOCK(keg); } /* See uma.h */ void uma_zone_set_fini(uma_zone_t zone, uma_fini fini) { uma_keg_t keg; keg = zone_first_keg(zone); KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type")); KEG_LOCK(keg); KASSERT(keg->uk_pages == 0, ("uma_zone_set_fini on non-empty keg")); keg->uk_fini = fini; KEG_UNLOCK(keg); } /* See uma.h */ void uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) { ZONE_LOCK(zone); KASSERT(zone_first_keg(zone)->uk_pages == 0, ("uma_zone_set_zinit on non-empty keg")); zone->uz_init = zinit; ZONE_UNLOCK(zone); } /* See uma.h */ void uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) { ZONE_LOCK(zone); KASSERT(zone_first_keg(zone)->uk_pages == 0, ("uma_zone_set_zfini on non-empty keg")); zone->uz_fini = zfini; ZONE_UNLOCK(zone); } /* See uma.h */ /* XXX uk_freef is not actually used with the zone locked */ void uma_zone_set_freef(uma_zone_t zone, uma_free freef) { uma_keg_t keg; keg = zone_first_keg(zone); KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type")); KEG_LOCK(keg); keg->uk_freef = freef; KEG_UNLOCK(keg); } /* See uma.h */ /* XXX uk_allocf is not actually used with the zone locked */ void uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) { uma_keg_t keg; keg = zone_first_keg(zone); KEG_LOCK(keg); keg->uk_allocf = allocf; KEG_UNLOCK(keg); } /* See uma.h */ void uma_zone_reserve(uma_zone_t zone, int items) { uma_keg_t keg; keg = zone_first_keg(zone); if (keg == NULL) return; KEG_LOCK(keg); keg->uk_reserve = items; KEG_UNLOCK(keg); return; } /* See uma.h */ int uma_zone_reserve_kva(uma_zone_t zone, int count) { uma_keg_t keg; vm_offset_t kva; u_int pages; keg = zone_first_keg(zone); if (keg == NULL) return (0); pages = count / keg->uk_ipers; if (pages * keg->uk_ipers < count) pages++; pages *= keg->uk_ppera; #ifdef UMA_MD_SMALL_ALLOC if (keg->uk_ppera > 1) { #else if (1) { #endif kva = kva_alloc((vm_size_t)pages * PAGE_SIZE); if (kva == 0) return (0); } else kva = 0; KEG_LOCK(keg); keg->uk_kva = kva; keg->uk_offset = 0; keg->uk_maxpages = pages; #ifdef UMA_MD_SMALL_ALLOC keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; #else keg->uk_allocf = noobj_alloc; #endif keg->uk_flags |= UMA_ZONE_NOFREE; KEG_UNLOCK(keg); return (1); } /* See uma.h */ void uma_prealloc(uma_zone_t zone, int items) { + struct vm_domainset_iter di; uma_domain_t dom; uma_slab_t slab; uma_keg_t keg; - int domain, slabs; + int domain, flags, slabs; keg = zone_first_keg(zone); if (keg == NULL) return; KEG_LOCK(keg); slabs = items / keg->uk_ipers; - domain = 0; if (slabs * keg->uk_ipers < items) slabs++; + flags = M_WAITOK; + vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain, &flags); while (slabs-- > 0) { - slab = keg_alloc_slab(keg, zone, domain, M_WAITOK); + slab = keg_alloc_slab(keg, zone, domain, flags); if (slab == NULL) return; MPASS(slab->us_keg == keg); dom = &keg->uk_domain[slab->us_domain]; LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link); - do { - domain = (domain + 1) % vm_ndomains; - } while (VM_DOMAIN_EMPTY(domain)); + if (vm_domainset_iter_policy(&di, &domain) != 0) + break; } KEG_UNLOCK(keg); } /* See uma.h */ static void uma_reclaim_locked(bool kmem_danger) { CTR0(KTR_UMA, "UMA: vm asked us to release pages!"); sx_assert(&uma_drain_lock, SA_XLOCKED); bucket_enable(); zone_foreach(zone_drain); if (vm_page_count_min() || kmem_danger) { cache_drain_safe(NULL); zone_foreach(zone_drain); } /* * Some slabs may have been freed but this zone will be visited early * we visit again so that we can free pages that are empty once other * zones are drained. We have to do the same for buckets. */ zone_drain(slabzone); bucket_zone_drain(); } void uma_reclaim(void) { sx_xlock(&uma_drain_lock); uma_reclaim_locked(false); sx_xunlock(&uma_drain_lock); } static volatile int uma_reclaim_needed; void uma_reclaim_wakeup(void) { if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0) wakeup(uma_reclaim); } void uma_reclaim_worker(void *arg __unused) { for (;;) { sx_xlock(&uma_drain_lock); while (atomic_load_int(&uma_reclaim_needed) == 0) sx_sleep(uma_reclaim, &uma_drain_lock, PVM, "umarcl", hz); sx_xunlock(&uma_drain_lock); EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM); sx_xlock(&uma_drain_lock); uma_reclaim_locked(true); atomic_store_int(&uma_reclaim_needed, 0); sx_xunlock(&uma_drain_lock); /* Don't fire more than once per-second. */ pause("umarclslp", hz); } } /* See uma.h */ int uma_zone_exhausted(uma_zone_t zone) { int full; ZONE_LOCK(zone); full = (zone->uz_flags & UMA_ZFLAG_FULL); ZONE_UNLOCK(zone); return (full); } int uma_zone_exhausted_nolock(uma_zone_t zone) { return (zone->uz_flags & UMA_ZFLAG_FULL); } void * uma_large_malloc_domain(vm_size_t size, int domain, int wait) { vm_offset_t addr; uma_slab_t slab; if (domain != UMA_ANYDOMAIN) { /* avoid allocs targeting empty domains */ if (VM_DOMAIN_EMPTY(domain)) domain = UMA_ANYDOMAIN; } slab = zone_alloc_item(slabzone, NULL, domain, wait); if (slab == NULL) return (NULL); if (domain == UMA_ANYDOMAIN) addr = kmem_malloc(size, wait); else addr = kmem_malloc_domain(domain, size, wait); if (addr != 0) { vsetslab(addr, slab); slab->us_data = (void *)addr; slab->us_flags = UMA_SLAB_KERNEL | UMA_SLAB_MALLOC; slab->us_size = size; slab->us_domain = vm_phys_domain(PHYS_TO_VM_PAGE( pmap_kextract(addr))); uma_total_inc(size); } else { zone_free_item(slabzone, slab, NULL, SKIP_NONE); } return ((void *)addr); } void * uma_large_malloc(vm_size_t size, int wait) { return uma_large_malloc_domain(size, UMA_ANYDOMAIN, wait); } void uma_large_free(uma_slab_t slab) { KASSERT((slab->us_flags & UMA_SLAB_KERNEL) != 0, ("uma_large_free: Memory not allocated with uma_large_malloc.")); kmem_free((vm_offset_t)slab->us_data, slab->us_size); uma_total_dec(slab->us_size); zone_free_item(slabzone, slab, NULL, SKIP_NONE); } static void uma_zero_item(void *item, uma_zone_t zone) { bzero(item, zone->uz_size); } unsigned long uma_limit(void) { return (uma_kmem_limit); } void uma_set_limit(unsigned long limit) { uma_kmem_limit = limit; } unsigned long uma_size(void) { return (uma_kmem_total); } long uma_avail(void) { return (uma_kmem_limit - uma_kmem_total); } void uma_print_stats(void) { zone_foreach(uma_print_zone); } static void slab_print(uma_slab_t slab) { printf("slab: keg %p, data %p, freecount %d\n", slab->us_keg, slab->us_data, slab->us_freecount); } static void cache_print(uma_cache_t cache) { printf("alloc: %p(%d), free: %p(%d)\n", cache->uc_allocbucket, cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, cache->uc_freebucket, cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); } static void uma_print_keg(uma_keg_t keg) { uma_domain_t dom; uma_slab_t slab; int i; printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " "out %d free %d limit %d\n", keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, keg->uk_ipers, keg->uk_ppera, (keg->uk_pages / keg->uk_ppera) * keg->uk_ipers - keg->uk_free, keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); for (i = 0; i < vm_ndomains; i++) { dom = &keg->uk_domain[i]; printf("Part slabs:\n"); LIST_FOREACH(slab, &dom->ud_part_slab, us_link) slab_print(slab); printf("Free slabs:\n"); LIST_FOREACH(slab, &dom->ud_free_slab, us_link) slab_print(slab); printf("Full slabs:\n"); LIST_FOREACH(slab, &dom->ud_full_slab, us_link) slab_print(slab); } } void uma_print_zone(uma_zone_t zone) { uma_cache_t cache; uma_klink_t kl; int i; printf("zone: %s(%p) size %d flags %#x\n", zone->uz_name, zone, zone->uz_size, zone->uz_flags); LIST_FOREACH(kl, &zone->uz_kegs, kl_link) uma_print_keg(kl->kl_keg); CPU_FOREACH(i) { cache = &zone->uz_cpu[i]; printf("CPU %d Cache:\n", i); cache_print(cache); } } #ifdef DDB /* * Generate statistics across both the zone and its per-cpu cache's. Return * desired statistics if the pointer is non-NULL for that statistic. * * Note: does not update the zone statistics, as it can't safely clear the * per-CPU cache statistic. * * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't * safe from off-CPU; we should modify the caches to track this information * directly so that we don't have to. */ static void uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, uint64_t *freesp, uint64_t *sleepsp) { uma_cache_t cache; uint64_t allocs, frees, sleeps; int cachefree, cpu; allocs = frees = sleeps = 0; cachefree = 0; CPU_FOREACH(cpu) { cache = &z->uz_cpu[cpu]; if (cache->uc_allocbucket != NULL) cachefree += cache->uc_allocbucket->ub_cnt; if (cache->uc_freebucket != NULL) cachefree += cache->uc_freebucket->ub_cnt; allocs += cache->uc_allocs; frees += cache->uc_frees; } allocs += z->uz_allocs; frees += z->uz_frees; sleeps += z->uz_sleeps; if (cachefreep != NULL) *cachefreep = cachefree; if (allocsp != NULL) *allocsp = allocs; if (freesp != NULL) *freesp = frees; if (sleepsp != NULL) *sleepsp = sleeps; } #endif /* DDB */ static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) { uma_keg_t kz; uma_zone_t z; int count; count = 0; rw_rlock(&uma_rwlock); LIST_FOREACH(kz, &uma_kegs, uk_link) { LIST_FOREACH(z, &kz->uk_zones, uz_link) count++; } rw_runlock(&uma_rwlock); return (sysctl_handle_int(oidp, &count, 0, req)); } static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) { struct uma_stream_header ush; struct uma_type_header uth; struct uma_percpu_stat *ups; uma_bucket_t bucket; uma_zone_domain_t zdom; struct sbuf sbuf; uma_cache_t cache; uma_klink_t kl; uma_keg_t kz; uma_zone_t z; uma_keg_t k; int count, error, i; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128, req); sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK); count = 0; rw_rlock(&uma_rwlock); LIST_FOREACH(kz, &uma_kegs, uk_link) { LIST_FOREACH(z, &kz->uk_zones, uz_link) count++; } /* * Insert stream header. */ bzero(&ush, sizeof(ush)); ush.ush_version = UMA_STREAM_VERSION; ush.ush_maxcpus = (mp_maxid + 1); ush.ush_count = count; (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); LIST_FOREACH(kz, &uma_kegs, uk_link) { LIST_FOREACH(z, &kz->uk_zones, uz_link) { bzero(&uth, sizeof(uth)); ZONE_LOCK(z); strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); uth.uth_align = kz->uk_align; uth.uth_size = kz->uk_size; uth.uth_rsize = kz->uk_rsize; LIST_FOREACH(kl, &z->uz_kegs, kl_link) { k = kl->kl_keg; uth.uth_maxpages += k->uk_maxpages; uth.uth_pages += k->uk_pages; uth.uth_keg_free += k->uk_free; uth.uth_limit = (k->uk_maxpages / k->uk_ppera) * k->uk_ipers; } /* * A zone is secondary is it is not the first entry * on the keg's zone list. */ if ((z->uz_flags & UMA_ZONE_SECONDARY) && (LIST_FIRST(&kz->uk_zones) != z)) uth.uth_zone_flags = UTH_ZONE_SECONDARY; for (i = 0; i < vm_ndomains; i++) { zdom = &z->uz_domain[i]; LIST_FOREACH(bucket, &zdom->uzd_buckets, ub_link) uth.uth_zone_free += bucket->ub_cnt; } uth.uth_allocs = z->uz_allocs; uth.uth_frees = z->uz_frees; uth.uth_fails = z->uz_fails; uth.uth_sleeps = z->uz_sleeps; /* * While it is not normally safe to access the cache * bucket pointers while not on the CPU that owns the * cache, we only allow the pointers to be exchanged * without the zone lock held, not invalidated, so * accept the possible race associated with bucket * exchange during monitoring. */ for (i = 0; i < mp_maxid + 1; i++) { bzero(&ups[i], sizeof(*ups)); if (kz->uk_flags & UMA_ZFLAG_INTERNAL || CPU_ABSENT(i)) continue; cache = &z->uz_cpu[i]; if (cache->uc_allocbucket != NULL) ups[i].ups_cache_free += cache->uc_allocbucket->ub_cnt; if (cache->uc_freebucket != NULL) ups[i].ups_cache_free += cache->uc_freebucket->ub_cnt; ups[i].ups_allocs = cache->uc_allocs; ups[i].ups_frees = cache->uc_frees; } ZONE_UNLOCK(z); (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); for (i = 0; i < mp_maxid + 1; i++) (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i])); } } rw_runlock(&uma_rwlock); error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); free(ups, M_TEMP); return (error); } int sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) { uma_zone_t zone = *(uma_zone_t *)arg1; int error, max; max = uma_zone_get_max(zone); error = sysctl_handle_int(oidp, &max, 0, req); if (error || !req->newptr) return (error); uma_zone_set_max(zone, max); return (0); } int sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) { uma_zone_t zone = *(uma_zone_t *)arg1; int cur; cur = uma_zone_get_cur(zone); return (sysctl_handle_int(oidp, &cur, 0, req)); } #ifdef INVARIANTS static uma_slab_t uma_dbg_getslab(uma_zone_t zone, void *item) { uma_slab_t slab; uma_keg_t keg; uint8_t *mem; mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); if (zone->uz_flags & UMA_ZONE_VTOSLAB) { slab = vtoslab((vm_offset_t)mem); } else { /* * It is safe to return the slab here even though the * zone is unlocked because the item's allocation state * essentially holds a reference. */ ZONE_LOCK(zone); keg = LIST_FIRST(&zone->uz_kegs)->kl_keg; if (keg->uk_flags & UMA_ZONE_HASH) slab = hash_sfind(&keg->uk_hash, mem); else slab = (uma_slab_t)(mem + keg->uk_pgoff); ZONE_UNLOCK(zone); } return (slab); } static bool uma_dbg_zskip(uma_zone_t zone, void *mem) { uma_keg_t keg; if ((keg = zone_first_keg(zone)) == NULL) return (true); return (uma_dbg_kskip(keg, mem)); } static bool uma_dbg_kskip(uma_keg_t keg, void *mem) { uintptr_t idx; if (dbg_divisor == 0) return (true); if (dbg_divisor == 1) return (false); idx = (uintptr_t)mem >> PAGE_SHIFT; if (keg->uk_ipers > 1) { idx *= keg->uk_ipers; idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize; } if ((idx / dbg_divisor) * dbg_divisor != idx) { counter_u64_add(uma_skip_cnt, 1); return (true); } counter_u64_add(uma_dbg_cnt, 1); return (false); } /* * Set up the slab's freei data such that uma_dbg_free can function. * */ static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) { uma_keg_t keg; int freei; if (slab == NULL) { slab = uma_dbg_getslab(zone, item); if (slab == NULL) panic("uma: item %p did not belong to zone %s\n", item, zone->uz_name); } keg = slab->us_keg; freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n", item, zone, zone->uz_name, slab, freei); BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); return; } /* * Verifies freed addresses. Checks for alignment, valid slab membership * and duplicate frees. * */ static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) { uma_keg_t keg; int freei; if (slab == NULL) { slab = uma_dbg_getslab(zone, item); if (slab == NULL) panic("uma: Freed item %p did not belong to zone %s\n", item, zone->uz_name); } keg = slab->us_keg; freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; if (freei >= keg->uk_ipers) panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n", item, zone, zone->uz_name, slab, freei); if (((freei * keg->uk_rsize) + slab->us_data) != item) panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n", item, zone, zone->uz_name, slab, freei); if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n", item, zone, zone->uz_name, slab, freei); BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); } #endif /* INVARIANTS */ #ifdef DDB DB_SHOW_COMMAND(uma, db_show_uma) { uma_bucket_t bucket; uma_keg_t kz; uma_zone_t z; uma_zone_domain_t zdom; uint64_t allocs, frees, sleeps; int cachefree, i; db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used", "Free", "Requests", "Sleeps", "Bucket"); LIST_FOREACH(kz, &uma_kegs, uk_link) { LIST_FOREACH(z, &kz->uk_zones, uz_link) { if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { allocs = z->uz_allocs; frees = z->uz_frees; sleeps = z->uz_sleeps; cachefree = 0; } else uma_zone_sumstat(z, &cachefree, &allocs, &frees, &sleeps); if (!((z->uz_flags & UMA_ZONE_SECONDARY) && (LIST_FIRST(&kz->uk_zones) != z))) cachefree += kz->uk_free; for (i = 0; i < vm_ndomains; i++) { zdom = &z->uz_domain[i]; LIST_FOREACH(bucket, &zdom->uzd_buckets, ub_link) cachefree += bucket->ub_cnt; } db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n", z->uz_name, (uintmax_t)kz->uk_size, (intmax_t)(allocs - frees), cachefree, (uintmax_t)allocs, sleeps, z->uz_count); if (db_pager_quit) return; } } } DB_SHOW_COMMAND(umacache, db_show_umacache) { uma_bucket_t bucket; uma_zone_t z; uma_zone_domain_t zdom; uint64_t allocs, frees; int cachefree, i; db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", "Requests", "Bucket"); LIST_FOREACH(z, &uma_cachezones, uz_link) { uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL); for (i = 0; i < vm_ndomains; i++) { zdom = &z->uz_domain[i]; LIST_FOREACH(bucket, &zdom->uzd_buckets, ub_link) cachefree += bucket->ub_cnt; } db_printf("%18s %8ju %8jd %8d %12ju %8u\n", z->uz_name, (uintmax_t)z->uz_size, (intmax_t)(allocs - frees), cachefree, (uintmax_t)allocs, z->uz_count); if (db_pager_quit) return; } } #endif /* DDB */ Index: head/sys/vm/vm_init.c =================================================================== --- head/sys/vm/vm_init.c (revision 339924) +++ head/sys/vm/vm_init.c (revision 339925) @@ -1,274 +1,280 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. 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. 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_init.c 8.1 (Berkeley) 6/11/93 * * * 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. */ /* * Initialize the Virtual Memory subsystem. */ #include __FBSDID("$FreeBSD$"); #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 extern void uma_startup1(void); extern void uma_startup2(void); extern void vm_radix_reserve_kva(void); long physmem; /* * System initialization */ static void vm_mem_init(void *); SYSINIT(vm_mem, SI_SUB_VM, SI_ORDER_FIRST, vm_mem_init, NULL); /* * vm_init initializes the virtual memory system. * This is done only by the first cpu up. * * The start and end address of physical memory is passed in. */ static void vm_mem_init(void *dummy) { /* * Initialize static domainsets, used by various allocators. */ domainset_init(); /* - * Initializes resident memory structures. From here on, all physical + * Initialize resident memory structures. From here on, all physical * memory is accounted for, and we use only virtual addresses. */ vm_set_page_size(); virtual_avail = vm_page_startup(virtual_avail); + + /* + * Set an initial domain policy for thread0 so that allocations + * can work. + */ + domainset_zero(); #ifdef UMA_MD_SMALL_ALLOC /* Announce page availability to UMA. */ uma_startup1(); #endif /* * Initialize other VM packages */ vmem_startup(); vm_object_init(); vm_map_startup(); kmem_init(virtual_avail, virtual_end); #ifndef UMA_MD_SMALL_ALLOC /* Set up radix zone to use noobj_alloc. */ vm_radix_reserve_kva(); #endif /* Announce full page availability to UMA. */ uma_startup2(); kmem_init_zero_region(); pmap_init(); vm_pager_init(); } void vm_ksubmap_init(struct kva_md_info *kmi) { vm_offset_t firstaddr; caddr_t v; vm_size_t size = 0; long physmem_est; vm_offset_t minaddr; vm_offset_t maxaddr; /* * Allocate space for system data structures. * The first available kernel virtual address is in "v". * As pages of kernel virtual memory are allocated, "v" is incremented. * As pages of memory are allocated and cleared, * "firstaddr" is incremented. */ /* * Make two passes. The first pass calculates how much memory is * needed and allocates it. The second pass assigns virtual * addresses to the various data structures. */ firstaddr = 0; again: v = (caddr_t)firstaddr; /* * Discount the physical memory larger than the size of kernel_map * to avoid eating up all of KVA space. */ physmem_est = lmin(physmem, btoc(vm_map_max(kernel_map) - vm_map_min(kernel_map))); v = kern_vfs_bio_buffer_alloc(v, physmem_est); /* * End of first pass, size has been calculated so allocate memory */ if (firstaddr == 0) { size = (vm_size_t)v; #ifdef VM_FREELIST_DMA32 /* * Try to protect 32-bit DMAable memory from the largest * early alloc of wired mem. */ firstaddr = kmem_alloc_attr(size, M_ZERO | M_NOWAIT, (vm_paddr_t)1 << 32, ~(vm_paddr_t)0, VM_MEMATTR_DEFAULT); if (firstaddr == 0) #endif firstaddr = kmem_malloc(size, M_ZERO | M_WAITOK); if (firstaddr == 0) panic("startup: no room for tables"); goto again; } /* * End of second pass, addresses have been assigned */ if ((vm_size_t)((char *)v - firstaddr) != size) panic("startup: table size inconsistency"); /* * Allocate the clean map to hold all of the paging and I/O virtual * memory. */ size = (long)nbuf * BKVASIZE + (long)nswbuf * MAXPHYS + (long)bio_transient_maxcnt * MAXPHYS; kmi->clean_sva = firstaddr = kva_alloc(size); kmi->clean_eva = firstaddr + size; /* * Allocate the buffer arena. * * Enable the quantum cache if we have more than 4 cpus. This * avoids lock contention at the expense of some fragmentation. */ size = (long)nbuf * BKVASIZE; kmi->buffer_sva = firstaddr; kmi->buffer_eva = kmi->buffer_sva + size; vmem_init(buffer_arena, "buffer arena", kmi->buffer_sva, size, PAGE_SIZE, (mp_ncpus > 4) ? BKVASIZE * 8 : 0, 0); firstaddr += size; /* * Now swap kva. */ swapbkva = firstaddr; size = (long)nswbuf * MAXPHYS; firstaddr += size; /* * And optionally transient bio space. */ if (bio_transient_maxcnt != 0) { size = (long)bio_transient_maxcnt * MAXPHYS; vmem_init(transient_arena, "transient arena", firstaddr, size, PAGE_SIZE, 0, 0); firstaddr += size; } if (firstaddr != kmi->clean_eva) panic("Clean map calculation incorrect"); /* * Allocate the pageable submaps. We may cache an exec map entry per * CPU, so we therefore need to reserve space for at least ncpu+1 * entries to avoid deadlock. The exec map is also used by some image * activators, so we leave a fixed number of pages for their use. */ #ifdef __LP64__ exec_map_entries = 8 * mp_ncpus; #else exec_map_entries = 2 * mp_ncpus + 4; #endif exec_map_entry_size = round_page(PATH_MAX + ARG_MAX); exec_map = kmem_suballoc(kernel_map, &minaddr, &maxaddr, exec_map_entries * exec_map_entry_size + 64 * PAGE_SIZE, FALSE); pipe_map = kmem_suballoc(kernel_map, &minaddr, &maxaddr, maxpipekva, FALSE); } Index: head/sys/vm/vm_kern.c =================================================================== --- head/sys/vm/vm_kern.c (revision 339924) +++ head/sys/vm/vm_kern.c (revision 339925) @@ -1,837 +1,836 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. 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. 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_kern.c 8.3 (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. */ /* * Kernel memory management. