/* * Copyright (c) 1998 Matthew Dillon, * Copyright (c) 1994 John S. Dyson * Copyright (c) 1990 University of Utah. * 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 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 * * $Id: swap_pager.c,v 1.115 1999/02/21 08:30:49 dillon Exp $ */ #include #include #include #include #include #include #include #include #include #include #include #ifndef MAX_PAGEOUT_CLUSTER #define MAX_PAGEOUT_CLUSTER 16 #endif #define SWB_NPAGES MAX_PAGEOUT_CLUSTER #include "opt_swap.h" #include #include #include #include #include #include #include #include #include #define SWM_FREE 0x02 /* free, period */ #define SWM_POP 0x04 /* pop out */ /* * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks * in the old system. */ extern int vm_swap_size; /* number of free swap blocks, in pages */ int swap_pager_full; /* swap space exhaustion (task killing) */ static int swap_pager_almost_full; /* 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 int sw_alloc_interlock; /* swap pager allocation interlock */ struct blist *swapblist; static struct swblock **swhash; static int swhash_mask; static int swap_async_max = 4; /* maximum in-progress async I/O's */ 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)(long)handle >> 4) & (NOBJLISTS-1)]) static struct pagerlst swap_pager_object_list[NOBJLISTS]; struct pagerlst swap_pager_un_object_list; vm_zone_t swap_zone; /* * 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 __P((void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t offset)); static void swap_pager_dealloc __P((vm_object_t object)); static int swap_pager_getpages __P((vm_object_t, vm_page_t *, int, int)); static void swap_pager_init __P((void)); static void swap_pager_unswapped __P((vm_page_t)); struct pagerops swappagerops = { swap_pager_init, /* early system initialization of pager */ swap_pager_alloc, /* allocate an OBJT_SWAP object */ swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ swap_pager_getpages, /* pagein */ swap_pager_putpages, /* pageout */ swap_pager_haspage, /* get backing store status for page */ 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. * * swap_*() routines are externally accessible. swp_*() routines are * internal. */ int dmmax; static int dmmax_mask; int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ static __inline void swp_sizecheck __P((void)); static void swp_pager_sync_iodone __P((struct buf *bp)); static void swp_pager_async_iodone __P((struct buf *bp)); /* * Swap bitmap functions */ static __inline void swp_pager_freeswapspace __P((daddr_t blk, int npages)); static __inline daddr_t swp_pager_getswapspace __P((int npages)); /* * Metadata functions */ static void swp_pager_meta_build __P((vm_object_t, daddr_t, daddr_t, int)); static void swp_pager_meta_free __P((vm_object_t, daddr_t, daddr_t)); static void swp_pager_meta_free_all __P((vm_object_t)); static daddr_t swp_pager_meta_ctl __P((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 __inline void swp_sizecheck() { if (vm_swap_size < 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 (vm_swap_size > nswap_hiwat) swap_pager_almost_full = 0; } } /* * 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() { /* * Initialize object lists */ int i; for (i = 0; i < NOBJLISTS; ++i) TAILQ_INIT(&swap_pager_object_list[i]); TAILQ_INIT(&swap_pager_un_object_list); /* * Device Stripe, in PAGE_SIZE'd blocks */ dmmax = SWB_NPAGES * 2; dmmax_mask = ~(dmmax - 1); } /* * 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() { int n; /* * 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); nsw_rcount = (nswbuf + 1) / 2; nsw_wcount_sync = (nswbuf + 3) / 4; nsw_wcount_async = 4; nsw_wcount_async_max = nsw_wcount_async; /* * 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. */ n = cnt.v_page_count * 2; swap_zone = zinit( "SWAPMETA", sizeof(struct swblock), n, ZONE_INTERRUPT, 1 ); /* * 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 < cnt.v_page_count / 4; n <<= 1) ; swhash = malloc(sizeof(struct swblock *) * n, M_VMPGDATA, M_WAITOK); bzero(swhash, sizeof(struct swblock *) * n); swhash_mask = n - 1; } /* * 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. */ 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; if (handle) { /* * 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. */ while (sw_alloc_interlock) { sw_alloc_interlock = -1; tsleep(&sw_alloc_interlock, PVM, "swpalc", 0); } sw_alloc_interlock = 1; object = vm_pager_object_lookup(NOBJLIST(handle), handle); if (object != NULL) { vm_object_reference(object); } else { object = vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(offset + PAGE_MASK + size)); object->handle = handle; swp_pager_meta_build( object, 0, SWAPBLK_NONE, 0 ); } if (sw_alloc_interlock < 0) wakeup(&sw_alloc_interlock); sw_alloc_interlock = 0; } else { object = vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(offset + PAGE_MASK + size)); swp_pager_meta_build( object, 0, SWAPBLK_NONE, 0 ); } 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(object) vm_object_t object; { /* * Remove from list right away so lookups will fail if we block for * pageout completion. */ if (object->handle == NULL) { TAILQ_REMOVE(&swap_pager_un_object_list, object, pager_object_list); } else { TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list); } 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(). */ static __inline daddr_t swp_pager_getswapspace(npages) int npages; { daddr_t blk; if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) { if (swap_pager_full != 2) { printf("swap_pager_getswapspace: failed\n"); swap_pager_full = 2; swap_pager_almost_full = 1; } } else { vm_swap_size -= npages; swp_sizecheck(); } return(blk); } /* * 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 __inline void swp_pager_freeswapspace(blk, npages) daddr_t blk; int npages; { blist_free(swapblist, blk, npages); vm_swap_size += npages; swp_sizecheck(); } /* * 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. */ void swap_pager_freespace(object, start, size) vm_object_t object; vm_pindex_t start; vm_size_t size; { swp_pager_meta_free(object, start, size); } /* * 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. * * 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(srcobject, dstobject, offset, destroysource) vm_object_t srcobject; vm_object_t dstobject; vm_pindex_t offset; int destroysource; { vm_pindex_t i; /* * If destroysource is set, we remove the source object from the * swap_pager internal queue now. */ if (destroysource) { if (srcobject->handle == NULL) { TAILQ_REMOVE( &swap_pager_un_object_list, srcobject, pager_object_list ); } else { TAILQ_REMOVE( NOBJLIST(srcobject->handle), srcobject, pager_object_list ); } } /* * 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(dstobject, i, srcaddr, 1); } 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 consistancy since we've removed the object from its * queues. */ srcobject->type = OBJT_DEFAULT; } return; } /* * 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. */ boolean_t swap_pager_haspage(object, pindex, before, after) vm_object_t object; vm_pindex_t pindex; int *before; int *after; { daddr_t blk0; /* * 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 & SWAPBLK_NONE) break; 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 & SWAPBLK_NONE) break; 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 */ static void swap_pager_unswapped(m) vm_page_t m; { 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(object, m, count, reqpage) vm_object_t object; vm_page_t *m; int count, reqpage; { struct buf *bp; vm_page_t mreq; int s; int i; int j; daddr_t blk; vm_offset_t kva; vm_pindex_t lastpindex; mreq = m[reqpage]; #if !defined(MAX_PERF) if (mreq->object != object) { panic("swap_pager_getpages: object mismatch %p/%p", object, mreq->object ); } #endif /* * Calculate range to retrieve. The pages have already been assigned * their swapblks. We require a *contiguous* range that falls entirely * within a single device stripe. If we do not supply it, bad things * happen. */ 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 (iblk & SWAPBLK_NONE) break; if ((blk ^ iblk) & dmmax_mask) break; 