X-Git-Url: https://git.sur5r.net/?a=blobdiff_plain;f=libraries%2Fliblmdb%2Fmdb.c;h=1e784ae11de1b8ad549037fd2b11aceed37a7cef;hb=29fd241fadc3dd49b3486f0e3556b029b716bcbf;hp=cc4dd6ddc9dedd6c49817e6873ab26bc536ffb81;hpb=7ef970436d83a734f56ddf31f8868eb6032d666a;p=openldap diff --git a/libraries/liblmdb/mdb.c b/libraries/liblmdb/mdb.c index cc4dd6ddc9..1e784ae11d 100644 --- a/libraries/liblmdb/mdb.c +++ b/libraries/liblmdb/mdb.c @@ -70,6 +70,16 @@ #include #endif +#if defined(__mips) && defined(__linux) +/* MIPS has cache coherency issues, requires explicit cache control */ +#include +extern int cacheflush(char *addr, int nbytes, int cache); +#define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache) +#else +#define CACHEFLUSH(addr, bytes, cache) +#endif + + #include #include #include @@ -80,6 +90,12 @@ #include #include +#if defined(__sun) +/* Most platforms have posix_memalign, older may only have memalign */ +#define HAVE_MEMALIGN 1 +#include +#endif + #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER)) #include #include /* defines BYTE_ORDER on HPUX and Solaris */ @@ -735,7 +751,7 @@ typedef struct MDB_page { /** The number of overflow pages needed to store the given size. */ #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1) - /** Link in #MDB_txn.%mt_loose_pages list */ + /** Link in #MDB_txn.%mt_loose_pgs list */ #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2)) /** Header for a single key/data pair within a page. @@ -940,6 +956,8 @@ struct MDB_txn { * in this transaction, linked through #NEXT_LOOSE_PAGE(page). */ MDB_page *mt_loose_pgs; + /* #Number of loose pages (#mt_loose_pgs) */ + int mt_loose_count; /** The sorted list of dirty pages we temporarily wrote to disk * because the dirty list was full. page numbers in here are * shifted left by 1, deleted slots have the LSB set. @@ -1065,6 +1083,7 @@ typedef struct MDB_xcursor { typedef struct MDB_pgstate { pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */ txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */ + txnid_t mf_pgoldest; /**< ID of oldest reader last time we looked */ } MDB_pgstate; /** The database environment. */ @@ -1092,6 +1111,7 @@ struct MDB_env { MDB_meta *me_metas[2]; /**< pointers to the two meta pages */ void *me_pbuf; /**< scratch area for DUPSORT put() */ MDB_txn *me_txn; /**< current write transaction */ + MDB_txn *me_txn0; /**< prealloc'd write transaction */ size_t me_mapsize; /**< size of the data memory map */ off_t me_size; /**< current file size */ pgno_t me_maxpg; /**< me_mapsize / me_psize */ @@ -1102,6 +1122,7 @@ struct MDB_env { MDB_pgstate me_pgstate; /**< state of old pages from freeDB */ # define me_pglast me_pgstate.mf_pglast # define me_pghead me_pgstate.mf_pghead +# define me_pgoldest me_pgstate.mf_pgoldest MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */ /** IDL of pages that became unused in a write txn */ MDB_IDL me_free_pgs; @@ -1622,10 +1643,11 @@ mdb_page_loose(MDB_cursor *mc, MDB_page *mp) { int loose = 0; pgno_t pgno = mp->mp_pgno; + MDB_txn *txn = mc->mc_txn; if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) { - if (mc->mc_txn->mt_parent) { - MDB_ID2 *dl = mc->mc_txn->mt_u.dirty_list; + if (txn->mt_parent) { + MDB_ID2 *dl = txn->mt_u.dirty_list; /* If txn has a parent, make sure the page is in our * dirty list. */ @@ -1634,7 +1656,7 @@ mdb_page_loose(MDB_cursor *mc, MDB_page *mp) if (x <= dl[0].mid && dl[x].mid == pgno) { if (mp != dl[x].mptr) { /* bad cursor? */ mc->mc_flags &= ~(C_INITIALIZED|C_EOF); - mc->mc_txn->mt_flags |= MDB_TXN_ERROR; + txn->mt_flags |= MDB_TXN_ERROR; return MDB_CORRUPTED; } /* ok, it's ours */ @@ -1649,11 +1671,12 @@ mdb_page_loose(MDB_cursor *mc, MDB_page *mp) if (loose) { DPRINTF(("loosen db %d page %"Z"u", DDBI(mc), mp->mp_pgno)); - NEXT_LOOSE_PAGE(mp) = mc->mc_txn->mt_loose_pgs; - mc->mc_txn->mt_loose_pgs = mp; + NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs; + txn->mt_loose_pgs = mp; + txn->mt_loose_count++; mp->mp_flags |= P_LOOSE; } else { - int rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, pgno); + int rc = mdb_midl_append(&txn->mt_free_pgs, pgno); if (rc) return rc; } @@ -1930,16 +1953,18 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp) MDB_txn *txn = mc->mc_txn; MDB_env *env = txn->mt_env; pgno_t pgno, *mop = env->me_pghead; - unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1; + unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1; MDB_page *np; txnid_t oldest = 0, last; MDB_cursor_op op; MDB_cursor m2; + int found_old = 0; /* If there are any loose pages, just use them */ if (num == 1 && txn->mt_loose_pgs) { np = txn->mt_loose_pgs; txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np); + txn->mt_loose_count--; DPRINTF(("db %d use loose page %"Z"u", DDBI(mc), np->mp_pgno)); *mp = np; @@ -1957,7 +1982,7 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp) for (op = MDB_FIRST;; op = MDB_NEXT) { MDB_val key, data; MDB_node *leaf; - pgno_t *idl, old_id, new_id; + pgno_t *idl; /* Seek a big enough contiguous page range. Prefer * pages at the tail, just truncating the list. @@ -1975,8 +2000,8 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp) if (op == MDB_FIRST) { /* 1st iteration */ /* Prepare to fetch more and coalesce */ - oldest = mdb_find_oldest(txn); last = env->me_pglast; + oldest = env->me_pgoldest; mdb_cursor_init(&m2, txn, FREE_DBI, NULL); if (last) { op = MDB_SET_RANGE; @@ -1991,8 +2016,15 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp) last++; /* Do not fetch more if the record will be too recent */ - if (oldest <= last) - break; + if (oldest <= last) { + if (!found_old) { + oldest = mdb_find_oldest(txn); + env->me_pgoldest = oldest; + found_old = 1; + } + if (oldest <= last) + break; + } rc = mdb_cursor_get(&m2, &key, NULL, op); if (rc) { if (rc == MDB_NOTFOUND) @@ -2000,8 +2032,15 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp) goto fail; } last = *(txnid_t*)key.mv_data; - if (oldest <= last) - break; + if (oldest <= last) { + if (!found_old) { + oldest = mdb_find_oldest(txn); + env->me_pgoldest = oldest; + found_old = 1; + } + if (oldest <= last) + break; + } np = m2.mc_pg[m2.mc_top]; leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]); if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS) @@ -2023,21 +2062,12 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp) #if (MDB_DEBUG) > 1 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u", last, txn->mt_dbs[FREE_DBI].md_root, i)); - for (k = i; k; k--) - DPRINTF(("IDL %"Z"u", idl[k])); + for (j = i; j; j--) + DPRINTF(("IDL %"Z"u", idl[j])); #endif /* Merge in descending sorted order */ - j = mop_len; - k = mop_len += i; - mop[0] = (pgno_t)-1; - old_id = mop[j]; - while (i) { - new_id = idl[i--]; - for (; old_id < new_id; old_id = mop[--j]) - mop[k--] = old_id; - mop[k--] = new_id; - } - mop[0] = mop_len; + mdb_midl_xmerge(mop, idl); + mop_len = mop[0]; } /* Use new pages from the map when nothing suitable in the freeDB */ @@ -2595,6 +2625,10 @@ mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret) } size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1); if (!