ref: 1ba67b57e8e863229f5eb19fceba5f7f7c224cf5
dir: /blk.c/
#include <u.h> #include <libc.h> #include <fcall.h> #include <avl.h> #include "dat.h" #include "fns.h" #include "atomic.h" static vlong blkalloc_lk(Arena*, int); static vlong blkalloc(int, uint); static int blkdealloc_lk(Arena*, vlong); static Blk* initblk(Blk*, vlong, vlong, int); int checkflag(Blk *b, int f) { long v; v = agetl(&b->flag); return (v & f) == f; } void setflag(Blk *b, int f) { long ov, nv; while(1){ ov = agetl(&b->flag); nv = ov | f; if(acasl(&b->flag, ov, nv) == ov) break; } } void clrflag(Blk *b, int f) { long ov, nv; while(1){ ov = agetl(&b->flag); nv = ov & ~f; if(acasl(&b->flag, ov, nv) == ov) break; } } int syncblk(Blk *b) { assert(checkflag(b, Bfinal)); assert(b->bp.addr >= 0); clrflag(b, Bdirty); return pwrite(fs->fd, b->buf, Blksz, b->bp.addr); } static Blk* readblk(vlong bp, int flg) { vlong off, rem, n; char *p; Blk *b; assert(bp != -1); if((b = cachepluck()) == nil) return nil; b->alloced = getcallerpc(&bp); off = bp; rem = Blksz; while(rem != 0){ n = pread(fs->fd, b->buf, rem, off); if(n <= 0){ dropblk(b); return nil; } off += n; rem -= n; } b->cnext = nil; b->cprev = nil; b->hnext = nil; b->flag = 0; b->bp.addr = bp; b->bp.hash = -1; b->bp.gen = -1; b->fnext = nil; b->nval = 0; b->valsz = 0; b->nbuf = 0; b->bufsz = 0; b->logsz = 0; p = b->buf + 2; b->type = (flg&GBraw) ? Tdat : UNPACK16(b->buf+0); switch(b->type){ default: fprint(2, "invalid block type %d @%llx\n", b->type, bp); abort(); break; case Tdat: case Tsuper: b->data = b->buf; break; case Tarena: b->data = p; break; case Tdlist: case Tlog: b->logsz = UNPACK16(p); p += 2; b->logh = UNPACK64(p); p += 8; b->logp = unpackbp(p, Ptrsz); p += Ptrsz; assert(p - b->buf == Loghdsz); b->data = p; break; case Tpivot: b->nval = UNPACK16(p); p += 2; b->valsz = UNPACK16(p); p += 2; b->nbuf = UNPACK16(p); p += 2; b->bufsz = UNPACK16(p); p += 2; assert(p - b->buf == Pivhdsz); b->data = p; break; case Tleaf: b->nval = UNPACK16(p); p += 2; b->valsz = UNPACK16(p); p += 2; assert(p - b->buf == Leafhdsz); b->data = p; break; } assert(b->magic == Magic); return b; } static Arena* pickarena(uint ty, uint hint, int tries) { uint n; n = hint + tries + ainc(&fs->roundrobin)/1024; if(ty == Tdat) n++; if(hint % fs->narena == 0) n++; return &fs->arenas[n%fs->narena]; } Arena* getarena(vlong b) { int i; i = b / fs->arenasz; if(i < 0 || i >= fs->narena){ werrstr("out of range block %lld", b); abort(); return nil; } return &fs->arenas[i]; } static int freerange(Avltree *t, vlong off, vlong len) { Arange *r, *s; assert(len % Blksz == 0); if((r = calloc(1, sizeof(Arange))) == nil) return -1; r->off = off; r->len = len; assert(avllookup(t, r, 0) == nil); avlinsert(t, r); Again: s = (Arange*)avlprev(r); if(s != nil && s->off+s->len == r->off){ avldelete(t, r); s->len = s->len + r->len; free(r); r = s; goto Again; } s = (Arange*)avlnext(r); if(s != nil && r->off+r->len == s->off){ avldelete(t, r); s->off = r->off; s->len = s->len + r->len; free(r); r = s; goto Again; } return 0; } static int grabrange(Avltree *t, vlong off, vlong len) { Arange *r, *s, q; vlong l; assert(len % Blksz == 0); q.off = off; q.len = len; r = (Arange*)avllookup(t, &q.Avl, -1); if(r == nil || off + len > r->off + r->len) abort(); if(off == r->off){ r->off += len; r->len -= len; }else if(off + len == r->off + r->len){ r->len -= len; }else if(off > r->off && off+len < r->off + r->len){ if((s = malloc(sizeof(Arange))) == nil) return -1; l = r->len; s->off = off + len; r->len = off - r->off; s->len = l - r->len - len; avlinsert(t, s); }else abort(); if(r->len == 0){ avldelete(t, r); free(r); } return 0; } static Blk* mklogblk(vlong o) { Blk *lb; if((lb = cachepluck()) == nil) return nil; initblk(lb, o, -1, Tlog); finalize(lb); if(syncblk(lb) == -1){ dropblk(lb); return nil; } return lb; } /* * Logs an allocation. Must be called * with arena lock held. Duplicates some * of the work in allocblk to prevent * recursion. */ static int logappend(Arena *a, vlong off, vlong len, int op) { vlong o; Blk *nl, *lb; char *p, *name; lb = a->logtl; assert((off & 0xff) == 0); assert(op == LogAlloc || op == LogFree || op == LogSync); if(op != LogSync){ assert(lb == nil || lb->type == Tlog); assert(off >= a->hd->bp.addr + Blksz); assert(off < a->tl->bp.addr); } switch(op){ case LogAlloc: name = "alloc"; break; case LogFree: name = "free"; break; case LogSync: name = "sync"; break; default: name = "???"; break; } assert(lb == nil || lb->logsz >= 0); dprint("logop %llx+%llx@%x: %s\n", off, len, lb?lb->logsz:-1, name); /* * move to the next block when we have * too little room in the log: * We're appending up to 16 bytes as * part of the operation, followed by * 16 bytes of new log entry allocation * and chaining. */ if(lb == nil || lb->logsz >= Logspc - Logslop){ if((o = blkalloc_lk(a, 1)) == -1) return -1; if((nl = mklogblk(o)) == nil) return -1; p = lb->data + lb->logsz; PACK64(p, o|LogAlloc1); lb->logsz += 8; lb->logp = nl->bp; setflag(lb, Bdirty); finalize(lb); if(syncblk(lb) == -1){ fs->broken = 1; return -1; } dropblk(lb); a->logtl = nl; a->nlog++; lb = nl; } setflag(lb, Bdirty); if(len == Blksz){ if(op == LogAlloc) op = LogAlloc1; else if(op == LogFree) op = LogFree1; } off |= op; p = lb->data + lb->logsz; PACK64(p, off); lb->logsz += 8; if(op >= Log2wide){ PACK64(p+8, len); lb->logsz += 8; } return 0; } int loadlog(Arena *a, Bptr bp) { vlong ent, off, len, gen; int op, i, n; char *d; Blk *b; dprint("loadlog %B\n", bp); while(1){ if((b = getblk(bp, 0)) == nil) return -1; dprint("\tload %B chain %B\n", bp, b->logp); /* the hash covers the log and offset */ for(i = 0; i < b->logsz; i += n){ d = b->data + i; ent = UNPACK64(d); op = ent & 0xff; off = ent & ~0xff; n = (op >= Log2wide) ? 