diff --git a/sys/amd64/include/vmparam.h b/sys/amd64/include/vmparam.h
index fc88296f754c..205848489644 100644
--- a/sys/amd64/include/vmparam.h
+++ b/sys/amd64/include/vmparam.h
@@ -1,310 +1,310 @@
/*-
* SPDX-License-Identifier: BSD-4-Clause
*
* Copyright (c) 1990 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 1994 John S. Dyson
* All rights reserved.
* Copyright (c) 2003 Peter Wemm
* All rights reserved.
*
* This code is derived from software contributed to Berkeley by
* William Jolitz.
*
* 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: @(#)vmparam.h 5.9 (Berkeley) 5/12/91
* $FreeBSD$
*/
#ifdef __i386__
#include <i386/vmparam.h>
#else /* !__i386__ */
#ifndef _MACHINE_VMPARAM_H_
#define _MACHINE_VMPARAM_H_ 1
/*
* Machine dependent constants for AMD64.
*/
/*
* Virtual memory related constants, all in bytes
*/
#define MAXTSIZ (32768UL*1024*1024) /* max text size */
#ifndef DFLDSIZ
#define DFLDSIZ (32768UL*1024*1024) /* initial data size limit */
#endif
#ifndef MAXDSIZ
#define MAXDSIZ (32768UL*1024*1024) /* max data size */
#endif
#ifndef DFLSSIZ
#define DFLSSIZ (8UL*1024*1024) /* initial stack size limit */
#endif
#ifndef MAXSSIZ
#define MAXSSIZ (512UL*1024*1024) /* max stack size */
#endif
#ifndef SGROWSIZ
#define SGROWSIZ (128UL*1024) /* amount to grow stack */
#endif
/*
* We provide a machine specific single page allocator through the use
* of the direct mapped segment. This uses 2MB pages for reduced
* TLB pressure.
*/
#if !defined(KASAN) && !defined(KMSAN)
#define UMA_MD_SMALL_ALLOC
#endif
/*
* The physical address space is densely populated.
*/
#define VM_PHYSSEG_DENSE
/*
* The number of PHYSSEG entries must be one greater than the number
* of phys_avail entries because the phys_avail entry that spans the
* largest physical address that is accessible by ISA DMA is split
* into two PHYSSEG entries.
*/
#define VM_PHYSSEG_MAX 63
/*
* Create two free page pools: VM_FREEPOOL_DEFAULT is the default pool
* from which physical pages are allocated and VM_FREEPOOL_DIRECT is
* the pool from which physical pages for page tables and small UMA
* objects are allocated.
*/
#define VM_NFREEPOOL 2
#define VM_FREEPOOL_DEFAULT 0
#define VM_FREEPOOL_DIRECT 1
/*
* Create up to three free page lists: VM_FREELIST_DMA32 is for physical pages
* that have physical addresses below 4G but are not accessible by ISA DMA,
* and VM_FREELIST_ISADMA is for physical pages that are accessible by ISA
* DMA.
*/
#define VM_NFREELIST 3
#define VM_FREELIST_DEFAULT 0
#define VM_FREELIST_DMA32 1
#define VM_FREELIST_LOWMEM 2
#define VM_LOWMEM_BOUNDARY (16 << 20) /* 16MB ISA DMA limit */
/*
* Create the DMA32 free list only if the number of physical pages above
* physical address 4G is at least 16M, which amounts to 64GB of physical
* memory.
*/
#define VM_DMA32_NPAGES_THRESHOLD 16777216
/*
* An allocation size of 16MB is supported in order to optimize the
* use of the direct map by UMA. Specifically, a cache line contains
* at most 8 PDEs, collectively mapping 16MB of physical memory. By
* reducing the number of distinct 16MB "pages" that are used by UMA,
* the physical memory allocator reduces the likelihood of both 2MB
* page TLB misses and cache misses caused by 2MB page TLB misses.
*/
#define VM_NFREEORDER 13
/*
* Enable superpage reservations: 1 level.
*/
#ifndef VM_NRESERVLEVEL
#define VM_NRESERVLEVEL 1
#endif
/*
* Level 0 reservations consist of 512 pages.
*/
#ifndef VM_LEVEL_0_ORDER
#define VM_LEVEL_0_ORDER 9
#endif
#ifdef SMP
#define PA_LOCK_COUNT 256
#endif
/*
* Kernel physical load address for non-UEFI boot and for legacy UEFI loader.
* Newer UEFI loader loads kernel anywhere below 4G, with memory allocated
* by boot services.
* Needs to be aligned at 2MB superpage boundary.
*/
#ifndef KERNLOAD
#define KERNLOAD 0x200000
#endif
/*
* Virtual addresses of things. Derived from the page directory and
* page table indexes from pmap.h for precision.
*
* 0x0000000000000000 - 0x00007fffffffffff user map
* 0x0000800000000000 - 0xffff7fffffffffff does not exist (hole)
* 0xffff800000000000 - 0xffff804020100fff recursive page table (512GB slot)
* 0xffff804020100fff - 0xffff807fffffffff unused
* 0xffff808000000000 - 0xffff847fffffffff large map (can be tuned up)
* 0xffff848000000000 - 0xfffff77fffffffff unused (large map extends there)
* 0xfffff60000000000 - 0xfffff7ffffffffff 2TB KMSAN origin map, optional
* 0xfffff78000000000 - 0xfffff7bfffffffff 512GB KASAN shadow map, optional
* 0xfffff80000000000 - 0xfffffbffffffffff 4TB direct map
* 0xfffffc0000000000 - 0xfffffdffffffffff 2TB KMSAN shadow map, optional
* 0xfffffe0000000000 - 0xffffffffffffffff 2TB kernel map
*
* Within the kernel map:
*
* 0xfffffe0000000000 vm_page_array
* 0xffffffff80000000 KERNBASE
*/
#define VM_MIN_KERNEL_ADDRESS KV4ADDR(KPML4BASE, 0, 0, 0)
#define VM_MAX_KERNEL_ADDRESS KV4ADDR(KPML4BASE + NKPML4E - 1, \
NPDPEPG-1, NPDEPG-1, NPTEPG-1)
#define DMAP_MIN_ADDRESS KV4ADDR(DMPML4I, 0, 0, 0)
#define DMAP_MAX_ADDRESS KV4ADDR(DMPML4I + NDMPML4E, 0, 0, 0)
#define KASAN_MIN_ADDRESS KV4ADDR(KASANPML4I, 0, 0, 0)
#define KASAN_MAX_ADDRESS KV4ADDR(KASANPML4I + NKASANPML4E, 0, 0, 0)
#define KMSAN_SHAD_MIN_ADDRESS KV4ADDR(KMSANSHADPML4I, 0, 0, 0)
#define KMSAN_SHAD_MAX_ADDRESS KV4ADDR(KMSANSHADPML4I + NKMSANSHADPML4E, \
0, 0, 0)
#define KMSAN_ORIG_MIN_ADDRESS KV4ADDR(KMSANORIGPML4I, 0, 0, 0)
#define KMSAN_ORIG_MAX_ADDRESS KV4ADDR(KMSANORIGPML4I + NKMSANORIGPML4E, \
0, 0, 0)
#define LARGEMAP_MIN_ADDRESS KV4ADDR(LMSPML4I, 0, 0, 0)
#define LARGEMAP_MAX_ADDRESS KV4ADDR(LMEPML4I + 1, 0, 0, 0)
/*
* Formally kernel mapping starts at KERNBASE, but kernel linker
* script leaves first PDE reserved. For legacy BIOS boot, kernel is
* loaded at KERNLOAD = 2M, and initial kernel page table maps
* physical memory from zero to KERNend starting at KERNBASE.
*
* KERNSTART is where the first actual kernel page is mapped, after
* the compatibility mapping.
*/
#define KERNBASE KV4ADDR(KPML4I, KPDPI, 0, 0)
#define KERNSTART (KERNBASE + NBPDR)
#define UPT_MAX_ADDRESS KV4ADDR(PML4PML4I, PML4PML4I, PML4PML4I, PML4PML4I)
#define UPT_MIN_ADDRESS KV4ADDR(PML4PML4I, 0, 0, 0)
#define VM_MAXUSER_ADDRESS_LA57 UVADDR(NUPML5E, 0, 0, 0, 0)
#define VM_MAXUSER_ADDRESS_LA48 UVADDR(0, NUP4ML4E, 0, 0, 0)
#define VM_MAXUSER_ADDRESS VM_MAXUSER_ADDRESS_LA57
#define SHAREDPAGE_LA57 (VM_MAXUSER_ADDRESS_LA57 - PAGE_SIZE)
#define SHAREDPAGE_LA48 (VM_MAXUSER_ADDRESS_LA48 - PAGE_SIZE)
#define USRSTACK_LA57 SHAREDPAGE_LA57
#define USRSTACK_LA48 SHAREDPAGE_LA48
#define USRSTACK USRSTACK_LA48
#define PS_STRINGS_LA57 (USRSTACK_LA57 - sizeof(struct ps_strings))
#define PS_STRINGS_LA48 (USRSTACK_LA48 - sizeof(struct ps_strings))
#define VM_MAX_ADDRESS UPT_MAX_ADDRESS
#define VM_MIN_ADDRESS (0)
/*
* XXX Allowing dmaplimit == 0 is a temporary workaround for vt(4) efifb's
* early use of PHYS_TO_DMAP before the mapping is actually setup. This works
* because the result is not actually accessed until later, but the early
* vt fb startup needs to be reworked.
*/
#define PHYS_IN_DMAP(pa) (dmaplimit == 0 || (pa) < dmaplimit)
#define VIRT_IN_DMAP(va) ((va) >= DMAP_MIN_ADDRESS && \
(va) < (DMAP_MIN_ADDRESS + dmaplimit))
#define PMAP_HAS_DMAP 1
#define PHYS_TO_DMAP(x) ({ \
KASSERT(PHYS_IN_DMAP(x), \
("physical address %#jx not covered by the DMAP", \
(uintmax_t)x)); \
(x) | DMAP_MIN_ADDRESS; })
#define DMAP_TO_PHYS(x) ({ \
KASSERT(VIRT_IN_DMAP(x), \
("virtual address %#jx not covered by the DMAP", \
(uintmax_t)x)); \
(x) & ~DMAP_MIN_ADDRESS; })
/*
* amd64 maps the page array into KVA so that it can be more easily
* allocated on the correct memory domains.
*/
#define PMAP_HAS_PAGE_ARRAY 1
/*
* How many physical pages per kmem arena virtual page.
*/
#ifndef VM_KMEM_SIZE_SCALE
#define VM_KMEM_SIZE_SCALE (1)
#endif
/*
* Optional ceiling (in bytes) on the size of the kmem arena: 60% of the
* kernel map.
*/
#ifndef VM_KMEM_SIZE_MAX
#define VM_KMEM_SIZE_MAX ((VM_MAX_KERNEL_ADDRESS - \
VM_MIN_KERNEL_ADDRESS + 1) * 3 / 5)
#endif
/* initial pagein size of beginning of executable file */
#ifndef VM_INITIAL_PAGEIN
#define VM_INITIAL_PAGEIN 16
#endif
#define ZERO_REGION_SIZE (2 * 1024 * 1024) /* 2MB */
/*
* Use a fairly large batch size since we expect amd64 systems to have lots of
* memory.
*/
-#define VM_BATCHQUEUE_SIZE 31
+#define VM_BATCHQUEUE_SIZE 63
/*
* The pmap can create non-transparent large page mappings.
*/
#define PMAP_HAS_LARGEPAGES 1
/*
* Need a page dump array for minidump.
*/
#define MINIDUMP_PAGE_TRACKING 1
#endif /* _MACHINE_VMPARAM_H_ */
#endif /* __i386__ */
diff --git a/sys/powerpc/include/vmparam.h b/sys/powerpc/include/vmparam.h
index 77457717a3fd..1b9873aede4a 100644
--- a/sys/powerpc/include/vmparam.h
+++ b/sys/powerpc/include/vmparam.h
@@ -1,338 +1,338 @@
/*-
* SPDX-License-Identifier: BSD-4-Clause
*
* Copyright (C) 1995, 1996 Wolfgang Solfrank.
* Copyright (C) 1995, 1996 TooLs GmbH.
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. All advertising materials mentioning features or use of this software
* must display the following acknowledgement:
* This product includes software developed by TooLs GmbH.
* 4. The name of TooLs GmbH may not be used to endorse or promote products
* derived from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
* OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
* IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
* WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
* OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
* ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
* $NetBSD: vmparam.h,v 1.11 2000/02/11 19:25:16 thorpej Exp $
* $FreeBSD$
*/
#ifndef _MACHINE_VMPARAM_H_
#define _MACHINE_VMPARAM_H_
#ifndef LOCORE
#include <machine/md_var.h>
#endif
#define USRSTACK SHAREDPAGE
#ifndef MAXTSIZ
#define MAXTSIZ (1*1024*1024*1024) /* max text size */
#endif
#ifndef DFLDSIZ
#define DFLDSIZ (128*1024*1024) /* default data size */
#endif
#ifndef MAXDSIZ
#ifdef __powerpc64__
#define MAXDSIZ (32UL*1024*1024*1024) /* max data size */
#else
#define MAXDSIZ (1*1024*1024*1024) /* max data size */
#endif
#endif
#ifndef DFLSSIZ
#define DFLSSIZ (8*1024*1024) /* default stack size */
#endif
#ifndef MAXSSIZ
#ifdef __powerpc64__
#define MAXSSIZ (512*1024*1024) /* max stack size */
#else
#define MAXSSIZ (64*1024*1024) /* max stack size */
#endif
#endif
#ifdef AIM
#define VM_MAXUSER_ADDRESS32 0xfffff000
#else
#define VM_MAXUSER_ADDRESS32 0x7ffff000
#endif
/*
* Would like to have MAX addresses = 0, but this doesn't (currently) work
*/
#ifdef __powerpc64__
/*
* Virtual addresses of things. Derived from the page directory and
* page table indexes from pmap.h for precision.
*
* kernel map should be able to start at 0xc008000000000000 -
* but at least the functional simulator doesn't like it
*
* 0x0000000000000000 - 0x000fffffffffffff user map
* 0xc000000000000000 - 0xc007ffffffffffff direct map
* 0xc008000000000000 - 0xc00fffffffffffff kernel map
*
*/
#define VM_MIN_ADDRESS 0x0000000000000000
#define VM_MAXUSER_ADDRESS 0x000fffffc0000000
#define VM_MAX_ADDRESS 0xc00fffffffffffff
#define VM_MIN_KERNEL_ADDRESS 0xc008000000000000
#define VM_MAX_KERNEL_ADDRESS 0xc0080007ffffffff
#define VM_MAX_SAFE_KERNEL_ADDRESS VM_MAX_KERNEL_ADDRESS
#else
#define VM_MIN_ADDRESS 0
#define VM_MAXUSER_ADDRESS VM_MAXUSER_ADDRESS32
#define VM_MAX_ADDRESS 0xffffffff
#endif
#define SHAREDPAGE (VM_MAXUSER_ADDRESS - PAGE_SIZE)
#define FREEBSD32_SHAREDPAGE (VM_MAXUSER_ADDRESS32 - PAGE_SIZE)
#define FREEBSD32_USRSTACK FREEBSD32_SHAREDPAGE
#define KERNBASE 0x00100100 /* start of kernel virtual */
#ifdef AIM
#ifndef __powerpc64__
#define VM_MIN_KERNEL_ADDRESS ((vm_offset_t)KERNEL_SR << ADDR_SR_SHFT)
#define VM_MAX_SAFE_KERNEL_ADDRESS (VM_MIN_KERNEL_ADDRESS + 2*SEGMENT_LENGTH -1)
#define VM_MAX_KERNEL_ADDRESS (VM_MIN_KERNEL_ADDRESS + 3*SEGMENT_LENGTH - 1)
#endif
/*
* Use the direct-mapped BAT registers for UMA small allocs. This
* takes pressure off the small amount of available KVA.
*/
#define UMA_MD_SMALL_ALLOC
#else /* Book-E */
/* Use the direct map for UMA small allocs on powerpc64. */
#ifdef __powerpc64__
#define UMA_MD_SMALL_ALLOC
#else
#define VM_MIN_KERNEL_ADDRESS 0xc0000000
#define VM_MAX_KERNEL_ADDRESS 0xffffefff
#define VM_MAX_SAFE_KERNEL_ADDRESS VM_MAX_KERNEL_ADDRESS
#endif
#endif /* AIM/E500 */
#if !defined(LOCORE)
struct pmap_physseg {
struct pv_entry *pvent;
char *attrs;
};
#endif
#ifdef __powerpc64__
#define VM_PHYSSEG_MAX 63 /* 1? */
#else
#define VM_PHYSSEG_MAX 16 /* 1? */
#endif
#define PHYS_AVAIL_SZ 256 /* Allows up to 16GB Ram on pSeries with
* logical memory block size of 64MB.
* For more Ram increase the lmb or this value.
*/
/* XXX This is non-sensical. Phys avail should hold contiguous regions. */
#define PHYS_AVAIL_ENTRIES PHYS_AVAIL_SZ
/*
* The physical address space is densely populated on 32-bit systems,
* but may not be on 64-bit ones.
*/
#ifdef __powerpc64__
#define VM_PHYSSEG_SPARSE
#else
#define VM_PHYSSEG_DENSE
#endif
/*
* Create two free page pools: VM_FREEPOOL_DEFAULT is the default pool
* from which physical pages are allocated and VM_FREEPOOL_DIRECT is
* the pool from which physical pages for small UMA objects are
* allocated.
*/
#define VM_NFREEPOOL 2
#define VM_FREEPOOL_DEFAULT 0
#define VM_FREEPOOL_DIRECT 1
/*
* Create one free page list.
*/
#define VM_NFREELIST 1
#define VM_FREELIST_DEFAULT 0
#ifdef __powerpc64__
/* The largest allocation size is 16MB. */
#define VM_NFREEORDER 13
#else
/* The largest allocation size is 4MB. */
#define VM_NFREEORDER 11
#endif
#ifndef VM_NRESERVLEVEL
#ifdef __powerpc64__
/* Enable superpage reservations: 1 level. */
#define VM_NRESERVLEVEL 1
#else
/* Disable superpage reservations. */
#define VM_NRESERVLEVEL 0
#endif
#endif
#ifndef VM_LEVEL_0_ORDER
/* Level 0 reservations consist of 512 (RPT) or 4096 (HPT) pages. */
#define VM_LEVEL_0_ORDER vm_level_0_order
#ifndef __ASSEMBLER__
extern int vm_level_0_order;
#endif
#endif
#ifndef VM_LEVEL_0_ORDER_MAX
#define VM_LEVEL_0_ORDER_MAX 12
#endif
#ifdef __powerpc64__
#ifdef SMP
#define PA_LOCK_COUNT 256
#endif
#endif
#ifndef VM_INITIAL_PAGEIN
#define VM_INITIAL_PAGEIN 16
#endif
#ifndef SGROWSIZ
#define SGROWSIZ (128UL*1024) /* amount to grow stack */
#endif
/*
* How many physical pages per kmem arena virtual page.
