/* * Mach Operating System * Copyright (c) 1993-1989 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 Mellon * the rights to redistribute these changes. */ /* * Default pager. Pages to paging partition. * * MUST BE ABLE TO ALLOCATE WIRED-DOWN MEMORY!!! */ #include #include #include #include #include #include #include #include #include #include #include #include #include "queue.h" #include "wiring.h" #include "kalloc.h" #include "default_pager.h" #include #include #include #include #include #include #include "memory_object_S.h" #include "memory_object_default_S.h" #include "default_pager_S.h" #include "exc_S.h" #include "priv.h" #define debug 0 static char my_name[] = "(default pager):"; static void __attribute__ ((format (printf, 1, 2), unused)) synchronized_printf (const char *fmt, ...) { static pthread_mutex_t printf_lock = PTHREAD_MUTEX_INITIALIZER; va_list ap; va_start (ap, fmt); pthread_mutex_lock (&printf_lock); vprintf (fmt, ap); fflush (stdout); pthread_mutex_unlock (&printf_lock); va_end (ap); } #if 0 #define dprintf(f, x...) synchronized_printf (f, ##x) #else #define dprintf(f, x...) (void) 0 #endif #if 0 #define ddprintf(f, x...) synchronized_printf (f, ##x) #else #define ddprintf(f, x...) (void) 0 #endif /* * parallel vs serial switch */ #define PARALLEL 1 #if 0 #define CHECKSUM 1 #endif #define USE_PRECIOUS 1 #define ptoa(p) ((p)*vm_page_size) #define atop(a) ((a)/vm_page_size) partition_t partition_of(x) int x; { if (x >= all_partitions.n_partitions || x < 0) panic("partition_of x%x", x); return all_partitions.partition_list[x]; } void set_partition_of(x, p) int x; partition_t p; { if (x >= all_partitions.n_partitions || x < 0) panic("set_partition_of x%x", x); all_partitions.partition_list[x] = p; } /* * Simple mapping from (file)NAME to id * Saves space, filenames can be long. */ unsigned int part_id(const char *name) { unsigned int id, xorid; size_t len; len = strlen(name); id = xorid = 0; while (len--) { xorid ^= *name; id += *name++; } return (id << 8) | xorid; } void partition_init() { pthread_mutex_init(&all_partitions.lock, NULL); all_partitions.n_partitions = 0; } static partition_t new_partition (const char *name, struct file_direct *fdp, int check_linux_signature) { partition_t part; vm_size_t size, bmsize; vm_offset_t raddr; mach_msg_type_number_t rsize; int rc; unsigned int id = part_id(name); pthread_mutex_lock(&all_partitions.lock); { unsigned int i; for (i = 0; i < all_partitions.n_partitions; i++) { part = partition_of(i); if (part && part->id == id) { pthread_mutex_unlock(&all_partitions.lock); return 0; } } } pthread_mutex_unlock(&all_partitions.lock); size = atop(fdp->fd_size * fdp->fd_bsize); bmsize = howmany(size, NB_BM) * sizeof(bm_entry_t); part = (partition_t) kalloc(sizeof(struct part)); pthread_mutex_init(&part->p_lock, NULL); part->total_size = size; part->free = size; part->id = id; part->bitmap = (bm_entry_t *)kalloc(bmsize); part->going_away= FALSE; part->file = fdp; bzero((char *)part->bitmap, bmsize); if (check_linux_signature < 0) { if (check_linux_signature != -3) printf("(default pager): " "Paging to raw partition %s (%uk paging space)\n", name, part->total_size * (vm_page_size / 1024)); return part; } #define LINUX_PAGE_SIZE 4096 /* size of pages in Linux swap partitions */ rc = page_read_file_direct(part->file, 0, LINUX_PAGE_SIZE, &raddr, &rsize); if (rc) panic("(default pager): cannot read first page of %s! rc=%#x\n", name, rc); while (rsize < LINUX_PAGE_SIZE) { /* Filesystem block size is smaller than page size, so we must do several reads to get the whole page. */ vm_address_t baddr; vm_size_t bsize; rc = page_read_file_direct(part->file, rsize, LINUX_PAGE_SIZE-rsize, &baddr, &bsize); if (rc) panic("(default pager): " "cannot read first page of %s! rc=%#x at %#x\n", name, rc, rsize); memcpy ((char *) raddr + rsize, (void *) baddr, bsize); rsize += bsize; vm_deallocate (mach_task_self (), baddr, bsize); } if (!memcmp("SWAP-SPACE", (char *) raddr + LINUX_PAGE_SIZE-10, 10)) { /* The partition's first page has a Linux swap signature. This means the beginning of the page contains a bitmap of good pages, and all others are bad. */ unsigned int i, j, bad, max; int waste; printf("(default pager): Found Linux 2.0 swap signature in %s\n", name); /* The first page, and the pages corresponding to the bits occupied by the signature in the final 10 bytes of the page, are always unavailable ("bad"). */ *(u_int32_t *)raddr &= ~(u_int32_t) 1; memset((char *) raddr + LINUX_PAGE_SIZE-10, 0, 10); max = LINUX_PAGE_SIZE / sizeof(u_int32_t); if (max > (part->total_size + 31) / 32) max = (part->total_size + 31) / 32; bad = 0; for (i = 0; i < max; ++i) { u_int32_t bm = ((u_int32_t *) raddr)[i]; if (bm == ~(u_int32_t) 0) continue; /* There are some zero bits in this word. */ for (j = 0; j < 32; ++j) if ((bm & (1 << j)) == 0) { unsigned int p = i*32 + j; if (p >= part->total_size) break; ++bad; part->bitmap[p / NB_BM] |= 1 << (p % NB_BM); } } part->free -= bad; --bad; /* Don't complain about first page. */ waste = part->total_size - (8 * (LINUX_PAGE_SIZE-10)); if (waste > 0) { /* The wasted pages were already marked "bad". */ bad -= waste; if (bad > 0) printf("\ (default pager): Paging to %s, %dk swap-space (%dk bad, %dk wasted at end)\n", name, part->free * (LINUX_PAGE_SIZE / 1024), bad * (LINUX_PAGE_SIZE / 1024), waste * (LINUX_PAGE_SIZE / 1024)); else printf("\ (default pager): Paging to %s, %dk swap-space (%dk wasted at end)\n", name, part->free * (LINUX_PAGE_SIZE / 1024), waste * (LINUX_PAGE_SIZE / 1024)); } else if (bad > 0) printf("\ (default pager): Paging to %s, %dk swap-space (excludes %dk marked bad)\n", name, part->free * (LINUX_PAGE_SIZE / 1024), bad * (LINUX_PAGE_SIZE / 1024)); else printf("\ (default pager): Paging to %s, %dk swap-space\n", name, part->free * (LINUX_PAGE_SIZE / 1024)); } else if (!memcmp("SWAPSPACE2", (char *) raddr + LINUX_PAGE_SIZE-10, 10)) { struct { u_int8_t bootbits[1024]; u_int32_t version; u_int32_t last_page; u_int32_t nr_badpages; u_int32_t padding[125]; u_int32_t badpages[1]; } *hdr = (void *) raddr; printf("\ (default pager): Found Linux 2.2 swap signature (v%u) in %s...", hdr->version, name); part->bitmap[0] |= 1; /* first page unusable */ part->free--; switch (hdr->version) { default: if (check_linux_signature) { printf ("version %u unknown! SKIPPING %s!\n", hdr->version, name); vm_deallocate(mach_task_self(), raddr, rsize); kfree(part->bitmap, bmsize); kfree(part, sizeof *part); return 0; } else printf ("version %u unknown! IGNORING SIGNATURE PAGE!" " %dk swap-space\n", hdr->version, part->free * (LINUX_PAGE_SIZE / 1024)); break; case 1: { unsigned int waste, i; if (hdr->last_page > part->total_size) { printf ("signature says %uk, partition has only %uk! ", hdr->last_page * (LINUX_PAGE_SIZE / 1024), part->total_size * (LINUX_PAGE_SIZE / 1024)); waste = 0; } else { waste = part->total_size - hdr->last_page; part->total_size = hdr->last_page; part->free = part->total_size - 1; } for (i = 0; i < hdr->nr_badpages; ++i) { const u_int32_t bad = hdr->badpages[i]; part->bitmap[bad / NB_BM] |= 1 << (bad % NB_BM); part->free--; } printf ("%uk swap-space", part->free * (LINUX_PAGE_SIZE / 1024)); if (hdr->nr_badpages != 0) printf (" (excludes %uk marked bad)", hdr->nr_badpages * (LINUX_PAGE_SIZE / 1024)); if (waste != 0) printf (" (excludes %uk at end of partition)", waste * (LINUX_PAGE_SIZE / 1024)); printf ("\n"); } } } else if (check_linux_signature) { printf ("(default pager): " "Cannot find Linux swap signature page! " "SKIPPING %s (%uk partition)!", name, part->total_size * (vm_page_size / 1024)); kfree(part->bitmap, bmsize); kfree(part, sizeof *part); part = 0; } else printf("(default pager): " "Paging to raw partition %s (%uk paging space)\n", name, part->total_size * (vm_page_size / 1024)); vm_deallocate(mach_task_self(), raddr, rsize); return part; } /* * Create a partition descriptor, * add it to the list of all such. * size is in BYTES. */ void create_paging_partition(const char *name, struct file_direct *fdp, int isa_file, int linux_signature) { partition_t part; part = new_partition (name, fdp, linux_signature); if (!part) return; pthread_mutex_lock(&all_partitions.lock); { int i; for (i = 0; i < all_partitions.n_partitions; i++) if (partition_of(i) == 0) break; if (i == all_partitions.n_partitions) { partition_t *new_list, *old_list; int n; n = i ? (i<<1) : 2; new_list = (partition_t *) kalloc( n * sizeof(partition_t) ); if (new_list == 0) no_paging_space(TRUE); bzero(new_list, n*sizeof(partition_t)); if (i) { old_list = all_partitions.partition_list; bcopy(old_list, new_list, i*sizeof(partition_t)); } all_partitions.partition_list = new_list; all_partitions.n_partitions = n; if (i) kfree(old_list, i*sizeof(partition_t)); } set_partition_of(i, part); } pthread_mutex_unlock(&all_partitions.lock); #if 0 dprintf("%s Added paging %s %s\n", my_name, (isa_file) ? "file" : "device", name); #endif overcommitted(TRUE, part->free); } /* * Choose the most appropriate default partition * for an object of SIZE bytes. * Return the partition locked, unless * the object has no CUR_PARTition. */ p_index_t choose_partition(size, cur_part) unsigned int size; p_index_t cur_part; { partition_t part; boolean_t found = FALSE; int i; pthread_mutex_lock(&all_partitions.lock); for (i = 0; i < all_partitions.n_partitions; i++) { /* the undesirable one ? */ if (i == cur_part) continue; ddprintf ("choose_partition(%x,%d,%d)\n",size,cur_part,i); /* one that was removed ? */ if ((part = partition_of(i)) == 0) continue; /* one that is being removed ? */ if (part->going_away) continue; /* is it big enough ? */ pthread_mutex_lock(&part->p_lock); if (ptoa(part->free) >= size) { if (cur_part != P_INDEX_INVALID) { pthread_mutex_unlock(&all_partitions.lock); return (p_index_t)i; } else found = TRUE; } pthread_mutex_unlock(&part->p_lock); if (found) break; } pthread_mutex_unlock(&all_partitions.lock); return (found) ? (p_index_t)i : P_INDEX_INVALID; } /* * Allocate a page in a paging partition * The partition is returned unlocked. */ vm_offset_t pager_alloc_page(pindex, lock_it) p_index_t pindex; boolean_t lock_it; { int bm_e; int bit; int limit; bm_entry_t *bm; partition_t part; static char here[] = "%spager_alloc_page"; if (no_partition(pindex)) return (NO_BLOCK); ddprintf ("pager_alloc_page(%d,%d)\n",pindex,lock_it); part = partition_of(pindex); /* unlikely, but possible deadlock against destroy_partition */ if (!part || part->going_away) return (NO_BLOCK); if (lock_it) pthread_mutex_lock(&part->p_lock); if (part->free == 0) { /* out of paging space */ pthread_mutex_unlock(&part->p_lock); return (NO_BLOCK); } limit = howmany(part->total_size, NB_BM); bm = part->bitmap; for (bm_e = 0; bm_e < limit; bm_e++, bm++) if (*bm != BM_MASK) break; if (bm_e == limit) panic(here,my_name); /* * Find and set the proper bit */ { bm_entry_t b = *bm; for (bit = 0; bit < NB_BM; bit++) if ((b & (1<free--; } pthread_mutex_unlock(&part->p_lock); return (bm_e*NB_BM+bit); } /* * Deallocate a page in a paging partition */ void pager_dealloc_page(pindex, page, lock_it) p_index_t pindex; vm_offset_t page; boolean_t lock_it; { partition_t part; int bit, bm_e; /* be paranoid */ if (no_partition(pindex)) panic("%sdealloc_page",my_name); ddprintf ("pager_dealloc_page(%d,%x,%d)\n",pindex,page,lock_it); part = partition_of(pindex); if (page >= part->total_size) panic("%sdealloc_page",my_name); bm_e = page / NB_BM; bit = page % NB_BM; if (lock_it) pthread_mutex_lock(&part->p_lock); part->bitmap[bm_e] &= ~(1<free++; if (lock_it) pthread_mutex_unlock(&part->p_lock); } /* * Object sizes are rounded up to the next power of 2, * unless they are bigger than a given maximum size. */ vm_size_t max_doubled_size = 4 * 1024 * 1024; /* 4 meg */ /* * Return first level map for pager. * If there is no such map, than allocate it. */ dp_map_t pager_get_direct_map(pager) dpager_t pager; { dp_map_t mapptr, emapptr; vm_size_t size = pager->size; if (pager->map) return pager->map; /* * Allocate and initialize the block map */ { vm_size_t alloc_size; dp_map_t init_value; if (INDIRECT_PAGEMAP(size)) { alloc_size = INDIRECT_PAGEMAP_SIZE(size); init_value = (dp_map_t)0; } else { alloc_size = PAGEMAP_SIZE(size); init_value = (dp_map_t)NO_BLOCK; } mapptr = (dp_map_t) kalloc(alloc_size); for (emapptr = &mapptr[(alloc_size-1) / sizeof(vm_offset_t)]; emapptr >= mapptr; emapptr--) emapptr->indirect = init_value; } pager->map = mapptr; return mapptr; } /* * Attach a new paging object to a paging partition */ void pager_alloc(pager, part, size) dpager_t pager; p_index_t part; vm_size_t size; /* in BYTES */ { int i; #ifdef CHECKSUM dp_map_t mapptr, emapptr; #endif pthread_mutex_init(&pager->lock, NULL); #if DEBUG_READER_CONFLICTS pager->readers = 0; pager->writer = FALSE; #endif pager->cur_partition = part; /* * Convert byte size to number of pages, then increase to the nearest * power of 2. */ size = atop(size); if (size <= atop(max_doubled_size)) { i = 1; while (i < size) i <<= 1; size = i; } else size = ROUNDUP_TO_PAGEMAP(size); pager->map = NULL; pager->size = size; pager->limit = (vm_size_t)-1; #ifdef CHECKSUM if (INDIRECT_PAGEMAP(size)) { mapptr = (vm_offset_t *) kalloc(INDIRECT_PAGEMAP_SIZE(size)); for (i = INDIRECT_PAGEMAP_ENTRIES(size); --i >= 0;) mapptr[i] = 0; } else { mapptr = (vm_offset_t *) kalloc(PAGEMAP_SIZE(size)); for (i = 0; i < size; i++) mapptr[i] = NO_CHECKSUM; } pager->checksum = mapptr; #endif /* CHECKSUM */ } /* * Return size (in bytes) of space actually allocated to this pager. * The pager is read-locked. */ vm_size_t pager_allocated(pager) dpager_t pager; { vm_size_t size; dp_map_t map, emap; vm_size_t asize; size = pager->size; /* in pages */ asize = 0; /* allocated, in pages */ map = pager_get_direct_map(pager); if (INDIRECT_PAGEMAP(size)) { for (emap = &map[INDIRECT_PAGEMAP_ENTRIES(size)]; map < emap; map++) { dp_map_t map2, emap2; if ((map2 = map->indirect) == 0) continue; for (emap2 = &map2[PAGEMAP_ENTRIES]; map2 < emap2; map2++) if ( ! no_block(*map2) ) asize++; } } else { for (emap = &map[size]; map < emap; map++) if ( ! no_block(*map) ) asize++; } return ptoa(asize); } /* * Find offsets (in the object) of pages actually allocated to this pager. * Returns the number of allocated pages, whether or not they all fit. * The pager is read-locked. */ unsigned int pager_pages(pager, pages, numpages) dpager_t pager; default_pager_page_t *pages; unsigned int numpages; { vm_size_t size; dp_map_t map, emap; unsigned int actual; vm_offset_t offset; size = pager->size; /* in pages */ map = pager_get_direct_map(pager); actual = 0; offset = 0; if (INDIRECT_PAGEMAP(size)) { for (emap = &map[INDIRECT_PAGEMAP_ENTRIES(size)]; map < emap; map++) { dp_map_t map2, emap2; if ((map2 = map->indirect) == 0) { offset += vm_page_size * PAGEMAP_ENTRIES; continue; } for (emap2 = &map2[PAGEMAP_ENTRIES]; map2 < emap2; map2++) if ( ! no_block(*map2) ) { if (actual++ < numpages) pages++->dpp_offset = offset; } offset += vm_page_size; } } else { for (emap = &map[size]; map < emap; map++) { if ( ! no_block(*map) ) { if (actual++ < numpages) pages++->dpp_offset = offset; } offset += vm_page_size; } } return actual; } /* * Extend the map for a paging object. * * XXX This implementation can allocate an arbitrary large amount * of wired memory when extending a big block map. Because vm-privileged * threads call pager_extend, this can crash the system by exhausting * system memory. */ void pager_extend(pager, new_size) dpager_t pager; vm_size_t new_size; /* in pages */ { dp_map_t new_mapptr; dp_map_t old_mapptr; int i; vm_size_t old_size; pthread_mutex_lock(&pager->lock); /* XXX lock_write */ #if DEBUG_READER_CONFLICTS pager->writer = TRUE; #endif /* * Double current size until we cover new size. * If object is 'too big' just use new size. */ old_size = pager->size; if (new_size <= atop(max_doubled_size)) { /* New size cannot be less than 1 */ i = old_size ? old_size : 1; while (i < new_size) i <<= 1; new_size = i; } else new_size = ROUNDUP_TO_PAGEMAP(new_size); if (INDIRECT_PAGEMAP(old_size)) { /* * Pager already uses two levels. Allocate * a larger indirect block. */ new_mapptr = (dp_map_t) kalloc(INDIRECT_PAGEMAP_SIZE(new_size)); old_mapptr = pager_get_direct_map(pager); for (i = 0; i < INDIRECT_PAGEMAP_ENTRIES(old_size); i++) new_mapptr[i] = old_mapptr[i]; for (; i < INDIRECT_PAGEMAP_ENTRIES(new_size); i++) new_mapptr[i].indirect = (dp_map_t)0; kfree((char *)old_mapptr, INDIRECT_PAGEMAP_SIZE(old_size)); pager->map = new_mapptr; pager->size = new_size; #ifdef CHECKSUM new_mapptr = (vm_offset_t *) kalloc(INDIRECT_PAGEMAP_SIZE(new_size)); old_mapptr = pager->checksum; for (i = 0; i < INDIRECT_PAGEMAP_ENTRIES(old_size); i++) new_mapptr[i] = old_mapptr[i]; for (; i < INDIRECT_PAGEMAP_ENTRIES(new_size); i++) new_mapptr[i] = 0; kfree((char *)old_mapptr, INDIRECT_PAGEMAP_SIZE(old_size)); pager->checksum = new_mapptr; #endif /* CHECKSUM */ #if DEBUG_READER_CONFLICTS pager->writer = FALSE; #endif pthread_mutex_unlock(&pager->lock); #if 0 ddprintf ("pager_extend 1 mapptr %x [3b] = %x\n", new_mapptr, new_mapptr[0x3b]); if (new_mapptr[0x3b].indirect > 0x10000 && new_mapptr[0x3b].indirect != NO_BLOCK) panic ("debug panic"); #endif return; } if (INDIRECT_PAGEMAP(new_size)) { /* * Changing from direct map to indirect map. * Allocate both indirect and direct map blocks, * since second-level (direct) block must be * full size (PAGEMAP_SIZE(PAGEMAP_ENTRIES)). */ /* * Allocate new second-level map first. */ new_mapptr = (dp_map_t) kalloc(PAGEMAP_SIZE(PAGEMAP_ENTRIES)); old_mapptr = pager_get_direct_map(pager); for (i = 0; i < old_size; i++) new_mapptr[i] = old_mapptr[i]; for (; i < PAGEMAP_ENTRIES; i++) invalidate_block(new_mapptr[i]); kfree((char *)old_mapptr, PAGEMAP_SIZE(old_size)); old_mapptr = new_mapptr; #if 0 ddprintf ("pager_extend 2 mapptr %x [3b] = %x\n", new_mapptr, new_mapptr[0x3b]); if (new_mapptr[0x3b].indirect > 0x10000 && new_mapptr[0x3b].indirect != NO_BLOCK) panic ("debug panic"); #endif /* * Now allocate indirect map. */ new_mapptr = (dp_map_t) kalloc(INDIRECT_PAGEMAP_SIZE(new_size)); new_mapptr[0].indirect = old_mapptr; for (i = 1; i < INDIRECT_PAGEMAP_ENTRIES(new_size); i++) new_mapptr[i].indirect = 0; pager->map = new_mapptr; pager->size = new_size; #ifdef CHECKSUM /* * Allocate new second-level map first. */ new_mapptr = (vm_offset_t *)kalloc(PAGEMAP_SIZE(PAGEMAP_ENTRIES)); old_mapptr = pager->checksum; for (i = 0; i < old_size; i++) new_mapptr[i] = old_mapptr[i]; for (; i < PAGEMAP_ENTRIES; i++) new_mapptr[i] = NO_CHECKSUM; kfree((char *)old_mapptr, PAGEMAP_SIZE(old_size)); old_mapptr = new_mapptr; /* * Now allocate indirect map. */ new_mapptr = (vm_offset_t *) kalloc(INDIRECT_PAGEMAP_SIZE(new_size)); new_mapptr[0] = (vm_offset_t) old_mapptr; for (i = 1; i < INDIRECT_PAGEMAP_ENTRIES(new_size); i++) new_mapptr[i] = 0; pager->checksum = new_mapptr; #endif /* CHECKSUM */ #if DEBUG_READER_CONFLICTS pager->writer = FALSE; #endif pthread_mutex_unlock(&pager->lock); return; } /* * Enlarging a direct block. */ new_mapptr = (dp_map_t) kalloc(PAGEMAP_SIZE(new_size)); old_mapptr = pager_get_direct_map(pager); for (i = 0; i < old_size; i++) new_mapptr[i] = old_mapptr[i]; for (; i < new_size; i++) invalidate_block(new_mapptr[i]); kfree((char *)old_mapptr, PAGEMAP_SIZE(old_size)); pager->map = new_mapptr; pager->size = new_size; #ifdef CHECKSUM new_mapptr = (vm_offset_t *) kalloc(PAGEMAP_SIZE(new_size)); old_mapptr = pager->checksum; for (i = 0; i < old_size; i++) new_mapptr[i] = old_mapptr[i]; for (; i < new_size; i++) new_mapptr[i] = NO_CHECKSUM; kfree((char *)old_mapptr, PAGEMAP_SIZE(old_size)); pager->checksum = new_mapptr; #endif /* CHECKSUM */ #if DEBUG_READER_CONFLICTS pager->writer = FALSE; #endif pthread_mutex_unlock(&pager->lock); } /* This deallocates the pages necessary to truncate a direct map previously of size NEW_SIZE to the smaller size OLD_SIZE. */ static void dealloc_direct (dp_map_t mapptr, vm_size_t old_size, vm_size_t new_size) { vm_size_t i; if (!mapptr) return; for (i = new_size; i < old_size; ++i) { const union dp_map entry = mapptr[i]; if (!no_block(entry)) { pager_dealloc_page(entry.block.p_index, entry.block.p_offset, TRUE); invalidate_block(mapptr[i]); } } } /* Truncate a memory object. First, any pages between the new size and the (larger) old size are deallocated. Then, the size of the pagemap may be reduced, an indirect map may be turned into a direct map. The pager must be locked by the caller. */ static void pager_truncate(dpager_t pager, vm_size_t new_size) /* in