From 4565fa16977a4fac5f98cb75e9f5559d67c6446f Mon Sep 17 00:00:00 2001 From: alc Date: Sat, 29 Dec 2007 19:53:04 +0000 Subject: Add the superpage reservation system. This is "part 2 of 2" of the machine-independent support for superpages. (The earlier part was the rewrite of the physical memory allocator.) The remainder of the code required for superpages support is machine-dependent and will be added to the various pmap implementations at a later date. Initially, I am only supporting one large page size per architecture. Moreover, I am only enabling the reservation system on amd64. (In an emergency, it can be disabled by setting VM_NRESERVLEVELS to 0 in amd64/include/vmparam.h or your kernel configuration file.) --- sys/conf/files | 1 + sys/kern/kern_exec.c | 7 + sys/vm/vm_fault.c | 13 + sys/vm/vm_object.c | 45 ++++ sys/vm/vm_page.c | 53 +++- sys/vm/vm_reserv.c | 668 +++++++++++++++++++++++++++++++++++++++++++++++++++ sys/vm/vm_reserv.h | 59 +++++ 7 files changed, 842 insertions(+), 4 deletions(-) create mode 100644 sys/vm/vm_reserv.c create mode 100644 sys/vm/vm_reserv.h diff --git a/sys/conf/files b/sys/conf/files index a3392eb..558c6d1 100644 --- a/sys/conf/files +++ b/sys/conf/files @@ -2108,6 +2108,7 @@ vm/vm_pageout.c standard vm/vm_pageq.c standard vm/vm_pager.c standard vm/vm_phys.c standard +vm/vm_reserv.c standard vm/vm_unix.c standard vm/vm_zeroidle.c standard vm/vnode_pager.c standard diff --git a/sys/kern/kern_exec.c b/sys/kern/kern_exec.c index 2878b37..b0a8dcc 100644 --- a/sys/kern/kern_exec.c +++ b/sys/kern/kern_exec.c @@ -30,6 +30,7 @@ __FBSDID("$FreeBSD$"); #include "opt_hwpmc_hooks.h" #include "opt_ktrace.h" #include "opt_mac.h" +#include "opt_vm.h" #include #include @@ -820,6 +821,12 @@ exec_map_first_page(imgp) if (object == NULL) return (EACCES); VM_OBJECT_LOCK(object); +#if VM_NRESERVLEVEL > 0 + if ((object->flags & OBJ_COLORED) == 0) { + object->flags |= OBJ_COLORED; + object->pg_color = 0; + } +#endif ma[0] = vm_page_grab(object, 0, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); if ((ma[0]->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) { initial_pagein = VM_INITIAL_PAGEIN; diff --git a/sys/vm/vm_fault.c b/sys/vm/vm_fault.c index 49a4077..d2a8220 100644 --- a/sys/vm/vm_fault.c +++ b/sys/vm/vm_fault.c @@ -74,6 +74,8 @@ #include __FBSDID("$FreeBSD$"); +#include "opt_vm.h" + #include #include #include @@ -429,6 +431,13 @@ RetryFault:; */ fs.m = NULL; if (!vm_page_count_severe()) { +#if VM_NRESERVLEVEL > 0 + if ((fs.object->flags & OBJ_COLORED) == 0) { + fs.object->flags |= OBJ_COLORED; + fs.object->pg_color = atop(vaddr) - + fs.pindex; + } +#endif fs.m = vm_page_alloc(fs.object, fs.pindex, (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO); } @@ -1107,6 +1116,10 @@ vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry) */ dst_object = vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(dst_entry->end - dst_entry->start)); +#if VM_NRESERVLEVEL > 0 + dst_object->flags |= OBJ_COLORED; + dst_object->pg_color = atop(dst_entry->start); +#endif VM_OBJECT_LOCK(dst_object); dst_entry->object.vm_object = dst_object; diff --git a/sys/vm/vm_object.c b/sys/vm/vm_object.c index 4c433d3..8f24a8a 100644 --- a/sys/vm/vm_object.c +++ b/sys/vm/vm_object.c @@ -65,6 +65,8 @@ #include __FBSDID("$FreeBSD$"); +#include "opt_vm.h" + #include #include #include @@ -90,6 +92,7 @@ __FBSDID("$FreeBSD$"); #include #include #include +#include #include #define EASY_SCAN_FACTOR 8 @@ -170,6 +173,11 @@ vm_object_zdtor(void *mem, int size, void *arg) KASSERT(TAILQ_EMPTY(&object->memq), ("object %p has resident pages", object)); +#if VM_NRESERVLEVEL > 0 + KASSERT(LIST_EMPTY(&object->rvq), + ("object %p has reservations", + object)); +#endif KASSERT(object->cache == NULL, ("object %p has cached pages", object)); @@ -220,6 +228,9 @@ _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object) object->handle = NULL; object->backing_object = NULL; object->backing_object_offset = (vm_ooffset_t) 0; +#if VM_NRESERVLEVEL > 0 + LIST_INIT(&object->rvq); +#endif object->cache = NULL; mtx_lock(&vm_object_list_mtx); @@ -241,10 +252,18 @@ vm_object_init(void) VM_OBJECT_LOCK_INIT(&kernel_object_store, "kernel object"); _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), kernel_object); +#if VM_NRESERVLEVEL > 0 + kernel_object->flags |= OBJ_COLORED; + kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); +#endif VM_OBJECT_LOCK_INIT(&kmem_object_store, "kmem object"); _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), kmem_object); +#if VM_NRESERVLEVEL > 0 + kmem_object->flags |= OBJ_COLORED; + kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); +#endif /* * The lock portion of struct vm_object must be type stable due @@ -652,6 +671,10 @@ vm_object_terminate(vm_object_t object) } vm_page_unlock_queues(); +#if VM_NRESERVLEVEL > 0 + if (__predict_false(!LIST_EMPTY(&object->rvq))) + vm_reserv_break_all(object); +#endif if (__predict_false(object->cache != NULL)) vm_page_cache_free(object, 0, 0); @@ -1248,7 +1271,13 @@ vm_object_shadow( LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list); source->shadow_count++; source->generation++; +#if VM_NRESERVLEVEL > 0 + result->flags |= source->flags & (OBJ_NEEDGIANT | OBJ_COLORED); + result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & + ((1 << (VM_NFREEORDER - 1)) - 1); +#else result->flags |= source->flags & OBJ_NEEDGIANT; +#endif VM_OBJECT_UNLOCK(source); } @@ -1560,6 +1589,14 @@ vm_object_backing_scan(vm_object_t object, int op) continue; } +#if VM_NRESERVLEVEL > 0 + /* + * Rename the reservation. + */ + vm_reserv_rename(p, object, backing_object, + backing_offset_index); +#endif + /* * Page does not exist in parent, rename the * page from the backing object to the main object. @@ -1662,6 +1699,14 @@ vm_object_collapse(vm_object_t object) */ vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT); +#if VM_NRESERVLEVEL > 0 + /* + * Break any reservations from backing_object. + */ + if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) + vm_reserv_break_all(backing_object); +#endif + /* * Move the pager from backing_object to object. */ diff --git a/sys/vm/vm_page.c b/sys/vm/vm_page.c index 622f5fd..485b738 100644 --- a/sys/vm/vm_page.c +++ b/sys/vm/vm_page.c @@ -99,6 +99,8 @@ #include __FBSDID("$FreeBSD$"); +#include "opt_vm.h" + #include #include #include @@ -118,6 +120,7 @@ __FBSDID("$FreeBSD$"); #include #include #include +#include #include #include #include @@ -318,6 +321,13 @@ vm_page_startup(vm_offset_t vaddr) mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); vm_page_array = (vm_page_t) mapped; +#if VM_NRESERVLEVEL > 0 + /* + * Allocate memory for the reservation management system's data + * structures. + */ + new_end = vm_reserv_startup(&vaddr, new_end, high_water); +#endif #ifdef __amd64__ /* * pmap_map on amd64 comes out of the direct-map, not kvm like i386, @@ -363,6 +373,12 @@ vm_page_startup(vm_offset_t vaddr) } } freeenv(list); +#if VM_NRESERVLEVEL > 0 + /* + * Initialize the reservation management system. + */ + vm_reserv_init(); +#endif return (vaddr); } @@ -1028,15 +1044,33 @@ vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) } if (vm_phys_unfree_page(m)) vm_phys_set_pool(VM_FREEPOOL_DEFAULT, m, 0); +#if VM_NRESERVLEVEL > 0 + else if (!vm_reserv_reactivate_page(m)) +#else else +#endif