Index: sys/kern/subr_blist.c =================================================================== --- sys/kern/subr_blist.c +++ sys/kern/subr_blist.c @@ -130,9 +130,10 @@ /* * static support functions */ -static daddr_t blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int count); -static daddr_t blst_meta_alloc(blmeta_t *scan, daddr_t cursor, daddr_t count, - u_daddr_t radix); +static daddr_t blst_leaf_alloc(blmeta_t *scan, daddr_t blk, + int *count, int maxcount); +static daddr_t blst_meta_alloc(blmeta_t *scan, daddr_t cursor, int *count, + int maxcount, u_daddr_t radix); static void blst_leaf_free(blmeta_t *scan, daddr_t relblk, int count); static void blst_meta_free(blmeta_t *scan, daddr_t freeBlk, daddr_t count, u_daddr_t radix); @@ -193,7 +194,6 @@ /* * Use binary search, or a faster method, to find the 1 bit in a u_daddr_t. - * Assumes that the argument has only one bit set. */ static inline int bitpos(u_daddr_t mask) @@ -210,7 +210,7 @@ hi = BLIST_BMAP_RADIX; while (lo + 1 < hi) { mid = (lo + hi) >> 1; - if ((mask >> mid) != 0) + if ((mask & (((u_daddr_t)1 << mid) - 1)) == 0) lo = mid; else hi = mid; @@ -284,11 +284,13 @@ * not be allocated. */ daddr_t -blist_alloc(blist_t bl, daddr_t count) +blist_alloc(blist_t bl, int *count, int maxcount) { daddr_t blk; - if (count > BLIST_MAX_ALLOC) + if (*count > maxcount) + panic("invalid parameters"); + if (maxcount > BLIST_MAX_ALLOC) panic("allocation too large"); /* @@ -299,10 +301,10 @@ */ for (;;) { blk = blst_meta_alloc(bl->bl_root, bl->bl_cursor, count, - bl->bl_radix); + maxcount, bl->bl_radix); if (blk != SWAPBLK_NONE) { - bl->bl_avail -= count; - bl->bl_cursor = blk + count; + bl->bl_avail -= *count; + bl->bl_cursor = blk + *count; if (bl->bl_cursor == bl->bl_blocks) bl->bl_cursor = 0; return (blk); @@ -649,13 +651,13 @@ * crosses a leaf boundary. */ static daddr_t -blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int count) +blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int *count, int maxcount) { u_daddr_t cursor_mask, mask; int count1, hi, lo, num_shifts, range1, range_ext; range1 = 0; - count1 = count - 1; + count1 = *count - 1; num_shifts = fls(count1); mask = scan->bm_bitmap; while ((-mask & ~mask) != 0 && num_shifts > 0) { @@ -716,7 +718,7 @@ mask &= -mask; lo = bitpos(mask); - hi = lo + count; + hi = lo + *count; if (hi > BLIST_BMAP_RADIX) { /* * An allocation within this leaf is impossible, so a successful @@ -759,7 +761,8 @@ * and we have a few optimizations strewn in as well. */ static daddr_t -blst_meta_alloc(blmeta_t *scan, daddr_t cursor, daddr_t count, u_daddr_t radix) +blst_meta_alloc(blmeta_t *scan, daddr_t cursor, int *count, + int maxcount, u_daddr_t radix) { daddr_t blk, i, r, skip; u_daddr_t bit, mask; @@ -767,7 +770,7 @@ int digit; if (radix == BLIST_BMAP_RADIX) - return (blst_leaf_alloc(scan, cursor, count)); + return (blst_leaf_alloc(scan, cursor, count, maxcount)); blk = cursor & -radix; scan_from_start = (cursor == blk); radix /= BLIST_META_RADIX; @@ -797,12 +800,12 @@ bit = mask & -mask; digit = bitpos(bit); i = 1 + digit * skip; - if (count <= scan[i].bm_bighint) { + if (*count <= scan[i].bm_bighint) { /* * The allocation might fit beginning in the i'th subtree. */ r = blst_meta_alloc(&scan[i], cursor + digit * radix, - count, radix); + count, maxcount, radix); if (r != SWAPBLK_NONE) { if (scan[i].bm_bitmap == 0) scan->bm_bitmap ^= bit; @@ -818,7 +821,7 @@ */ if (scan_from_start && !