Index: head/sys/vm/swap_pager.c =================================================================== --- head/sys/vm/swap_pager.c (revision 163621) +++ head/sys/vm/swap_pager.c (revision 163622) @@ -1,2574 +1,2574 @@ /*- * Copyright (c) 1998 Matthew Dillon, * Copyright (c) 1994 John S. Dyson * Copyright (c) 1990 University of Utah. * Copyright (c) 1982, 1986, 1989, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department. * * 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. * * New Swap System * Matthew Dillon * * Radix Bitmap 'blists'. * * - The new swapper uses the new radix bitmap code. This should scale * to arbitrarily small or arbitrarily large swap spaces and an almost * arbitrary degree of fragmentation. * * Features: * * - on the fly reallocation of swap during putpages. The new system * does not try to keep previously allocated swap blocks for dirty * pages. * * - on the fly deallocation of swap * * - No more garbage collection required. Unnecessarily allocated swap * blocks only exist for dirty vm_page_t's now and these are already * cycled (in a high-load system) by the pager. We also do on-the-fly * removal of invalidated swap blocks when a page is destroyed * or renamed. * * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ * * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_mac.h" #include "opt_swap.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * SWB_NPAGES must be a power of 2. It may be set to 1, 2, 4, 8, or 16 * pages per allocation. We recommend you stick with the default of 8. * The 16-page limit is due to the radix code (kern/subr_blist.c). */ #ifndef MAX_PAGEOUT_CLUSTER #define MAX_PAGEOUT_CLUSTER 16 #endif #if !defined(SWB_NPAGES) #define SWB_NPAGES MAX_PAGEOUT_CLUSTER #endif /* * Piecemeal swap metadata structure. Swap is stored in a radix tree. * * If SWB_NPAGES is 8 and sizeof(char *) == sizeof(daddr_t), our radix * is basically 8. Assuming PAGE_SIZE == 4096, one tree level represents * 32K worth of data, two levels represent 256K, three levels represent * 2 MBytes. This is acceptable. * * Overall memory utilization is about the same as the old swap structure. */ #define SWCORRECT(n) (sizeof(void *) * (n) / sizeof(daddr_t)) #define SWAP_META_PAGES (SWB_NPAGES * 2) #define SWAP_META_MASK (SWAP_META_PAGES - 1) typedef int32_t swblk_t; /* * swap offset. This is the type used to * address the "virtual swap device" and * therefore the maximum swap space is * 2^32 pages. */ struct swdevt; typedef void sw_strategy_t(struct buf *bp, struct swdevt *sw); typedef void sw_close_t(struct thread *td, struct swdevt *sw); /* * Swap device table */ struct swdevt { int sw_flags; int sw_nblks; int sw_used; dev_t sw_dev; struct vnode *sw_vp; void *sw_id; swblk_t sw_first; swblk_t sw_end; struct blist *sw_blist; TAILQ_ENTRY(swdevt) sw_list; sw_strategy_t *sw_strategy; sw_close_t *sw_close; }; #define SW_CLOSING 0x04 struct swblock { struct swblock *swb_hnext; vm_object_t swb_object; vm_pindex_t swb_index; int swb_count; daddr_t swb_pages[SWAP_META_PAGES]; }; static struct mtx sw_dev_mtx; static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); static struct swdevt *swdevhd; /* Allocate from here next */ static int nswapdev; /* Number of swap devices */ int swap_pager_avail; static int swdev_syscall_active = 0; /* serialize swap(on|off) */ static void swapdev_strategy(struct buf *, struct swdevt *sw); #define SWM_FREE 0x02 /* free, period */ #define SWM_POP 0x04 /* pop out */ int swap_pager_full = 2; /* swap space exhaustion (task killing) */ static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ static int nsw_rcount; /* free read buffers */ static int nsw_wcount_sync; /* limit write buffers / synchronous */ static int nsw_wcount_async; /* limit write buffers / asynchronous */ static int nsw_wcount_async_max;/* assigned maximum */ static int nsw_cluster_max; /* maximum VOP I/O allowed */ static struct swblock **swhash; static int swhash_mask; static struct mtx swhash_mtx; static int swap_async_max = 4; /* maximum in-progress async I/O's */ static struct sx sw_alloc_sx; SYSCTL_INT(_vm, OID_AUTO, swap_async_max, CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops"); /* * "named" and "unnamed" anon region objects. Try to reduce the overhead * of searching a named list by hashing it just a little. */ #define NOBJLISTS 8 #define NOBJLIST(handle) \ (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) static struct mtx sw_alloc_mtx; /* protect list manipulation */ static struct pagerlst swap_pager_object_list[NOBJLISTS]; static uma_zone_t swap_zone; static struct vm_object swap_zone_obj; /* * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure * calls hooked from other parts of the VM system and do not appear here. * (see vm/swap_pager.h). */ static vm_object_t swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t offset); static void swap_pager_dealloc(vm_object_t object); static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int); static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); static boolean_t swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); static void swap_pager_init(void); static void swap_pager_unswapped(vm_page_t); static void swap_pager_swapoff(struct swdevt *sp); struct pagerops swappagerops = { .pgo_init = swap_pager_init, /* early system initialization of pager */ .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ .pgo_getpages = swap_pager_getpages, /* pagein */ .pgo_putpages = swap_pager_putpages, /* pageout */ .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ }; /* * dmmax is in page-sized chunks with the new swap system. It was * dev-bsized chunks in the old. dmmax is always a power of 2. * * swap_*() routines are externally accessible. swp_*() routines are * internal. */ static int dmmax; static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &dmmax, 0, "Maximum size of a swap block"); static void swp_sizecheck(void); static void swp_pager_async_iodone(struct buf *bp); static int swapongeom(struct thread *, struct vnode *); static int swaponvp(struct thread *, struct vnode *, u_long); static int swapoff_one(struct swdevt *sp, struct thread *td); /* * Swap bitmap functions */ static void swp_pager_freeswapspace(daddr_t blk, int npages); static daddr_t swp_pager_getswapspace(int npages); /* * Metadata functions */ static struct swblock **swp_pager_hash(vm_object_t object, vm_pindex_t index); static void swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); static void swp_pager_meta_free(vm_object_t, vm_pindex_t, daddr_t); static void swp_pager_meta_free_all(vm_object_t); static daddr_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int); /* * SWP_SIZECHECK() - update swap_pager_full indication * * update the swap_pager_almost_full indication and warn when we are * about to run out of swap space, using lowat/hiwat hysteresis. * * Clear swap_pager_full ( task killing ) indication when lowat is met. * * No restrictions on call * This routine may not block. * This routine must be called at splvm() */ static void swp_sizecheck(void) { if (swap_pager_avail < nswap_lowat) { if (swap_pager_almost_full == 0) { printf("swap_pager: out of swap space\n"); swap_pager_almost_full = 1; } } else { swap_pager_full = 0; if (swap_pager_avail > nswap_hiwat) swap_pager_almost_full = 0; } } /* * SWP_PAGER_HASH() - hash swap meta data * * This is an helper function which hashes the swapblk given * the object and page index. It returns a pointer to a pointer * to the object, or a pointer to a NULL pointer if it could not * find a swapblk. * * This routine must be called at splvm(). */ static struct swblock ** swp_pager_hash(vm_object_t object, vm_pindex_t index) { struct swblock **pswap; struct swblock *swap; index &= ~(vm_pindex_t)SWAP_META_MASK; pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask]; while ((swap = *pswap) != NULL) { if (swap->swb_object == object && swap->swb_index == index ) { break; } pswap = &swap->swb_hnext; } return (pswap); } /* * SWAP_PAGER_INIT() - initialize the swap pager! * * Expected to be started from system init. NOTE: This code is run * before much else so be careful what you depend on. Most of the VM * system has yet to be initialized at this point. */ static void swap_pager_init(void) { /* * Initialize object lists */ int i; for (i = 0; i < NOBJLISTS; ++i) TAILQ_INIT(&swap_pager_object_list[i]); mtx_init(&sw_alloc_mtx, "swap_pager list", NULL, MTX_DEF); mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); /* * Device Stripe, in PAGE_SIZE'd blocks */ dmmax = SWB_NPAGES * 2; } /* * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process * * Expected to be started from pageout process once, prior to entering * its main loop. */ void swap_pager_swap_init(void) { int n, n2; /* * Number of in-transit swap bp operations. Don't * exhaust the pbufs completely. Make sure we * initialize workable values (0 will work for hysteresis * but it isn't very efficient). * * The nsw_cluster_max is constrained by the bp->b_pages[] * array (MAXPHYS/PAGE_SIZE) and our locally defined * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are * constrained by the swap device interleave stripe size. * * Currently we hardwire nsw_wcount_async to 4. This limit is * designed to prevent other I/O from having high latencies due to * our pageout I/O. The value 4 works well for one or two active swap * devices but is probably a little low if you have more. Even so, * a higher value would probably generate only a limited improvement * with three or four active swap devices since the system does not * typically have to pageout at extreme bandwidths. We will want * at least 2 per swap devices, and 4 is a pretty good value if you * have one NFS swap device due to the command/ack latency over NFS. * So it all works out pretty well. */ nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER); mtx_lock(&pbuf_mtx); nsw_rcount = (nswbuf + 1) / 2; nsw_wcount_sync = (nswbuf + 3) / 4; nsw_wcount_async = 4; nsw_wcount_async_max = nsw_wcount_async; mtx_unlock(&pbuf_mtx); /* * Initialize our zone. Right now I'm just guessing on the number * we need based on the number of pages in the system. Each swblock * can hold 16 pages, so this is probably overkill. This reservation * is typically limited to around 32MB by default. */ n = cnt.v_page_count / 2; if (maxswzone && n > maxswzone / sizeof(struct swblock)) n = maxswzone / sizeof(struct swblock); n2 = n; swap_zone = uma_zcreate("SWAPMETA", sizeof(struct swblock), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); if (swap_zone == NULL) panic("failed to create swap_zone."); do { if (uma_zone_set_obj(swap_zone, &swap_zone_obj, n)) break; /* * if the allocation failed, try a zone two thirds the * size of the previous attempt. */ n -= ((n + 2) / 3); } while (n > 0); if (n2 != n) printf("Swap zone entries reduced from %d to %d.\n", n2, n); n2 = n; /* * Initialize our meta-data hash table. The swapper does not need to * be quite as efficient as the VM system, so we do not use an * oversized hash table. * * n: size of hash table, must be power of 2 * swhash_mask: hash table index mask */ for (n = 1; n < n2 / 8; n *= 2) ; swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK | M_ZERO); swhash_mask = n - 1; mtx_init(&swhash_mtx, "swap_pager swhash", NULL, MTX_DEF); } /* * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate * its metadata structures. * * This routine is called from the mmap and fork code to create a new * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object * and then converting it with swp_pager_meta_build(). * * This routine may block in vm_object_allocate() and create a named * object lookup race, so we must interlock. We must also run at * splvm() for the object lookup to handle races with interrupts, but * we do not have to maintain splvm() in between the lookup and the * add because (I believe) it is not possible to attempt to create * a new swap object w/handle when a default object with that handle * already exists. * * MPSAFE */ static vm_object_t swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t offset) { vm_object_t object; vm_pindex_t pindex; pindex = OFF_TO_IDX(offset + PAGE_MASK + size); if (handle) { mtx_lock(&Giant); /* * Reference existing named region or allocate new one. There * should not be a race here against swp_pager_meta_build() * as called from vm_page_remove() in regards to the lookup * of the handle. */ sx_xlock(&sw_alloc_sx); object = vm_pager_object_lookup(NOBJLIST(handle), handle); if (object != NULL) { vm_object_reference(object); } else { object = vm_object_allocate(OBJT_DEFAULT, pindex); object->handle = handle; VM_OBJECT_LOCK(object); swp_pager_meta_build(object, 0, SWAPBLK_NONE); VM_OBJECT_UNLOCK(object); } sx_xunlock(&sw_alloc_sx); mtx_unlock(&Giant); } else { object = vm_object_allocate(OBJT_DEFAULT, pindex); VM_OBJECT_LOCK(object); swp_pager_meta_build(object, 0, SWAPBLK_NONE); VM_OBJECT_UNLOCK(object); } return (object); } /* * SWAP_PAGER_DEALLOC() - remove swap metadata from object * * The swap backing for the object is destroyed. The code is * designed such that we can reinstantiate it later, but this * routine is typically called only when the entire object is * about to be destroyed. * * This routine may block, but no longer does. * * The object must be locked or unreferenceable. */ static void swap_pager_dealloc(vm_object_t object) { /* * Remove from list right away so lookups will fail if we block for * pageout completion. */ if (object->handle != NULL) { mtx_lock(&sw_alloc_mtx); TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list); mtx_unlock(&sw_alloc_mtx); } VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); vm_object_pip_wait(object, "swpdea"); /* * Free all remaining metadata. We only bother to free it from * the swap meta data. We do not attempt to free swapblk's still * associated with vm_page_t's for this object. We do not care * if paging is still in progress on some objects. */ swp_pager_meta_free_all(object); } /************************************************************************ * SWAP PAGER BITMAP ROUTINES * ************************************************************************/ /* * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space * * Allocate swap for the requested number of pages. The starting * swap block number (a page index) is returned or SWAPBLK_NONE * if the allocation failed. * * Also has the side effect of advising that somebody made a mistake * when they configured swap and didn't configure enough. * * Must be called at splvm() to avoid races with bitmap frees from * vm_page_remove() aka swap_pager_page_removed(). * * This routine may not block * This routine must be called at splvm(). * * We allocate in round-robin fashion from the configured devices. */ static daddr_t swp_pager_getswapspace(int npages) { daddr_t blk; struct swdevt *sp; int i; blk = SWAPBLK_NONE; mtx_lock(&sw_dev_mtx); sp = swdevhd; for (i = 0; i < nswapdev; i++) { if (sp == NULL) sp = TAILQ_FIRST(&swtailq); if (!