pages */ { int i; vm_size_t old_size; pthread_mutex_lock(&pager->lock); /* XXX lock_write */ if (!pager->map) goto done; old_size = pager->size; if (INDIRECT_PAGEMAP(old_size)) { /* First handle the entire second-levels blocks that are being freed. */ for (i = INDIRECT_PAGEMAP_ENTRIES(new_size); i < INDIRECT_PAGEMAP_ENTRIES(old_size); ++i) { const dp_map_t mapptr = pager->map[i].indirect; pager->map[i].indirect = (dp_map_t)0; dealloc_direct (mapptr, PAGEMAP_ENTRIES, 0); kfree ((char *)mapptr, PAGEMAP_SIZE(PAGEMAP_ENTRIES)); } /* Now truncate what's now the final nonempty direct block. */ dealloc_direct (pager->map[(new_size - 1) / PAGEMAP_ENTRIES].indirect, old_size & (PAGEMAP_ENTRIES - 1), new_size & (PAGEMAP_ENTRIES - 1)); if (INDIRECT_PAGEMAP (new_size)) { const dp_map_t old_mapptr = pager->map; pager->map = (dp_map_t) kalloc (INDIRECT_PAGEMAP_SIZE(new_size)); memcpy (pager->map, old_mapptr, INDIRECT_PAGEMAP_SIZE(new_size)); kfree ((char *) old_mapptr, INDIRECT_PAGEMAP_SIZE (old_size)); } else { /* We are truncating to a size small enough that it goes to using a one-level map. We already have that map, as the first and only nonempty element in our indirect map. */ const dp_map_t mapptr = pager->map[0].indirect; kfree((char *)pager->map, INDIRECT_PAGEMAP_SIZE(old_size)); pager->map = mapptr; } } if (! INDIRECT_PAGEMAP(old_size)) { /* First deallocate pages in the truncated region. */ dealloc_direct (pager->map, old_size, new_size); /* Now reduce the size of the direct map itself. We don't bother with kalloc/kfree if it's not shrinking enough that kalloc.c would actually use less. */ if (PAGEMAP_SIZE (new_size) <= PAGEMAP_SIZE (old_size) / 2) { const dp_map_t old_mapptr = pager->map; pager->map = (dp_map_t) kalloc (PAGEMAP_SIZE (new_size)); memcpy (pager->map, old_mapptr, PAGEMAP_SIZE (new_size)); kfree ((char *) old_mapptr, PAGEMAP_SIZE (old_size)); } } done: pager->size = new_size; pthread_mutex_unlock(&pager->lock); #ifdef CHECKSUM #error write me #endif /* CHECKSUM */ } /* * Given an offset within a paging object, find the * corresponding block within the paging partition. * Return NO_BLOCK if none allocated. */ union dp_map pager_read_offset(pager, offset) dpager_t pager; vm_offset_t offset; { vm_offset_t f_page; union dp_map pager_offset; f_page = atop(offset); #if DEBUG_READER_CONFLICTS if (pager->readers > 0) default_pager_read_conflicts++; /* would have proceeded with read/write lock */ #endif pthread_mutex_lock(&pager->lock); /* XXX lock_read */ #if DEBUG_READER_CONFLICTS pager->readers++; #endif if (f_page >= pager->size) { ddprintf ("%spager_read_offset pager %x: bad page %d >= size %d", my_name, pager, f_page, pager->size); pthread_mutex_unlock(&pager->lock); return (union dp_map) (union dp_map *) NO_BLOCK; #if 0 panic("%spager_read_offset",my_name); #endif } invalidate_block(pager_offset); if (INDIRECT_PAGEMAP(pager->size)) { dp_map_t mapptr; if (pager->map) { mapptr = pager->map[f_page/PAGEMAP_ENTRIES].indirect; if (mapptr) pager_offset = mapptr[f_page%PAGEMAP_ENTRIES]; } } else { if (pager->map) pager_offset = pager->map[f_page]; } #if DEBUG_READER_CONFLICTS pager->readers--; #endif pthread_mutex_unlock(&pager->lock); return (pager_offset); } #if USE_PRECIOUS /* * Release a single disk block. */ void pager_release_offset(pager, offset) dpager_t pager; vm_offset_t offset; { union dp_map entry; offset = atop(offset); pthread_mutex_lock(&pager->lock); /* XXX lock_read */ assert (pager->map); if (INDIRECT_PAGEMAP(pager->size)) { dp_map_t mapptr; mapptr = pager->map[offset / PAGEMAP_ENTRIES].indirect; entry = mapptr[offset % PAGEMAP_ENTRIES]; invalidate_block(mapptr[offset % PAGEMAP_ENTRIES]); } else { entry = pager->map[offset]; invalidate_block(pager->map[offset]); } pthread_mutex_unlock(&pager->lock); pager_dealloc_page(entry.block.p_index, entry.block.p_offset, TRUE); } #endif /*USE_PRECIOUS*/ /* * Move a page from one partition to another * New partition is locked, old partition is * locked unless LOCK_OLD sez otherwise. */ union dp_map pager_move_page(block) union dp_map block; { partition_t old_part, new_part; p_index_t old_pindex, new_pindex; union dp_map ret; vm_size_t size; vm_offset_t raddr, offset, new_offset; kern_return_t rc; static char here[] = "%spager_move_page"; old_pindex = block.block.p_index; invalidate_block(ret); /* See if we have room to put it anywhere else */ new_pindex = choose_partition( ptoa(1), old_pindex); if (no_partition(new_pindex)) return ret; /* this unlocks the new partition */ new_offset = pager_alloc_page(new_pindex, FALSE); if (new_offset == NO_BLOCK) panic(here,my_name); /* * Got the resources, now move the data */ ddprintf ("pager_move_page(%x,%d,%d)\n",block.block.p_offset,old_pindex,new_pindex); old_part = partition_of(old_pindex); offset = ptoa(block.block.p_offset); rc = page_read_file_direct (old_part->file, offset, vm_page_size, &raddr, &size); if (rc != 0) panic(here,my_name); /* release old */ pager_dealloc_page(old_pindex, block.block.p_offset, FALSE); new_part = partition_of(new_pindex); offset = ptoa(new_offset); rc = page_write_file_direct (new_part->file, offset, raddr, size, &size); if (rc != 0) panic(here,my_name); (void) vm_deallocate( mach_task_self(), raddr, size); ret.block.p_offset = new_offset; ret.block.p_index = new_pindex; return ret; } #ifdef CHECKSUM /* * Return the checksum for a block. */ int pager_get_checksum(pager, offset) dpager_t pager; vm_offset_t offset; { vm_offset_t f_page; int checksum; f_page = atop(offset); pthread_mutex_lock(&pager->lock); /* XXX lock_read */ if (f_page >= pager->size) panic("%spager_get_checksum",my_name); if (INDIRECT_PAGEMAP(pager->size)) { vm_offset_t *mapptr; mapptr = (vm_offset_t *)pager->checksum[f_page/PAGEMAP_ENTRIES]; if (mapptr == 0) checksum = NO_CHECKSUM; else checksum = mapptr[f_page%PAGEMAP_ENTRIES]; } else { checksum = pager->checksum[f_page]; } pthread_mutex_unlock(&pager->lock); return (checksum); } /* * Remember the checksum for a block. */ int pager_put_checksum(pager, offset, checksum) dpager_t pager; vm_offset_t offset; int checksum; { vm_offset_t f_page; static char here[] = "%spager_put_checksum"; f_page = atop(offset); pthread_mutex_lock(&pager->lock); /* XXX lock_read */ if (f_page >= pager->size) panic(here,my_name); if (INDIRECT_PAGEMAP(pager->size)) { vm_offset_t *mapptr; mapptr = (vm_offset_t *)pager->checksum[f_page/PAGEMAP_ENTRIES]; if (mapptr == 0) panic(here,my_name); mapptr[f_page%PAGEMAP_ENTRIES] = checksum; } else { pager->checksum[f_page] = checksum; } pthread_mutex_unlock(&pager->lock); } /* * Compute a checksum - XOR each 32-bit word. */ int compute_checksum(addr, size) vm_offset_t addr; vm_size_t size; { int checksum = NO_CHECKSUM; int *ptr; int count; ptr = (int *)addr; count = size / sizeof(int); while (--count >= 0) checksum ^= *ptr++; return (checksum); } #endif /* CHECKSUM */ /* * Given an offset within a paging object, find the * corresponding block within the paging partition. * Allocate a new block if necessary. * * WARNING: paging objects apparently may be extended * without notice! */ union dp_map pager_write_offset(pager, offset) dpager_t pager; vm_offset_t offset; { vm_offset_t f_page; dp_map_t mapptr; union dp_map block; invalidate_block(block); f_page = atop(offset); #if DEBUG_READER_CONFLICTS if (pager->readers > 0) default_pager_read_conflicts++; /* would have proceeded with read/write lock */ #endif pthread_mutex_lock(&pager->lock); /* XXX lock_read */ #if DEBUG_READER_CONFLICTS pager->readers++; #endif /* Catch the case where we had no initial fit partition for this object, but one was added later on */ if (no_partition(pager->cur_partition)) { p_index_t new_part; vm_size_t size; size = (f_page > pager->size) ? f_page : pager->size; new_part = choose_partition(ptoa(size), P_INDEX_INVALID); if (no_partition(new_part)) new_part = choose_partition(ptoa(1), P_INDEX_INVALID); if (no_partition(new_part)) /* give up right now to avoid confusion */ goto out; else pager->cur_partition = new_part; } while (f_page >= pager->size) { ddprintf ("pager_write_offset: extending: %x %x\n", f_page, pager->size); /* * Paging object must be extended. * Remember that offset is 0-based, but size is 1-based. */ #if DEBUG_READER_CONFLICTS pager->readers--; #endif pthread_mutex_unlock(&pager->lock); pager_extend(pager, f_page + 1); #if DEBUG_READER_CONFLICTS if (pager->readers > 0) default_pager_read_conflicts++; /* would have proceeded with read/write lock */ #endif pthread_mutex_lock(&pager->lock); /* XXX lock_read */ #if DEBUG_READER_CONFLICTS pager->readers++; #endif ddprintf ("pager_write_offset: done extending: %x %x\n", f_page, pager->size); } if (INDIRECT_PAGEMAP(pager->size)) { ddprintf ("pager_write_offset: indirect\n"); mapptr = pager_get_direct_map(pager); mapptr = mapptr[f_page/PAGEMAP_ENTRIES].indirect; if (mapptr == 0) { /* * Allocate the indirect block */ int i; ddprintf ("pager_write_offset: allocating indirect\n"); mapptr = (dp_map_t) kalloc(PAGEMAP_SIZE(PAGEMAP_ENTRIES)); if (mapptr == 0) { /* out of space! */ no_paging_space(TRUE); goto out; } pager->map[f_page/PAGEMAP_ENTRIES].indirect = mapptr; for (i = 0; i < PAGEMAP_ENTRIES; i++) invalidate_block(mapptr[i]); #ifdef CHECKSUM { vm_offset_t *cksumptr; int j; cksumptr = (vm_offset_t *) kalloc(PAGEMAP_SIZE(PAGEMAP_ENTRIES)); if (cksumptr == 0) { /* out of space! */ no_paging_space(TRUE); goto out; } pager->checksum[f_page/PAGEMAP_ENTRIES] = (vm_offset_t)cksumptr; for (j = 0; j < PAGEMAP_ENTRIES; j++) cksumptr[j] = NO_CHECKSUM; } #endif /* CHECKSUM */ } f_page %= PAGEMAP_ENTRIES; } else { mapptr = pager_get_direct_map(pager); } block = mapptr[f_page]; ddprintf ("pager_write_offset: block starts as %x[%x] %x\n", mapptr, f_page, block); if (no_block(block)) { vm_offset_t off; /* get room now */ off = pager_alloc_page(pager->cur_partition, TRUE); if (off == NO_BLOCK) { /* * Before giving up, try all other partitions. */ p_index_t new_part; ddprintf ("pager_write_offset: could not allocate block\n"); /* returns it locked (if any one is non-full) */ new_part = choose_partition( ptoa(1), pager->cur_partition); if ( ! no_partition(new_part) ) { #if debug dprintf("%s partition %x filled,", my_name, pager->cur_partition); dprintf("extending object %x (size %x) to %x.