*/
#ifndef VM_KMEM_SIZE_SCALE
#define VM_KMEM_SIZE_SCALE (3)
#endif
/*
* Optional floor (in bytes) on the size of the kmem arena.
*/
#ifndef VM_KMEM_SIZE_MIN
#define VM_KMEM_SIZE_MIN (12 * 1024 * 1024)
#endif
/*
* Optional ceiling (in bytes) on the size of the kmem arena: 40% of the
* usable KVA space.
*/
#ifndef VM_KMEM_SIZE_MAX
#define VM_KMEM_SIZE_MAX ((VM_MAX_SAFE_KERNEL_ADDRESS - \
VM_MIN_KERNEL_ADDRESS + 1) * 2 / 5)
#endif
#ifdef __powerpc64__
#define ZERO_REGION_SIZE (2 * 1024 * 1024) /* 2MB */
#else
#define ZERO_REGION_SIZE (64 * 1024) /* 64KB */
#endif
/*
* Use a fairly large batch size since we expect ppc64 systems to have lots of
* memory.
*/
#ifdef __powerpc64__
-#define VM_BATCHQUEUE_SIZE 31
+#define VM_BATCHQUEUE_SIZE 63
#endif
/*
* On 32-bit OEA, the only purpose for which sf_buf is used is to implement
* an opaque pointer required by the machine-independent parts of the kernel.
* That pointer references the vm_page that is "mapped" by the sf_buf. The
* actual mapping is provided by the direct virtual-to-physical mapping.
*
* On OEA64 and Book-E, we need to do something a little more complicated. Use
* the runtime-detected hw_direct_map to pick between the two cases. Our
* friends in vm_machdep.c will do the same to ensure nothing gets confused.
*/
#define SFBUF
#define SFBUF_NOMD
/*
* We (usually) have a direct map of all physical memory, so provide
* a macro to use to get the kernel VA address for a given PA. Check the
* value of PMAP_HAS_PMAP before using.
*/
#ifndef LOCORE
#ifdef __powerpc64__
#define DMAP_BASE_ADDRESS 0xc000000000000000UL
#define DMAP_MIN_ADDRESS DMAP_BASE_ADDRESS
#define DMAP_MAX_ADDRESS 0xc007ffffffffffffUL
#else
#define DMAP_BASE_ADDRESS 0x00000000UL
#define DMAP_MAX_ADDRESS 0xbfffffffUL
#endif
#endif
#if defined(__powerpc64__) || defined(BOOKE)
/*
* powerpc64 and Book-E will provide their own page array allocators.
*
* On AIM, this will allocate a single virtual array, with pages from the
* correct memory domains.
* On Book-E this will let us put the array in TLB1, removing the need for TLB
* thrashing.
*
* VM_MIN_KERNEL_ADDRESS is just a dummy. It will get set by the MMU driver.
*/
#define PA_MIN_ADDRESS VM_MIN_KERNEL_ADDRESS
#define PMAP_HAS_PAGE_ARRAY 1
#endif
#if defined(__powerpc64__)
/*
* Need a page dump array for minidump.
*/
#define MINIDUMP_PAGE_TRACKING 1
#else
/*
* No minidump with 32-bit powerpc.
*/
#define MINIDUMP_PAGE_TRACKING 0
#endif
#define PMAP_HAS_DMAP (hw_direct_map)
#define PHYS_TO_DMAP(x) ({ \
KASSERT(hw_direct_map, ("Direct map not provided by PMAP")); \
(x) | DMAP_BASE_ADDRESS; })
#define DMAP_TO_PHYS(x) ({ \
KASSERT(hw_direct_map, ("Direct map not provided by PMAP")); \
(x) &~ DMAP_BASE_ADDRESS; })
/*
* No non-transparent large page support in the pmap.
*/
#define PMAP_HAS_LARGEPAGES 0
#endif /* _MACHINE_VMPARAM_H_ */
diff --git a/sys/vm/vm_page.c b/sys/vm/vm_page.c
index 2b7bc6a5b66e..797207205f42 100644
--- a/sys/vm/vm_page.c
+++ b/sys/vm/vm_page.c
@@ -1,5609 +1,5622 @@
/*-
* 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.
*/
/*
* Resident memory management module.
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include "opt_vm.h"
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/counter.h>
#include <sys/domainset.h>
#include <sys/kernel.h>
#include <sys/limits.h>
#include <sys/linker.h>
#include <sys/lock.h>
#include <sys/malloc.h>
#include <sys/mman.h>
#include <sys/msgbuf.h>
#include <sys/mutex.h>
#include <sys/proc.h>
#include <sys/rwlock.h>
#include <sys/sleepqueue.h>
#include <sys/sbuf.h>
#include <sys/sched.h>
#include <sys/smp.h>
#include <sys/sysctl.h>
#include <sys/vmmeter.h>
#include <sys/vnode.h>
#include <vm/vm.h>
#include <vm/pmap.h>
#include <vm/vm_param.h>
#include <vm/vm_domainset.h>
#include <vm/vm_kern.h>
#include <vm/vm_map.h>
#include <vm/vm_object.h>
#include <vm/vm_page.h>
#include <vm/vm_pageout.h>
#include <vm/vm_phys.h>
#include <vm/vm_pagequeue.h>
#include <vm/vm_pager.h>
#include <vm/vm_radix.h>
#include <vm/vm_reserv.h>
#include <vm/vm_extern.h>
#include <vm/vm_dumpset.h>
#include <vm/uma.h>
#include <vm/uma_int.h>
#include <machine/md_var.h>
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;
static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
"VM page statistics");
static COUNTER_U64_DEFINE_EARLY(pqstate_commit_retries);
SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries,
CTLFLAG_RD, &pqstate_commit_retries,
"Number of failed per-page atomic queue state updates");
static COUNTER_U64_DEFINE_EARLY(queue_ops);
SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops,
CTLFLAG_RD, &queue_ops,
"Number of batched queue operations");
static COUNTER_U64_DEFINE_EARLY(queue_nops);
SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops,
CTLFLAG_RD, &queue_nops,
"Number of batched queue operations with no effects");
/*
* 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;
struct bitset *vm_page_dump;
long vm_page_dump_pages;
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 bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m,
vm_pindex_t pindex, const char *wmesg, int allocflags, bool locked);
static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
static void vm_page_enqueue(vm_page_t m, uint8_t queue);
static bool vm_page_free_prep(vm_page_t m);
static void vm_page_free_toq(vm_page_t m);
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 void vm_page_mvqueue(vm_page_t m, const uint8_t queue,
const uint16_t nflag);
static int vm_page_reclaim_run(int req_class, int domain, u_long npages,
vm_page_t m_run, vm_paddr_t high);
static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse);
static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object,
int req);
static int vm_page_zone_import(void *arg, void **store, int cnt, int domain,
int flags);
static void vm_page_zone_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);
bogus_page = vm_page_alloc_noobj(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;
struct vm_pgcache *pgcache;
int cache, domain, maxcache, pool;
maxcache = 0;
TUNABLE_INT_FETCH("vm.pgcache_zone_max_pcpu", &maxcache);
maxcache *= mp_ncpus;
for (domain = 0; domain < vm_ndomains; domain++) {
vmd = VM_DOMAIN(domain);
for (pool = 0; pool < VM_NFREEPOOL; pool++) {
pgcache = &vmd->vmd_pgcache[pool];
pgcache->domain = domain;
pgcache->pool = pool;
pgcache->zone = uma_zcache_create("vm pgcache",
PAGE_SIZE, NULL, NULL, NULL, NULL,
vm_page_zone_import, vm_page_zone_release, pgcache,
UMA_ZONE_VM);
/*
* Limit each pool's zone to 0.1% of the pages in the
* domain.
*/
cache = maxcache != 0 ? maxcache :
vmd->vmd_page_count / 1000;
uma_zone_set_maxcache(pgcache->zone, cache);
}
}
}
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
/*
* 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 != 0) {
vm_domain_freecnt_inc(vmd, -1);
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 the page as a safety precaution.
*/
void
vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags)
{
bzero(marker, sizeof(*marker));
marker->flags = PG_MARKER;
marker->a.flags = aflags;
marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
marker->a.queue = queue;
}
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(const char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
"vm inactive pagequeue";
*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
"vm active pagequeue";
*__DECONST(const char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) =
"vm laundry pagequeue";
*__DECONST(const 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.
*/
void
vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind)
{
m->object = NULL;
m->ref_count = 0;
m->busy_lock = VPB_FREED;
m->flags = m->a.flags = 0;
m->phys_addr = pa;
m->a.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);
}
#ifndef PMAP_HAS_PAGE_ARRAY
static vm_paddr_t
vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range)
{
vm_paddr_t 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));
vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end,
VM_PROT_READ | VM_PROT_WRITE);
vm_page_array_size = page_range;
return (new_end);
}
#endif
/*
* 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;
struct vm_domain *vmd;
vm_page_t m;
char *list, *listend;
vm_paddr_t end, high_avail, low_avail, new_end, size;
vm_paddr_t page_range __unused;
vm_paddr_t last_pa, pa, startp, endp;
u_long pagecount;
#if MINIDUMP_PAGE_TRACKING
u_long vm_page_dump_size;
#endif
int biggestone, i, segind;
#ifdef WITNESS
vm_offset_t mapped;
int witness_size;
#endif
#if defined(__i386__) && defined(VM_PHYSSEG_DENSE)
long ii;
#endif
vaddr = round_page(vaddr);
vm_phys_early_startup();
biggestone = vm_phys_avail_largest();
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);
new_end = end;
#ifdef WITNESS
witness_size = round_page(witness_startup_count());
new_end -= witness_size;
mapped = pmap_map(&vaddr, new_end, new_end + witness_size,
VM_PROT_READ | VM_PROT_WRITE);
bzero((void *)mapped, witness_size);
witness_startup((void *)mapped);
#endif
#if MINIDUMP_PAGE_TRACKING
/*
* 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;
vm_page_dump_pages = 0;
for (i = 0; dump_avail[i + 1] != 0; i += 2) {
vm_page_dump_pages += howmany(dump_avail[i + 1], PAGE_SIZE) -
dump_avail[i] / PAGE_SIZE;
if (dump_avail[i + 1] > last_pa)
last_pa = dump_avail[i + 1];
}
vm_page_dump_size = round_page(BITSET_SIZE(vm_page_dump_pages));
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(__riscv) || defined(__powerpc64__)
/*
* Include the UMA bootstrap pages, witness 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 PMAP_HAS_PAGE_ARRAY
pmap_page_array_startup(size / PAGE_SIZE);
biggestone = vm_phys_avail_largest();
end = new_end = phys_avail[biggestone + 1];
#else
#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;
new_end = vm_page_array_alloc(&vaddr, end, page_range);
#endif
#if VM_NRESERVLEVEL > 0
/*
* Allocate physical memory for the reservation management system's
* data structures, and map it.
*/
new_end = vm_reserv_startup(&vaddr, new_end);
#endif
#if defined(__aarch64__) || defined(__amd64__) || \
defined(__riscv) || defined(__powerpc64__)
/*
* 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)
if (vm_phys_avail_size(i) != 0)
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's pages that are covered by one of
* phys_avail's ranges to the free lists.
*/
for (i = 0; phys_avail[i + 1] != 0; i += 2) {
if (seg->end <= phys_avail[i] ||
seg->start >= phys_avail[i + 1])
continue;
startp = MAX(seg->start, phys_avail[i]);
endp = MIN(seg->end, phys_avail[i + 1]);
pagecount = (u_long)atop(endp - startp);
if (pagecount == 0)
continue;
m = seg->first_page + atop(startp - seg->start);
vmd = VM_DOMAIN(seg->domain);
vm_domain_free_lock(vmd);
vm_phys_enqueue_contig(m, pagecount);
vm_domain_free_unlock(vmd);
vm_domain_freecnt_inc(vmd, pagecount);
vm_cnt.v_page_count += (u_int)pagecount;
vmd->vmd_page_count += (u_int)pagecount;
vmd->vmd_segs |= 1UL << segind;
}
}
/*
* 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
return (vaddr);
}
void
vm_page_reference(vm_page_t m)
{
vm_page_aflag_set(m, PGA_REFERENCED);
}
/*
* vm_page_trybusy
*
* Helper routine for grab functions to trylock busy.
*
* Returns true on success and false on failure.
*/
static bool
vm_page_trybusy(vm_page_t m, int allocflags)
{
if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0)
return (vm_page_trysbusy(m));
else
return (vm_page_tryxbusy(m));
}
/*
* vm_page_tryacquire
*
* Helper routine for grab functions to trylock busy and wire.
*
* Returns true on success and false on failure.
*/
static inline bool
vm_page_tryacquire(vm_page_t m, int allocflags)
{
bool locked;
locked = vm_page_trybusy(m, allocflags);
if (locked && (allocflags & VM_ALLOC_WIRED) != 0)
vm_page_wire(m);
return (locked);
}
/*
* vm_page_busy_acquire:
*
* Acquire the busy lock as described by VM_ALLOC_* flags. Will loop
* and drop the object lock if necessary.
*/
bool
vm_page_busy_acquire(vm_page_t m, int allocflags)
{
vm_object_t obj;
bool locked;
/*
* 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 = atomic_load_ptr(&m->object);
for (;;) {
if (vm_page_tryacquire(m, allocflags))
return (true);
if ((allocflags & VM_ALLOC_NOWAIT) != 0)
return (false);
if (obj != NULL)
locked = VM_OBJECT_WOWNED(obj);
else
locked = false;
MPASS(locked || vm_page_wired(m));
if (_vm_page_busy_sleep(obj, m, m->pindex, "vmpba", allocflags,
locked) && locked)
VM_OBJECT_WLOCK(obj);
if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
return (false);
KASSERT(m->object == obj || m->object == NULL,
("vm_page_busy_acquire: page %p does not belong to %p",
m, obj));
}
}
/*
* 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;
vm_page_assert_xbusied(m);
x = vm_page_busy_fetch(m);
for (;;) {
if (atomic_fcmpset_rel_int(&m->busy_lock,
&x, VPB_SHARERS_WORD(1)))
break;
}
if ((x & VPB_BIT_WAITERS) != 0)
wakeup(m);
}
/*
*
* vm_page_busy_tryupgrade:
*
* Attempt to upgrade a single shared busy into an exclusive busy.
*/
int
vm_page_busy_tryupgrade(vm_page_t m)
{
u_int ce, x;
vm_page_assert_sbusied(m);
x = vm_page_busy_fetch(m);
ce = VPB_CURTHREAD_EXCLUSIVE;
for (;;) {
if (VPB_SHARERS(x) > 1)
return (0);
KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
("vm_page_busy_tryupgrade: invalid lock state"));
if (!atomic_fcmpset_acq_int(&m->busy_lock, &x,
ce | (x & VPB_BIT_WAITERS)))
continue;
return (1);
}
}
/*
* 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 = vm_page_busy_fetch(m);
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_assert_sbusied(m);
x = vm_page_busy_fetch(m);
for (;;) {
KASSERT(x != VPB_FREED,
("vm_page_sunbusy: Unlocking freed page."));
if (VPB_SHARERS(x) > 1) {
if (atomic_fcmpset_int(&m->busy_lock, &x,
x - VPB_ONE_SHARER))
break;
continue;
}
KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1),
("vm_page_sunbusy: invalid lock state"));
if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
continue;
if ((x & VPB_BIT_WAITERS) == 0)
break;
wakeup(m);
break;
}
}
/*
* vm_page_busy_sleep:
*
* Sleep if the page is busy, using the page pointer as wchan.
* This is used to implement the hard-path of the busying mechanism.
*
* If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
* will not sleep if the page is shared-busy.
*
* The object lock must be held on entry.
*
* Returns true if it slept and dropped the object lock, or false
* if there was no sleep and the lock is still held.
*/
bool
vm_page_busy_sleep(vm_page_t m, const char *wmesg, int allocflags)
{
vm_object_t obj;
obj = m->object;
VM_OBJECT_ASSERT_LOCKED(obj);
return (_vm_page_busy_sleep(obj, m, m->pindex, wmesg, allocflags,
true));
}
/*
* vm_page_busy_sleep_unlocked:
*
* Sleep if the page is busy, using the page pointer as wchan.
* This is used to implement the hard-path of busying mechanism.
*
* If VM_ALLOC_IGN_SBUSY is specified in allocflags, the function
* will not sleep if the page is shared-busy.
*
* The object lock must not be held on entry. The operation will
* return if the page changes identity.
*/
void
vm_page_busy_sleep_unlocked(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
const char *wmesg, int allocflags)
{
VM_OBJECT_ASSERT_UNLOCKED(obj);
(void)_vm_page_busy_sleep(obj, m, pindex, wmesg, allocflags, false);
}
/*
* _vm_page_busy_sleep:
*
* Internal busy sleep function. Verifies the page identity and
* lockstate against parameters. Returns true if it sleeps and
* false otherwise.
*
* allocflags uses VM_ALLOC_* flags to specify the lock required.
*
* If locked is true the lock will be dropped for any true returns
* and held for any false returns.
*/
static bool
_vm_page_busy_sleep(vm_object_t obj, vm_page_t m, vm_pindex_t pindex,
const char *wmesg, int allocflags, bool locked)
{
bool xsleep;
u_int x;
/*
* If the object is busy we must wait for that to drain to zero
* before trying the page again.
*/
if (obj != NULL && vm_object_busied(obj)) {
if (locked)
VM_OBJECT_DROP(obj);
vm_object_busy_wait(obj, wmesg);
return (true);
}
if (!vm_page_busied(m))
return (false);
xsleep = (allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0;
sleepq_lock(m);
x = vm_page_busy_fetch(m);
do {
/*
* If the page changes objects or becomes unlocked we can
* simply return.
*/
if (x == VPB_UNBUSIED ||
(xsleep && (x & VPB_BIT_SHARED) != 0) ||
m->object != obj || m->pindex != pindex) {
sleepq_release(m);
return (false);
}
if ((x & VPB_BIT_WAITERS) != 0)
break;
} while (!atomic_fcmpset_int(&m->busy_lock, &x, x | VPB_BIT_WAITERS));
if (locked)
VM_OBJECT_DROP(obj);
DROP_GIANT();
sleepq_add(m, NULL, wmesg, 0, 0);
sleepq_wait(m, PVM);
PICKUP_GIANT();
return (true);
}
/*
* 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)
{
vm_object_t obj;
u_int x;
obj = m->object;
x = vm_page_busy_fetch(m);
for (;;) {
if ((x & VPB_BIT_SHARED) == 0)
return (0);
/*
* Reduce the window for transient busies that will trigger
* false negatives in vm_page_ps_test().
*/
if (obj != NULL && vm_object_busied(obj))
return (0);
if (atomic_fcmpset_acq_int(&m->busy_lock, &x,
x + VPB_ONE_SHARER))
break;
}
/* Refetch the object now that we're guaranteed that it is stable. */
obj = m->object;
if (obj != NULL && vm_object_busied(obj)) {
vm_page_sunbusy(m);
return (0);
}
return (1);
}
/*
* vm_page_tryxbusy:
*
* Try to exclusive busy a page.