panic("vm_page_alloc: cache page %p is missing" " from the free queue", m); } else if ((req & VM_ALLOC_IFCACHED) != 0) { mtx_unlock(&vm_page_queue_free_mtx); return (NULL); - } else +#if VM_NRESERVLEVEL > 0 + } else if (object == NULL || + (object->flags & OBJ_COLORED) == 0 || + (m = vm_reserv_alloc_page(object, pindex)) == NULL) { +#else + } else { +#endif m = vm_phys_alloc_pages(object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0); +#if VM_NRESERVLEVEL > 0 + if (m == NULL && vm_reserv_reclaim()) { + m = vm_phys_alloc_pages(object != NULL ? + VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, + 0); + } +#endif + } } else { /* * Not allocatable, give up. @@ -1301,7 +1335,12 @@ vm_page_free_toq(vm_page_t m) m->flags |= PG_FREE; mtx_lock(&vm_page_queue_free_mtx); cnt.v_free_count++; - vm_phys_free_pages(m, 0); +#if VM_NRESERVLEVEL > 0 + if (!vm_reserv_free_page(m)) +#else + if (TRUE) +#endif + vm_phys_free_pages(m, 0); if ((m->flags & PG_ZERO) != 0) ++vm_page_zero_count; else @@ -1548,7 +1587,6 @@ vm_page_cache(vm_page_t m) vm_page_flag_set(m, PG_CACHED); vm_page_flag_clear(m, PG_ZERO); mtx_lock(&vm_page_queue_free_mtx); - vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); cnt.v_cache_count++; root = object->cache; if (root == NULL) { @@ -1569,7 +1607,14 @@ vm_page_cache(vm_page_t m) } } object->cache = m; - vm_phys_free_pages(m, 0); +#if VM_NRESERVLEVEL > 0 + if (!vm_reserv_free_page(m)) { +#else + if (TRUE) { +#endif + vm_phys_set_pool(VM_FREEPOOL_CACHE, m, 0); + vm_phys_free_pages(m, 0); + } vm_page_free_wakeup(); mtx_unlock(&vm_page_queue_free_mtx); diff --git a/sys/vm/vm_reserv.c b/sys/vm/vm_reserv.c new file mode 100644 index 0000000..3d0690c --- /dev/null +++ b/sys/vm/vm_reserv.c @@ -0,0 +1,668 @@ +/*- + * Copyright (c) 2002-2006 Rice University + * Copyright (c) 2007 Alan L. Cox + * All rights reserved. + * + * This software was developed for the FreeBSD Project by Alan L. Cox, + * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS 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 COPYRIGHT + * HOLDERS 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. + */ + +/* + * Superpage reservation management module + */ + +#include +__FBSDID("$FreeBSD$"); + +#include "opt_vm.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include + +/* + * The reservation system supports the speculative allocation of large physical + * pages ("superpages"). Speculative allocation enables the fully-automatic + * utilization of superpages by the virtual memory system. In other words, no + * programmatic directives are required to use superpages. + */ + +#if VM_NRESERVLEVEL > 0 + +/* + * The number of small pages that are contained in a level 0 reservation + */ +#define VM_LEVEL_0_NPAGES (1 << VM_LEVEL_0_ORDER) + +/* + * The number of bits by which a physical address is shifted to obtain the + * reservation number + */ +#define VM_LEVEL_0_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT) + +/* + * The size of a level 0 reservation in bytes + */ +#define VM_LEVEL_0_SIZE (1 << VM_LEVEL_0_SHIFT) + +/* + * Computes the index of the small page underlying the given (object, pindex) + * within the reservation's array of small pages. + */ +#define VM_RESERV_INDEX(object, pindex) \ + (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1)) + +/* + * The reservation structure + * + * A reservation structure is constructed whenever a large physical page is + * speculatively allocated to an object. The reservation provides the small + * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets + * within that object. The reservation's "popcnt" tracks the number of these + * small physical pages