(flags & MDB_RDONLY)) { + if (!parent) { + txn = env->me_txn0; + goto ok; + } size += env->me_maxdbs * sizeof(MDB_cursor *); /* child txns use parent's dbiseqs */ if (!parent) @@ -2622,6 +2656,7 @@ mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret) } txn->mt_env = env; +ok: if (parent) { unsigned int i; txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE); @@ -2664,9 +2699,10 @@ mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret) } else { rc = mdb_txn_renew0(txn); } - if (rc) - free(txn); - else { + if (rc) { + if (txn != env->me_txn0) + free(txn); + } else { *ret = txn; DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u", txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', @@ -2793,7 +2829,8 @@ mdb_txn_abort(MDB_txn *txn) if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader) txn->mt_u.reader->mr_pid = 0; - free(txn); + if (txn != txn->mt_env->me_txn0) + free(txn); } /** Save the freelist as of this transaction to the freeDB. @@ -2822,30 +2859,17 @@ mdb_freelist_save(MDB_txn *txn) return rc; } - /* Dispose of loose pages. Usually they will have all - * been used up by the time we get here. - */ - if (txn->mt_loose_pgs) { + if (!env->me_pghead && txn->mt_loose_pgs) { + /* Put loose page numbers in mt_free_pgs, since + * we may be unable to return them to me_pghead. + */ MDB_page *mp = txn->mt_loose_pgs; - /* Just return them to freeDB */ - if (env->me_pghead) { - int i, j; - mop = env->me_pghead; - for (; mp; mp = NEXT_LOOSE_PAGE(mp)) { - pgno_t pg = mp->mp_pgno; - j = mop[0] + 1; - for (i = mop[0]; i && mop[i] < pg; i--) - mop[j--] = mop[i]; - mop[j] = pg; - mop[0] += 1; - } - } else { - /* Oh well, they were wasted. Put on freelist */ - for (; mp; mp = NEXT_LOOSE_PAGE(mp)) { - mdb_midl_append(&txn->mt_free_pgs, mp->mp_pgno); - } - } + if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0) + return rc; + for (; mp; mp = NEXT_LOOSE_PAGE(mp)) + mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno); txn->mt_loose_pgs = NULL; + txn->mt_loose_count = 0; } /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */ @@ -2909,7 +2933,7 @@ mdb_freelist_save(MDB_txn *txn) } mop = env->me_pghead; - mop_len = mop ? mop[0] : 0; + mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count; /* Reserve records for me_pghead[]. Split it if multi-page, * to avoid searching freeDB for a page range. Use keys in @@ -2949,6 +2973,28 @@ mdb_freelist_save(MDB_txn *txn) total_room += head_room; } + /* Return loose page numbers to me_pghead, though usually none are + * left at this point. The pages themselves remain in dirty_list. + */ + if (txn->mt_loose_pgs) { + MDB_page *mp = txn->mt_loose_pgs; + unsigned count = txn->mt_loose_count; + MDB_IDL loose; + /* Room for loose pages + temp IDL with same */ + if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0) + return rc; + mop = env->me_pghead; + loose = mop + MDB_IDL_ALLOCLEN(mop) - count; + for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp)) + loose[ ++count ] = mp->mp_pgno; + loose[0] = count; + mdb_midl_sort(loose); + mdb_midl_xmerge(mop, loose); + txn->mt_loose_pgs = NULL; + txn->mt_loose_count = 0; + mop_len = mop[0]; + } + /* Fill in the reserved me_pghead records */ rc = MDB_SUCCESS; if (mop_len) { @@ -3101,6 +3147,12 @@ mdb_page_flush(MDB_txn *txn, int keep) #endif /* _WIN32 */ } + /* MIPS has cache coherency issues, this is a no-op everywhere else + * Note: for any size >= on-chip cache size, entire on-chip cache is + * flushed. + */ + CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE); + for (i = keep; ++i <= pagecount; ) { dp = dl[i].mptr; /* This is a page we skipped above */ @@ -3257,6 +3309,7 @@ mdb_txn_commit(MDB_txn *txn) for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp)) ; *lp = txn->mt_loose_pgs; + parent->mt_loose_count += txn->mt_loose_count; parent->mt_child = NULL; mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead); @@ -3319,6 +3372,10 @@ mdb_txn_commit(MDB_txn *txn) (rc = mdb_env_write_meta(txn))) goto fail; + /* Free P_LOOSE pages left behind in dirty_list */ + if (!(env->me_flags & MDB_WRITEMAP)) + mdb_dlist_free(txn); + done: env->me_pglast = 0; env->me_txn = NULL; @@ -3326,7 +3383,8 @@ done: if (env->me_txns) UNLOCK_MUTEX_W(env); - free(txn); + if (txn != env->me_txn0) + free(txn); return MDB_SUCCESS; @@ -3474,6 +3532,7 @@ mdb_env_write_meta(MDB_txn *txn) { MDB_env *env; MDB_meta meta, metab, *mp; + size_t mapsize; off_t off; int rc, len, toggle; char *ptr; @@ -3490,11 +3549,13 @@ mdb_env_write_meta(MDB_txn *txn) env = txn->mt_env; mp = env->me_metas[toggle]; + mapsize = env->me_metas[toggle ^ 1]->mm_mapsize; + /* Persist any increases of mapsize config */ + if (mapsize < env->me_mapsize) + mapsize = env->me_mapsize; if (env->me_flags & MDB_WRITEMAP) { - /* Persist any increases of mapsize config */ - if (env->me_mapsize > mp->mm_mapsize) - mp->mm_mapsize = env->me_mapsize; + mp->mm_mapsize = mapsize; mp->mm_dbs[0] = txn->mt_dbs[0]; mp->mm_dbs[1] = txn->mt_dbs[1]; mp->mm_last_pg = txn->mt_next_pgno - 1; @@ -3521,22 +3582,15 @@ mdb_env_write_meta(MDB_txn *txn) metab.mm_txnid = env->me_metas[toggle]->mm_txnid; metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg; - ptr = (char *)&meta; - if (env->me_mapsize > mp->mm_mapsize) { - /* Persist any increases of mapsize config */ - meta.mm_mapsize = env->me_mapsize; - off = offsetof(MDB_meta, mm_mapsize); - } else { - off = offsetof(MDB_meta, mm_dbs[0].md_depth); - } - len = sizeof(MDB_meta) - off; - - ptr += off; + meta.mm_mapsize = mapsize; meta.mm_dbs[0] = txn->mt_dbs[0]; meta.mm_dbs[1] = txn->mt_dbs[1]; meta.mm_last_pg = txn->mt_next_pgno - 1; meta.mm_txnid = txn->mt_txnid; + off = offsetof(MDB_meta, mm_mapsize); + ptr = (char *)&meta + off; + len = sizeof(MDB_meta) - off; if (toggle) off += env->me_psize; off += PAGEHDRSZ; @@ -3575,6 +3629,8 @@ fail: env->me_flags |= MDB_FATAL_ERROR; return rc; } + /* MIPS has cache coherency issues, this is a no-op everywhere else */ + CACHEFLUSH(env->me_map + off, len, DCACHE); done: /* Memory ordering issues are irrelevant; since the entire writer * is wrapped by wmutex, all of these changes will become visible @@ -4459,6 +4515,22 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode if (!((flags & MDB_RDONLY) || (env->me_pbuf = calloc(1, env->me_psize)))) rc = ENOMEM; + if (!(flags & MDB_RDONLY)) { + MDB_txn *txn; + int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs * + (sizeof(MDB_db)+sizeof(MDB_cursor)+sizeof(unsigned int)+1); + txn = calloc(1, size); + if (txn) { + txn->mt_dbs = (MDB_db *)((char *)txn + tsize); + txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs); + txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs); + txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs); + txn->mt_env = env; + env->me_txn0 = txn; + } else { + rc = ENOMEM; + } + } } leave: @@ -5345,8 +5417,10 @@ mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op) if (op == MDB_PREV || op == MDB_PREV_DUP) { rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV); if (op != MDB_PREV || rc != MDB_NOTFOUND) { - if (rc == MDB_SUCCESS) + if (rc == MDB_SUCCESS) { MDB_GET_KEY(leaf, key); + mc->mc_flags &= ~C_EOF; + } return rc; } } else { @@ -5823,6 +5897,14 @@ fetchm: rc = MDB_INCOMPATIBLE; break; } + { + MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); + if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) { + MDB_GET_KEY(leaf, key); + rc = mdb_node_read(mc->mc_txn, leaf, data); + break; + } + } if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) { rc = EINVAL; break; @@ -6036,6 +6118,24 @@ mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data, return MDB_BAD_VALSIZE; ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize); memcpy(ptr, key->mv_data, ksize); +fix_parent: + /* if overwriting slot 0 of leaf, need to + * update branch key if there is a parent page + */ + if (mc->mc_top && !mc->mc_ki[mc->mc_top]) { + unsigned short top = mc->mc_top; + mc->mc_top--; + /* slot 0 is always an empty key, find real slot */ + while (mc->mc_top && !mc->mc_ki[mc->mc_top]) + mc->mc_top--; + if (mc->mc_ki[mc->mc_top]) + rc2 = mdb_update_key(mc, key); + else + rc2 = MDB_SUCCESS; + mc->mc_top = top; + if (rc2) + return rc2; + } return MDB_SUCCESS; } @@ -6241,8 +6341,10 @@ current: data->mv_data = olddata.mv_data; else if (!(mc->mc_flags & C_SUB)) memcpy(olddata.mv_data, data->mv_data, data->mv_size); - else + else { memcpy(NODEKEY(leaf), key->mv_data, key->mv_size); + goto fix_parent; + } return MDB_SUCCESS; } mdb_node_del(mc, 0); @@ -6966,6 +7068,9 @@ mdb_cursor_count(MDB_cursor *mc, size_t *countp) if (!(mc->mc_flags & C_INITIALIZED)) return EINVAL; + if (!mc->mc_snum || (mc->mc_flags & C_EOF)) + return MDB_NOTFOUND; + leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]); if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) { *countp = 1; @@ -8414,15 +8519,21 @@ mdb_env_copyfd1(MDB_env *env, HANDLE fd) #ifdef _WIN32 my.mc_mutex = CreateMutex(NULL, FALSE, NULL); my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL); - my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_psize); + my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize); if (my.mc_wbuf[0] == NULL) return errno; #else pthread_mutex_init(&my.mc_mutex, NULL); pthread_cond_init(&my.mc_cond, NULL); - rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_psize, MDB_WBUF*2); +#ifdef HAVE_MEMALIGN + my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2); + if (my.mc_wbuf[0] == NULL) + return errno; +#else + rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2); if (rc) return rc; +#endif #endif memset(my.mc_wbuf[0], 0, MDB_WBUF*2); my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF; @@ -8658,6 +8769,7 @@ mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags) goto leave; } + if (env->me_psize >= env->me_os_psize) { #ifdef O_DIRECT /* Set O_DIRECT if the file system supports it */ if ((rc = fcntl(newfd, F_GETFL)) != -1) @@ -8670,6 +8782,7 @@ mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags) goto leave; } #endif + } rc = mdb_env_copyfd2(env, newfd, flags); @@ -8840,6 +8953,7 @@ int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *db MDB_val key, data; MDB_dbi i; MDB_cursor mc; + MDB_db dummy; int rc, dbflag, exact; unsigned int unused = 0, seq; size_t len; @@ -8909,7 +9023,6 @@ int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *db return MDB_INCOMPATIBLE; } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) { /* Create if requested */ - MDB_db dummy; data.mv_size = sizeof(MDB_db); data.mv_data = &dummy; memset(&dummy, 0, sizeof(dummy));