* If the operation succeeds 1 is returned otherwise 0.
* The operation never sleeps.
*/
int
vm_page_tryxbusy(vm_page_t m)
{
vm_object_t obj;
if (atomic_cmpset_acq_int(&m->busy_lock, VPB_UNBUSIED,
VPB_CURTHREAD_EXCLUSIVE) == 0)
return (0);
obj = m->object;
if (obj != NULL && vm_object_busied(obj)) {
vm_page_xunbusy(m);
return (0);
}
return (1);
}
static void
vm_page_xunbusy_hard_tail(vm_page_t m)
{
atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
/* Wake the waiter. */
wakeup(m);
}
/*
* vm_page_xunbusy_hard:
*
* Called when unbusy has failed because there is a waiter.
*/
void
vm_page_xunbusy_hard(vm_page_t m)
{
vm_page_assert_xbusied(m);
vm_page_xunbusy_hard_tail(m);
}
void
vm_page_xunbusy_hard_unchecked(vm_page_t m)
{
vm_page_assert_xbusied_unchecked(m);
vm_page_xunbusy_hard_tail(m);
}
static void
vm_page_busy_free(vm_page_t m)
{
u_int x;
atomic_thread_fence_rel();
x = atomic_swap_int(&m->busy_lock, VPB_FREED);
if ((x & VPB_BIT_WAITERS) != 0)
wakeup(m);
}
/*
* 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)
{
for (; count != 0; count--) {
vm_page_unwire(*ma, PQ_ACTIVE);
ma++;
}
}
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->a.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_CURTHREAD_EXCLUSIVE;
/* Fictitious pages are unevictable. */
m->ref_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));
vm_page_assert_xbusied(m);
vm_page_busy_free(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(!vm_page_none_valid(m), ("%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.
*/
if ((vm_page_busy_fetch(m) & VPB_BIT_WAITERS) != 0)
vm_page_activate(m);
else
vm_page_deactivate(m);
vm_page_xunbusy_unchecked(m);
}
/*
* Destroy the identity of an invalid page and free it if possible.
* This is intended to be used when reading a page from backing store fails.
*/
void
vm_page_free_invalid(vm_page_t m)
{
KASSERT(vm_page_none_valid(m), ("page %p is valid", m));
KASSERT(!pmap_page_is_mapped(m), ("page %p is mapped", m));
KASSERT(m->object != NULL, ("page %p has no object", m));
VM_OBJECT_ASSERT_WLOCKED(m->object);
/*
* We may be attempting to free the page as part of the handling for an
* I/O error, in which case the page was xbusied by a different thread.
*/
vm_page_xbusy_claim(m);
/*
* If someone has wired this page while the object lock
* was not held, then the thread that unwires is responsible
* for freeing the page. Otherwise just free the page now.
* The wire count of this unmapped page cannot change while
* we have the page xbusy and the page's object wlocked.
*/
if (vm_page_remove(m))
vm_page_free(m);
}
/*
* 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(vm_page_all_valid(m), ("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;
m->ref_count |= VPRC_OBJREF;
/*
* Now link into the object's ordered list of backed pages.
*/
if (vm_radix_insert(&object->rtree, m)) {
m->object = NULL;
m->pindex = 0;
m->ref_count &= ~VPRC_OBJREF;
return (1);
}
vm_page_insert_radixdone(m, object, mpred);
vm_pager_page_inserted(object, m);
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));
KASSERT((m->ref_count & VPRC_OBJREF) != 0,
("vm_page_insert_radixdone: page %p is missing object ref", m));
if (mpred != NULL) {
KASSERT(mpred->object == object,
("vm_page_insert_radixdone: object doesn't contain mpred"));
KASSERT(mpred->pindex < m->pindex,
("vm_page_insert_radixdone: 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 generation count.
*/
if (pmap_page_is_write_mapped(m))
vm_object_set_writeable_dirty(object);
}
/*
* Do the work to remove a page from its object. The caller is responsible for
* updating the page's fields to reflect this removal.
*/
static void
vm_page_object_remove(vm_page_t m)
{
vm_object_t object;
vm_page_t mrem __diagused;
vm_page_assert_xbusied(m);
object = m->object;
VM_OBJECT_ASSERT_WLOCKED(object);
KASSERT((m->ref_count & VPRC_OBJREF) != 0,
("page %p is missing its object ref", m));
/* Deferred free of swap space. */
if ((m->a.flags & PGA_SWAP_FREE) != 0)
vm_pager_page_unswapped(m);
vm_pager_page_removed(object, m);
m->object = NULL;
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);
}
/*
* vm_page_remove:
*
* Removes the specified page from its containing object, but does not
* invalidate any backing storage. Returns true if the object's reference
* was the last reference to the page, and false otherwise.
*
* The object must be locked and the page must be exclusively busied.
* The exclusive busy will be released on return. If this is not the
* final ref and the caller does not hold a wire reference it may not
* continue to access the page.
*/
bool
vm_page_remove(vm_page_t m)
{
bool dropped;
dropped = vm_page_remove_xbusy(m);
vm_page_xunbusy(m);
return (dropped);
}
/*
* vm_page_remove_xbusy
*
* Removes the page but leaves the xbusy held. Returns true if this
* removed the final ref and false otherwise.
*/
bool
vm_page_remove_xbusy(vm_page_t m)
{
vm_page_object_remove(m);
return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF);
}
/*
* 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_lookup_unlocked:
*
* Returns the page associated with the object/offset pair specified;
* if none is found, NULL is returned. The page may be no longer be
* present in the object at the time that this function returns. Only
* useful for opportunistic checks such as inmem().
*/
vm_page_t
vm_page_lookup_unlocked(vm_object_t object, vm_pindex_t pindex)
{
return (vm_radix_lookup_unlocked(&object->rtree, pindex));
}
/*
* vm_page_relookup:
*
* Returns a page that must already have been busied by
* the caller. Used for bogus page replacement.
*/
vm_page_t
vm_page_relookup(vm_object_t object, vm_pindex_t pindex)
{
vm_page_t m;
m = vm_radix_lookup_unlocked(&object->rtree, pindex);
KASSERT(m != NULL && (vm_page_busied(m) || vm_page_wired(m)) &&
m->object == object && m->pindex == pindex,
("vm_page_relookup: Invalid page %p", m));
return (m);
}
/*
* This should only be used by lockless functions for releasing transient
* incorrect acquires. The page may have been freed after we acquired a
* busy lock. In this case busy_lock == VPB_FREED and we have nothing
* further to do.
*/
static void
vm_page_busy_release(vm_page_t m)
{
u_int x;
x = vm_page_busy_fetch(m);
for (;;) {
if (x == VPB_FREED)
break;
if ((x & VPB_BIT_SHARED) != 0 && VPB_SHARERS(x) > 1) {
if (atomic_fcmpset_int(&m->busy_lock, &x,
x - VPB_ONE_SHARER))
break;
continue;
}
KASSERT((x & VPB_BIT_SHARED) != 0 ||
(x & ~VPB_BIT_WAITERS) == VPB_CURTHREAD_EXCLUSIVE,
("vm_page_busy_release: %p xbusy not owned.", m));
if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED))
continue;
if ((x & VPB_BIT_WAITERS) != 0)
wakeup(m);
break;
}
}
/*
* 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.
*
* Both pages must be exclusively busied on enter. The old page is
* unbusied on exit.
*
* A return value of true means mold is now free. If this is not the
* final ref and the caller does not hold a wire reference it may not
* continue to access the page.
*/
static bool
vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
vm_page_t mold)
{
vm_page_t mret __diagused;
bool dropped;
VM_OBJECT_ASSERT_WLOCKED(object);
vm_page_assert_xbusied(mold);
KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0,
("vm_page_replace: page %p already in object", 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;
atomic_set_int(&mnew->ref_count, VPRC_OBJREF);
mret = vm_radix_replace(&object->rtree, mnew);
KASSERT(mret == mold,
("invalid page replacement, mold=%p, mret=%p", mold, mret));
KASSERT((mold->oflags & VPO_UNMANAGED) ==
(mnew->oflags & VPO_UNMANAGED),
("vm_page_replace: mismatched VPO_UNMANAGED"));
/* 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;
/*
* The object's resident_page_count does not change because we have
* swapped one page for another, but the generation count should
* change if the page is dirty.
*/
if (pmap_page_is_write_mapped(mnew))
vm_object_set_writeable_dirty(object);
dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF;
vm_page_xunbusy(mold);
return (dropped);
}
void
vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex,
vm_page_t mold)
{
vm_page_assert_xbusied(mnew);
if (vm_page_replace_hold(mnew, object, pindex, mold))
vm_page_free(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);
KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m));
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_object_remove(m);
/* Return back to the new pindex to complete vm_page_insert(). */
m->pindex = new_pindex;
m->object = new_object;
vm_page_insert_radixdone(m, new_object, mpred);
vm_page_dirty(m);
vm_pager_page_inserted(new_object, 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_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,
vm_radix_lookup_le(&object->rtree, pindex)));
}
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,
vm_radix_lookup_le(&object->rtree, pindex)));
}
/*
* 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.
*/
static int
_vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages)
{
u_int limit, old, new;
if (req_class == VM_ALLOC_INTERRUPT)
limit = 0;
else if (req_class == 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);
}
int
vm_domain_allocate(struct vm_domain *vmd, int req, int npages)
{
int req_class;
/*
* The page daemon is allowed to dig deeper into the free page list.
*/
req_class = req & VM_ALLOC_CLASS_MASK;
if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
req_class = VM_ALLOC_SYSTEM;
return (_vm_domain_allocate(vmd, req_class, npages));
}
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;
#define VPA_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \
VM_ALLOC_NOWAIT | VM_ALLOC_NOBUSY | \
VM_ALLOC_SBUSY | VM_ALLOC_WIRED | \
VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK)
KASSERT((req & ~VPA_FLAGS) == 0,
("invalid request %#x", req));
KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
("invalid request %#x", req));
KASSERT(mpred == NULL || mpred->pindex < pindex,
("mpred %p doesn't precede pindex 0x%jx", mpred,
(uintmax_t)pindex));
VM_OBJECT_ASSERT_WLOCKED(object);
flags = 0;
m = NULL;
if (!vm_pager_can_alloc_page(object, pindex))
return (NULL);
again:
#if VM_NRESERVLEVEL > 0
/*
* Can we allocate the page from a reservation?
*/
if (vm_object_reserv(object) &&
(m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) !=
NULL) {
goto found;
}
#endif
vmd = VM_DOMAIN(domain);
if (vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone != NULL) {
m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DEFAULT].zone,
M_NOWAIT | M_NOVM);
if (m != NULL) {
flags |= PG_PCPU_CACHE;
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, VM_FREEPOOL_DEFAULT, 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.
*/
found:
vm_page_dequeue(m);
vm_page_alloc_check(m);
/*
* Initialize the page. Only the PG_ZERO flag is inherited.
*/
flags |= m->flags & PG_ZERO;
if ((req & VM_ALLOC_NODUMP) != 0)
flags |= PG_NODUMP;
m->flags = flags;
m->a.flags = 0;
m->oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
m->busy_lock = VPB_CURTHREAD_EXCLUSIVE;
else if ((req & VM_ALLOC_SBUSY) != 0)
m->busy_lock = VPB_SHARERS_WORD(1);
else
m->busy_lock = VPB_UNBUSIED;
if (req & VM_ALLOC_WIRED) {
vm_wire_add(1);
m->ref_count = 1;
}
m->a.act_count = 0;
if (vm_page_insert_after(m, object, pindex, mpred)) {
if (req & VM_ALLOC_WIRED) {
vm_wire_sub(1);
m->ref_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);
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_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);
}
static vm_page_t
vm_page_find_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_page_t m_ret;
/*
* Can we allocate the pages without the number of free pages falling
* below the lower bound for the allocation class?
*/
vmd = VM_DOMAIN(domain);
if (!vm_domain_allocate(vmd, req, npages))
return (NULL);
/*
* Try to allocate the pages 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)
return (m_ret);
#if VM_NRESERVLEVEL > 0
/*
* Try to break a reservation to allocate the pages.
*/
if ((req & VM_ALLOC_NORECLAIM) == 0) {
m_ret = vm_reserv_reclaim_contig(domain, npages, low,
high, alignment, boundary);
if (m_ret != NULL)
return (m_ret);
}
#endif
vm_domain_freecnt_inc(vmd, npages);
return (NULL);
}
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)
{
vm_page_t m, m_ret, mpred;
u_int busy_lock, flags, oflags;
#define VPAC_FLAGS (VPA_FLAGS | VM_ALLOC_NORECLAIM)
KASSERT((req & ~VPAC_FLAGS) == 0,
("invalid request %#x", req));
KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
("invalid request %#x", req));
KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
(VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
("invalid request %#x", req));
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"));
mpred = vm_radix_lookup_le(&object->rtree, pindex);
KASSERT(mpred == NULL || mpred->pindex != pindex,
("vm_page_alloc_contig: pindex already allocated"));
for (;;) {
#if VM_NRESERVLEVEL > 0
/*
* Can we allocate the pages from a reservation?
*/
if (vm_object_reserv(object) &&
(m_ret = vm_reserv_alloc_contig(object, pindex, domain, req,
mpred, npages, low, high, alignment, boundary)) != NULL) {
break;
}
#endif
if ((m_ret = vm_page_find_contig_domain(domain, req, npages,
low, high, alignment, boundary)) != NULL)
break;
if (!vm_domain_alloc_fail(VM_DOMAIN(domain), object, req))
return (NULL);
}
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 = PG_ZERO;
if ((req & VM_ALLOC_NODUMP) != 0)
flags |= PG_NODUMP;
oflags = (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0;
if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
busy_lock = VPB_CURTHREAD_EXCLUSIVE;
else if ((req & VM_ALLOC_SBUSY) != 0)
busy_lock = VPB_SHARERS_WORD(1);
else
busy_lock = VPB_UNBUSIED;
if ((req & VM_ALLOC_WIRED) != 0)
vm_wire_add(npages);
if (object->memattr != VM_MEMATTR_DEFAULT &&
memattr == VM_MEMATTR_DEFAULT)
memattr = object->memattr;
for (m = m_ret; m < &m_ret[npages]; m++) {
m->a.flags = 0;
m->flags = (m->flags | PG_NODUMP) & flags;
m->busy_lock = busy_lock;
if ((req & VM_ALLOC_WIRED) != 0)
m->ref_count = 1;
m->a.act_count = 0;
m->oflags = oflags;
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->ref_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;
if (memattr != VM_MEMATTR_DEFAULT)
pmap_page_set_memattr(m, memattr);
pindex++;
}
return (m_ret);
}
/*
* Allocate a physical page that is not intended to be inserted into a VM
* object. If the "freelist" parameter is not equal to VM_NFREELIST, then only
* pages from the specified vm_phys freelist will be returned.
*/
static __always_inline vm_page_t
_vm_page_alloc_noobj_domain(int domain, const int freelist, int req)
{
struct vm_domain *vmd;
vm_page_t m;
int flags;
#define VPAN_FLAGS (VM_ALLOC_CLASS_MASK | VM_ALLOC_WAITFAIL | \
VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | \
VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | \
VM_ALLOC_NODUMP | VM_ALLOC_ZERO | VM_ALLOC_COUNT_MASK)
KASSERT((req & ~VPAN_FLAGS) == 0,
("invalid request %#x", req));
flags = (req & VM_ALLOC_NODUMP) != 0 ? PG_NODUMP : 0;
vmd = VM_DOMAIN(domain);
again:
if (freelist == VM_NFREELIST &&
vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone != NULL) {
m = uma_zalloc(vmd->vmd_pgcache[VM_FREEPOOL_DIRECT].zone,
M_NOWAIT | M_NOVM);
if (m != NULL) {
flags |= PG_PCPU_CACHE;
goto found;
}
}
if (vm_domain_allocate(vmd, req, 1)) {
vm_domain_free_lock(vmd);
if (freelist == VM_NFREELIST)
m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DIRECT, 0);
else
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 VM_NRESERVLEVEL > 0
if (freelist == VM_NFREELIST &&
vm_reserv_reclaim_inactive(domain))
goto again;
#endif
}
}
if (m == NULL) {
if (vm_domain_alloc_fail(vmd, NULL, req))
goto again;
return (NULL);
}
found:
vm_page_dequeue(m);
vm_page_alloc_check(m);
/*
* Consumers should not rely on a useful default pindex value.
*/
m->pindex = 0xdeadc0dedeadc0de;
m->flags = (m->flags & PG_ZERO) | flags;
m->a.flags = 0;
m->oflags = VPO_UNMANAGED;
m->busy_lock = VPB_UNBUSIED;
if ((req & VM_ALLOC_WIRED) != 0) {
vm_wire_add(1);
m->ref_count = 1;
}
if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
pmap_zero_page(m);
return (m);
}
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)
{
KASSERT(freelist >= 0 && freelist < VM_NFREELIST,
("%s: invalid freelist %d", __func__, freelist));
return (_vm_page_alloc_noobj_domain(domain, freelist, req));
}
vm_page_t
vm_page_alloc_noobj(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_noobj_domain(domain, req);
if (m != NULL)
break;
} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
return (m);
}
vm_page_t
vm_page_alloc_noobj_domain(int domain, int req)
{
return (_vm_page_alloc_noobj_domain(domain, VM_NFREELIST, req));
}
vm_page_t
vm_page_alloc_noobj_contig(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, NULL, 0, &domain, &req);
do {
m = vm_page_alloc_noobj_contig_domain(domain, req, npages, low,
high, alignment, boundary, memattr);
if (m != NULL)
break;
} while (vm_domainset_iter_page(&di, NULL, &domain) == 0);
return (m);
}
vm_page_t
vm_page_alloc_noobj_contig_domain(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)
{
vm_page_t m, m_ret;
u_int flags;
#define VPANC_FLAGS (VPAN_FLAGS | VM_ALLOC_NORECLAIM)
KASSERT((req & ~VPANC_FLAGS) == 0,
("invalid request %#x", req));
KASSERT((req & (VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM)) !=
(VM_ALLOC_WAITOK | VM_ALLOC_NORECLAIM),
("invalid request %#x", req));
KASSERT(((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
("invalid request %#x", req));
KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
while ((m_ret = vm_page_find_contig_domain(domain, req, npages,
low, high, alignment, boundary)) == NULL) {
if (!vm_domain_alloc_fail(VM_DOMAIN(domain), NULL, req))
return (NULL);
}
/*
* Initialize the pages. Only the PG_ZERO flag is inherited.
*/
flags = PG_ZERO;
if ((req & VM_ALLOC_NODUMP) != 0)
flags |= PG_NODUMP;
if ((req & VM_ALLOC_WIRED) != 0)
vm_wire_add(npages);
for (m = m_ret; m < &m_ret[npages]; m++) {
vm_page_dequeue(m);
vm_page_alloc_check(m);
/*
* Consumers should not rely on a useful default pindex value.
