📄 lock_deadlock.c
字号:
lip->dd_id = DD_INVALID_ID; } /* * We only need consider objects that have waiters, so we use * the list of objects with waiters (dd_objs) instead of traversing * the entire hash table. For each object, we traverse the waiters * list and add an entry in the waitsfor matrix for each waiter/holder * combination. */obj_loop: for (op = SH_TAILQ_FIRST(®ion->dd_objs, __db_lockobj); op != NULL; op = SH_TAILQ_NEXT(op, dd_links, __db_lockobj)) { if (expire_only) goto look_waiters; CLEAR_MAP(tmpmap, nentries); /* * First we go through and create a bit map that * represents all the holders of this object. */ for (lp = SH_TAILQ_FIRST(&op->holders, __db_lock); lp != NULL; lp = SH_TAILQ_NEXT(lp, links, __db_lock)) { LOCKER_LOCK(lt, region, lp->holder, ndx); if ((ret = __lock_getlocker(lt, lp->holder, ndx, 0, &lockerp)) != 0) continue; if (lockerp->dd_id == DD_INVALID_ID) { dd = ((DB_LOCKER *)R_ADDR(<->reginfo, lockerp->master_locker))->dd_id; lockerp->dd_id = dd; switch (atype) { case DB_LOCK_MINLOCKS: case DB_LOCK_MAXLOCKS: id_array[dd].count += lockerp->nlocks; break; case DB_LOCK_MINWRITE: case DB_LOCK_MAXWRITE: id_array[dd].count += lockerp->nwrites; break; } if (F_ISSET(lockerp, DB_LOCKER_INABORT)) id_array[dd].in_abort = 1; } else dd = lockerp->dd_id; id_array[dd].valid = 1; /* * If the holder has already been aborted, then * we should ignore it for now. */ if (lp->status == DB_LSTAT_HELD) SET_MAP(tmpmap, dd); } /* * Next, for each waiter, we set its row in the matrix * equal to the map of holders we set up above. */look_waiters: for (is_first = 1, lp = SH_TAILQ_FIRST(&op->waiters, __db_lock); lp != NULL; is_first = 0, lp = SH_TAILQ_NEXT(lp, links, __db_lock)) { LOCKER_LOCK(lt, region, lp->holder, ndx); if ((ret = __lock_getlocker(lt, lp->holder, ndx, 0, &lockerp)) != 0) continue; if (lp->status == DB_LSTAT_WAITING) { if (__lock_expired(dbenv, &now, &lockerp->lk_expire)) { lp->status = DB_LSTAT_EXPIRED; MUTEX_UNLOCK(dbenv, &lp->mutex); continue; } if (LOCK_TIME_GREATER( &min_timeout, &lockerp->lk_expire)) min_timeout = lockerp->lk_expire; } if (expire_only) continue; if (lockerp->dd_id == DD_INVALID_ID) { dd = ((DB_LOCKER *)R_ADDR(<->reginfo, lockerp->master_locker))->dd_id; lockerp->dd_id = dd; switch (atype) { case DB_LOCK_MINLOCKS: case DB_LOCK_MAXLOCKS: id_array[dd].count += lockerp->nlocks; break; case DB_LOCK_MINWRITE: case DB_LOCK_MAXWRITE: id_array[dd].count += lockerp->nwrites; break; } } else dd = lockerp->dd_id; id_array[dd].valid = 1; /* * If the transaction is pending abortion, then * ignore it on this iteration. */ if (lp->status != DB_LSTAT_WAITING) continue; entryp = bitmap + (nentries * dd); OR_MAP(entryp, tmpmap, nentries); /* * If this is the first waiter on the queue, * then we remove the waitsfor relationship * with oneself. However, if it's anywhere * else on the queue, then we have to keep * it and we have an automatic deadlock. */ if (is_first) { if (ISSET_MAP(entryp, dd)) id_array[dd].self_wait = 1; CLR_MAP(entryp, dd); } } } if (LOCK_TIME_ISVALID(®ion->next_timeout)) { if (LOCK_TIME_ISMAX(&min_timeout)) LOCK_SET_TIME_INVALID(®ion->next_timeout); else region->next_timeout = min_timeout; } if (expire_only) return (0); /* Now for each locker; record its last lock. */ for (id = 0; id < count; id++) { if (!id_array[id].valid) continue; LOCKER_LOCK(lt, region, id_array[id].id, ndx); if ((ret = __lock_getlocker(lt, id_array[id].id, ndx, 0, &lockerp)) != 0) { __db_err(dbenv, "No locks for locker %lu", (u_long)id_array[id].id); continue; } /* * If this is a master transaction, try to * find one of its children's locks first, * as they are probably more recent. */ child = SH_LIST_FIRST(&lockerp->child_locker, __db_locker); if (child != NULL) { do { lp = SH_LIST_FIRST(&child->heldby, __db_lock); if (lp != NULL && lp->status == DB_LSTAT_WAITING) { id_array[id].last_locker_id = child->id; goto get_lock; } child = SH_LIST_NEXT( child, child_link, __db_locker); } while (child != NULL); } lp = SH_LIST_FIRST(&lockerp->heldby, __db_lock); if (lp != NULL) { id_array[id].last_locker_id = lockerp->id;get_lock: id_array[id].last_lock = R_OFFSET(<->reginfo, lp); id_array[id].last_obj = lp->obj; lo = (DB_LOCKOBJ *)((u_int8_t *)lp + lp->obj); pptr = SH_DBT_PTR(&lo->lockobj); if (lo->lockobj.size >= sizeof(db_pgno_t)) memcpy(&id_array[id].pgno, pptr, sizeof(db_pgno_t)); else id_array[id].pgno = 0; } } /* * Pass complete, reset the deadlock detector bit. */ region->need_dd = 0; /* * Now we can release everything except the bitmap matrix that we * created. */ *nlockers = id; *idmap = id_array; *bmp = bitmap; *allocp = nentries; __os_free(dbenv, tmpmap); return (0);}static int__dd_find(dbenv, bmp, idmap, nlockers, nalloc, deadp) DB_ENV *dbenv; u_int32_t *bmp, nlockers, nalloc; locker_info *idmap; u_int32_t ***deadp;{ u_int32_t i, j, k, *mymap, *tmpmap, **retp; u_int ndead, ndeadalloc; int ret;#undef INITIAL_DEAD_ALLOC#define INITIAL_DEAD_ALLOC 8 ndeadalloc = INITIAL_DEAD_ALLOC; ndead = 0; if ((ret = __os_malloc(dbenv, ndeadalloc * sizeof(u_int32_t *), &retp)) != 0) return (ret); /* * For each locker, OR in the bits from the lockers on which that * locker is waiting. */ for (mymap = bmp, i = 0; i < nlockers; i++, mymap += nalloc) { if (!idmap[i].valid) continue; for (j = 0; j < nlockers; j++) { if (!ISSET_MAP(mymap, j)) continue; /* Find the map for this bit. */ tmpmap = bmp + (nalloc * j); OR_MAP(mymap, tmpmap, nalloc); if (!ISSET_MAP(mymap, i)) continue; /* Make sure we leave room for NULL. */ if (ndead + 2 >= ndeadalloc) { ndeadalloc <<= 1; /* * If the alloc fails, then simply return the * deadlocks that we already have. */ if (__os_realloc(dbenv, ndeadalloc * sizeof(u_int32_t), &retp) != 0) { retp[ndead] = NULL; *deadp = retp; return (0); } } retp[ndead++] = mymap; /* Mark all participants in this deadlock invalid. */ for (k = 0; k < nlockers; k++) if (ISSET_MAP(mymap, k)) idmap[k].valid = 0; break; } } retp[ndead] = NULL; *deadp = retp; return (0);}static int__dd_abort(dbenv, info) DB_ENV *dbenv; locker_info *info;{ struct __db_lock *lockp; DB_LOCKER *lockerp; DB_LOCKOBJ *sh_obj; DB_LOCKREGION *region; DB_LOCKTAB *lt; u_int32_t ndx; int ret; lt = dbenv->lk_handle; region = lt->reginfo.primary; LOCKREGION(dbenv, lt); /* * Get the locker. If its gone or was aborted while * we were detecting return that. */ LOCKER_LOCK(lt, region, info->last_locker_id, ndx); if ((ret = __lock_getlocker(lt, info->last_locker_id, ndx, 0, &lockerp)) != 0 || lockerp == NULL || F_ISSET(lockerp, DB_LOCKER_INABORT)) { if (ret == 0) ret = DB_ALREADY_ABORTED; goto out; } /* * Find the locker's last lock. * It is possible for this lock to have been freed, * either though a timeout or another detector run. */ if ((lockp = SH_LIST_FIRST(&lockerp->heldby, __db_lock)) == NULL) { ret = DB_ALREADY_ABORTED; goto out; } if (R_OFFSET(<->reginfo, lockp) != info->last_lock || lockp->holder != lockerp->id || lockp->obj != info->last_obj || lockp->status != DB_LSTAT_WAITING) { ret = DB_ALREADY_ABORTED; goto out; } sh_obj = (DB_LOCKOBJ *)((u_int8_t *)lockp + lockp->obj); /* Abort lock, take it off list, and wake up this lock. */ SHOBJECT_LOCK(lt, region, sh_obj, ndx); lockp->status = DB_LSTAT_ABORTED; SH_TAILQ_REMOVE(&sh_obj->waiters, lockp, links, __db_lock); /* * Either the waiters list is now empty, in which case we remove * it from dd_objs, or it is not empty, in which case we need to * do promotion. */ if (SH_TAILQ_FIRST(&sh_obj->waiters, __db_lock) == NULL) SH_TAILQ_REMOVE(®ion->dd_objs, sh_obj, dd_links, __db_lockobj); else