(sp->sw_flags & SW_CLOSING)) { blk = blist_alloc(sp->sw_blist, npages); if (blk != SWAPBLK_NONE) { blk += sp->sw_first; sp->sw_used += npages; swap_pager_avail -= npages; swp_sizecheck(); swdevhd = TAILQ_NEXT(sp, sw_list); goto done; } } sp = TAILQ_NEXT(sp, sw_list); } if (swap_pager_full != 2) { printf("swap_pager_getswapspace(%d): failed\n", npages); swap_pager_full = 2; swap_pager_almost_full = 1; } swdevhd = NULL; done: mtx_unlock(&sw_dev_mtx); return (blk); } static int swp_pager_isondev(daddr_t blk, struct swdevt *sp) { return (blk >= sp->sw_first && blk < sp->sw_end); } static void swp_pager_strategy(struct buf *bp) { struct swdevt *sp; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (bp->b_blkno >= sp->sw_first && bp->b_blkno < sp->sw_end) { mtx_unlock(&sw_dev_mtx); sp->sw_strategy(bp, sp); return; } } panic("Swapdev not found"); } /* * SWP_PAGER_FREESWAPSPACE() - free raw swap space * * This routine returns the specified swap blocks back to the bitmap. * * Note: This routine may not block (it could in the old swap code), * and through the use of the new blist routines it does not block. * * We must be called at splvm() to avoid races with bitmap frees from * vm_page_remove() aka swap_pager_page_removed(). * * This routine may not block * This routine must be called at splvm(). */ static void swp_pager_freeswapspace(daddr_t blk, int npages) { struct swdevt *sp; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (blk >= sp->sw_first && blk < sp->sw_end) { sp->sw_used -= npages; /* * If we are attempting to stop swapping on * this device, we don't want to mark any * blocks free lest they be reused. */ if ((sp->sw_flags & SW_CLOSING) == 0) { blist_free(sp->sw_blist, blk - sp->sw_first, npages); swap_pager_avail += npages; swp_sizecheck(); } mtx_unlock(&sw_dev_mtx); return; } } panic("Swapdev not found"); } /* * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page * range within an object. * * This is a globally accessible routine. * * This routine removes swapblk assignments from swap metadata. * * The external callers of this routine typically have already destroyed * or renamed vm_page_t's associated with this range in the object so * we should be ok. * * This routine may be called at any spl. We up our spl to splvm temporarily * in order to perform the metadata removal. */ void swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) { VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); swp_pager_meta_free(object, start, size); } /* * SWAP_PAGER_RESERVE() - reserve swap blocks in object * * Assigns swap blocks to the specified range within the object. The * swap blocks are not zerod. Any previous swap assignment is destroyed. * * Returns 0 on success, -1 on failure. */ int swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) { int n = 0; daddr_t blk = SWAPBLK_NONE; vm_pindex_t beg = start; /* save start index */ VM_OBJECT_LOCK(object); while (size) { if (n == 0) { n = BLIST_MAX_ALLOC; while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) { n >>= 1; if (n == 0) { swp_pager_meta_free(object, beg, start - beg); VM_OBJECT_UNLOCK(object); return (-1); } } } swp_pager_meta_build(object, start, blk); --size; ++start; ++blk; --n; } swp_pager_meta_free(object, start, n); VM_OBJECT_UNLOCK(object); return (0); } /* * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager * and destroy the source. * * Copy any valid swapblks from the source to the destination. In * cases where both the source and destination have a valid swapblk, * we keep the destination's. * * This routine is allowed to block. It may block allocating metadata * indirectly through swp_pager_meta_build() or if paging is still in * progress on the source. * * This routine can be called at any spl * * XXX vm_page_collapse() kinda expects us not to block because we * supposedly do not need to allocate memory, but for the moment we * *may* have to get a little memory from the zone allocator, but * it is taken from the interrupt memory. We should be ok. * * The source object contains no vm_page_t's (which is just as well) * * The source object is of type OBJT_SWAP. * * The source and destination objects must be locked or * inaccessible (XXX are they ?) */ void swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, vm_pindex_t offset, int destroysource) { vm_pindex_t i; VM_OBJECT_LOCK_ASSERT(srcobject, MA_OWNED); VM_OBJECT_LOCK_ASSERT(dstobject, MA_OWNED); /* * If destroysource is set, we remove the source object from the * swap_pager internal queue now. */ if (destroysource) { if (srcobject->handle != NULL) { mtx_lock(&sw_alloc_mtx); TAILQ_REMOVE( NOBJLIST(srcobject->handle), srcobject, pager_object_list ); mtx_unlock(&sw_alloc_mtx); } } /* * transfer source to destination. */ for (i = 0; i < dstobject->size; ++i) { daddr_t dstaddr; /* * Locate (without changing) the swapblk on the destination, * unless it is invalid in which case free it silently, or * if the destination is a resident page, in which case the * source is thrown away. */ dstaddr = swp_pager_meta_ctl(dstobject, i, 0); if (dstaddr == SWAPBLK_NONE) { /* * Destination has no swapblk and is not resident, * copy source. */ daddr_t srcaddr; srcaddr = swp_pager_meta_ctl( srcobject, i + offset, SWM_POP ); if (srcaddr != SWAPBLK_NONE) { /* * swp_pager_meta_build() can sleep. */ vm_object_pip_add(srcobject, 1); VM_OBJECT_UNLOCK(srcobject); vm_object_pip_add(dstobject, 1); swp_pager_meta_build(dstobject, i, srcaddr); vm_object_pip_wakeup(dstobject); VM_OBJECT_LOCK(srcobject); vm_object_pip_wakeup(srcobject); } } else { /* * Destination has valid swapblk or it is represented * by a resident page. We destroy the sourceblock. */ swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE); } } /* * Free left over swap blocks in source. * * We have to revert the type to OBJT_DEFAULT so we do not accidently * double-remove the object from the swap queues. */ if (destroysource) { swp_pager_meta_free_all(srcobject); /* * Reverting the type is not necessary, the caller is going * to destroy srcobject directly, but I'm doing it here * for consistency since we've removed the object from its * queues. */ srcobject->type = OBJT_DEFAULT; } } /* * SWAP_PAGER_HASPAGE() - determine if we have good backing store for * the requested page. * * We determine whether good backing store exists for the requested * page and return TRUE if it does, FALSE if it doesn't. * * If TRUE, we also try to determine how much valid, contiguous backing * store exists before and after the requested page within a reasonable * distance. We do not try to restrict it to the swap device stripe * (that is handled in getpages/putpages). It probably isn't worth * doing here. */ static boolean_t swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) { daddr_t blk0; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); /* * do we have good backing store at the requested index ? */ blk0 = swp_pager_meta_ctl(object, pindex, 0); if (blk0 == SWAPBLK_NONE) { if (before) *before = 0; if (after) *after = 0; return (FALSE); } /* * find backwards-looking contiguous good backing store */ if (before != NULL) { int i; for (i = 1; i < (SWB_NPAGES/2); ++i) { daddr_t blk; if (i > pindex) break; blk = swp_pager_meta_ctl(object, pindex - i, 0); if (blk != blk0 - i) break; } *before = (i - 1); } /* * find forward-looking contiguous good backing store */ if (after != NULL) { int i; for (i = 1; i < (SWB_NPAGES/2); ++i) { daddr_t blk; blk = swp_pager_meta_ctl(object, pindex + i, 0); if (blk != blk0 + i) break; } *after = (i - 1); } return (TRUE); } /* * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page * * This removes any associated swap backing store, whether valid or * not, from the page. * * This routine is typically called when a page is made dirty, at * which point any associated swap can be freed. MADV_FREE also * calls us in a special-case situation * * NOTE!!! If the page is clean and the swap was valid, the caller * should make the page dirty before calling this routine. This routine * does NOT change the m->dirty status of the page. Also: MADV_FREE * depends on it. * * This routine may not block * This routine must be called at splvm() */ static void swap_pager_unswapped(vm_page_t m) { VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE); } /* * SWAP_PAGER_GETPAGES() - bring pages in from swap * * Attempt to retrieve (m, count) pages from backing store, but make * sure we retrieve at least m[reqpage]. We try to load in as large * a chunk surrounding m[reqpage] as is contiguous in swap and which * belongs to the same object. * * The code is designed for asynchronous operation and * immediate-notification of 'reqpage' but tends not to be * used that way. Please do not optimize-out this algorithmic * feature, I intend to improve on it in the future. * * The parent has a single vm_object_pip_add() reference prior to * calling us and we should return with the same. * * The parent has BUSY'd the pages. We should return with 'm' * left busy, but the others adjusted. */ static int swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage) { struct buf *bp; vm_page_t mreq; int i; int j; daddr_t blk; mreq = m[reqpage]; KASSERT(mreq->object == object, ("swap_pager_getpages: object mismatch %p/%p", object, mreq->object)); /* * Calculate range to retrieve. The pages have already been assigned * their swapblks. We require a *contiguous* range but we know it to * not span devices. If we do not supply it, bad things * happen. Note that blk, iblk & jblk can be SWAPBLK_NONE, but the * loops are set up such that the case(s) are handled implicitly. * * The swp_*() calls must be made at splvm(). vm_page_free() does * not need to be, but it will go a little faster if it is. */ blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0); for (i = reqpage - 1; i >= 0; --i) { daddr_t iblk; iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0); if (blk != iblk + (reqpage - i)) break; } ++i; for (j = reqpage + 1; j < count; ++j) { daddr_t jblk; jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0); if (blk != jblk - (j - reqpage)) break; } /* * free pages outside our collection range. Note: we never free * mreq, it must remain busy throughout. */ if (0 < i || j < count) { int k; vm_page_lock_queues(); for (k = 0; k < i; ++k) vm_page_free(m[k]); for (k = j; k < count; ++k) vm_page_free(m[k]); vm_page_unlock_queues(); } /* * Return VM_PAGER_FAIL if we have nothing to do. Return mreq * still busy, but the others unbusied. */ if (blk == SWAPBLK_NONE) return (VM_PAGER_FAIL); /* * Getpbuf() can sleep. */ VM_OBJECT_UNLOCK(object); /* * Get a swap buffer header to perform the IO */ bp = getpbuf(&nsw_rcount); bp->b_flags |= B_PAGING; /* * map our page(s) into kva for input */ pmap_qenter((vm_offset_t)bp->b_data, m + i, j - i); bp->b_iocmd = BIO_READ; bp->b_iodone = swp_pager_async_iodone; bp->b_rcred = crhold(thread0.td_ucred); bp->b_wcred = crhold(thread0.td_ucred); bp->b_blkno = blk - (reqpage - i); bp->b_bcount = PAGE_SIZE * (j - i); bp->b_bufsize = PAGE_SIZE * (j - i); bp->b_pager.pg_reqpage = reqpage - i; VM_OBJECT_LOCK(object); { int k; for (k = i; k < j; ++k) { bp->b_pages[k - i] = m[k]; m[k]->oflags |= VPO_SWAPINPROG; } } bp->b_npages = j - i; cnt.v_swapin++; cnt.v_swappgsin += bp->b_npages; /* * We still hold the lock on mreq, and our automatic completion routine * does not remove it. */ vm_object_pip_add(object, bp->b_npages); VM_OBJECT_UNLOCK(object); /* * perform the I/O. NOTE!!! bp cannot be considered valid after * this point because we automatically release it on completion. * Instead, we look at the one page we are interested in which we * still hold a lock on even through the I/O completion. * * The other pages in our m[] array are also released on completion, * so we cannot assume they are valid anymore either. * * NOTE: b_blkno is destroyed by the call to swapdev_strategy */ BUF_KERNPROC(bp); swp_pager_strategy(bp); /* * wait for the page we want to complete. VPO_SWAPINPROG is always * cleared on completion. If an I/O error occurs, SWAPBLK_NONE * is set in the meta-data. */ VM_OBJECT_LOCK(object); while ((mreq->oflags & VPO_SWAPINPROG) != 0) { mreq->oflags |= VPO_WANTED; vm_page_lock_queues(); vm_page_flag_set(mreq, PG_REFERENCED); vm_page_unlock_queues(); cnt.v_intrans++; if (msleep(mreq, VM_OBJECT_MTX(object), PSWP, "swread", hz*20)) { printf( "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); } } /* * mreq is left busied after completion, but all the other pages * are freed. If we had an unrecoverable read error the page will * not be valid. */ if (mreq->valid != VM_PAGE_BITS_ALL) { return (VM_PAGER_ERROR); } else { return (VM_PAGER_OK); } /* * A final note: in a low swap situation, we cannot deallocate swap * and mark a page dirty here because the caller is likely to mark * the page clean when we return, causing the page to possibly revert * to all-zero's later. */ } /* * swap_pager_putpages: * * Assign swap (if necessary) and initiate I/O on the specified pages. * * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects * are automatically converted to SWAP objects. * * In a low memory situation we may block in VOP_STRATEGY(), but the new * vm_page reservation system coupled with properly written VFS devices * should ensure that no low-memory deadlock occurs. This is an area * which needs work. * * The parent has N vm_object_pip_add() references prior to * calling us and will remove references for rtvals[] that are * not set to VM_PAGER_PEND. We need to remove the rest on I/O * completion. * * The parent has soft-busy'd the pages it passes us and will unbusy * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. * We need to unbusy the rest on I/O completion. */ void swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, boolean_t sync, int *rtvals) { int i; int n = 0; GIANT_REQUIRED; if (count && m[0]->object != object) { panic("swap_pager_getpages: object mismatch %p/%p", object, m[0]->object ); } /* * Step 1 * * Turn object into OBJT_SWAP * check for bogus sysops * force sync if not pageout process */ if (object->type != OBJT_SWAP) swp_pager_meta_build(object, 0, SWAPBLK_NONE); VM_OBJECT_UNLOCK(object); if (curproc != pageproc) sync = TRUE; /* * Step 2 * * Update nsw parameters from swap_async_max sysctl values. * Do not let the sysop crash the machine with bogus numbers. */ mtx_lock(&pbuf_mtx); if (swap_async_max != nsw_wcount_async_max) { int n; /* * limit range */ if ((n = swap_async_max) > nswbuf / 2) n = nswbuf / 2; if (n < 1) n = 1; swap_async_max = n; /* * Adjust difference ( if possible ). If the current async * count is too low, we may not be able to make the adjustment * at this time. */ n -= nsw_wcount_async_max; if (nsw_wcount_async + n >= 0) { nsw_wcount_async += n; nsw_wcount_async_max += n; wakeup(&nsw_wcount_async); } } mtx_unlock(&pbuf_mtx); /* * Step 3 * * Assign swap blocks and issue I/O. We reallocate swap on the fly. * The page is left dirty until the pageout operation completes * successfully. */ for (i = 0; i < count; i += n) { int j; struct buf *bp; daddr_t blk; /* * Maximum I/O size is limited by a number of factors. */ n = min(BLIST_MAX_ALLOC, count - i); n = min(n, nsw_cluster_max); /* * Get biggest block of swap we can. If we fail, fall * back and try to allocate a smaller block. Don't go * overboard trying to allocate space if it would overly * fragment swap. */ while ( (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE && n > 4 ) { n >>= 1; } if (blk == SWAPBLK_NONE) { for (j = 0; j < n; ++j) rtvals[i+j] = VM_PAGER_FAIL; continue; } /* * All I/O parameters have been satisfied, build the I/O * request and assign the swap space. */ if (sync == TRUE) { bp = getpbuf(&nsw_wcount_sync); } else { bp = getpbuf(&nsw_wcount_async); bp->b_flags = B_ASYNC; } bp->b_flags |= B_PAGING; bp->b_iocmd = BIO_WRITE; pmap_qenter((vm_offset_t)bp->b_data, &m[i], n); bp->b_rcred = crhold(thread0.td_ucred); bp->b_wcred = crhold(thread0.td_ucred); bp->b_bcount = PAGE_SIZE * n; bp->b_bufsize = PAGE_SIZE * n; bp->b_blkno = blk; VM_OBJECT_LOCK(object); for (j = 0; j < n; ++j) { vm_page_t mreq = m[i+j]; swp_pager_meta_build( mreq->object, mreq->pindex, blk + j ); vm_page_dirty(mreq); rtvals[i+j] = VM_PAGER_OK; mreq->oflags |= VPO_SWAPINPROG; bp->b_pages[j] = mreq; } VM_OBJECT_UNLOCK(object); bp->b_npages = n; /* * Must set dirty range for NFS to work. */ bp->b_dirtyoff = 0; bp->b_dirtyend = bp->b_bcount; cnt.v_swapout++; cnt.v_swappgsout += bp->b_npages; /* * asynchronous * * NOTE: b_blkno is destroyed by the call to swapdev_strategy */ if (sync == FALSE) { bp->b_iodone = swp_pager_async_iodone; BUF_KERNPROC(bp); swp_pager_strategy(bp); for (j = 0; j < n; ++j) rtvals[i+j] = VM_PAGER_PEND; /* restart outter loop */ continue; } /* * synchronous * * NOTE: b_blkno is destroyed by the call to swapdev_strategy */ bp->b_iodone = bdone; swp_pager_strategy(bp); /* * Wait for the sync I/O to complete, then update rtvals. * We just set the rtvals[] to VM_PAGER_PEND so we can call * our async completion routine at the end, thus avoiding a * double-free. */ bwait(bp, PVM, "swwrt"); for (j = 0; j < n; ++j) rtvals[i+j] = VM_PAGER_PEND; /* * Now that we are through with the bp, we can call the * normal async completion, which frees everything up. */ swp_pager_async_iodone(bp); } VM_OBJECT_LOCK(object); } /* * swp_pager_async_iodone: * * Completion routine for asynchronous reads and writes from/to swap. * Also called manually by synchronous code to finish up a bp. * * For READ operations, the pages are PG_BUSY'd. For WRITE operations, * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY * unbusy all pages except the 'main' request page. For WRITE * operations, we vm_page_t->busy'd unbusy all pages ( we can do this * because we marked them all VM_PAGER_PEND on return from putpages ). * * This routine may not block. * This routine is called at splbio() or better * * We up ourselves to splvm() as required for various vm_page related * calls. */ static void swp_pager_async_iodone(struct buf *bp) { int i; vm_object_t object = NULL; /* * report error */ if (bp->b_ioflags & BIO_ERROR) { printf( "swap_pager: I/O error - %s failed; blkno %ld," "size %ld, error %d\n", ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), (long)bp->b_blkno, (long)bp->b_bcount, bp->b_error ); } /* * remove the mapping for kernel virtual */ pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); if (bp->b_npages) { object = bp->b_pages[0]->object; VM_OBJECT_LOCK(object); } vm_page_lock_queues(); /* * cleanup pages. If an error occurs writing to swap, we are in * very serious trouble. If it happens to be a disk error, though, * we may be able to recover by reassigning the swap later on. So * in this case we remove the m->swapblk assignment for the page * but do not free it in the rlist. The errornous block(s) are thus * never reallocated as swap. Redirty the page and continue. */ for (i = 0; i < bp->b_npages; ++i) { vm_page_t m = bp->b_pages[i]; m->oflags &= ~VPO_SWAPINPROG; if (bp->b_ioflags & BIO_ERROR) { /* * If an error occurs I'd love to throw the swapblk * away without freeing it back to swapspace, so it * can never be used again. But I can't from an * interrupt. */ if (bp->b_iocmd == BIO_READ) { /* * When reading, reqpage needs to stay * locked for the parent, but all other * pages can be freed. We still want to * wakeup the parent waiting on the page, * though. ( also: pg_reqpage can be -1 and * not match anything ). * * We have to wake specifically requested pages * up too because we cleared VPO_SWAPINPROG and * someone may be waiting for that. * * NOTE: for reads, m->dirty will probably * be overridden by the original caller of * getpages so don't play cute tricks here. */ m->valid = 0; if (i != bp->b_pager.pg_reqpage) vm_page_free(m); else vm_page_flash(m); /* * If i == bp->b_pager.pg_reqpage, do not wake * the page up. The caller needs to. */ } else { /* * If a write error occurs, reactivate page * so it doesn't clog the inactive list, * then finish the I/O. */ vm_page_dirty(m); vm_page_activate(m); vm_page_io_finish(m); } } else if (bp->b_iocmd == BIO_READ) { /* * For read success, clear dirty bits. Nobody should * have this page mapped but don't take any chances, * make sure the pmap modify bits are also cleared. * * NOTE: for reads, m->dirty will probably be * overridden by the original caller of getpages so * we cannot set them in order to free the underlying * swap in a low-swap situation. I don't think we'd * want to do that anyway, but it was an optimization * that existed in the old swapper for a time before * it got ripped out due to precisely this problem. * * If not the requested page then deactivate it. * * Note that the requested page, reqpage, is left * busied, but we still have to wake it up. The * other pages are released (unbusied) by * vm_page_wakeup(). We do not set reqpage's * valid bits here, it is up to the caller. */ pmap_clear_modify(m); m->valid = VM_PAGE_BITS_ALL; vm_page_undirty(m); /* * We have to wake specifically requested pages * up too because we cleared VPO_SWAPINPROG and * could be waiting for it in getpages. However, * be sure to not unbusy getpages specifically * requested page - getpages expects it to be * left busy. */ if (i != bp->b_pager.pg_reqpage) { vm_page_deactivate(m); vm_page_wakeup(m); } else { vm_page_flash(m); } } else { /* * For write success, clear the modify and dirty * status, then finish the I/O ( which decrements the * busy count and possibly wakes waiter's up ). */ pmap_clear_modify(m); vm_page_undirty(m); vm_page_io_finish(m); if (vm_page_count_severe()) vm_page_try_to_cache(m); } } vm_page_unlock_queues(); /* * adjust pip. NOTE: the original parent may still have its own * pip refs on the object. */ if (object != NULL) { vm_object_pip_wakeupn(object, bp->b_npages); VM_OBJECT_UNLOCK(object); } /* * swapdev_strategy() manually sets b_vp and b_bufobj before calling * bstrategy(). Set them back to NULL now we're done with it, or we'll * trigger a KASSERT in relpbuf(). */ if (bp->b_vp) { bp->b_vp = NULL; bp->b_bufobj = NULL; } /* * release the physical I/O buffer */ relpbuf( bp, ((bp->b_iocmd == BIO_READ) ? &nsw_rcount : ((bp->b_flags & B_ASYNC) ? &nsw_wcount_async : &nsw_wcount_sync ) ) ); } /* * swap_pager_isswapped: * * Return 1 if at least one page in the given object is paged * out to the given swap device. * * This routine may not block. */ int swap_pager_isswapped(vm_object_t object, struct swdevt *sp) { daddr_t index = 0; int bcount; int i; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); if (object->type != OBJT_SWAP) return (0); mtx_lock(&swhash_mtx); for (bcount = 0; bcount < object->un_pager.swp.swp_bcount; bcount++) { struct swblock *swap; if ((swap = *swp_pager_hash(object, index)) != NULL) { for (i = 0; i < SWAP_META_PAGES; ++i) { if (swp_pager_isondev(swap->swb_pages[i], sp)) { mtx_unlock(&swhash_mtx); return (1); } } } index += SWAP_META_PAGES; if (index > 0x20000000) panic("swap_pager_isswapped: failed to locate all swap meta blocks"); } mtx_unlock(&swhash_mtx); return (0); } /* * SWP_PAGER_FORCE_PAGEIN() - force a swap block to be paged in * * This routine dissociates the page at the given index within a * swap block from its backing store, paging it in if necessary. * If the page is paged in, it is placed in the inactive queue, * since it had its backing store ripped out from under it. * We also attempt to swap in all other pages in the swap block, * we only guarantee that the one at the specified index is * paged in. * * XXX - The code to page the whole block in doesn't work, so we * revert to the one-by-one behavior for now. Sigh. */ static inline void swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex) { vm_page_t m; vm_object_pip_add(object, 1); m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL|VM_ALLOC_RETRY); if (m->valid == VM_PAGE_BITS_ALL) { vm_object_pip_subtract(object, 1); vm_page_lock_queues(); vm_page_activate(m); vm_page_dirty(m); - vm_page_wakeup(m); vm_page_unlock_queues(); + vm_page_wakeup(m); vm_pager_page_unswapped(m); return; } if (swap_pager_getpages(object, &m, 1, 0) != VM_PAGER_OK) panic("swap_pager_force_pagein: read from swap failed");/*XXX*/ vm_object_pip_subtract(object, 1); vm_page_lock_queues(); vm_page_dirty(m); vm_page_dontneed(m); - vm_page_wakeup(m); vm_page_unlock_queues(); + vm_page_wakeup(m); vm_pager_page_unswapped(m); } /* * swap_pager_swapoff: * * Page in all of the pages that have been paged out to the * given device. The corresponding blocks in the bitmap must be * marked as allocated and the device must be flagged SW_CLOSING. * There may be no processes swapped out to the device. * * This routine may block. */ static void swap_pager_swapoff(struct swdevt *sp) { struct swblock *swap; int i, j, retries; GIANT_REQUIRED; retries = 0; full_rescan: mtx_lock(&swhash_mtx); for (i = 0; i <= swhash_mask; i++) { /* '<=' is correct here */ restart: for (swap = swhash[i]; swap != NULL; swap = swap->swb_hnext) { vm_object_t object = swap->swb_object; vm_pindex_t pindex = swap->swb_index; for (j = 0; j < SWAP_META_PAGES; ++j) { if (swp_pager_isondev(swap->swb_pages[j], sp)) { /* avoid deadlock */ if (!VM_OBJECT_TRYLOCK(object)) { break; } else { mtx_unlock(&swhash_mtx); swp_pager_force_pagein(object, pindex + j); VM_OBJECT_UNLOCK(object); mtx_lock(&swhash_mtx); goto restart; } } } } } mtx_unlock(&swhash_mtx); if (sp->sw_used) { int dummy; /* * Objects may be locked or paging to the device being * removed, so we will miss their pages and need to * make another pass. We have marked this device as * SW_CLOSING, so the activity should finish soon. */ retries++; if (retries > 100) { panic("swapoff: failed to locate %d swap blocks", sp->sw_used); } tsleep(&dummy, PVM, "swpoff", hz / 20); goto full_rescan; } } /************************************************************************ * SWAP META DATA * ************************************************************************ * * These routines manipulate the swap metadata stored in the * OBJT_SWAP object. All swp_*() routines must be called at * splvm() because swap can be freed up by the low level vm_page * code which might be called from interrupts beyond what splbio() covers. * * Swap metadata is implemented with a global hash and not directly * linked into the object. Instead the object simply contains * appropriate tracking counters. */ /* * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object * * We first convert the object to a swap object if it is a default * object. * * The specified swapblk is added to the object's swap metadata. If * the swapblk is not valid, it is freed instead. Any previously * assigned swapblk is freed. * * This routine must be called at splvm(), except when used to convert * an OBJT_DEFAULT object into an OBJT_SWAP object. */ static void swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) { struct swblock *swap; struct swblock **pswap; int idx; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); /* * Convert default object to swap object if necessary */ if (object->type != OBJT_SWAP) { object->type = OBJT_SWAP; object->un_pager.swp.swp_bcount = 0; if (object->handle != NULL) { mtx_lock(&sw_alloc_mtx); TAILQ_INSERT_TAIL( NOBJLIST(object->handle), object, pager_object_list ); mtx_unlock(&sw_alloc_mtx); } } /* * Locate hash entry. If not found create, but if we aren't adding * anything just return. If we run out of space in the map we wait * and, since the hash table may have changed, retry. */ retry: mtx_lock(&swhash_mtx); pswap = swp_pager_hash(object, pindex); if ((swap = *pswap) == NULL) { int i; if (swapblk == SWAPBLK_NONE) goto done; swap = *pswap = uma_zalloc(swap_zone, M_NOWAIT); if (swap == NULL) { mtx_unlock(&swhash_mtx); VM_OBJECT_UNLOCK(object); VM_WAIT; VM_OBJECT_LOCK(object); goto retry; } swap->swb_hnext = NULL; swap->swb_object = object; swap->swb_index = pindex & ~(vm_pindex_t)SWAP_META_MASK; swap->swb_count = 0; ++object->un_pager.swp.swp_bcount; for (i = 0; i < SWAP_META_PAGES; ++i) swap->swb_pages[i] = SWAPBLK_NONE; } /* * Delete prior contents of metadata */ idx = pindex & SWAP_META_MASK; if (swap->swb_pages[idx] != SWAPBLK_NONE) { swp_pager_freeswapspace(swap->swb_pages[idx], 1); --swap->swb_count; } /* * Enter block into metadata */ swap->swb_pages[idx] = swapblk; if (swapblk != SWAPBLK_NONE) ++swap->swb_count; done: mtx_unlock(&swhash_mtx); } /* * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata * * The requested range of blocks is freed, with any associated swap * returned to the swap bitmap. * * This routine will free swap metadata structures as they are cleaned * out. This routine does *NOT* operate on swap metadata associated * with resident pages. * * This routine must be called at splvm() */ static void swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count) { VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); if (object->type != OBJT_SWAP) return; while (count > 0) { struct swblock **pswap; struct swblock *swap; mtx_lock(&swhash_mtx); pswap = swp_pager_hash(object, index); if ((swap = *pswap) != NULL) { daddr_t v = swap->swb_pages[index & SWAP_META_MASK]; if (v != SWAPBLK_NONE) { swp_pager_freeswapspace(v, 1); swap->swb_pages[index & SWAP_META_MASK] = SWAPBLK_NONE; if (--swap->swb_count == 0) { *pswap = swap->swb_hnext; uma_zfree(swap_zone, swap); --object->un_pager.swp.swp_bcount; } } --count; ++index; } else { int n = SWAP_META_PAGES - (index & SWAP_META_MASK); count -= n; index += n; } mtx_unlock(&swhash_mtx); } } /* * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object * * This routine locates and destroys all swap metadata associated with * an object. * * This routine must be called at splvm() */ static void swp_pager_meta_free_all(vm_object_t object) { daddr_t index = 0; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); if (object->type != OBJT_SWAP) return; while (object->un_pager.swp.swp_bcount) { struct swblock **pswap; struct swblock *swap; mtx_lock(&swhash_mtx); pswap = swp_pager_hash(object, index); if ((swap = *pswap) != NULL) { int i; for (i = 0; i < SWAP_META_PAGES; ++i) { daddr_t v = swap->swb_pages[i]; if (v != SWAPBLK_NONE) { --swap->swb_count; swp_pager_freeswapspace(v, 1); } } if (swap->swb_count != 0) panic("swap_pager_meta_free_all: swb_count != 0"); *pswap = swap->swb_hnext; uma_zfree(swap_zone, swap); --object->un_pager.swp.swp_bcount; } mtx_unlock(&swhash_mtx); index += SWAP_META_PAGES; if (index > 0x20000000) panic("swp_pager_meta_free_all: failed to locate all swap meta blocks"); } } /* * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. * * This routine is capable of looking up, popping, or freeing * swapblk assignments in the swap