X-Git-Url: http://git.indexdata.com/?a=blobdiff_plain;f=isamc%2Fmerge.c;h=e0b7fce8e07602125d8bb00ce7f42f0350572f7b;hb=ef696645cc3b7e0f4027008d1dc589c0f0f90c1f;hp=bb89fa10093ecb2593e200d269897fef386dd96b;hpb=7af74fb4a98eb4baeffb6a1a648e837563fe7b4f;p=idzebra-moved-to-github.git diff --git a/isamc/merge.c b/isamc/merge.c index bb89fa1..e0b7fce 100644 --- a/isamc/merge.c +++ b/isamc/merge.c @@ -1,10 +1,35 @@ /* - * 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 1.4 1996-11-08 11:15:31 adam + * Revision 1.11 1999-05-26 07:49:14 adam + * C++ compilation. + * + * Revision 1.10 1998/03/19 12:22:09 adam + * Minor change. + * + * 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. * @@ -36,6 +61,34 @@ 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, int *numkeys) @@ -44,8 +97,30 @@ static void flush_blocks (ISAMC is, struct isc_merge_block *mb, int ptr, 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 */ - - if (!mb[i].block) - mb[i].block = isc_alloc_block (is, cat); - - /* 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 */ - ssize = mb[i+1].offset - mb[i].offset; - assert (ssize); if (!*firstpos) { @@ -82,16 +145,16 @@ static void flush_blocks (ISAMC is, struct isc_merge_block *mb, int ptr, memcpy (src+sizeof(int)+sizeof(ssize), numkeys, sizeof(*numkeys)); if (is->method->debug > 2) - logf (LOG_LOG, "isc: flush numk=%d size=%d nextpos=%d", - *numkeys, (int) ssize, mb[i+1].block); + 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 size=%d nextpos=%d", - (int) ssize, mb[i+1].block); + 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)); @@ -124,6 +187,8 @@ 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[200]; @@ -148,7 +213,7 @@ 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 */ @@ -173,35 +238,40 @@ 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 (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) 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, &numKeys); - + 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, @@ -218,6 +288,7 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) } border = get_border (is, mb, ptr, cat, firstpos); } + last_dirty = 0; if (!f_more) cmp = -1; else if (!i_more) @@ -235,7 +306,7 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) /* 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); @@ -247,7 +318,8 @@ 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 @@ -279,7 +351,8 @@ 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, @@ -298,7 +371,7 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) 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 */ @@ -311,11 +384,10 @@ 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, &numKeys); - + 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, @@ -354,26 +426,29 @@ 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_1 + mb[j].offset; - if (is->method->debug > 3) + 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) @@ -386,15 +461,16 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) { /* 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) @@ -403,7 +479,7 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) has changed */ if (numKeys != isc_pp_num (pp)) { - if (is->method->debug > 2) + 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, @@ -422,7 +498,7 @@ ISAMC_P isc_merge (ISAMC is, ISAMC_P ipos, ISAMC_I data) (*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;