(digit == BLIST_META_RADIX - 1 && scan[i].bm_bighint == BLIST_MAX_ALLOC)) - scan->bm_bighint = count - 1; + scan->bm_bighint = *count - 1; return (SWAPBLK_NONE); } @@ -1077,7 +1080,7 @@ for (;;) { char buf[1024]; long long da = 0; - long long count = 0; + int count = 0, maxcount = 0; printf("%lld/%lld/%lld> ", (long long)blist_avail(bl), (long long)size, (long long)bl->bl_radix); @@ -1086,7 +1089,7 @@ break; switch(buf[0]) { case 'r': - if (sscanf(buf + 1, "%lld", &count) == 1) { + if (sscanf(buf + 1, "%d", &count) == 1) { blist_resize(&bl, count, 1, M_WAITOK); } else { printf("?\n"); @@ -1102,22 +1105,23 @@ sbuf_delete(s); break; case 'a': - if (sscanf(buf + 1, "%lld", &count) == 1) { - daddr_t blk = blist_alloc(bl, count); - printf(" R=%08llx\n", (long long)blk); + if (sscanf(buf + 1, "%d%d", &count, &maxcount) == 2) { + daddr_t blk = blist_alloc(bl, &count, maxcount); + printf(" R=%08llx, c=%08d\n", + (long long)blk, count); } else { printf("?\n"); } break; case 'f': - if (sscanf(buf + 1, "%llx %lld", &da, &count) == 2) { + if (sscanf(buf + 1, "%llx %d", &da, &count) == 2) { blist_free(bl, da, count); } else { printf("?\n"); } break; case 'l': - if (sscanf(buf + 1, "%llx %lld", &da, &count) == 2) { + if (sscanf(buf + 1, "%llx %d", &da, &count) == 2) { printf(" n=%jd\n", (intmax_t)blist_fill(bl, da, count)); } else { @@ -1129,17 +1133,22 @@ puts( "p -print\n" "s -stats\n" - "a %d -allocate\n" + "a %d %d -allocate\n" "f %x %d -free\n" "l %x %d -fill\n" "r %d -resize\n" - "h/? -help" + "h/? -help\n" + "q -quit" ); break; + case 'q': + break; default: printf("?\n"); break; } + if (buf[0] == 'q') + break; } return(0); } Index: sys/sys/blist.h =================================================================== --- sys/sys/blist.h +++ sys/sys/blist.h @@ -33,7 +33,7 @@ * Usage: * blist = blist_create(blocks, flags) * (void) blist_destroy(blist) - * blkno = blist_alloc(blist, count) + * blkno = blist_alloc(blist, &count, maxcount) * (void) blist_free(blist, blkno, count) * nblks = blist_fill(blist, blkno, count) * (void) blist_resize(&blist, count, freeextra, flags) @@ -69,11 +69,11 @@ #define SWAPBLK_NONE ((daddr_t)((u_daddr_t)SWAPBLK_MASK + 1))/* flag */ /* - * Both blmeta and bmu_bitmap MUST be a power of 2 in size. + * Both blmeta and bm_bitmap MUST be a power of 2 in size. */ typedef struct blmeta { - u_daddr_t bm_bitmap; /* bitmap if we are a leaf */ + u_daddr_t bm_bitmap; /* bitmap marking unfilled block sets*/ daddr_t bm_bighint; /* biggest contiguous block hint*/ } blmeta_t; @@ -92,7 +92,7 @@ struct sbuf; -daddr_t blist_alloc(blist_t blist, daddr_t count); +daddr_t blist_alloc(blist_t blist, int *count, int maxcount); daddr_t blist_avail(blist_t blist); blist_t blist_create(daddr_t blocks, int flags); void blist_destroy(blist_t blist); Index: sys/vm/swap_pager.c =================================================================== --- sys/vm/swap_pager.c +++ sys/vm/swap_pager.c @@ -407,7 +407,7 @@ * Swap bitmap functions */ static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages); -static daddr_t swp_pager_getswapspace(int npages); +static daddr_t swp_pager_getswapspace(int *npages, int limit); /* * Metadata