meta data or in the vm_page_t. * The routine typically returns the swapblk being looked-up, or popped, * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block * was invalid. This routine will automatically free any invalid * meta-data swapblks. * * It is not possible to store invalid swapblks in the swap meta data * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. * * When acting on a busy resident page and paging is in progress, we * have to wait until paging is complete but otherwise can act on the * busy page. * * This routine must be called at splvm(). * * SWM_FREE remove and free swap block from metadata * SWM_POP remove from meta data but do not free.. pop it out */ static daddr_t swp_pager_meta_ctl(vm_object_t object, vm_pindex_t pindex, int flags) { struct swblock **pswap; struct swblock *swap; daddr_t r1; int idx; VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); /* * The meta data only exists of the object is OBJT_SWAP * and even then might not be allocated yet. */ if (object->type != OBJT_SWAP) return (SWAPBLK_NONE); r1 = SWAPBLK_NONE; mtx_lock(&swhash_mtx); pswap = swp_pager_hash(object, pindex); if ((swap = *pswap) != NULL) { idx = pindex & SWAP_META_MASK; r1 = swap->swb_pages[idx]; if (r1 != SWAPBLK_NONE) { if (flags & SWM_FREE) { swp_pager_freeswapspace(r1, 1); r1 = SWAPBLK_NONE; } if (flags & (SWM_FREE|SWM_POP)) { swap->swb_pages[idx] = SWAPBLK_NONE; if (--swap->swb_count == 0) { *pswap = swap->swb_hnext; uma_zfree(swap_zone, swap); --object->un_pager.swp.swp_bcount; } } } } mtx_unlock(&swhash_mtx); return (r1); } /* * System call swapon(name) enables swapping on device name, * which must be in the swdevsw. Return EBUSY * if already swapping on this device. */ #ifndef _SYS_SYSPROTO_H_ struct swapon_args { char *name; }; #endif /* * MPSAFE */ /* ARGSUSED */ int swapon(struct thread *td, struct swapon_args *uap) { struct vattr attr; struct vnode *vp; struct nameidata nd; int error; mtx_lock(&Giant); error = suser(td); if (error) goto done2; while (swdev_syscall_active) tsleep(&swdev_syscall_active, PUSER - 1, "swpon", 0); swdev_syscall_active = 1; /* * Swap metadata may not fit in the KVM if we have physical * memory of >1GB. */ if (swap_zone == NULL) { error = ENOMEM; goto done; } NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW, UIO_USERSPACE, uap->name, td); error = namei(&nd); if (error) goto done; NDFREE(&nd, NDF_ONLY_PNBUF); vp = nd.ni_vp; if (vn_isdisk(vp, &error)) { error = swapongeom(td, vp); } else if (vp->v_type == VREG && (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && (error = VOP_GETATTR(vp, &attr, td->td_ucred, td)) == 0) { /* * Allow direct swapping to NFS regular files in the same * way that nfs_mountroot() sets up diskless swapping. */ error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); } if (error) vrele(vp); done: swdev_syscall_active = 0; wakeup_one(&swdev_syscall_active); done2: mtx_unlock(&Giant); return (error); } static void swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev) { struct swdevt *sp, *tsp; swblk_t dvbase; u_long mblocks; /* * If we go beyond this, we get overflows in the radix * tree bitmap code. */ mblocks = 0x40000000 / BLIST_META_RADIX; if (nblks > mblocks) { printf("WARNING: reducing size to maximum of %lu blocks per swap unit\n", mblocks); nblks = mblocks; } /* * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. * First chop nblks off to page-align it, then convert. * * sw->sw_nblks is in page-sized chunks now too. */ nblks &= ~(ctodb(1) - 1); nblks = dbtoc(nblks); sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); sp->sw_vp = vp; sp->sw_id = id; sp->sw_dev = dev; sp->sw_flags = 0; sp->sw_nblks = nblks; sp->sw_used = 0; sp->sw_strategy = strategy; sp->sw_close = close; sp->sw_blist = blist_create(nblks); /* * Do not free the first two block in order to avoid overwriting * any bsd label at the front of the partition */ blist_free(sp->sw_blist, 2, nblks - 2); dvbase = 0; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(tsp, &swtailq, sw_list) { if (tsp->sw_end >= dvbase) { /* * We put one uncovered page between the devices * in order to definitively prevent any cross-device * I/O requests */ dvbase = tsp->sw_end + 1; } } sp->sw_first = dvbase; sp->sw_end = dvbase + nblks; TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); nswapdev++; swap_pager_avail += nblks; swp_sizecheck(); mtx_unlock(&sw_dev_mtx); } /* * SYSCALL: swapoff(devname) * * Disable swapping on the given device. * * XXX: Badly designed system call: it should use a device index * rather than filename as specification. We keep sw_vp around * only to make this work. */ #ifndef _SYS_SYSPROTO_H_ struct swapoff_args { char *name; }; #endif /* * MPSAFE */ /* ARGSUSED */ int swapoff(struct thread *td, struct swapoff_args *uap) { struct vnode *vp; struct nameidata nd; struct swdevt *sp; int error; error = suser(td); if (error) return (error); mtx_lock(&Giant); while (swdev_syscall_active) tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0); swdev_syscall_active = 1; NDINIT(&nd, LOOKUP, FOLLOW, UIO_USERSPACE, uap->name, td); error = namei(&nd); if (error) goto done; NDFREE(&nd, NDF_ONLY_PNBUF); vp = nd.ni_vp; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (sp->sw_vp == vp) break; } mtx_unlock(&sw_dev_mtx); if (sp == NULL) { error = EINVAL; goto done; } error = swapoff_one(sp, td); done: swdev_syscall_active = 0; wakeup_one(&swdev_syscall_active); mtx_unlock(&Giant); return (error); } static int swapoff_one(struct swdevt *sp, struct thread *td) { u_long nblks, dvbase; #ifdef MAC int error; #endif mtx_assert(&Giant, MA_OWNED); #ifdef MAC (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY, td); error = mac_check_system_swapoff(td->td_ucred, sp->sw_vp); (void) VOP_UNLOCK(sp->sw_vp, 0, td); if (error != 0) return (error); #endif nblks = sp->sw_nblks; /* * We can turn off this swap device safely only if the * available virtual memory in the system will fit the amount * of data we will have to page back in, plus an epsilon so * the system doesn't become critically low on swap space. */ if (cnt.v_free_count + cnt.v_cache_count + swap_pager_avail < nblks + nswap_lowat) { return (ENOMEM); } /* * Prevent further allocations on this device. */ mtx_lock(&sw_dev_mtx); sp->sw_flags |= SW_CLOSING; for (dvbase = 0; dvbase < sp->sw_end; dvbase += dmmax) { swap_pager_avail -= blist_fill(sp->sw_blist, dvbase, dmmax); } mtx_unlock(&sw_dev_mtx); /* * Page in the contents of the device and close it. */ swap_pager_swapoff(sp); sp->sw_close(td, sp); sp->sw_id = NULL; mtx_lock(&sw_dev_mtx); TAILQ_REMOVE(&swtailq, sp, sw_list); nswapdev--; if (nswapdev == 0) { swap_pager_full = 2; swap_pager_almost_full = 1; } if (swdevhd == sp) swdevhd = NULL; mtx_unlock(&sw_dev_mtx); blist_destroy(sp->sw_blist); free(sp, M_VMPGDATA); return (0); } void swapoff_all(void) { struct swdevt *sp, *spt; const char *devname; int error; mtx_lock(&Giant); while (swdev_syscall_active) tsleep(&swdev_syscall_active, PUSER - 1, "swpoff", 0); swdev_syscall_active = 1; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { mtx_unlock(&sw_dev_mtx); if (vn_isdisk(sp->sw_vp, NULL)) devname = sp->sw_vp->v_rdev->si_name; else devname = "[file]"; error = swapoff_one(sp, &thread0); if (error != 0) { printf("Cannot remove swap device %s (error=%d), " "skipping.\n", devname, error); } else if (bootverbose) { printf("Swap device %s removed.\n", devname); } mtx_lock(&sw_dev_mtx); } mtx_unlock(&sw_dev_mtx); swdev_syscall_active = 0; wakeup_one(&swdev_syscall_active); mtx_unlock(&Giant); } void swap_pager_status(int *total, int *used) { struct swdevt *sp; *total = 0; *used = 0; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { *total += sp->sw_nblks; *used += sp->sw_used; } mtx_unlock(&sw_dev_mtx); } static int sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) { int *name = (int *)arg1; int error, n; struct xswdev xs; struct swdevt *sp; if (arg2 != 1) /* name length */ return (EINVAL); n = 0; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (n == *name) { mtx_unlock(&sw_dev_mtx); xs.xsw_version = XSWDEV_VERSION; xs.xsw_dev = sp->sw_dev; xs.xsw_flags = sp->sw_flags; xs.xsw_nblks = sp->sw_nblks; xs.xsw_used = sp->sw_used; error = SYSCTL_OUT(req, &xs, sizeof(xs)); return (error); } n++; } mtx_unlock(&sw_dev_mtx); return (ENOENT); } SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, "Number of swap devices"); SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD, sysctl_vm_swap_info, "Swap statistics by device"); /* * vmspace_swap_count() - count the approximate swap useage in pages for a * vmspace. * * The map must be locked. * * Swap useage is determined by taking the proportional swap used by * VM objects backing the VM map. To make up for fractional losses, * if the VM object has any swap use at all the associated map entries * count for at least 1 swap page. */ int vmspace_swap_count(struct vmspace *vmspace) { vm_map_t map = &vmspace->vm_map; vm_map_entry_t cur; int count = 0; for (cur = map->header.next; cur != &map->header; cur = cur->next) { vm_object_t object; if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 && (object = cur->object.vm_object) != NULL) { VM_OBJECT_LOCK(object); if (object->type == OBJT_SWAP && object->un_pager.swp.swp_bcount != 0) { int n = (cur->end - cur->start) / PAGE_SIZE; count += object->un_pager.swp.swp_bcount * SWAP_META_PAGES * n / object->size + 1; } VM_OBJECT_UNLOCK(object); } } return (count); } /* * GEOM backend * * Swapping onto disk devices. * */ static g_orphan_t swapgeom_orphan; static struct g_class g_swap_class = { .name = "SWAP", .version = G_VERSION, .orphan = swapgeom_orphan, }; DECLARE_GEOM_CLASS(g_swap_class, g_class); static void swapgeom_done(struct bio *bp2) { struct buf *bp; bp = bp2->bio_caller2; bp->b_ioflags = bp2->bio_flags; if (bp2->bio_error) bp->b_ioflags |= BIO_ERROR; bp->b_resid = bp->b_bcount - bp2->bio_completed; bp->b_error = bp2->bio_error; bufdone(bp); g_destroy_bio(bp2); } static void swapgeom_strategy(struct buf *bp, struct swdevt *sp) { struct bio *bio; struct g_consumer *cp; cp = sp->sw_id; if (cp == NULL) { bp->b_error = ENXIO; bp->b_ioflags |= BIO_ERROR; bufdone(bp); return; } bio = g_alloc_bio(); #if 0 /* * XXX: We shouldn't really sleep here when we run out of buffers * XXX: but the alternative is worse right now. */ if (bio == NULL) { bp->b_error = ENOMEM; bp->b_ioflags |= BIO_ERROR; bufdone(bp); return; } #endif bio->bio_caller2 = bp; bio->bio_cmd = bp->b_iocmd; bio->bio_data = bp->b_data; bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; bio->bio_length = bp->b_bcount; bio->bio_done = swapgeom_done; g_io_request(bio, cp); return; } static void swapgeom_orphan(struct g_consumer *cp) { struct swdevt *sp; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) if (sp->sw_id == cp) sp->sw_id = NULL; mtx_unlock(&sw_dev_mtx); } static void swapgeom_close_ev(void *arg, int flags) { struct g_consumer *cp; cp = arg; g_access(cp, -1, -1, 0); g_detach(cp); g_destroy_consumer(cp); } static void swapgeom_close(struct thread *td, struct swdevt *sw) { /* XXX: direct call when Giant untangled */ g_waitfor_event(swapgeom_close_ev, sw->sw_id, M_WAITOK, NULL); } struct swh0h0 { struct cdev *dev; struct vnode *vp; int error; }; static void swapongeom_ev(void *arg, int flags) { struct swh0h0 *swh; struct g_provider *pp; struct g_consumer *cp; static struct g_geom *gp; struct swdevt *sp; u_long nblks; int error; swh = arg; swh->error = 0; pp = g_dev_getprovider(swh->dev); if (pp == NULL) { swh->error = ENODEV; return; } mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { cp = sp->sw_id; if (cp != NULL && cp->provider == pp) { mtx_unlock(&sw_dev_mtx); swh->error = EBUSY; return; } } mtx_unlock(&sw_dev_mtx); if (gp == NULL) gp = g_new_geomf(&g_swap_class, "swap", NULL); cp = g_new_consumer(gp); g_attach(cp, pp); /* * XXX: Everytime you think you can improve the margin for * footshooting, somebody depends on the ability to do so: * savecore(8) wants to write to our swapdev so we cannot * set an exclusive count :-( */ error = g_access(cp, 1, 1, 0); if (error) { g_detach(cp); g_destroy_consumer(cp); swh->error = error; return; } nblks = pp->mediasize / DEV_BSIZE; swaponsomething(swh->vp, cp, nblks, swapgeom_strategy, swapgeom_close, dev2udev(swh->dev)); swh->error = 0; return; } static int swapongeom(struct thread *td, struct vnode *vp) { int error; struct swh0h0 swh; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); swh.dev = vp->v_rdev; swh.vp = vp; swh.error = 0; /* XXX: direct call when Giant untangled */ error = g_waitfor_event(swapongeom_ev, &swh, M_WAITOK, NULL); if (!error) error = swh.error; VOP_UNLOCK(vp, 0, td); return (error); } /* * VNODE backend * * This is used mainly for network filesystem (read: probably only tested * with NFS) swapfiles. * */ static void swapdev_strategy(struct buf *bp, struct swdevt *sp) { struct vnode *vp2; bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); vp2 = sp->sw_id; vhold(vp2); if (bp->b_iocmd == BIO_WRITE) { if (bp->b_bufobj) bufobj_wdrop(bp->b_bufobj); bufobj_wref(&vp2->v_bufobj); } if (bp->b_bufobj != &vp2->v_bufobj) bp->b_bufobj = &vp2->v_bufobj; bp->b_vp = vp2; bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); return; } static void swapdev_close(struct thread *td, struct swdevt *sp) { VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); vrele(sp->sw_vp); } static int swaponvp(struct thread *td, struct vnode *vp, u_long nblks) { struct swdevt *sp; int error; if (nblks == 0) return (ENXIO); mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (sp->sw_id == vp) { mtx_unlock(&sw_dev_mtx); return (EBUSY); } } mtx_unlock(&sw_dev_mtx); (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td); #ifdef MAC error = mac_check_system_swapon(td->td_ucred, vp); if (error == 0) #endif error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, -1); (void) VOP_UNLOCK(vp, 0, td); if (error) return (error); swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, NODEV); return (0); } Index: head/sys/vm/vm_fault.c =================================================================== --- head/sys/vm/vm_fault.c (revision 163621) +++ head/sys/vm/vm_fault.c (revision 163622) @@ -1,1348 +1,1348 @@ /*- * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Page fault handling module. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* XXX Temporary for VFS_LOCK_GIANT() */ #define PFBAK 4 #define PFFOR 4 #define PAGEORDER_SIZE (PFBAK+PFFOR) static int prefault_pageorder[] = { -1 * PAGE_SIZE, 1 * PAGE_SIZE, -2 * PAGE_SIZE, 2 * PAGE_SIZE, -3 * PAGE_SIZE, 3 * PAGE_SIZE, -4 * PAGE_SIZE, 4 * PAGE_SIZE }; static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *); static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t); #define VM_FAULT_READ_AHEAD 8 #define VM_FAULT_READ_BEHIND 7 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1) struct faultstate { vm_page_t m; vm_object_t object; vm_pindex_t pindex; vm_page_t first_m; vm_object_t first_object; vm_pindex_t first_pindex; vm_map_t map; vm_map_entry_t entry; int lookup_still_valid; struct vnode *vp; }; static inline void release_page(struct faultstate *fs) { - vm_page_lock_queues(); vm_page_wakeup(fs->m); + vm_page_lock_queues(); vm_page_deactivate(fs->m); vm_page_unlock_queues(); fs->m = NULL; } static inline void unlock_map(struct faultstate *fs) { if (fs->lookup_still_valid) { vm_map_lookup_done(fs->map, fs->entry); fs->lookup_still_valid = FALSE; } } static void unlock_and_deallocate(struct faultstate *fs) { vm_object_pip_wakeup(fs->object); VM_OBJECT_UNLOCK(fs->object); if (fs->object != fs->first_object) { VM_OBJECT_LOCK(fs->first_object); vm_page_lock_queues(); vm_page_free(fs->first_m); vm_page_unlock_queues(); vm_object_pip_wakeup(fs->first_object); VM_OBJECT_UNLOCK(fs->first_object); fs->first_m = NULL; } vm_object_deallocate(fs->first_object); unlock_map(fs); if (fs->vp != NULL) { int vfslocked; vfslocked = VFS_LOCK_GIANT(fs->vp->v_mount); vput(fs->vp); fs->vp = NULL; VFS_UNLOCK_GIANT(vfslocked); } } /* * TRYPAGER - used by vm_fault to calculate whether the pager for the * current object *might* contain the page. * * default objects are zero-fill, there is no real pager. */ #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \ (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired)) /* * vm_fault: * * Handle a page fault occurring at the given address, * requiring the given permissions, in the map specified. * If successful, the page is inserted into the * associated physical map. * * NOTE: the given address should be truncated to the * proper page address. * * KERN_SUCCESS is returned if the page fault is handled; otherwise, * a standard error specifying why the fault is fatal is returned. * * * The map in question must be referenced, and remains so. * Caller may hold no locks. */ int vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) { vm_prot_t prot; int is_first_object_locked, result; boolean_t growstack, wired; int map_generation; vm_object_t next_object; vm_page_t marray[VM_FAULT_READ]; int hardfault; int faultcount; struct faultstate fs; hardfault = 0; growstack = TRUE; atomic_add_int(&cnt.v_vm_faults, 1); RetryFault:; /* * Find the backing store object and offset into it to begin the * search. */ fs.map = map; result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); if (result != KERN_SUCCESS) { if (result != KERN_PROTECTION_FAILURE || (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) { if (growstack && result == KERN_INVALID_ADDRESS && map != kernel_map && curproc != NULL) { result = vm_map_growstack(curproc, vaddr); if (result != KERN_SUCCESS) return (KERN_FAILURE); growstack = FALSE; goto RetryFault; } return (result); } /* * If we are user-wiring a r/w segment, and it is COW, then * we need to do the COW operation. Note that we don't COW * currently RO sections now, because it is NOT desirable * to COW .text. We simply keep .text from ever being COW'ed * and take the heat that one cannot debug wired .text sections. */ result = vm_map_lookup(&fs.map, vaddr, VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE, &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); if (result != KERN_SUCCESS) return (result); /* * If we don't COW now, on a user wire, the user will never * be able to write to the mapping. If we don't make this * restriction, the bookkeeping would be nearly impossible. * * XXX The following assignment modifies the map without * holding a write lock on it. */ if ((fs.entry->protection & VM_PROT_WRITE) == 0) fs.entry->max_protection &= ~VM_PROT_WRITE; } map_generation = fs.map->timestamp; if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { panic("vm_fault: fault on nofault entry, addr: %lx", (u_long)vaddr); } /* * Make a reference to this object to prevent its disposal while we * are messing with it. Once we have the reference, the map is free * to be diddled. Since objects reference their shadows (and copies), * they will stay around as well. * * Bump the paging-in-progress count to prevent size changes (e.g. * truncation operations) during I/O. This must be done after * obtaining the vnode lock in order to avoid possible deadlocks. * * XXX vnode_pager_lock() can block without releasing the map lock. */ if (fs.first_object->flags & OBJ_NEEDGIANT) mtx_lock(&Giant); VM_OBJECT_LOCK(fs.first_object); vm_object_reference_locked(fs.first_object); fs.vp = vnode_pager_lock(fs.first_object); KASSERT(fs.vp == NULL || !fs.map->system_map, ("vm_fault: vnode-backed object mapped by system map")); KASSERT((fs.first_object->flags & OBJ_NEEDGIANT) == 0 || !fs.map->system_map, ("vm_fault: Object requiring giant mapped by system map")); if (fs.first_object->flags & OBJ_NEEDGIANT) mtx_unlock(&Giant); vm_object_pip_add(fs.first_object, 1); fs.lookup_still_valid = TRUE; if (wired) fault_type = prot; fs.first_m = NULL; /* * Search for the page at object/offset. */ fs.object = fs.first_object; fs.pindex = fs.first_pindex; while (TRUE) { /* * If the object is dead, we stop here */ if (fs.object->flags & OBJ_DEAD) { unlock_and_deallocate(&fs); return (KERN_PROTECTION_FAILURE); } /* * See if page is resident */ fs.m = vm_page_lookup(fs.object, fs.pindex); if (fs.m != NULL) { int queue; /* * check for page-based copy on write. * We check fs.object == fs.first_object so * as to ensure the legacy COW mechanism is * used when the page in question is part of * a shadow object. Otherwise, vm_page_cowfault() * removes the page from the backing object, * which is not what we want. */ vm_page_lock_queues(); if ((fs.m->cow) && (fault_type & VM_PROT_WRITE) && (fs.object == fs.first_object)) { vm_page_cowfault(fs.m); vm_page_unlock_queues(); unlock_and_deallocate(&fs); goto RetryFault; } /* * Wait/Retry if the page is busy. We have to do this * if the page is busy via either VPO_BUSY or * vm_page_t->busy because the vm_pager may be using * vm_page_t->busy for pageouts ( and even pageins if * it is the vnode pager ), and we could end up trying * to pagein and pageout the same page simultaneously. * * We can theoretically allow the busy case on a read * fault if the page is marked valid, but since such * pages are typically already pmap'd, putting that * special case in might be more effort then it is * worth. We cannot under any circumstances mess * around with a vm_page_t->busy page except, perhaps, * to pmap it. */ if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) { vm_page_unlock_queues(); VM_OBJECT_UNLOCK(fs.object); if (fs.object != fs.first_object) { VM_OBJECT_LOCK(fs.first_object); vm_page_lock_queues(); vm_page_free(fs.first_m); vm_page_unlock_queues(); vm_object_pip_wakeup(fs.first_object); VM_OBJECT_UNLOCK(fs.first_object); fs.first_m = NULL; } unlock_map(&fs); if (fs.vp != NULL) { int vfslck; vfslck = VFS_LOCK_GIANT(fs.vp->v_mount); vput(fs.vp); fs.vp = NULL; VFS_UNLOCK_GIANT(vfslck); } VM_OBJECT_LOCK(fs.object); if (fs.m == vm_page_lookup(fs.object, fs.pindex)) { vm_page_sleep_if_busy(fs.m, TRUE, "vmpfw"); } vm_object_pip_wakeup(fs.object); VM_OBJECT_UNLOCK(fs.object); atomic_add_int(&cnt.v_intrans, 1); vm_object_deallocate(fs.first_object); goto RetryFault; } queue = fs.m->queue; vm_pageq_remove_nowakeup(fs.m); if (VM_PAGE_RESOLVEQUEUE(fs.m, queue) == PQ_CACHE && vm_page_count_severe()) { vm_page_activate(fs.m); vm_page_unlock_queues(); unlock_and_deallocate(&fs); VM_WAITPFAULT; goto RetryFault; } /* * Mark page busy for other processes, and the * pagedaemon. If it still isn't completely valid * (readable), jump to readrest, else break-out ( we * found the page ). */ vm_page_busy(fs.m); vm_page_unlock_queues(); if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && fs.m->object != kernel_object && fs.m->object != kmem_object) { goto readrest; } break; } /* * Page is not resident, If this is the search termination * or the pager might contain the page, allocate a new page. */ if (TRYPAGER || fs.object == fs.first_object) { if (fs.pindex >= fs.object->size) { unlock_and_deallocate(&fs); return (KERN_PROTECTION_FAILURE); } /* * Allocate a new page for this object/offset pair. */ fs.m = NULL; if (!vm_page_count_severe()) { fs.m = vm_page_alloc(fs.object, fs.pindex, (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO); } if (fs.m == NULL) { unlock_and_deallocate(&fs); VM_WAITPFAULT; goto RetryFault; } } readrest: /* * We have found a valid page or we have allocated a new page. * The page thus may not be valid or may not be entirely * valid. * * Attempt to fault-in the page if there is a chance that the * pager has it, and potentially fault in additional pages * at the same time. */ if (TRYPAGER) { int rv; int reqpage; int ahead, behind; u_char behavior = vm_map_entry_behavior(fs.entry); if (behavior == MAP_ENTRY_BEHAV_RANDOM) { ahead = 0; behind = 0; } else { behind = (vaddr - fs.entry->start) >> PAGE_SHIFT; if (behind > VM_FAULT_READ_BEHIND) behind = VM_FAULT_READ_BEHIND; ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1; if (ahead > VM_FAULT_READ_AHEAD) ahead = VM_FAULT_READ_AHEAD; } is_first_object_locked = FALSE; if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || (behavior != MAP_ENTRY_BEHAV_RANDOM && fs.pindex >= fs.entry->lastr && fs.pindex < fs.entry->lastr + VM_FAULT_READ)) && (fs.first_object == fs.object || (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) && fs.first_object->type != OBJT_DEVICE) { vm_pindex_t firstpindex, tmppindex; if (fs.first_pindex < 2 * VM_FAULT_READ) firstpindex = 0; else firstpindex = fs.first_pindex - 2 * VM_FAULT_READ; vm_page_lock_queues(); /* * note: partially valid pages cannot be * included in the lookahead - NFS piecemeal * writes will barf on it badly. */ for (tmppindex = fs.first_pindex - 1; tmppindex >= firstpindex; --tmppindex) { vm_page_t mt; mt = vm_page_lookup(fs.first_object, tmppindex); if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) break; if (mt->busy || (mt->oflags & VPO_BUSY) || (mt->flags & (PG_FICTITIOUS | PG_UNMANAGED)) || mt->hold_count || mt->wire_count) continue; pmap_remove_all(mt); if (mt->dirty) { vm_page_deactivate(mt); } else { vm_page_cache(mt); } } vm_page_unlock_queues(); ahead += behind; behind = 0; } if (is_first_object_locked) VM_OBJECT_UNLOCK(fs.first_object); /* * now we find out if any other pages should be paged * in at this time this routine checks to see if the * pages surrounding this fault reside in the same * object as the page for this fault. If they do, * then they are faulted in also into the object. The * array "marray" returned contains an array of * vm_page_t structs where one of them is the * vm_page_t passed to the routine. The reqpage * return value is the index into the marray for the * vm_page_t passed to the routine. * * fs.m plus the additional pages are VPO_BUSY'd. * * XXX vm_fault_additional_pages() can block * without releasing the map lock. */ faultcount = vm_fault_additional_pages( fs.m, behind, ahead, marray, &reqpage); /* * update lastr imperfectly (we do not know how much * getpages will actually read), but good enough. * * XXX The following assignment modifies the map * without holding a write lock on it. */ fs.entry->lastr = fs.pindex + faultcount - behind; /* * Call the pager to retrieve the data, if any, after * releasing the lock on the map. We hold a ref on * fs.object and the pages are VPO_BUSY'd. */ unlock_map(&fs); rv = faultcount ? vm_pager_get_pages(fs.object, marray, faultcount, reqpage) : VM_PAGER_FAIL; if (rv == VM_PAGER_OK) { /* * Found the page. Leave it busy while we play * with it. */ /* * Relookup in case pager changed page. Pager * is responsible for disposition of old page * if moved. */ fs.m = vm_page_lookup(fs.object, fs.pindex); if (!fs.m) { unlock_and_deallocate(&fs); goto RetryFault; } hardfault++; break; /* break to PAGE HAS BEEN FOUND */ } /* * Remove the bogus page (which does not exist at this * object/offset); before doing so, we must get back * our object lock to preserve our invariant. * * Also wake up any other process that may want to bring * in this page. * * If this is the top-level object, we must leave the * busy page to prevent another process from rushing * past us, and inserting the page in that object at * the same time that we are. */ if (rv == VM_PAGER_ERROR) printf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm); /* * Data outside the range of the pager or an I/O error */ /* * XXX - the check for kernel_map is a kludge to work * around having the machine panic on a kernel space * fault w/ I/O error. */ if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) { vm_page_lock_queues(); vm_page_free(fs.m); vm_page_unlock_queues(); fs.m = NULL; unlock_and_deallocate(&fs); return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); } if (fs.object != fs.first_object) { vm_page_lock_queues(); vm_page_free(fs.m); vm_page_unlock_queues(); fs.m = NULL; /* * XXX - we cannot just fall out at this * point, m has been freed and is invalid! */ } } /* * We get here if the object has default pager (or unwiring) * or the pager doesn't have the page. */ if (fs.object == fs.first_object) fs.first_m = fs.m; /* * Move on to the next object. Lock the next object before * unlocking the current one. */ fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); next_object = fs.object->backing_object; if (next_object == NULL) { /* * If there's no object left, fill the page in the top * object with zeros. */ if (fs.object != fs.first_object) { vm_object_pip_wakeup(fs.object); VM_OBJECT_UNLOCK(fs.object); fs.object = fs.first_object; fs.pindex = fs.first_pindex; fs.m = fs.first_m; VM_OBJECT_LOCK(fs.object); } fs.first_m = NULL; /* * Zero the page if necessary and mark it valid. */ if ((fs.m->flags & PG_ZERO) == 0) { pmap_zero_page(fs.m); } else { atomic_add_int(&cnt.v_ozfod, 1); } atomic_add_int(&cnt.v_zfod, 1); fs.m->valid = VM_PAGE_BITS_ALL; break; /* break to PAGE HAS BEEN FOUND */ } else { KASSERT(fs.object != next_object, ("object loop %p", next_object)); VM_OBJECT_LOCK(next_object); vm_object_pip_add(next_object, 1); if (fs.object != fs.first_object) vm_object_pip_wakeup(fs.object); VM_OBJECT_UNLOCK(fs.object); fs.object = next_object; } } KASSERT((fs.m->oflags & VPO_BUSY) != 0, ("vm_fault: not busy after main loop")); /* * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock * is held.] */ /* * If the page is being written, but isn't already owned by the * top-level object, we have to copy it into a new page owned by the * top-level object. */ if (fs.object != fs.first_object) { /* * We only really need to copy if we want to write it. */ if (fault_type & VM_PROT_WRITE) { /* * This allows pages to be virtually copied from a * backing_object into the first_object, where the * backing object has no other refs to it, and cannot * gain any more refs. Instead of a bcopy, we just * move the page from the backing object to the * first object. Note that we must mark the page * dirty in the first object so that it will go out * to swap when needed. */ is_first_object_locked = FALSE; if ( /* * Only one shadow object */ (fs.object->shadow_count == 1) && /* * No COW refs, except us */ (fs.object->ref_count == 1) && /* * No one else can look this object up */ (fs.object->handle == NULL) && /* * No other ways to look the object up */ ((fs.object->type == OBJT_DEFAULT) || (fs.object->type == OBJT_SWAP)) && (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) && /* * We don't chase down the shadow chain */ fs.object == fs.first_object->backing_object) { vm_page_lock_queues(); /* * get rid of the