\n", pager, pager->size, new_part); #endif /* this one tastes better */ pager->cur_partition = new_part; /* this unlocks the partition too */ off = pager_alloc_page(pager->cur_partition, FALSE); } if (off == NO_BLOCK) { /* * Oh well. */ overcommitted(FALSE, 1); goto out; } ddprintf ("pager_write_offset: decided to allocate block\n"); } block.block.p_offset = off; block.block.p_index = pager->cur_partition; mapptr[f_page] = block; } out: #if DEBUG_READER_CONFLICTS pager->readers--; #endif pthread_mutex_unlock(&pager->lock); return (block); } /* * Deallocate all of the blocks belonging to a paging object. * No locking needed because no other operations can be in progress. */ void pager_dealloc(pager) dpager_t pager; { int i, j; dp_map_t mapptr; union dp_map block; if (!pager->map) return; if (INDIRECT_PAGEMAP(pager->size)) { for (i = INDIRECT_PAGEMAP_ENTRIES(pager->size); --i >= 0; ) { mapptr = pager->map[i].indirect; if (mapptr != 0) { for (j = 0; j < PAGEMAP_ENTRIES; j++) { block = mapptr[j]; if ( ! no_block(block) ) pager_dealloc_page(block.block.p_index, block.block.p_offset, TRUE); } kfree((char *)mapptr, PAGEMAP_SIZE(PAGEMAP_ENTRIES)); pager->map[i].indirect = (dp_map_t) 0; } } kfree((char *)pager->map, INDIRECT_PAGEMAP_SIZE(pager->size)); pager->map = (dp_map_t) 0; #ifdef CHECKSUM for (i = INDIRECT_PAGEMAP_ENTRIES(pager->size); --i >= 0; ) { mapptr = (vm_offset_t *)pager->checksum[i]; if (mapptr) { kfree((char *)mapptr, PAGEMAP_SIZE(PAGEMAP_ENTRIES)); } } kfree((char *)pager->checksum, INDIRECT_PAGEMAP_SIZE(pager->size)); #endif /* CHECKSUM */ } else { mapptr = pager->map; for (i = 0; i < pager->size; i++ ) { block = mapptr[i]; if ( ! no_block(block) ) pager_dealloc_page(block.block.p_index, block.block.p_offset, TRUE); } kfree((char *)pager->map, PAGEMAP_SIZE(pager->size)); pager->map = (dp_map_t) 0; #ifdef CHECKSUM kfree((char *)pager->checksum, PAGEMAP_SIZE(pager->size)); #endif /* CHECKSUM */ } } /* * Move all the pages of a PAGER that live in a * partition PINDEX somewhere else. * Pager should be write-locked, partition too. * Returns FALSE if it could not do it, but * some pages might have been moved nonetheless. */ boolean_t pager_realloc(pager, pindex) dpager_t pager; p_index_t pindex; { dp_map_t map, emap; vm_size_t size; union dp_map block; if (!pager->map) return TRUE; size = pager->size; /* in pages */ map = pager->map; if (INDIRECT_PAGEMAP(size)) { for (emap = &map[INDIRECT_PAGEMAP_ENTRIES(size)]; map < emap; map++) { dp_map_t map2, emap2; if ((map2 = map->indirect) == 0) continue; for (emap2 = &map2[PAGEMAP_ENTRIES]; map2 < emap2; map2++) if ( map2->block.p_index == pindex) { block = pager_move_page(*map2); if (!no_block(block)) *map2 = block; else return FALSE; } } goto ok; } /* A small one */ for (emap = &map[size]; map < emap; map++) if (map->block.p_index == pindex) { block = pager_move_page(*map); if (!no_block(block)) *map = block; else return FALSE; } ok: pager->cur_partition = choose_partition(0, P_INDEX_INVALID); return TRUE; } /* * Read/write routines. */ #define PAGER_SUCCESS 0 #define PAGER_ABSENT 1 #define PAGER_ERROR 2 /* * Read data from a default pager. Addr is the address of a buffer * to fill. Out_addr returns the buffer that contains the data; * if it is different from , it must be deallocated after use. */ int default_read(ds, addr, size, offset, out_addr, deallocate, external) dpager_t ds; vm_offset_t addr; /* pointer to block to fill */ vm_size_t size; vm_offset_t offset; vm_offset_t *out_addr; /* returns pointer to data */ boolean_t deallocate; boolean_t external; { union dp_map block; vm_offset_t raddr; vm_size_t rsize; int rc; boolean_t first_time; partition_t part; #ifdef CHECKSUM vm_size_t original_size = size; #endif /* CHECKSUM */ vm_offset_t original_offset = offset; /* * Find the block in the paging partition */ block = pager_read_offset(ds, offset); if ( no_block(block) ) { if (external) { /* * An external object is requesting unswapped data, * zero fill the page and return. */ bzero((char *) addr, vm_page_size); *out_addr = addr; return (PAGER_SUCCESS); } return (PAGER_ABSENT); } /* * Read it, trying for the entire page. */ offset = ptoa(block.block.p_offset); ddprintf ("default_read(%x,%x,%x,%d)\n",addr,size,offset,block.block.p_index); part = partition_of(block.block.p_index); first_time = TRUE; *out_addr = addr; do { rc = page_read_file_direct(part->file, offset, size, &raddr, &rsize); if (rc != 0) return (PAGER_ERROR); /* * If we got the entire page on the first read, return it. */ if (first_time && rsize == size) { *out_addr = raddr; break; } /* * Otherwise, copy the data into the * buffer we were passed, and try for * the next piece. */ first_time = FALSE; bcopy((char *)raddr, (char *)addr, rsize); addr += rsize; offset += rsize; size -= rsize; } while (size != 0); #if USE_PRECIOUS if (deallocate) pager_release_offset(ds, original_offset); #endif /*USE_PRECIOUS*/ #ifdef CHECKSUM { int write_checksum, read_checksum; write_checksum = pager_get_checksum(ds, original_offset); read_checksum = compute_checksum(*out_addr, original_size); if (write_checksum != read_checksum) { panic( "PAGER CHECKSUM ERROR: offset 0x%x, written 0x%x, read 0x%x", original_offset, write_checksum, read_checksum); } } #endif /* CHECKSUM */ return (PAGER_SUCCESS); } int default_write(ds, addr, size, offset) dpager_t ds; vm_offset_t addr; vm_size_t size; vm_offset_t offset; { union dp_map block; partition_t part; vm_size_t wsize; int rc; ddprintf ("default_write: pager offset %x\n", offset); /* * Find block in paging partition */ block = pager_write_offset(ds, offset); if ( no_block(block) ) return (PAGER_ERROR); #ifdef CHECKSUM /* * Save checksum */ { int checksum; checksum = compute_checksum(addr, size); pager_put_checksum(ds, offset, checksum); } #endif /* CHECKSUM */ offset = ptoa(block.block.p_offset); ddprintf ("default_write(%x,%x,%x,%d)\n",addr,size,offset,block.block.p_index); part = partition_of(block.block.p_index); /* * There are various assumptions made here,we * will not get into the next disk 'block' by * accident. It might well be non-contiguous. */ do { rc = page_write_file_direct(part->file, offset, addr, size, &wsize); if (rc != 0) { dprintf("*** PAGER ERROR: default_write: "); dprintf("ds=0x%x addr=0x%x size=0x%x offset=0x%x resid=0x%x\n", ds, addr, size, offset, wsize); return (PAGER_ERROR); } addr += wsize; offset += wsize; size -= wsize; } while (size != 0); return (PAGER_SUCCESS); } boolean_t default_has_page(ds, offset) dpager_t ds; vm_offset_t offset; { return ( ! no_block(pager_read_offset(ds, offset)) ); } #if PARALLEL #define dstruct_lock_init(ds) pthread_mutex_init(&ds->lock, NULL) #define dstruct_lock(ds) pthread_mutex_lock(&ds->lock) #define dstruct_unlock(ds) pthread_mutex_unlock(&ds->lock) #else /* PARALLEL */ #define dstruct_lock_init(ds) #define dstruct_lock(ds) #define dstruct_unlock(ds) #endif /* PARALLEL */ struct pager_port all_pagers; #define pager_port_list_init() \ { \ pthread_mutex_init(&all_pagers.lock, NULL); \ hurd_ihash_init (&all_pagers.htable, \ offsetof (struct dstruct, htable_locp)); \ queue_init(&all_pagers.leak_queue); \ } void pager_port_list_insert(port, ds) mach_port_t port; default_pager_t ds; { pthread_mutex_lock(&all_pagers.lock); hurd_ihash_add (&all_pagers.htable, (hurd_ihash_key_t) port, (hurd_ihash_value_t) ds); pthread_mutex_unlock(&all_pagers.lock); /* Try to set a protected payload. This is an optimization, if it fails we degrade gracefully. */ mach_port_set_protected_payload (mach_task_self (), port, (unsigned long) ds); } void pager_port_list_delete(ds) default_pager_t ds; { pthread_mutex_lock(&all_pagers.lock); hurd_ihash_locp_remove (&all_pagers.htable, ds->htable_locp); pthread_mutex_unlock(&all_pagers.lock); mach_port_clear_protected_payload (mach_task_self (), ds->pager); } /* * Destroy a paging partition. * XXX this is not re-entrant XXX */ kern_return_t destroy_paging_partition(name, pp_private) char *name; void **pp_private; { unsigned int id = part_id(name); partition_t part = NULL; boolean_t all_ok = TRUE; default_pager_t entry; int pindex; /* * Find and take partition out of list * This prevents choose_partition from * getting in the way. */ pthread_mutex_lock(&all_partitions.lock); for (pindex = 0; pindex < all_partitions.n_partitions; pindex++) { part = partition_of(pindex); if (part && (part->id == id)) break; } if (! part) { pthread_mutex_unlock(&all_partitions.lock); return KERN_INVALID_ARGUMENT; } part->going_away = TRUE; pthread_mutex_unlock(&all_partitions.lock); /* * This might take a while.. */ all_over_again: #if debug dprintf("Partition x%x (id x%x) for %s, all_ok %d\n", part, id, name, all_ok); #endif all_ok = TRUE; pthread_mutex_lock(&part->p_lock); pthread_mutex_lock(&all_pagers.lock); HURD_IHASH_ITERATE (&all_pagers.htable, val) { entry = (default_pager_t) val; dstruct_lock(entry); if (pthread_mutex_trylock(&entry->dpager.lock)) { dstruct_unlock(entry); pthread_mutex_unlock(&all_pagers.lock); pthread_mutex_unlock(&part->p_lock); /* yield the processor */ (void) thread_switch(MACH_PORT_NULL, SWITCH_OPTION_NONE, 0); goto all_over_again; } /* * See if we can relocate all the pages of this object * currently on this partition on some other partition */ all_ok = pager_realloc(&entry->dpager, pindex); pthread_mutex_unlock(&entry->dpager.lock); dstruct_unlock(entry); if (!all_ok) break; } pthread_mutex_unlock(&all_pagers.lock); if (all_ok) { /* No need to unlock partition, there are no refs left */ set_partition_of(pindex, 0); *pp_private = part->file; kfree(part->bitmap, howmany(part->total_size, NB_BM) * sizeof(bm_entry_t)); kfree(part, sizeof(struct part)); dprintf("%s Removed paging partition %s\n", my_name, name); return KERN_SUCCESS; } /* * Put partition back in. */ part->going_away = FALSE; return KERN_FAILURE; } /* * We use the sequence numbers on requests to regulate * our parallelism. In general, we allow multiple reads and writes * to proceed in parallel, with the exception that reads must * wait for previous writes to finish. (Because the kernel might * generate a data-request for a page on the heels of a data-write * for the same page, and we must avoid returning stale data.) * terminate requests wait for proceeding reads and writes to finish. */ unsigned int default_pager_total = 0; /* debugging */ unsigned int default_pager_wait_seqno = 0; /* debugging */ unsigned int default_pager_wait_read = 0; /* debugging */ unsigned int default_pager_wait_write = 0; /* debugging */ unsigned int default_pager_wait_refs = 0; /* debugging */ #if PARALLEL /* * Waits for correct sequence number. Leaves pager locked. */ void pager_port_lock(ds, seqno) default_pager_t ds; mach_port_seqno_t seqno; { default_pager_total++; dstruct_lock(ds); while (ds->seqno != seqno) { default_pager_wait_seqno++; pthread_cond_wait(&ds->waiting_seqno, &ds->lock); } } /* * Increments sequence number and unlocks pager. */ void pager_port_unlock(ds) default_pager_t ds; { ds->seqno++; dstruct_unlock(ds); pthread_cond_broadcast(&ds->waiting_seqno); } /* * Start a read - one more reader. Pager must be locked. */ void pager_port_start_read(ds) default_pager_t ds; { ds->readers++; } /* * Wait for readers. Unlocks and relocks pager if wait needed. */ void pager_port_wait_for_readers(ds) default_pager_t ds; { while (ds->readers != 0) { default_pager_wait_read++; pthread_cond_wait(&ds->waiting_read, &ds->lock); } } /* * Finish a read. Pager is unlocked and returns unlocked. */ void pager_port_finish_read(ds) default_pager_t ds; { dstruct_lock(ds); if (--ds->readers == 0) { dstruct_unlock(ds); pthread_cond_broadcast(&ds->waiting_read); } else { dstruct_unlock(ds); } } /* * Start a write - one more writer. Pager must be locked. */ void pager_port_start_write(ds) default_pager_t ds; { ds->writers++; } /* * Wait for writers. Unlocks and relocks pager if wait needed. */ void pager_port_wait_for_writers(ds) default_pager_t ds; { while (ds->writers != 0) { default_pager_wait_write++; pthread_cond_wait(&ds->waiting_write, &ds->lock); } } /* * Finish a write. Pager is unlocked and returns unlocked. */ void pager_port_finish_write(ds) default_pager_t ds; { dstruct_lock(ds); if (--ds->writers == 0) { dstruct_unlock(ds); pthread_cond_broadcast(&ds->waiting_write); } else { dstruct_unlock(ds); } } /* * Wait for concurrent default_pager_objects. * Unlocks and relocks pager if wait needed. */ void pager_port_wait_for_refs(ds) default_pager_t ds; { while (ds->name_refs == 0) { default_pager_wait_refs++; pthread_cond_wait(&ds->waiting_refs, &ds->lock); } } /* * Finished creating name refs - wake up waiters. */ void pager_port_finish_refs(ds) default_pager_t ds; { pthread_cond_broadcast(&ds->waiting_refs); } #else /* PARALLEL */ #define pager_port_lock(ds,seqno) #define pager_port_unlock(ds) #define pager_port_start_read(ds) #define pager_port_wait_for_readers(ds) #define pager_port_finish_read(ds) #define pager_port_start_write(ds) #define pager_port_wait_for_writers(ds) #define pager_port_finish_write(ds) #define pager_port_wait_for_refs(ds) #define pager_port_finish_refs(ds) #endif /* PARALLEL */ /* * Default pager. */ task_t default_pager_self; /* Our task port. */ mach_port_t default_pager_default_port; /* Port for memory_object_create. */ /* We catch exceptions on ourself & startup using this port. */ mach_port_t default_pager_exception_port; mach_port_t default_pager_internal_set; /* Port set for internal objects. */ mach_port_t default_pager_external_set; /* Port set for external objects. */ mach_port_t default_pager_default_set; /* Port set for "default" thread. */ typedef struct default_pager_thread { pthread_t dpt_thread; /* Server thread. */ vm_offset_t dpt_buffer; /* Read buffer. */ boolean_t dpt_internal; /* Do we handle internal objects? */ } default_pager_thread_t; #if PARALLEL /* determine number of threads at run time */ #define DEFAULT_PAGER_INTERNAL_COUNT (0) #else /* PARALLEL */ #define DEFAULT_PAGER_INTERNAL_COUNT (1) #endif /* PARALLEL */ /* Memory created by default_pager_object_create should mostly be resident. */ #define DEFAULT_PAGER_EXTERNAL_COUNT (1) unsigned int default_pager_internal_count = DEFAULT_PAGER_INTERNAL_COUNT; /* Number of "internal" threads. */ unsigned int default_pager_external_count = DEFAULT_PAGER_EXTERNAL_COUNT; /* Number of "external" threads. */ default_pager_t pager_port_alloc(size) vm_size_t size; { default_pager_t ds; p_index_t part; ds = (default_pager_t) kalloc(sizeof *ds); if (ds == DEFAULT_PAGER_NULL) panic("%spager_port_alloc",my_name); bzero((char *) ds, sizeof *ds); dstruct_lock_init(ds); /* * Get a suitable partition. If none big enough * just pick one and overcommit. If no partitions * at all.. well just fake one so that we will * kill specific objects on pageouts rather than * panicing the system now. */ part = choose_partition(size, P_INDEX_INVALID); if (no_partition(part)) { overcommitted(FALSE, atop(size)); part = choose_partition(0,P_INDEX_INVALID); #if debug if (no_partition(part)) dprintf("%s No paging space at all !!\n", my_name); #endif } pager_alloc(&ds->dpager, part, size); return ds; } mach_port_urefs_t default_pager_max_urefs = 10000; /* * Check user reference count on pager_request port. * Pager must be locked. * Unlocks and re-locks pager if needs to call kernel. */ void pager_port_check_request(ds, pager_request) default_pager_t ds; mach_port_t pager_request; { mach_port_delta_t delta; kern_return_t kr; assert(ds->pager_request == pager_request); if (++ds->request_refs > default_pager_max_urefs) { delta = 1 - ds->request_refs; ds->request_refs = 1; dstruct_unlock(ds); /* * Deallocate excess user references. */ kr = mach_port_mod_refs(default_pager_self, pager_request, MACH_PORT_RIGHT_SEND, delta); if (kr != KERN_SUCCESS) panic("%spager_port_check_request",my_name); dstruct_lock(ds); } } void default_pager_add(ds, internal) default_pager_t ds; boolean_t internal; { mach_port_t pager = ds->pager; mach_port_t pset; mach_port_mscount_t sync; mach_port_t previous; kern_return_t kr; static char here[] = "%sdefault_pager_add"; /* * The port currently has a make-send count of zero, * because either we just created the port or we just * received the port in a memory_object_create request. */ if (internal) { /* possibly generate an immediate no-senders notification */ sync = 0; pset = default_pager_internal_set; ds->external = FALSE; } else { /* delay notification till send right is created */ sync = 1; pset = default_pager_external_set; ds->external = TRUE; } kr = mach_port_request_notification(default_pager_self, pager, MACH_NOTIFY_NO_SENDERS, sync, pager, MACH_MSG_TYPE_MAKE_SEND_ONCE, &previous); if ((kr != KERN_SUCCESS) || (previous != MACH_PORT_NULL)) panic(here,my_name); kr = mach_port_move_member(default_pager_self, pager, pset); if (kr != KERN_SUCCESS) panic(here,my_name); } /* * Routine: memory_object_create * Purpose: * Handle requests for memory objects from the * kernel. * Notes: * Because we only give out the default memory * manager port to the kernel, we don't have to * be so paranoid about the contents. */ kern_return_t seqnos_memory_object_create(old_pager, seqno, new_pager, new_size, new_pager_request, new_pager_name, new_page_size) mach_port_t old_pager; mach_port_seqno_t seqno; mach_port_t new_pager; vm_size_t new_size; mach_port_t new_pager_request; mach_port_t new_pager_name; vm_size_t new_page_size; { default_pager_t ds; assert(old_pager == default_pager_default_port); assert(MACH_PORT_VALID(new_pager_request)); assert(MACH_PORT_VALID(new_pager_name)); assert(new_page_size == vm_page_size); ds = pager_port_alloc(new_size); /* * Set up associations between these ports * and this default_pager structure */ ds->pager = new_pager; ds->pager_request = new_pager_request; ds->request_refs = 1; ds->pager_name = new_pager_name; ds->name_refs = 1; /* * After this, other threads might receive requests * for this memory object or find it in the port list. */ pager_port_list_insert(new_pager, ds); default_pager_add(ds, TRUE); return(KERN_SUCCESS); } memory_object_copy_strategy_t default_pager_copy_strategy = MEMORY_OBJECT_COPY_DELAY; kern_return_t seqnos_memory_object_init(ds, seqno, pager_request, pager_name, pager_page_size) default_pager_t ds; mach_port_seqno_t seqno; mach_port_t pager_request; mach_port_t pager_name; vm_size_t pager_page_size; { kern_return_t kr; static char here[] = "%sinit"; assert(MACH_PORT_VALID(pager_request)); assert(MACH_PORT_VALID(pager_name)); assert(pager_page_size == vm_page_size); if (ds == DEFAULT_PAGER_NULL) panic(here, my_name); pager_port_lock(ds, seqno); if (ds->pager_request != MACH_PORT_NULL) panic(here, my_name); ds->pager_request = pager_request; ds->request_refs = 1; ds->pager_name = pager_name; ds->name_refs = 1; /* * Even if the kernel immediately terminates the object, * the pager_request port won't be destroyed until * we process the terminate request, which won't happen * until we unlock the object. */ kr = memory_object_ready(pager_request, FALSE, /* Do not cache */ default_pager_copy_strategy); if (kr != KERN_SUCCESS) panic(here, my_name); pager_port_unlock(ds); return(KERN_SUCCESS); } kern_return_t seqnos_memory_object_terminate(ds, seqno, pager_request, pager_name) default_pager_t ds; mach_port_seqno_t seqno; mach_port_t pager_request; mach_port_t pager_name; { static char here[] = "%sterminate"; /* * pager_request and pager_name are receive rights, * not send rights. */ if (ds == DEFAULT_PAGER_NULL) panic(here, my_name); ddprintf ("seqnos_memory_object_terminate <%p>: pager_port_lock: <%p>[s:%d,r:%d,w:%d,l:%d], %d\n", &ds, ds, ds->seqno, ds->readers, ds->writers, ds->lock.held, seqno); pager_port_lock(ds, seqno); /* * Wait for read and write requests to terminate. */ pager_port_wait_for_readers(ds); pager_port_wait_for_writers(ds); /* * After memory_object_terminate both memory_object_init * and a no-senders notification are possible, so we need * to clean up the request and name ports but leave * the pager port. * * A concurrent default_pager_objects might be allocating * more references for the name port. In this case, * we must first wait for it to finish. */ pager_port_wait_for_refs(ds); if (ds->external) pager_request = ds->pager_request; ds->pager_request = MACH_PORT_NULL; ds->request_refs = 0; assert(ds->pager_name == pager_name); ds->pager_name = MACH_PORT_NULL; ds->name_refs = 0; ddprintf ("seqnos_memory_object_terminate <%p>: pager_port_unlock: <%p>[s:%d,r:%d,w:%d,l:%d]\n", &ds, ds, ds->seqno, ds->readers, ds->writers, ds->lock.held); pager_port_unlock(ds); /* * Now we destroy our port