that are in use at any given time. When and if the + * reservation is not fully utilized, it appears in the queue of partially- + * populated reservations. The reservation always appears on the containing + * object's list of reservations. + * + * A partially-populated reservation can be broken and reclaimed at any time. + */ +struct vm_reserv { + TAILQ_ENTRY(vm_reserv) partpopq; + LIST_ENTRY(vm_reserv) objq; + vm_object_t object; /* containing object */ + vm_pindex_t pindex; /* offset within object */ + vm_page_t pages; /* first page of a superpage */ + int popcnt; /* # of pages in use */ + char inpartpopq; +}; + +/* + * The reservation array + * + * This array is analoguous in function to vm_page_array. It differs in the + * respect that it may contain a greater number of useful reservation + * structures than there are (physical) superpages. These "invalid" + * reservation structures exist to trade-off space for time in the + * implementation of vm_reserv_from_page(). Invalid reservation structures are + * distinguishable from "valid" reservation structures by inspecting the + * reservation's "pages" field. Invalid reservation structures have a NULL + * "pages" field. + * + * vm_reserv_from_page() maps a small (physical) page to an element of this + * array by computing a physical reservation number from the page's physical + * address. The physical reservation number is used as the array index. + * + * An "active" reservation is a valid reservation structure that has a non-NULL + * "object" field and a non-zero "popcnt" field. In other words, every active + * reservation belongs to a particular object. Moreover, every active + * reservation has an entry in the containing object's list of reservations. + */ +static vm_reserv_t vm_reserv_array; + +/* + * The partially-populated reservation queue + * + * This queue enables the fast recovery of an unused cached or free small page + * from a partially-populated reservation. The head of this queue is either + * the least-recently-populated or most-recently-depopulated reservation. + * + * Access to this queue is synchronized by the free page queue lock. + */ +static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop = + TAILQ_HEAD_INITIALIZER(vm_rvq_partpop); + +static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info"); + +static long vm_reserv_broken; +SYSCTL_LONG(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD, + &vm_reserv_broken, 0, "Cumulative number of broken reservations"); + +static long vm_reserv_freed; +SYSCTL_LONG(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD, + &vm_reserv_freed, 0, "Cumulative number of freed reservations"); + +static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS); + +SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0, + sysctl_vm_reserv_partpopq, "A", "Partially-populated reservation queues"); + +static long vm_reserv_reclaimed; +SYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD, + &vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations"); + +static void vm_reserv_depopulate(vm_reserv_t rv); +static vm_reserv_t vm_reserv_from_page(vm_page_t m); +static boolean_t vm_reserv_has_pindex(vm_reserv_t rv, + vm_pindex_t pindex); +static void vm_reserv_populate(vm_reserv_t rv); + +/* + * Describes the current state of the partially-populated reservation queue. + */ +static int +sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS) +{ + struct sbuf sbuf; + vm_reserv_t rv; + char *cbuf; + const int cbufsize = (VM_NRESERVLEVEL + 1) * 81; + int counter, error, level, unused_pages; + + cbuf = malloc(cbufsize, M_TEMP, M_WAITOK | M_ZERO); + sbuf_new(&sbuf, cbuf, cbufsize, SBUF_FIXEDLEN); + sbuf_printf(&sbuf, "\nLEVEL