*/
m->pindex = 0xdeadc0dedeadc0de;
m->a.flags = 0;
m->flags = (m->flags | PG_NODUMP) & flags;
m->busy_lock = VPB_UNBUSIED;
if ((req & VM_ALLOC_WIRED) != 0)
m->ref_count = 1;
m->a.act_count = 0;
m->oflags = VPO_UNMANAGED;
/*
* Zero the page before updating any mappings since the page is
* not yet shared with any devices which might require the
* non-default memory attribute. pmap_page_set_memattr()
* flushes data caches before returning.
*/
if ((req & VM_ALLOC_ZERO) != 0 && (m->flags & PG_ZERO) == 0)
pmap_zero_page(m);
if (memattr != VM_MEMATTR_DEFAULT)
pmap_page_set_memattr(m, memattr);
}
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->a.queue == PQ_NONE &&
(m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
("page %p has unexpected queue %d, flags %#x",
m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK)));
KASSERT(m->ref_count == 0, ("page %p has references", m));
KASSERT(vm_page_busy_freed(m), ("page %p is not freed", 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(vm_page_none_valid(m), ("free page %p is valid", m));
pmap_vm_page_alloc_check(m);
}
static int
vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags)
{
struct vm_domain *vmd;
struct vm_pgcache *pgcache;
int i;
pgcache = arg;
vmd = VM_DOMAIN(pgcache->domain);
/*
* The page daemon should avoid creating extra memory pressure since its
* main purpose is to replenish the store of free pages.
*/
if (vmd->vmd_severeset || curproc == pageproc ||
!_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, pgcache->pool, 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_zone_release(void *arg, void **store, int cnt)
{
struct vm_domain *vmd;
struct vm_pgcache *pgcache;
vm_page_t m;
int i;
pgcache = arg;
vmd = VM_DOMAIN(pgcache->domain);
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)
{
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;
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->ref_count >= 1,
("fictitious page %p has invalid ref 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 (!vm_addr_align_ok(pa, alignment)) {
m_inc = atop(roundup2(pa, alignment) - pa);
continue;
}
if (!vm_addr_bound_ok(pa, ptoa(npages), boundary)) {
m_inc = atop(roundup2(pa, boundary) - pa);
continue;
}
} else
KASSERT(m_run != NULL, ("m_run == NULL"));
retry:
m_inc = 1;
if (vm_page_wired(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 = atomic_load_ptr(&m->object)) != NULL) {
/*
* The page is considered eligible for relocation if
* and only if it could be laundered or reclaimed by
* the page daemon.
*/
VM_OBJECT_RLOCK(object);
if (object != m->object) {
VM_OBJECT_RUNLOCK(object);
goto retry;
}
/* Don't care: PG_NODUMP, PG_ZERO. */
if ((object->flags & OBJ_SWAP) == 0 &&
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: PGA_NOSYNC. */
run_ext = 1;
} else
run_ext = 0;
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 busy 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 (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 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;
for (; error == 0 && m < m_end; m++) {
KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
("page %p is PG_FICTITIOUS or PG_MARKER", m));
/*
* Racily check for wirings. Races are handled once the object
* lock is held and the page is unmapped.
*/
if (vm_page_wired(m))
error = EBUSY;
else if ((object = atomic_load_ptr(&m->object)) != NULL) {
/*
* The page is relocated if and only if it could be
* laundered or reclaimed by the page daemon.
*/
VM_OBJECT_WLOCK(object);
/* Don't care: PG_NODUMP, PG_ZERO. */
if (m->object != object ||
((object->flags & OBJ_SWAP) == 0 &&
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_tryxbusy(m) != 0) {
if (vm_page_wired(m)) {
vm_page_xunbusy(m);
error = EBUSY;
goto unlock;
}
KASSERT(pmap_page_get_memattr(m) ==
VM_MEMATTR_DEFAULT,
("page %p has an unexpected memattr", m));
KASSERT(m->oflags == 0,
("page %p has unexpected oflags", m));
/* Don't care: PGA_NOSYNC. */
if (!vm_page_none_valid(m)) {
/*
* 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;
if ((m->flags & PG_NODUMP) != 0)
req |= VM_ALLOC_NODUMP;
if (trunc_page(high) !=
~(vm_paddr_t)PAGE_MASK) {
m_new =
vm_page_alloc_noobj_contig(
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_noobj_contig(
req, 1, 0, pa - 1,
PAGE_SIZE, 0,
VM_MEMATTR_DEFAULT);
}
if (m_new == NULL) {
pa += ptoa(npages);
m_new =
vm_page_alloc_noobj_contig(
req, 1, pa, high, PAGE_SIZE,
0, VM_MEMATTR_DEFAULT);
}
if (m_new == NULL) {
vm_page_xunbusy(m);
error = ENOMEM;
goto unlock;
}
/*
* Unmap the page and check for new
* wirings that may have been acquired
* through a pmap lookup.
*/
if (object->ref_count != 0 &&
!vm_page_try_remove_all(m)) {
vm_page_xunbusy(m);
vm_page_free(m_new);
error = EBUSY;
goto unlock;
}
/*
* 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.
*/
m_new->a.flags = m->a.flags &
~PGA_QUEUE_STATE_MASK;
KASSERT(m_new->oflags == VPO_UNMANAGED,
("page %p is managed", m_new));
m_new->oflags = 0;
pmap_copy_page(m, m_new);
m_new->valid = m->valid;
m_new->dirty = m->dirty;
m->flags &= ~PG_ZERO;
vm_page_dequeue(m);
if (vm_page_replace_hold(m_new, object,
m->pindex, m) &&
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_deactivate(m_new);
} else {
m->flags &= ~PG_ZERO;
vm_page_dequeue(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_page_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 = SLIST_FIRST(&free)) != NULL) {
int cnt;
vmd = VM_DOMAIN(domain);
cnt = 0;
vm_domain_free_lock(vmd);
do {
MPASS(vm_page_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, minalign, 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"));
/*
* The caller will attempt an allocation after some runs have been
* reclaimed and added to the vm_phys buddy lists. Due to limitations
* of vm_phys_alloc_contig(), round up the requested length to the next
* power of two or maximum chunk size, and ensure that each run is
* suitably aligned.
*/
minalign = 1ul << imin(flsl(npages - 1), VM_NFREEORDER - 1);
npages = roundup2(npages, minalign);
if (alignment < ptoa(minalign))
alignment = ptoa(minalign);
/*
* The page daemon is allowed to dig deeper into the free page list.
*/
req_class = req & VM_ALLOC_CLASS_MASK;
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);
}
int
vm_wait_doms(const domainset_t *wdoms, int mflags)
{
int error;
error = 0;
/*
* 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++;
error = msleep(&vm_pageproc_waiters, &vm_domainset_lock,
PVM | PDROP | mflags, "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)) {
if (pageproc == NULL)
panic("vm_wait in early boot");
vm_min_waiters++;
error = msleep(&vm_min_domains, &vm_domainset_lock,
PVM | PDROP | mflags, "vmwait", 0);
} else
mtx_unlock(&vm_domainset_lock);
}
return (error);
}
/*
* 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 {
DOMAINSET_ZERO(&wdom);
DOMAINSET_SET(vmd->vmd_domain, &wdom);
vm_wait_doms(&wdom, 0);
}
}
static int
vm_wait_flags(vm_object_t obj, int mflags)
{
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;
return (vm_wait_doms(&d->ds_mask, mflags));
}
/*
* 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)
{
(void)vm_wait_flags(obj, 0);
}
int
vm_wait_intr(vm_object_t obj)
{
return (vm_wait_flags(obj, PCATCH));
}
/*
* 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, int timo)
{
/*
* 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", timo);
} else
mtx_unlock(&vm_domainset_lock);
}
static struct vm_pagequeue *
_vm_page_pagequeue(vm_page_t m, uint8_t queue)
{
return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]);
}
#ifdef INVARIANTS
static struct vm_pagequeue *
vm_page_pagequeue(vm_page_t m)
{
return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue));
}
#endif
static __always_inline bool
vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
{
vm_page_astate_t tmp;
tmp = *old;
do {
if (__predict_true(vm_page_astate_fcmpset(m, old, new)))
return (true);
counter_u64_add(pqstate_commit_retries, 1);
} while (old->_bits == tmp._bits);
return (false);
}
/*
* Do the work of committing a queue state update that moves the page out of
* its current queue.
*/
static bool
_vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m,
vm_page_astate_t *old, vm_page_astate_t new)
{
vm_page_t next;
vm_pagequeue_assert_locked(pq);
KASSERT(vm_page_pagequeue(m) == pq,
("%s: queue %p does not match page %p", __func__, pq, m));
KASSERT(old->queue != PQ_NONE && new.queue != old->queue,
("%s: invalid queue indices %d %d",
__func__, old->queue, new.queue));
/*
* Once the queue index of the page changes there is nothing
* synchronizing with further updates to the page's physical
* queue state. Therefore we must speculatively remove the page
* from the queue now and be prepared to roll back if the queue
* state update fails. If the page is not physically enqueued then
* we just update its queue index.
*/
if ((old->flags & PGA_ENQUEUED) != 0) {
new.flags &= ~PGA_ENQUEUED;
next = TAILQ_NEXT(m, plinks.q);
TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
vm_pagequeue_cnt_dec(pq);
if (!vm_page_pqstate_fcmpset(m, old, new)) {
if (next == NULL)
TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
else
TAILQ_INSERT_BEFORE(next, m, plinks.q);
vm_pagequeue_cnt_inc(pq);
return (false);
} else {
return (true);
}
} else {
return (vm_page_pqstate_fcmpset(m, old, new));
}
}
static bool
vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old,
vm_page_astate_t new)
{
struct vm_pagequeue *pq;
vm_page_astate_t as;
bool ret;
pq = _vm_page_pagequeue(m, old->queue);
/*
* The queue field and PGA_ENQUEUED flag are stable only so long as the
* corresponding page queue lock is held.
*/
vm_pagequeue_lock(pq);
as = vm_page_astate_load(m);
if (__predict_false(as._bits != old->_bits)) {
*old = as;
ret = false;
} else {
ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new);
}
vm_pagequeue_unlock(pq);
return (ret);
}
/*
* Commit a queue state update that enqueues or requeues a page.
*/
static bool
_vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m,
vm_page_astate_t *old, vm_page_astate_t new)
{
struct vm_domain *vmd;
vm_pagequeue_assert_locked(pq);
KASSERT(old->queue != PQ_NONE && new.queue == old->queue,
("%s: invalid queue indices %d %d",
__func__, old->queue, new.queue));
new.flags |= PGA_ENQUEUED;
if (!vm_page_pqstate_fcmpset(m, old, new))
return (false);
if ((old->flags & PGA_ENQUEUED) != 0)
TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
else
vm_pagequeue_cnt_inc(pq);
/*
* Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if
* both flags are set in close succession, only PGA_REQUEUE_HEAD will be
* applied, even if it was set first.
*/
if ((old->flags & PGA_REQUEUE_HEAD) != 0) {
vmd = vm_pagequeue_domain(m);
KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE],
("%s: invalid page queue for page %p", __func__, m));
TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
} else {
TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
}
return (true);
}
/*
* Commit a queue state update that encodes a request for a deferred queue
* operation.
*/
static bool
vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old,
vm_page_astate_t new)
{
KASSERT(old->queue == new.queue || new.queue != PQ_NONE,
("%s: invalid state, queue %d flags %x",
__func__, new.queue, new.flags));
if (old->_bits != new._bits &&
!vm_page_pqstate_fcmpset(m, old, new))
return (false);
vm_page_pqbatch_submit(m, new.queue);
return (true);
}
/*
* A generic queue state update function. This handles more cases than the
* specialized functions above.
*/
bool
vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new)
{
if (old->_bits == new._bits)
return (true);
if (old->queue != PQ_NONE && new.queue != old->queue) {
if (!vm_page_pqstate_commit_dequeue(m, old, new))
return (false);
if (new.queue != PQ_NONE)
vm_page_pqbatch_submit(m, new.queue);
} else {
if (!vm_page_pqstate_fcmpset(m, old, new))
return (false);
if (new.queue != PQ_NONE &&
((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0)
vm_page_pqbatch_submit(m, new.queue);
}
return (true);
}
/*
* Apply deferred queue state updates to a page.
*/
static inline void
vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue)
{
vm_page_astate_t new, old;
CRITICAL_ASSERT(curthread);
vm_pagequeue_assert_locked(pq);
KASSERT(queue < PQ_COUNT,
("%s: invalid queue index %d", __func__, queue));
KASSERT(pq == _vm_page_pagequeue(m, queue),
("%s: page %p does not belong to queue %p", __func__, m, pq));
for (old = vm_page_astate_load(m);;) {
if (__predict_false(old.queue != queue ||
(old.flags & PGA_QUEUE_OP_MASK) == 0)) {
counter_u64_add(queue_nops, 1);
break;
}
KASSERT((m->oflags & VPO_UNMANAGED) == 0,
("%s: page %p is unmanaged", __func__, m));
new = old;
if ((old.flags & PGA_DEQUEUE) != 0) {
new.flags &= ~PGA_QUEUE_OP_MASK;
new.queue = PQ_NONE;
if (__predict_true(_vm_page_pqstate_commit_dequeue(pq,
m, &old, new))) {
counter_u64_add(queue_ops, 1);
break;
}
} else {
new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD);
if (__predict_true(_vm_page_pqstate_commit_requeue(pq,
m, &old, new))) {
counter_u64_add(queue_ops, 1);
break;
}
}
}
}
static void
vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq,
uint8_t queue)
{
int i;
for (i = 0; i < bq->bq_cnt; i++)
vm_pqbatch_process_page(pq, bq->bq_pa[i], queue);
vm_batchqueue_init(bq);
}
/*
* vm_page_pqbatch_submit: [ internal use only ]
*
* Enqueue a page in the specified page queue's batched work queue.
* The caller must have encoded the requested operation in the page
* structure's a.flags field.
*/
void
vm_page_pqbatch_submit(vm_page_t m, uint8_t queue)
{
struct vm_batchqueue *bq;
struct vm_pagequeue *pq;
- int domain;
+ int domain, slots_remaining;
KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue));
domain = vm_page_domain(m);
critical_enter();
bq = DPCPU_PTR(pqbatch[domain][queue]);
- if (vm_batchqueue_insert(bq, m)) {
+ slots_remaining = vm_batchqueue_insert(bq, m);
+ if (slots_remaining > (VM_BATCHQUEUE_SIZE >> 1)) {
+ /* keep building the bq */
+ critical_exit();
+ return;
+ } else if (slots_remaining > 0 ) {
+ /* Try to process the bq if we can get the lock */
+ pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
+ if (vm_pagequeue_trylock(pq)) {
+ vm_pqbatch_process(pq, bq, queue);
+ vm_pagequeue_unlock(pq);
+ }
critical_exit();
return;
}
critical_exit();
+ /* if we make it here, the bq is full so wait for the lock */
+
pq = &VM_DOMAIN(domain)->vmd_pagequeues[queue];
vm_pagequeue_lock(pq);
critical_enter();
bq = DPCPU_PTR(pqbatch[domain][queue]);
vm_pqbatch_process(pq, bq, queue);
vm_pqbatch_process_page(pq, m, queue);
vm_pagequeue_unlock(pq);
critical_exit();
}
/*
* vm_page_pqbatch_drain: [ 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_pqbatch_drain(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);
}
/*
* 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.
*/
void
vm_page_dequeue_deferred(vm_page_t m)
{
vm_page_astate_t new, old;
old = vm_page_astate_load(m);
do {
if (old.queue == PQ_NONE) {
KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
("%s: page %p has unexpected queue state",
__func__, m));
break;
}
new = old;
new.flags |= PGA_DEQUEUE;
} while (!vm_page_pqstate_commit_request(m, &old, new));
}
/*
* vm_page_dequeue:
*
* Remove the page from whichever page queue it's in, if any, before
* returning.
*/
void
vm_page_dequeue(vm_page_t m)
{
vm_page_astate_t new, old;
old = vm_page_astate_load(m);
do {
if (old.queue == PQ_NONE) {
KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0,
("%s: page %p has unexpected queue state",
__func__, m));
break;
}
new = old;
new.flags &= ~PGA_QUEUE_OP_MASK;
new.queue = PQ_NONE;
} while (!vm_page_pqstate_commit_dequeue(m, &old, new));
}
/*
* 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)
{
KASSERT(m->a.queue == PQ_NONE &&
(m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
("%s: page %p is already enqueued", __func__, m));
KASSERT(m->ref_count > 0,
("%s: page %p does not carry any references", __func__, m));
m->a.queue = queue;
if ((m->a.flags & PGA_REQUEUE) == 0)
vm_page_aflag_set(m, PGA_REQUEUE);
vm_page_pqbatch_submit(m, queue);
}
/*
* 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, if it exists, must be locked, and then the page must
* be xbusy. Otherwise the page must be not busied. A managed
* page must be unmapped.
*/
static bool
vm_page_free_prep(vm_page_t m)
{
/*
* Synchronize with threads that have dropped a reference to this
* page.
*/
atomic_thread_fence_acq();
#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) {
KASSERT(!pmap_page_is_mapped(m),
("vm_page_free_prep: freeing mapped page %p", m));
KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0,
("vm_page_free_prep: mapping flags set in page %p", m));
} else {
KASSERT(m->a.queue == PQ_NONE,
("vm_page_free_prep: unmanaged page %p is queued", m));
}
VM_CNT_INC(v_tfree);
if (m->object != NULL) {
KASSERT(((m->oflags & VPO_UNMANAGED) != 0) ==
((m->object->flags & OBJ_UNMANAGED) != 0),
("vm_page_free_prep: managed flag mismatch for page %p",
m));
vm_page_assert_xbusied(m);
/*
* The object reference can be released without an atomic
* operation.
*/
KASSERT((m->flags & PG_FICTITIOUS) != 0 ||
m->ref_count == VPRC_OBJREF,
("vm_page_free_prep: page %p has unexpected ref_count %u",
m, m->ref_count));
vm_page_object_remove(m);
m->ref_count -= VPRC_OBJREF;
} else
vm_page_assert_unbusied(m);
vm_page_busy_free(m);
/*
* If fictitious remove object association and
* return.
*/
if ((m->flags & PG_FICTITIOUS) != 0) {
KASSERT(m->ref_count == 1,
("fictitious page %p is referenced", m));
KASSERT(m->a.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->ref_count != 0)
panic("vm_page_free_prep: page %p has references", m);
/*
* 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
/*
* Determine whether the page belongs to a reservation. If the page was
* allocated from a per-CPU cache, it cannot belong to a reservation, so
* as an optimization, we avoid the check in that case.
*/
if ((m->flags & PG_PCPU_CACHE) == 0 && 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 exclusively busied if it
* belongs to an object.
*/
static void
vm_page_free_toq(vm_page_t m)
{
struct vm_domain *vmd;
uma_zone_t zone;
if (!vm_page_free_prep(m))
return;
vmd = vm_pagequeue_domain(m);
zone = vmd->vmd_pgcache[m->pool].zone;
if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) {
uma_zfree(zone, 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.
*/
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);
}
/*
* Mark this page as wired down. For managed pages, this prevents reclamation
* by the page daemon, or when the containing object, if any, is destroyed.