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include /* for ticks and hz */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include vm_map_t kernel_map; vm_map_t exec_map; vm_map_t pipe_map; const void *zero_region; CTASSERT((ZERO_REGION_SIZE & PAGE_MASK) == 0); /* NB: Used by kernel debuggers. */ const u_long vm_maxuser_address = VM_MAXUSER_ADDRESS; u_int exec_map_entry_size; u_int exec_map_entries; SYSCTL_ULONG(_vm, OID_AUTO, min_kernel_address, CTLFLAG_RD, SYSCTL_NULL_ULONG_PTR, VM_MIN_KERNEL_ADDRESS, "Min kernel address"); SYSCTL_ULONG(_vm, OID_AUTO, max_kernel_address, CTLFLAG_RD, #if defined(__arm__) || defined(__sparc64__) &vm_max_kernel_address, 0, #else SYSCTL_NULL_ULONG_PTR, VM_MAX_KERNEL_ADDRESS, #endif "Max kernel address"); #if VM_NRESERVLEVEL > 0 #define KVA_QUANTUM_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT) #else /* On non-superpage architectures want large import sizes. */ #define KVA_QUANTUM_SHIFT (10 + PAGE_SHIFT) #endif #define KVA_QUANTUM (1 << KVA_QUANTUM_SHIFT) /* * kva_alloc: * * Allocate a virtual address range with no underlying object and * no initial mapping to physical memory. Any mapping from this * range to physical memory must be explicitly created prior to * its use, typically with pmap_qenter(). Any attempt to create * a mapping on demand through vm_fault() will result in a panic. */ vm_offset_t kva_alloc(vm_size_t size) { vm_offset_t addr; size = round_page(size); if (vmem_alloc(kernel_arena, size, M_BESTFIT | M_NOWAIT, &addr)) return (0); return (addr); } /* * kva_free: * * Release a region of kernel virtual memory allocated * with kva_alloc, and return the physical pages * associated with that region. * * This routine may not block on kernel maps. */ void kva_free(vm_offset_t addr, vm_size_t size) { size = round_page(size); vmem_free(kernel_arena, addr, size); } /* * Allocates a region from the kernel address map and physical pages * within the specified address range to the kernel object. Creates a * wired mapping from this region to these pages, and returns the * region's starting virtual address. The allocated pages are not * necessarily physically contiguous. If M_ZERO is specified through the * given flags, then the pages are zeroed before they are mapped. */ vm_offset_t kmem_alloc_attr_domain(int domain, vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr) { vmem_t *vmem; vm_object_t object = kernel_object; vm_offset_t addr, i, offset; vm_page_t m; int pflags, tries; size = round_page(size); vmem = vm_dom[domain].vmd_kernel_arena; if (vmem_alloc(vmem, size, M_BESTFIT | flags, &addr)) return (0); offset = addr - VM_MIN_KERNEL_ADDRESS; pflags = malloc2vm_flags(flags) | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED; pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); pflags |= VM_ALLOC_NOWAIT; VM_OBJECT_WLOCK(object); for (i = 0; i < size; i += PAGE_SIZE) { tries = 0; retry: m = vm_page_alloc_contig_domain(object, atop(offset + i), domain, pflags, 1, low, high, PAGE_SIZE, 0, memattr); if (m == NULL) { VM_OBJECT_WUNLOCK(object); if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) { if (!vm_page_reclaim_contig_domain(domain, pflags, 1, low, high, PAGE_SIZE, 0) && (flags & M_WAITOK) != 0) vm_wait_domain(domain); VM_OBJECT_WLOCK(object); tries++; goto retry; } kmem_unback(object, addr, i); vmem_free(vmem, addr, size); return (0); } KASSERT(vm_phys_domain(m) == domain, ("kmem_alloc_attr_domain: Domain mismatch %d != %d", vm_phys_domain(m), domain)); if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); m->valid = VM_PAGE_BITS_ALL; pmap_enter(kernel_pmap, addr + i, m, VM_PROT_RW, VM_PROT_RW | PMAP_ENTER_WIRED, 0); } VM_OBJECT_WUNLOCK(object); return (addr); } vm_offset_t kmem_alloc_attr(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, vm_memattr_t memattr) { struct vm_domainset_iter di; vm_offset_t addr; int domain; vm_domainset_iter_policy_init(&di, DOMAINSET_RR(), &domain, &flags); do { addr = kmem_alloc_attr_domain(domain, size, flags, low, high, memattr); if (addr != 0) break; } while (vm_domainset_iter_policy(&di, &domain) == 0); return (addr); } /* * Allocates a region from the kernel address map and physically * contiguous pages within the specified address range to the kernel * object. Creates a wired mapping from this region to these pages, and * returns the region's starting virtual address. If M_ZERO is specified * through the given flags, then the pages are zeroed before they are * mapped. */ vm_offset_t kmem_alloc_contig_domain(int domain, vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { vmem_t *vmem; vm_object_t object = kernel_object; vm_offset_t addr, offset, tmp; vm_page_t end_m, m; u_long npages; int pflags, tries; size = round_page(size); vmem = vm_dom[domain].vmd_kernel_arena; if (vmem_alloc(vmem, size, flags | M_BESTFIT, &addr)) return (0); offset = addr - VM_MIN_KERNEL_ADDRESS; pflags = malloc2vm_flags(flags) | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED; pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); pflags |= VM_ALLOC_NOWAIT; npages = atop(size); VM_OBJECT_WLOCK(object); tries = 0; retry: m = vm_page_alloc_contig_domain(object, atop(offset), domain, pflags, npages, low, high, alignment, boundary, memattr); if (m == NULL) { VM_OBJECT_WUNLOCK(object); if (tries < ((flags & M_NOWAIT) != 0 ? 1 : 3)) { if (!vm_page_reclaim_contig_domain(domain, pflags, npages, low, high, alignment, boundary) && (flags & M_WAITOK) != 0) vm_wait_domain(domain); VM_OBJECT_WLOCK(object); tries++; goto retry; } vmem_free(vmem, addr, size); return (0); } KASSERT(vm_phys_domain(m) == domain, ("kmem_alloc_contig_domain: Domain mismatch %d != %d", vm_phys_domain(m), domain)); end_m = m + npages; tmp = addr; for (; m < end_m; m++) { if ((flags & M_ZERO) && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); m->valid = VM_PAGE_BITS_ALL; pmap_enter(kernel_pmap, tmp, m, VM_PROT_RW, VM_PROT_RW | PMAP_ENTER_WIRED, 0); tmp += PAGE_SIZE; } VM_OBJECT_WUNLOCK(object); return (addr); } vm_offset_t kmem_alloc_contig(vm_size_t size, int flags, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domainset_iter di; vm_offset_t addr; int domain; vm_domainset_iter_policy_init(&di, DOMAINSET_RR(), &domain, &flags); do { addr = kmem_alloc_contig_domain(domain, size, flags, low, high, alignment, boundary, memattr); if (addr != 0) break; } while (vm_domainset_iter_policy(&di, &domain) == 0); return (addr); } /* * kmem_suballoc: * * Allocates a map to manage a subrange * of the kernel virtual address space. * * Arguments are as follows: * * parent Map to take range from * min, max Returned endpoints of map * size Size of range to find * superpage_align Request that min is superpage aligned */ vm_map_t kmem_suballoc(vm_map_t parent, vm_offset_t *min, vm_offset_t *max, vm_size_t size, boolean_t superpage_align) { int ret; vm_map_t result; size = round_page(size); *min = vm_map_min(parent); ret = vm_map_find(parent, NULL, 0, min, size, 0, superpage_align ? VMFS_SUPER_SPACE : VMFS_ANY_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_ACC_NO_CHARGE); if (ret != KERN_SUCCESS) panic("kmem_suballoc: bad status return of %d", ret); *max = *min + size; result = vm_map_create(vm_map_pmap(parent), *min, *max); if (result == NULL) panic("kmem_suballoc: cannot create submap"); if (vm_map_submap(parent, *min, *max, result) != KERN_SUCCESS) panic("kmem_suballoc: unable to change range to submap"); return (result); } /* * kmem_malloc: * * Allocate wired-down pages in the kernel's address space. */ vm_offset_t kmem_malloc_domain(int domain, vm_size_t size, int flags) { vmem_t *arena; vm_offset_t addr; int rv; #if VM_NRESERVLEVEL > 0 if (__predict_true((flags & M_EXEC) == 0)) arena = vm_dom[domain].vmd_kernel_arena; else arena = vm_dom[domain].vmd_kernel_rwx_arena; #else arena = vm_dom[domain].vmd_kernel_arena; #endif size = round_page(size); if (vmem_alloc(arena, size, flags | M_BESTFIT, &addr)) return (0); rv = kmem_back_domain(domain, kernel_object, addr, size, flags); if (rv != KERN_SUCCESS) { vmem_free(arena, addr, size); return (0); } return (addr); } vm_offset_t kmem_malloc(vm_size_t size, int flags) { struct vm_domainset_iter di; vm_offset_t addr; int domain; vm_domainset_iter_policy_init(&di, DOMAINSET_RR(), &domain, &flags); do { addr = kmem_malloc_domain(domain, size, flags); if (addr != 0) break; } while (vm_domainset_iter_policy(&di, &domain) == 0); return (addr); } /* * kmem_back_domain: * * Allocate physical pages from the specified domain for the specified * virtual address range. */ int kmem_back_domain(int domain, vm_object_t object, vm_offset_t addr, vm_size_t size, int flags) { vm_offset_t offset, i; vm_page_t m, mpred; vm_prot_t prot; int pflags; KASSERT(object == kernel_object, ("kmem_back_domain: only supports kernel object.")); offset = addr - VM_MIN_KERNEL_ADDRESS; pflags = malloc2vm_flags(flags) | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED; pflags &= ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); if (flags & M_WAITOK) pflags |= VM_ALLOC_WAITFAIL; prot = (flags & M_EXEC) != 0 ? VM_PROT_ALL : VM_PROT_RW; i = 0; VM_OBJECT_WLOCK(object); retry: mpred = vm_radix_lookup_le(&object->rtree, atop(offset + i)); for (; i < size; i += PAGE_SIZE, mpred = m) { m = vm_page_alloc_domain_after(object, atop(offset + i), domain, pflags, mpred); /* * Ran out of space, free everything up and return. Don't need * to lock page queues here as we know that the pages we got * aren't on any queues. */ if (m == NULL) { if ((flags & M_NOWAIT) == 0) goto retry; VM_OBJECT_WUNLOCK(object); kmem_unback(object, addr, i); return (KERN_NO_SPACE); } KASSERT(vm_phys_domain(m) == domain, ("kmem_back_domain: Domain mismatch %d != %d", vm_phys_domain(m), domain)); if (flags & M_ZERO && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("kmem_malloc: page %p is managed", m)); m->valid = VM_PAGE_BITS_ALL; pmap_enter(kernel_pmap, addr + i, m, prot, prot | PMAP_ENTER_WIRED, 0); #if VM_NRESERVLEVEL > 0 if (__predict_false((prot & VM_PROT_EXECUTE) != 0)) m->oflags |= VPO_KMEM_EXEC; #endif } VM_OBJECT_WUNLOCK(object); return (KERN_SUCCESS); } /* * kmem_back: * * Allocate physical pages for the specified virtual address range. */ int kmem_back(vm_object_t object, vm_offset_t addr, vm_size_t size, int flags) { vm_offset_t end, next, start; int domain, rv; KASSERT(object == kernel_object, ("kmem_back: only supports kernel object.")); for (start = addr, end = addr + size; addr < end; addr = next) { /* * We must ensure that pages backing a given large virtual page * all come from the same physical domain. */ if (vm_ndomains > 1) { domain = (addr >> KVA_QUANTUM_SHIFT) % vm_ndomains; while (VM_DOMAIN_EMPTY(domain)) domain++; next = roundup2(addr + 1, KVA_QUANTUM); if (next > end || next < start) next = end; } else { domain = 0; next = end; } rv = kmem_back_domain(domain, object, addr, next - addr, flags); if (rv != KERN_SUCCESS) { kmem_unback(object, start, addr - start); break; } } return (rv); } /* * kmem_unback: * * Unmap and free the physical pages underlying the specified virtual * address range. * * A physical page must exist within the specified object at each index * that is being unmapped. */ static struct vmem * _kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size) { struct vmem *arena; vm_page_t m, next; vm_offset_t end, offset; int domain; KASSERT(object == kernel_object, ("kmem_unback: only supports kernel object.")); if (size == 0) return (NULL); pmap_remove(kernel_pmap, addr, addr + size); offset = addr - VM_MIN_KERNEL_ADDRESS; end = offset + size; VM_OBJECT_WLOCK(object); m = vm_page_lookup(object, atop(offset)); domain = vm_phys_domain(m); #if VM_NRESERVLEVEL > 0 if (__predict_true((m->oflags & VPO_KMEM_EXEC) == 0)) arena = vm_dom[domain].vmd_kernel_arena; else arena = vm_dom[domain].vmd_kernel_rwx_arena; #else arena = vm_dom[domain].vmd_kernel_arena; #endif for (; offset < end; offset += PAGE_SIZE, m = next) { next = vm_page_next(m); vm_page_unwire(m, PQ_NONE); vm_page_free(m); } VM_OBJECT_WUNLOCK(object); return (arena); } void kmem_unback(vm_object_t object, vm_offset_t addr, vm_size_t size) { (void)_kmem_unback(object, addr, size); } /* * kmem_free: * * Free memory allocated with kmem_malloc. The size must match the * original allocation. */ void kmem_free(vm_offset_t addr, vm_size_t size) { struct vmem *arena; size = round_page(size); arena = _kmem_unback(kernel_object, addr, size); if (arena != NULL) vmem_free(arena, addr, size); } /* * kmap_alloc_wait: * * Allocates pageable memory from a sub-map of the kernel. If the submap * has no room, the caller sleeps waiting for more memory in the submap. * * This routine may block. */ vm_offset_t kmap_alloc_wait(vm_map_t map, vm_size_t size) { vm_offset_t addr; size = round_page(size); if (!swap_reserve(size)) return (0); for (;;) { /* * To make this work for more than one map, use the map's lock * to lock out sleepers/wakers. */ vm_map_lock(map); if (vm_map_findspace(map, vm_map_min(map), size, &addr) == 0) break; /* no space now; see if we can ever get space */ if (vm_map_max(map) - vm_map_min(map) < size) { vm_map_unlock(map); swap_release(size); return (0); } map->needs_wakeup = TRUE; vm_map_unlock_and_wait(map, 0); } vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_ALL, VM_PROT_ALL, MAP_ACC_CHARGED); vm_map_unlock(map); return (addr); } /* * kmap_free_wakeup: * * Returns memory to a submap of the kernel, and wakes up any processes * waiting for memory in that map. */ void kmap_free_wakeup(vm_map_t map, vm_offset_t addr, vm_size_t size) { vm_map_lock(map); (void) vm_map_delete(map, trunc_page(addr), round_page(addr + size)); if (map->needs_wakeup) { map->needs_wakeup = FALSE; vm_map_wakeup(map); } vm_map_unlock(map); } void kmem_init_zero_region(void) { vm_offset_t addr, i; vm_page_t m; /* * Map a single physical page of zeros to a larger virtual range. * This requires less looping in places that want large amounts of * zeros, while not using much more physical resources. */ addr = kva_alloc(ZERO_REGION_SIZE); m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); for (i = 0; i < ZERO_REGION_SIZE; i += PAGE_SIZE) pmap_qenter(addr + i, &m, 1); pmap_protect(kernel_pmap, addr, addr + ZERO_REGION_SIZE, VM_PROT_READ); zero_region = (const void *)addr; } /* * Import KVA from the kernel map into the kernel arena. */ static int kva_import(void *unused, vmem_size_t size, int flags, vmem_addr_t *addrp) { vm_offset_t addr; int result; KASSERT((size % KVA_QUANTUM) == 0, ("kva_import: Size %jd is not a multiple of %d", (intmax_t)size, (int)KVA_QUANTUM)); addr = vm_map_min(kernel_map); result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0, VMFS_SUPER_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); if (result != KERN_SUCCESS) return (ENOMEM); *addrp = addr; return (0); } /* * Import KVA from a parent arena into a per-domain arena. Imports must be * KVA_QUANTUM-aligned and a multiple of KVA_QUANTUM in size. */ static int kva_import_domain(void *arena, vmem_size_t size, int flags, vmem_addr_t *addrp) { KASSERT((size % KVA_QUANTUM) == 0, ("kva_import_domain: Size %jd is not a multiple of %d", (intmax_t)size, (int)KVA_QUANTUM)); return (vmem_xalloc(arena, size, KVA_QUANTUM, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, flags, addrp)); } /* * kmem_init: * * Create the kernel map; insert a mapping covering kernel text, * data, bss, and all space allocated thus far (`boostrap' data). The * new map will thus map the range between VM_MIN_KERNEL_ADDRESS and * `start' as allocated, and the range between `start' and `end' as free. * Create the kernel vmem arena and its per-domain children. */ void kmem_init(vm_offset_t start, vm_offset_t end) { vm_map_t m; int domain; m = vm_map_create(kernel_pmap, VM_MIN_KERNEL_ADDRESS, end); m->system_map = 1; vm_map_lock(m); /* N.B.: cannot use kgdb to debug, starting with this assignment ... */ kernel_map = m; (void) vm_map_insert(m, NULL, (vm_ooffset_t) 0, #ifdef __amd64__ KERNBASE, #else VM_MIN_KERNEL_ADDRESS, #endif start, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); /* ... and ending with the completion of the above `insert' */ vm_map_unlock(m); /* * Initialize the kernel_arena. This can grow on demand. */ vmem_init(kernel_arena, "kernel arena", 0, 0, PAGE_SIZE, 0, 0); vmem_set_import(kernel_arena, kva_import, NULL, NULL, KVA_QUANTUM); for (domain = 0; domain < vm_ndomains; domain++) { /* * Initialize the per-domain arenas. These are used to color * the KVA space in a way that ensures that virtual large pages * are backed by memory from the same physical domain, * maximizing the potential for superpage promotion. */ vm_dom[domain].vmd_kernel_arena = vmem_create( "kernel arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK); vmem_set_import(vm_dom[domain].vmd_kernel_arena, kva_import_domain, NULL, kernel_arena, KVA_QUANTUM); /* * In architectures with superpages, maintain separate arenas * for allocations with permissions that differ from the * "standard" read/write permissions used for kernel memory, * so as not to inhibit superpage promotion. */ #if VM_NRESERVLEVEL > 0 vm_dom[domain].vmd_kernel_rwx_arena = vmem_create( "kernel rwx arena domain", 0, 0, PAGE_SIZE, 0, M_WAITOK); vmem_set_import(vm_dom[domain].vmd_kernel_rwx_arena, kva_import_domain, (vmem_release_t *)vmem_xfree, kernel_arena, KVA_QUANTUM); #endif } } /* * kmem_bootstrap_free: * * Free pages backing preloaded data (e.g., kernel modules) to the * system. Currently only supported on platforms that create a * vm_phys segment for preloaded data. */ void kmem_bootstrap_free(vm_offset_t start, vm_size_t size) { #if defined(__i386__) || defined(__amd64__) struct vm_domain *vmd; vm_offset_t end, va; vm_paddr_t pa; vm_page_t m; end = trunc_page(start + size); start = round_page(start); for (va = start; va < end; va += PAGE_SIZE) { pa = pmap_kextract(va); m = PHYS_TO_VM_PAGE(pa); vmd = vm_pagequeue_domain(m); vm_domain_free_lock(vmd); vm_phys_free_pages(m, 0); - vmd->vmd_page_count++; vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, 1); vm_cnt.v_page_count++; } pmap_remove(kernel_pmap, start, end); (void)vmem_add(kernel_arena, start, end - start, M_WAITOK); #endif } #ifdef DIAGNOSTIC /* * Allow userspace to directly trigger the VM drain routine for testing * purposes. */ static int debug_vm_lowmem(SYSCTL_HANDLER_ARGS) { int error, i; i = 0; error = sysctl_handle_int(oidp, &i, 0, req); if (error) return (error); if ((i & ~(VM_LOW_KMEM | VM_LOW_PAGES)) != 0) return (EINVAL); if (i != 0) EVENTHANDLER_INVOKE(vm_lowmem, i); return (0); } SYSCTL_PROC(_debug, OID_AUTO, vm_lowmem, CTLTYPE_INT | CTLFLAG_RW, 0, 0, debug_vm_lowmem, "I", "set to trigger vm_lowmem event with given flags"); #endif Index: head/sys/vm/vm_page.c =================================================================== --- head/sys/vm/vm_page.c (revision 339924) +++ head/sys/vm/vm_page.c (revision 339925) @@ -1,4519 +1,4514 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1998 Matthew Dillon. 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. 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_page.c 7.4 (Berkeley) 5/7/91 */ /*- * 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. */ /* * GENERAL RULES ON VM_PAGE MANIPULATION * * - A page queue lock is required when adding or removing a page from a * page queue regardless of other locks or the busy state of a page. * * * In general, no thread besides the page daemon can acquire or * hold more than one page queue lock at a time. * * * The page daemon can acquire and hold any pair of page queue * locks in any order. * * - The object lock is required when inserting or removing * pages from an object (vm_page_insert() or vm_page_remove()). * */ /* * Resident memory management module. */ #include __FBSDID("$FreeBSD$"); #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 #include #include #include #include #include #include #include #include #include #include #include #include #include extern int uma_startup_count(int); extern void uma_startup(void *, int); extern int vmem_startup_count(void); struct vm_domain vm_dom[MAXMEMDOM]; DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; struct mtx_padalign __exclusive_cache_line vm_domainset_lock; /* The following fields are protected by the domainset lock. */ domainset_t __exclusive_cache_line vm_min_domains; domainset_t __exclusive_cache_line vm_severe_domains; static int vm_min_waiters; static int vm_severe_waiters; static int vm_pageproc_waiters; /* * bogus page -- for I/O to/from partially complete buffers, * or for paging into sparsely invalid regions. */ vm_page_t bogus_page; vm_page_t vm_page_array; long vm_page_array_size; long first_page; static int boot_pages; SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &boot_pages, 0, "number of pages allocated for bootstrapping the VM system"); static int pa_tryrelock_restart; SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD, &pa_tryrelock_restart, 0, "Number of tryrelock restarts"); static TAILQ_HEAD(, vm_page) blacklist_head; static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); static uma_zone_t fakepg_zone; static void vm_page_alloc_check(vm_page_t m); static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); static void vm_page_dequeue_complete(vm_page_t m); static void vm_page_enqueue(vm_page_t m, uint8_t queue); static void vm_page_init(void *dummy); static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred); static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred); static int vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, vm_paddr_t high); static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req); static int vm_page_import(void *arg, void **store, int cnt, int domain, int flags); static void vm_page_release(void *arg, void **store, int cnt); SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); static void vm_page_init(void *dummy) { fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | VM_ALLOC_NORMAL | VM_ALLOC_WIRED); } /* * The cache page zone is initialized later since we need to be able to allocate * pages before UMA is fully initialized. */ static void vm_page_init_cache_zones(void *dummy __unused) { struct vm_domain *vmd; int i; for (i = 0; i < vm_ndomains; i++) { vmd = VM_DOMAIN(i); /* * Don't allow the page cache to take up more than .25% of * memory. */ if (vmd->vmd_page_count / 400 < 256 * mp_ncpus) continue; vmd->vmd_pgcache = uma_zcache_create("vm pgcache", sizeof(struct vm_page), NULL, NULL, NULL, NULL, vm_page_import, vm_page_release, vmd, UMA_ZONE_NOBUCKETCACHE | UMA_ZONE_MAXBUCKET | UMA_ZONE_VM); } } SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ #if PAGE_SIZE == 32768 #ifdef CTASSERT CTASSERT(sizeof(u_long) >= 8); #endif #endif /* * Try to acquire a physical address lock while a pmap is locked. If we * fail to trylock we unlock and lock the pmap directly and cache the * locked pa in *locked. The caller should then restart their loop in case * the virtual to physical mapping has changed. */ int vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked) { vm_paddr_t lockpa; lockpa = *locked; *locked = pa; if (lockpa) { PA_LOCK_ASSERT(lockpa, MA_OWNED); if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa)) return (0); PA_UNLOCK(lockpa); } if (PA_TRYLOCK(pa)) return (0); PMAP_UNLOCK(pmap); atomic_add_int(&pa_tryrelock_restart, 1); PA_LOCK(pa); PMAP_LOCK(pmap); return (EAGAIN); } /* * vm_set_page_size: * * Sets the page size, perhaps based upon the memory * size. Must be called before any use of page-size * dependent functions. */ void vm_set_page_size(void) { if (vm_cnt.v_page_size == 0) vm_cnt.v_page_size = PAGE_SIZE; if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) panic("vm_set_page_size: page size not a power of two"); } /* * vm_page_blacklist_next: * * Find the next entry in the provided string of blacklist * addresses. Entries are separated by space, comma, or newline. * If an invalid integer is encountered then the rest of the * string is skipped. Updates the list pointer to the next * character, or NULL if the string is exhausted or invalid. */ static vm_paddr_t vm_page_blacklist_next(char **list, char *end) { vm_paddr_t bad; char *cp, *pos; if (list == NULL || *list == NULL) return (0); if (**list =='\0') { *list = NULL; return (0); } /* * If there's no end pointer then the buffer is coming from * the kenv and we know it's null-terminated. */ if (end == NULL) end = *list + strlen(*list); /* Ensure that strtoq() won't walk off the end */ if (*end != '\0') { if (*end == '\n' || *end == ' ' || *end == ',') *end = '\0'; else { printf("Blacklist not terminated, skipping\n"); *list = NULL; return (0); } } for (pos = *list; *pos != '\0'; pos = cp) { bad = strtoq(pos, &cp, 0); if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { if (bad == 0) { if (++cp < end) continue; else break; } } else break; if (*cp == '\0' || ++cp >= end) *list = NULL; else *list = cp; return (trunc_page(bad)); } printf("Garbage in RAM blacklist, skipping\n"); *list = NULL; return (0); } bool vm_page_blacklist_add(vm_paddr_t pa, bool verbose) { struct vm_domain *vmd; vm_page_t m; int ret; m = vm_phys_paddr_to_vm_page(pa); if (m == NULL) return (true); /* page does not exist, no failure */ vmd = vm_pagequeue_domain(m); vm_domain_free_lock(vmd); ret = vm_phys_unfree_page(m); vm_domain_free_unlock(vmd); if (ret) { TAILQ_INSERT_TAIL(&blacklist_head, m, listq); if (verbose) printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); } return (ret); } /* * vm_page_blacklist_check: * * Iterate through the provided string of blacklist addresses, pulling * each entry out of the physical allocator free list and putting it * onto a list for reporting via the vm.page_blacklist sysctl. */ static void vm_page_blacklist_check(char *list, char *end) { vm_paddr_t pa; char *next; next = list; while (next != NULL) { if ((pa = vm_page_blacklist_next(&next, end)) == 0) continue; vm_page_blacklist_add(pa, bootverbose); } } /* * vm_page_blacklist_load: * * Search for a special module named "ram_blacklist". It'll be a * plain text file provided by the user via the loader directive * of the same name. */ static void vm_page_blacklist_load(char **list, char **end) { void *mod; u_char *ptr; u_int len; mod = NULL; ptr = NULL; mod = preload_search_by_type("ram_blacklist"); if (mod != NULL) { ptr = preload_fetch_addr(mod); len = preload_fetch_size(mod); } *list = ptr; if (ptr != NULL) *end = ptr + len; else *end = NULL; return; } static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) { vm_page_t m; struct sbuf sbuf; int error, first; first = 1; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128, req); TAILQ_FOREACH(m, &blacklist_head, listq) { sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", (uintmax_t)m->phys_addr); first = 0; } error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } /* * Initialize