SIZE NUMBER\n\n"); + for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) { + counter = 0; + unused_pages = 0; + mtx_lock(&vm_page_queue_free_mtx); + TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) { + counter++; + unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt; + } + mtx_unlock(&vm_page_queue_free_mtx); + sbuf_printf(&sbuf, "%5.5d: %6.6dK, %6.6d\n", level, + unused_pages * (PAGE_SIZE / 1024), counter); + } + sbuf_finish(&sbuf); + error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf)); + sbuf_delete(&sbuf); + free(cbuf, M_TEMP); + return (error); +} + +/* + * Reduces the given reservation's population count. If the population count + * becomes zero, the reservation is destroyed. Additionally, moves the + * reservation to the head of the partially-populated reservations queue if the + * population count is non-zero. + * + * The free page queue lock must be held. + */ +static void +vm_reserv_depopulate(vm_reserv_t rv) +{ + + mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); + KASSERT(rv->object != NULL, + ("vm_reserv_depopulate: reserv %p is free", rv)); + KASSERT(rv->popcnt > 0, + ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv)); + if (rv->inpartpopq) { + TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq); + rv->inpartpopq = FALSE; + } + rv->popcnt--; + if (rv->popcnt == 0) { + LIST_REMOVE(rv, objq); + rv->object = NULL; + vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER); + vm_reserv_freed++; + } else { + rv->inpartpopq = TRUE; + TAILQ_INSERT_HEAD(&vm_rvq_partpop, rv, partpopq); + } +} + +/* + * Returns the reservation to which the given page might belong. + */ +static __inline vm_reserv_t +vm_reserv_from_page(vm_page_t m) +{ + + return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]); +} + +/* + * Returns TRUE if the given reservation contains the given page index and + * FALSE otherwise. + */ +static __inline boolean_t +vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex) +{ + + return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0); +} + +/* + * Increases the given reservation's population count. Moves the reservation + * to the tail of the partially-populated reservation queue. + * + * The free page queue must be locked. + */ +static void +vm_reserv_populate(vm_reserv_t rv) +{ + + mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); + KASSERT(rv->object != NULL, + ("vm_reserv_populate: reserv %p is free", rv)); + KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES, + ("vm_reserv_populate: reserv %p is already full", rv)); + if (rv->inpartpopq) { + TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq); + rv->inpartpopq = FALSE; + } + rv->popcnt++; + if (rv->popcnt < VM_LEVEL_0_NPAGES) { + rv->inpartpopq = TRUE; + TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq); + } +} + +/* + * Allocates a page from an existing or newly-created reservation. + * + * The object and free page queue must be locked. + */ +vm_page_t +vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex) +{ + vm_page_t m, mpred, msucc; + vm_pindex_t first, leftcap, rightcap; + vm_reserv_t rv; + + mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); + + /* + * Is a reservation fundamentally not possible? + */ + VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); + if (pindex < VM_RESERV_INDEX(object, pindex) || + pindex >= object->size) + return (NULL); + + /* + * Look for an existing reservation. + */ + msucc = NULL; + mpred = object->root; + while (mpred != NULL) { + KASSERT(mpred->pindex != pindex, + ("vm_reserv_alloc_page: pindex already allocated")); + rv = vm_reserv_from_page(mpred); + if (rv->object == object && vm_reserv_has_pindex(rv, pindex)) { + m = &rv->pages[VM_RESERV_INDEX(object, pindex)]; + /* Handle vm_page_rename(m, new_object, ...). */ + if ((m->flags & (PG_CACHED | PG_FREE)) == 0) + return (NULL); + vm_reserv_populate(rv); + return (m); + } else if (mpred->pindex < pindex) { + if (msucc != NULL || + (msucc = TAILQ_NEXT(mpred, listq)) == NULL) + break; + KASSERT(msucc->pindex != pindex, + ("vm_reserv_alloc_page: pindex already allocated")); + rv = vm_reserv_from_page(msucc); + if (rv->object == object && + vm_reserv_has_pindex(rv, pindex)) { + m = &rv->pages[VM_RESERV_INDEX(object, pindex)]; + /* Handle vm_page_rename(m, new_object, ...). */ + if ((m->flags & (PG_CACHED | PG_FREE)) == 0) + return (NULL); + vm_reserv_populate(rv); + return (m); + } else if (pindex < msucc->pindex) + break; + } else if (msucc == NULL) { + msucc = mpred; + mpred = TAILQ_PREV(msucc, pglist, listq); + continue; + } + msucc = NULL; + mpred = object->root = vm_page_splay(pindex, object->root); + } + + /* + * Determine the first index to the left that can be used. + */ + if (mpred == NULL) + leftcap = 0; + else if ((rv = vm_reserv_from_page(mpred))->object != object) + leftcap = mpred->pindex + 1; + else + leftcap = rv->pindex + VM_LEVEL_0_NPAGES; + + /* + * Determine the first index to the right that cannot be used. + */ + if (msucc == NULL) + rightcap = pindex + VM_LEVEL_0_NPAGES; + else if ((rv = vm_reserv_from_page(msucc))->object != object) + rightcap = msucc->pindex; + else + rightcap = rv->pindex; + + /* + * Determine if a reservation fits between the first index to + * the left that can be used and the first index to the right + * that cannot be used. + */ + first = pindex - VM_RESERV_INDEX(object, pindex); + if (first < leftcap || first + VM_LEVEL_0_NPAGES > rightcap) + return (NULL); + + /* + * Would a new reservation extend past the end of the given object? + */ + if (object->size < first + VM_LEVEL_0_NPAGES) { + /* + * Don't allocate a new reservation if the object is a vnode or + * backed by another object that is a vnode. + */ + if (object->type == OBJT_VNODE || + (object->backing_object != NULL && + object->backing_object->type == OBJT_VNODE)) + return (NULL); + /* Speculate that the object may grow. */ + } + + /* + * Allocate a new reservation. + */ + m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER); + if (m != NULL) { + rv = vm_reserv_from_page(m); + KASSERT(rv->pages == m, + ("vm_reserv_alloc_page: reserv %p's pages is corrupted", + rv)); + KASSERT(rv->object == NULL, + ("vm_reserv_alloc_page: reserv %p isn't free", rv)); + LIST_INSERT_HEAD(&object->rvq, rv, objq); + rv->object = object; + rv->pindex = first; + KASSERT(rv->popcnt == 0, + ("vm_reserv_alloc_page: reserv %p's popcnt is corrupted", + rv)); + KASSERT(!rv->inpartpopq, + ("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE", + rv)); + vm_reserv_populate(rv); + m = &rv->pages[VM_RESERV_INDEX(object, pindex)]; + } + return (m); +} + +/* + * Breaks all reservations belonging to the given object. + */ +void +vm_reserv_break_all(vm_object_t object) +{ + vm_reserv_t rv; + int i; + + mtx_lock(&vm_page_queue_free_mtx); + while ((rv = LIST_FIRST(&object->rvq)) != NULL) { + KASSERT(rv->object == object, + ("vm_reserv_break_all: reserv %p is corrupted", rv)); + if (rv->inpartpopq) { + TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq); + rv->inpartpopq = FALSE; + } + LIST_REMOVE(rv, objq); + rv->object = NULL; + for (i = 0; i < VM_LEVEL_0_NPAGES; i++) { + if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) + vm_phys_free_pages(&rv->pages[i], 0); + else + rv->popcnt--; + } + KASSERT(rv->popcnt == 