*/
void
vm_page_wire(vm_page_t m)
{
u_int old;
#ifdef INVARIANTS
if (m->object != NULL && !vm_page_busied(m) &&
!vm_object_busied(m->object))
VM_OBJECT_ASSERT_LOCKED(m->object);
#endif
KASSERT((m->flags & PG_FICTITIOUS) == 0 ||
VPRC_WIRE_COUNT(m->ref_count) >= 1,
("vm_page_wire: fictitious page %p has zero wirings", m));
old = atomic_fetchadd_int(&m->ref_count, 1);
KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX,
("vm_page_wire: counter overflow for page %p", m));
if (VPRC_WIRE_COUNT(old) == 0) {
if ((m->oflags & VPO_UNMANAGED) == 0)
vm_page_aflag_set(m, PGA_DEQUEUE);
vm_wire_add(1);
}
}
/*
* Attempt to wire a mapped page following a pmap lookup of that page.
* This may fail if a thread is concurrently tearing down mappings of the page.
* The transient failure is acceptable because it translates to the
* failure of the caller pmap_extract_and_hold(), which should be then
* followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages().
*/
bool
vm_page_wire_mapped(vm_page_t m)
{
u_int old;
old = m->ref_count;
do {
KASSERT(old > 0,
("vm_page_wire_mapped: wiring unreferenced page %p", m));
if ((old & VPRC_BLOCKED) != 0)
return (false);
} while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1));
if (VPRC_WIRE_COUNT(old) == 0) {
if ((m->oflags & VPO_UNMANAGED) == 0)
vm_page_aflag_set(m, PGA_DEQUEUE);
vm_wire_add(1);
}
return (true);
}
/*
* Release a wiring reference to a managed page. If the page still belongs to
* an object, update its position in the page queues to reflect the reference.
* If the wiring was the last reference to the page, free the page.
*/
static void
vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse)
{
u_int old;
KASSERT((m->oflags & VPO_UNMANAGED) == 0,
("%s: page %p is unmanaged", __func__, m));
/*
* Update LRU state before releasing the wiring reference.
* Use a release store when updating the reference count to
* synchronize with vm_page_free_prep().
*/
old = m->ref_count;
do {
KASSERT(VPRC_WIRE_COUNT(old) > 0,
("vm_page_unwire: wire count underflow for page %p", m));
if (old > VPRC_OBJREF + 1) {
/*
* The page has at least one other wiring reference. An
* earlier iteration of this loop may have called
* vm_page_release_toq() and cleared PGA_DEQUEUE, so
* re-set it if necessary.
*/
if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0)
vm_page_aflag_set(m, PGA_DEQUEUE);
} else if (old == VPRC_OBJREF + 1) {
/*
* This is the last wiring. Clear PGA_DEQUEUE and
* update the page's queue state to reflect the
* reference. If the page does not belong to an object
* (i.e., the VPRC_OBJREF bit is clear), we only need to
* clear leftover queue state.
*/
vm_page_release_toq(m, nqueue, noreuse);
} else if (old == 1) {
vm_page_aflag_clear(m, PGA_DEQUEUE);
}
} while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1));
if (VPRC_WIRE_COUNT(old) == 1) {
vm_wire_sub(1);
if (old == 1)
vm_page_free(m);
}
}
/*
* Release one wiring of the specified page, potentially allowing it to be
* paged out.
*
* 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. If the released wiring
* represented the last reference to the page, the page is freed.
*/
void
vm_page_unwire(vm_page_t m, uint8_t nqueue)
{
KASSERT(nqueue < PQ_COUNT,
("vm_page_unwire: invalid queue %u request for page %p",
nqueue, m));
if ((m->oflags & VPO_UNMANAGED) != 0) {
if (vm_page_unwire_noq(m) && m->ref_count == 0)
vm_page_free(m);
return;
}
vm_page_unwire_managed(m, nqueue, false);
}
/*
* 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 involving managed pages, vm_page_unwire() should be used
* instead.
*/
bool
vm_page_unwire_noq(vm_page_t m)
{
u_int old;
old = vm_page_drop(m, 1);
KASSERT(VPRC_WIRE_COUNT(old) != 0,
("%s: counter underflow for page %p", __func__, m));
KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1,
("%s: missing ref on fictitious page %p", __func__, m));
if (VPRC_WIRE_COUNT(old) > 1)
return (false);
if ((m->oflags & VPO_UNMANAGED) == 0)
vm_page_aflag_clear(m, PGA_DEQUEUE);
vm_wire_sub(1);
return (true);
}
/*
* Ensure that the page ends up in the specified page queue. If the page is
* active or being moved to the active queue, ensure that its act_count is
* at least ACT_INIT but do not otherwise mess with it.
*/
static __always_inline void
vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag)
{
vm_page_astate_t old, new;
KASSERT(m->ref_count > 0,
("%s: page %p does not carry any references", __func__, m));
KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD,
("%s: invalid flags %x", __func__, nflag));
if ((m->oflags & VPO_UNMANAGED) != 0 || vm_page_wired(m))
return;
old = vm_page_astate_load(m);
do {
if ((old.flags & PGA_DEQUEUE) != 0)
break;
new = old;
new.flags &= ~PGA_QUEUE_OP_MASK;
if (nqueue == PQ_ACTIVE)
new.act_count = max(old.act_count, ACT_INIT);
if (old.queue == nqueue) {
/*
* There is no need to requeue pages already in the
* active queue.
*/
if (nqueue != PQ_ACTIVE ||
(old.flags & PGA_ENQUEUED) == 0)
new.flags |= nflag;
} else {
new.flags |= nflag;
new.queue = nqueue;
}
} while (!vm_page_pqstate_commit(m, &old, new));
}
/*
* Put the specified page on the active list (if appropriate).
*/
void
vm_page_activate(vm_page_t m)
{
vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE);
}
/*
* Move the specified page to the tail of the inactive queue, or requeue
* the page if it is already in the inactive queue.
*/
void
vm_page_deactivate(vm_page_t m)
{
vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE);
}
void
vm_page_deactivate_noreuse(vm_page_t m)
{
vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD);
}
/*
* Put a page in the laundry, or requeue it if it is already there.
*/
void
vm_page_launder(vm_page_t m)
{
vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE);
}
/*
* Put a page in the PQ_UNSWAPPABLE holding queue.
*/
void
vm_page_unswappable(vm_page_t m)
{
VM_OBJECT_ASSERT_LOCKED(m->object);
KASSERT((m->oflags & VPO_UNMANAGED) == 0,
("page %p already unswappable", m));
vm_page_dequeue(m);
vm_page_enqueue(m, PQ_UNSWAPPABLE);
}
/*
* Release a page back to the page queues in preparation for unwiring.
*/
static void
vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse)
{
vm_page_astate_t old, new;
uint16_t nflag;
/*
* Use a check of the valid bits to determine whether we should
* accelerate reclamation of the page. The object lock might not be
* held here, in which case the check is racy. At worst we will either
* accelerate reclamation of a valid page and violate LRU, or
* unnecessarily defer reclamation of an invalid page.
*
* If we were asked to not cache the page, place it near the head of the
* inactive queue so that is reclaimed sooner.
*/
if (noreuse || vm_page_none_valid(m)) {
nqueue = PQ_INACTIVE;
nflag = PGA_REQUEUE_HEAD;
} else {
nflag = PGA_REQUEUE;
}
old = vm_page_astate_load(m);
do {
new = old;
/*
* If the page is already in the active queue and we are not
* trying to accelerate reclamation, simply mark it as
* referenced and avoid any queue operations.
*/
new.flags &= ~PGA_QUEUE_OP_MASK;
if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE &&
(old.flags & PGA_ENQUEUED) != 0)
new.flags |= PGA_REFERENCED;
else {
new.flags |= nflag;
new.queue = nqueue;
}
} while (!vm_page_pqstate_commit(m, &old, new));
}
/*
* Unwire a page and either attempt to free it or re-add it to the page queues.
*/
void
vm_page_release(vm_page_t m, int flags)
{
vm_object_t object;
KASSERT((m->oflags & VPO_UNMANAGED) == 0,
("vm_page_release: page %p is unmanaged", m));
if ((flags & VPR_TRYFREE) != 0) {
for (;;) {
object = atomic_load_ptr(&m->object);
if (object == NULL)
break;
/* Depends on type-stability. */
if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object))
break;
if (object == m->object) {
vm_page_release_locked(m, flags);
VM_OBJECT_WUNLOCK(object);
return;
}
VM_OBJECT_WUNLOCK(object);
}
}
vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0);
}
/* See vm_page_release(). */
void
vm_page_release_locked(vm_page_t m, int flags)
{
VM_OBJECT_ASSERT_WLOCKED(m->object);
KASSERT((m->oflags & VPO_UNMANAGED) == 0,
("vm_page_release_locked: page %p is unmanaged", m));
if (vm_page_unwire_noq(m)) {
if ((flags & VPR_TRYFREE) != 0 &&
(m->object->ref_count == 0 || !pmap_page_is_mapped(m)) &&
m->dirty == 0 && vm_page_tryxbusy(m)) {
/*
* An unlocked lookup may have wired the page before the
* busy lock was acquired, in which case the page must
* not be freed.
*/
if (__predict_true(!vm_page_wired(m))) {
vm_page_free(m);
return;
}
vm_page_xunbusy(m);
} else {
vm_page_release_toq(m, PQ_INACTIVE, flags != 0);
}
}
}
static bool
vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t))
{
u_int old;
KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0,
("vm_page_try_blocked_op: page %p has no object", m));
KASSERT(vm_page_busied(m),
("vm_page_try_blocked_op: page %p is not busy", m));
VM_OBJECT_ASSERT_LOCKED(m->object);
old = m->ref_count;
do {
KASSERT(old != 0,
("vm_page_try_blocked_op: page %p has no references", m));
if (VPRC_WIRE_COUNT(old) != 0)
return (false);
} while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED));
(op)(m);
/*
* If the object is read-locked, new wirings may be created via an
* object lookup.
*/
old = vm_page_drop(m, VPRC_BLOCKED);
KASSERT(!VM_OBJECT_WOWNED(m->object) ||
old == (VPRC_BLOCKED | VPRC_OBJREF),
("vm_page_try_blocked_op: unexpected refcount value %u for %p",
old, m));
return (true);
}
/*
* Atomically check for wirings and remove all mappings of the page.
*/
bool
vm_page_try_remove_all(vm_page_t m)
{
return (vm_page_try_blocked_op(m, pmap_remove_all));
}
/*
* Atomically check for wirings and remove all writeable mappings of the page.
*/
bool
vm_page_try_remove_write(vm_page_t m)
{
return (vm_page_try_blocked_op(m, pmap_remove_write));
}
/*
* vm_page_advise
*
* Apply the specified advice to the given page.
*/
void
vm_page_advise(vm_page_t m, int advice)
{
VM_OBJECT_ASSERT_WLOCKED(m->object);
vm_page_assert_xbusied(m);
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;
}
if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
vm_page_dirty(m);
/*
* Clear any references to the page. Otherwise, the page daemon will
* immediately reactivate the page.
*/
vm_page_aflag_clear(m, PGA_REFERENCED);
/*
* 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);
}
/*
* vm_page_grab_release
*
* Helper routine for grab functions to release busy on return.
*/
static inline void
vm_page_grab_release(vm_page_t m, int allocflags)
{
if ((allocflags & VM_ALLOC_NOBUSY) != 0) {
if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
vm_page_sunbusy(m);
else
vm_page_xunbusy(m);
}
}
/*
* vm_page_grab_sleep
*
* Sleep for busy according to VM_ALLOC_ parameters. Returns true
* if the caller should retry and false otherwise.
*
* If the object is locked on entry the object will be unlocked with
* false returns and still locked but possibly having been dropped
* with true returns.
*/
static bool
vm_page_grab_sleep(vm_object_t object, vm_page_t m, vm_pindex_t pindex,
const char *wmesg, int allocflags, bool locked)
{
if ((allocflags & VM_ALLOC_NOWAIT) != 0)
return (false);
/*
* Reference the page before unlocking and sleeping so that
* the page daemon is less likely to reclaim it.
*/
if (locked && (allocflags & VM_ALLOC_NOCREAT) == 0)
vm_page_reference(m);
if (_vm_page_busy_sleep(object, m, pindex, wmesg, allocflags, locked) &&
locked)
VM_OBJECT_WLOCK(object);
if ((allocflags & VM_ALLOC_WAITFAIL) != 0)
return (false);
return (true);
}
/*
* Assert that the grab flags are valid.
*/
static inline void
vm_page_grab_check(int allocflags)
{
KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 ||
(allocflags & VM_ALLOC_WIRED) != 0,
("vm_page_grab*: the pages must be busied or wired"));
KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
(allocflags & VM_ALLOC_IGN_SBUSY) != 0,
("vm_page_grab*: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
}
/*
* Calculate the page allocation flags for grab.
*/
static inline int
vm_page_grab_pflags(int allocflags)
{
int pflags;
pflags = allocflags &
~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL |
VM_ALLOC_NOBUSY | VM_ALLOC_IGN_SBUSY);
if ((allocflags & VM_ALLOC_NOWAIT) == 0)
pflags |= VM_ALLOC_WAITFAIL;
if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0)
pflags |= VM_ALLOC_SBUSY;
return (pflags);
}
/*
* 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;
VM_OBJECT_ASSERT_WLOCKED(object);
vm_page_grab_check(allocflags);
retrylookup:
if ((m = vm_page_lookup(object, pindex)) != NULL) {
if (!vm_page_tryacquire(m, allocflags)) {
if (vm_page_grab_sleep(object, m, pindex, "pgrbwt",
allocflags, true))
goto retrylookup;
return (NULL);
}
goto out;
}
if ((allocflags & VM_ALLOC_NOCREAT) != 0)
return (NULL);
m = vm_page_alloc(object, pindex, vm_page_grab_pflags(allocflags));
if (m == NULL) {
if ((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL)) != 0)
return (NULL);
goto retrylookup;
}
if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
pmap_zero_page(m);
out:
vm_page_grab_release(m, allocflags);
return (m);
}
/*
* Locklessly attempt to acquire a page given a (object, pindex) tuple
* and an optional previous page to avoid the radix lookup. The resulting
* page will be validated against the identity tuple and busied or wired
* as requested. A NULL *mp return guarantees that the page was not in
* radix at the time of the call but callers must perform higher level
* synchronization or retry the operation under a lock if they require
* an atomic answer. This is the only lock free validation routine,
* other routines can depend on the resulting page state.
*
* The return value indicates whether the operation failed due to caller
* flags. The return is tri-state with mp:
*
* (true, *mp != NULL) - The operation was successful.
* (true, *mp == NULL) - The page was not found in tree.
* (false, *mp == NULL) - WAITFAIL or NOWAIT prevented acquisition.
*/
static bool
vm_page_acquire_unlocked(vm_object_t object, vm_pindex_t pindex,
vm_page_t prev, vm_page_t *mp, int allocflags)
{
vm_page_t m;
vm_page_grab_check(allocflags);
MPASS(prev == NULL || vm_page_busied(prev) || vm_page_wired(prev));
*mp = NULL;
for (;;) {
/*
* We may see a false NULL here because the previous page
* has been removed or just inserted and the list is loaded
* without barriers. Switch to radix to verify.
*/
if (prev == NULL || (m = TAILQ_NEXT(prev, listq)) == NULL ||
QMD_IS_TRASHED(m) || m->pindex != pindex ||
atomic_load_ptr(&m->object) != object) {
prev = NULL;
/*
* This guarantees the result is instantaneously
* correct.
*/
m = vm_radix_lookup_unlocked(&object->rtree, pindex);
}
if (m == NULL)
return (true);
if (vm_page_trybusy(m, allocflags)) {
if (m->object == object && m->pindex == pindex)
break;
/* relookup. */
vm_page_busy_release(m);
cpu_spinwait();
continue;
}
if (!vm_page_grab_sleep(object, m, pindex, "pgnslp",
allocflags, false))
return (false);
}
if ((allocflags & VM_ALLOC_WIRED) != 0)
vm_page_wire(m);
vm_page_grab_release(m, allocflags);
*mp = m;
return (true);
}
/*
* Try to locklessly grab a page and fall back to the object lock if NOCREAT
* is not set.
*/
vm_page_t
vm_page_grab_unlocked(vm_object_t object, vm_pindex_t pindex, int allocflags)
{
vm_page_t m;
vm_page_grab_check(allocflags);
if (!vm_page_acquire_unlocked(object, pindex, NULL, &m, allocflags))
return (NULL);
if (m != NULL)
return (m);
/*
* The radix lockless lookup should never return a false negative
* errors. If the user specifies NOCREAT they are guaranteed there
* was no page present at the instant of the call. A NOCREAT caller
* must handle create races gracefully.
*/
if ((allocflags & VM_ALLOC_NOCREAT) != 0)
return (NULL);
VM_OBJECT_WLOCK(object);
m = vm_page_grab(object, pindex, allocflags);
VM_OBJECT_WUNLOCK(object);
return (m);
}
/*
* Grab a page and make it valid, paging in if necessary. Pages missing from
* their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied
* and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought
* in simultaneously. Additional pages will be left on a paging queue but
* will neither be wired nor busy regardless of allocflags.
*/
int
vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags)
{
vm_page_t m;
vm_page_t ma[VM_INITIAL_PAGEIN];
int after, i, pflags, rv;
KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
(allocflags & VM_ALLOC_IGN_SBUSY) != 0,
("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
KASSERT((allocflags &
(VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
("vm_page_grab_valid: Invalid flags 0x%X", allocflags));
VM_OBJECT_ASSERT_WLOCKED(object);
pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY |
VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY);
pflags |= VM_ALLOC_WAITFAIL;
retrylookup:
if ((m = vm_page_lookup(object, pindex)) != NULL) {
/*
* If the page is fully valid it can only become invalid
* with the object lock held. If it is not valid it can
* become valid with the busy lock held. Therefore, we
* may unnecessarily lock the exclusive busy here if we
* race with I/O completion not using the object lock.
* However, we will not end up with an invalid page and a
* shared lock.