rights. */ mach_port_destroy(mach_task_self(), pager_request); mach_port_destroy(mach_task_self(), pager_name); return (KERN_SUCCESS); } void default_pager_no_senders(pager, seqno, mscount) memory_object_t pager; mach_port_seqno_t seqno; mach_port_mscount_t mscount; { default_pager_t ds; kern_return_t kr; static char here[] = "%sno_senders"; /* * Because we don't give out multiple send rights * for a memory object, there can't be a race * between getting a no-senders notification * and creating a new send right for the object. * Hence we don't keep track of mscount. */ ds = begin_using_default_pager(pager); if (ds == DEFAULT_PAGER_NULL) panic(here,my_name); pager_port_lock(ds, seqno); /* * We shouldn't get a no-senders notification * when the kernel has the object cached. */ if (ds->pager_request != MACH_PORT_NULL) panic(here,my_name); /* * Unlock the pager (though there should be no one * waiting for it). */ dstruct_unlock(ds); /* * Remove the memory object port association, and then * the destroy the port itself. We must remove the object * from the port list before deallocating the pager, * because of default_pager_objects. */ pager_port_list_delete(ds); pager_dealloc(&ds->dpager); kr = mach_port_mod_refs(default_pager_self, pager, MACH_PORT_RIGHT_RECEIVE, -1); if (kr != KERN_SUCCESS) panic(here,my_name); /* * Do this *after* deallocating the port name */ kfree((char *) ds, sizeof(*ds)); /* * Recover memory that we might have wasted because * of name conflicts */ pthread_mutex_lock(&all_pagers.lock); while (!queue_empty(&all_pagers.leak_queue)) { ds = (default_pager_t) queue_first(&all_pagers.leak_queue); queue_remove_first(&all_pagers.leak_queue, ds, default_pager_t, links); kfree((char *) ds, sizeof(*ds)); } pthread_mutex_unlock(&all_pagers.lock); } int default_pager_pagein_count = 0; int default_pager_pageout_count = 0; static __thread default_pager_thread_t *dpt; kern_return_t seqnos_memory_object_data_request(ds, seqno, reply_to, offset, length, protection_required) default_pager_t ds; mach_port_seqno_t seqno; mach_port_t reply_to; vm_offset_t offset; vm_size_t length; vm_prot_t protection_required; { vm_offset_t addr; unsigned int errors; kern_return_t rc; static char here[] = "%sdata_request"; if (length != vm_page_size) panic(here,my_name); if (ds == DEFAULT_PAGER_NULL) panic(here,my_name); ddprintf ("seqnos_memory_object_data_request <%p>: pager_port_lock: <%p>[s:%d,r:%d,w:%d,l:%d], %d\n", &ds, ds, ds->seqno, ds->readers, ds->writers, ds->lock.held, seqno); pager_port_lock(ds, seqno); pager_port_check_request(ds, reply_to); pager_port_wait_for_writers(ds); pager_port_start_read(ds); /* * Get error count while pager locked. */ errors = ds->errors; ddprintf ("seqnos_memory_object_data_request <%p>: pager_port_unlock: <%p>[s:%d,r:%d,w:%d,l:%d]\n", &ds, ds, ds->seqno, ds->readers, ds->writers, ds->lock.held); pager_port_unlock(ds); if (errors) { dprintf("%s %s\n", my_name, "dropping data_request because of previous paging errors"); (void) memory_object_data_error(reply_to, offset, vm_page_size, KERN_FAILURE); goto done; } if (offset >= ds->dpager.limit) rc = PAGER_ERROR; else rc = default_read(&ds->dpager, dpt->dpt_buffer, vm_page_size, offset, &addr, protection_required & VM_PROT_WRITE, ds->external); switch (rc) { case PAGER_SUCCESS: if (addr != dpt->dpt_buffer) { /* * Deallocates data buffer */ (void) memory_object_data_supply( reply_to, offset, addr, vm_page_size, TRUE, VM_PROT_NONE, FALSE, MACH_PORT_NULL); } else { (void) memory_object_data_supply( reply_to, offset, addr, vm_page_size, FALSE, VM_PROT_NONE, FALSE, MACH_PORT_NULL); } break; case PAGER_ABSENT: (void) memory_object_data_unavailable( reply_to, offset, vm_page_size); break; case PAGER_ERROR: (void) memory_object_data_error( reply_to, offset, vm_page_size, KERN_FAILURE); break; } default_pager_pagein_count++; done: pager_port_finish_read(ds); return(KERN_SUCCESS); } /* * memory_object_data_initialize: check whether we already have each page, and * write it if we do not. The implementation is far from optimized, and * also assumes that the default_pager is single-threaded. */ kern_return_t seqnos_memory_object_data_initialize(ds, seqno, pager_request, offset, addr, data_cnt) default_pager_t ds; mach_port_seqno_t seqno; mach_port_t pager_request; register vm_offset_t offset; register pointer_t addr; vm_size_t data_cnt; { vm_offset_t amount_sent; static char here[] = "%sdata_initialize"; #ifdef lint pager_request++; #endif /* lint */ if (ds == DEFAULT_PAGER_NULL) panic(here,my_name); ddprintf ("seqnos_memory_object_data_initialize <%p>: pager_port_lock: <%p>[s:%d,r:%d,w:%d,l:%d], %d\n", &ds, ds, ds->seqno, ds->readers, ds->writers, ds->lock.held, seqno); pager_port_lock(ds, seqno); pager_port_check_request(ds, pager_request); pager_port_start_write(ds); ddprintf ("seqnos_memory_object_data_initialize <%p>: pager_port_unlock: <%p>[s:%d,r:%d,w:%d,l:%d]\n", &ds, ds, ds->seqno, ds->readers, ds->writers, ds->lock.held); pager_port_unlock(ds); for (amount_sent = 0; amount_sent < data_cnt; amount_sent += vm_page_size) { if (!default_has_page(&ds->dpager, offset + amount_sent)) { if (default_write(&ds->dpager, addr + amount_sent, vm_page_size, offset + amount_sent) != PAGER_SUCCESS) { dprintf("%s%s write error\n", my_name, here); dstruct_lock(ds); ds->errors++; dstruct_unlock(ds); } } } pager_port_finish_write(ds); if (vm_deallocate(default_pager_self, addr, data_cnt) != KERN_SUCCESS) panic(here,my_name); return(KERN_SUCCESS); } /* * memory_object_data_write: split up the stuff coming in from * a memory_object_data_write call * into individual pages and pass them off to default_write. */ kern_return_t seqnos_memory_object_data_write(ds, seqno, pager_request, offset, addr, data_cnt) default_pager_t ds; mach_port_seqno_t seqno; mach_port_t pager_request; register vm_offset_t offset; register pointer_t addr; vm_size_t data_cnt; { register vm_size_t amount_sent; static char here[] = "%sdata_write"; int err; #ifdef lint pager_request++; #endif /* lint */ if ((data_cnt % vm_page_size) != 0) panic(here,my_name); if (ds == DEFAULT_PAGER_NULL) panic(here,my_name); pager_port_lock(ds, seqno); pager_port_start_write(ds); vm_size_t limit = ds->dpager.byte_limit; pager_port_unlock(ds); if ((limit != round_page(limit)) && (trunc_page(limit) == offset)) { assert(trunc_page(limit) == offset); assert(data_cnt == vm_page_size); vm_offset_t tail = addr + limit - trunc_page(limit); vm_size_t tail_size = round_page(limit) - limit; memset((void *) tail, 0, tail_size); memory_object_data_supply(pager_request, trunc_page(limit), addr, vm_page_size, TRUE, VM_PROT_NONE, TRUE, MACH_PORT_NULL); dstruct_lock(ds); ds->dpager.byte_limit = round_page(limit); dstruct_unlock(ds); pager_port_finish_write(ds); return(KERN_SUCCESS); } for (amount_sent = 0; amount_sent < data_cnt; amount_sent += vm_page_size) { int result; result = default_write(&ds->dpager, addr + amount_sent, vm_page_size, offset + amount_sent); if (result != KERN_SUCCESS) { dstruct_lock(ds); ds->errors++; dstruct_unlock(ds); } default_pager_pageout_count++; } pager_port_finish_write(ds); err = vm_deallocate(default_pager_self, addr, data_cnt); if (err != KERN_SUCCESS) { panic(here,my_name); } return(KERN_SUCCESS); } /*ARGSUSED*/ kern_return_t seqnos_memory_object_copy(old_memory_object, seqno, old_memory_control, offset, length, new_memory_object) default_pager_t old_memory_object; mach_port_seqno_t seqno; memory_object_control_t old_memory_control; vm_offset_t offset; vm_size_t length; memory_object_t new_memory_object; { panic("%scopy", my_name); return KERN_FAILURE; } /* We get this when our memory_object_lock_request has completed after we truncated an object. */ kern_return_t seqnos_memory_object_lock_completed (default_pager_t ds, mach_port_seqno_t seqno, mach_port_t pager_request, vm_offset_t offset, vm_size_t length) { panic("%slock_completed",my_name); return KERN_FAILURE; } kern_return_t seqnos_memory_object_data_unlock(pager, seqno, pager_request, offset, length, protection_required) default_pager_t pager; mach_port_seqno_t seqno; mach_port_t pager_request; vm_offset_t offset; vm_size_t length; vm_prot_t protection_required; { panic("%sdata_unlock",my_name); return(KERN_FAILURE); } kern_return_t seqnos_memory_object_supply_completed(ds, seqno, pager_request, offset, length, result, error_offset) default_pager_t ds; mach_port_seqno_t seqno; mach_port_t pager_request; vm_offset_t offset; vm_size_t length; kern_return_t result; vm_offset_t error_offset; { panic("%ssupply_completed",my_name); return(KERN_FAILURE); } /* * memory_object_data_return: split up the stuff coming in from * a memory_object_data_write call * into individual pages and pass them off to default_write. */ kern_return_t seqnos_memory_object_data_return(ds, seqno, pager_request, offset, addr, data_cnt, dirty, kernel_copy) default_pager_t ds; mach_port_seqno_t seqno; mach_port_t pager_request; vm_offset_t offset; pointer_t addr; vm_size_t data_cnt; boolean_t dirty; boolean_t kernel_copy; { return seqnos_memory_object_data_write (ds, seqno, pager_request, offset, addr, data_cnt); } kern_return_t seqnos_memory_object_change_completed(ds, seqno, may_cache, copy_strategy) default_pager_t ds; mach_port_seqno_t seqno; boolean_t may_cache; memory_object_copy_strategy_t copy_strategy; { panic("%schange_completed",my_name); return(KERN_FAILURE); } boolean_t default_pager_notify_server(in, out) mach_msg_header_t *in, *out; { mach_no_senders_notification_t *n = (mach_no_senders_notification_t *) in; /* * The only send-once rights we create are for * receiving no-more-senders notifications. * Hence, if we receive a message directed to * a send-once right, we can assume it is * a genuine no-senders notification from the kernel. */ if ((n->not_header.msgh_bits != MACH_MSGH_BITS(0, MACH_MSG_TYPE_PORT_SEND_ONCE)) || (n->not_header.msgh_id != MACH_NOTIFY_NO_SENDERS)) return FALSE; assert(n->not_header.msgh_size == sizeof *n); assert(n->not_header.msgh_remote_port == MACH_PORT_NULL); assert(n->not_type.msgt_name == MACH_MSG_TYPE_INTEGER_32); assert(n->not_type.msgt_size == 32); assert(n->not_type.msgt_number == 1); assert(n->not_type.msgt_inline); assert(! n->not_type.msgt_longform); default_pager_no_senders(n->not_header.msgh_local_port, n->not_header.msgh_seqno, n->not_count); out->msgh_remote_port = MACH_PORT_NULL; return TRUE; } extern boolean_t seqnos_memory_object_default_server(); extern boolean_t default_pager_server(); extern boolean_t exc_server(); extern boolean_t bootstrap_server(); extern void bootstrap_compat(); mach_msg_size_t default_pager_msg_size_object = 128; /* Fill in default response. */ static void mig_reply_setup ( const mach_msg_header_t *in, mach_msg_header_t *out) { static const mach_msg_type_t RetCodeType = { /* msgt_name = */ MACH_MSG_TYPE_INTEGER_32, /* msgt_size = */ 32, /* msgt_number = */ 1, /* msgt_inline = */ TRUE, /* msgt_longform = */ FALSE, /* msgt_deallocate = */ FALSE, /* msgt_unused = */ 0 }; #define InP (in) #define OutP ((mig_reply_header_t *) out) OutP->Head.msgh_bits = MACH_MSGH_BITS(MACH_MSGH_BITS_REMOTE(InP->msgh_bits), 0); OutP->Head.msgh_size = sizeof *OutP; OutP->Head.msgh_remote_port = InP->msgh_remote_port; OutP->Head.msgh_local_port = MACH_PORT_NULL; OutP->Head.msgh_seqno = 0; OutP->Head.msgh_id = InP->msgh_id + 100; OutP->RetCodeType = RetCodeType; OutP->RetCode = MIG_BAD_ID; #undef InP #undef OutP } boolean_t default_pager_demux_object(in, out) mach_msg_header_t *in; mach_msg_header_t *out; { /* * We receive memory_object_data_initialize messages in * the memory_object_default interface. */ int rval = FALSE; ddprintf ("DPAGER DEMUX OBJECT <%p>: %d\n", in, in->msgh_id); mig_reply_setup (in, out); mig_routine_t routine; if ((routine = seqnos_memory_object_server_routine (in)) || (routine = seqnos_memory_object_default_server_routine (in)) || (routine = NULL, default_pager_notify_server (in, out)) || (routine = default_pager_server_routine (in))) { if (routine) (*routine) (in, out); rval = TRUE; } ddprintf ("DPAGER DEMUX OBJECT DONE <%p>: %d\n", in, in->msgh_id); return rval; } mach_msg_size_t default_pager_msg_size_default = 8 * 1024; boolean_t default_pager_demux_default(in, out) mach_msg_header_t *in; mach_msg_header_t *out; { if (in->msgh_local_port == default_pager_default_port) { /* * We receive memory_object_create messages in * the memory_object_default interface. */ int rval; ddprintf ("DPAGER DEMUX DEFAULT <%p>: %d\n", in, in->msgh_id); rval = (seqnos_memory_object_default_server(in, out) || default_pager_server(in, out)); ddprintf ("DPAGER DEMUX DEFAULT DONE <%p>: %d\n", in, in->msgh_id); return rval; } else if (in->msgh_local_port == default_pager_exception_port) { /* * We receive exception messages for * ourself and the startup task. */ return exc_server(in, out); } else { panic(my_name); return FALSE; } } /* * We use multiple threads, for two reasons. * * First, memory objects created by default_pager_object_create * are "external", instead of "internal". This means the kernel * sends data (memory_object_data_write) to the object pageable. * To prevent deadlocks, the external and internal objects must * be managed by different threads. * * Second, the default pager uses synchronous IO operations. * Spreading requests across multiple threads should * recover some of the performance loss from synchronous IO. * * We have 3+ threads. * One receives memory_object_create and * default_pager_object_create requests. * One or more manage internal objects. * One or more manage external objects. */ void default_pager_thread_privileges() { /* * Set thread privileges. */ wire_thread(); /* grab a kernel stack and memory allocation privileges */ } void * default_pager_default_thread(void *arg) { kern_return_t kr; default_pager_thread_privileges (); for (;;) { kr = mach_msg_server(default_pager_demux_default, default_pager_msg_size_default, default_pager_default_set); panic(my_name, kr); } } void * default_pager_thread(void *arg) { mach_port_t pset; kern_return_t kr; dpt = (default_pager_thread_t *) arg; /* * Threads handling external objects cannot have * privileges. Otherwise a burst of data-requests for an * external object could empty the free-page queue, * because the fault code only reserves real pages for * requests sent to internal objects. */ if (dpt->dpt_internal) { default_pager_thread_privileges(); pset = default_pager_internal_set; } else { pset = default_pager_external_set; } for (;;) { kr = mach_msg_server(default_pager_demux_object, default_pager_msg_size_object, pset); panic(my_name, kr); } } void start_default_pager_thread(internal) boolean_t internal; { default_pager_thread_t *ndpt; kern_return_t kr; error_t err; ndpt = (default_pager_thread_t *) kalloc(sizeof *ndpt); if (ndpt == 0) panic(my_name); ndpt->dpt_internal = internal; kr = vm_allocate(default_pager_self, &ndpt->dpt_buffer, vm_page_size, TRUE); if (kr != KERN_SUCCESS) panic(my_name); wire_memory(ndpt->dpt_buffer, vm_page_size, VM_PROT_READ|VM_PROT_WRITE); err = pthread_create(&ndpt->dpt_thread, NULL, default_pager_thread, ndpt); if (!err) pthread_detach (ndpt->dpt_thread); else { errno = err; perror ("pthread_create"); } } void default_pager_initialize(host_port) mach_port_t host_port; { memory_object_t DMM; kern_return_t kr; /* * This task will become the default pager. */ default_pager_self = mach_task_self(); /* * Initialize the "default pager" port. */ kr = mach_port_allocate(default_pager_self, MACH_PORT_RIGHT_RECEIVE, &default_pager_default_port); if (kr != KERN_SUCCESS) panic(my_name); DMM = default_pager_default_port; kr = vm_set_default_memory_manager(host_port, &DMM); if ((kr != KERN_SUCCESS) || MACH_PORT_VALID(DMM)) panic(my_name); /* * Initialize the exception port. */ kr = mach_port_allocate(default_pager_self, MACH_PORT_RIGHT_RECEIVE, &default_pager_exception_port); if (kr != KERN_SUCCESS) panic(my_name); /* * Arrange for wiring privileges. */ wire_setup(host_port); /* * Find out how many CPUs we have, to determine the number * of threads to create. */ if (default_pager_internal_count == 0) { host_basic_info_data_t h_info; natural_t h_info_count; h_info_count = HOST_BASIC_INFO_COUNT; (void) host_info(host_port, HOST_BASIC_INFO, (host_info_t)&h_info, &h_info_count); /* * Random computation to get more parallelism on * multiprocessors. */ default_pager_internal_count = (h_info.avail_cpus > 32 ? 32 : h_info.avail_cpus) / 4 + 3; } } /* * Initialize and Run the default pager */ void default_pager() { kern_return_t kr; int i; default_pager_thread_privileges(); /* * Wire down code, data, stack */ wire_all_memory(); /* * Initialize the list of all pagers. */ pager_port_list_init(); kr = mach_port_allocate(default_pager_self, MACH_PORT_RIGHT_PORT_SET, &default_pager_internal_set); if (kr != KERN_SUCCESS) panic(my_name); kr = mach_port_allocate(default_pager_self, MACH_PORT_RIGHT_PORT_SET, &default_pager_external_set); if (kr != KERN_SUCCESS) panic(my_name); kr = mach_port_allocate(default_pager_self, MACH_PORT_RIGHT_PORT_SET, &default_pager_default_set); if (kr != KERN_SUCCESS) panic(my_name); kr = mach_port_move_member(default_pager_self, default_pager_default_port, default_pager_default_set); if (kr != KERN_SUCCESS) panic(my_name); kr = mach_port_move_member(default_pager_self, default_pager_exception_port, default_pager_default_set); if (kr != KERN_SUCCESS) panic(my_name); /* * Now we create the threads that will actually * manage objects. */ for (i = 0; i < default_pager_internal_count; i++) start_default_pager_thread(TRUE); for (i = 0; i < default_pager_external_count; i++) start_default_pager_thread(FALSE); default_pager_default_thread(0); /* Become the default_pager server */ #if 0 cthread_fork (default_pager_default_thread, 0); /* cthread_exit (cthread_self ()); */ thread_suspend (mach_thread_self ()); #endif } /* * Create an external object. */ kern_return_t S_default_pager_object_create (mach_port_t pager, mach_port_t *mem_obj, mach_msg_type_name_t *mem_obj_type, vm_size_t size) { default_pager_t ds; mach_port_t port; kern_return_t result; if (pager != default_pager_default_port) return KERN_INVALID_ARGUMENT; ds = pager_port_alloc(size); result = mach_port_allocate (default_pager_self, MACH_PORT_RIGHT_RECEIVE, &port); if (result != KERN_SUCCESS) { kfree ((char *) ds, sizeof *ds); return result; } /* * Set up associations between these ports * and this default_pager structure */ ds->pager = port; ds->dpager.limit = size; pager_port_list_insert(port, ds); default_pager_add(ds, FALSE); *mem_obj = port; *mem_obj_type = MACH_MSG_TYPE_MAKE_SEND; return (KERN_SUCCESS); } kern_return_t S_default_pager_info (mach_port_t pager, default_pager_info_t *infop) { vm_size_t total, free; if (pager != default_pager_default_port) return KERN_INVALID_ARGUMENT; pthread_mutex_lock(&all_partitions.lock); paging_space_info(&total, &free); pthread_mutex_unlock(&all_partitions.lock); infop->dpi_total_space = ptoa(total); infop->dpi_free_space = ptoa(free); infop->dpi_page_size = vm_page_size; return KERN_SUCCESS; } kern_return_t S_default_pager_objects (mach_port_t pager, default_pager_object_array_t *objectsp, natural_t *ocountp, mach_port_array_t *portsp, natural_t *pcountp) { vm_offset_t oaddr = 0; /* memory for objects */ vm_size_t osize = 0; /* current size */ default_pager_object_t *objects; natural_t opotential; vm_offset_t paddr = 0; /* memory for ports */ vm_size_t psize = 0; /* current size */ mach_port_t *ports; natural_t ppotential; unsigned int actual; unsigned int num_pagers; kern_return_t kr; default_pager_t entry; if (pager != default_pager_default_port) return KERN_INVALID_ARGUMENT; /* start with the inline memory */ num_pagers = 0; objects = *objectsp; opotential = *ocountp; ports = *portsp; ppotential = *pcountp; pthread_mutex_lock(&all_pagers.lock); /* * We will send no more than this many */ actual = all_pagers.htable.nr_items; pthread_mutex_unlock(&all_pagers.lock); if (opotential < actual) { vm_offset_t newaddr; vm_size_t newsize; newsize = 2 * round_page(actual * sizeof *objects); kr = vm_allocate(default_pager_self, &newaddr, newsize, TRUE); if (kr != KERN_SUCCESS) goto nomemory; oaddr = newaddr; osize = newsize; opotential = osize/sizeof *objects; objects = (default_pager_object_t *) oaddr; } if (ppotential < actual) { vm_offset_t newaddr; vm_size_t newsize; newsize = 2 * round_page(actual * sizeof *ports); kr = vm_allocate(default_pager_self, &newaddr, newsize, TRUE); if (kr != KERN_SUCCESS) goto nomemory; paddr = newaddr; psize = newsize; ppotential = psize/sizeof *ports; ports = (mach_port_t *) paddr; } /* * Now scan the list. */ pthread_mutex_lock(&all_pagers.lock); num_pagers = 0; HURD_IHASH_ITERATE (&all_pagers.htable, val) { entry = (default_pager_t) val; mach_port_t port; vm_size_t size; if ((num_pagers >= opotential) || (num_pagers >= ppotential)) { /* * This should be rare. In any case, * we will only miss recent objects, * because they are added at the end. */ break; } /* * Avoid interfering with normal operations */ if (pthread_mutex_trylock(&entry->dpager.lock)) goto not_this_one; size = pager_allocated(&entry->dpager); pthread_mutex_unlock(&entry->dpager.lock); dstruct_lock(entry); port = entry->pager_name; if (port == MACH_PORT_NULL) { /* * The object is waiting for no-senders * or memory_object_init. */ dstruct_unlock(entry); goto not_this_one; } /* * We need a reference for the reply message. * While we are unlocked, the bucket queue * can change and the object might be terminated. * memory_object_terminate will wait for us, * preventing deallocation of the entry. */ if (--entry->name_refs == 0) { dstruct_unlock(entry); /* keep the list locked, wont take long */ kr = mach_port_mod_refs(default_pager_self, port, MACH_PORT_RIGHT_SEND, default_pager_max_urefs); if (kr != KERN_SUCCESS) panic("%sdefault_pager_objects",my_name); dstruct_lock(entry); entry->name_refs += default_pager_max_urefs; pager_port_finish_refs(entry); } dstruct_unlock(entry); /* the arrays are wired, so no deadlock worries */ objects[num_pagers].dpo_object = (vm_offset_t) entry; objects[num_pagers].dpo_size = size; ports [num_pagers++] = port; continue; not_this_one: /* * Do not return garbage */ objects[num_pagers].dpo_object = (vm_offset_t) 0; objects[num_pagers].dpo_size = 0; ports [num_pagers++] = MACH_PORT_NULL; } pthread_mutex_unlock(&all_pagers.lock); /* * Deallocate and clear unused memory. * (Returned memory will automagically become pageable.) */ if (objects == *objectsp) { /* * Our returned information fit inline. * Nothing to deallocate. */ *ocountp = num_pagers; } else if (actual == 0) { (void) vm_deallocate(default_pager_self, oaddr, osize); /* return zero items inline */ *ocountp = 0; } else { vm_offset_t used; used = round_page(actual * sizeof *objects); if (used != osize) (void) vm_deallocate(default_pager_self, oaddr + used, osize - used); *objectsp = objects; *ocountp = num_pagers; } if (ports == *portsp) { /* * Our returned information fit inline. * Nothing to deallocate. */ *pcountp = num_pagers; } else if (actual == 0) { (void) vm_deallocate(default_pager_self, paddr, psize); /* return zero items inline */ *pcountp = 0; } else { vm_offset_t used; used = round_page(actual * sizeof *ports); if (used != psize) (void) vm_deallocate(default_pager_self, paddr + used, psize - used); *portsp = ports; *pcountp = num_pagers; } return KERN_SUCCESS; nomemory: { int i; for (i = 0; i < num_pagers; i++) (void) mach_port_deallocate(default_pager_self, ports[i]); } if (objects != *objectsp) (void) vm_deallocate(default_pager_self, oaddr, osize); if (ports != *portsp) (void) vm_deallocate(default_pager_self, paddr, psize); return KERN_RESOURCE_SHORTAGE; } kern_return_t S_default_pager_object_pages (mach_port_t pager, mach_port_t object, default_pager_page_array_t *pagesp, natural_t *countp) { vm_offset_t addr = 0; /* memory for page offsets */ vm_size_t size = 0; /* current memory size */ default_pager_page_t *pages; natural_t potential, actual; kern_return_t kr; if (pager != default_pager_default_port) return KERN_INVALID_ARGUMENT; /* we start with the inline space */ pages = *pagesp; potential = *countp; for (;;) { default_pager_t entry; pthread_mutex_lock(&all_pagers.lock); HURD_IHASH_ITERATE (&all_pagers.htable, val) { entry = (default_pager_t) val; dstruct_lock(entry); if (entry->pager_name == object) { pthread_mutex_unlock(&all_pagers.lock); goto found_object; } dstruct_unlock(entry); } pthread_mutex_unlock(&all_pagers.lock); /* did not find the object */ if (pages != *pagesp) (void) vm_deallocate(default_pager_self, addr, size); return KERN_INVALID_ARGUMENT; found_object: if (pthread_mutex_trylock(&entry->dpager.lock)) { /* oh well bad luck */ dstruct_unlock(entry); /* yield the processor */ (void) thread_switch(MACH_PORT_NULL, SWITCH_OPTION_NONE, 0); continue; } actual = pager_pages(&entry->dpager, pages, potential); pthread_mutex_unlock(&entry->dpager.lock); dstruct_unlock(entry); if (actual <= potential) break; /* allocate more memory */ if (pages != *pagesp) (void) vm_deallocate(default_pager_self, addr, size); size = round_page(actual * sizeof *pages); kr = vm_allocate(default_pager_self, &addr, size, TRUE); if (kr != KERN_SUCCESS) return kr; pages = (default_pager_page_t *) addr; potential = size/sizeof *pages; } /* * Deallocate and clear unused memory. * (Returned memory will automagically become pageable.) */ if (pages == *pagesp) { /* * Our returned information fit inline. * Nothing to deallocate. */ *countp = actual; } else if (actual == 0) { (void) vm_deallocate(default_pager_self, addr, size); /* return zero items inline */ *countp = 0; } else { vm_offset_t used; used = round_page(actual * sizeof *pages); if (used != size) (void) vm_deallocate(default_pager_self, addr + used, size - used); *pagesp = pages; *countp = actual; } return KERN_SUCCESS; } kern_return_t S_default_pager_object_set_size (default_pager_t ds, mach_port_seqno_t seqno, vm_size_t limit) { kern_return_t kr = KERN_SUCCESS; if (ds == DEFAULT_PAGER_NULL) return KERN_INVALID_ARGUMENT; pager_port_lock(ds, seqno); pager_port_wait_for_readers(ds); pager_port_wait_for_writers(ds); vm_size_t rounded_limit = round_page (limit); vm_size_t trunc_limit = trunc_page (limit); if (ds->dpager.limit < rounded_limit) { /* The limit has not been exceeded heretofore. Just change it. */ ds->dpager.limit = rounded_limit; /* Byte limit is used for truncation of file, which aren't rounded to page boundary. But by enlarging of file we are free to increase this value*/ ds->dpager.byte_limit = rounded_limit; kr = memory_object_lock_request(ds->pager_request, 0, rounded_limit, MEMORY_OBJECT_RETURN_NONE, FALSE, VM_PROT_NONE, MACH_PORT_NULL); if (kr != KERN_SUCCESS) panic ("memory_object_lock_request: %d", kr); } else { if (ds->dpager.limit != rounded_limit) { kr = memory_object_lock_request(ds->pager_request, rounded_limit, ds->dpager.limit - rounded_limit, MEMORY_OBJECT_RETURN_NONE, TRUE, VM_PROT_ALL, MACH_PORT_NULL); if (kr != KERN_SUCCESS) panic ("memory_object_lock_request: %d", kr); ds->dpager.limit = rounded_limit; } /* Deallocate the old backing store pages and shrink the page map. */ if (ds->dpager.size > ds->dpager.limit / vm_page_size) pager_truncate (&ds->dpager, ds->dpager.limit / vm_page_size); /* If memory object size isn't page aligned, fill the tail of last page with zeroes */ if ((limit != rounded_limit) && (ds->dpager.limit > limit)) { /* Clean part of last page which isn't part of file. For file sizes that aren't multiple of vm_page_size */ ds->dpager.byte_limit = limit; kr = memory_object_lock_request(ds->pager_request, trunc_limit, vm_page_size, MEMORY_OBJECT_RETURN_ALL, TRUE, VM_PROT_NONE, MACH_PORT_NULL); } } pager_port_unlock(ds); return kr; } /* * Add/remove extra paging space */ extern mach_port_t bootstrap_master_device_port; extern mach_port_t bootstrap_master_host_port; kern_return_t S_default_pager_paging_file (pager, mdport, file_name, add) mach_port_t pager; mach_port_t mdport; default_pager_filename_t file_name; boolean_t add; { kern_return_t kr; if (pager != default_pager_default_port) return KERN_INVALID_ARGUMENT; #if 0 dprintf("bmd %x md %x\n", bootstrap_master_device_port, mdport); #endif if (add) { kr = add_paging_file(bootstrap_master_device_port, file_name, 0); } else { kr = remove_paging_file(file_name); } /* XXXX more code needed */ if (mdport != bootstrap_master_device_port) mach_port_deallocate( mach_task_self(), mdport); return kr; } kern_return_t default_pager_register_fileserver(pager, fileserver) mach_port_t pager; mach_port_t fileserver; { if (pager != default_pager_default_port) return KERN_INVALID_ARGUMENT; #if notyet mach_port_deallocate(mach_task_self(), fileserver); if (0) dp_helper_paging_space(0,0,0);/*just linkit*/ #endif return KERN_SUCCESS; } /* * When things do not quite workout... */ void no_paging_space(out_of_memory) boolean_t out_of_memory; { static char here[] = "%s *** NOT ENOUGH PAGING SPACE ***"; if (out_of_memory) dprintf("*** OUT OF MEMORY *** "); panic(here, my_name); } void overcommitted(got_more_space, space) boolean_t got_more_space; vm_size_t space; /* in pages */ { vm_size_t pages_free, pages_total; static boolean_t user_warned = FALSE; static vm_size_t pages_shortage = 0; paging_space_info(&pages_total, &pages_free); /* * If user added more space, see if it is enough */ if (got_more_space) { pages_free -= pages_shortage; if (pages_free > 0) { pages_shortage = 0; if (user_warned) dprintf("%s paging space ok now.\n", my_name); } else pages_shortage = pages_free; user_warned = FALSE; return; } /* * We ran out of gas, let user know. */ pages_free -= space; pages_shortage = (pages_free > 0) ? 0 : -pages_free; if (!user_warned && pages_shortage) { user_warned = TRUE; dprintf("%s paging space over-committed.\n", my_name); } #if debug user_warned = FALSE; dprintf("%s paging space over-committed [+%d (%d) pages].\n", my_name, space, pages_shortage); #endif } void paging_space_info(totp, freep) vm_size_t *totp, *freep; { vm_size_t total, free; partition_t part; int i; total = free = 0; for (i = 0; i < all_partitions.n_partitions; i++) { if ((part = partition_of(i)) == 0) continue; /* no need to lock: by the time this data gets back to any remote requestor it will be obsolete anyways */ total += part->total_size; free += part->free; #if debug dprintf("Partition %d: x%x total, x%x free\n", i, part->total_size, part->free); #endif } *totp = total; *freep = free; } /* * Catch exceptions. */ kern_return_t catch_exception_raise(exception_port, thread, task, exception, code, subcode) mach_port_t exception_port; mach_port_t thread, task; int exception, code, subcode; { ddprintf ("(default_pager)catch_exception_raise(%d,%d,%d)\n", exception, code, subcode); panic(my_name); /* mach_msg_server will deallocate thread/task for us */ return KERN_FAILURE; }