unnecessary page */ vm_page_free(fs.first_m); /* * grab the page and put it into the * process'es object. The page is * automatically made dirty. */ vm_page_rename(fs.m, fs.first_object, fs.first_pindex); vm_page_busy(fs.m); vm_page_unlock_queues(); fs.first_m = fs.m; fs.m = NULL; atomic_add_int(&cnt.v_cow_optim, 1); } else { /* * Oh, well, lets copy it. */ pmap_copy_page(fs.m, fs.first_m); fs.first_m->valid = VM_PAGE_BITS_ALL; } if (fs.m) { /* * We no longer need the old page or object. */ release_page(&fs); } /* * fs.object != fs.first_object due to above * conditional */ vm_object_pip_wakeup(fs.object); VM_OBJECT_UNLOCK(fs.object); /* * Only use the new page below... */ fs.object = fs.first_object; fs.pindex = fs.first_pindex; fs.m = fs.first_m; if (!is_first_object_locked) VM_OBJECT_LOCK(fs.object); atomic_add_int(&cnt.v_cow_faults, 1); } else { prot &= ~VM_PROT_WRITE; } } /* * We must verify that the maps have not changed since our last * lookup. */ if (!fs.lookup_still_valid) { vm_object_t retry_object; vm_pindex_t retry_pindex; vm_prot_t retry_prot; if (!vm_map_trylock_read(fs.map)) { release_page(&fs); unlock_and_deallocate(&fs); goto RetryFault; } fs.lookup_still_valid = TRUE; if (fs.map->timestamp != map_generation) { result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); /* * If we don't need the page any longer, put it on the inactive * list (the easiest thing to do here). If no one needs it, * pageout will grab it eventually. */ if (result != KERN_SUCCESS) { release_page(&fs); unlock_and_deallocate(&fs); /* * If retry of map lookup would have blocked then * retry fault from start. */ if (result == KERN_FAILURE) goto RetryFault; return (result); } if ((retry_object != fs.first_object) || (retry_pindex != fs.first_pindex)) { release_page(&fs); unlock_and_deallocate(&fs); goto RetryFault; } /* * Check whether the protection has changed or the object has * been copied while we left the map unlocked. Changing from * read to write permission is OK - we leave the page * write-protected, and catch the write fault. Changing from * write to read permission means that we can't mark the page * write-enabled after all. */ prot &= retry_prot; } } if (prot & VM_PROT_WRITE) { vm_page_lock_queues(); vm_page_flag_set(fs.m, PG_WRITEABLE); vm_page_unlock_queues(); vm_object_set_writeable_dirty(fs.object); /* * If the fault is a write, we know that this page is being * written NOW so dirty it explicitly to save on * pmap_is_modified() calls later. * * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC * if the page is already dirty to prevent data written with * the expectation of being synced from not being synced. * Likewise if this entry does not request NOSYNC then make * sure the page isn't marked NOSYNC. Applications sharing * data should use the same flags to avoid ping ponging. * * Also tell the backing pager, if any, that it should remove * any swap backing since the page is now dirty. */ if (fs.entry->eflags & MAP_ENTRY_NOSYNC) { if (fs.m->dirty == 0) fs.m->oflags |= VPO_NOSYNC; } else { fs.m->oflags &= ~VPO_NOSYNC; } if (fault_flags & VM_FAULT_DIRTY) { vm_page_dirty(fs.m); vm_pager_page_unswapped(fs.m); } } /* * Page had better still be busy */ KASSERT(fs.m->oflags & VPO_BUSY, ("vm_fault: page %p not busy!", fs.m)); /* * Sanity check: page must be completely valid or it is not fit to * map into user space. vm_pager_get_pages() ensures this. */ if (fs.m->valid != VM_PAGE_BITS_ALL) { vm_page_zero_invalid(fs.m, TRUE); printf("Warning: page %p partially invalid on fault\n", fs.m); } VM_OBJECT_UNLOCK(fs.object); /* * Put this page into the physical map. We had to do the unlock above * because pmap_enter() may sleep. We don't put the page * back on the active queue until later so that the pageout daemon * won't find it (yet). */ pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired); if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) { vm_fault_prefault(fs.map->pmap, vaddr, fs.entry); } VM_OBJECT_LOCK(fs.object); vm_page_lock_queues(); vm_page_flag_set(fs.m, PG_REFERENCED); /* * If the page is not wired down, then put it where the pageout daemon * can find it. */ if (fault_flags & VM_FAULT_WIRE_MASK) { if (wired) vm_page_wire(fs.m); else vm_page_unwire(fs.m, 1); } else { vm_page_activate(fs.m); } - vm_page_wakeup(fs.m); vm_page_unlock_queues(); + vm_page_wakeup(fs.m); /* * Unlock everything, and return */ unlock_and_deallocate(&fs); PROC_LOCK(curproc); if ((curproc->p_sflag & PS_INMEM) && curproc->p_stats) { if (hardfault) { curproc->p_stats->p_ru.ru_majflt++; } else { curproc->p_stats->p_ru.ru_minflt++; } } PROC_UNLOCK(curproc); return (KERN_SUCCESS); } /* * vm_fault_prefault provides a quick way of clustering * pagefaults into a processes address space. It is a "cousin" * of vm_map_pmap_enter, except it runs at page fault time instead * of mmap time. */ static void vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry) { int i; vm_offset_t addr, starta; vm_pindex_t pindex; vm_page_t m; vm_object_t object; if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) return; object = entry->object.vm_object; starta = addra - PFBAK * PAGE_SIZE; if (starta < entry->start) { starta = entry->start; } else if (starta > addra) { starta = 0; } for (i = 0; i < PAGEORDER_SIZE; i++) { vm_object_t backing_object, lobject; addr = addra + prefault_pageorder[i]; if (addr > addra + (PFFOR * PAGE_SIZE)) addr = 0; if (addr < starta || addr >= entry->end) continue; if (!pmap_is_prefaultable(pmap, addr)) continue; pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; lobject = object; VM_OBJECT_LOCK(lobject); while ((m = vm_page_lookup(lobject, pindex)) == NULL && lobject->type == OBJT_DEFAULT && (backing_object = lobject->backing_object) != NULL) { if (lobject->backing_object_offset & PAGE_MASK) break; pindex += lobject->backing_object_offset >> PAGE_SHIFT; VM_OBJECT_LOCK(backing_object); VM_OBJECT_UNLOCK(lobject); lobject = backing_object; } /* * give-up when a page is not in memory */ if (m == NULL) { VM_OBJECT_UNLOCK(lobject); break; } if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && (m->busy == 0) && (m->flags & PG_FICTITIOUS) == 0) { vm_page_lock_queues(); if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) vm_page_deactivate(m); pmap_enter_quick(pmap, addr, m, entry->protection); vm_page_unlock_queues(); } VM_OBJECT_UNLOCK(lobject); } } /* * vm_fault_quick: * * Ensure that the requested virtual address, which may be in userland, * is valid. Fault-in the page if necessary. Return -1 on failure. */ int vm_fault_quick(caddr_t v, int prot) { int r; if (prot & VM_PROT_WRITE) r = subyte(v, fubyte(v)); else r = fubyte(v); return(r); } /* * vm_fault_wire: * * Wire down a range of virtual addresses in a map. */ int vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, boolean_t user_wire, boolean_t fictitious) { vm_offset_t va; int rv; /* * We simulate a fault to get the page and enter it in the physical * map. For user wiring, we only ask for read access on currently * read-only sections. */ for (va = start; va < end; va += PAGE_SIZE) { rv = vm_fault(map, va, user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE, user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING); if (rv) { if (va != start) vm_fault_unwire(map, start, va, fictitious); return (rv); } } return (KERN_SUCCESS); } /* * vm_fault_unwire: * * Unwire a range of virtual addresses in a map. */ void vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, boolean_t fictitious) { vm_paddr_t pa; vm_offset_t va; pmap_t pmap; pmap = vm_map_pmap(map); /* * Since the pages are wired down, we must be able to get their * mappings from the physical map system. */ for (va = start; va < end; va += PAGE_SIZE) { pa = pmap_extract(pmap, va); if (pa != 0) { pmap_change_wiring(pmap, va, FALSE); if (!fictitious) { vm_page_lock_queues(); vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); vm_page_unlock_queues(); } } } } /* * Routine: * vm_fault_copy_entry * Function: * Copy all of the pages from a wired-down map entry to another. * * In/out conditions: * The source and destination maps must be locked for write. * The source map entry must be wired down (or be a sharing map * entry corresponding to a main map entry that is wired down). */ void vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry) vm_map_t dst_map; vm_map_t src_map; vm_map_entry_t dst_entry; vm_map_entry_t src_entry; { vm_object_t backing_object, dst_object, object; vm_object_t src_object; vm_ooffset_t dst_offset; vm_ooffset_t src_offset; vm_pindex_t pindex; vm_prot_t prot; vm_offset_t vaddr; vm_page_t dst_m; vm_page_t src_m; #ifdef lint src_map++; #endif /* lint */ src_object = src_entry->object.vm_object; src_offset = src_entry->offset; /* * Create the top-level object for the destination entry. (Doesn't * actually shadow anything - we copy the pages directly.) */ dst_object = vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(dst_entry->end - dst_entry->start)); VM_OBJECT_LOCK(dst_object); dst_entry->object.vm_object = dst_object; dst_entry->offset = 0; prot = dst_entry->max_protection; /* * Loop through all of the pages in the entry's range, copying each * one from the source object (it should be there) to the destination * object. */ for (vaddr = dst_entry->start, dst_offset = 0; vaddr < dst_entry->end; vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { /* * Allocate a page in the destination object */ do { dst_m = vm_page_alloc(dst_object, OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); if (dst_m == NULL) { VM_OBJECT_UNLOCK(dst_object); VM_WAIT; VM_OBJECT_LOCK(dst_object); } } while (dst_m == NULL); /* * Find the page in the source object, and copy it in. * (Because the source is wired down, the page will be in * memory.) */ VM_OBJECT_LOCK(src_object); object = src_object; pindex = 0; while ((src_m = vm_page_lookup(object, pindex + OFF_TO_IDX(dst_offset + src_offset))) == NULL && (src_entry->protection & VM_PROT_WRITE) == 0 && (backing_object = object->backing_object) != NULL) { /* * Allow fallback to backing objects if we are reading. */ VM_OBJECT_LOCK(backing_object); pindex += OFF_TO_IDX(object->backing_object_offset); VM_OBJECT_UNLOCK(object); object = backing_object; } if (src_m == NULL) panic("vm_fault_copy_wired: page missing"); pmap_copy_page(src_m, dst_m); VM_OBJECT_UNLOCK(object); dst_m->valid = VM_PAGE_BITS_ALL; VM_OBJECT_UNLOCK(dst_object); /* * Enter it in the pmap... */ pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE); VM_OBJECT_LOCK(dst_object); vm_page_lock_queues(); if ((prot & VM_PROT_WRITE) != 0) vm_page_flag_set(dst_m, PG_WRITEABLE); /* * Mark it no longer busy, and put it on the active list. */ vm_page_activate(dst_m); - vm_page_wakeup(dst_m); vm_page_unlock_queues(); + vm_page_wakeup(dst_m); } VM_OBJECT_UNLOCK(dst_object); } /* * This routine checks around the requested page for other pages that * might be able to be faulted in. This routine brackets the viable * pages for the pages to be paged in. * * Inputs: * m, rbehind, rahead * * Outputs: * marray (array of vm_page_t), reqpage (index of requested page) * * Return value: * number of pages in marray * * This routine can't block. */ static int vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage) vm_page_t m; int rbehind; int rahead; vm_page_t *marray; int *reqpage; { int i,j; vm_object_t object; vm_pindex_t pindex, startpindex, endpindex, tpindex; vm_page_t rtm; int cbehind, cahead; VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); object = m->object; pindex = m->pindex; /* * we don't fault-ahead for device pager */ if (object->type == OBJT_DEVICE) { *reqpage = 0; marray[0] = m; return 1; } /* * if the requested page is not available, then give up now */ if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { return 0; } if ((cbehind == 0) && (cahead == 0)) { *reqpage = 0; marray[0] = m; return 1; } if (rahead > cahead) { rahead = cahead; } if (rbehind > cbehind) { rbehind = cbehind; } /* * try to do any readahead that we might have free pages for. */ if ((rahead + rbehind) > ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) { pagedaemon_wakeup(); marray[0] = m; *reqpage = 0; return 1; } /* * scan backward for the read behind pages -- in memory */ if (pindex > 0) { if (rbehind > pindex) { rbehind = pindex; startpindex = 0; } else { startpindex = pindex - rbehind; } if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL && rtm->pindex >= startpindex) startpindex = rtm->pindex + 1; for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) { rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); if (rtm == NULL) { vm_page_lock_queues(); for (j = 0; j < i; j++) { vm_page_free(marray[j]); } vm_page_unlock_queues(); marray[0] = m; *reqpage = 0; return 1; } marray[i] = rtm; } } else { startpindex = 0; i = 0; } marray[i] = m; /* page offset of the required page */ *reqpage = i; tpindex = pindex + 1; i++; /* * scan forward for the read ahead pages */ endpindex = tpindex + rahead; if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex) endpindex = rtm->pindex; if (endpindex > object->size) endpindex = object->size; for (; tpindex < endpindex; i++, tpindex++) { rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); if (rtm == NULL) { break; } marray[i] = rtm; } /* return number of bytes of pages */ return i; } Index: head/sys/vm/vm_glue.c =================================================================== --- head/sys/vm/vm_glue.c (revision 163621) +++ head/sys/vm/vm_glue.c (revision 163622) @@ -1,1016 +1,1016 @@ /*- * 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. * 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_glue.c 8.6 (Berkeley) 1/5/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * 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. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include "opt_kstack_pages.h" #include "opt_kstack_max_pages.