a dummy page for use in scans of the specified paging queue. * In principle, this function only needs to set the flag PG_MARKER. * Nonetheless, it write busies and initializes the hold count to one as * safety precautions. */ static void vm_page_init_marker(vm_page_t marker, int queue, uint8_t aflags) { bzero(marker, sizeof(*marker)); marker->flags = PG_MARKER; marker->aflags = aflags; marker->busy_lock = VPB_SINGLE_EXCLUSIVER; marker->queue = queue; marker->hold_count = 1; } static void vm_page_domain_init(int domain) { struct vm_domain *vmd; struct vm_pagequeue *pq; int i; vmd = VM_DOMAIN(domain); bzero(vmd, sizeof(*vmd)); *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = "vm inactive pagequeue"; *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = "vm active pagequeue"; *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = "vm laundry pagequeue"; *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = "vm unswappable pagequeue"; vmd->vmd_domain = domain; vmd->vmd_page_count = 0; vmd->vmd_free_count = 0; vmd->vmd_segs = 0; vmd->vmd_oom = FALSE; for (i = 0; i < PQ_COUNT; i++) { pq = &vmd->vmd_pagequeues[i]; TAILQ_INIT(&pq->pq_pl); mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", MTX_DEF | MTX_DUPOK); pq->pq_pdpages = 0; vm_page_init_marker(&vmd->vmd_markers[i], i, 0); } mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); /* * inacthead is used to provide FIFO ordering for LRU-bypassing * insertions. */ vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, &vmd->vmd_inacthead, plinks.q); /* * The clock pages are used to implement active queue scanning without * requeues. Scans start at clock[0], which is advanced after the scan * ends. When the two clock hands meet, they are reset and scanning * resumes from the head of the queue. */ vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, &vmd->vmd_clock[1], plinks.q); } /* * Initialize a physical page in preparation for adding it to the free * lists. */ static void vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) { m->object = NULL; m->wire_count = 0; m->busy_lock = VPB_UNBUSIED; m->hold_count = 0; m->flags = m->aflags = 0; m->phys_addr = pa; m->queue = PQ_NONE; m->psind = 0; m->segind = segind; m->order = VM_NFREEORDER; m->pool = VM_FREEPOOL_DEFAULT; m->valid = m->dirty = 0; pmap_page_init(m); } /* * vm_page_startup: * * Initializes the resident memory module. Allocates physical memory for * bootstrapping UMA and some data structures that are used to manage * physical pages. Initializes these structures, and populates the free * page queues. */ vm_offset_t vm_page_startup(vm_offset_t vaddr) { struct vm_phys_seg *seg; vm_page_t m; char *list, *listend; vm_offset_t mapped; vm_paddr_t end, high_avail, low_avail, new_end, page_range, size; vm_paddr_t biggestsize, last_pa, pa; u_long pagecount; int biggestone, i, segind; #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) long ii; #endif biggestsize = 0; biggestone = 0; vaddr = round_page(vaddr); for (i = 0; phys_avail[i + 1]; i += 2) { phys_avail[i] = round_page(phys_avail[i]); phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); } for (i = 0; phys_avail[i + 1]; i += 2) { size = phys_avail[i + 1] - phys_avail[i]; if (size > biggestsize) { biggestone = i; biggestsize = size; } } end = phys_avail[biggestone+1]; /* * Initialize the page and queue locks. */ mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); for (i = 0; i < PA_LOCK_COUNT; i++) mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); for (i = 0; i < vm_ndomains; i++) vm_page_domain_init(i); /* * Allocate memory for use when boot strapping the kernel memory * allocator. Tell UMA how many zones we are going to create * before going fully functional. UMA will add its zones. * * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP, * KMAP ENTRY, MAP ENTRY, VMSPACE. */ boot_pages = uma_startup_count(8); #ifndef UMA_MD_SMALL_ALLOC /* vmem_startup() calls uma_prealloc(). */ boot_pages += vmem_startup_count(); /* vm_map_startup() calls uma_prealloc(). */ boot_pages += howmany(MAX_KMAP, UMA_SLAB_SPACE / sizeof(struct vm_map)); /* * Before going fully functional kmem_init() does allocation * from "KMAP ENTRY" and vmem_create() does allocation from "vmem". */ boot_pages += 2; #endif /* * CTFLAG_RDTUN doesn't work during the early boot process, so we must * manually fetch the value. */ TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); new_end = end - (boot_pages * UMA_SLAB_SIZE); new_end = trunc_page(new_end); mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)mapped, end - new_end); uma_startup((void *)mapped, boot_pages); #ifdef WITNESS end = new_end; new_end = end - round_page(witness_startup_count()); mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)mapped, end - new_end); witness_startup((void *)mapped); #endif #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ defined(__i386__) || defined(__mips__) /* * Allocate a bitmap to indicate that a random physical page * needs to be included in a minidump. * * The amd64 port needs this to indicate which direct map pages * need to be dumped, via calls to dump_add_page()/dump_drop_page(). * * However, i386 still needs this workspace internally within the * minidump code. In theory, they are not needed on i386, but are * included should the sf_buf code decide to use them. */ last_pa = 0; for (i = 0; dump_avail[i + 1] != 0; i += 2) if (dump_avail[i + 1] > last_pa) last_pa = dump_avail[i + 1]; page_range = last_pa / PAGE_SIZE; vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); new_end -= vm_page_dump_size; vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)vm_page_dump, vm_page_dump_size); #else (void)last_pa; #endif #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) /* * Include the UMA bootstrap pages and vm_page_dump in a crash dump. * When pmap_map() uses the direct map, they are not automatically * included. */ for (pa = new_end; pa < end; pa += PAGE_SIZE) dump_add_page(pa); #endif phys_avail[biggestone + 1] = new_end; #ifdef __amd64__ /* * Request that the physical pages underlying the message buffer be * included in a crash dump. Since the message buffer is accessed * through the direct map, they are not automatically included. */ pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); last_pa = pa + round_page(msgbufsize); while (pa < last_pa) { dump_add_page(pa); pa += PAGE_SIZE; } #endif /* * Compute the number of pages of memory that will be available for * use, taking into account the overhead of a page structure per page. * In other words, solve * "available physical memory" - round_page(page_range * * sizeof(struct vm_page)) = page_range * PAGE_SIZE * for page_range. */ low_avail = phys_avail[0]; high_avail = phys_avail[1]; for (i = 0; i < vm_phys_nsegs; i++) { if (vm_phys_segs[i].start < low_avail) low_avail = vm_phys_segs[i].start; if (vm_phys_segs[i].end > high_avail) high_avail = vm_phys_segs[i].end; } /* Skip the first chunk. It is already accounted for. */ for (i = 2; phys_avail[i + 1] != 0; i += 2) { if (phys_avail[i] < low_avail) low_avail = phys_avail[i]; if (phys_avail[i + 1] > high_avail) high_avail = phys_avail[i + 1]; } first_page = low_avail / PAGE_SIZE; #ifdef VM_PHYSSEG_SPARSE size = 0; for (i = 0; i < vm_phys_nsegs; i++) size += vm_phys_segs[i].end - vm_phys_segs[i].start; for (i = 0; phys_avail[i + 1] != 0; i += 2) size += phys_avail[i + 1] - phys_avail[i]; #elif defined(VM_PHYSSEG_DENSE) size = high_avail - low_avail; #else #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." #endif #ifdef VM_PHYSSEG_DENSE /* * In the VM_PHYSSEG_DENSE case, the number of pages can account for * the overhead of a page structure per page only if vm_page_array is * allocated from the last physical memory chunk. Otherwise, we must * allocate page structures representing the physical memory * underlying vm_page_array, even though they will not be used. */ if (new_end != high_avail) page_range = size / PAGE_SIZE; else #endif { page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); /* * If the partial bytes remaining are large enough for * a page (PAGE_SIZE) without a corresponding * 'struct vm_page', then new_end will contain an * extra page after subtracting the length of the VM * page array. Compensate by subtracting an extra * page from new_end. */ if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { if (new_end == high_avail) high_avail -= PAGE_SIZE; new_end -= PAGE_SIZE; } } end = new_end; /* * Reserve an unmapped guard page to trap access to vm_page_array[-1]. * However, because this page is allocated from KVM, out-of-bounds * accesses using the direct map will not be trapped. */ vaddr += PAGE_SIZE; /* * Allocate physical memory for the page structures, and map it. */ new_end = trunc_page(end - page_range * sizeof(struct vm_page)); mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); vm_page_array = (vm_page_t)mapped; vm_page_array_size = page_range; #if VM_NRESERVLEVEL > 0 /* * Allocate physical memory for the reservation management system's * data structures, and map it. */ if (high_avail == end) high_avail = new_end; new_end = vm_reserv_startup(&vaddr, new_end, high_avail); #endif #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) /* * Include vm_page_array and vm_reserv_array in a crash dump. */ for (pa = new_end; pa < end; pa += PAGE_SIZE) dump_add_page(pa); #endif phys_avail[biggestone + 1] = new_end; /* * Add physical memory segments corresponding to the available * physical pages. */ for (i = 0; phys_avail[i + 1] != 0; i += 2) vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); /* * Initialize the physical memory allocator. */ vm_phys_init(); /* * Initialize the page structures and add every available page to the * physical memory allocator's free lists. */ #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) for (ii = 0; ii < vm_page_array_size; ii++) { m = &vm_page_array[ii]; vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); m->flags = PG_FICTITIOUS; } #endif vm_cnt.v_page_count = 0; for (segind = 0; segind < vm_phys_nsegs; segind++) { seg = &vm_phys_segs[segind]; for (m = seg->first_page, pa = seg->start; pa < seg->end; m++, pa += PAGE_SIZE) vm_page_init_page(m, pa, segind); /* * Add the segment to the free lists only if it is covered by * one of the ranges in phys_avail. Because we've added the * ranges to the vm_phys_segs array, we can assume that each * segment is either entirely contained in one of the ranges, * or doesn't overlap any of them. */ for (i = 0; phys_avail[i + 1] != 0; i += 2) { struct vm_domain *vmd; if (seg->start < phys_avail[i] || seg->end > phys_avail[i + 1]) continue; m = seg->first_page; pagecount = (u_long)atop(seg->end - seg->start); vmd = VM_DOMAIN(seg->domain); vm_domain_free_lock(vmd); vm_phys_free_contig(m, pagecount); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, pagecount); vm_cnt.v_page_count += (u_int)pagecount; vmd = VM_DOMAIN(seg->domain); vmd->vmd_page_count += (u_int)pagecount; vmd->vmd_segs |= 1UL << m->segind; break; } } /* * Remove blacklisted pages from the physical memory allocator. */ TAILQ_INIT(&blacklist_head); vm_page_blacklist_load(&list, &listend); vm_page_blacklist_check(list, listend); list = kern_getenv("vm.blacklist"); vm_page_blacklist_check(list, NULL); freeenv(list); #if VM_NRESERVLEVEL > 0 /* * Initialize the reservation management system. */ vm_reserv_init(); #endif - /* - * Set an initial domain policy for thread0 so that allocations - * can work. - */ - domainset_zero(); return (vaddr); } void vm_page_reference(vm_page_t m) { vm_page_aflag_set(m, PGA_REFERENCED); } /* * vm_page_busy_downgrade: * * Downgrade an exclusive busy page into a single shared busy page. */ void vm_page_busy_downgrade(vm_page_t m) { u_int x; bool locked; vm_page_assert_xbusied(m); locked = mtx_owned(vm_page_lockptr(m)); for (;;) { x = m->busy_lock; x &= VPB_BIT_WAITERS; if (x != 0 && !locked) vm_page_lock(m); if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1))) break; if (x != 0 && !locked) vm_page_unlock(m); } if (x != 0) { wakeup(m); if (!locked) vm_page_unlock(m); } } /* * vm_page_sbusied: * * Return a positive value if the page is shared busied, 0 otherwise. */ int vm_page_sbusied(vm_page_t m) { u_int x; x = m->busy_lock; return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); } /* * vm_page_sunbusy: * * Shared unbusy a page. */ void vm_page_sunbusy(vm_page_t m) { u_int x; vm_page_lock_assert(m, MA_NOTOWNED); vm_page_assert_sbusied(m); for (;;) { x = m->busy_lock; if (VPB_SHARERS(x) > 1) { if (atomic_cmpset_int(&m->busy_lock, x, x - VPB_ONE_SHARER)) break; continue; } if ((x & VPB_BIT_WAITERS) == 0) { KASSERT(x == VPB_SHARERS_WORD(1), ("vm_page_sunbusy: invalid lock state")); if (atomic_cmpset_int(&m->busy_lock, VPB_SHARERS_WORD(1), VPB_UNBUSIED)) break; continue; } KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS), ("vm_page_sunbusy: invalid lock state for waiters")); vm_page_lock(m); if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) { vm_page_unlock(m); continue; } wakeup(m); vm_page_unlock(m); break; } } /* * vm_page_busy_sleep: * * Sleep and release the page lock, using the page pointer as wchan. * This is used to implement the hard-path of busying mechanism. * * The given page must be locked. * * If nonshared is true, sleep only if the page is xbusy. */ void vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) { u_int x; vm_page_assert_locked(m); x = m->busy_lock; if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || ((x & VPB_BIT_WAITERS) == 0 && !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { vm_page_unlock(m); return; } msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0); } /* * vm_page_trysbusy: * * Try to shared busy a page. * If the operation succeeds 1 is returned otherwise 0. * The operation never sleeps. */ int vm_page_trysbusy(vm_page_t m) { u_int x; for (;;) { x = m->busy_lock; if ((x & VPB_BIT_SHARED) == 0) return (0); if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER)) return (1); } } static void vm_page_xunbusy_locked(vm_page_t m) { vm_page_assert_xbusied(m); vm_page_assert_locked(m); atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); /* There is a waiter, do wakeup() instead of vm_page_flash(). */ wakeup(m); } void vm_page_xunbusy_maybelocked(vm_page_t m) { bool lockacq; vm_page_assert_xbusied(m); /* * Fast path for unbusy. If it succeeds, we know that there * are no waiters, so we do not need a wakeup. */ if (atomic_cmpset_rel_int(&m->busy_lock, VPB_SINGLE_EXCLUSIVER, VPB_UNBUSIED)) return; lockacq = !mtx_owned(vm_page_lockptr(m)); if (lockacq) vm_page_lock(m); vm_page_xunbusy_locked(m); if (lockacq) vm_page_unlock(m); } /* * vm_page_xunbusy_hard: * * Called after the first try the exclusive unbusy of a page failed. * It is assumed that the waiters bit is on. */ void vm_page_xunbusy_hard(vm_page_t m) { vm_page_assert_xbusied(m); vm_page_lock(m); vm_page_xunbusy_locked(m); vm_page_unlock(m); } /* * vm_page_flash: * * Wakeup anyone waiting for the page. * The ownership bits do not change. * * The given page must be locked. */ void vm_page_flash(vm_page_t m) { u_int x; vm_page_lock_assert(m, MA_OWNED); for (;;) { x = m->busy_lock; if ((x & VPB_BIT_WAITERS) == 0) return; if (atomic_cmpset_int(&m->busy_lock, x, x & (~VPB_BIT_WAITERS))) break; } wakeup(m); } /* * Avoid releasing and reacquiring the same page lock. */ void vm_page_change_lock(vm_page_t m, struct mtx **mtx) { struct mtx *mtx1; mtx1 = vm_page_lockptr(m); if (*mtx == mtx1) return; if (*mtx != NULL) mtx_unlock(*mtx); *mtx = mtx1; mtx_lock(mtx1); } /* * Keep page from being freed by the page daemon * much of the same effect as wiring, except much lower * overhead and should be used only for *very* temporary * holding ("wiring"). */ void vm_page_hold(vm_page_t mem) { vm_page_lock_assert(mem, MA_OWNED); mem->hold_count++; } void vm_page_unhold(vm_page_t mem) { vm_page_lock_assert(mem, MA_OWNED); KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!")); --mem->hold_count; if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0) vm_page_free_toq(mem); } /* * vm_page_unhold_pages: * * Unhold each of the pages that is referenced by the given array. */ void vm_page_unhold_pages(vm_page_t *ma, int count) { struct mtx *mtx; mtx = NULL; for (; count != 0; count--) { vm_page_change_lock(*ma, &mtx); vm_page_unhold(*ma); ma++; } if (mtx != NULL) mtx_unlock(mtx); } vm_page_t PHYS_TO_VM_PAGE(vm_paddr_t pa) { vm_page_t m; #ifdef VM_PHYSSEG_SPARSE m = vm_phys_paddr_to_vm_page(pa); if (m == NULL) m = vm_phys_fictitious_to_vm_page(pa); return (m); #elif defined(VM_PHYSSEG_DENSE) long pi; pi = atop(pa); if (pi >= first_page && (pi - first_page) < vm_page_array_size) { m = &vm_page_array[pi - first_page]; return (m); } return (vm_phys_fictitious_to_vm_page(pa)); #else #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." #endif } /* * vm_page_getfake: * * Create a fictitious page with the specified physical address and * memory attribute. The memory attribute is the only the machine- * dependent aspect of a fictitious page that must be initialized. */ vm_page_t vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) { vm_page_t m; m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); vm_page_initfake(m, paddr, memattr); return (m); } void vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) { if ((m->flags & PG_FICTITIOUS) != 0) { /* * The page's memattr might have changed since the * previous initialization. Update the pmap to the * new memattr. */ goto memattr; } m->phys_addr = paddr; m->queue = PQ_NONE; /* Fictitious pages don't use "segind". */ m->flags = PG_FICTITIOUS; /* Fictitious pages don't use "order" or "pool". */ m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_SINGLE_EXCLUSIVER; m->wire_count = 1; pmap_page_init(m); memattr: pmap_page_set_memattr(m, memattr); } /* * vm_page_putfake: * * Release a fictitious page. */ void vm_page_putfake(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); KASSERT((m->flags & PG_FICTITIOUS) != 0, ("vm_page_putfake: bad page %p", m)); uma_zfree(fakepg_zone, m); } /* * vm_page_updatefake: * * Update the given fictitious page to the specified physical address and * memory attribute. */ void vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) { KASSERT((m->flags & PG_FICTITIOUS) != 0, ("vm_page_updatefake: bad page %p", m)); m->phys_addr = paddr; pmap_page_set_memattr(m, memattr); } /* * vm_page_free: * * Free a page. */ void vm_page_free(vm_page_t m) { m->flags &= ~PG_ZERO; vm_page_free_toq(m); } /* * vm_page_free_zero: * * Free a page to the zerod-pages queue */ void vm_page_free_zero(vm_page_t m) { m->flags |= PG_ZERO; vm_page_free_toq(m); } /* * Unbusy and handle the page queueing for a page from a getpages request that * was optionally read ahead or behind. */ void vm_page_readahead_finish(vm_page_t m) { /* We shouldn't put invalid pages on queues. */ KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m)); /* * Since the page is not the actually needed one, whether it should * be activated or deactivated is not obvious. Empirical results * have shown that deactivating the page is usually the best choice, * unless the page is wanted by another thread. */ vm_page_lock(m); if ((m->busy_lock & VPB_BIT_WAITERS) != 0) vm_page_activate(m); else vm_page_deactivate(m); vm_page_unlock(m); vm_page_xunbusy(m); } /* * vm_page_sleep_if_busy: * * Sleep and release the page queues lock if the page is busied. * Returns TRUE if the thread slept. * * The given page must be unlocked and object containing it must * be locked. */ int vm_page_sleep_if_busy(vm_page_t m, const char *msg) { vm_object_t obj; vm_page_lock_assert(m, MA_NOTOWNED); VM_OBJECT_ASSERT_WLOCKED(m->object); if (vm_page_busied(m)) { /* * The page-specific object must be cached because page * identity can change during the sleep, causing the * re-lock of a different object. * It is assumed that a reference to the object is already * held by the callers. */ obj = m->object; vm_page_lock(m); VM_OBJECT_WUNLOCK(obj); vm_page_busy_sleep(m, msg, false); VM_OBJECT_WLOCK(obj); return (TRUE); } return (FALSE); } /* * vm_page_dirty_KBI: [ internal use only ] * * Set all bits in the page's dirty field. * * The object containing the specified page must be locked if the * call is made from the machine-independent layer. * * See vm_page_clear_dirty_mask(). * * This function should only be called by vm_page_dirty(). */ void vm_page_dirty_KBI(vm_page_t m) { /* Refer to this operation by its public name. */ KASSERT(m->valid == VM_PAGE_BITS_ALL, ("vm_page_dirty: page is invalid!")); m->dirty = VM_PAGE_BITS_ALL; } /* * vm_page_insert: [ internal use only ] * * Inserts the given mem entry into the object and object list. * * The object must be locked. */ int vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) { vm_page_t mpred; VM_OBJECT_ASSERT_WLOCKED(object); mpred = vm_radix_lookup_le(&object->rtree, pindex); return (vm_page_insert_after(m, object, pindex, mpred)); } /* * vm_page_insert_after: * * Inserts the page "m" into the specified object at offset "pindex". * * The page "mpred" must immediately precede the offset "pindex" within * the specified object. * * The object must be locked. */ static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred) { vm_page_t msucc; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(m->object == NULL, ("vm_page_insert_after: page already inserted")); if (mpred != NULL) { KASSERT(mpred->object == object, ("vm_page_insert_after: object doesn't contain mpred")); KASSERT(mpred->pindex < pindex, ("vm_page_insert_after: mpred doesn't precede pindex")); msucc = TAILQ_NEXT(mpred, listq); } else msucc = TAILQ_FIRST(&object->memq); if (msucc != NULL) KASSERT(msucc->pindex > pindex, ("vm_page_insert_after: msucc doesn't succeed pindex")); /* * Record the object/offset pair in this page */ m->object = object; m->pindex = pindex; /* * Now link into the object's ordered list of backed pages. */ if (vm_radix_insert(&object->rtree, m)) { m->object = NULL; m->pindex = 0; return (1); } vm_page_insert_radixdone(m, object, mpred); return (0); } /* * vm_page_insert_radixdone: * * Complete page "m" insertion into the specified object after the * radix trie hooking. * * The page "mpred" must precede the offset "m->pindex" within the * specified object. * * The object must be locked. */ static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object != NULL && m->object == object, ("vm_page_insert_radixdone: page %p has inconsistent object", m)); if (mpred != NULL) { KASSERT(mpred->object == object, ("vm_page_insert_after: object doesn't contain mpred")); KASSERT(mpred->pindex < m->pindex, ("vm_page_insert_after: mpred doesn't precede pindex")); } if (mpred != NULL) TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); else TAILQ_INSERT_HEAD(&object->memq, m, listq); /* * Show that the object has one more resident page. */ object->resident_page_count++; /* * Hold the vnode until the last page is released. */ if (object->resident_page_count == 1 && object->type == OBJT_VNODE) vhold(object->handle); /* * Since we are inserting a new and possibly dirty page, * update the object's OBJ_MIGHTBEDIRTY flag. */ if (pmap_page_is_write_mapped(m)) vm_object_set_writeable_dirty(object); } /* * vm_page_remove: * * Removes the specified page from its containing object, but does not * invalidate any backing storage. * * The object must be locked. The page must be locked if it is managed. */ void vm_page_remove(vm_page_t m) { vm_object_t object; vm_page_t mrem; if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_assert_locked(m); if ((object = m->object) == NULL) return; VM_OBJECT_ASSERT_WLOCKED(object); if (vm_page_xbusied(m)) vm_page_xunbusy_maybelocked(m); mrem = vm_radix_remove(&object->rtree, m->pindex); KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); /* * Now remove from the object's list of backed pages. */ TAILQ_REMOVE(&object->memq, m, listq); /* * And show that the object has one fewer resident page. */ object->resident_page_count--; /* * The vnode may now be recycled. */ if (object->resident_page_count == 0 && object->type == OBJT_VNODE) vdrop(object->handle); m->object = NULL; } /* * vm_page_lookup: * * Returns the page associated with the object/offset * pair specified; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_lookup(vm_object_t object, vm_pindex_t pindex) { VM_OBJECT_ASSERT_LOCKED(object); return (vm_radix_lookup(&object->rtree, pindex)); } /* * vm_page_find_least: * * Returns the page associated with the object with least pindex * greater than or equal to the parameter pindex, or NULL. * * The object must be locked. */ vm_page_t vm_page_find_least(vm_object_t object, vm_pindex_t pindex) { vm_page_t m; VM_OBJECT_ASSERT_LOCKED(object); if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) m = vm_radix_lookup_ge(&object->rtree, pindex); return (m); } /* * Returns the given page's successor (by pindex) within the object if it is * resident; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_next(vm_page_t m) { vm_page_t next; VM_OBJECT_ASSERT_LOCKED(m->object); if ((next = TAILQ_NEXT(m, listq)) != NULL) { MPASS(next->object == m->object); if (next->pindex != m->pindex + 1) next = NULL; } return (next); } /* * Returns the given page's predecessor (by pindex) within the object if it is * resident; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_prev(vm_page_t m) { vm_page_t prev; VM_OBJECT_ASSERT_LOCKED(m->object); if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { MPASS(prev->object == m->object); if (prev->pindex != m->pindex - 1) prev = NULL; } return (prev); } /* * Uses the page mnew as a replacement for an existing page at index * pindex which must be already present in the object. * * The existing page must not be on a paging queue. */ vm_page_t vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex) { vm_page_t mold; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(mnew->object == NULL, ("vm_page_replace: page %p already in object", mnew)); KASSERT(mnew->queue == PQ_NONE, ("vm_page_replace: new page %p is on a paging queue", mnew)); /* * This function mostly follows vm_page_insert() and * vm_page_remove() without the radix, object count and vnode * dance. Double check such functions for more comments. */ mnew->object = object; mnew->pindex = pindex; mold = vm_radix_replace(&object->rtree, mnew); KASSERT(mold->queue == PQ_NONE, ("vm_page_replace: old page %p is on a paging queue", mold)); /* Keep the resident page list in sorted order. */ TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); TAILQ_REMOVE(&object->memq, mold, listq); mold->object = NULL; vm_page_xunbusy_maybelocked(mold); /* * The object's resident_page_count does not change because we have * swapped one page for another, but OBJ_MIGHTBEDIRTY. */ if (pmap_page_is_write_mapped(mnew)) vm_object_set_writeable_dirty(object); return (mold); } /* * vm_page_rename: * * Move the given memory entry from its * current object to the specified target object/offset. * * Note: swap associated with the page must be invalidated by the move. We * have to do this for several reasons: (1) we aren't freeing the * page, (2) we are dirtying the page, (3) the VM system is probably * moving the page from object A to B, and will then later move * the backing store from A to B and we can't have a conflict. * * Note: we *always* dirty the page. It is necessary both for the * fact that we moved it, and because we may be invalidating * swap. * * The objects must be locked. */ int vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) { vm_page_t mpred; vm_pindex_t opidx; VM_OBJECT_ASSERT_WLOCKED(new_object); mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); KASSERT(mpred == NULL || mpred->pindex != new_pindex, ("vm_page_rename: pindex already renamed")); /* * Create a custom version of vm_page_insert() which does not depend * by m_prev and can cheat on the implementation aspects of the * function. */ opidx = m->pindex; m->pindex = new_pindex; if (vm_radix_insert(&new_object->rtree, m)) { m->pindex = opidx; return (1); } /* * The operation cannot fail anymore. The removal must happen before * the listq iterator is tainted. */ m->pindex = opidx; vm_page_lock(m); vm_page_remove(m); /* Return back to the new pindex to complete vm_page_insert(). */ m->pindex = new_pindex; m->object = new_object; vm_page_unlock(m); vm_page_insert_radixdone(m, new_object, mpred); vm_page_dirty(m); return (0); } /* * vm_page_alloc: * * Allocate and return a page that is associated with the specified * object and offset pair. By default, this page is exclusive busied. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_COUNT(number) the number of additional pages that the caller * intends to allocate * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NODUMP do not include the page in a kernel core dump * VM_ALLOC_NOOBJ page is not associated with an object and * should not be exclusive busy * VM_ALLOC_SBUSY shared busy the allocated page * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) { return (vm_page_alloc_after(object, pindex, req, object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : NULL)); } vm_page_t vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req) { return (vm_page_alloc_domain_after(object, pindex, domain, req, object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : NULL)); } /* * Allocate a page in the specified object with the given page index. To * optimize insertion of the page into the object, the caller must also specifiy * the resident page in the object with largest index smaller than the given * page index, or NULL if no such page exists. */ vm_page_t vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, int req, vm_page_t mpred) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); do { m = vm_page_alloc_domain_after(object, pindex, domain, req, mpred); if (m != NULL) break; } while (vm_domainset_iter_page(&di, object, &domain) == 0); return (m); } /* * Returns true if the number of free pages exceeds the minimum * for the request class and false otherwise. */ int vm_domain_allocate(struct vm_domain *vmd, int req, int npages) { u_int limit, old, new; req = req & VM_ALLOC_CLASS_MASK; /* * The page daemon is allowed to dig deeper into the free page list. */ if (curproc == pageproc && req != VM_ALLOC_INTERRUPT) req = VM_ALLOC_SYSTEM; if (req == VM_ALLOC_INTERRUPT) limit = 0; else if (req == VM_ALLOC_SYSTEM) limit = vmd->vmd_interrupt_free_min; else limit = vmd->vmd_free_reserved; /* * Attempt to reserve the pages. Fail if we're below the limit. */ limit += npages; old = vmd->vmd_free_count; do { if (old < limit) return (0); new = old - npages; } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); /* Wake the page daemon if we've crossed the threshold. */ if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) pagedaemon_wakeup(vmd->vmd_domain); /* Only update bitsets on transitions. */ if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) vm_domain_set(vmd); return (1); } vm_page_t vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, int req, vm_page_t mpred) { struct vm_domain *vmd; vm_page_t m; int flags; KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("inconsistent object(%p)/req(%x)", object, req)); KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, ("Can't sleep and retry object insertion.")); KASSERT(mpred == NULL || mpred->pindex < pindex, ("mpred %p doesn't precede pindex 0x%jx", mpred, (uintmax_t)pindex)); if (object != NULL) VM_OBJECT_ASSERT_WLOCKED(object); again: m = NULL; #if VM_NRESERVLEVEL > 0 /* * Can we allocate the page from a reservation? */ if (vm_object_reserv(object) && ((m = vm_reserv_extend(req, object, pindex, domain, mpred)) != NULL || (m = vm_reserv_alloc_page(req, object, pindex, domain, mpred)) != NULL)) { domain = vm_phys_domain(m); vmd = VM_DOMAIN(domain); goto found; } #endif vmd = VM_DOMAIN(domain); if (object != NULL && vmd->vmd_pgcache != NULL) { m = uma_zalloc(vmd->vmd_pgcache, M_NOWAIT); if (m != NULL) goto found; } if (vm_domain_allocate(vmd, req, 1)) { /* * If not, allocate it from the free page queues. */ vm_domain_free_lock(vmd); m = vm_phys_alloc_pages(domain, object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); vm_domain_free_unlock(vmd); if (m == NULL) { vm_domain_freecnt_inc(vmd, 1); #if VM_NRESERVLEVEL > 0 if (vm_reserv_reclaim_inactive(domain)) goto again; #endif } } if (m == NULL) { /* * Not allocatable, give up. */ if (vm_domain_alloc_fail(vmd, object, req)) goto again; return (NULL); } /* * At this point we had better have found a good page. */ KASSERT(m != NULL, ("missing page")); found: vm_page_dequeue(m); vm_page_alloc_check(m); /* * Initialize the page. Only the PG_ZERO flag is inherited. */ flags = 0; if ((req & VM_ALLOC_ZERO) != 0) flags = PG_ZERO; flags &= m->flags; if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; m->flags = flags; m->aflags = 0; m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; m->busy_lock = VPB_UNBUSIED; if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) m->busy_lock = VPB_SINGLE_EXCLUSIVER; if ((req & VM_ALLOC_SBUSY) != 0) m->busy_lock = VPB_SHARERS_WORD(1); if (req & VM_ALLOC_WIRED) { /* * The page lock is not required for wiring a page until that * page is inserted into the object. */ vm_wire_add(1); m->wire_count = 1; } m->act_count = 0; if (object != NULL) { if (vm_page_insert_after(m, object, pindex, mpred)) { if (req & VM_ALLOC_WIRED) { vm_wire_sub(1); m->wire_count = 0; } KASSERT(m->object == NULL, ("page %p has object", m)); m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; /* Don't change PG_ZERO. */ vm_page_free_toq(m); if (req & VM_ALLOC_WAITFAIL) { VM_OBJECT_WUNLOCK(object); vm_radix_wait(); VM_OBJECT_WLOCK(object); } return (NULL); } /* Ignore device objects; the pager sets "memattr" for them. */ if (object->memattr != VM_MEMATTR_DEFAULT && (object->flags & OBJ_FICTITIOUS) == 0) pmap_page_set_memattr(m, object->memattr); } else m->pindex = pindex; return (m); } /* * vm_page_alloc_contig: * * Allocate a contiguous set of physical pages of the given size "npages" * from the free lists. All of the physical pages must be at or above * the given physical address "low" and below the given physical address * "high". The given value "alignment" determines the alignment of the * first physical page in the set. If the given value "boundary" is * non-zero, then the set of physical pages cannot cross any physical * address boundary that is a multiple of that value. Both "alignment" * and "boundary" must be a power of two. * * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, * then the memory attribute setting for the physical pages is configured * to the object's memory attribute setting. Otherwise, the memory * attribute setting for the physical pages is configured to "memattr", * overriding the object's memory attribute setting. However, if the * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the * memory attribute setting for the physical pages cannot be configured * to VM_MEMATTR_DEFAULT. * * The specified object may not contain fictitious pages. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NODUMP do not include the page in a kernel core dump * VM_ALLOC_NOOBJ page is not associated with an object and * should not be exclusive busy * VM_ALLOC_SBUSY shared busy the allocated page * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); do { m = vm_page_alloc_contig_domain(object, pindex, domain, req, npages, low, high, alignment, boundary, memattr); if (m != NULL) break; } while (vm_domainset_iter_page(&di, object, &domain) == 0); return (m); } vm_page_t vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domain *vmd; vm_page_t m, m_ret, mpred; u_int busy_lock, flags, oflags; mpred = NULL; /* XXX: pacify gcc */ KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, req)); KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, ("Can't sleep and retry object insertion.")); if (object != NULL) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_FICTITIOUS) == 0, ("vm_page_alloc_contig: object %p has fictitious pages", object)); } KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); if (object != NULL) { mpred = vm_radix_lookup_le(&object->rtree, pindex); KASSERT(mpred == NULL || mpred->pindex != pindex, ("vm_page_alloc_contig: pindex already allocated")); } /* * Can we allocate the pages without the number of free pages falling * below the lower bound for the allocation class? */ again: #if VM_NRESERVLEVEL > 0 /* * Can we allocate the pages from a reservation? */ if (vm_object_reserv(object) && ((m_ret = vm_reserv_extend_contig(req, object, pindex, domain, npages, low, high, alignment, boundary, mpred)) != NULL || (m_ret = vm_reserv_alloc_contig(req, object, pindex, domain, npages, low, high, alignment, boundary, mpred)) != NULL)) { domain = vm_phys_domain(m_ret); vmd = VM_DOMAIN(domain); goto found; } #endif m_ret = NULL; vmd = VM_DOMAIN(domain); if (vm_domain_allocate(vmd, req, npages)) { /* * allocate them from the free page queues. */ vm_domain_free_lock(vmd); m_ret = vm_phys_alloc_contig(domain, npages, low, high, alignment, boundary); vm_domain_free_unlock(vmd); if (m_ret == NULL) { vm_domain_freecnt_inc(vmd, npages); #if VM_NRESERVLEVEL > 0 if (vm_reserv_reclaim_contig(domain, npages, low, high, alignment, boundary)) goto again; #endif } } if (m_ret == NULL) { if (vm_domain_alloc_fail(vmd, object, req)) goto again; return (NULL); } #if VM_NRESERVLEVEL > 0 found: #endif for (m = m_ret; m < &m_ret[npages]; m++) { vm_page_dequeue(m); vm_page_alloc_check(m); } /* * Initialize the pages. Only the PG_ZERO flag is inherited. */ flags = 0; if ((req & VM_ALLOC_ZERO) != 0) flags = PG_ZERO; if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; busy_lock = VPB_UNBUSIED; if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) busy_lock = VPB_SINGLE_EXCLUSIVER; if ((req & VM_ALLOC_SBUSY) != 0) busy_lock = VPB_SHARERS_WORD(1); if ((req & VM_ALLOC_WIRED) != 0) vm_wire_add(npages); if (object != NULL) { if (object->memattr != VM_MEMATTR_DEFAULT && memattr == VM_MEMATTR_DEFAULT) memattr = object->memattr; } for (m = m_ret; m < &m_ret[npages]; m++) { m->aflags = 0; m->flags = (m->flags | PG_NODUMP) & flags; m->busy_lock = busy_lock; if ((req & VM_ALLOC_WIRED) != 0) m->wire_count = 1; m->act_count = 0; m->oflags = oflags; if (object != NULL) { if (vm_page_insert_after(m, object, pindex, mpred)) { if ((req & VM_ALLOC_WIRED) != 0) vm_wire_sub(npages); KASSERT(m->object == NULL, ("page %p has object", m)); mpred = m; for (m = m_ret; m < &m_ret[npages]; m++) { if (m <= mpred && (req & VM_ALLOC_WIRED) != 0) m->wire_count = 0; m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; /* Don't change PG_ZERO. */ vm_page_free_toq(m); } if (req & VM_ALLOC_WAITFAIL) { VM_OBJECT_WUNLOCK(object); vm_radix_wait(); VM_OBJECT_WLOCK(object); } return (NULL); } mpred = m; } else m->pindex = pindex; if (memattr != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, memattr); pindex++; } return (m_ret); } /* * Check a page that has been freshly dequeued from a freelist. */ static void vm_page_alloc_check(vm_page_t m) { KASSERT(m->object == NULL, ("page %p has object", m)); KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, ("page %p has unexpected queue %d, flags %#x", m, m->queue, (m->aflags & PGA_QUEUE_STATE_MASK))); KASSERT(!vm_page_held(m), ("page %p is held", m)); KASSERT(!vm_page_busied(m), ("page %p is busy", m)); KASSERT(m->dirty == 0, ("page %p is dirty", m)); KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has unexpected memattr %d", m, pmap_page_get_memattr(m))); KASSERT(m->valid == 0, ("free page %p is valid", m)); } /* * vm_page_alloc_freelist: * * Allocate a physical page from the specified free page list. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_COUNT(number) the number of additional pages that the caller * intends to allocate * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc_freelist(int freelist, int req) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { m = vm_page_alloc_freelist_domain(domain, freelist, req); if (m != NULL) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (m); } vm_page_t vm_page_alloc_freelist_domain(int domain, int freelist, int req) { struct vm_domain *vmd; vm_page_t m; u_int flags; m = NULL; vmd = VM_DOMAIN(domain); again: if (vm_domain_allocate(vmd, req, 1)) { vm_domain_free_lock(vmd); m = vm_phys_alloc_freelist_pages(domain, freelist, VM_FREEPOOL_DIRECT, 0); vm_domain_free_unlock(vmd); if (m == NULL) vm_domain_freecnt_inc(vmd, 1); } if (m == NULL) { if (vm_domain_alloc_fail(vmd, NULL, req)) goto again; return (NULL); } vm_page_dequeue(m); vm_page_alloc_check(m); /* * Initialize the page. Only the PG_ZERO flag is inherited. */ m->aflags = 0; flags = 0; if ((req & VM_ALLOC_ZERO) != 0) flags = PG_ZERO; m->flags &= flags; if ((req & VM_ALLOC_WIRED) != 0) { /* * The page lock is not required for wiring a page that does * not belong to an object. */ vm_wire_add(1); m->wire_count = 1; } /* Unmanaged pages don't use "act_count". */ m->oflags = VPO_UNMANAGED; return (m); } static int vm_page_import(void *arg, void **store, int cnt, int domain, int flags) { struct vm_domain *vmd; int i; vmd = arg; /* Only import if we can bring in a full bucket. */ if (cnt == 1 || !vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) return (0); domain = vmd->vmd_domain; vm_domain_free_lock(vmd); i = vm_phys_alloc_npages(domain, VM_FREEPOOL_DEFAULT, cnt, (vm_page_t *)store); vm_domain_free_unlock(vmd); if (cnt != i) vm_domain_freecnt_inc(vmd, cnt - i); return (i); } static void vm_page_release(void *arg, void **store, int cnt) { struct vm_domain *vmd; vm_page_t m; int i; vmd = arg; vm_domain_free_lock(vmd); for (i = 0; i < cnt; i++) { m = (vm_page_t)store[i]; vm_phys_free_pages(m, 0); } vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, cnt); } #define VPSC_ANY 0 /* No restrictions. */ #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ #define VPSC_NOSUPER 2 /* Skip superpages. */ /* * vm_page_scan_contig: * * Scan vm_page_array[] between the specified entries "m_start" and * "m_end" for a run of contiguous physical pages that satisfy the * specified conditions, and return the lowest page in the run. The * specified "alignment" determines the alignment of the lowest physical * page in the run. If the specified "boundary" is non-zero, then the * run of physical pages cannot span a physical address that is a * multiple of "boundary". * * "m_end" is never dereferenced, so it need not point to a vm_page * structure within vm_page_array[]. * * "npages" must be greater than zero. "m_start" and "m_end" must not * span a hole (or discontiguity) in the physical address space. Both * "alignment" and "boundary" must be a power of two. */ vm_page_t vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, u_long alignment, vm_paddr_t boundary, int options) { struct mtx *m_mtx; vm_object_t object; vm_paddr_t pa; vm_page_t m, m_run; #if VM_NRESERVLEVEL > 0 int level; #endif int m_inc, order, run_ext, run_len; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); m_run = NULL; run_len = 0; m_mtx = NULL; for (m = m_start; m < m_end && run_len < npages; m += m_inc) { KASSERT((m->flags & PG_MARKER) == 0, ("page %p is PG_MARKER", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->wire_count == 1, ("fictitious page %p has invalid wire count", m)); /* * If the current page would be the start of a run, check its * physical address against the end, alignment, and boundary * conditions. If it doesn't satisfy these conditions, either * terminate the scan or advance to the next page that * satisfies the failed condition. */ if (run_len == 0) { KASSERT(m_run == NULL, ("m_run != NULL")); if (m + npages > m_end) break; pa = VM_PAGE_TO_PHYS(m); if ((pa & (alignment - 1)) != 0) { m_inc = atop(roundup2(pa, alignment) - pa); continue; } if (rounddown2(pa ^ (pa + ptoa(npages) - 1), boundary) != 0) { m_inc = atop(roundup2(pa, boundary) - pa); continue; } } else KASSERT(m_run != NULL, ("m_run == NULL")); vm_page_change_lock(m, &m_mtx); m_inc = 1; retry: if (vm_page_held(m)) run_ext = 0; #if VM_NRESERVLEVEL > 0 else if ((level = vm_reserv_level(m)) >= 0 && (options & VPSC_NORESERV) != 0) { run_ext = 0; /* Advance to the end of the reservation. */ pa = VM_PAGE_TO_PHYS(m); m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - pa); } #endif else if ((object = m->object) != NULL) { /* * The page is considered eligible for relocation if * and only if it could be laundered or reclaimed by * the page daemon. */ if (!VM_OBJECT_TRYRLOCK(object)) { mtx_unlock(m_mtx); VM_OBJECT_RLOCK(object); mtx_lock(m_mtx); if (m->object != object) { /* * The page may have been freed. */ VM_OBJECT_RUNLOCK(object); goto retry; } else if (vm_page_held(m)) { run_ext = 0; goto unlock; } } KASSERT((m->flags & PG_UNHOLDFREE) == 0, ("page %p is PG_UNHOLDFREE", m)); /* Don't care: PG_NODUMP, PG_ZERO. */ if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP && object->type != OBJT_VNODE) { run_ext = 0; #if VM_NRESERVLEVEL > 0 } else if ((options & VPSC_NOSUPER) != 0 && (level = vm_reserv_level_iffullpop(m)) >= 0) { run_ext = 0; /* Advance to the end of the superpage. */ pa = VM_PAGE_TO_PHYS(m); m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - pa); #endif } else if (object->memattr == VM_MEMATTR_DEFAULT && vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { /* * The page is allocated but eligible for * relocation. Extend the current run by one * page. */ KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has an unexpected memattr", m)); KASSERT((m->oflags & (VPO_SWAPINPROG | VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, ("page %p has unexpected oflags", m)); /* Don't care: VPO_NOSYNC. */ run_ext = 1; } else run_ext = 0; unlock: VM_OBJECT_RUNLOCK(object); #if VM_NRESERVLEVEL > 0 } else if (level >= 0) { /* * The page is reserved but not yet allocated. In * other words, it is still free. Extend the current * run by one page. */ run_ext = 1; #endif } else if ((order = m->order) < VM_NFREEORDER) { /* * The page is enqueued in the physical memory * allocator's free page queues. Moreover, it is the * first page in a power-of-two-sized run of * contiguous free pages. Add these pages to the end * of the current run, and jump ahead. */ run_ext = 1 << order; m_inc = 1 << order; } else { /* * Skip the page for one of the following reasons: (1) * It is enqueued in the physical memory allocator's * free page queues. However, it is not the first * page in a run of contiguous free pages. (This case * rarely occurs because the scan is performed in * ascending order.) (2) It is not reserved, and it is * transitioning from free to allocated. (Conversely, * the transition from allocated to free for managed * pages is blocked by the page lock.) (3) It is * allocated but not contained by an object and not * wired, e.g., allocated by Xen's balloon driver. */ run_ext = 0; } /* * Extend or reset the current run of pages. */ if (run_ext > 0) { if (run_len == 0) m_run = m; run_len += run_ext; } else { if (run_len > 0) { m_run = NULL; run_len = 0; } } } if (m_mtx != NULL) mtx_unlock(m_mtx); if (run_len >= npages) return (m_run); return (NULL); } /* * vm_page_reclaim_run: * * Try to relocate each of the allocated virtual pages within the * specified run of physical pages to a new physical address. Free the * physical pages underlying the relocated virtual pages. A virtual page * is relocatable if and only if it could be laundered or reclaimed by * the page daemon. Whenever possible, a virtual page is relocated to a * physical address above "high". * * Returns 0 if every physical page within the run was already free or * just freed by a successful relocation. Otherwise, returns a non-zero * value indicating why the last attempt to relocate a virtual page was * unsuccessful. * * "req_class" must be an allocation class. */ static int vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, vm_paddr_t high) { struct vm_domain *vmd; struct mtx *m_mtx; struct spglist free; vm_object_t object; vm_paddr_t pa; vm_page_t m, m_end, m_new; int error, order, req; KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, ("req_class is not an allocation class")); SLIST_INIT(&free); error = 0; m = m_run; m_end = m_run + npages; m_mtx = NULL; for (; error == 0 && m < m_end; m++) { KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, ("page %p is PG_FICTITIOUS or PG_MARKER", m)); /* * Avoid releasing and reacquiring the same page lock. */ vm_page_change_lock(m, &m_mtx); retry: if (vm_page_held(m)) error = EBUSY; else if ((object = m->object) != NULL) { /* * The page is relocated if and only if it could be * laundered or reclaimed by the page daemon. */ if (!VM_OBJECT_TRYWLOCK(object)) { mtx_unlock(m_mtx); VM_OBJECT_WLOCK(object); mtx_lock(m_mtx); if (m->object != object) { /* * The page may have been freed. */ VM_OBJECT_WUNLOCK(object); goto retry; } else if (vm_page_held(m)) { error = EBUSY; goto unlock; } } KASSERT((m->flags & PG_UNHOLDFREE) == 0, ("page %p is PG_UNHOLDFREE", m)); /* Don't care: PG_NODUMP, PG_ZERO. */ if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP && object->type != OBJT_VNODE) error = EINVAL; else if (object->memattr != VM_MEMATTR_DEFAULT) error = EINVAL; else if (vm_page_queue(m) != PQ_NONE && !vm_page_busied(m)) { KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has an unexpected memattr", m)); KASSERT((m->oflags & (VPO_SWAPINPROG | VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, ("page %p has unexpected oflags", m)); /* Don't care: VPO_NOSYNC. */ if (m->valid != 0) { /* * First, try to allocate a new page * that is above "high". Failing * that, try to allocate a new page * that is below "m_run". Allocate * the new page between the end of * "m_run" and "high" only as a last * resort. */ req = req_class | VM_ALLOC_NOOBJ; if ((m->flags & PG_NODUMP) != 0) req |= VM_ALLOC_NODUMP; if (trunc_page(high) != ~(vm_paddr_t)PAGE_MASK) { m_new = vm_page_alloc_contig( NULL, 0, req, 1, round_page(high), ~(vm_paddr_t)0, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } else m_new = NULL; if (m_new == NULL) { pa = VM_PAGE_TO_PHYS(m_run); m_new = vm_page_alloc_contig( NULL, 0, req, 1, 0, pa - 1, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } if (m_new == NULL) { pa += ptoa(npages); m_new = vm_page_alloc_contig( NULL, 0, req, 1, pa, high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } if (m_new == NULL) { error = ENOMEM; goto unlock; } KASSERT(m_new->wire_count == 0, ("page %p is wired", m_new)); /* * Replace "m" with the new page. For * vm_page_replace(), "m" must be busy * and dequeued. Finally, change "m" * as if vm_page_free() was called. */ if (object->ref_count != 0) pmap_remove_all(m); m_new->aflags = m->aflags & ~PGA_QUEUE_STATE_MASK; KASSERT(m_new->oflags == VPO_UNMANAGED, ("page %p is managed", m_new)); m_new->oflags = m->oflags & VPO_NOSYNC; pmap_copy_page(m, m_new); m_new->valid = m->valid; m_new->dirty = m->dirty; m->flags &= ~PG_ZERO; vm_page_xbusy(m); vm_page_dequeue(m); vm_page_replace_checked(m_new, object, m->pindex, m); if (vm_page_free_prep(m)) SLIST_INSERT_HEAD(&free, m, plinks.s.ss); /* * The new page must be deactivated * before the object is unlocked. */ vm_page_change_lock(m_new, &m_mtx); vm_page_deactivate(m_new); } else { m->flags &= ~PG_ZERO; vm_page_dequeue(m); vm_page_remove(m); if (vm_page_free_prep(m)) SLIST_INSERT_HEAD(&free, m, plinks.s.ss); KASSERT(m->dirty == 0, ("page %p is dirty", m)); } } else error = EBUSY; unlock: VM_OBJECT_WUNLOCK(object); } else { MPASS(vm_phys_domain(m) == domain); vmd = VM_DOMAIN(domain); vm_domain_free_lock(vmd); order = m->order; if (order < VM_NFREEORDER) { /* * The page is enqueued in the physical memory * allocator's free page queues. Moreover, it * is the first page in a power-of-two-sized * run of contiguous free pages. Jump ahead * to the last page within that run, and * continue from there. */ m += (1 << order) - 1; } #if VM_NRESERVLEVEL > 0 else if (vm_reserv_is_page_free(m)) order = 0; #endif vm_domain_free_unlock(vmd); if (order == VM_NFREEORDER) error = EINVAL; } } if (m_mtx != NULL) mtx_unlock(m_mtx); if ((m = SLIST_FIRST(&free)) != NULL) { int cnt; vmd = VM_DOMAIN(domain); cnt = 0; vm_domain_free_lock(vmd); do { MPASS(vm_phys_domain(m) == domain); SLIST_REMOVE_HEAD(&free, plinks.s.ss); vm_phys_free_pages(m, 0); cnt++; } while ((m = SLIST_FIRST(&free)) != NULL); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, cnt); } return (error); } #define NRUNS 16 CTASSERT(powerof2(NRUNS)); #define RUN_INDEX(count) ((count) & (NRUNS - 1)) #define MIN_RECLAIM 8 /* * vm_page_reclaim_contig: * * Reclaim allocated, contiguous physical memory satisfying the specified * conditions by relocating the virtual pages using that physical memory. * Returns true if reclamation is successful and false otherwise. Since * relocation requires the allocation of physical pages, reclamation may * fail due to a shortage of free pages. When reclamation fails, callers * are expected to perform vm_wait() before retrying a failed allocation * operation, e.g., vm_page_alloc_contig(). * * The caller must always specify an allocation class through "req". * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * The optional allocation flags are ignored. * * "npages" must be greater than zero. Both "alignment" and "boundary" * must be a power of two. */ bool vm_page_reclaim_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_domain *vmd; vm_paddr_t curr_low; vm_page_t m_run, m_runs[NRUNS]; u_long count, reclaimed; int error, i, options, req_class; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); req_class = req & VM_ALLOC_CLASS_MASK; /* * The page daemon is allowed to dig deeper into the free page list. */ if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) req_class = VM_ALLOC_SYSTEM; /* * Return if the number of free pages cannot satisfy the requested * allocation. */ vmd = VM_DOMAIN(domain); count = vmd->vmd_free_count; if (count < npages + vmd->vmd_free_reserved || (count < npages + vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || (count < npages && req_class == VM_ALLOC_INTERRUPT)) return (false); /* * Scan up to three times, relaxing the restrictions ("options") on * the reclamation of reservations and superpages each time. */ for (options = VPSC_NORESERV;;) { /* * Find the highest runs that satisfy the given constraints * and restrictions, and record them in "m_runs". */ curr_low = low; count = 0; for (;;) { m_run = vm_phys_scan_contig(domain, npages, curr_low, high, alignment, boundary, options); if (m_run == NULL) break; curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); m_runs[RUN_INDEX(count)] = m_run; count++; } /* * Reclaim the highest runs in LIFO (descending) order until * the number of reclaimed pages, "reclaimed", is at least * MIN_RECLAIM. Reset "reclaimed" each time because each * reclamation is idempotent, and runs will (likely) recur * from one scan to the next as restrictions are relaxed. */ reclaimed = 0; for (i = 0; count > 0 && i < NRUNS; i++) { count--; m_run = m_runs[RUN_INDEX(count)]; error = vm_page_reclaim_run(req_class, domain, npages, m_run, high); if (error == 0) { reclaimed += npages; if (reclaimed >= MIN_RECLAIM) return (true); } } /* * Either relax the restrictions on the next scan or return if * the last scan had no restrictions. */ if (options == VPSC_NORESERV) options = VPSC_NOSUPER; else if (options == VPSC_NOSUPER) options = VPSC_ANY; else if (options == VPSC_ANY) return (reclaimed != 0); } } bool vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_domainset_iter di; int domain; bool ret; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { ret = vm_page_reclaim_contig_domain(domain, req, npages, low, high, alignment, boundary); if (ret) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (ret); } /* * Set the domain in the appropriate page level domainset. */ void vm_domain_set(struct vm_domain *vmd) { mtx_lock(&vm_domainset_lock); if (!vmd->vmd_minset && vm_paging_min(vmd)) { vmd->vmd_minset = 1; DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); } if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { vmd->vmd_severeset = 1; DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); } mtx_unlock(&vm_domainset_lock); } /* * Clear the domain from the appropriate page level domainset. */ void vm_domain_clear(struct vm_domain *vmd) { mtx_lock(&vm_domainset_lock); if (vmd->vmd_minset && !vm_paging_min(vmd)) { vmd->vmd_minset = 0; DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); if (vm_min_waiters != 0) { vm_min_waiters = 0; wakeup(&vm_min_domains); } } if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { vmd->vmd_severeset = 0; DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); if (vm_severe_waiters != 0) { vm_severe_waiters = 0; wakeup(&vm_severe_domains); } } /* * If pageout daemon needs pages, then tell it that there are * some free. */ if (vmd->vmd_pageout_pages_needed && vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { wakeup(&vmd->vmd_pageout_pages_needed); vmd->vmd_pageout_pages_needed = 0; } /* See comments in vm_wait_doms(). */ if (vm_pageproc_waiters) { vm_pageproc_waiters = 0; wakeup(&vm_pageproc_waiters); } mtx_unlock(&vm_domainset_lock); } /* * Wait for free pages to exceed the min threshold globally. */ void vm_wait_min(void) { mtx_lock(&vm_domainset_lock); while (vm_page_count_min()) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); } mtx_unlock(&vm_domainset_lock); } /* * Wait for free pages to exceed the severe threshold globally. */ void vm_wait_severe(void) { mtx_lock(&vm_domainset_lock); while (vm_page_count_severe()) { vm_severe_waiters++; msleep(&vm_severe_domains, &vm_domainset_lock, PVM, "vmwait", 0); } mtx_unlock(&vm_domainset_lock); } u_int vm_wait_count(void) { return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); } void vm_wait_doms(const domainset_t *wdoms) { /* * We use racey wakeup synchronization to avoid expensive global * locking for the pageproc when sleeping with a non-specific vm_wait. * To handle this, we only sleep for one tick in this instance. It * is expected that most allocations for the pageproc will come from * kmem or vm_page_grab* which will use the more specific and * race-free vm_wait_domain(). */ if (curproc == pageproc) { mtx_lock(&vm_domainset_lock); vm_pageproc_waiters++; msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP, "pageprocwait", 1); } else { /* * XXX Ideally we would wait only until the allocation could * be satisfied. This condition can cause new allocators to * consume all freed pages while old allocators wait. */ mtx_lock(&vm_domainset_lock); if (vm_page_count_min_set(wdoms)) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PVM | PDROP, "vmwait", 0); } else mtx_unlock(&vm_domainset_lock); } } /* * vm_wait_domain: * * Sleep until free pages are available for allocation. * - Called in various places after failed memory allocations. */ void vm_wait_domain(int domain) { struct vm_domain *vmd; domainset_t wdom; vmd = VM_DOMAIN(domain); vm_domain_free_assert_unlocked(vmd); if (curproc == pageproc) { mtx_lock(&vm_domainset_lock); if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { vmd->vmd_pageout_pages_needed = 1; msleep(&vmd->vmd_pageout_pages_needed, &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); } else mtx_unlock(&vm_domainset_lock); } else { if (pageproc == NULL) panic("vm_wait in early boot"); DOMAINSET_ZERO(&wdom); DOMAINSET_SET(vmd->vmd_domain, &wdom); vm_wait_doms(&wdom); } } /* * vm_wait: * * Sleep until free pages are available for allocation in the * affinity domains of the obj. If obj is NULL, the domain set * for the calling thread is used. * Called in various places after failed memory allocations. */ void vm_wait(vm_object_t obj) { struct domainset *d; d = NULL; /* * Carefully fetch pointers only once: the struct domainset * itself is ummutable but the pointer might change. */ if (obj != NULL) d = obj->domain.dr_policy; if (d == NULL) d = curthread->td_domain.dr_policy; vm_wait_doms(&d->ds_mask); } /* * vm_domain_alloc_fail: * * Called when a page allocation function fails. Informs the * pagedaemon and performs the requested wait. Requires the * domain_free and object lock on entry. Returns with the * object lock held and free lock released. Returns an error when * retry is necessary. * */ static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) { vm_domain_free_assert_unlocked(vmd); atomic_add_int(&vmd->vmd_pageout_deficit, max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { if (object != NULL) VM_OBJECT_WUNLOCK(object); vm_wait_domain(vmd->vmd_domain); if (object != NULL) VM_OBJECT_WLOCK(object); if (req & VM_ALLOC_WAITOK) return (EAGAIN); } return (0); } /* * vm_waitpfault: * * Sleep until free pages are available for allocation. * - Called only in vm_fault so that processes page faulting * can be easily tracked. * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing * processes will be able to grab memory first. Do not change * this balance without careful testing first. */ void vm_waitpfault(struct domainset *dset) { /* * XXX Ideally we would wait only until the allocation could * be satisfied. This condition can cause new allocators to * consume all freed pages while old allocators wait. */ mtx_lock(&vm_domainset_lock); if (vm_page_count_min_set(&dset->ds_mask)) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, "pfault", 0); } else mtx_unlock(&vm_domainset_lock); } struct vm_pagequeue * vm_page_pagequeue(vm_page_t m) { return (&vm_pagequeue_domain(m)->vmd_pagequeues[m->queue]); } static struct mtx * vm_page_pagequeue_lockptr(vm_page_t m) { uint8_t queue; if ((queue = atomic_load_8(&m->queue)) == PQ_NONE) return (NULL); return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue].pq_mutex); } static inline void vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m) { struct vm_domain *vmd; uint8_t qflags; CRITICAL_ASSERT(curthread); vm_pagequeue_assert_locked(pq); /* * The page daemon is allowed to set m->queue = PQ_NONE without * the page queue lock held. In this case it is about to free the page, * which must not have any queue state. */ qflags = atomic_load_8(&m->aflags) & PGA_QUEUE_STATE_MASK; KASSERT(pq == vm_page_pagequeue(m) || qflags == 0, ("page %p doesn't belong to queue %p but has queue state %#x", m, pq, qflags)); if ((qflags & PGA_DEQUEUE) != 0) { if (__predict_true((qflags & PGA_ENQUEUED) != 0)) { TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); vm_pagequeue_cnt_dec(pq); } vm_page_dequeue_complete(m); } else if ((qflags & (PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) { if ((qflags & PGA_ENQUEUED) != 0) TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); else { vm_pagequeue_cnt_inc(pq); vm_page_aflag_set(m, PGA_ENQUEUED); } if ((qflags & PGA_REQUEUE_HEAD) != 0) { KASSERT(m->queue == PQ_INACTIVE, ("head enqueue not supported for page %p", m)); vmd = vm_pagequeue_domain(m); TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); } else TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); /* * PGA_REQUEUE and PGA_REQUEUE_HEAD must be cleared after * setting PGA_ENQUEUED in order to synchronize with the * page daemon. */ vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD); } } static void vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, uint8_t queue) { vm_page_t m; int i; for (i = 0; i < bq->bq_cnt; i++) { m = bq->bq_pa[i]; if (__predict_false(m->queue != queue)) continue; vm_pqbatch_process_page(pq, m); } vm_batchqueue_init(bq); } static void vm_pqbatch_submit_page(vm_page_t m, uint8_t queue) { struct vm_batchqueue *bq; struct vm_pagequeue *pq; int domain; vm_page_assert_locked(m); KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); domain = vm_phys_domain(m); pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue]; critical_enter(); bq = DPCPU_PTR(pqbatch[domain][queue]); if (vm_batchqueue_insert(bq, m)) { critical_exit(); return; } if (!vm_pagequeue_trylock(pq)) { critical_exit(); vm_pagequeue_lock(pq); critical_enter(); bq = DPCPU_PTR(pqbatch[domain][queue]); } vm_pqbatch_process(pq, bq, queue); /* * The page may have been logically dequeued before we acquired the * page queue lock. In this case, the page lock prevents the page * from being logically enqueued elsewhere. */ if (__predict_true(m->queue == queue)) vm_pqbatch_process_page(pq, m); else { KASSERT(m->queue == PQ_NONE, ("invalid queue transition for page %p", m)); KASSERT((m->aflags & PGA_ENQUEUED) == 0, ("page %p is enqueued with invalid queue index", m)); vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK); } vm_pagequeue_unlock(pq); critical_exit(); } /* * vm_page_drain_pqbatch: [ internal use only ] * * Force all per-CPU page queue batch queues to be drained. This is * intended for use in severe memory shortages, to ensure that pages * do not remain stuck in the batch queues. */ void vm_page_drain_pqbatch(void) { struct thread *td; struct vm_domain *vmd; struct vm_pagequeue *pq; int cpu, domain, queue; td = curthread; CPU_FOREACH(cpu) { thread_lock(td); sched_bind(td, cpu); thread_unlock(td); for (domain = 0; domain < vm_ndomains; domain++) { vmd = VM_DOMAIN(domain); for (queue = 0; queue < PQ_COUNT; queue++) { pq = &vmd->vmd_pagequeues[queue]; vm_pagequeue_lock(pq); critical_enter(); vm_pqbatch_process(pq, DPCPU_PTR(pqbatch[domain][queue]), queue); critical_exit(); vm_pagequeue_unlock(pq); } } } thread_lock(td); sched_unbind(td); thread_unlock(td); } /* * Complete the logical removal of a page from a page queue. We must be * careful to synchronize with the page daemon, which may be concurrently * examining the page with only the page lock held. The page must not be * in a state where it appears to be logically enqueued. */ static void vm_page_dequeue_complete(vm_page_t m) { m->queue = PQ_NONE; atomic_thread_fence_rel(); vm_page_aflag_clear(m, PGA_QUEUE_STATE_MASK); } /* * vm_page_dequeue_deferred: [ internal use only ] * * Request removal of the given page from its current page * queue. Physical removal from the queue may be deferred * indefinitely. * * The page must be locked. */ void vm_page_dequeue_deferred(vm_page_t m) { int queue; vm_page_assert_locked(m); queue = atomic_load_8(&m->queue); if (queue == PQ_NONE) { KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0, ("page %p has queue state", m)); return; } if ((m->aflags & PGA_DEQUEUE) == 0) vm_page_aflag_set(m, PGA_DEQUEUE); vm_pqbatch_submit_page(m, queue); } /* * vm_page_dequeue: * * Remove the page from whichever page queue it's in, if any. * The page must either be locked or unallocated. This constraint * ensures that the queue state of the page will remain consistent * after this function returns. */ void vm_page_dequeue(vm_page_t m) { struct mtx *lock, *lock1; struct vm_pagequeue *pq; uint8_t aflags; KASSERT(mtx_owned(vm_page_lockptr(m)) || m->order == VM_NFREEORDER, ("page %p is allocated and unlocked", m)); for (;;) { lock = vm_page_pagequeue_lockptr(m); if (lock == NULL) { /* * A thread may be concurrently executing * vm_page_dequeue_complete(). Ensure that all queue * state is cleared before we return. */ aflags = atomic_load_8(&m->aflags); if ((aflags & PGA_QUEUE_STATE_MASK) == 0) return; KASSERT((aflags & PGA_DEQUEUE) != 0, ("page %p has unexpected queue state flags %#x", m, aflags)); /* * Busy wait until the thread updating queue state is * finished. Such a thread must be executing in a * critical section. */ cpu_spinwait(); continue; } mtx_lock(lock); if ((lock1 = vm_page_pagequeue_lockptr(m)) == lock) break; mtx_unlock(lock); lock = lock1; } KASSERT(lock == vm_page_pagequeue_lockptr(m), ("%s: page %p migrated directly between queues", __func__, m)); KASSERT((m->aflags & PGA_DEQUEUE) != 0 || mtx_owned(vm_page_lockptr(m)), ("%s: queued unlocked page %p", __func__, m)); if ((m->aflags & PGA_ENQUEUED) != 0) { pq = vm_page_pagequeue(m); TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); vm_pagequeue_cnt_dec(pq); } vm_page_dequeue_complete(m); mtx_unlock(lock); } /* * Schedule the given page for insertion into the specified page queue. * Physical insertion of the page may be deferred indefinitely. */ static void vm_page_enqueue(vm_page_t m, uint8_t queue) { vm_page_assert_locked(m); KASSERT(m->queue == PQ_NONE && (m->aflags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p is already enqueued", __func__, m)); m->queue = queue; if ((m->aflags & PGA_REQUEUE) == 0) vm_page_aflag_set(m, PGA_REQUEUE); vm_pqbatch_submit_page(m, queue); } /* * vm_page_requeue: [ internal use only ] * * Schedule a requeue of the given page. * * The page must be locked. */ void vm_page_requeue(vm_page_t m) { vm_page_assert_locked(m); KASSERT(m->queue != PQ_NONE, ("%s: page %p is not logically enqueued", __func__, m)); if ((m->aflags & PGA_REQUEUE) == 0) vm_page_aflag_set(m, PGA_REQUEUE); vm_pqbatch_submit_page(m, atomic_load_8(&m->queue)); } /* * vm_page_activate: * * Put the specified page on the active list (if appropriate). * Ensure that act_count is at least ACT_INIT but do not otherwise * mess with it. * * The page must be locked. */ void vm_page_activate(vm_page_t m) { vm_page_assert_locked(m); if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0) return; if (vm_page_queue(m) == PQ_ACTIVE) { if (m->act_count < ACT_INIT) m->act_count = ACT_INIT; return; } vm_page_dequeue(m); if (m->act_count < ACT_INIT) m->act_count = ACT_INIT; vm_page_enqueue(m, PQ_ACTIVE); } /* * vm_page_free_prep: * * Prepares the given page to be put on the free list, * disassociating it from any VM object. The caller may return * the page to the free list only if this function returns true. * * The object must be locked. The page must be locked if it is * managed. */ bool vm_page_free_prep(vm_page_t m) { #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { uint64_t *p; int i; p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", m, i, (uintmax_t)*p)); } #endif if ((m->oflags & VPO_UNMANAGED) == 0) { vm_page_lock_assert(m, MA_OWNED); KASSERT(!pmap_page_is_mapped(m), ("vm_page_free_prep: freeing mapped page %p", m)); } else KASSERT(m->queue == PQ_NONE, ("vm_page_free_prep: unmanaged page %p is queued", m)); VM_CNT_INC(v_tfree); if (vm_page_sbusied(m)) panic("vm_page_free_prep: freeing busy page %p", m); vm_page_remove(m); /* * If fictitious remove object association and * return. */ if ((m->flags & PG_FICTITIOUS) != 0) { KASSERT(m->wire_count == 1, ("fictitious page %p is not wired", m)); KASSERT(m->queue == PQ_NONE, ("fictitious page %p is queued", m)); return (false); } /* * Pages need not be dequeued before they are returned to the physical * memory allocator, but they must at least be marked for a deferred * dequeue. */ if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_dequeue_deferred(m); m->valid = 0; vm_page_undirty(m); if (m->wire_count != 0) panic("vm_page_free_prep: freeing wired page %p", m); if (m->hold_count != 0) { m->flags &= ~PG_ZERO; KASSERT((m->flags & PG_UNHOLDFREE) == 0, ("vm_page_free_prep: freeing PG_UNHOLDFREE page %p", m)); m->flags |= PG_UNHOLDFREE; return (false); } /* * Restore the default memory attribute to the page. */ if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); #if VM_NRESERVLEVEL > 0 if (vm_reserv_free_page(m)) return (false); #endif return (true); } /* * vm_page_free_toq: * * Returns the given page to the free list, disassociating it * from any VM object. * * The object must be locked. The page must be locked if it is * managed. */ void vm_page_free_toq(vm_page_t m) { struct vm_domain *vmd; if (!vm_page_free_prep(m)) return; vmd = vm_pagequeue_domain(m); if (m->pool == VM_FREEPOOL_DEFAULT && vmd->vmd_pgcache != NULL) { uma_zfree(vmd->vmd_pgcache, m); return; } vm_domain_free_lock(vmd); vm_phys_free_pages(m, 0); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, 1); } /* * vm_page_free_pages_toq: * * Returns a list of pages to the free list, disassociating it * from any VM object. In other words, this is equivalent to * calling vm_page_free_toq() for each page of a list of VM objects. * * The objects must be locked. The pages must be locked if it is * managed. */ void vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) { vm_page_t m; int count; if (SLIST_EMPTY(free)) return; count = 0; while ((m = SLIST_FIRST(free)) != NULL) { count++; SLIST_REMOVE_HEAD(free, plinks.s.ss); vm_page_free_toq(m); } if (update_wire_count) vm_wire_sub(count); } /* * vm_page_wire: * * Mark this page as wired down. If the page is fictitious, then * its wire count must remain one. * * The page must be locked. */ void vm_page_wire(vm_page_t m) { vm_page_assert_locked(m); if ((m->flags & PG_FICTITIOUS) != 0) { KASSERT(m->wire_count == 1, ("vm_page_wire: fictitious page %p's wire count isn't one", m)); return; } if (m->wire_count == 0) { KASSERT((m->oflags & VPO_UNMANAGED) == 0 || m->queue == PQ_NONE, ("vm_page_wire: unmanaged page %p is queued", m)); vm_wire_add(1); } m->wire_count++; KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); } /* * vm_page_unwire: * * Release one wiring of the specified page, potentially allowing it to be * paged out. Returns TRUE if the number of wirings transitions to zero and * FALSE otherwise. * * Only managed pages belonging to an object can be paged out. If the number * of wirings transitions to zero and the page is eligible for page out, then * the page is added to the specified paging queue (unless PQ_NONE is * specified, in which case the page is dequeued if it belongs to a paging * queue). * * If a page is fictitious, then its wire count must always be one. * * A managed page must be locked. */ bool vm_page_unwire(vm_page_t