*/
if (!vm_page_trybusy(m,
vm_page_all_valid(m) ? allocflags : 0)) {
(void)vm_page_grab_sleep(object, m, pindex, "pgrbwt",
allocflags, true);
goto retrylookup;
}
if (vm_page_all_valid(m))
goto out;
if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
vm_page_busy_release(m);
*mp = NULL;
return (VM_PAGER_FAIL);
}
} else if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
*mp = NULL;
return (VM_PAGER_FAIL);
} else if ((m = vm_page_alloc(object, pindex, pflags)) == NULL) {
if (!vm_pager_can_alloc_page(object, pindex))
return (VM_PAGER_AGAIN);
goto retrylookup;
}
vm_page_assert_xbusied(m);
if (vm_pager_has_page(object, pindex, NULL, &after)) {
after = MIN(after, VM_INITIAL_PAGEIN);
after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT);
after = MAX(after, 1);
ma[0] = m;
for (i = 1; i < after; i++) {
if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) {
if (vm_page_any_valid(ma[i]) ||
!vm_page_tryxbusy(ma[i]))
break;
} else {
ma[i] = vm_page_alloc(object, m->pindex + i,
VM_ALLOC_NORMAL);
if (ma[i] == NULL)
break;
}
}
after = i;
vm_object_pip_add(object, after);
VM_OBJECT_WUNLOCK(object);
rv = vm_pager_get_pages(object, ma, after, NULL, NULL);
VM_OBJECT_WLOCK(object);
vm_object_pip_wakeupn(object, after);
/* Pager may have replaced a page. */
m = ma[0];
if (rv != VM_PAGER_OK) {
for (i = 0; i < after; i++) {
if (!vm_page_wired(ma[i]))
vm_page_free(ma[i]);
else
vm_page_xunbusy(ma[i]);
}
*mp = NULL;
return (rv);
}
for (i = 1; i < after; i++)
vm_page_readahead_finish(ma[i]);
MPASS(vm_page_all_valid(m));
} else {
vm_page_zero_invalid(m, TRUE);
}
out:
if ((allocflags & VM_ALLOC_WIRED) != 0)
vm_page_wire(m);
if ((allocflags & VM_ALLOC_SBUSY) != 0 && vm_page_xbusied(m))
vm_page_busy_downgrade(m);
else if ((allocflags & VM_ALLOC_NOBUSY) != 0)
vm_page_busy_release(m);
*mp = m;
return (VM_PAGER_OK);
}
/*
* Locklessly grab a valid page. If the page is not valid or not yet
* allocated this will fall back to the object lock method.
*/
int
vm_page_grab_valid_unlocked(vm_page_t *mp, vm_object_t object,
vm_pindex_t pindex, int allocflags)
{
vm_page_t m;
int flags;
int error;
KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
(allocflags & VM_ALLOC_IGN_SBUSY) != 0,
("vm_page_grab_valid_unlocked: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY "
"mismatch"));
KASSERT((allocflags &
(VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0,
("vm_page_grab_valid_unlocked: Invalid flags 0x%X", allocflags));
/*
* Attempt a lockless lookup and busy. We need at least an sbusy
* before we can inspect the valid field and return a wired page.
*/
flags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_WIRED);
if (!vm_page_acquire_unlocked(object, pindex, NULL, mp, flags))
return (VM_PAGER_FAIL);
if ((m = *mp) != NULL) {
if (vm_page_all_valid(m)) {
if ((allocflags & VM_ALLOC_WIRED) != 0)
vm_page_wire(m);
vm_page_grab_release(m, allocflags);
return (VM_PAGER_OK);
}
vm_page_busy_release(m);
}
if ((allocflags & VM_ALLOC_NOCREAT) != 0) {
*mp = NULL;
return (VM_PAGER_FAIL);
}
VM_OBJECT_WLOCK(object);
error = vm_page_grab_valid(mp, object, pindex, allocflags);
VM_OBJECT_WUNLOCK(object);
return (error);
}
/*
* 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;
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(count > 0,
("vm_page_grab_pages: invalid page count %d", count));
vm_page_grab_check(allocflags);
pflags = vm_page_grab_pflags(allocflags);
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) {
if (!vm_page_tryacquire(m, allocflags)) {
if (vm_page_grab_sleep(object, m, pindex + i,
"grbmaw", allocflags, true))
goto retrylookup;
break;
}
} else {
if ((allocflags & VM_ALLOC_NOCREAT) != 0)
break;
m = vm_page_alloc_after(object, pindex + i,
pflags | VM_ALLOC_COUNT(count - i), mpred);
if (m == NULL) {
if ((allocflags & (VM_ALLOC_NOWAIT |
VM_ALLOC_WAITFAIL)) != 0)
break;
goto retrylookup;
}
}
if (vm_page_none_valid(m) &&
(allocflags & VM_ALLOC_ZERO) != 0) {
if ((m->flags & PG_ZERO) == 0)
pmap_zero_page(m);
vm_page_valid(m);
}
vm_page_grab_release(m, allocflags);
ma[i] = mpred = m;
m = vm_page_next(m);
}
return (i);
}
/*
* Unlocked variant of vm_page_grab_pages(). This accepts the same flags
* and will fall back to the locked variant to handle allocation.
*/
int
vm_page_grab_pages_unlocked(vm_object_t object, vm_pindex_t pindex,
int allocflags, vm_page_t *ma, int count)
{
vm_page_t m, pred;
int flags;
int i;
KASSERT(count > 0,
("vm_page_grab_pages_unlocked: invalid page count %d", count));
vm_page_grab_check(allocflags);
/*
* Modify flags for lockless acquire to hold the page until we
* set it valid if necessary.
*/
flags = allocflags & ~VM_ALLOC_NOBUSY;
pred = NULL;
for (i = 0; i < count; i++, pindex++) {
if (!vm_page_acquire_unlocked(object, pindex, pred, &m, flags))
return (i);
if (m == NULL)
break;
if ((flags & VM_ALLOC_ZERO) != 0 && vm_page_none_valid(m)) {
if ((m->flags & PG_ZERO) == 0)
pmap_zero_page(m);
vm_page_valid(m);
}
/* m will still be wired or busy according to flags. */
vm_page_grab_release(m, allocflags);
pred = ma[i] = m;
}
if (i == count || (allocflags & VM_ALLOC_NOCREAT) != 0)
return (i);
count -= i;
VM_OBJECT_WLOCK(object);
i += vm_page_grab_pages(object, pindex, allocflags, &ma[i], count);
VM_OBJECT_WUNLOCK(object);
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));
}
void
vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set)
{
#if PAGE_SIZE == 32768
atomic_set_64((uint64_t *)bits, set);
#elif PAGE_SIZE == 16384
atomic_set_32((uint32_t *)bits, set);
#elif (PAGE_SIZE == 8192) && defined(atomic_set_16)
atomic_set_16((uint16_t *)bits, set);
#elif (PAGE_SIZE == 4096) && defined(atomic_set_8)
atomic_set_8((uint8_t *)bits, set);
#else /* PAGE_SIZE <= 8192 */
uintptr_t addr;
int shift;
addr = (uintptr_t)bits;
/*
* Use a trick to perform a 32-bit atomic on the
* containing aligned word, to not depend on the existence
* of atomic_{set, clear}_{8, 16}.
*/
shift = addr & (sizeof(uint32_t) - 1);
#if BYTE_ORDER == BIG_ENDIAN
shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
#else
shift *= NBBY;
#endif
addr &= ~(sizeof(uint32_t) - 1);
atomic_set_32((uint32_t *)addr, set << shift);
#endif /* PAGE_SIZE */
}
static inline void
vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear)
{
#if PAGE_SIZE == 32768
atomic_clear_64((uint64_t *)bits, clear);
#elif PAGE_SIZE == 16384
atomic_clear_32((uint32_t *)bits, clear);
#elif (PAGE_SIZE == 8192) && defined(atomic_clear_16)
atomic_clear_16((uint16_t *)bits, clear);
#elif (PAGE_SIZE == 4096) && defined(atomic_clear_8)
atomic_clear_8((uint8_t *)bits, clear);
#else /* PAGE_SIZE <= 8192 */
uintptr_t addr;
int shift;
addr = (uintptr_t)bits;
/*
* Use a trick to perform a 32-bit atomic on the
* containing aligned word, to not depend on the existence
* of atomic_{set, clear}_{8, 16}.
*/
shift = addr & (sizeof(uint32_t) - 1);
#if BYTE_ORDER == BIG_ENDIAN
shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
#else
shift *= NBBY;
#endif
addr &= ~(sizeof(uint32_t) - 1);
atomic_clear_32((uint32_t *)addr, clear << shift);
#endif /* PAGE_SIZE */
}
static inline vm_page_bits_t
vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits)
{
#if PAGE_SIZE == 32768
uint64_t old;
old = *bits;
while (atomic_fcmpset_64(bits, &old, newbits) == 0);
return (old);
#elif PAGE_SIZE == 16384
uint32_t old;
old = *bits;
while (atomic_fcmpset_32(bits, &old, newbits) == 0);
return (old);
#elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16)
uint16_t old;
old = *bits;
while (atomic_fcmpset_16(bits, &old, newbits) == 0);
return (old);
#elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8)
uint8_t old;
old = *bits;
while (atomic_fcmpset_8(bits, &old, newbits) == 0);
return (old);
#else /* PAGE_SIZE <= 4096*/
uintptr_t addr;
uint32_t old, new, mask;
int shift;
addr = (uintptr_t)bits;
/*
* Use a trick to perform a 32-bit atomic on the
* containing aligned word, to not depend on the existence
* of atomic_{set, swap, clear}_{8, 16}.
*/
shift = addr & (sizeof(uint32_t) - 1);
#if BYTE_ORDER == BIG_ENDIAN
shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY;
#else
shift *= NBBY;
#endif
addr &= ~(sizeof(uint32_t) - 1);
mask = VM_PAGE_BITS_ALL << shift;
old = *bits;
do {
new = old & ~mask;
new |= newbits << shift;
} while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0);
return (old >> shift);
#endif /* PAGE_SIZE */
}
/*
* 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_page_bits_t pagebits;
vm_page_assert_busied(m);
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.
*/
pagebits = vm_page_bits(base, size);
if (vm_page_xbusied(m))
m->valid |= pagebits;
else
vm_page_bits_set(m, &m->valid, pagebits);
}
/*
* Set the page dirty bits and free the invalid swap space if
* present. Returns the previous dirty bits.
*/
vm_page_bits_t
vm_page_set_dirty(vm_page_t m)
{
vm_page_bits_t old;
VM_PAGE_OBJECT_BUSY_ASSERT(m);
if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) {
old = m->dirty;
m->dirty = VM_PAGE_BITS_ALL;
} else
old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL);
if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0)
vm_pager_page_unswapped(m);
return (old);
}
/*
* 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)
{
vm_page_assert_busied(m);
/*
* If the page is xbusied and not write mapped we are the
* only thread that can modify dirty bits. Otherwise, The pmap
* layer can call vm_page_dirty() without holding a distinguished
* lock. The combination of page busy and atomic operations
* suffice to guarantee consistency of the page dirty field.
*/
if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
m->dirty &= ~pagebits;
else
vm_page_bits_clear(m, &m->dirty, pagebits);
}
/*
* 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_page_assert_busied(m);
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 PGA_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);
if (vm_page_xbusied(m))
m->valid |= pagebits;
else
vm_page_bits_set(m, &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;
vm_page_aflag_clear(m, PGA_NOSYNC);
} else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m))
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;
/*
* The object lock is required so that pages can't be mapped
* read-only while we're in the process of invalidating them.
*/
object = m->object;
VM_OBJECT_ASSERT_WLOCKED(object);
vm_page_assert_busied(m);
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 && vm_page_all_valid(m) && bits != 0)
pmap_remove_all(m);
KASSERT((bits == 0 && vm_page_all_valid(m)) ||
!pmap_page_is_mapped(m),
("vm_page_set_invalid: page %p is mapped", m));
if (vm_page_xbusied(m)) {
m->valid &= ~bits;
m->dirty &= ~bits;
} else {
vm_page_bits_clear(m, &m->valid, bits);
vm_page_bits_clear(m, &m->dirty, bits);
}
}
/*
* vm_page_invalid:
*
* Invalidates the entire page. The page must be busy, unmapped, and
* the enclosing object must be locked. The object locks protects
* against concurrent read-only pmap enter which is done without
* busy.
*/
void
vm_page_invalid(vm_page_t m)
{
vm_page_assert_busied(m);
VM_OBJECT_ASSERT_WLOCKED(m->object);
MPASS(!pmap_page_is_mapped(m));
if (vm_page_xbusied(m))
m->valid = 0;
else
vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL);
}
/*
* 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;
/*
* 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 consistency
* issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
*/
if (setvalid)
vm_page_valid(m);
}
/*
* 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.
*
* Some callers envoke this routine without the busy lock held and
* handle races via higher level locks. Typical callers should
* hold a busy lock to prevent invalidation.
*/
int
vm_page_is_valid(vm_page_t m, int base, int size)
{
vm_page_bits_t bits;
bits = vm_page_bits(base, size);
return (vm_page_any_valid(m) && (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_page_assert_busied(m);
if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
vm_page_dirty(m);
}
void
vm_page_valid(vm_page_t m)
{
vm_page_assert_busied(m);
if (vm_page_xbusied(m))
m->valid = VM_PAGE_BITS_ALL;
else
vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL);
}
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_busy_assert(vm_page_t m)
{
/*
* Certain of the page's fields may only be modified by the
* holder of a page or object busy.
*/
if (m->object != NULL && !vm_page_busied(m))
VM_OBJECT_ASSERT_BUSY(m->object);
}
void
vm_page_assert_pga_writeable(vm_page_t m, uint16_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_BUSY(m->object);
}
#endif
#include "opt_ddb.h"
#ifdef DDB
#include <sys/kernel.h>
#include <ddb/ddb.h>
DB_SHOW_COMMAND_FLAGS(page, vm_page_print_page_info, DB_CMD_MEMSAFE)
{
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_FLAGS(pageq, vm_page_print_pageq_info, DB_CMD_MEMSAFE)
{
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, virt;
if (!have_addr) {
db_printf("show pginfo addr\n");
return;
}
phys = strchr(modif, 'p') != NULL;
virt = strchr(modif, 'v') != NULL;
if (virt)
m = PHYS_TO_VM_PAGE(pmap_kextract(addr));
else 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 ref 0x%x\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->a.queue, m->ref_count, m->a.flags, m->oflags,
m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty);
}
#endif /* DDB */
diff --git a/sys/vm/vm_pageout.c b/sys/vm/vm_pageout.c
index bb12a7e335d5..2945b53835c6 100644
--- a/sys/vm/vm_pageout.c
+++ b/sys/vm/vm_pageout.c
@@ -1,2419 +1,2419 @@
/*-
* 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 <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include "opt_vm.h"
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/blockcount.h>
#include <sys/eventhandler.h>
#include <sys/lock.h>
#include <sys/mutex.h>
#include <sys/proc.h>
#include <sys/kthread.h>
#include <sys/ktr.h>
#include <sys/mount.h>
#include <sys/racct.h>
#include <sys/resourcevar.h>
#include <sys/sched.h>
#include <sys/sdt.h>
#include <sys/signalvar.h>
#include <sys/smp.h>
#include <sys/time.h>
#include <sys/vnode.h>
#include <sys/vmmeter.h>
#include <sys/rwlock.h>
#include <sys/sx.h>
#include <sys/sysctl.h>
#include <vm/vm.h>
#include <vm/vm_param.h>
#include <vm/vm_object.h>
#include <vm/vm_page.h>
#include <vm/vm_map.h>
#include <vm/vm_pageout.h>
#include <vm/vm_pager.h>
#include <vm/vm_phys.h>
#include <vm/vm_pagequeue.h>
#include <vm/swap_pager.h>
#include <vm/vm_extern.h>
#include <vm/uma.h>
/*
* 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 swapdev_enabled;
int vm_pageout_page_count = 32;
static int vm_panic_on_oom = 0;
SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
"Panic on the given number of out-of-memory errors instead of "
"killing the largest process");
static int vm_pageout_update_period;
SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
CTLFLAG_RWTUN, &vm_pageout_update_period, 0,
"Maximum active LRU update period");
static int pageout_cpus_per_thread = 16;
SYSCTL_INT(_vm, OID_AUTO, pageout_cpus_per_thread, CTLFLAG_RDTUN,
&pageout_cpus_per_thread, 0,
"Number of CPUs per pagedaemon worker thread");
static int lowmem_period = 10;
SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0,
"Low memory callback period");
static int disable_swap_pageouts;
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");
static int vm_pageout_oom_seq = 12;
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");
u_long vm_page_max_user_wired;
SYSCTL_ULONG(_vm, OID_AUTO, max_user_wired, CTLFLAG_RW,
&vm_page_max_user_wired, 0,
"system-wide limit to user-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->a.flags & 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->a.flags & 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_pageout_defer() must be used to
* determine whether the page has been logically dequeued since the batch was
* collected.
*/
static __always_inline void
vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue)
{
struct vm_pagequeue *pq;
vm_page_t m, marker, n;
marker = ss->marker;
pq = ss->pq;
KASSERT((marker->a.flags & 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 = n, ss->scanned++) {
n = TAILQ_NEXT(m, plinks.q);
if ((m->flags & PG_MARKER) == 0) {
KASSERT((m->a.flags & 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));
}
/*
* Determine whether processing of a page should be deferred and ensure that any
* outstanding queue operations are processed.
*/
static __always_inline bool
vm_pageout_defer(vm_page_t m, const uint8_t queue, const bool enqueued)
{
vm_page_astate_t as;
as = vm_page_astate_load(m);
if (__predict_false(as.queue != queue ||
((as.flags & PGA_ENQUEUED) != 0) != enqueued))
return (true);
if ((as.flags & PGA_QUEUE_OP_MASK) != 0) {
vm_page_pqbatch_submit(m, queue);
return (true);
}
return (false);
}
/*
* 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;
object = m->object;
VM_OBJECT_ASSERT_WLOCKED(object);
pindex = m->pindex;
vm_page_assert_xbusied(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_tryxbusy(p) == 0) {
ib = 0;
break;
}
if (vm_page_wired(p)) {
ib = 0;
vm_page_xunbusy(p);
break;
}
vm_page_test_dirty(p);
if (p->dirty == 0) {
ib = 0;
vm_page_xunbusy(p);
break;
}
if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
vm_page_xunbusy(p);
ib = 0;
break;
}
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_tryxbusy(p) == 0)
break;
if (vm_page_wired(p)) {
vm_page_xunbusy(p);
break;
}
vm_page_test_dirty(p);
if (p->dirty == 0) {
vm_page_xunbusy(p);
break;
}
if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
vm_page_xunbusy(p);
break;
}
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 shared 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(vm_page_all_valid(mc[i]),
("vm_pageout_flush: partially invalid page %p index %d/%d",
mc[i], i, count));
KASSERT((mc[i]->a.flags & PGA_WRITEABLE) == 0,
("vm_pageout_flush: writeable page %p", mc[i]));
vm_page_busy_downgrade(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:
/*
* The page may have moved since laundering started, in
* which case it should be left alone.
*/
if (vm_page_in_laundry(mt))
vm_page_deactivate_noreuse(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);
if (vm_page_in_laundry(mt))
vm_page_deactivate_noreuse(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
* because 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.)
*/
if ((object->flags & OBJ_SWAP) != 0 &&
pageout_status[i] == VM_PAGER_FAIL) {
vm_page_unswappable(mt);
numpagedout++;
} else
vm_page_activate(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;
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_xunbusy(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);
if (vget(vp, vn_lktype_write(NULL, vp) | LK_TIMELOCK) != 0) {
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;
}
/*
* While the object was 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) {
error = ENXIO;
goto unlock_all;
}
/*
* The page may have been busied while the object lock was
* released.
*/
if (vm_page_tryxbusy(m) == 0) {
error = EBUSY;
goto unlock_all;
}
}
/*
* Remove all writeable mappings, failing if the page is wired.