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include extern int maxslp; /* * System initialization * * Note: proc0 from proc.h */ static void vm_init_limits(void *); SYSINIT(vm_limits, SI_SUB_VM_CONF, SI_ORDER_FIRST, vm_init_limits, &proc0) /* * THIS MUST BE THE LAST INITIALIZATION ITEM!!! * * Note: run scheduling should be divorced from the vm system. */ static void scheduler(void *); SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_ANY, scheduler, NULL) #ifndef NO_SWAPPING static void swapout(struct proc *); #endif static volatile int proc0_rescan; /* * MPSAFE * * WARNING! This code calls vm_map_check_protection() which only checks * the associated vm_map_entry range. It does not determine whether the * contents of the memory is actually readable or writable. In most cases * just checking the vm_map_entry is sufficient within the kernel's address * space. */ int kernacc(addr, len, rw) void *addr; int len, rw; { boolean_t rv; vm_offset_t saddr, eaddr; vm_prot_t prot; KASSERT((rw & ~VM_PROT_ALL) == 0, ("illegal ``rw'' argument to kernacc (%x)\n", rw)); if ((vm_offset_t)addr + len > kernel_map->max_offset || (vm_offset_t)addr + len < (vm_offset_t)addr) return (FALSE); prot = rw; saddr = trunc_page((vm_offset_t)addr); eaddr = round_page((vm_offset_t)addr + len); vm_map_lock_read(kernel_map); rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot); vm_map_unlock_read(kernel_map); return (rv == TRUE); } /* * MPSAFE * * WARNING! This code calls vm_map_check_protection() which only checks * the associated vm_map_entry range. It does not determine whether the * contents of the memory is actually readable or writable. vmapbuf(), * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be * used in conjuction with this call. */ int useracc(addr, len, rw) void *addr; int len, rw; { boolean_t rv; vm_prot_t prot; vm_map_t map; KASSERT((rw & ~VM_PROT_ALL) == 0, ("illegal ``rw'' argument to useracc (%x)\n", rw)); prot = rw; map = &curproc->p_vmspace->vm_map; if ((vm_offset_t)addr + len > vm_map_max(map) || (vm_offset_t)addr + len < (vm_offset_t)addr) { return (FALSE); } vm_map_lock_read(map); rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len), prot); vm_map_unlock_read(map); return (rv == TRUE); } int vslock(void *addr, size_t len) { vm_offset_t end, last, start; vm_size_t npages; int error; last = (vm_offset_t)addr + len; start = trunc_page((vm_offset_t)addr); end = round_page(last); if (last < (vm_offset_t)addr || end < (vm_offset_t)addr) return (EINVAL); npages = atop(end - start); if (npages > vm_page_max_wired) return (ENOMEM); PROC_LOCK(curproc); if (ptoa(npages + pmap_wired_count(vm_map_pmap(&curproc->p_vmspace->vm_map))) > lim_cur(curproc, RLIMIT_MEMLOCK)) { PROC_UNLOCK(curproc); return (ENOMEM); } PROC_UNLOCK(curproc); #if 0 /* * XXX - not yet * * The limit for transient usage of wired pages should be * larger than for "permanent" wired pages (mlock()). * * Also, the sysctl code, which is the only present user * of vslock(), does a hard loop on EAGAIN. */ if (npages + cnt.v_wire_count > vm_page_max_wired) return (EAGAIN); #endif error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end, VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); /* * Return EFAULT on error to match copy{in,out}() behaviour * rather than returning ENOMEM like mlock() would. */ return (error == KERN_SUCCESS ? 0 : EFAULT); } void vsunlock(void *addr, size_t len) { /* Rely on the parameter sanity checks performed by vslock(). */ (void)vm_map_unwire(&curproc->p_vmspace->vm_map, trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len), VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); } /* * Pin the page contained within the given object at the given offset. If the * page is not resident, allocate and load it using the given object's pager. * Return the pinned page if successful; otherwise, return NULL. */ static vm_page_t vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset) { vm_page_t m, ma[1]; vm_pindex_t pindex; int rv; VM_OBJECT_LOCK(object); pindex = OFF_TO_IDX(offset); m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); if ((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) { ma[0] = m; rv = vm_pager_get_pages(object, ma, 1, 0); m = vm_page_lookup(object, pindex); if (m == NULL) goto out; if (m->valid == 0 || rv != VM_PAGER_OK) { vm_page_lock_queues(); vm_page_free(m); vm_page_unlock_queues(); m = NULL; goto out; } } vm_page_lock_queues(); vm_page_hold(m); - vm_page_wakeup(m); vm_page_unlock_queues(); + vm_page_wakeup(m); out: VM_OBJECT_UNLOCK(object); return (m); } /* * Return a CPU private mapping to the page at the given offset within the * given object. The page is pinned before it is mapped. */ struct sf_buf * vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset) { vm_page_t m; m = vm_imgact_hold_page(object, offset); if (m == NULL) return (NULL); sched_pin(); return (sf_buf_alloc(m, SFB_CPUPRIVATE)); } /* * Destroy the given CPU private mapping and unpin the page that it mapped. */ void vm_imgact_unmap_page(struct sf_buf *sf) { vm_page_t m; m = sf_buf_page(sf); sf_buf_free(sf); sched_unpin(); vm_page_lock_queues(); vm_page_unhold(m); vm_page_unlock_queues(); } #ifndef KSTACK_MAX_PAGES #define KSTACK_MAX_PAGES 32 #endif /* * Create the kernel stack (including pcb for i386) for a new thread. * This routine directly affects the fork perf for a process and * create performance for a thread. */ void vm_thread_new(struct thread *td, int pages) { vm_object_t ksobj; vm_offset_t ks; vm_page_t m, ma[KSTACK_MAX_PAGES]; int i; /* Bounds check */ if (pages <= 1) pages = KSTACK_PAGES; else if (pages > KSTACK_MAX_PAGES) pages = KSTACK_MAX_PAGES; /* * Allocate an object for the kstack. */ ksobj = vm_object_allocate(OBJT_DEFAULT, pages); td->td_kstack_obj = ksobj; /* * Get a kernel virtual address for this thread's kstack. */ ks = kmem_alloc_nofault(kernel_map, (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE); if (ks == 0) panic("vm_thread_new: kstack allocation failed"); if (KSTACK_GUARD_PAGES != 0) { pmap_qremove(ks, KSTACK_GUARD_PAGES); ks += KSTACK_GUARD_PAGES * PAGE_SIZE; } td->td_kstack = ks; /* * Knowing the number of pages allocated is useful when you * want to deallocate them. */ td->td_kstack_pages = pages; /* * For the length of the stack, link in a real page of ram for each * page of stack. */ VM_OBJECT_LOCK(ksobj); for (i = 0; i < pages; i++) { /* * Get a kernel stack page. */ m = vm_page_grab(ksobj, i, VM_ALLOC_NOBUSY | VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED); ma[i] = m; m->valid = VM_PAGE_BITS_ALL; } VM_OBJECT_UNLOCK(ksobj); pmap_qenter(ks, ma, pages); } /* * Dispose of a thread's kernel stack. */ void vm_thread_dispose(struct thread *td) { vm_object_t ksobj; vm_offset_t ks; vm_page_t m; int i, pages; pages = td->td_kstack_pages; ksobj = td->td_kstack_obj; ks = td->td_kstack; pmap_qremove(ks, pages); VM_OBJECT_LOCK(ksobj); for (i = 0; i < pages; i++) { m = vm_page_lookup(ksobj, i); if (m == NULL) panic("vm_thread_dispose: kstack already missing?"); vm_page_lock_queues(); vm_page_unwire(m, 0); vm_page_free(m); vm_page_unlock_queues(); } VM_OBJECT_UNLOCK(ksobj); vm_object_deallocate(ksobj); kmem_free(kernel_map, ks - (KSTACK_GUARD_PAGES * PAGE_SIZE), (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE); } /* * Allow a thread's kernel stack to be paged out. */ void vm_thread_swapout(struct thread *td) { vm_object_t ksobj; vm_page_t m; int i, pages; cpu_thread_swapout(td); pages = td->td_kstack_pages; ksobj = td->td_kstack_obj; pmap_qremove(td->td_kstack, pages); VM_OBJECT_LOCK(ksobj); for (i = 0; i < pages; i++) { m = vm_page_lookup(ksobj, i); if (m == NULL) panic("vm_thread_swapout: kstack already missing?"); vm_page_lock_queues(); vm_page_dirty(m); vm_page_unwire(m, 0); vm_page_unlock_queues(); } VM_OBJECT_UNLOCK(ksobj); } /* * Bring the kernel stack for a specified thread back in. */ void vm_thread_swapin(struct thread *td) { vm_object_t ksobj; vm_page_t m, ma[KSTACK_MAX_PAGES]; int i, pages, rv; pages = td->td_kstack_pages; ksobj = td->td_kstack_obj; VM_OBJECT_LOCK(ksobj); for (i = 0; i < pages; i++) { m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY); if (m->valid != VM_PAGE_BITS_ALL) { rv = vm_pager_get_pages(ksobj, &m, 1, 0); if (rv != VM_PAGER_OK) panic("vm_thread_swapin: cannot get kstack for proc: %d", td->td_proc->p_pid); m = vm_page_lookup(ksobj, i); m->valid = VM_PAGE_BITS_ALL; } ma[i] = m; vm_page_lock_queues(); vm_page_wire(m); - vm_page_wakeup(m); vm_page_unlock_queues(); + vm_page_wakeup(m); } VM_OBJECT_UNLOCK(ksobj); pmap_qenter(td->td_kstack, ma, pages); cpu_thread_swapin(td); } /* * Set up a variable-sized alternate kstack. */ void vm_thread_new_altkstack(struct thread *td, int pages) { td->td_altkstack = td->td_kstack; td->td_altkstack_obj = td->td_kstack_obj; td->td_altkstack_pages = td->td_kstack_pages; vm_thread_new(td, pages); } /* * Restore the original kstack. */ void vm_thread_dispose_altkstack(struct thread *td) { vm_thread_dispose(td); td->td_kstack = td->td_altkstack; td->td_kstack_obj = td->td_altkstack_obj; td->td_kstack_pages = td->td_altkstack_pages; td->td_altkstack = 0; td->td_altkstack_obj = NULL; td->td_altkstack_pages = 0; } /* * Implement fork's actions on an address space. * Here we arrange for the address space to be copied or referenced, * allocate a user struct (pcb and kernel stack), then call the * machine-dependent layer to fill those in and make the new process * ready to run. The new process is set up so that it returns directly * to user mode to avoid stack copying and relocation problems. */ void vm_forkproc(td, p2, td2, flags) struct thread *td; struct proc *p2; struct thread *td2; int flags; { struct proc *p1 = td->td_proc; if ((flags & RFPROC) == 0) { /* * Divorce the memory, if it is shared, essentially * this changes shared memory amongst threads, into * COW locally. */ if ((flags & RFMEM) == 0) { if (p1->p_vmspace->vm_refcnt > 1) { vmspace_unshare(p1); } } cpu_fork(td, p2, td2, flags); return; } if (flags & RFMEM) { p2->p_vmspace = p1->p_vmspace; atomic_add_int(&p1->p_vmspace->vm_refcnt, 1); } while (vm_page_count_severe()) { VM_WAIT; } if ((flags & RFMEM) == 0) { p2->p_vmspace = vmspace_fork(p1->p_vmspace); if (p1->p_vmspace->vm_shm) shmfork(p1, p2); } /* * cpu_fork will copy and update the pcb, set up the kernel stack, * and make the child ready to run. */ cpu_fork(td, p2, td2, flags); } /* * Called after process has been wait(2)'ed apon and is being reaped. * The idea is to reclaim resources that we could not reclaim while * the process was still executing. */ void vm_waitproc(p) struct proc *p; { vmspace_exitfree(p); /* and clean-out the vmspace */ } /* * Set default limits for VM system. * Called for proc 0, and then inherited by all others. * * XXX should probably act directly on proc0. */ static void vm_init_limits(udata) void *udata; { struct proc *p = udata; struct plimit *limp; int rss_limit; /* * Set up the initial limits on process VM. Set the maximum resident * set size to be half of (reasonably) available memory. Since this * is a soft limit, it comes into effect only when the system is out * of memory - half of main memory helps to favor smaller processes, * and reduces thrashing of the object cache. */ limp = p->p_limit; limp->pl_rlimit[RLIMIT_STACK].rlim_cur = dflssiz; limp->pl_rlimit[RLIMIT_STACK].rlim_max = maxssiz; limp->pl_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz; limp->pl_rlimit[RLIMIT_DATA].rlim_max = maxdsiz; /* limit the limit to no less than 2MB */ rss_limit = max(cnt.v_free_count, 512); limp->pl_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit); limp->pl_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY; } void faultin(p) struct proc *p; { #ifdef NO_SWAPPING PROC_LOCK_ASSERT(p, MA_OWNED); if ((p->p_sflag & PS_INMEM) == 0) panic("faultin: proc swapped out with NO_SWAPPING!"); #else /* !NO_SWAPPING */ struct thread *td; PROC_LOCK_ASSERT(p, MA_OWNED); /* * If another process is swapping in this process, * just wait until it finishes. */ if (p->p_sflag & PS_SWAPPINGIN) msleep(&p->p_sflag, &p->p_mtx, PVM, "faultin", 0); else if ((p->p_sflag & PS_INMEM) == 0) { /* * Don't let another thread swap process p out while we are * busy swapping it in. */ ++p->p_lock; mtx_lock_spin(&sched_lock); p->p_sflag |= PS_SWAPPINGIN; mtx_unlock_spin(&sched_lock); PROC_UNLOCK(p); FOREACH_THREAD_IN_PROC(p, td) vm_thread_swapin(td); PROC_LOCK(p); mtx_lock_spin(&sched_lock); p->p_sflag &= ~PS_SWAPPINGIN; p->p_sflag |= PS_INMEM; FOREACH_THREAD_IN_PROC(p, td) { TD_CLR_SWAPPED(td); if (TD_CAN_RUN(td)) setrunnable(td); } mtx_unlock_spin(&sched_lock); wakeup(&p->p_sflag); /* Allow other threads to swap p out now. */ --p->p_lock; } #endif /* NO_SWAPPING */ } /* * This swapin algorithm attempts to swap-in processes only if there * is enough space for them. Of course, if a process waits for a long * time, it will be swapped in anyway. * * XXXKSE - process with the thread with highest priority counts.. * * Giant is held on entry. */ /* ARGSUSED*/ static void scheduler(dummy) void *dummy; { struct proc *p; struct thread *td; int pri; struct proc *pp; int ppri; mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED); mtx_unlock(&Giant); loop: if (vm_page_count_min()) { VM_WAIT; mtx_lock_spin(&sched_lock); proc0_rescan = 0; mtx_unlock_spin(&sched_lock); goto loop; } pp = NULL; ppri = INT_MIN; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { struct ksegrp *kg; if (p->p_sflag & (PS_INMEM | PS_SWAPPINGOUT | PS_SWAPPINGIN)) { continue; } mtx_lock_spin(&sched_lock); FOREACH_THREAD_IN_PROC(p, td) { /* * An otherwise runnable thread of a process * swapped out has only the TDI_SWAPPED bit set. * */ if (td->td_inhibitors == TDI_SWAPPED) { kg = td->td_ksegrp; pri = p->p_swtime + kg->kg_slptime; if ((p->p_sflag & PS_SWAPINREQ) == 0) { pri -= p->p_nice * 8; } /* * if this ksegrp is higher priority * and there is enough space, then select * this process instead of the previous * selection. */ if (pri > ppri) { pp = p; ppri = pri; } } } mtx_unlock_spin(&sched_lock); } sx_sunlock(&allproc_lock); /* * Nothing to do, back to sleep. */ if ((p = pp) == NULL) { mtx_lock_spin(&sched_lock); if (!proc0_rescan) { TD_SET_IWAIT(&thread0); mi_switch(SW_VOL, NULL); } proc0_rescan = 0; mtx_unlock_spin(&sched_lock); goto loop; } PROC_LOCK(p); /* * Another process may be bringing or may have already * brought this process in while we traverse all threads. * Or, this process may even be being swapped out again. */ if (p->p_sflag & (PS_INMEM | PS_SWAPPINGOUT | PS_SWAPPINGIN)) { PROC_UNLOCK(p); mtx_lock_spin(&sched_lock); proc0_rescan = 0; mtx_unlock_spin(&sched_lock); goto loop; } mtx_lock_spin(&sched_lock); p->p_sflag &= ~PS_SWAPINREQ; mtx_unlock_spin(&sched_lock); /* * We would like to bring someone in. (only if there is space). * [What checks the space? ] */ faultin(p); PROC_UNLOCK(p); mtx_lock_spin(&sched_lock); p->p_swtime = 0; proc0_rescan = 0; mtx_unlock_spin(&sched_lock); goto loop; } void kick_proc0(void) { struct thread *td = &thread0; if (TD_AWAITING_INTR(td)) { CTR2(KTR_INTR, "%s: setrunqueue %d", __func__, 0); TD_CLR_IWAIT(td); setrunqueue(td, SRQ_INTR); } else { proc0_rescan = 1; CTR2(KTR_INTR, "%s: state %d", __func__, td->td_state); } } #ifndef NO_SWAPPING /* * Swap_idle_threshold1 is the guaranteed swapped in time for a process */ static int swap_idle_threshold1 = 2; SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process"); /* * Swap_idle_threshold2 is the time that a process can be idle before * it will be swapped out, if idle swapping is enabled. */ static int swap_idle_threshold2 = 10; SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, &swap_idle_threshold2, 0, "Time before a process will be swapped out"); /* * Swapout is driven by the pageout daemon. Very simple, we find eligible * procs and unwire their u-areas. We try to always "swap" at least one * process in case we need the room for a swapin. * If any procs have been sleeping/stopped for at least maxslp seconds, * they are swapped. Else, we swap the longest-sleeping or stopped process, * if any, otherwise the longest-resident process. */ void swapout_procs(action) int action; { struct proc *p; struct thread *td; struct ksegrp *kg; int didswap = 0; retry: sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { struct vmspace *vm; int minslptime = 100000; /* * Watch out for a process in * creation. It may have no * address space or lock yet. */ mtx_lock_spin(&sched_lock); if (p->p_state == PRS_NEW) { mtx_unlock_spin(&sched_lock); continue; } mtx_unlock_spin(&sched_lock); /* * An