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 (jblk & SWAPBLK_NONE) break; if ((blk ^ jblk) & dmmax_mask) break; if (blk != jblk - (j - reqpage)) break; } /* * If blk itself is bad, well, we can't do any I/O. This should * already be covered as a side effect, but I'm making sure. */ if (blk & SWAPBLK_NONE) { i = reqpage; j = reqpage + 1; } /* * free pages outside our collection range. Note: we never free * mreq, it must remain busy throughout. */ { int k; for (k = 0; k < i; ++k) { vm_page_free(m[k]); } for (k = j; k < count; ++k) { vm_page_free(m[k]); } } /* * 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); /* * Get a swap buffer header to perform the IO */ bp = getpbuf(&nsw_rcount); kva = (vm_offset_t) bp->b_data; /* * map our page(s) into kva for input * * NOTE: B_PAGING is set by pbgetvp() */ pmap_qenter(kva, m + i, j - i); bp->b_flags = B_BUSY | B_READ | B_CALL; bp->b_iodone = swp_pager_async_iodone; bp->b_proc = &proc0; /* XXX (but without B_PHYS set this is ok) */ bp->b_rcred = bp->b_wcred = bp->b_proc->p_ucred; crhold(bp->b_rcred); crhold(bp->b_wcred); bp->b_data = (caddr_t) kva; /* * b_blkno is in page-sized chunks. swapblk is valid, too, so * we don't have to mask it against SWAPBLK_MASK. */ 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; { int k; for (k = i; k < j; ++k) { bp->b_pages[k - i] = m[k]; vm_page_flag_set(m[k], PG_SWAPINPROG); } } bp->b_npages = j - i; pbgetvp(swapdev_vp, bp); 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(mreq->object, bp->b_npages); lastpindex = m[j-1]->pindex; /* * 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 VOP_STRATEGY */ VOP_STRATEGY(bp->b_vp, bp); /* * wait for the page we want to complete. PG_SWAPINPROG is always * cleared on completion. If an I/O error occurs, SWAPBLK_NONE * is set in the meta-data. */ s = splvm(); while ((mreq->flags & PG_SWAPINPROG) != 0) { vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED); cnt.v_intrans++; if (tsleep(mreq, PSWP, "swread", hz*20)) { printf( "swap_pager: indefinite wait buffer: device:" " %#lx, blkno: %ld, size: %ld\n", (u_long)bp->b_dev, (long)bp->b_blkno, (long)bp->b_bcount ); } } splx(s); /* * mreq is left bussied 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 { mreq->object->last_read = lastpindex; 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(object, m, count, sync, rtvals) vm_object_t object; vm_page_t *m; int count; boolean_t sync; int *rtvals; { int i; int n = 0; #if !defined(MAX_PERF) if (count && m[0]->object != object) { panic("swap_pager_getpages: object mismatch %p/%p", object, m[0]->object ); } #endif /* * 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, 0); } 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. */ if (swap_async_max != nsw_wcount_async_max) { int n; int s; /* * 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. */ s = splvm(); n -= nsw_wcount_async_max; if (nsw_wcount_async + n >= 0) { nsw_wcount_async += n; nsw_wcount_async_max += n; wakeup(&nsw_wcount_async); } splx(s); } /* * 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 s; 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; } /* * Oops, too big if it crosses a stripe * * 1111000000 * 111111 * 1000001 */ if ((blk ^ (blk + n)) & dmmax_mask) { j = ((blk + dmmax) & dmmax_mask) - blk; swp_pager_freeswapspace(blk + j, n - j); n = j; } /* * All I/O parameters have been satisfied, build the I/O * request and assign the swap space. * * NOTE: B_PAGING is set by pbgetvp() */ if (sync == TRUE) { bp = getpbuf(&nsw_wcount_sync); bp->b_flags = B_BUSY; } else { bp = getpbuf(&nsw_wcount_async); bp->b_flags = B_BUSY | B_ASYNC; } bp->b_spc = NULL; /* not used, but NULL-out anyway */ pmap_qenter((vm_offset_t)bp->b_data, &m[i], n); bp->b_proc = &proc0; /* XXX (but without B_PHYS this is ok) */ bp->b_rcred = bp->b_wcred = bp->b_proc->p_ucred; if (bp->b_rcred != NOCRED) crhold(bp->b_rcred); if (bp->b_wcred != NOCRED) crhold(bp->b_wcred); pbgetvp(swapdev_vp, bp); bp->b_bcount = PAGE_SIZE * n; bp->b_bufsize = PAGE_SIZE * n; bp->b_blkno = blk; s = splvm(); for (j = 0; j < n; ++j) { vm_page_t mreq = m[i+j]; swp_pager_meta_build( mreq->object, mreq->pindex, blk + j, 0 ); vm_page_dirty(mreq); rtvals[i+j] = VM_PAGER_OK; vm_page_flag_set(mreq, PG_SWAPINPROG); bp->b_pages[j] = mreq; } bp->b_flags |= B_CALL; bp->b_npages = n; cnt.v_swapout++; cnt.v_swappgsout += bp->b_npages; swapdev_vp->v_numoutput++; /* * asynchronous * * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY */ if (sync == FALSE) { bp->b_iodone = swp_pager_async_iodone; bp->b_dirtyoff = 0; bp->b_dirtyend = bp->b_bcount; VOP_STRATEGY(bp->b_vp, bp); for (j = 0; j < n; ++j) rtvals[i+j] = VM_PAGER_PEND; splx(s); continue; } /* * synchronous * * NOTE: b_blkno is destroyed by the call to VOP_STRATEGY */ bp->b_iodone = swp_pager_sync_iodone; VOP_STRATEGY(bp->b_vp, 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. */ while ((bp->b_flags & B_DONE) == 0) { tsleep(bp, PVM, "swwrt", 0); } 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); splx(s); } } /* * swap_pager_sync_iodone: * * Completion routine for synchronous reads and writes from/to swap. * We just mark the bp is complete and wake up anyone waiting on it. * * This routine may not block. */ static void swp_pager_sync_iodone(bp) struct buf *bp; { bp->b_flags |= B_DONE; bp->b_flags &= ~B_ASYNC; wakeup(bp); } /* * 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. * * WARNING! This routine may be called from an interrupt. We cannot * mess with swap metadata unless we want to run all our other routines * at splbio() too, which I'd rather not do. We up ourselves * to splvm() because we may call vm_page_free(), which can unlink a * page from an object. * * XXX currently I do not believe any object routines protect * object->memq at splvm(). The code must be gone over to determine * the actual state of the problem. * * 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() */ static void swp_pager_async_iodone(bp) register struct buf *bp; { int s; int i; vm_object_t object = NULL; s = splvm(); bp->b_flags |= B_DONE; /* * report error */ if (bp->b_flags & B_ERROR) { printf( "swap_pager: I/O error - %s failed; blkno %ld," "size %ld, error %d\n", ((bp->b_flags & B_READ) ? "pagein" : "pageout"), (long)bp->b_blkno, (long)bp->b_bcount, bp->b_error ); } /* * set object. */ if (bp->b_npages) object = bp->b_pages[0]->object; /* * remove the mapping for kernel virtual */ pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); /* * 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]; vm_page_flag_clear(m, PG_SWAPINPROG); if (bp->b_flags & B_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_flags & B_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 PG_SWAPINPROG and * someone may be waiting for that. * * NOTE: for reads, m->dirty will probably * be overriden by the original caller of * getpages so don't play cute tricks here. * * XXX it may not be legal to free the page * here as this messes with the object->memq's. */ m->valid = 0; vm_page_flag_clear(m, PG_ZERO); 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_flags & B_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 * overriden 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. * * clear PG_ZERO in page. * * 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(VM_PAGE_TO_PHYS(m)); m->valid = VM_PAGE_BITS_ALL; m->dirty = 0; vm_page_flag_clear(m, PG_ZERO); /* * We have to wake specifically requested pages * up too because we cleared PG_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 ). */ vm_page_protect(m, VM_PROT_READ); pmap_clear_modify(VM_PAGE_TO_PHYS(m)); m->dirty = 0; vm_page_io_finish(m); } } /* * adjust pip. NOTE: the original parent may still have its own * pip refs on the object. */ if (object) vm_object_pip_wakeupn(object, bp->b_npages); /* * release the physical I/O buffer */ relpbuf( bp, ((bp->b_flags & B_READ) ? &nsw_rcount : ((bp->b_flags & B_ASYNC) ? &nsw_wcount_async : &nsw_wcount_sync ) ) ); splx(s); } /************************************************************************ * SWAP META DATA * ************************************************************************ * * These routines manipulate the swap metadata stored in