m, uint8_t queue) { bool unwired; KASSERT(queue < PQ_COUNT || queue == PQ_NONE, ("vm_page_unwire: invalid queue %u request for page %p", queue, m)); if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_assert_locked(m); unwired = vm_page_unwire_noq(m); if (!unwired || (m->oflags & VPO_UNMANAGED) != 0 || m->object == NULL) return (unwired); if (vm_page_queue(m) == queue) { if (queue == PQ_ACTIVE) vm_page_reference(m); else if (queue != PQ_NONE) vm_page_requeue(m); } else { vm_page_dequeue(m); if (queue != PQ_NONE) { vm_page_enqueue(m, queue); if (queue == PQ_ACTIVE) /* Initialize act_count. */ vm_page_activate(m); } } return (unwired); } /* * * vm_page_unwire_noq: * * Unwire a page without (re-)inserting it into a page queue. It is up * to the caller to enqueue, requeue, or free the page as appropriate. * In most cases, vm_page_unwire() should be used instead. */ bool vm_page_unwire_noq(vm_page_t m) { if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_assert_locked(m); if ((m->flags & PG_FICTITIOUS) != 0) { KASSERT(m->wire_count == 1, ("vm_page_unwire: fictitious page %p's wire count isn't one", m)); return (false); } if (m->wire_count == 0) panic("vm_page_unwire: page %p's wire count is zero", m); m->wire_count--; if (m->wire_count == 0) { vm_wire_sub(1); return (true); } else return (false); } /* * Move the specified page to the tail of the inactive queue, or requeue * the page if it is already in the inactive queue. * * The page must be locked. */ void vm_page_deactivate(vm_page_t m) { vm_page_assert_locked(m); if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0) return; if (!vm_page_inactive(m)) { vm_page_dequeue(m); vm_page_enqueue(m, PQ_INACTIVE); } else vm_page_requeue(m); } /* * Move the specified page close to the head of the inactive queue, * bypassing LRU. A marker page is used to maintain FIFO ordering. * As with regular enqueues, we use a per-CPU batch queue to reduce * contention on the page queue lock. * * The page must be locked. */ void vm_page_deactivate_noreuse(vm_page_t m) { vm_page_assert_locked(m); if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0) return; if (!vm_page_inactive(m)) { vm_page_dequeue(m); m->queue = PQ_INACTIVE; } if ((m->aflags & PGA_REQUEUE_HEAD) == 0) vm_page_aflag_set(m, PGA_REQUEUE_HEAD); vm_pqbatch_submit_page(m, PQ_INACTIVE); } /* * vm_page_launder * * Put a page in the laundry, or requeue it if it is already there. */ void vm_page_launder(vm_page_t m) { vm_page_assert_locked(m); if (m->wire_count > 0 || (m->oflags & VPO_UNMANAGED) != 0) return; if (vm_page_in_laundry(m)) vm_page_requeue(m); else { vm_page_dequeue(m); vm_page_enqueue(m, PQ_LAUNDRY); } } /* * vm_page_unswappable * * Put a page in the PQ_UNSWAPPABLE holding queue. */ void vm_page_unswappable(vm_page_t m) { vm_page_assert_locked(m); KASSERT(m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0, ("page %p already unswappable", m)); vm_page_dequeue(m); vm_page_enqueue(m, PQ_UNSWAPPABLE); } /* * Attempt to free the page. If it cannot be freed, do nothing. Returns true * if the page is freed and false otherwise. * * The page must be managed. The page and its containing object must be * locked. */ bool vm_page_try_to_free(vm_page_t m) { vm_page_assert_locked(m); VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m)); if (m->dirty != 0 || vm_page_held(m) || vm_page_busied(m)) return (false); if (m->object->ref_count != 0) { pmap_remove_all(m); if (m->dirty != 0) return (false); } vm_page_free(m); return (true); } /* * vm_page_advise * * Apply the specified advice to the given page. * * The object and page must be locked. */ void vm_page_advise(vm_page_t m, int advice) { vm_page_assert_locked(m); VM_OBJECT_ASSERT_WLOCKED(m->object); if (advice == MADV_FREE) /* * Mark the page clean. This will allow the page to be freed * without first paging it out. MADV_FREE pages are often * quickly reused by malloc(3), so we do not do anything that * would result in a page fault on a later access. */ vm_page_undirty(m); else if (advice != MADV_DONTNEED) { if (advice == MADV_WILLNEED) vm_page_activate(m); return; } /* * Clear any references to the page. Otherwise, the page daemon will * immediately reactivate the page. */ vm_page_aflag_clear(m, PGA_REFERENCED); if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) vm_page_dirty(m); /* * Place clean pages near the head of the inactive queue rather than * the tail, thus defeating the queue's LRU operation and ensuring that * the page will be reused quickly. Dirty pages not already in the * laundry are moved there. */ if (m->dirty == 0) vm_page_deactivate_noreuse(m); else if (!vm_page_in_laundry(m)) vm_page_launder(m); } /* * Grab a page, waiting until we are waken up due to the page * changing state. We keep on waiting, if the page continues * to be in the object. If the page doesn't exist, first allocate it * and then conditionally zero it. * * This routine may sleep. * * The object must be locked on entry. The lock will, however, be released * and reacquired if the routine sleeps. */ vm_page_t vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; int sleep; int pflags; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL); if ((allocflags & VM_ALLOC_NOWAIT) == 0) pflags |= VM_ALLOC_WAITFAIL; retrylookup: if ((m = vm_page_lookup(object, pindex)) != NULL) { sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? vm_page_xbusied(m) : vm_page_busied(m); if (sleep) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (NULL); /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ vm_page_aflag_set(m, PGA_REFERENCED); vm_page_lock(m); VM_OBJECT_WUNLOCK(object); vm_page_busy_sleep(m, "pgrbwt", (allocflags & VM_ALLOC_IGN_SBUSY) != 0); VM_OBJECT_WLOCK(object); goto retrylookup; } else { if ((allocflags & VM_ALLOC_WIRED) != 0) { vm_page_lock(m); vm_page_wire(m); vm_page_unlock(m); } if ((allocflags & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) vm_page_xbusy(m); if ((allocflags & VM_ALLOC_SBUSY) != 0) vm_page_sbusy(m); return (m); } } m = vm_page_alloc(object, pindex, pflags); if (m == NULL) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (NULL); goto retrylookup; } if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); return (m); } /* * Return the specified range of pages from the given object. For each * page offset within the range, if a page already exists within the object * at that offset and it is busy, then wait for it to change state. If, * instead, the page doesn't exist, then allocate it. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs the pages * * The caller must always specify that the pages are to be busied and/or * wired. * * optional allocation flags: * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NOWAIT do not sleep * VM_ALLOC_SBUSY set page to sbusy state * VM_ALLOC_WIRED wire the pages * VM_ALLOC_ZERO zero and validate any invalid pages * * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it * may return a partial prefix of the requested range. */ int vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count) { vm_page_t m, mpred; int pflags; int i; bool sleep; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || (allocflags & VM_ALLOC_WIRED) != 0, ("vm_page_grab_pages: the pages must be busied or wired")); KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab_pages: VM_ALLOC_SBUSY/IGN_SBUSY mismatch")); if (count == 0) return (0); pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | VM_ALLOC_IGN_SBUSY); if ((allocflags & VM_ALLOC_NOWAIT) == 0) pflags |= VM_ALLOC_WAITFAIL; i = 0; retrylookup: m = vm_radix_lookup_le(&object->rtree, pindex + i); if (m == NULL || m->pindex != pindex + i) { mpred = m; m = NULL; } else mpred = TAILQ_PREV(m, pglist, listq); for (; i < count; i++) { if (m != NULL) { sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ? vm_page_xbusied(m) : vm_page_busied(m); if (sleep) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) break; /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ vm_page_aflag_set(m, PGA_REFERENCED); vm_page_lock(m); VM_OBJECT_WUNLOCK(object); vm_page_busy_sleep(m, "grbmaw", (allocflags & VM_ALLOC_IGN_SBUSY) != 0); VM_OBJECT_WLOCK(object); goto retrylookup; } if ((allocflags & VM_ALLOC_WIRED) != 0) { vm_page_lock(m); vm_page_wire(m); vm_page_unlock(m); } if ((allocflags & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0) vm_page_xbusy(m); if ((allocflags & VM_ALLOC_SBUSY) != 0) vm_page_sbusy(m); } else { m = vm_page_alloc_after(object, pindex + i, pflags | VM_ALLOC_COUNT(count - i), mpred); if (m == NULL) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) break; goto retrylookup; } } if (m->valid == 0 && (allocflags & VM_ALLOC_ZERO) != 0) { if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); m->valid = VM_PAGE_BITS_ALL; } ma[i] = mpred = m; m = vm_page_next(m); } return (i); } /* * Mapping function for valid or dirty bits in a page. * * Inputs are required to range within a page. */ vm_page_bits_t vm_page_bits(int base, int size) { int first_bit; int last_bit; KASSERT( base + size <= PAGE_SIZE, ("vm_page_bits: illegal base/size %d/%d", base, size) ); if (size == 0) /* handle degenerate case */ return (0); first_bit = base >> DEV_BSHIFT; last_bit = (base + size - 1) >> DEV_BSHIFT; return (((vm_page_bits_t)2 << last_bit) - ((vm_page_bits_t)1 << first_bit)); } /* * vm_page_set_valid_range: * * Sets portions of a page valid. The arguments are expected * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive * of any partial chunks touched by the range. The invalid portion of * such chunks will be zeroed. * * (base + size) must be less then or equal to PAGE_SIZE. */ void vm_page_set_valid_range(vm_page_t m, int base, int size) { int endoff, frag; VM_OBJECT_ASSERT_WLOCKED(m->object); if (size == 0) /* handle degenerate case */ return; /* * If the base is not DEV_BSIZE aligned and the valid * bit is clear, we have to zero out a portion of the * first block. */ if ((frag = rounddown2(base, DEV_BSIZE)) != base && (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, frag, base - frag); /* * If the ending offset is not DEV_BSIZE aligned and the * valid bit is clear, we have to zero out a portion of * the last block. */ endoff = base + size; if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, endoff, DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); /* * Assert that no previously invalid block that is now being validated * is already dirty. */ KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, ("vm_page_set_valid_range: page %p is dirty", m)); /* * Set valid bits inclusive of any overlap. */ m->valid |= vm_page_bits(base, size); } /* * Clear the given bits from the specified page's dirty field. */ static __inline void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) { uintptr_t addr; #if PAGE_SIZE < 16384 int shift; #endif /* * If the object is locked and the page is neither exclusive busy nor * write mapped, then the page's dirty field cannot possibly be * set by a concurrent pmap operation. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) m->dirty &= ~pagebits; else { /* * The pmap layer can call vm_page_dirty() without * holding a distinguished lock. The combination of * the object's lock and an atomic operation suffice * to guarantee consistency of the page dirty field. * * For PAGE_SIZE == 32768 case, compiler already * properly aligns the dirty field, so no forcible * alignment is needed. Only require existence of * atomic_clear_64 when page size is 32768. */ addr = (uintptr_t)&m->dirty; #if PAGE_SIZE == 32768 atomic_clear_64((uint64_t *)addr, pagebits); #elif PAGE_SIZE == 16384 atomic_clear_32((uint32_t *)addr, pagebits); #else /* PAGE_SIZE <= 8192 */ /* * Use a trick to perform a 32-bit atomic on the * containing aligned word, to not depend on the existence * of atomic_clear_{8, 16}. */ shift = addr & (sizeof(uint32_t) - 1); #if BYTE_ORDER == BIG_ENDIAN shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY; #else shift *= NBBY; #endif addr &= ~(sizeof(uint32_t) - 1); atomic_clear_32((uint32_t *)addr, pagebits << shift); #endif /* PAGE_SIZE */ } } /* * vm_page_set_validclean: * * Sets portions of a page valid and clean. The arguments are expected * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive * of any partial chunks touched by the range. The invalid portion of * such chunks will be zero'd. * * (base + size) must be less then or equal to PAGE_SIZE. */ void vm_page_set_validclean(vm_page_t m, int base, int size) { vm_page_bits_t oldvalid, pagebits; int endoff, frag; VM_OBJECT_ASSERT_WLOCKED(m->object); if (size == 0) /* handle degenerate case */ return; /* * If the base is not DEV_BSIZE aligned and the valid * bit is clear, we have to zero out a portion of the * first block. */ if ((frag = rounddown2(base, DEV_BSIZE)) != base && (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, frag, base - frag); /* * If the ending offset is not DEV_BSIZE aligned and the * valid bit is clear, we have to zero out a portion of * the last block. */ endoff = base + size; if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, endoff, DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); /* * Set valid, clear dirty bits. If validating the entire * page we can safely clear the pmap modify bit. We also * use this opportunity to clear the VPO_NOSYNC flag. If a process * takes a write fault on a MAP_NOSYNC memory area the flag will * be set again. * * We set valid bits inclusive of any overlap, but we can only * clear dirty bits for DEV_BSIZE chunks that are fully within * the range. */ oldvalid = m->valid; pagebits = vm_page_bits(base, size); m->valid |= pagebits; #if 0 /* NOT YET */ if ((frag = base & (DEV_BSIZE - 1)) != 0) { frag = DEV_BSIZE - frag; base += frag; size -= frag; if (size < 0) size = 0; } pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); #endif if (base == 0 && size == PAGE_SIZE) { /* * The page can only be modified within the pmap if it is * mapped, and it can only be mapped if it was previously * fully valid. */ if (oldvalid == VM_PAGE_BITS_ALL) /* * Perform the pmap_clear_modify() first. Otherwise, * a concurrent pmap operation, such as * pmap_protect(), could clear a modification in the * pmap and set the dirty field on the page before * pmap_clear_modify() had begun and after the dirty * field was cleared here. */ pmap_clear_modify(m); m->dirty = 0; m->oflags &= ~VPO_NOSYNC; } else if (oldvalid != VM_PAGE_BITS_ALL) m->dirty &= ~pagebits; else vm_page_clear_dirty_mask(m, pagebits); } void vm_page_clear_dirty(vm_page_t m, int base, int size) { vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); } /* * vm_page_set_invalid: * * Invalidates DEV_BSIZE'd chunks within a page. Both the * valid and dirty bits for the effected areas are cleared. */ void vm_page_set_invalid(vm_page_t m, int base, int size) { vm_page_bits_t bits; vm_object_t object; object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + size >= object->un_pager.vnp.vnp_size) bits = VM_PAGE_BITS_ALL; else bits = vm_page_bits(base, size); if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL && bits != 0) pmap_remove_all(m); KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) || !pmap_page_is_mapped(m), ("vm_page_set_invalid: page %p is mapped", m)); m->valid &= ~bits; m->dirty &= ~bits; } /* * vm_page_zero_invalid() * * The kernel assumes that the invalid portions of a page contain * garbage, but such pages can be mapped into memory by user code. * When this occurs, we must zero out the non-valid portions of the * page so user code sees what it expects. * * Pages are most often semi-valid when the end of a file is mapped * into memory and the file's size is not page aligned. */ void vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) { int b; int i; VM_OBJECT_ASSERT_WLOCKED(m->object); /* * Scan the valid bits looking for invalid sections that * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the * valid bit may be set ) have already been zeroed by * vm_page_set_validclean(). */ for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { if (i == (PAGE_SIZE / DEV_BSIZE) || (m->valid & ((vm_page_bits_t)1 << i))) { if (i > b) { pmap_zero_page_area(m, b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); } b = i + 1; } } /* * setvalid is TRUE when we can safely set the zero'd areas * as being valid. We can do this if there are no cache consistancy * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. */ if (setvalid) m->valid = VM_PAGE_BITS_ALL; } /* * vm_page_is_valid: * * Is (partial) page valid? Note that the case where size == 0 * will return FALSE in the degenerate case where the page is * entirely invalid, and TRUE otherwise. */ int vm_page_is_valid(vm_page_t m, int base, int size) { vm_page_bits_t bits; VM_OBJECT_ASSERT_LOCKED(m->object); bits = vm_page_bits(base, size); return (m->valid != 0 && (m->valid & bits) == bits); } /* * Returns true if all of the specified predicates are true for the entire * (super)page and false otherwise. */ bool vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) { vm_object_t object; int i, npages; object = m->object; if (skip_m != NULL && skip_m->object != object) return (false); VM_OBJECT_ASSERT_LOCKED(object); npages = atop(pagesizes[m->psind]); /* * The physically contiguous pages that make up a superpage, i.e., a * page with a page size index ("psind") greater than zero, will * occupy adjacent entries in vm_page_array[]. */ for (i = 0; i < npages; i++) { /* Always test object consistency, including "skip_m". */ if (m[i].object != object) return (false); if (&m[i] == skip_m) continue; if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) return (false); if ((flags & PS_ALL_DIRTY) != 0) { /* * Calling vm_page_test_dirty() or pmap_is_modified() * might stop this case from spuriously returning * "false". However, that would require a write lock * on the object containing "m[i]". */ if (m[i].dirty != VM_PAGE_BITS_ALL) return (false); } if ((flags & PS_ALL_VALID) != 0 && m[i].valid != VM_PAGE_BITS_ALL) return (false); } return (true); } /* * Set the page's dirty bits if the page is modified. */ void vm_page_test_dirty(vm_page_t m) { VM_OBJECT_ASSERT_WLOCKED(m->object); if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) vm_page_dirty(m); } void vm_page_lock_KBI(vm_page_t m, const char *file, int line) { mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); } void vm_page_unlock_KBI(vm_page_t m, const char *file, int line) { mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); } int vm_page_trylock_KBI(vm_page_t m, const char *file, int line) { return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); } #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) { vm_page_lock_assert_KBI(m, MA_OWNED, file, line); } void vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) { mtx_assert_(vm_page_lockptr(m), a, file, line); } #endif #ifdef INVARIANTS void vm_page_object_lock_assert(vm_page_t m) { /* * Certain of the page's fields may only be modified by the * holder of the containing object's lock or the exclusive busy. * holder. Unfortunately, the holder of the write busy is * not recorded, and thus cannot be checked here. */ if (m->object != NULL && !vm_page_xbusied(m)) VM_OBJECT_ASSERT_WLOCKED(m->object); } void vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits) { if ((bits & PGA_WRITEABLE) == 0) return; /* * The PGA_WRITEABLE flag can only be set if the page is * managed, is exclusively busied or the object is locked. * Currently, this flag is only set by pmap_enter(). */ KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("PGA_WRITEABLE on unmanaged page")); if (!vm_page_xbusied(m)) VM_OBJECT_ASSERT_LOCKED(m->object); } #endif #include "opt_ddb.h" #ifdef DDB #include #include DB_SHOW_COMMAND(page, vm_page_print_page_info) { db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); } DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) { int dom; db_printf("pq_free %d\n", vm_free_count()); for (dom = 0; dom < vm_ndomains; dom++) { db_printf( "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", dom, vm_dom[dom].vmd_page_count, vm_dom[dom].vmd_free_count, vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); } } DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) { vm_page_t m; boolean_t phys; if (!have_addr) { db_printf("show pginfo addr\n"); return; } phys = strchr(modif, 'p') != NULL; if (phys) m = PHYS_TO_VM_PAGE(addr); else m = (vm_page_t)addr; db_printf( "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n" " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags, m->flags, m->act_count, m->busy_lock, m->valid, m->dirty); } #endif /* DDB */ Index: head/sys/vm/vm_pageout.c =================================================================== --- head/sys/vm/vm_pageout.c (revision 339924) +++ head/sys/vm/vm_pageout.c (revision 339925) @@ -1,2131 +1,2131 @@ /*- * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU) * * 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) 2005 Yahoo! Technologies Norway AS * 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_pageout.c 7.4 (Berkeley) 5/7/91 * * * 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. */ /* * The proverbial page-out daemon. */ #include __FBSDID("$FreeBSD$"); #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 #include #include #include #include #include #include #include #include #include #include /* * System initialization */ /* the kernel process "vm_pageout"*/ static void vm_pageout(void); static void vm_pageout_init(void); static int vm_pageout_clean(vm_page_t m, int *numpagedout); static int vm_pageout_cluster(vm_page_t m); static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, int starting_page_shortage); SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init, NULL); struct proc *pageproc; static struct kproc_desc page_kp = { "pagedaemon", vm_pageout, &pageproc }; SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &page_kp); SDT_PROVIDER_DEFINE(vm); SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan); /* Pagedaemon activity rates, in subdivisions of one second. */ #define VM_LAUNDER_RATE 10 #define VM_INACT_SCAN_RATE 10 static int vm_pageout_oom_seq = 12; static int vm_pageout_update_period; static int disable_swap_pageouts; static int lowmem_period = 10; static int swapdev_enabled; static int vm_panic_on_oom = 0; SYSCTL_INT(_vm, OID_AUTO, panic_on_oom, CTLFLAG_RWTUN, &vm_panic_on_oom, 0, "panic on out of memory instead of killing the largest process"); SYSCTL_INT(_vm, OID_AUTO, pageout_update_period, CTLFLAG_RWTUN, &vm_pageout_update_period, 0, "Maximum active LRU update period"); SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0, "Low memory callback period"); SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); static int pageout_lock_miss; SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq, CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0, "back-to-back calls to oom detector to start OOM"); static int act_scan_laundry_weight = 3; SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN, &act_scan_laundry_weight, 0, "weight given to clean vs. dirty pages in active queue scans"); static u_int vm_background_launder_rate = 4096; SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN, &vm_background_launder_rate, 0, "background laundering rate, in kilobytes per second"); static u_int vm_background_launder_max = 20 * 1024; SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN, &vm_background_launder_max, 0, "background laundering cap, in kilobytes"); int vm_pageout_page_count = 32; int vm_page_max_wired; /* XXX max # of wired pages system-wide */ SYSCTL_INT(_vm, OID_AUTO, max_wired, CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count"); static u_int isqrt(u_int num); static int vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall); static void vm_pageout_laundry_worker(void *arg); struct scan_state { struct vm_batchqueue bq; struct vm_pagequeue *pq; vm_page_t marker; int maxscan; int scanned; }; static void vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq, vm_page_t marker, vm_page_t after, int maxscan) { vm_pagequeue_assert_locked(pq); KASSERT((marker->aflags & PGA_ENQUEUED) == 0, ("marker %p already enqueued", marker)); if (after == NULL) TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q); else TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q); vm_page_aflag_set(marker, PGA_ENQUEUED); vm_batchqueue_init(&ss->bq); ss->pq = pq; ss->marker = marker; ss->maxscan = maxscan; ss->scanned = 0; vm_pagequeue_unlock(pq); } static void vm_pageout_end_scan(struct scan_state *ss) { struct vm_pagequeue *pq; pq = ss->pq; vm_pagequeue_assert_locked(pq); KASSERT((ss->marker->aflags & PGA_ENQUEUED) != 0, ("marker %p not enqueued", ss->marker)); TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q); vm_page_aflag_clear(ss->marker, PGA_ENQUEUED); pq->pq_pdpages += ss->scanned; } /* * Add a small number of queued pages to a batch queue for later processing * without the corresponding queue lock held. The caller must have enqueued a * marker page at the desired start point for the scan. Pages will be * physically dequeued if the caller so requests. Otherwise, the returned * batch may contain marker pages, and it is up to the caller to handle them. * * When processing the batch queue, vm_page_queue() must be used to * determine whether the page has been logically dequeued by another thread. * Once this check is performed, the page lock guarantees that the page will * not be disassociated from the queue. */ static __always_inline void vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue) { struct vm_pagequeue *pq; vm_page_t m, marker; marker = ss->marker; pq = ss->pq; KASSERT((marker->aflags & PGA_ENQUEUED) != 0, ("marker %p not enqueued", ss->marker)); vm_pagequeue_lock(pq); for (m = TAILQ_NEXT(marker, plinks.q); m != NULL && ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE; m = TAILQ_NEXT(m, plinks.q), ss->scanned++) { if ((m->flags & PG_MARKER) == 0) { KASSERT((m->aflags & PGA_ENQUEUED) != 0, ("page %p not enqueued", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0, ("Fictitious page %p cannot be in page queue", m)); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("Unmanaged page %p cannot be in page queue", m)); } else if (dequeue) continue; (void)vm_batchqueue_insert(&ss->bq, m); if (dequeue) { TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); vm_page_aflag_clear(m, PGA_ENQUEUED); } } TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q); if (__predict_true(m != NULL)) TAILQ_INSERT_BEFORE(m, marker, plinks.q); else TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q); if (dequeue) vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt); vm_pagequeue_unlock(pq); } /* Return the next page to be scanned, or NULL if the scan is complete. */ static __always_inline vm_page_t vm_pageout_next(struct scan_state *ss, const bool dequeue) { if (ss->bq.bq_cnt == 0) vm_pageout_collect_batch(ss, dequeue); return (vm_batchqueue_pop(&ss->bq)); } /* * Scan for pages at adjacent offsets within the given page's object that are * eligible for laundering, form a cluster of these pages and the given page, * and launder that cluster. */ static int vm_pageout_cluster(vm_page_t m) { vm_object_t object; vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps; vm_pindex_t pindex; int ib, is, page_base, pageout_count; vm_page_assert_locked(m); object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); pindex = m->pindex; vm_page_assert_unbusied(m); KASSERT(!vm_page_held(m), ("page %p is held", m)); pmap_remove_write(m); vm_page_unlock(m); mc[vm_pageout_page_count] = pb = ps = m; pageout_count = 1; page_base = vm_pageout_page_count; ib = 1; is = 1; /* * We can cluster only if the page is not clean, busy, or held, and * the page is in the laundry queue. * * During heavy mmap/modification loads the pageout * daemon can really fragment the underlying file * due to flushing pages out of order and not trying to * align the clusters (which leaves sporadic out-of-order * holes). To solve this problem we do the reverse scan * first and attempt to align our cluster, then do a * forward scan if room remains. */ more: while (ib != 0 && pageout_count < vm_pageout_page_count) { if (ib > pindex) { ib = 0; break; } if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) { ib = 0; break; } vm_page_test_dirty(p); if (p->dirty == 0) { ib = 0; break; } vm_page_lock(p); if (vm_page_held(p) || !vm_page_in_laundry(p)) { vm_page_unlock(p); ib = 0; break; } pmap_remove_write(p); vm_page_unlock(p); mc[--page_base] = pb = p; ++pageout_count; ++ib; /* * We are at an alignment boundary. Stop here, and switch * directions. Do not clear ib. */ if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) break; } while (pageout_count < vm_pageout_page_count && pindex + is < object->size) { if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p)) break; vm_page_test_dirty(p); if (p->dirty == 0) break; vm_page_lock(p); if (vm_page_held(p) || !vm_page_in_laundry(p)) { vm_page_unlock(p); break; } pmap_remove_write(p); vm_page_unlock(p); mc[page_base + pageout_count] = ps = p; ++pageout_count; ++is; } /* * If we exhausted our forward scan, continue with the reverse scan * when possible, even past an alignment boundary. This catches * boundary conditions. */ if (ib != 0 && pageout_count < vm_pageout_page_count) goto more; return (vm_pageout_flush(&mc[page_base], pageout_count, VM_PAGER_PUT_NOREUSE, 0, NULL, NULL)); } /* * vm_pageout_flush() - launder the given pages * * The given pages are laundered. Note that we setup for the start of * I/O ( i.e. busy the page ), mark it read-only, and bump the object * reference count all in here rather then in the parent. If we want * the parent to do more sophisticated things we may have to change * the ordering. * * Returned runlen is the count of pages between mreq and first * page after mreq with status VM_PAGER_AGAIN. * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL * for any page in runlen set. */ int vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, boolean_t *eio) { vm_object_t object = mc[0]->object; int pageout_status[count]; int numpagedout = 0; int i, runlen; VM_OBJECT_ASSERT_WLOCKED(object); /* * Initiate I/O. Mark the pages busy and verify that they're valid * and read-only. * * We do not have to fixup the clean/dirty bits here... we can * allow the pager to do it after the I/O completes. * * NOTE! mc[i]->dirty may be partial or fragmented due to an * edge case with file fragments. */ for (i = 0; i < count; i++) { KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush: partially invalid page %p index %d/%d", mc[i], i, count)); KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0, ("vm_pageout_flush: writeable page %p", mc[i])); vm_page_sbusy(mc[i]); } vm_object_pip_add(object, count); vm_pager_put_pages(object, mc, count, flags, pageout_status); runlen = count - mreq; if (eio != NULL) *eio = FALSE; for (i = 0; i < count; i++) { vm_page_t mt = mc[i]; KASSERT(pageout_status[i] == VM_PAGER_PEND || !pmap_page_is_write_mapped(mt), ("vm_pageout_flush: page %p is not write protected", mt)); switch (pageout_status[i]) { case VM_PAGER_OK: vm_page_lock(mt); if (vm_page_in_laundry(mt)) vm_page_deactivate_noreuse(mt); vm_page_unlock(mt); /* FALLTHROUGH */ case VM_PAGER_PEND: numpagedout++; break; case VM_PAGER_BAD: /* * The page is outside the object's range. We pretend * that the page out worked and clean the page, so the * changes will be lost if the page is reclaimed by * the page daemon. */ vm_page_undirty(mt); vm_page_lock(mt); if (vm_page_in_laundry(mt)) vm_page_deactivate_noreuse(mt); vm_page_unlock(mt); break; case VM_PAGER_ERROR: case VM_PAGER_FAIL: /* * If the page couldn't be paged out to swap because the * pager wasn't able to find space, place the page in * the PQ_UNSWAPPABLE holding queue. This is an * optimization that prevents the page daemon from * wasting CPU cycles on pages that cannot be reclaimed * becase no swap device is configured. * * Otherwise, reactivate the page so that it doesn't * clog the laundry and inactive queues. (We will try * paging it out again later.) */ vm_page_lock(mt); if (object->type == OBJT_SWAP && pageout_status[i] == VM_PAGER_FAIL) { vm_page_unswappable(mt); numpagedout++; } else vm_page_activate(mt); vm_page_unlock(mt); if (eio != NULL && i >= mreq && i - mreq < runlen) *eio = TRUE; break; case VM_PAGER_AGAIN: if (i >= mreq && i - mreq < runlen) runlen = i - mreq; break; } /* * If the operation is still going, leave the page busy to * block all other accesses. Also, leave the paging in * progress indicator set so that we don't attempt an object * collapse. */ if (pageout_status[i] != VM_PAGER_PEND) { vm_object_pip_wakeup(object); vm_page_sunbusy(mt); } } if (prunlen != NULL) *prunlen = runlen; return (numpagedout); } static void vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused) { atomic_store_rel_int(&swapdev_enabled, 1); } static void vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused) { if (swap_pager_nswapdev() == 1) atomic_store_rel_int(&swapdev_enabled, 0); } /* * Attempt to acquire all of the necessary locks to launder a page and * then call through the clustering layer to PUTPAGES. Wait a short * time for a vnode lock. * * Requires the page and object lock on entry, releases both before return. * Returns 0 on success and an errno otherwise. */ static int vm_pageout_clean(vm_page_t m, int *numpagedout) { struct vnode *vp; struct mount *mp; vm_object_t object; vm_pindex_t pindex; int error, lockmode; vm_page_assert_locked(m); object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); error = 0; vp = NULL; mp = NULL; /* * The object is already known NOT to be dead. It * is possible for the vget() to block the whole * pageout daemon, but the new low-memory handling * code should prevent it. * * We can't wait forever for the vnode lock, we might * deadlock due to a vn_read() getting stuck in * vm_wait while holding this vnode. We skip the * vnode if we can't get it in a reasonable amount * of time. */ if (object->type == OBJT_VNODE) { vm_page_unlock(m); vp = object->handle; if (vp->v_type == VREG && vn_start_write(vp, &mp, V_NOWAIT) != 0) { mp = NULL; error = EDEADLK; goto unlock_all; } KASSERT(mp != NULL, ("vp %p with NULL v_mount", vp)); vm_object_reference_locked(object); pindex = m->pindex; VM_OBJECT_WUNLOCK(object); lockmode = MNT_SHARED_WRITES(vp->v_mount) ? LK_SHARED : LK_EXCLUSIVE; if (vget(vp, lockmode | LK_TIMELOCK, curthread)) { vp = NULL; error = EDEADLK; goto unlock_mp; } VM_OBJECT_WLOCK(object); /* * Ensure that the object and vnode were not disassociated * while locks were dropped. */ if (vp->v_object != object) { error = ENOENT; goto unlock_all; } vm_page_lock(m); /* * While the object and page were unlocked, the page * may have been: * (1) moved to a different queue, * (2) reallocated to a different object, * (3) reallocated to a different offset, or * (4) cleaned. */ if (!vm_page_in_laundry(m) || m->object != object || m->pindex != pindex || m->dirty == 0) { vm_page_unlock(m); error = ENXIO; goto unlock_all; } /* * The page may have been busied or referenced while the object * and page locks were released. */ if (vm_page_busied(m) || vm_page_held(m)) { vm_page_unlock(m); error = EBUSY; goto unlock_all; } } /* * If a page is dirty, then it is either being washed * (but not yet cleaned) or it is still in the * laundry. If it is still in the laundry, then we * start the cleaning operation. */ if ((*numpagedout = vm_pageout_cluster(m)) == 0) error = EIO; unlock_all: VM_OBJECT_WUNLOCK(object); unlock_mp: vm_page_lock_assert(m, MA_NOTOWNED); if (mp != NULL) { if (vp != NULL) vput(vp); vm_object_deallocate(object); vn_finished_write(mp); } return (error); } /* * Attempt to launder the specified number of pages. * * Returns the number of pages successfully laundered. */ static int vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall) { struct scan_state ss; struct vm_pagequeue *pq; struct mtx *mtx; vm_object_t object; vm_page_t m, marker; int act_delta, error, numpagedout, queue, starting_target; int vnodes_skipped; bool obj_locked, pageout_ok; mtx = NULL; obj_locked = false; object = NULL; starting_target = launder; vnodes_skipped = 0; /* * Scan the laundry queues for pages eligible to be laundered. We stop * once the target number of dirty pages have been laundered, or once * we've reached the end of the queue. A single iteration of this loop * may cause more than one page to be laundered because of clustering. * * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no * swap devices are configured. */ if (atomic_load_acq_int(&swapdev_enabled)) queue = PQ_UNSWAPPABLE; else queue = PQ_LAUNDRY; scan: marker = &vmd->vmd_markers[queue]; pq = &vmd->vmd_pagequeues[queue]; vm_pagequeue_lock(pq); vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt); while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) { if (__predict_false((m->flags & PG_MARKER) != 0)) continue; vm_page_change_lock(m, &mtx); recheck: /* * The page may have been disassociated from the queue * while locks were dropped. */ if (vm_page_queue(m) != queue) continue; /* * A requeue was requested, so this page gets a second * chance. */ if ((m->aflags & PGA_REQUEUE) != 0) { vm_page_requeue(m); continue; } /* * Held pages are essentially stuck in the queue. * * Wired pages may not be freed. Complete their removal * from the queue now to avoid needless revisits during * future scans. */ if (m->hold_count != 0) continue; if (m->wire_count != 0) { vm_page_dequeue_deferred(m); continue; } if (object != m->object) { if (obj_locked) { VM_OBJECT_WUNLOCK(object); obj_locked = false; } object = m->object; } if (!obj_locked) { if (!VM_OBJECT_TRYWLOCK(object)) { mtx_unlock(mtx); /* Depends on type-stability. */ VM_OBJECT_WLOCK(object); obj_locked = true; mtx_lock(mtx); goto recheck; } else obj_locked = true; } if (vm_page_busied(m)) continue; /* * Invalid pages can be easily freed. They cannot be * mapped; vm_page_free() asserts this. */ if (m->valid == 0) goto free_page; /* * If the page has been referenced and the object is not dead, * reactivate or requeue the page depending on whether the * object is mapped. * * Test PGA_REFERENCED after calling pmap_ts_referenced() so * that a reference from a concurrently destroyed mapping is * observed here and now. */ if (object->ref_count != 0) act_delta = pmap_ts_referenced(m); else { KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); act_delta = 0; } if ((m->aflags & PGA_REFERENCED) != 0) { vm_page_aflag_clear(m, PGA_REFERENCED); act_delta++; } if (act_delta != 0) { if (object->ref_count != 0) { VM_CNT_INC(v_reactivated); vm_page_activate(m); /* * Increase the activation count if the page * was referenced while in the laundry queue. * This makes it less likely that the page will * be returned prematurely to the inactive * queue. */ m->act_count += act_delta + ACT_ADVANCE; /* * If this was a background laundering, count * activated pages towards our target. The * purpose of background laundering is to ensure * that pages are eventually cycled through the * laundry queue, and an activation is a valid * way out. */ if (!in_shortfall) launder--; continue; } else if ((object->flags & OBJ_DEAD) == 0) { vm_page_requeue(m); continue; } } /* * If the page appears to be clean at the machine-independent * layer, then remove all of its mappings from the pmap in * anticipation of freeing it. If, however, any of the page's * mappings allow write access, then the page may still be * modified until the last of those mappings are removed. */ if (object->ref_count != 0) { vm_page_test_dirty(m); if (m->dirty == 0) pmap_remove_all(m); } /* * Clean pages are freed, and dirty pages are paged out unless * they belong to a dead object. Requeueing dirty pages from * dead objects is pointless, as they are being paged out and * freed by the thread that destroyed the object. */ if (m->dirty == 0) { free_page: vm_page_free(m); VM_CNT_INC(v_dfree); } else if ((object->flags & OBJ_DEAD) == 0) { if (object->type != OBJT_SWAP && object->type != OBJT_DEFAULT) pageout_ok = true; else if (disable_swap_pageouts) pageout_ok = false; else pageout_ok = true; if (!pageout_ok) { vm_page_requeue(m); continue; } /* * Form a cluster with adjacent, dirty pages from the * same object, and page out that entire cluster. * * The adjacent, dirty pages must also be in the * laundry. However, their mappings are not checked * for new references. Consequently, a recently * referenced page may be paged out. However, that * page will not be prematurely reclaimed. After page * out, the page will be placed in the inactive queue, * where any new references will be detected and the * page reactivated. */ error = vm_pageout_clean(m, &numpagedout); if (error == 0) { launder -= numpagedout; ss.scanned += numpagedout; } else if (error == EDEADLK) { pageout_lock_miss++; vnodes_skipped++; } mtx = NULL; obj_locked = false; } } if (mtx != NULL) { mtx_unlock(mtx); mtx = NULL; } if (obj_locked) { VM_OBJECT_WUNLOCK(object); obj_locked = false; } vm_pagequeue_lock(pq); vm_pageout_end_scan(&ss); vm_pagequeue_unlock(pq); if (launder > 0 && queue == PQ_UNSWAPPABLE) { queue = PQ_LAUNDRY; goto scan; } /* * Wakeup the sync daemon if we skipped a vnode in a writeable object * and we didn't launder enough pages. */ if (vnodes_skipped > 0 && launder > 0) (void)speedup_syncer(); return (starting_target - launder); } /* * Compute the integer square root. */ static u_int isqrt(u_int num) { u_int bit, root, tmp; bit = 1u << ((NBBY * sizeof(u_int)) - 2); while (bit > num) bit >>= 2; root = 0; while (bit != 0) { tmp = root + bit; root >>= 1; if (num >= tmp) { num -= tmp; root += bit; } bit >>= 2; } return (root); } /* * Perform the work of the laundry thread: periodically wake up and determine * whether any pages need to be laundered. If so, determine the number of pages * that need to be laundered, and launder them. */ static void vm_pageout_laundry_worker(void *arg) { struct vm_domain *vmd; struct vm_pagequeue *pq; uint64_t nclean, ndirty, nfreed; int domain, last_target, launder, shortfall, shortfall_cycle, target; bool in_shortfall; domain = (uintptr_t)arg; vmd = VM_DOMAIN(domain); pq = &vmd->vmd_pagequeues[PQ_LAUNDRY]; KASSERT(vmd->vmd_segs != 0, ("domain without segments")); shortfall = 0; in_shortfall = false; shortfall_cycle = 0; target = 0; nfreed = 0; /* * Calls to these handlers are serialized by the swap syscall lock. */ (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd, EVENTHANDLER_PRI_ANY); (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd, EVENTHANDLER_PRI_ANY); /* * The pageout laundry worker is never done, so loop forever. */ for (;;) { KASSERT(target >= 0, ("negative target %d", target)); KASSERT(shortfall_cycle >= 0, ("negative cycle %d", shortfall_cycle)); launder = 0; /* * First determine whether we need to launder pages to meet a * shortage of free pages. */ if (shortfall > 0) { in_shortfall = true; shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE; target = shortfall; } else if (!in_shortfall) goto trybackground; else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) { /* * We recently entered shortfall and began laundering * pages. If we have completed that laundering run * (and we are no longer in shortfall) or we have met * our laundry target through other activity, then we * can stop laundering pages. */ in_shortfall = false; target = 0; goto trybackground; } launder = target / shortfall_cycle--; goto dolaundry; /* * There's no immediate need to launder any pages; see if we * meet the conditions to perform background laundering: * * 1. The ratio of dirty to clean inactive pages exceeds the * background laundering threshold, or * 2. we haven't yet reached the target of the current * background laundering run. * * The background laundering threshold is not a constant. * Instead, it is a slowly growing function of the number of * clean pages freed by the page daemon since the last * background laundering. Thus, as the ratio of dirty to * clean inactive pages grows, the amount of memory pressure * required to trigger laundering decreases. We ensure * that the threshold is non-zero after an inactive queue * scan, even if that scan failed to free a single clean page. */ trybackground: nclean = vmd->vmd_free_count + vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt; ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt; if (target == 0 && ndirty * isqrt(howmany(nfreed + 1, vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) { target = vmd->vmd_background_launder_target; } /* * We have a non-zero background laundering target. If we've * laundered up to our maximum without observing a page daemon * request, just stop. This is a safety belt that ensures we * don't launder an excessive amount if memory pressure is low * and the ratio of dirty to clean pages is large. Otherwise, * proceed at the background laundering rate. */ if (target > 0) { if (nfreed > 0) { nfreed = 0; last_target = target; } else if (last_target - target >= vm_background_launder_max * PAGE_SIZE / 1024) { target = 0; } launder = vm_background_launder_rate * PAGE_SIZE / 1024; launder /= VM_LAUNDER_RATE; if (launder > target) launder = target; } dolaundry: if (launder > 0) { /* * Because of I/O clustering, the number of laundered * pages could exceed "target" by the maximum size of * a cluster minus one. */ target -= min(vm_pageout_launder(vmd, launder, in_shortfall), target); pause("laundp", hz / VM_LAUNDER_RATE); } /* * If we're not currently laundering pages and the page daemon * hasn't posted a new request, sleep until the page daemon * kicks us. */ vm_pagequeue_lock(pq); if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE) (void)mtx_sleep(&vmd->vmd_laundry_request, vm_pagequeue_lockptr(pq), PVM, "launds", 0); /* * If the pagedaemon has indicated that it's in shortfall, start * a shortfall laundering unless we're already in the middle of * one. This may preempt a background laundering. */ if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL && (!in_shortfall || shortfall_cycle == 0)) { shortfall = vm_laundry_target(vmd) + vmd->vmd_pageout_deficit; target = 0; } else shortfall = 0; if (target == 0) vmd->vmd_laundry_request = VM_LAUNDRY_IDLE; nfreed += vmd->vmd_clean_pages_freed; vmd->vmd_clean_pages_freed = 0; vm_pagequeue_unlock(pq); } } /* * Compute the number of pages we want to try to move from the * active queue to either the inactive or laundry queue. * * When scanning active pages during a shortage, we make clean pages * count more heavily towards the page shortage than dirty pages. * This is because dirty pages must be laundered before they can be * reused and thus have less utility when attempting to quickly * alleviate a free page shortage. However, this weighting also * causes the scan to deactivate dirty pages more aggressively, * improving the effectiveness of clustering. */ static int vm_pageout_active_target(struct vm_domain *vmd) { int shortage; shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) - (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt + vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight); shortage *= act_scan_laundry_weight; return (shortage); } /* * Scan the active queue. If there is no shortage of inactive pages, scan a * small portion of the queue in order to maintain quasi-LRU. */ static void vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage) { struct scan_state ss; struct mtx *mtx; vm_page_t m, marker; struct vm_pagequeue *pq; long min_scan; int act_delta, max_scan, scan_tick; marker = &vmd->vmd_markers[PQ_ACTIVE]; pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; vm_pagequeue_lock(pq); /* * If we're just idle polling attempt to visit every * active page within 'update_period' seconds. */ scan_tick = ticks; if (vm_pageout_update_period != 0) { min_scan = pq->pq_cnt; min_scan *= scan_tick - vmd->vmd_last_active_scan; min_scan /= hz * vm_pageout_update_period; } else min_scan = 0; if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0)) vmd->vmd_last_active_scan = scan_tick; /* * Scan the active queue for pages that can be deactivated. Update * the per-page activity counter and use it to identify deactivation * candidates. Held pages may be deactivated. * * To avoid requeuing each page that remains in the active queue, we * implement the CLOCK algorithm. To keep the implementation of the * enqueue operation consistent for all page queues, we use two hands, * represented by marker pages. Scans begin at the first hand, which * precedes the second hand in the queue. When the two hands meet, * they are moved back to the head and tail of the queue, respectively, * and scanning resumes. */ max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan; mtx = NULL; act_scan: vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan); while ((m = vm_pageout_next(&ss, false)) != NULL) { if (__predict_false(m == &vmd->vmd_clock[1])) { vm_pagequeue_lock(pq); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q); TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q); max_scan -= ss.scanned; vm_pageout_end_scan(&ss); goto act_scan; } if (__predict_false((m->flags & PG_MARKER) != 0)) continue; vm_page_change_lock(m, &mtx); /* * The page may have been disassociated from the queue * while locks were dropped. */ if (vm_page_queue(m) != PQ_ACTIVE) continue; /* * Wired pages are dequeued lazily. */ if (m->wire_count != 0) { vm_page_dequeue_deferred(m); continue; } /* * Check to see "how much" the page has been used. * * Test PGA_REFERENCED after calling pmap_ts_referenced() so * that a reference from a concurrently destroyed mapping is * observed here and now. * * Perform an unsynchronized object ref count check. While * the page lock ensures that the page is not reallocated to * another object, in particular, one with unmanaged mappings * that cannot support pmap_ts_referenced(), two races are, * nonetheless, possible: * 1) The count was transitioning to zero, but we saw a non- * zero value. pmap_ts_referenced() will return zero * because the page is not mapped. * 2) The count was transitioning to one, but we saw zero. * This race delays the detection of a new reference. At * worst, we will deactivate and reactivate the page. */ if (m->object->ref_count != 0) act_delta = pmap_ts_referenced(m); else act_delta = 0; if ((m->aflags & PGA_REFERENCED) != 0) { vm_page_aflag_clear(m, PGA_REFERENCED); act_delta++; } /* * Advance or decay the act_count based on recent usage. */ if (act_delta != 0) { m->act_count += ACT_ADVANCE + act_delta; if (m->act_count > ACT_MAX) m->act_count = ACT_MAX; } else m->act_count -= min(m->act_count, ACT_DECLINE); if (m->act_count == 0) { /* * When not short for inactive pages, let dirty pages go * through the inactive queue before moving to the * laundry queues. This gives them some extra time to * be reactivated, potentially avoiding an expensive * pageout. However, during a page shortage, the * inactive queue is necessarily small, and so dirty * pages would only spend a trivial amount of time in * the inactive queue. Therefore, we might as well * place them directly in the laundry queue to reduce * queuing overhead. */ if (page_shortage <= 0) vm_page_deactivate(m); else { /* * Calling vm_page_test_dirty() here would * require acquisition of the object's write * lock. However, during a page shortage, * directing dirty pages into the laundry * queue is only an optimization and not a * requirement. Therefore, we simply rely on * the opportunistic updates to the page's * dirty field by the pmap. */ if (m->dirty == 0) { vm_page_deactivate(m); page_shortage -= act_scan_laundry_weight; } else { vm_page_launder(m); page_shortage--; } } } } if (mtx != NULL) { mtx_unlock(mtx); mtx = NULL; } vm_pagequeue_lock(pq); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q); vm_pageout_end_scan(&ss); vm_pagequeue_unlock(pq); } static int vm_pageout_reinsert_inactive_page(struct scan_state *ss, vm_page_t m) { struct vm_domain *vmd; if (m->queue != PQ_INACTIVE || (m->aflags & PGA_ENQUEUED) != 0) return (0); vm_page_aflag_set(m, PGA_ENQUEUED); if ((m->aflags & PGA_REQUEUE_HEAD) != 0) { vmd = vm_pagequeue_domain(m); TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD); } else if ((m->aflags & PGA_REQUEUE) != 0) { TAILQ_INSERT_TAIL(&ss->pq->pq_pl, m, plinks.q); vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD); } else TAILQ_INSERT_BEFORE(ss->marker, m, plinks.q); return (1); } /* * Re-add stuck pages to the inactive queue. We will examine them again * during the next scan. If the queue state of a page has changed since * it was physically removed from the page queue in * vm_pageout_collect_batch(), don't do anything with that page. */ static void vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq, vm_page_t m) { struct vm_pagequeue *pq; int delta; delta = 0; pq = ss->pq; if (m != NULL) { if (vm_batchqueue_insert(bq, m)) return; vm_pagequeue_lock(pq); delta += vm_pageout_reinsert_inactive_page(ss, m); } else vm_pagequeue_lock(pq); while ((m = vm_batchqueue_pop(bq)) != NULL) delta += vm_pageout_reinsert_inactive_page(ss, m); vm_pagequeue_cnt_add(pq, delta); vm_pagequeue_unlock(pq); vm_batchqueue_init(bq); } /* * Attempt to reclaim the requested number of pages from the inactive queue. * Returns true if the shortage was addressed. */ static int vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage, int *addl_shortage) { struct scan_state ss; struct vm_batchqueue rq; struct mtx *mtx; vm_page_t m, marker; struct vm_pagequeue *pq; vm_object_t object; int act_delta, addl_page_shortage, deficit, page_shortage; int starting_page_shortage; bool obj_locked; /* * The addl_page_shortage is an estimate of the number of temporarily * stuck pages in the inactive queue. In other words, the * number of pages from the inactive count that should be * discounted in setting the target for the active queue scan. */ addl_page_shortage = 0; /* * vmd_pageout_deficit counts the number of pages requested in * allocations that failed because of a free page shortage. We assume * that the allocations will be reattempted and thus include the deficit * in our scan target. */ deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit); starting_page_shortage = page_shortage = shortage + deficit; mtx = NULL; obj_locked = false; object = NULL; vm_batchqueue_init(&rq); /* * Start scanning the inactive queue for pages that we can free. The * scan will stop when we reach the target or we have scanned the * entire queue. (Note that m->act_count is not used to make * decisions for the inactive queue, only for the active queue.) */ marker = &vmd->vmd_markers[PQ_INACTIVE]; pq = &vmd->vmd_pagequeues[PQ_INACTIVE]; vm_pagequeue_lock(pq); vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt); while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) { KASSERT((m->flags & PG_MARKER) == 0, ("marker page %p was dequeued", m)); vm_page_change_lock(m, &mtx); recheck: /* * The page may have been disassociated from the queue * while locks were dropped. */ if (vm_page_queue(m) != PQ_INACTIVE) { addl_page_shortage++; continue; } /* * The page was re-enqueued after the page queue lock was * dropped, or a requeue was requested. This page gets a second * chance. */ if ((m->aflags & (PGA_ENQUEUED | PGA_REQUEUE | PGA_REQUEUE_HEAD)) != 0) goto reinsert; /* * Held pages are essentially stuck in the queue. So, * they ought to be discounted from the inactive count. * See the description of addl_page_shortage above. * * Wired pages may not be freed. Complete their removal * from the queue now to avoid needless revisits during * future scans. */ if (m->hold_count != 0) { addl_page_shortage++; goto reinsert; } if (m->wire_count != 0) { vm_page_dequeue_deferred(m); continue; } if (object != m->object) { if (obj_locked) { VM_OBJECT_WUNLOCK(object); obj_locked = false; } object = m->object; } if (!obj_locked) { if (!VM_OBJECT_TRYWLOCK(object)) { mtx_unlock(mtx); /* Depends on type-stability. */ VM_OBJECT_WLOCK(object); obj_locked = true; mtx_lock(mtx); goto recheck; } else obj_locked = true; } if (vm_page_busied(m)) { /* * Don't mess with busy pages. Leave them at * the front of the queue. Most likely, they * are being paged out and will leave the * queue shortly after the scan finishes. So, * they ought to be discounted from the * inactive count. */ addl_page_shortage++; goto reinsert; } /* * Invalid pages can be easily freed. They cannot be * mapped, vm_page_free() asserts this. */ if (m->valid == 0) goto free_page; /* * If the page has been referenced and the object is not dead, * reactivate or requeue the page depending on whether the * object is mapped. * * Test PGA_REFERENCED after calling pmap_ts_referenced() so * that a reference from a concurrently destroyed mapping is * observed here and now. */ if (object->ref_count != 0) act_delta = pmap_ts_referenced(m); else { KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m)); act_delta = 0; } if ((m->aflags & PGA_REFERENCED) != 0) { vm_page_aflag_clear(m, PGA_REFERENCED); act_delta++; } if (act_delta != 0) { if (object->ref_count != 0) { VM_CNT_INC(v_reactivated); vm_page_activate(m); /* * Increase the activation count if the page * was referenced while in the inactive queue. * This makes it less likely that the page will * be returned prematurely to the inactive * queue. */ m->act_count += act_delta + ACT_ADVANCE; continue; } else if ((object->flags & OBJ_DEAD) == 0) { vm_page_aflag_set(m, PGA_REQUEUE); goto reinsert; } } /* * If the page appears to be clean at the machine-independent * layer, then remove all of its mappings from the pmap in * anticipation of freeing it. If, however, any of the page's * mappings allow write access, then the page may still be * modified until the last of those mappings are removed. */ if (object->ref_count != 0) { vm_page_test_dirty(m); if (m->dirty == 0) pmap_remove_all(m); } /* * Clean pages can be freed, but dirty pages must be sent back * to the laundry, unless they belong to a dead object. * Requeueing dirty pages from dead objects is pointless, as * they are being paged out and freed by the thread that * destroyed the object. */ if (m->dirty == 0) { free_page: /* * Because we dequeued the page and have already * checked for concurrent dequeue and enqueue * requests, we can safely disassociate the page * from the inactive queue. */ KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0, ("page %p has queue state", m)); m->queue = PQ_NONE; vm_page_free(m); page_shortage--; } else if ((object->flags & OBJ_DEAD) == 0) vm_page_launder(m); continue; reinsert: vm_pageout_reinsert_inactive(&ss, &rq, m); } if (mtx != NULL) { mtx_unlock(mtx); mtx = NULL; } if (obj_locked) { VM_OBJECT_WUNLOCK(object); obj_locked = false; } vm_pageout_reinsert_inactive(&ss, &rq, NULL); vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL); vm_pagequeue_lock(pq); vm_pageout_end_scan(&ss); vm_pagequeue_unlock(pq); VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage); /* * Wake up the laundry thread so that it can perform any needed * laundering. If we didn't meet our target, we're in shortfall and * need to launder more aggressively. If PQ_LAUNDRY is empty and no * swap devices are configured, the laundry thread has no work to do, so * don't bother waking it up. * * The laundry thread uses the number of inactive queue scans elapsed * since the last laundering to determine whether to launder again, so * keep count. */ if (starting_page_shortage > 0) { pq = &vmd->vmd_pagequeues[PQ_LAUNDRY]; vm_pagequeue_lock(pq); if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE && (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) { if (page_shortage > 0) { vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL; VM_CNT_INC(v_pdshortfalls); } else if (vmd->vmd_laundry_request != VM_LAUNDRY_SHORTFALL) vmd->vmd_laundry_request = VM_LAUNDRY_BACKGROUND; wakeup(&vmd->vmd_laundry_request); } vmd->vmd_clean_pages_freed += starting_page_shortage - page_shortage; vm_pagequeue_unlock(pq); } /* * Wakeup the swapout daemon if we didn't free the targeted number of * pages. */ if (page_shortage > 0) vm_swapout_run(); /* * If the inactive queue scan fails repeatedly to meet its * target, kill the largest process. */ vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage); /* * Reclaim pages by swapping out idle processes, if configured to do so. */ vm_swapout_run_idle(); /* * See the description of addl_page_shortage above. */ *addl_shortage = addl_page_shortage + deficit; return (page_shortage <= 0); } static int vm_pageout_oom_vote; /* * The pagedaemon threads randlomly select one to perform the * OOM. Trying to kill processes before all pagedaemons * failed to reach free target is premature. */ static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, int starting_page_shortage) { int old_vote; if (starting_page_shortage <= 0 || starting_page_shortage != page_shortage) vmd->vmd_oom_seq = 0; else vmd->vmd_oom_seq++; if (vmd->vmd_oom_seq < vm_pageout_oom_seq) { if (vmd->vmd_oom) { vmd->vmd_oom = FALSE; atomic_subtract_int(&vm_pageout_oom_vote, 1); } return; } /* * Do not follow the call sequence until OOM condition is * cleared. */ vmd->vmd_oom_seq = 0; if (vmd->vmd_oom) return; vmd->vmd_oom = TRUE; old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1); if (old_vote != vm_ndomains - 1) return; /* * The current pagedaemon thread is the last in the quorum to * start OOM. Initiate the selection and signaling of the * victim. */ vm_pageout_oom(VM_OOM_MEM); /* * After one round of OOM terror, recall our vote. On the * next pass, current pagedaemon would vote again if the low * memory condition is still there, due to vmd_oom being * false. */ vmd->vmd_oom = FALSE; atomic_subtract_int(&vm_pageout_oom_vote, 1); } /* * The OOM killer is the page daemon's action of last resort when * memory allocation requests have been stalled for a prolonged period * of time because it cannot reclaim memory. This function computes * the approximate number of physical pages that could be reclaimed if * the specified address space is destroyed. * * Private, anonymous memory owned by the address space is the * principal resource that we expect to recover after an OOM kill. * Since the physical pages mapped by the address space's COW entries * are typically shared pages, they are unlikely to be released and so * they are not counted. * * To get to the point where the page daemon runs the OOM killer, its * efforts to write-back vnode-backed pages may have stalled. This * could be caused by a memory allocation deadlock in the write path * that might be resolved by an OOM kill. Therefore, physical pages * belonging to vnode-backed objects are counted, because they might * be freed without being written out first if the address space holds * the last reference to an unlinked vnode. * * Similarly, physical pages belonging to OBJT_PHYS objects are * counted because the address space might hold the last reference to * the object. */ static long vm_pageout_oom_pagecount(struct vmspace *vmspace) { vm_map_t map; vm_map_entry_t entry; vm_object_t obj; long res; map = &vmspace->vm_map; KASSERT(!map->system_map, ("system map")); sx_assert(&map->lock, SA_LOCKED); res = 0; for (entry = map->header.next; entry != &map->header; entry = entry->next) { if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) continue; obj = entry->object.vm_object; if (obj == NULL) continue; if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 && obj->ref_count != 1) continue; switch (obj->type) { case OBJT_DEFAULT: case OBJT_SWAP: case OBJT_PHYS: case OBJT_VNODE: res += obj->resident_page_count; break; } } return (res); } void vm_pageout_oom(int shortage) { struct proc *p, *bigproc; vm_offset_t size, bigsize; struct thread *td; struct vmspace *vm; bool breakout; /* * We keep the process bigproc locked once we find it to keep anyone * from messing with it; however, there is a possibility of * deadlock if process B is bigproc and one of its child processes * attempts to propagate a signal to B while we are waiting for A's * lock while walking this list. To avoid this, we don't block on * the process lock but just skip a process if it is already locked. */ bigproc = NULL; bigsize = 0; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); /* * If this is a system, protected or killed process, skip it. */ if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 || p->p_pid == 1 || P_KILLED(p) || (p->p_pid < 48 && swap_pager_avail != 0)) { PROC_UNLOCK(p); continue; } /* * If the process is in a non-running type state, * don't touch it. Check all the threads individually. */ breakout = false; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (!TD_ON_RUNQ(td) && !TD_IS_RUNNING(td) && !TD_IS_SLEEPING(td) && !TD_IS_SUSPENDED(td) && !TD_IS_SWAPPED(td)) { thread_unlock(td); breakout = true; break; } thread_unlock(td); } if (breakout) { PROC_UNLOCK(p); continue; } /* * get the process size */ vm = vmspace_acquire_ref(p); if (vm == NULL) { PROC_UNLOCK(p); continue; } _PHOLD_LITE(p); PROC_UNLOCK(p); sx_sunlock(&allproc_lock); if (!vm_map_trylock_read(&vm->vm_map)) { vmspace_free(vm); sx_slock(&allproc_lock); PRELE(p); continue; } size = vmspace_swap_count(vm); if (shortage == VM_OOM_MEM) size += vm_pageout_oom_pagecount(vm); vm_map_unlock_read(&vm->vm_map); vmspace_free(vm); sx_slock(&allproc_lock); /* * If this process is bigger than the biggest one, * remember it. */ if (size > bigsize) { if (bigproc != NULL) PRELE(bigproc); bigproc = p; bigsize = size; } else { PRELE(p); } } sx_sunlock(&allproc_lock); if (bigproc != NULL) { if (vm_panic_on_oom != 0) panic("out of swap space"); PROC_LOCK(bigproc); killproc(bigproc, "out of swap space"); sched_nice(bigproc, PRIO_MIN); _PRELE(bigproc); PROC_UNLOCK(bigproc); } } static bool vm_pageout_lowmem(void) { static int lowmem_ticks = 0; int last; last = atomic_load_int(&lowmem_ticks); while ((u_int)(ticks - last) / hz >= lowmem_period) { if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0) continue; /* * Decrease registered cache sizes. */ SDT_PROBE0(vm, , , vm__lowmem_scan); EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES); /* * We do this explicitly after the caches have been * drained above. */ uma_reclaim(); return (true); } return (false); } static void vm_pageout_worker(void *arg) { struct vm_domain *vmd; u_int ofree; int addl_shortage, domain, shortage; bool target_met; domain = (uintptr_t)arg; vmd = VM_DOMAIN(domain); shortage = 0; target_met = true; /* * XXXKIB It could be useful to bind pageout daemon threads to * the cores belonging to the domain, from which vm_page_array * is allocated. */ KASSERT(vmd->vmd_segs != 0, ("domain without segments")); vmd->vmd_last_active_scan = ticks; /* * The pageout daemon worker is never done, so loop forever. */ while (TRUE) { vm_domain_pageout_lock(vmd); /* * We need to clear wanted before we check the limits. This * prevents races with wakers who will check wanted after they * reach the limit. */ atomic_store_int(&vmd->vmd_pageout_wanted, 0); /* * Might the page daemon need to run again? */ if (vm_paging_needed(vmd, vmd->vmd_free_count)) { /* * Yes. If the scan failed to produce enough free * pages, sleep uninterruptibly for some time in the * hope that the laundry thread will clean some pages. */ vm_domain_pageout_unlock(vmd); if (!target_met) pause("pwait", hz / VM_INACT_SCAN_RATE); } else { /* * No, sleep until the next wakeup or until pages * need to have their reference stats updated. */ if (mtx_sleep(&vmd->vmd_pageout_wanted, vm_domain_pageout_lockptr(vmd), PDROP | PVM, "psleep", hz / VM_INACT_SCAN_RATE) == 0) VM_CNT_INC(v_pdwakeups); } /* Prevent spurious wakeups by ensuring that wanted is set. */ atomic_store_int(&vmd->vmd_pageout_wanted, 1); /* * Use the controller to calculate how many pages to free in * this interval, and scan the inactive queue. If the lowmem * handlers appear to have freed up some pages, subtract the * difference from the inactive queue scan target. */ shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count); if (shortage > 0) { ofree = vmd->vmd_free_count; if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree) shortage -= min(vmd->vmd_free_count - ofree, (u_int)shortage); target_met = vm_pageout_scan_inactive(vmd, shortage, &addl_shortage); } else addl_shortage = 0; /* * Scan the active queue. A positive value for shortage * indicates that we must aggressively deactivate pages to avoid * a shortfall. */ shortage = vm_pageout_active_target(vmd) + addl_shortage; vm_pageout_scan_active(vmd, shortage); } } /* * vm_pageout_init initialises basic pageout daemon settings. */ static void vm_pageout_init_domain(int domain) { struct vm_domain *vmd; struct sysctl_oid *oid; vmd = VM_DOMAIN(domain); vmd->vmd_interrupt_free_min = 2; /* * v_free_reserved needs to include enough for the largest * swap pager structures plus enough for any pv_entry structs * when paging. */ if (vmd->vmd_page_count > 1024) vmd->vmd_free_min = 4 + (vmd->vmd_page_count - 1024) / 200; else vmd->vmd_free_min = 4; vmd->vmd_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + vmd->vmd_interrupt_free_min; vmd->vmd_free_reserved = vm_pageout_page_count + vmd->vmd_pageout_free_min + (vmd->vmd_page_count / 768); vmd->vmd_free_severe = vmd->vmd_free_min / 2; vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved; vmd->vmd_free_min += vmd->vmd_free_reserved; vmd->vmd_free_severe += vmd->vmd_free_reserved; vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2; if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3) vmd->vmd_inactive_target = vmd->vmd_free_count / 3; /* * Set the default wakeup threshold to be 10% below the paging * target. This keeps the steady state out of shortfall. */ vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9; /* * Target amount of memory to move out of the laundry queue during a * background laundering. This is proportional to the amount of system * memory. */ vmd->vmd_background_launder_target = (vmd->vmd_free_target - vmd->vmd_free_min) / 10; /* Initialize the pageout daemon pid controller. */ pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE, vmd->vmd_free_target, PIDCTRL_BOUND, PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD); oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO, "pidctrl", CTLFLAG_RD, NULL, ""); pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid)); } static void vm_pageout_init(void) { u_int freecount; int i; /* * Initialize some paging parameters. */ if (vm_cnt.v_page_count < 2000) vm_pageout_page_count = 8; freecount = 0; for (i = 0; i < vm_ndomains; i++) { struct vm_domain *vmd; vm_pageout_init_domain(i); vmd = VM_DOMAIN(i); vm_cnt.v_free_reserved += vmd->vmd_free_reserved; vm_cnt.v_free_target += vmd->vmd_free_target; vm_cnt.v_free_min += vmd->vmd_free_min; vm_cnt.v_inactive_target += vmd->vmd_inactive_target; vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min; vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min; vm_cnt.v_free_severe += vmd->vmd_free_severe; freecount += vmd->vmd_free_count; } /* * Set interval in seconds for active scan. We want to visit each * page at least once every ten minutes. This is to prevent worst * case paging behaviors with stale active LRU. */ if (vm_pageout_update_period == 0) vm_pageout_update_period = 600; if (vm_page_max_wired == 0) vm_page_max_wired = freecount / 3; } /* * vm_pageout is the high level pageout daemon. */ static void vm_pageout(void) { - int error; - int i; + struct proc *p; + struct thread *td; + int error, first, i; + p = curproc; + td = curthread; + swap_pager_swap_init(); - snprintf(curthread->td_name, sizeof(curthread->td_name), "dom0"); - error = kthread_add(vm_pageout_laundry_worker, NULL, curproc, NULL, - 0, 0, "laundry: dom0"); - if (error != 0) - panic("starting laundry for domain 0, error %d", error); - for (i = 1; i < vm_ndomains; i++) { + for (first = -1, i = 0; i < vm_ndomains; i++) { if (VM_DOMAIN_EMPTY(i)) { if (bootverbose) printf("domain %d empty; skipping pageout\n", i); continue; } - - error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i, - curproc, NULL, 0, 0, "dom%d", i); - if (error != 0) { - panic("starting pageout for domain %d, error %d\n", - i, error); + if (first == -1) + first = i; + else { + error = kthread_add(vm_pageout_worker, + (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i); + if (error != 0) + panic("starting pageout for domain %d: %d\n", + i, error); } error = kthread_add(vm_pageout_laundry_worker, - (void *)(uintptr_t)i, curproc, NULL, 0, 0, - "laundry: dom%d", i); + (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i); if (error != 0) - panic("starting laundry for domain %d, error %d", - i, error); + panic("starting laundry for domain %d: %d", i, error); } - error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL, - 0, 0, "uma"); + error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma"); if (error != 0) panic("starting uma_reclaim helper, error %d\n", error); - vm_pageout_worker((void *)(uintptr_t)0); + + snprintf(td->td_name, sizeof(td->td_name), "dom%d", first); + vm_pageout_worker((void *)(uintptr_t)first); } /* * Perform an advisory wakeup of the page daemon. */ void pagedaemon_wakeup(int domain) { struct vm_domain *vmd; vmd = VM_DOMAIN(domain); vm_domain_pageout_assert_unlocked(vmd); if (curproc == pageproc) return; if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) { vm_domain_pageout_lock(vmd); atomic_store_int(&vmd->vmd_pageout_wanted, 1); wakeup(&vmd->vmd_pageout_wanted); vm_domain_pageout_unlock(vmd); } }