ret = __lock_promote(lt, sh_obj, 0); MUTEX_UNLOCK(dbenv, &lockp->mutex); region->stat.st_ndeadlocks++; UNLOCKREGION(dbenv, lt); return (0);out: UNLOCKREGION(dbenv, lt); return (ret);}#ifdef DIAGNOSTICstatic void__dd_debug(dbenv, idmap, bitmap, nlockers, nalloc) DB_ENV *dbenv; locker_info *idmap; u_int32_t *bitmap, nlockers, nalloc;{ DB_MSGBUF mb; u_int32_t i, j, *mymap; DB_MSGBUF_INIT(&mb); __db_msg(dbenv, "Waitsfor array\nWaiter:\tWaiting on:"); for (mymap = bitmap, i = 0; i < nlockers; i++, mymap += nalloc) { if (!idmap[i].valid) continue; __db_msgadd(dbenv, &mb, /* Waiter. */ "%lx/%lu:\t", (u_long)idmap[i].id, (u_long)idmap[i].pgno); for (j = 0; j < nlockers; j++) if (ISSET_MAP(mymap, j)) __db_msgadd(dbenv, &mb, " %lx", (u_long)idmap[j].id); __db_msgadd(dbenv, &mb, " %lu", (u_long)idmap[i].last_lock); DB_MSGBUF_FLUSH(dbenv, &mb); }}#endif/* * Given a bitmap that contains a deadlock, verify that the bit * specified in the which parameter indicates a transaction that * is actually deadlocked. Return 1 if really deadlocked, 0 otherwise. * deadmap -- the array that identified the deadlock. * tmpmap -- a copy of the initial bitmaps from the dd_build phase. * origmap -- a temporary bit map into which we can OR things. * nlockers -- the number of actual lockers under consideration. * nalloc -- the number of words allocated for the bitmap. * which -- the locker in question. */static int__dd_verify(idmap, deadmap, tmpmap, origmap, nlockers, nalloc, which) locker_info *idmap; u_int32_t *deadmap, *tmpmap, *origmap; u_int32_t nlockers, nalloc, which;{ u_int32_t *tmap; u_int32_t j; int count; memset(tmpmap, 0, sizeof(u_int32_t) * nalloc); /* * In order for "which" to be actively involved in * the deadlock, removing him from the evaluation * must remove the deadlock. So, we OR together everyone * except which; if all the participants still have their * bits set, then the deadlock persists and which does * not participate. If the deadlock does not persist * then "which" does participate. */ count = 0; for (j = 0; j < nlockers; j++) { if (!ISSET_MAP(deadmap, j) || j == which) continue; /* Find the map for this bit. */ tmap = origmap + (nalloc * j); /* * We special case the first waiter who is also a holder, so * we don't automatically call that a deadlock. However, if * it really is a deadlock, we need the bit set now so that * we treat the first waiter like other waiters. */ if (idmap[j].self_wait) SET_MAP(tmap, j); OR_MAP(tmpmap, tmap, nalloc); count++; } if (count == 1) return (1); /* * Now check the resulting map and see whether * all participants still have their bit set. */ for (j = 0; j < nlockers; j++) { if (!ISSET_MAP(deadmap, j) || j == which) continue; if (!ISSET_MAP(tmpmap, j)) return (1); } return (0);}/* * __dd_isolder -- * * Figure out the relative age of two lockers. We make all lockers * older than all transactions, because that's how it's worked * historically (because lockers are lower ids). */static int__dd_isolder(a, b, lock_max, txn_max) u_int32_t a, b; u_int32_t lock_max, txn_max;{ u_int32_t max; /* Check for comparing lock-id and txnid. */ if (a <= DB_LOCK_MAXID && b > DB_LOCK_MAXID) return (1); if (b <= DB_LOCK_MAXID && a > DB_LOCK_MAXID) return (0); /* In the same space; figure out which one. */ max = txn_max; if (a <= DB_LOCK_MAXID) max = lock_max; /* * We can't get a 100% correct ordering, because we don't know * where the current interval started and if there were older * lockers outside the interval. We do the best we can. */ /* * Check for a wrapped case with ids above max. */ if (a > max && b < max) return (1); if (b > max && a < max) return (0); return (a < b);}
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -