X-Git-Url: http://git.indexdata.com/?a=blobdiff_plain;f=isamc%2Fmerge.c;h=b0d4315de57b08255deb99073e70c1e84bbed2ba;hb=c0826c3b142ac40494080a5c45c42abbd8ffc3f8;hp=12334c67877ce884ba0b428fc398a3bb67d60300;hpb=7107ed62f445c83a9c79b375158442f6b03f3ce6;p=idzebra-moved-to-github.git diff --git a/isamc/merge.c b/isamc/merge.c index 12334c6..b0d4315 100644 --- a/isamc/merge.c +++ b/isamc/merge.c @@ -1,13 +1,39 @@ /* - * Copyright (c) 1996, Index Data. + * Copyright (c) 1996-1998, Index Data. * See the file LICENSE for details. * Sebastian Hammer, Adam Dickmeiss * * $Log: merge.c,v $ - * Revision 1.2 1996-11-01 13:36:46 adam + * Revision 1.9 1998-03-19 10:04:38 adam + * Minor changes. + * + * Revision 1.8 1998/03/18 09:23:55 adam + * Blocks are stored in chunks on free list - up to factor 2 in speed. + * Fixed bug that could occur in block category rearrangemen. + * + * Revision 1.7 1998/03/11 11:18:18 adam + * Changed the isc_merge to take into account the mfill (minimum-fill). + * + * Revision 1.6 1998/03/06 13:54:03 adam + * Fixed two nasty bugs in isc_merge. + * + * Revision 1.5 1997/02/12 20:42:43 adam + * Bug fix: during isc_merge operations, some pages weren't marked dirty + * even though they should be. At this point the merge operation marks + * a page dirty if the previous page changed at all. A better approach is + * to mark it dirty if the last key written changed in previous page. + * + * Revision 1.4 1996/11/08 11:15:31 adam + * Number of keys in chain are stored in first block and the function + * to retrieve this information, isc_pp_num is implemented. + * + * Revision 1.3 1996/11/04 14:08:59 adam + * Optimized free block usage. + * + * Revision 1.2 1996/11/01 13:36:46 adam * New element, max_blocks_mem, that control how many blocks of max size * to store in memory during isc_merge. - * Function isc_merge now ignoreds delete/update of identical keys and + * Function isc_merge now ignores delete/update of identical keys and * the proper blocks are then non-dirty and not written in flush_blocks. * * Revision 1.1 1996/11/01 08:59:15 adam @@ -29,13 +55,66 @@ struct isc_merge_block { int dirty; /* block is different from that on file */ }; +static void opt_blocks (ISAMC is, struct isc_merge_block *mb, int ptr, int last) +{ + int i, no_dirty = 0; + for (i = 0; i mb[j].block) + j_min = j; + } + assert (j_min >= 0); + tmp = mb[j_min].block; + mb[j_min].block = mb[i].block; + mb[i].block = tmp; + mb[i].dirty = 1; + } + if (!last) + mb[i].dirty = 1; +} + static void flush_blocks (ISAMC is, struct isc_merge_block *mb, int ptr, - char *r_buf, int *firstpos, int cat, int last) + char *r_buf, int *firstpos, int cat, int last, + int *numkeys) { int i; for (i = 0; imethod->debug > 2) - logf (LOG_LOG, "isc: skip block %d %d", cat, mb[i].block); + logf (LOG_LOG, "isc: skip ptr=%d size=%d %d %d", + i, ssize, cat, mb[i].block); ++(is->files[cat].no_skip_writes); continue; } - /* consider this block number */ + /* write block */ - if (!mb[i].block) - mb[i].block = isc_alloc_block (is, cat); if (!*firstpos) + { *firstpos = mb[i].block; + src = r_buf + mb[i].offset - ISAMC_BLOCK_OFFSET_1; + ssize += ISAMC_BLOCK_OFFSET_1; - /* consider next block pointer */ - if (last && i == ptr-1) - mb[i+1].block = 0; - else if (!mb[i+1].block) - mb[i+1].block = isc_alloc_block (is, cat); - - /* write block */ - assert (mb[i+1].offset > mb[i].offset); - isc_write_dblock (is, cat, mb[i].block, r_buf + mb[i].offset, - mb[i+1].block, mb[i+1].offset - mb[i].offset); + memcpy (src+sizeof(int)+sizeof(ssize), numkeys, + sizeof(*numkeys)); + if (is->method->debug > 2) + logf (LOG_LOG, "isc: flush ptr=%d numk=%d size=%d nextpos=%d", + i, *numkeys, (int) ssize, mb[i+1].block); + } + else + { + src = r_buf + mb[i].offset - ISAMC_BLOCK_OFFSET_N; + ssize += ISAMC_BLOCK_OFFSET_N; + if (is->method->debug > 2) + logf (LOG_LOG, "isc: flush ptr=%d size=%d nextpos=%d", + i, (int) ssize, mb[i+1].block); + } + memcpy (src, &mb[i+1].block, sizeof(int)); + memcpy (src+sizeof(int), &ssize, sizeof(ssize)); + isc_write_block (is, cat, mb[i].block, src); } } +static int get_border (ISAMC is, struct isc_merge_block *mb, int ptr, + int cat, int firstpos) +{ + /* Border set to initial fill or block size depending on + whether we are creating a new one or updating and old one. + */ + + int fill = mb[ptr].block ? is->method->filecat[cat].bsize : + is->method->filecat[cat].ifill; + int off = (ptr||firstpos) ? ISAMC_BLOCK_OFFSET_N : ISAMC_BLOCK_OFFSET_1; + + assert (ptr < 199); + + return mb[ptr].offset + fill - off; +} + ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) { @@ -76,8 +180,11 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) ISAMC_PP pp; char f_item[128], *f_item_ptr; int f_more; + int last_dirty = 0; + int debug = is->method->debug; - struct isc_merge_block mb[100]; + struct isc_merge_block mb[200]; + int firstpos = 0; int cat = 0; char r_item_buf[128]; /* temporary result output */ @@ -85,9 +192,11 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) int r_offset = 0; /* current offset in r_buf */ int ptr = 0; /* pointer */ void *r_clientData; /* encode client data */ + int border; + int numKeys = 0; r_clientData = (*is->method->code_start)(ISAMC_ENCODE); - r_buf = is->merge_buf + ISAMC_BLOCK_OFFSET; + r_buf = is->merge_buf + 128; pp = isc_pp_open (is, ipos); /* read first item from file. make sure f_more indicates no boundary */ @@ -97,7 +206,7 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) f_more = 1; cat = pp->cat; - if (is->method->debug > 1) + if (debug > 1) logf (LOG_LOG, "isc: isc_merge begin %d %d", cat, pp->pos); /* read first item from i */ @@ -108,6 +217,7 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) mb[ptr].dirty = 0; mb[ptr].offset = 0; + border = get_border (is, mb, ptr, cat, firstpos); while (i_more || f_more) { char *r_item = r_item_buf; @@ -121,44 +231,57 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) { /* the resulting output is of the same category as the the original - */ - if (mb[ptr].offset == r_offset) + */ + if (r_offset <= mb[ptr].offset +is->method->filecat[cat].mfill) { - /* the resulting output block is empty. Delete + /* the resulting output block is too small/empty. Delete the original (if any) - */ - if (is->method->debug > 3) + */ + if (debug > 3) logf (LOG_LOG, "isc: release A"); if (mb[ptr].block) isc_release_block (is, pp->cat, mb[ptr].block); mb[ptr].block = pp->pos; - mb[ptr].dirty = 0; + if (!mb[ptr].dirty) + mb[ptr].dirty = 1; + if (ptr > 0) + mb[ptr-1].dirty = 1; } else { /* indicate new boundary based on the original file */ mb[++ptr].block = pp->pos; - mb[ptr].dirty = 0; + mb[ptr].dirty = last_dirty; mb[ptr].offset = r_offset; -#if 0 - if (r_item && cat == is->max_cat) + if (debug > 3) + logf (LOG_LOG, "isc: bound ptr=%d,offset=%d", + ptr, r_offset); + if (cat==is->max_cat && ptr >= is->method->max_blocks_mem) { /* We are dealing with block(s) of max size. Block(s) - will be flushed. Note: the block(s) are surely not - the last one(s). + except 1 will be flushed. */ - if (is->method->debug > 2) + if (debug > 2) logf (LOG_LOG, "isc: flush A %d sections", ptr); - flush_blocks (is, mb, ptr-1, r_buf, &firstpos, cat, 0); - ptr = 0; - mb[ptr].block = pp->pos; - mb[ptr].dirty = 0; - mb[ptr].offset = 0; + flush_blocks (is, mb, ptr-1, r_buf, &firstpos, cat, + 0, &pp->numKeys); + mb[0].block = mb[ptr-1].block; + mb[0].dirty = mb[ptr-1].dirty; + memcpy (r_buf, r_buf + mb[ptr-1].offset, + mb[ptr].offset - mb[ptr-1].offset); + mb[0].offset = 0; + + mb[1].block = mb[ptr].block; + mb[1].dirty = mb[ptr].dirty; + mb[1].offset = mb[ptr].offset - mb[ptr-1].offset; + ptr = 1; + r_offset = mb[ptr].offset; } -#endif } } + border = get_border (is, mb, ptr, cat, firstpos); } + last_dirty = 0; if (!f_more) cmp = -1; else if (!i_more) @@ -173,10 +296,10 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) i_item_ptr = i_item; i_more = (*data->read_item)(data->clientData, &i_item_ptr, &i_mode); - /* it next input item the same as current except + /* is next input item the same as current except for the delete flag? */ cmp = (*is->method->compare_item)(i_item, f_item); - if (!cmp && i_mode) + if (!cmp && i_mode) /* delete/insert nop? */ { /* yes! insert as if it was an insert only */ memcpy (r_item, i_item, i_item_ptr - i_item); @@ -188,7 +311,8 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) { /* no! delete the item */ r_item = NULL; - mb[ptr].dirty = 1; + last_dirty = 1; + mb[ptr].dirty = 2; } } else @@ -220,7 +344,8 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) abort (); } memcpy (r_item, i_item, i_item_ptr - i_item); - mb[ptr].dirty = 1; + mb[ptr].dirty = 2; + last_dirty = 1; /* move i */ i_item_ptr = i_item; i_more = (*data->read_item)(data->clientData, &i_item_ptr, @@ -230,25 +355,16 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) { char *r_out_ptr = r_buf + r_offset; int new_offset; - int border; - - /* border set to initial fill or block size depending on - whether we are creating a new one or updating and old one - */ - if (mb[ptr].block) - border = mb[ptr].offset + is->method->filecat[cat].bsize - -ISAMC_BLOCK_OFFSET; - else - border = mb[ptr].offset + is->method->filecat[cat].ifill - -ISAMC_BLOCK_OFFSET; (*is->method->code_item)(ISAMC_ENCODE, r_clientData, &r_out_ptr, &r_item); new_offset = r_out_ptr - r_buf; - if (border >= r_offset && border < new_offset) + numKeys++; + + if (border < new_offset && border >= r_offset) { - if (is->method->debug > 2) + if (debug > 2) logf (LOG_LOG, "isc: border %d %d", ptr, border); /* Max size of current block category reached ... make new virtual block entry */ @@ -261,10 +377,10 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) except one will be flushed. Note: the block(s) are surely not the last one(s). */ - if (is->method->debug > 2) + if (debug > 2) logf (LOG_LOG, "isc: flush B %d sections", ptr-1); - flush_blocks (is, mb, ptr-1, r_buf, &firstpos, cat, 0); - + flush_blocks (is, mb, ptr-1, r_buf, &firstpos, cat, + 0, &pp->numKeys); mb[0].block = mb[ptr-1].block; mb[0].dirty = mb[ptr-1].dirty; memcpy (r_buf, r_buf + mb[ptr-1].offset, @@ -278,7 +394,8 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) new_offset - r_offset); new_offset = (new_offset - r_offset) + mb[1].offset; ptr = 1; - } + } + border = get_border (is, mb, ptr, cat, firstpos); } r_offset = new_offset; } @@ -302,56 +419,81 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) cat++; mb[j].dirty = 1; mb[j].block = 0; + mb[ptr].offset = r_offset; for (i = 1; i < ptr; i++) { int border = is->method->filecat[cat].ifill - - ISAMC_BLOCK_OFFSET + mb[j].offset; - if (is->method->debug > 3) + ISAMC_BLOCK_OFFSET_1 + mb[j].offset; + if (debug > 3) logf (LOG_LOG, "isc: remap %d border=%d", i, border); if (mb[i+1].offset > border && mb[i].offset <= border) { - if (is->method->debug > 3) + if (debug > 3) logf (LOG_LOG, "isc: to %d %d", j, mb[i].offset); mb[++j].dirty = 1; mb[j].block = 0; mb[j].offset = mb[i].offset; } } - if (is->method->debug > 2) + if (debug > 2) logf (LOG_LOG, "isc: remap from %d to %d sections to cat %d", ptr, j, cat); ptr = j; + border = get_border (is, mb, ptr, cat, firstpos); + if (debug > 3) + logf (LOG_LOG, "isc: border=%d r_offset=%d", border, r_offset); } } if (mb[ptr].offset < r_offset) { /* make the final boundary offset */ - mb[++ptr].dirty = 1; /* ignored by flush_blocks */ - mb[ptr].block = 0; /* ignored by flush_blocks */ + mb[++ptr].dirty = 1; + mb[ptr].block = 0; mb[ptr].offset = r_offset; } else { /* empty output. Release last block if any */ if (cat == pp->cat && mb[ptr].block) { - if (is->method->debug > 3) + if (debug > 3) logf (LOG_LOG, "isc: release C"); isc_release_block (is, pp->cat, mb[ptr].block); mb[ptr].block = 0; - mb[ptr].dirty = 1; + if (ptr > 0) + mb[ptr-1].dirty = 1; } } - if (is->method->debug > 2) + if (debug > 2) logf (LOG_LOG, "isc: flush C, %d sections", ptr); + if (firstpos) + { + /* we have to patch initial block with num keys if that + has changed */ + if (numKeys != isc_pp_num (pp)) + { + if (debug > 2) + logf (LOG_LOG, "isc: patch num keys firstpos=%d num=%d", + firstpos, numKeys); + bf_write (is->files[cat].bf, firstpos, ISAMC_BLOCK_OFFSET_N, + sizeof(numKeys), &numKeys); + } + } + else if (ptr > 0) + { /* we haven't flushed initial block yet and there surely are some + blocks to flush. Make first block dirty if numKeys differ */ + if (numKeys != isc_pp_num (pp)) + mb[0].dirty = 1; + } /* flush rest of block(s) in r_buf */ - flush_blocks (is, mb, ptr, r_buf, &firstpos, cat, 1); + flush_blocks (is, mb, ptr, r_buf, &firstpos, cat, 1, &numKeys); (*is->method->code_stop)(ISAMC_ENCODE, r_clientData); if (!firstpos) cat = 0; - if (is->method->debug > 1) + if (debug > 1) logf (LOG_LOG, "isc: isc_merge return %d %d", cat, firstpos); + isc_pp_close (pp); return cat + firstpos * 8; }