aio daemon switches its * address space while running. * Perform a quick check whether * a process has P_SYSTEM. */ if ((p->p_flag & P_SYSTEM) != 0) continue; /* * Do not swapout a process that * is waiting for VM data * structures as there is a possible * deadlock. Test this first as * this may block. * * Lock the map until swapout * finishes, or a thread of this * process may attempt to alter * the map. */ vm = vmspace_acquire_ref(p); if (vm == NULL) continue; if (!vm_map_trylock(&vm->vm_map)) goto nextproc1; PROC_LOCK(p); if (p->p_lock != 0 || (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT) ) != 0) { goto nextproc2; } /* * only aiod changes vmspace, however it will be * skipped because of the if statement above checking * for P_SYSTEM */ if ((p->p_sflag & (PS_INMEM|PS_SWAPPINGOUT|PS_SWAPPINGIN)) != PS_INMEM) goto nextproc2; switch (p->p_state) { default: /* Don't swap out processes in any sort * of 'special' state. */ break; case PRS_NORMAL: mtx_lock_spin(&sched_lock); /* * do not swapout a realtime process * Check all the thread groups.. */ FOREACH_KSEGRP_IN_PROC(p, kg) { if (PRI_IS_REALTIME(kg->kg_pri_class)) goto nextproc; /* * Guarantee swap_idle_threshold1 * time in memory. */ if (kg->kg_slptime < swap_idle_threshold1) goto nextproc; /* * Do not swapout a process if it is * waiting on a critical event of some * kind or there is a thread whose * pageable memory may be accessed. * * This could be refined to support * swapping out a thread. */ FOREACH_THREAD_IN_GROUP(kg, td) { if ((td->td_priority) < PSOCK || !thread_safetoswapout(td)) goto nextproc; } /* * If the system is under memory stress, * or if we are swapping * idle processes >= swap_idle_threshold2, * then swap the process out. */ if (((action & VM_SWAP_NORMAL) == 0) && (((action & VM_SWAP_IDLE) == 0) || (kg->kg_slptime < swap_idle_threshold2))) goto nextproc; if (minslptime > kg->kg_slptime) minslptime = kg->kg_slptime; } /* * If the pageout daemon didn't free enough pages, * or if this process is idle and the system is * configured to swap proactively, swap it out. */ if ((action & VM_SWAP_NORMAL) || ((action & VM_SWAP_IDLE) && (minslptime > swap_idle_threshold2))) { swapout(p); didswap++; mtx_unlock_spin(&sched_lock); PROC_UNLOCK(p); vm_map_unlock(&vm->vm_map); vmspace_free(vm); sx_sunlock(&allproc_lock); goto retry; } nextproc: mtx_unlock_spin(&sched_lock); } nextproc2: PROC_UNLOCK(p); vm_map_unlock(&vm->vm_map); nextproc1: vmspace_free(vm); continue; } sx_sunlock(&allproc_lock); /* * If we swapped something out, and another process needed memory, * then wakeup the sched process. */ if (didswap) wakeup(&proc0); } static void swapout(p) struct proc *p; { struct thread *td; PROC_LOCK_ASSERT(p, MA_OWNED); mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); #if defined(SWAP_DEBUG) printf("swapping out %d\n", p->p_pid); #endif /* * The states of this process and its threads may have changed * by now. Assuming that there is only one pageout daemon thread, * this process should still be in memory. */ KASSERT((p->p_sflag & (PS_INMEM|PS_SWAPPINGOUT|PS_SWAPPINGIN)) == PS_INMEM, ("swapout: lost a swapout race?")); #if defined(INVARIANTS) /* * Make sure that all threads are safe to be swapped out. * * Alternatively, we could swap out only safe threads. */ FOREACH_THREAD_IN_PROC(p, td) { KASSERT(thread_safetoswapout(td), ("swapout: there is a thread not safe for swapout")); } #endif /* INVARIANTS */ ++p->p_stats->p_ru.ru_nswap; /* * remember the process resident count */ p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); p->p_sflag &= ~PS_INMEM; p->p_sflag |= PS_SWAPPINGOUT; PROC_UNLOCK(p); FOREACH_THREAD_IN_PROC(p, td) TD_SET_SWAPPED(td); mtx_unlock_spin(&sched_lock); FOREACH_THREAD_IN_PROC(p, td) vm_thread_swapout(td); PROC_LOCK(p); mtx_lock_spin(&sched_lock); p->p_sflag &= ~PS_SWAPPINGOUT; p->p_swtime = 0; } #endif /* !NO_SWAPPING */ Index: head/sys/vm/vm_kern.c =================================================================== --- head/sys/vm/vm_kern.c (revision 163621) +++ head/sys/vm/vm_kern.c (revision 163622) @@ -1,498 +1,498 @@ /*- * 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. * 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_kern.c 8.3 (Berkeley) 1/12/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Kernel memory management. */ #include __FBSDID("$FreeBSD$"); #include #include #include /* for ticks and hz */ #include #include #include #include #include #include #include #include #include #include #include #include vm_map_t kernel_map=0; vm_map_t kmem_map=0; vm_map_t exec_map=0; vm_map_t pipe_map; vm_map_t buffer_map=0; /* * kmem_alloc_nofault: * * Allocate a virtual address range with no underlying object and * no initial mapping to physical memory. Any mapping from this * range to physical memory must be explicitly created prior to * its use, typically with pmap_qenter(). Any attempt to create * a mapping on demand through vm_fault() will result in a panic. */ vm_offset_t kmem_alloc_nofault(map, size) vm_map_t map; vm_size_t size; { vm_offset_t addr; int result; size = round_page(size); addr = vm_map_min(map); result = vm_map_find(map, NULL, 0, &addr, size, TRUE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); if (result != KERN_SUCCESS) { return (0); } return (addr); } /* * Allocate wired-down memory in the kernel's address map * or a submap. */ vm_offset_t kmem_alloc(map, size) vm_map_t map; vm_size_t size; { vm_offset_t addr; vm_offset_t offset; vm_offset_t i; size = round_page(size); /* * Use the kernel object for wired-down kernel pages. Assume that no * region of the kernel object is referenced more than once. */ /* * Locate sufficient space in the map. This will give us the final * virtual address for the new memory, and thus will tell us the * offset within the kernel map. */ vm_map_lock(map); if (vm_map_findspace(map, vm_map_min(map), size, &addr)) { vm_map_unlock(map); return (0); } offset = addr - VM_MIN_KERNEL_ADDRESS; vm_object_reference(kernel_object); vm_map_insert(map, kernel_object, offset, addr, addr + size, VM_PROT_ALL, VM_PROT_ALL, 0); vm_map_unlock(map); /* * Guarantee that there are pages already in this object before * calling vm_map_wire. This is to prevent the following * scenario: * * 1) Threads have swapped out, so that there is a pager for the * kernel_object. 2) The kmsg zone is empty, and so we are * kmem_allocing a new page for it. 3) vm_map_wire calls vm_fault; * there is no page, but there is a pager, so we call * pager_data_request. But the kmsg zone is empty, so we must * kmem_alloc. 4) goto 1 5) Even if the kmsg zone is not empty: when * we get the data back from the pager, it will be (very stale) * non-zero data. kmem_alloc is defined to return zero-filled memory. * * We're intentionally not activating the pages we allocate to prevent a * race with page-out. vm_map_wire will wire the pages. */ VM_OBJECT_LOCK(kernel_object); for (i = 0; i < size; i += PAGE_SIZE) { vm_page_t mem; mem = vm_page_grab(kernel_object, OFF_TO_IDX(offset + i), VM_ALLOC_NOBUSY | VM_ALLOC_ZERO | VM_ALLOC_RETRY); mem->valid = VM_PAGE_BITS_ALL; vm_page_lock_queues(); vm_page_unmanage(mem); vm_page_unlock_queues(); } VM_OBJECT_UNLOCK(kernel_object); /* * And finally, mark the data as non-pageable. */ (void) vm_map_wire(map, addr, addr + size, VM_MAP_WIRE_SYSTEM|VM_MAP_WIRE_NOHOLES); return (addr); } /* * kmem_free: * * Release a region of kernel virtual memory allocated * with kmem_alloc, and return the physical pages * associated with that region. * * This routine may not block on kernel maps. */ void kmem_free(map, addr, size) vm_map_t map; vm_offset_t addr; vm_size_t size; { (void) vm_map_remove(map, trunc_page(addr), round_page(addr + size)); } /* * kmem_suballoc: * * Allocates a map to manage a subrange * of the kernel virtual address space. * * Arguments are as follows: * * parent Map to take range from * min, max Returned endpoints of map * size Size of range to find */ vm_map_t kmem_suballoc(parent, min, max, size) vm_map_t parent; vm_offset_t *min, *max; vm_size_t size; { int ret; vm_map_t result; size = round_page(size); *min = (vm_offset_t) vm_map_min(parent); ret = vm_map_find(parent, NULL, (vm_offset_t) 0, min, size, TRUE, VM_PROT_ALL, VM_PROT_ALL, 0); if (ret != KERN_SUCCESS) { printf("kmem_suballoc: bad status return of %d.\n", ret); panic("kmem_suballoc"); } *max = *min + size; result = vm_map_create(vm_map_pmap(parent), *min, *max); if (result == NULL) panic("kmem_suballoc: cannot create submap"); if (vm_map_submap(parent, *min, *max, result) != KERN_SUCCESS) panic("kmem_suballoc: unable to change range to submap"); return (result); } /* * kmem_malloc: * * Allocate wired-down memory in the kernel's address map for the higher * level kernel memory allocator (kern/kern_malloc.c). We cannot use * kmem_alloc() because we may need to allocate memory at interrupt * level where we cannot block (canwait == FALSE). * * This routine has its own private kernel submap (kmem_map) and object * (kmem_object). This, combined with the fact that only malloc uses * this routine, ensures that we will never block in map or object waits. * * Note that this still only works in a uni-processor environment and * when called at splhigh(). * * We don't worry about expanding the map (adding entries) since entries * for wired maps are statically allocated. * * NOTE: This routine is not supposed to block if M_NOWAIT is set, but * I have not verified that it actually does not block. * * `map' is ONLY allowed to be kmem_map or one of the mbuf submaps to * which we never free. */ vm_offset_t kmem_malloc(map, size, flags) vm_map_t map; vm_size_t size; int flags; { vm_offset_t offset, i; vm_map_entry_t entry; vm_offset_t addr; vm_page_t m; int pflags; size = round_page(size); addr = vm_map_min(map); /* * Locate sufficient space in the map. This will give us the final * virtual address for the new memory, and thus will tell us the * offset within the kernel map. */ vm_map_lock(map); if (vm_map_findspace(map, vm_map_min(map), size, &addr)) { vm_map_unlock(map); if ((flags & M_NOWAIT) == 0) panic("kmem_malloc(%ld): kmem_map too small: %ld total allocated", (long)size, (long)map->size); return (0); } offset = addr - VM_MIN_KERNEL_ADDRESS; vm_object_reference(kmem_object); vm_map_insert(map, kmem_object, offset, addr, addr + size, VM_PROT_ALL, VM_PROT_ALL, 0); /* * Note: if M_NOWAIT specified alone, allocate from * interrupt-safe queues only (just the free list). If * M_USE_RESERVE is also specified, we can also * allocate from the cache. Neither of the latter two * flags may be specified from an interrupt since interrupts * are not allowed to mess with the cache queue. */ if ((flags & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT) pflags = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED; else pflags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED; if (flags & M_ZERO) pflags |= VM_ALLOC_ZERO; VM_OBJECT_LOCK(kmem_object); for (i = 0; i < size; i += PAGE_SIZE) { retry: m = vm_page_alloc(kmem_object, OFF_TO_IDX(offset + i), pflags); /* * Ran out of space, free everything up and return. Don't need * to lock page queues here as we know that the pages we got * aren't on any queues. */ if (m == NULL) { if ((flags & M_NOWAIT) == 0) { VM_OBJECT_UNLOCK(kmem_object); vm_map_unlock(map); VM_WAIT; vm_map_lock(map); VM_OBJECT_LOCK(kmem_object); goto retry; } /* * Free the pages before removing the map entry. * They are already marked busy. Calling * vm_map_delete before the pages has been freed or * unbusied will cause a deadlock. */ while (i != 0) { i -= PAGE_SIZE; m = vm_page_lookup(kmem_object, OFF_TO_IDX(offset + i)); vm_page_lock_queues(); vm_page_unwire(m, 0); vm_page_free(m); vm_page_unlock_queues(); } VM_OBJECT_UNLOCK(kmem_object); vm_map_delete(map, addr, addr + size); vm_map_unlock(map); return (0); } if (flags & M_ZERO && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); m->valid = VM_PAGE_BITS_ALL; vm_page_lock_queues(); vm_page_unmanage(m); vm_page_unlock_queues(); } VM_OBJECT_UNLOCK(kmem_object); /* * Mark map entry as non-pageable. Assert: vm_map_insert() will never * be able to extend the previous entry so there will be a new entry * exactly corresponding to this address range and it will have * wired_count == 0. */ if (!vm_map_lookup_entry(map, addr, &entry) || entry->start != addr || entry->end != addr + size || entry->wired_count != 0) panic("kmem_malloc: entry not found or misaligned"); entry->wired_count = 1; /* * At this point, the kmem_object must be unlocked because * vm_map_simplify_entry() calls vm_object_deallocate(), which * locks the kmem_object. */ vm_map_simplify_entry(map, entry); /* * Loop thru pages, entering them in the pmap. (We cannot add them to * the wired count without wrapping the vm_page_queue_lock in * splimp...) */ VM_OBJECT_LOCK(kmem_object); for (i = 0; i < size; i += PAGE_SIZE) { m = vm_page_lookup(kmem_object, OFF_TO_IDX(offset + i)); /* * Because this is kernel_pmap, this call will not block. */ pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1); vm_page_lock_queues(); vm_page_flag_set(m, PG_WRITEABLE | PG_REFERENCED); - vm_page_wakeup(m); vm_page_unlock_queues(); + vm_page_wakeup(m); } VM_OBJECT_UNLOCK(kmem_object); vm_map_unlock(map); return (addr); } /* * kmem_alloc_wait: * * Allocates pageable memory from a sub-map of the kernel. If the submap * has no room, the caller sleeps waiting for more memory in the submap. * * This routine may block. */ vm_offset_t kmem_alloc_wait(map, size) vm_map_t map; vm_size_t size; { vm_offset_t addr; size = round_page(size); for (;;) { /* * To make this work for more than one map, use the map's lock * to lock out sleepers/wakers. */ vm_map_lock(map); if (vm_map_findspace(map, vm_map_min(map), size, &addr) == 0) break; /* no space now; see if we can ever get space */ if (vm_map_max(map) - vm_map_min(map) < size) { vm_map_unlock(map); return (0); } map->needs_wakeup = TRUE; vm_map_unlock_and_wait(map, FALSE); } vm_map_insert(map, NULL, 0, addr, addr + size, VM_PROT_ALL, VM_PROT_ALL, 0); vm_map_unlock(map); return (addr); } /* * kmem_free_wakeup: * * Returns memory to a submap of the kernel, and wakes up any processes * waiting for memory in that map. */ void kmem_free_wakeup(map, addr, size) vm_map_t map; vm_offset_t addr; vm_size_t size; { vm_map_lock(map); (void) vm_map_delete(map, trunc_page(addr), round_page(addr + size)); if (map->needs_wakeup) { map->needs_wakeup = FALSE; vm_map_wakeup(map); } vm_map_unlock(map); } /* * kmem_init: * * Create the kernel map; insert a mapping covering kernel text, * data, bss, and all space allocated thus far (`boostrap' data). The * new map will thus map the range between VM_MIN_KERNEL_ADDRESS and * `start' as allocated, and the range between `start' and `end' as free. */ void kmem_init(start, end) vm_offset_t start, end; { vm_map_t m; m = vm_map_create(kernel_pmap, VM_MIN_KERNEL_ADDRESS, end); m->system_map = 1; vm_map_lock(m); /* N.B.: cannot use kgdb to debug, starting with this assignment ... */ kernel_map = m; (void) vm_map_insert(m, NULL, (vm_ooffset_t) 0, VM_MIN_KERNEL_ADDRESS, start, VM_PROT_ALL, VM_PROT_ALL, 0); /* ... and ending with the completion of the above `insert' */ vm_map_unlock(m); }