the * OBJT_SWAP object. * * In fact, we just have a few counters in the vm_object_t. The * metadata is actually stored in a hash table. */ /* * SWP_PAGER_HASH() - hash swap meta data * * This is an inline helper function which hash 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. */ static __inline struct swblock ** swp_pager_hash(vm_object_t object, daddr_t index) { struct swblock **pswap; struct swblock *swap; index &= ~SWAP_META_MASK; pswap = &swhash[(index ^ (int)(long)object) & swhash_mask]; while ((swap = *pswap) != NULL) { if (swap->swb_object == object && swap->swb_index == index ) { break; } pswap = &swap->swb_hnext; } return(pswap); } /* * 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. */ static void swp_pager_meta_build( vm_object_t object, daddr_t index, daddr_t swapblk, int waitok ) { struct swblock *swap; struct swblock **pswap; /* * 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) { TAILQ_INSERT_TAIL( NOBJLIST(object->handle), object, pager_object_list ); } else { TAILQ_INSERT_TAIL( &swap_pager_un_object_list, object, pager_object_list ); } } /* * Wait for free memory when waitok is TRUE prior to calling the * zone allocator. */ while (waitok && cnt.v_free_count == 0) { VM_WAIT; } /* * If swapblk being added is invalid, just free it. */ if (swapblk & SWAPBLK_NONE) { if (swapblk != SWAPBLK_NONE) { swp_pager_freeswapspace( index, 1 ); swapblk = SWAPBLK_NONE; } } /* * Locate hash entry. If not found create, but if we aren't adding * anything just return. */ pswap = swp_pager_hash(object, index); if ((swap = *pswap) == NULL) { int i; if (swapblk == SWAPBLK_NONE) return; swap = *pswap = zalloc(swap_zone); swap->swb_hnext = NULL; swap->swb_object = object; swap->swb_index = index & ~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 */ index &= SWAP_META_MASK; if (swap->swb_pages[index] != SWAPBLK_NONE) { swp_pager_freeswapspace( swap->swb_pages[index] & SWAPBLK_MASK, 1 ); --swap->swb_count; } /* * Enter block into metadata */ swap->swb_pages[index] = swapblk; ++swap->swb_count; } /* * 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, daddr_t index, daddr_t count) { if (object->type != OBJT_SWAP) return; while (count > 0) { struct swblock **pswap; struct swblock *swap; 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; zfree(swap_zone, swap); --object->un_pager.swp.swp_bcount; } } --count; ++index; } else { daddr_t n = SWAP_META_PAGES - (index & SWAP_META_MASK); count -= n; index += n; } } } /* * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object * * This routine locates and destroys all swap metadata associated with * an object. */ static void swp_pager_meta_free_all(vm_object_t object) { daddr_t index = 0; if (object->type != OBJT_SWAP) return; while (object->un_pager.swp.swp_bcount) { struct swblock **pswap; struct swblock *swap; 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) { #if !defined(MAX_PERF) --swap->swb_count; #endif swp_pager_freeswapspace( v, 1 ); } } #if !defined(MAX_PERF) if (swap->swb_count != 0) panic("swap_pager_meta_free_all: swb_count != 0"); #endif *pswap = swap->swb_hnext; zfree(swap_zone, swap); --object->un_pager.swp.swp_bcount; } index += SWAP_META_PAGES; #if !defined(MAX_PERF) if (index > 0x20000000) panic("swp_pager_meta_free_all: failed to locate all swap meta blocks"); #endif } } /* * 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. * * 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 index, int flags ) { /* * The meta data only exists of the object is OBJT_SWAP * and even then might not be allocated yet. */ if ( object->type != OBJT_SWAP || object->un_pager.swp.swp_bcount == 0 ) { return(SWAPBLK_NONE); } { struct swblock **pswap; struct swblock *swap; daddr_t r1 = SWAPBLK_NONE; pswap = swp_pager_hash(object, index); index &= SWAP_META_MASK; if ((swap = *pswap) != NULL) { r1 = swap->swb_pages[index]; 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[index] = SWAPBLK_NONE; if (--swap->swb_count == 0) { *pswap = swap->swb_hnext; zfree(swap_zone, swap); --object->un_pager.swp.swp_bcount; } } } } return(r1); } /* not reached */ }