*/
if (!vm_page_try_remove_write(m)) {
vm_page_xunbusy(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:
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;
vm_object_t object;
vm_page_t m, marker;
vm_page_astate_t new, old;
int act_delta, error, numpagedout, queue, refs, starting_target;
int vnodes_skipped;
bool pageout_ok;
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;
/*
* Don't touch a page that was removed from the queue after the
* page queue lock was released. Otherwise, ensure that any
* pending queue operations, such as dequeues for wired pages,
* are handled.
*/
if (vm_pageout_defer(m, queue, true))
continue;
/*
* Lock the page's object.
*/
if (object == NULL || object != m->object) {
if (object != NULL)
VM_OBJECT_WUNLOCK(object);
object = atomic_load_ptr(&m->object);
if (__predict_false(object == NULL))
/* The page is being freed by another thread. */
continue;
/* Depends on type-stability. */
VM_OBJECT_WLOCK(object);
if (__predict_false(m->object != object)) {
VM_OBJECT_WUNLOCK(object);
object = NULL;
continue;
}
}
if (vm_page_tryxbusy(m) == 0)
continue;
/*
* Check for wirings now that we hold the object lock and have
* exclusively busied the page. If the page is mapped, it may
* still be wired by pmap lookups. The call to
* vm_page_try_remove_all() below atomically checks for such
* wirings and removes mappings. If the page is unmapped, the
* wire count is guaranteed not to increase after this check.
*/
if (__predict_false(vm_page_wired(m)))
goto skip_page;
/*
* Invalid pages can be easily freed. They cannot be
* mapped; vm_page_free() asserts this.
*/
if (vm_page_none_valid(m))
goto free_page;
refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
for (old = vm_page_astate_load(m);;) {
/*
* Check to see if the page has been removed from the
* queue since the first such check. Leave it alone if
* so, discarding any references collected by
* pmap_ts_referenced().
*/
if (__predict_false(_vm_page_queue(old) == PQ_NONE))
goto skip_page;
new = old;
act_delta = refs;
if ((old.flags & PGA_REFERENCED) != 0) {
new.flags &= ~PGA_REFERENCED;
act_delta++;
}
if (act_delta == 0) {
;
} else if (object->ref_count != 0) {
/*
* 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 laundry queue.
*/
new.act_count += ACT_ADVANCE +
act_delta;
if (new.act_count > ACT_MAX)
new.act_count = ACT_MAX;
new.flags &= ~PGA_QUEUE_OP_MASK;
new.flags |= PGA_REQUEUE;
new.queue = PQ_ACTIVE;
if (!vm_page_pqstate_commit(m, &old, new))
continue;
/*
* 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--;
VM_CNT_INC(v_reactivated);
goto skip_page;
} else if ((object->flags & OBJ_DEAD) == 0) {
new.flags |= PGA_REQUEUE;
if (!vm_page_pqstate_commit(m, &old, new))
continue;
goto skip_page;
}
break;
}
/*
* 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 && !vm_page_try_remove_all(m))
goto skip_page;
}
/*
* 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:
/*
* Now we are guaranteed that no other threads are
* manipulating the page, check for a last-second
* reference.
*/
if (vm_pageout_defer(m, queue, true))
goto skip_page;
vm_page_free(m);
VM_CNT_INC(v_dfree);
} else if ((object->flags & OBJ_DEAD) == 0) {
if ((object->flags & OBJ_SWAP) != 0)
pageout_ok = disable_swap_pageouts == 0;
else
pageout_ok = true;
if (!pageout_ok) {
vm_page_launder(m);
goto skip_page;
}
/*
* 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++;
}
object = NULL;
} else {
skip_page:
vm_page_xunbusy(m);
}
}
if (object != NULL) {
VM_OBJECT_WUNLOCK(object);
object = NULL;
}
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 = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0;
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;
last_target = 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;
vm_object_t object;
vm_page_t m, marker;
struct vm_pagequeue *pq;
vm_page_astate_t old, new;
long min_scan;
int act_delta, max_scan, ps_delta, refs, scan_tick;
uint8_t nqueue;
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;
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;
/*
* Don't touch a page that was removed from the queue after the
* page queue lock was released. Otherwise, ensure that any
* pending queue operations, such as dequeues for wired pages,
* are handled.
*/
if (vm_pageout_defer(m, PQ_ACTIVE, true))
continue;
/*
* A page's object pointer may be set to NULL before
* the object lock is acquired.
*/
object = atomic_load_ptr(&m->object);
if (__predict_false(object == NULL))
/*
* The page has been removed from its object.
*/
continue;
/* Deferred free of swap space. */
if ((m->a.flags & PGA_SWAP_FREE) != 0 &&
VM_OBJECT_TRYWLOCK(object)) {
if (m->object == object)
vm_pager_page_unswapped(m);
VM_OBJECT_WUNLOCK(object);
}
/*
* 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.
*/
refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
old = vm_page_astate_load(m);
do {
/*
* Check to see if the page has been removed from the
* queue since the first such check. Leave it alone if
* so, discarding any references collected by
* pmap_ts_referenced().
*/
if (__predict_false(_vm_page_queue(old) == PQ_NONE)) {
ps_delta = 0;
break;
}
/*
* Advance or decay the act_count based on recent usage.
*/
new = old;
act_delta = refs;
if ((old.flags & PGA_REFERENCED) != 0) {
new.flags &= ~PGA_REFERENCED;
act_delta++;
}
if (act_delta != 0) {
new.act_count += ACT_ADVANCE + act_delta;
if (new.act_count > ACT_MAX)
new.act_count = ACT_MAX;
} else {
new.act_count -= min(new.act_count,
ACT_DECLINE);
}
if (new.act_count > 0) {
/*
* Adjust the activation count and keep the page
* in the active queue. The count might be left
* unchanged if it is saturated. The page may
* have been moved to a different queue since we
* started the scan, in which case we move it
* back.
*/
ps_delta = 0;
if (old.queue != PQ_ACTIVE) {
new.flags &= ~PGA_QUEUE_OP_MASK;
new.flags |= PGA_REQUEUE;
new.queue = PQ_ACTIVE;
}
} else {
/*
* When not short for inactive pages, let dirty
* pages go through the inactive queue before
* moving to the laundry queue. 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.
*
* 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 (page_shortage <= 0) {
nqueue = PQ_INACTIVE;
ps_delta = 0;
} else if (m->dirty == 0) {
nqueue = PQ_INACTIVE;
ps_delta = act_scan_laundry_weight;
} else {
nqueue = PQ_LAUNDRY;
ps_delta = 1;
}
new.flags &= ~PGA_QUEUE_OP_MASK;
new.flags |= PGA_REQUEUE;
new.queue = nqueue;
}
} while (!vm_page_pqstate_commit(m, &old, new));
page_shortage -= ps_delta;
}
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 vm_pagequeue *pq, vm_page_t marker,
vm_page_t m)
{
vm_page_astate_t as;
vm_pagequeue_assert_locked(pq);
as = vm_page_astate_load(m);
if (as.queue != PQ_INACTIVE || (as.flags & PGA_ENQUEUED) != 0)
return (0);
vm_page_aflag_set(m, PGA_ENQUEUED);
TAILQ_INSERT_BEFORE(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;
vm_page_t marker;
int delta;
delta = 0;
marker = ss->marker;
pq = ss->pq;
if (m != NULL) {
- if (vm_batchqueue_insert(bq, m))
+ if (vm_batchqueue_insert(bq, m) != 0)
return;
vm_pagequeue_lock(pq);
delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
} else
vm_pagequeue_lock(pq);
while ((m = vm_batchqueue_pop(bq)) != NULL)
delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
vm_pagequeue_cnt_add(pq, delta);
vm_pagequeue_unlock(pq);
vm_batchqueue_init(bq);
}
static void
vm_pageout_scan_inactive(struct vm_domain *vmd, int page_shortage)
{
struct timeval start, end;
struct scan_state ss;
struct vm_batchqueue rq;
struct vm_page marker_page;
vm_page_t m, marker;
struct vm_pagequeue *pq;
vm_object_t object;
vm_page_astate_t old, new;
int act_delta, addl_page_shortage, starting_page_shortage, refs;
object = NULL;
vm_batchqueue_init(&rq);
getmicrouptime(&start);
/*
* 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;
/*
* 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->a.act_count is not used to make
* decisions for the inactive queue, only for the active queue.)
*/
starting_page_shortage = page_shortage;
marker = &marker_page;
vm_page_init_marker(marker, PQ_INACTIVE, 0);
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));
/*
* Don't touch a page that was removed from the queue after the
* page queue lock was released. Otherwise, ensure that any
* pending queue operations, such as dequeues for wired pages,
* are handled.
*/
if (vm_pageout_defer(m, PQ_INACTIVE, false))
continue;
/*
* Lock the page's object.
*/
if (object == NULL || object != m->object) {
if (object != NULL)
VM_OBJECT_WUNLOCK(object);
object = atomic_load_ptr(&m->object);
if (__predict_false(object == NULL))
/* The page is being freed by another thread. */
continue;
/* Depends on type-stability. */
VM_OBJECT_WLOCK(object);
if (__predict_false(m->object != object)) {
VM_OBJECT_WUNLOCK(object);
object = NULL;
goto reinsert;
}
}
if (vm_page_tryxbusy(m) == 0) {
/*
* 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;
}
/* Deferred free of swap space. */
if ((m->a.flags & PGA_SWAP_FREE) != 0)
vm_pager_page_unswapped(m);
/*
* Check for wirings now that we hold the object lock and have
* exclusively busied the page. If the page is mapped, it may
* still be wired by pmap lookups. The call to
* vm_page_try_remove_all() below atomically checks for such
* wirings and removes mappings. If the page is unmapped, the
* wire count is guaranteed not to increase after this check.
*/
if (__predict_false(vm_page_wired(m)))
goto skip_page;
/*
* Invalid pages can be easily freed. They cannot be
* mapped, vm_page_free() asserts this.
*/
if (vm_page_none_valid(m))
goto free_page;
refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
for (old = vm_page_astate_load(m);;) {
/*
* Check to see if the page has been removed from the
* queue since the first such check. Leave it alone if
* so, discarding any references collected by
* pmap_ts_referenced().
*/
if (__predict_false(_vm_page_queue(old) == PQ_NONE))
goto skip_page;
new = old;
act_delta = refs;
if ((old.flags & PGA_REFERENCED) != 0) {
new.flags &= ~PGA_REFERENCED;
act_delta++;
}
if (act_delta == 0) {
;
} else if (object->ref_count != 0) {
/*
* 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.
*/
new.act_count += ACT_ADVANCE +
act_delta;
if (new.act_count > ACT_MAX)
new.act_count = ACT_MAX;
new.flags &= ~PGA_QUEUE_OP_MASK;
new.flags |= PGA_REQUEUE;
new.queue = PQ_ACTIVE;
if (!vm_page_pqstate_commit(m, &old, new))
continue;
VM_CNT_INC(v_reactivated);
goto skip_page;
} else if ((object->flags & OBJ_DEAD) == 0) {
new.queue = PQ_INACTIVE;
new.flags |= PGA_REQUEUE;
if (!vm_page_pqstate_commit(m, &old, new))
continue;
goto skip_page;
}
break;
}
/*
* 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 && !vm_page_try_remove_all(m))
goto skip_page;
}
/*
* 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:
/*
* Now we are guaranteed that no other threads are
* manipulating the page, check for a last-second
* reference that would save it from doom.
*/
if (vm_pageout_defer(m, PQ_INACTIVE, false))
goto skip_page;
/*
* Because we dequeued the page and have already checked
* for pending dequeue and enqueue requests, we can
* safely disassociate the page from the inactive queue
* without holding the queue lock.
*/
m->a.queue = PQ_NONE;
vm_page_free(m);
page_shortage--;
continue;
}
if ((object->flags & OBJ_DEAD) == 0)
vm_page_launder(m);
skip_page:
vm_page_xunbusy(m);
continue;
reinsert:
vm_pageout_reinsert_inactive(&ss, &rq, m);
}
if (object != NULL)
VM_OBJECT_WUNLOCK(object);
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);
/*
* Record the remaining shortage and the progress and rate it was made.
*/
atomic_add_int(&vmd->vmd_addl_shortage, addl_page_shortage);
getmicrouptime(&end);
timevalsub(&end, &start);
atomic_add_int(&vmd->vmd_inactive_us,
end.tv_sec * 1000000 + end.tv_usec);
atomic_add_int(&vmd->vmd_inactive_freed,
starting_page_shortage - page_shortage);
}
/*
* Dispatch a number of inactive threads according to load and collect the
* results to present a coherent view of paging activity on this domain.
*/
static int
vm_pageout_inactive_dispatch(struct vm_domain *vmd, int shortage)
{
u_int freed, pps, slop, threads, us;
vmd->vmd_inactive_shortage = shortage;
slop = 0;
/*
* If we have more work than we can do in a quarter of our interval, we
* fire off multiple threads to process it.
*/
threads = vmd->vmd_inactive_threads;
if (threads > 1 && vmd->vmd_inactive_pps != 0 &&
shortage > vmd->vmd_inactive_pps / VM_INACT_SCAN_RATE / 4) {
vmd->vmd_inactive_shortage /= threads;
slop = shortage % threads;
vm_domain_pageout_lock(vmd);
blockcount_acquire(&vmd->vmd_inactive_starting, threads - 1);
blockcount_acquire(&vmd->vmd_inactive_running, threads - 1);
wakeup(&vmd->vmd_inactive_shortage);
vm_domain_pageout_unlock(vmd);
}
/* Run the local thread scan. */
vm_pageout_scan_inactive(vmd, vmd->vmd_inactive_shortage + slop);
/*
* Block until helper threads report results and then accumulate
* totals.
*/
blockcount_wait(&vmd->vmd_inactive_running, NULL, "vmpoid", PVM);
freed = atomic_readandclear_int(&vmd->vmd_inactive_freed);
VM_CNT_ADD(v_dfree, freed);
/*
* Calculate the per-thread paging rate with an exponential decay of
* prior results. Careful to avoid integer rounding errors with large
* us values.
*/
us = max(atomic_readandclear_int(&vmd->vmd_inactive_us), 1);
if (us > 1000000)
/* Keep rounding to tenths */
pps = (freed * 10) / ((us * 10) / 1000000);
else
pps = (1000000 / us) * freed;
vmd->vmd_inactive_pps = (vmd->vmd_inactive_pps / 2) + (pps / 2);
return (shortage - freed);
}
/*
* Attempt to reclaim the requested number of pages from the inactive queue.
* Returns true if the shortage was addressed.
*/
static int
vm_pageout_inactive(struct vm_domain *vmd, int shortage, int *addl_shortage)
{
struct vm_pagequeue *pq;
u_int addl_page_shortage, deficit, page_shortage;
u_int starting_page_shortage;
/*
* 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 = shortage + deficit;
/*
* Run the inactive scan on as many threads as is necessary.
*/
page_shortage = vm_pageout_inactive_dispatch(vmd, starting_page_shortage);
addl_page_shortage = atomic_readandclear_int(&vmd->vmd_addl_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;
VM_MAP_ENTRY_FOREACH(entry, map) {
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;
if (obj->type == OBJT_PHYS || obj->type == OBJT_VNODE ||
(obj->flags & OBJ_SWAP) != 0)
res += obj->resident_page_count;
}
return (res);
}
static int vm_oom_ratelim_last;
static int vm_oom_pf_secs = 10;
SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0,
"");
static struct mtx vm_oom_ratelim_mtx;
void
vm_pageout_oom(int shortage)
{
const char *reason;
struct proc *p, *bigproc;
vm_offset_t size, bigsize;
struct thread *td;
struct vmspace *vm;
int now;
bool breakout;
/*
* For OOM requests originating from vm_fault(), there is a high
* chance that a single large process faults simultaneously in
* several threads. Also, on an active system running many
* processes of middle-size, like buildworld, all of them
* could fault almost simultaneously as well.
*
* To avoid killing too many processes, rate-limit OOMs
* initiated by vm_fault() time-outs on the waits for free
* pages.
*/
mtx_lock(&vm_oom_ratelim_mtx);
now = ticks;
if (shortage == VM_OOM_MEM_PF &&
(u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) {
mtx_unlock(&vm_oom_ratelim_mtx);
return;
}
vm_oom_ratelim_last = now;
mtx_unlock(&vm_oom_ratelim_mtx);
/*
* 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 || shortage == VM_OOM_MEM_PF)
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) {
switch (shortage) {
case VM_OOM_MEM:
reason = "failed to reclaim memory";
break;
case VM_OOM_MEM_PF:
reason = "a thread waited too long to allocate a page";
break;
case VM_OOM_SWAPZ:
reason = "out of swap space";
break;
default:
panic("unknown OOM reason %d", shortage);
}
if (vm_panic_on_oom != 0 && --vm_panic_on_oom == 0)
panic("%s", reason);
PROC_LOCK(bigproc);
killproc(bigproc, reason);
sched_nice(bigproc, PRIO_MIN);
_PRELE(bigproc);
PROC_UNLOCK(bigproc);
}
}
/*
* Signal a free page shortage to subsystems that have registered an event
* handler. Reclaim memory from UMA in the event of a severe shortage.
* Return true if the free page count should be re-evaluated.
*/
static bool
vm_pageout_lowmem(void)
{
static int lowmem_ticks = 0;
int last;
bool ret;
ret = false;
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(UMA_RECLAIM_TRIM);
ret = true;
break;
}
/*
* Kick off an asynchronous reclaim of cached memory if one of the
* page daemons is failing to keep up with demand. Use the "severe"
* threshold instead of "min" to ensure that we do not blow away the
* caches if a subset of the NUMA domains are depleted by kernel memory
* allocations; the domainset iterators automatically skip domains
* below the "min" threshold on the first pass.
*
* UMA reclaim worker has its own rate-limiting mechanism, so don't
* worry about kicking it too often.
*/
if (vm_page_count_severe())
uma_reclaim_wakeup();
return (ret);
}
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_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_helper runs additional pageout daemons in times of high paging
* activity.
*/
static void
vm_pageout_helper(void *arg)
{
struct vm_domain *vmd;
int domain;
domain = (uintptr_t)arg;
vmd = VM_DOMAIN(domain);
vm_domain_pageout_lock(vmd);
for (;;) {
msleep(&vmd->vmd_inactive_shortage,
vm_domain_pageout_lockptr(vmd), PVM, "psleep", 0);
blockcount_release(&vmd->vmd_inactive_starting, 1);
vm_domain_pageout_unlock(vmd);
vm_pageout_scan_inactive(vmd, vmd->vmd_inactive_shortage);
vm_domain_pageout_lock(vmd);
/*
* Release the running count while the pageout lock is held to
* prevent wakeup races.
*/
blockcount_release(&vmd->vmd_inactive_running, 1);
}
}
static int
get_pageout_threads_per_domain(const struct vm_domain *vmd)
{
unsigned total_pageout_threads, eligible_cpus, domain_cpus;
if (VM_DOMAIN_EMPTY(vmd->vmd_domain))
return (0);
/*
* Semi-arbitrarily constrain pagedaemon threads to less than half the
* total number of CPUs in the system as an upper limit.