0, + ("vm_reserv_break_all: reserv %p's popcnt is corrupted", + rv)); + vm_reserv_broken++; + } + mtx_unlock(&vm_page_queue_free_mtx); +} + +/* + * Frees the given page if it belongs to a reservation. Returns TRUE if the + * page is freed and FALSE otherwise. + * + * The free page queue lock must be held. + */ +boolean_t +vm_reserv_free_page(vm_page_t m) +{ + vm_reserv_t rv; + + mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); + rv = vm_reserv_from_page(m); + if (rv->object != NULL) { + vm_reserv_depopulate(rv); + return (TRUE); + } + return (FALSE); +} + +/* + * Initializes the reservation management system. Specifically, initializes + * the reservation array. + * + * Requires that vm_page_array and first_page are initialized! + */ +void +vm_reserv_init(void) +{ + vm_paddr_t paddr; + int i; + + /* + * Initialize the reservation array. Specifically, initialize the + * "pages" field for every element that has an underlying superpage. + */ + for (i = 0; phys_avail[i + 1] != 0; i += 2) { + paddr = roundup2(phys_avail[i], VM_LEVEL_0_SIZE); + while (paddr + VM_LEVEL_0_SIZE <= phys_avail[i + 1]) { + vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages = + PHYS_TO_VM_PAGE(paddr); + paddr += VM_LEVEL_0_SIZE; + } + } +} + +/* + * Returns a reservation level if the given page belongs to a fully-populated + * reservation and -1 otherwise. + */ +int +vm_reserv_level_iffullpop(vm_page_t m) +{ + vm_reserv_t rv; + + rv = vm_reserv_from_page(m); + return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1); +} + +/* + * Prepare for the reactivation of a cached page. + * + * First, suppose that the given page "m" was allocated individually, i.e., not + * as part of a reservation, and cached. Then, suppose a reservation + * containing "m" is allocated by the same object. Although "m" and the + * reservation belong to the same object, "m"'s pindex may not match the + * reservation's. + * + * The free page queue must be locked. + */ +boolean_t +vm_reserv_reactivate_page(vm_page_t m) +{ + vm_reserv_t rv; + int i, m_index; + + mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); + rv = vm_reserv_from_page(m); + if (rv->object == NULL) + return (FALSE); + KASSERT((m->flags & PG_CACHED) != 0, + ("vm_reserv_uncache_page: page %p is not cached", m)); + if (m->object == rv->object && + m->pindex - rv->pindex == VM_RESERV_INDEX(m->object, m->pindex)) + vm_reserv_populate(rv); + else { + KASSERT(rv->inpartpopq, + ("vm_reserv_uncache_page: reserv %p's inpartpopq is FALSE", + rv)); + TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq); + rv->inpartpopq = FALSE; + LIST_REMOVE(rv, objq); + rv->object = NULL; + /* Don't vm_phys_free_pages(m, 0). */ + m_index = m - rv->pages; + for (i = 0; i < m_index; i++) { + if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) + vm_phys_free_pages(&rv->pages[i], 0); + else + rv->popcnt--; + } + for (i++; i < VM_LEVEL_0_NPAGES; i++) { + if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) + vm_phys_free_pages(&rv->pages[i], 0); + else + rv->popcnt--; + } + KASSERT(rv->popcnt == 0, + ("vm_reserv_uncache_page: reserv %p's popcnt is corrupted", + rv)); + vm_reserv_broken++; + } + return (TRUE); +} + +/* + * Breaks the reservation at the head of the partially-populated reservation + * queue, releasing its cached and free pages to the physical memory + * allocator. Returns TRUE if a reservation is broken and FALSE otherwise. + * + * The free page queue lock must be held. + */ +boolean_t +vm_reserv_reclaim(void) +{ + vm_reserv_t rv; + int i; + + mtx_assert(&vm_page_queue_free_mtx, MA_OWNED); + if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) { + KASSERT(rv->inpartpopq, + ("vm_reserv_reclaim: reserv %p's inpartpopq is corrupted", + rv)); + TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq); + rv->inpartpopq = FALSE; + KASSERT(rv->object != NULL, + ("vm_reserv_reclaim: reserv %p is free", rv)); + LIST_REMOVE(rv, objq); + rv->object = NULL; + for (i = 0; i < VM_LEVEL_0_NPAGES; i++) { + if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0) + vm_phys_free_pages(&rv->pages[i], 0); + else + rv->popcnt--; + } + KASSERT(rv->popcnt == 0, + ("vm_reserv_reclaim: reserv %p's popcnt is corrupted", + rv)); + vm_reserv_reclaimed++; + return (TRUE); + } + return (FALSE); +} + +/* + * Transfers the reservation underlying the given page to a new object. + * + * The object must be locked. + */ +void +vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object, + vm_pindex_t old_object_offset) +{ + vm_reserv_t rv; + + VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED); + rv = vm_reserv_from_page(m); + if (rv->object == old_object) { + mtx_lock(&vm_page_queue_free_mtx); + if (rv->object == old_object) { + LIST_REMOVE(rv, objq); + LIST_INSERT_HEAD(&new_object->rvq, rv, objq); + rv->object = new_object; + rv->pindex -= old_object_offset; + } + mtx_unlock(&vm_page_queue_free_mtx); + } +} + +/* + * Allocates the virtual and physical memory required by the reservation + * management system's data structures, in particular, the reservation array. + */ +vm_paddr_t +vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water) +{ + vm_paddr_t new_end; + size_t size; + + /* + * Calculate the size (in bytes) of the reservation array. Round up + * from "high_water" because every small page is mapped to an element + * in the reservation array based on its physical address. Thus, the + * number of elements in the reservation array can be greater than the + * number of superpages. + */ + size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv); + + /* + * Allocate and map the physical memory for the reservation array. The + * next available virtual address is returned by reference. + */ + new_end = end - round_page(size); + vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end, + VM_PROT_READ | VM_PROT_WRITE); + bzero(vm_reserv_array, size); + + /* + * Return the next available physical address. + */ + return (new_end); +} + +#endif /* VM_NRESERVLEVEL > 0 */ diff --git a/sys/vm/vm_reserv.h b/sys/vm/vm_reserv.h new file mode 100644 index 0000000..7cf814b --- /dev/null +++ b/sys/vm/vm_reserv.h @@ -0,0 +1,59 @@ +/*- + * Copyright (c) 2002-2006 Rice University + * Copyright (c) 2007 Alan L. Cox + * All rights reserved. + * + * This software was developed for the FreeBSD Project by Alan L. Cox, + * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS 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 COPYRIGHT + * HOLDERS 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. + * + * $FreeBSD$ + */ + +/* + * Superpage reservation management definitions + */ + +#ifndef _VM_RESERV_H_ +#define _VM_RESERV_H_ + +#ifdef _KERNEL + +#if VM_NRESERVLEVEL > 0 + +vm_page_t vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex); +void vm_reserv_break_all(vm_object_t object); +boolean_t vm_reserv_free_page(vm_page_t m); +void vm_reserv_init(void); +int vm_reserv_level_iffullpop(vm_page_t m); +boolean_t vm_reserv_reactivate_page(vm_page_t m); +boolean_t vm_reserv_reclaim(void); +void vm_reserv_rename(vm_page_t m, vm_object_t new_object, + vm_object_t old_object, vm_pindex_t old_object_offset); +vm_paddr_t vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, + vm_paddr_t high_water); + +#endif /* VM_NRESERVLEVEL > 0 */ +#endif /* _KERNEL */ +#endif /* !_VM_RESERV_H_ */ -- cgit v1.1