functions @@ -708,9 +708,10 @@ /* * 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. + * Allocate swap for up to the requested number of pages, and at + * least a minimum 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. @@ -720,38 +721,49 @@ * We allocate in round-robin fashion from the configured devices. */ static daddr_t -swp_pager_getswapspace(int npages) +swp_pager_getswapspace(int *io_npages, int limit) { daddr_t blk; struct swdevt *sp; - int i; + int mpages, npages; blk = SWAPBLK_NONE; + npages = mpages = *io_npages; mtx_lock(&sw_dev_mtx); + if (TAILQ_EMPTY(&swtailq)) { + mtx_unlock(&sw_dev_mtx); + return (blk); + } sp = swdevhd; - for (i = 0; i < nswapdev; i++) { + while (mpages >= limit) { if (sp == NULL) sp = TAILQ_FIRST(&swtailq); - if (!(sp->sw_flags & SW_CLOSING)) { - blk = blist_alloc(sp->sw_blist, npages); - if (blk != SWAPBLK_NONE) { - blk += sp->sw_first; - sp->sw_used += npages; - swap_pager_avail -= npages; - swp_sizecheck(); - swdevhd = TAILQ_NEXT(sp, sw_list); - goto done; - } - } + if ((sp->sw_flags & SW_CLOSING) == 0) + blk = blist_alloc(sp->sw_blist, &npages, mpages); + if (blk != SWAPBLK_NONE) + break; sp = TAILQ_NEXT(sp, sw_list); + if (swdevhd == sp) { + mpages = npages - 1; + npages >>= 1; + } } - if (swap_pager_full != 2) { - printf("swap_pager_getswapspace(%d): failed\n", npages); - swap_pager_full = 2; - swap_pager_almost_full = 1; + if (blk != SWAPBLK_NONE) { + *io_npages = npages; + blk += sp->sw_first; + sp->sw_used += npages; + swap_pager_avail -= npages; + swp_sizecheck(); + swdevhd = TAILQ_NEXT(sp, sw_list); + } else { + if (swap_pager_full != 2) { + printf("swp_pager_getswapspace(%d): failed\n", + *io_npages); + swap_pager_full = 2; + swap_pager_almost_full = 1; + } + swdevhd = NULL; } - swdevhd = NULL; -done: mtx_unlock(&sw_dev_mtx); return (blk); } @@ -886,35 +898,30 @@ int swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) { - int n = 0; - daddr_t blk = SWAPBLK_NONE; + int i, n; + daddr_t blk; vm_pindex_t beg = start; /* save start index */ daddr_t addr, n_free, s_free; swp_pager_init_freerange(&s_free, &n_free); VM_OBJECT_WLOCK(object); - while (size) { - if (n == 0) { - n = BLIST_MAX_ALLOC; - while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) { - n >>= 1; - if (n == 0) { - swp_pager_meta_free(object, beg, start - beg); - VM_OBJECT_WUNLOCK(object); - return (-1); - } - } + while (beg + size > start) { + n = min(BLIST_MAX_ALLOC, beg + size - start); + blk = swp_pager_getswapspace(&n, 1); + if (blk == SWAPBLK_NONE) { + swp_pager_meta_free(object, beg, start - beg); + VM_OBJECT_WUNLOCK(object); + return (-1); } - addr = swp_pager_meta_build(object, start, blk); - if (addr != SWAPBLK_NONE) + for (i = 0; i < n; ++i) { + addr = swp_pager_meta_build(object, start + i, blk + i); + if (addr == SWAPBLK_NONE) + continue; swp_pager_update_freerange(&s_free, &n_free, addr); - --size; - ++start; - ++blk; - --n; + } + start += n; } swp_pager_freeswapspace(s_free, n_free); - swp_pager_meta_free(object, start, n); VM_OBJECT_WUNLOCK(object); return (0); } @@ -1384,18 +1391,8 @@ 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; - } + /* Get a block of swap of size up to size n. */ + blk = swp_pager_getswapspace(&n, 4); if (blk == SWAPBLK_NONE) { for (j = 0; j < n; ++j) rtvals[i+j] = VM_PAGER_FAIL;