*/
if (pageout_cpus_per_thread < 2)
pageout_cpus_per_thread = 2;
else if (pageout_cpus_per_thread > mp_ncpus)
pageout_cpus_per_thread = mp_ncpus;
total_pageout_threads = howmany(mp_ncpus, pageout_cpus_per_thread);
domain_cpus = CPU_COUNT(&cpuset_domain[vmd->vmd_domain]);
/* Pagedaemons are not run in empty domains. */
eligible_cpus = mp_ncpus;
for (unsigned i = 0; i < vm_ndomains; i++)
if (VM_DOMAIN_EMPTY(i))
eligible_cpus -= CPU_COUNT(&cpuset_domain[i]);
/*
* Assign a portion of the total pageout threads to this domain
* corresponding to the fraction of pagedaemon-eligible CPUs in the
* domain. In asymmetric NUMA systems, domains with more CPUs may be
* allocated more threads than domains with fewer CPUs.
*/
return (howmany(total_pageout_threads * domain_cpus, eligible_cpus));
}
/*
* Initialize basic pageout daemon settings. See the comment above the
* definition of vm_domain for some explanation of how these thresholds are
* used.
*/
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.
*/
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_min = vmd->vmd_page_count / 200;
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 | CTLFLAG_MPSAFE, NULL, "");
pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
vmd->vmd_inactive_threads = get_pageout_threads_per_domain(vmd);
}
static void
vm_pageout_init(void)
{
u_long 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;
/*
* Set the maximum number of user-wired virtual pages. Historically the
* main source of such pages was mlock(2) and mlockall(2). Hypervisors
* may also request user-wired memory.
*/
if (vm_page_max_user_wired == 0)
vm_page_max_user_wired = 4 * freecount / 5;
}
/*
* vm_pageout is the high level pageout daemon.
*/
static void
vm_pageout(void)
{
struct proc *p;
struct thread *td;
int error, first, i, j, pageout_threads;
p = curproc;
td = curthread;
mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF);
swap_pager_swap_init();
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;
}
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);
}
pageout_threads = VM_DOMAIN(i)->vmd_inactive_threads;
for (j = 0; j < pageout_threads - 1; j++) {
error = kthread_add(vm_pageout_helper,
(void *)(uintptr_t)i, p, NULL, 0, 0,
"dom%d helper%d", i, j);
if (error != 0)
panic("starting pageout helper %d for domain "
"%d: %d\n", j, i, error);
}
error = kthread_add(vm_pageout_laundry_worker,
(void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i);
if (error != 0)
panic("starting laundry for domain %d: %d", i, error);
}
error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma");
if (error != 0)
panic("starting uma_reclaim helper, error %d\n", error);
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);
}
}
diff --git a/sys/vm/vm_pagequeue.h b/sys/vm/vm_pagequeue.h
index a9d4c920e5be..268d53a391db 100644
--- a/sys/vm/vm_pagequeue.h
+++ b/sys/vm/vm_pagequeue.h
@@ -1,468 +1,470 @@
/*-
* 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_page.h 8.2 (Berkeley) 12/13/93
*
*
* 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.
*
* $FreeBSD$
*/
#ifndef _VM_PAGEQUEUE_
#define _VM_PAGEQUEUE_
#ifdef _KERNEL
struct vm_pagequeue {
struct mtx pq_mutex;
struct pglist pq_pl;
int pq_cnt;
const char * const pq_name;
uint64_t pq_pdpages;
} __aligned(CACHE_LINE_SIZE);
#ifndef VM_BATCHQUEUE_SIZE
-#define VM_BATCHQUEUE_SIZE 7
+#define VM_BATCHQUEUE_SIZE 15
#endif
struct vm_batchqueue {
vm_page_t bq_pa[VM_BATCHQUEUE_SIZE];
int bq_cnt;
} __aligned(CACHE_LINE_SIZE);
#include <vm/uma.h>
#include <sys/_blockcount.h>
#include <sys/pidctrl.h>
struct sysctl_oid;
/*
* One vm_domain per NUMA domain. Contains pagequeues, free page structures,
* and accounting.
*
* Lock Key:
* f vmd_free_mtx
* p vmd_pageout_mtx
* d vm_domainset_lock
* a atomic
* c const after boot
* q page queue lock
*
* A unique page daemon thread manages each vm_domain structure and is
* responsible for ensuring that some free memory is available by freeing
* inactive pages and aging active pages. To decide how many pages to process,
* it uses thresholds derived from the number of pages in the domain:
*
* vmd_page_count
* ---
* |
* |-> vmd_inactive_target (~3%)
* | - The active queue scan target is given by
* | (vmd_inactive_target + vmd_free_target - vmd_free_count).
* |
* |
* |-> vmd_free_target (~2%)
* | - Target for page reclamation.
* |
* |-> vmd_pageout_wakeup_thresh (~1.8%)
* | - Threshold for waking up the page daemon.
* |
* |
* |-> vmd_free_min (~0.5%)
* | - First low memory threshold.
* | - Causes per-CPU caching to be lazily disabled in UMA.
* | - vm_wait() sleeps below this threshold.
* |
* |-> vmd_free_severe (~0.25%)
* | - Second low memory threshold.
* | - Triggers aggressive UMA reclamation, disables delayed buffer
* | writes.
* |
* |-> vmd_free_reserved (~0.13%)
* | - Minimum for VM_ALLOC_NORMAL page allocations.
* |-> vmd_pageout_free_min (32 + 2 pages)
* | - Minimum for waking a page daemon thread sleeping in vm_wait().
* |-> vmd_interrupt_free_min (2 pages)
* | - Minimum for VM_ALLOC_SYSTEM page allocations.
* ---
*
*--
* Free page count regulation:
*
* The page daemon attempts to ensure that the free page count is above the free
* target. It wakes up periodically (every 100ms) to input the current free
* page shortage (free_target - free_count) to a PID controller, which in
* response outputs the number of pages to attempt to reclaim. The shortage's
* current magnitude, rate of change, and cumulative value are together used to
* determine the controller's output. The page daemon target thus adapts
* dynamically to the system's demand for free pages, resulting in less
* burstiness than a simple hysteresis loop.
*
* When the free page count drops below the wakeup threshold,
* vm_domain_allocate() proactively wakes up the page daemon. This helps ensure
* that the system responds promptly to a large instantaneous free page
* shortage.
*
* The page daemon also attempts to ensure that some fraction of the system's
* memory is present in the inactive (I) and laundry (L) page queues, so that it
* can respond promptly to a sudden free page shortage. In particular, the page
* daemon thread aggressively scans active pages so long as the following
* condition holds:
*
* len(I) + len(L) + free_target - free_count < inactive_target
*
* Otherwise, when the inactive target is met, the page daemon periodically
* scans a small portion of the active queue in order to maintain up-to-date
* per-page access history. Unreferenced pages in the active queue thus
* eventually migrate to the inactive queue.
*
* The per-domain laundry thread periodically launders dirty pages based on the
* number of clean pages freed by the page daemon since the last laundering. If
* the page daemon fails to meet its scan target (i.e., the PID controller
* output) because of a shortage of clean inactive pages, the laundry thread
* attempts to launder enough pages to meet the free page target.
*
*--
* Page allocation priorities:
*
* The system defines three page allocation priorities: VM_ALLOC_NORMAL,
* VM_ALLOC_SYSTEM and VM_ALLOC_INTERRUPT. An interrupt-priority allocation can
* claim any free page. This priority is used in the pmap layer when attempting
* to allocate a page for the kernel page tables; in such cases an allocation
* failure will usually result in a kernel panic. The system priority is used
* for most other kernel memory allocations, for instance by UMA's slab
* allocator or the buffer cache. Such allocations will fail if the free count
* is below interrupt_free_min. All other allocations occur at the normal
* priority, which is typically used for allocation of user pages, for instance
* in the page fault handler or when allocating page table pages or pv_entry
* structures for user pmaps. Such allocations fail if the free count is below
* the free_reserved threshold.
*
*--
* Free memory shortages:
*
* The system uses the free_min and free_severe thresholds to apply
* back-pressure and give the page daemon a chance to recover. When a page
* allocation fails due to a shortage and the allocating thread cannot handle
* failure, it may call vm_wait() to sleep until free pages are available.
* vm_domain_freecnt_inc() wakes sleeping threads once the free page count rises
* above the free_min threshold; the page daemon and laundry threads are given
* priority and will wake up once free_count reaches the (much smaller)
* pageout_free_min threshold.
*
* On NUMA systems, the domainset iterators always prefer NUMA domains where the
* free page count is above the free_min threshold. This means that given the
* choice between two NUMA domains, one above the free_min threshold and one
* below, the former will be used to satisfy the allocation request regardless
* of the domain selection policy.
*
* In addition to reclaiming memory from the page queues, the vm_lowmem event
* fires every ten seconds so long as the system is under memory pressure (i.e.,
* vmd_free_count < vmd_free_target). This allows kernel subsystems to register
* for notifications of free page shortages, upon which they may shrink their
* caches. Following a vm_lowmem event, UMA's caches are pruned to ensure that
* they do not contain an excess of unused memory. When a domain is below the
* free_min threshold, UMA limits the population of per-CPU caches. When a
* domain falls below the free_severe threshold, UMA's caches are completely
* drained.
*
* If the system encounters a global memory shortage, it may resort to the
* out-of-memory (OOM) killer, which selects a process and delivers SIGKILL in a
* last-ditch attempt to free up some pages. Either of the two following
* conditions will activate the OOM killer:
*
* 1. The page daemons collectively fail to reclaim any pages during their
* inactive queue scans. After vm_pageout_oom_seq consecutive scans fail,
* the page daemon thread votes for an OOM kill, and an OOM kill is
* triggered when all page daemons have voted. This heuristic is strict and
* may fail to trigger even when the system is effectively deadlocked.
*
* 2. Threads in the user fault handler are repeatedly unable to make progress
* while allocating a page to satisfy the fault. After
* vm_pfault_oom_attempts page allocation failures with intervening
* vm_wait() calls, the faulting thread will trigger an OOM kill.
*/
struct vm_domain {
struct vm_pagequeue vmd_pagequeues[PQ_COUNT];
struct mtx_padalign vmd_free_mtx;
struct mtx_padalign vmd_pageout_mtx;
struct vm_pgcache {
int domain;
int pool;
uma_zone_t zone;
} vmd_pgcache[VM_NFREEPOOL];
struct vmem *vmd_kernel_arena; /* (c) per-domain kva R/W arena. */
struct vmem *vmd_kernel_rwx_arena; /* (c) per-domain kva R/W/X arena. */
u_int vmd_domain; /* (c) Domain number. */
u_int vmd_page_count; /* (c) Total page count. */
long vmd_segs; /* (c) bitmask of the segments */
u_int __aligned(CACHE_LINE_SIZE) vmd_free_count; /* (a,f) free page count */
u_int vmd_pageout_deficit; /* (a) Estimated number of pages deficit */
uint8_t vmd_pad[CACHE_LINE_SIZE - (sizeof(u_int) * 2)];
/* Paging control variables, used within single threaded page daemon. */
struct pidctrl vmd_pid; /* Pageout controller. */
boolean_t vmd_oom;
u_int vmd_inactive_threads;
u_int vmd_inactive_shortage; /* Per-thread shortage. */
blockcount_t vmd_inactive_running; /* Number of inactive threads. */
blockcount_t vmd_inactive_starting; /* Number of threads started. */
volatile u_int vmd_addl_shortage; /* Shortage accumulator. */
volatile u_int vmd_inactive_freed; /* Successful inactive frees. */
volatile u_int vmd_inactive_us; /* Microseconds for above. */
u_int vmd_inactive_pps; /* Exponential decay frees/second. */
int vmd_oom_seq;
int vmd_last_active_scan;
struct vm_page vmd_markers[PQ_COUNT]; /* (q) markers for queue scans */
struct vm_page vmd_inacthead; /* marker for LRU-defeating insertions */
struct vm_page vmd_clock[2]; /* markers for active queue scan */
int vmd_pageout_wanted; /* (a, p) pageout daemon wait channel */
int vmd_pageout_pages_needed; /* (d) page daemon waiting for pages? */
bool vmd_minset; /* (d) Are we in vm_min_domains? */
bool vmd_severeset; /* (d) Are we in vm_severe_domains? */
enum {
VM_LAUNDRY_IDLE = 0,
VM_LAUNDRY_BACKGROUND,
VM_LAUNDRY_SHORTFALL
} vmd_laundry_request;
/* Paging thresholds and targets. */
u_int vmd_clean_pages_freed; /* (q) accumulator for laundry thread */
u_int vmd_background_launder_target; /* (c) */
u_int vmd_free_reserved; /* (c) pages reserved for deadlock */
u_int vmd_free_target; /* (c) pages desired free */
u_int vmd_free_min; /* (c) pages desired free */
u_int vmd_inactive_target; /* (c) pages desired inactive */
u_int vmd_pageout_free_min; /* (c) min pages reserved for kernel */
u_int vmd_pageout_wakeup_thresh;/* (c) min pages to wake pagedaemon */
u_int vmd_interrupt_free_min; /* (c) reserved pages for int code */
u_int vmd_free_severe; /* (c) severe page depletion point */
/* Name for sysctl etc. */
struct sysctl_oid *vmd_oid;
char vmd_name[sizeof(__XSTRING(MAXMEMDOM))];
} __aligned(CACHE_LINE_SIZE);
extern struct vm_domain vm_dom[MAXMEMDOM];
#define VM_DOMAIN(n) (&vm_dom[(n)])
#define VM_DOMAIN_EMPTY(n) (vm_dom[(n)].vmd_page_count == 0)
#define vm_pagequeue_assert_locked(pq) mtx_assert(&(pq)->pq_mutex, MA_OWNED)
#define vm_pagequeue_lock(pq) mtx_lock(&(pq)->pq_mutex)
#define vm_pagequeue_lockptr(pq) (&(pq)->pq_mutex)
#define vm_pagequeue_trylock(pq) mtx_trylock(&(pq)->pq_mutex)
#define vm_pagequeue_unlock(pq) mtx_unlock(&(pq)->pq_mutex)
#define vm_domain_free_assert_locked(n) \
mtx_assert(vm_domain_free_lockptr((n)), MA_OWNED)
#define vm_domain_free_assert_unlocked(n) \
mtx_assert(vm_domain_free_lockptr((n)), MA_NOTOWNED)
#define vm_domain_free_lock(d) \
mtx_lock(vm_domain_free_lockptr((d)))
#define vm_domain_free_lockptr(d) \
(&(d)->vmd_free_mtx)
#define vm_domain_free_trylock(d) \
mtx_trylock(vm_domain_free_lockptr((d)))
#define vm_domain_free_unlock(d) \
mtx_unlock(vm_domain_free_lockptr((d)))
#define vm_domain_pageout_lockptr(d) \
(&(d)->vmd_pageout_mtx)
#define vm_domain_pageout_assert_locked(n) \
mtx_assert(vm_domain_pageout_lockptr((n)), MA_OWNED)
#define vm_domain_pageout_assert_unlocked(n) \
mtx_assert(vm_domain_pageout_lockptr((n)), MA_NOTOWNED)
#define vm_domain_pageout_lock(d) \
mtx_lock(vm_domain_pageout_lockptr((d)))
#define vm_domain_pageout_unlock(d) \
mtx_unlock(vm_domain_pageout_lockptr((d)))
static __inline void
vm_pagequeue_cnt_add(struct vm_pagequeue *pq, int addend)
{
vm_pagequeue_assert_locked(pq);
pq->pq_cnt += addend;
}
#define vm_pagequeue_cnt_inc(pq) vm_pagequeue_cnt_add((pq), 1)
#define vm_pagequeue_cnt_dec(pq) vm_pagequeue_cnt_add((pq), -1)
static inline void
vm_pagequeue_remove(struct vm_pagequeue *pq, vm_page_t m)
{
TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
vm_pagequeue_cnt_dec(pq);
}
static inline void
vm_batchqueue_init(struct vm_batchqueue *bq)
{
bq->bq_cnt = 0;
}
-static inline bool
+static inline int
vm_batchqueue_insert(struct vm_batchqueue *bq, vm_page_t m)
{
+ int slots_free;
- if (bq->bq_cnt < nitems(bq->bq_pa)) {
+ slots_free = nitems(bq->bq_pa) - bq->bq_cnt;
+ if (slots_free > 0) {
bq->bq_pa[bq->bq_cnt++] = m;
- return (true);
+ return (slots_free);
}
- return (false);
+ return (slots_free);
}
static inline vm_page_t
vm_batchqueue_pop(struct vm_batchqueue *bq)
{
if (bq->bq_cnt == 0)
return (NULL);
return (bq->bq_pa[--bq->bq_cnt]);
}
void vm_domain_set(struct vm_domain *vmd);
void vm_domain_clear(struct vm_domain *vmd);
int vm_domain_allocate(struct vm_domain *vmd, int req, int npages);
/*
* vm_pagequeue_domain:
*
* Return the memory domain the page belongs to.
*/
static inline struct vm_domain *
vm_pagequeue_domain(vm_page_t m)
{
return (VM_DOMAIN(vm_page_domain(m)));
}
/*
* Return the number of pages we need to free-up or cache
* A positive number indicates that we do not have enough free pages.
*/
static inline int
vm_paging_target(struct vm_domain *vmd)
{
return (vmd->vmd_free_target - vmd->vmd_free_count);
}
/*
* Returns TRUE if the pagedaemon needs to be woken up.
*/
static inline int
vm_paging_needed(struct vm_domain *vmd, u_int free_count)
{
return (free_count < vmd->vmd_pageout_wakeup_thresh);
}
/*
* Returns TRUE if the domain is below the min paging target.
*/
static inline int
vm_paging_min(struct vm_domain *vmd)
{
return (vmd->vmd_free_min > vmd->vmd_free_count);
}
/*
* Returns TRUE if the domain is below the severe paging target.
*/
static inline int
vm_paging_severe(struct vm_domain *vmd)
{
return (vmd->vmd_free_severe > vmd->vmd_free_count);
}
/*
* Return the number of pages we need to launder.
* A positive number indicates that we have a shortfall of clean pages.
*/
static inline int
vm_laundry_target(struct vm_domain *vmd)
{
return (vm_paging_target(vmd));
}
void pagedaemon_wakeup(int domain);
static inline void
vm_domain_freecnt_inc(struct vm_domain *vmd, int adj)
{
u_int old, new;
old = atomic_fetchadd_int(&vmd->vmd_free_count, adj);
new = old + adj;
/*
* Only update bitsets on transitions. Notice we short-circuit the
* rest of the checks if we're above min already.
*/
if (old < vmd->vmd_free_min && (new >= vmd->vmd_free_min ||
(old < vmd->vmd_free_severe && new >= vmd->vmd_free_severe) ||
(old < vmd->vmd_pageout_free_min &&
new >= vmd->vmd_pageout_free_min)))
vm_domain_clear(vmd);
}
#endif /* _KERNEL */
#endif /* !_VM_PAGEQUEUE_ */