16 : 8; switch(op){ case LogSync: gen = ent >> 8; dprint("\tlog@%d: sync %llx\n", i, gen); if(gen >= fs->qgen){ if(a->logtl == nil){ b->logsz = i; a->logtl = holdblk(b); return 0; } dropblk(b); return 0; } break; case LogAlloc: case LogAlloc1: len = (op >= Log2wide) ? UNPACK64(d+8) : Blksz; dprint("\tlog@%d alloc: %llx+%llx\n", i, off, len); if(grabrange(a->free, off & ~0xff, len) == -1) return -1; a->used += len; break; case LogFree: case LogFree1: len = (op >= Log2wide) ? UNPACK64(d+8) : Blksz; dprint("\tlog@%d free: %llx+%llx\n", i, off, len); if(freerange(a->free, off & ~0xff, len) == -1){ werrstr("invalid free: %r"); return -1; } a->used -= len; break; default: n = 0; dprint("\tlog@%d: log op %d\n", i, op); abort(); break; } } if(b->logp.addr == -1){ a->logtl = b; return 0; } bp = b->logp; dropblk(b); } } int compresslog(Arena *a) { int i, nr, nblks; vlong sz, *blks; Blk *b, *nb; Arange *r; Bptr hd; char *p; if(a->logtl != nil && checkflag(a->logtl, Bdirty)) enqueue(a->logtl); /* * Prepare what we're writing back. * Arenas must be sized so that we can * keep the merged log in memory for * a rewrite. */ sz = 0; nr = 0; a->nlog = 0; for(r = (Arange*)avlmin(a->free); r != nil; r = (Arange*)avlnext(r)){ sz += 16; nr++; } /* * Make a pessimistic estimate of the number of blocks * needed to store the ranges, as well as the blocks * used to store the range allocations. * * This does modify the tree, but it's safe because * we can only be removing entries from the tree, not * splitting or inserting new ones. */ nblks = (sz+Logspc)/(Logspc - Logslop) + 16*nr/(Logspc-Logslop) + 1; if((blks = calloc(nblks, sizeof(vlong))) == nil) return -1; for(i = 0; i < nblks; i++){ blks[i] = blkalloc_lk(a, 1); if(blks[i] == -1) return -1; } /* fill up the log with the ranges from the tree */ i = 0; hd = (Bptr){blks[0], -1, -1}; if((b = cachepluck()) == nil) return -1; initblk(b, blks[i++], -1, Tlog); for(r = (Arange*)avlmin(a->free); r != nil; r = (Arange*)avlnext(r)){ if(b->logsz >= Logspc - Logslop){ if((nb = cachepluck()) == nil) return -1; a->nlog++; initblk(nb, blks[i++], -1, Tlog); b->logp = nb->bp; setflag(b, Bdirty); finalize(b); enqueue(b); dropblk(b); b = nb; } p = b->data + b->logsz; PACK64(p+0, r->off|LogFree); PACK64(p+8, r->len); setflag(b, Bdirty); b->logsz += 16; setflag(b, Bdirty); } finalize(b); enqueue(b); /* * now we have a valid freelist, and we can start * appending stuff to it. Clean up the old logs * and the eagerly allocated extra blocks. */ dropblk(a->logtl); a->loghd = hd; a->logtl = b; for(; i < nblks; i++){ cachedel(b->bp.addr); blkdealloc_lk(a, blks[i]); } return 0; } int syncbarrier(Arena *a, vlong gen) { if(logappend(a, gen<<8, 0, LogSync) == -1) return -1; if(a->loghd.addr == -1) a->loghd = a->logtl->bp; return 0; } /* * Allocate from an arena, with lock * held. May be called multiple times * per operation, to alloc space for * the alloc log. */ static vlong blkalloc_lk(Arena *a, int force) { Avltree *t; Arange *r; vlong b; t = a->free; r = (Arange*)t->root; if(!force && a->size - a->used <= a->reserve) return -1; if(r == nil){ fprint(2, "out of space"); abort(); } /* * A bit of sleight of hand here: * while we're changing the sorting * key, but we know it won't change * the sort order because the tree * covers disjoint ranges */ b = r->off; r->len -= Blksz; r->off += Blksz; if(r->len == 0){ avldelete(t, r); free(r); } a->used += Blksz; return b; } static int blkdealloc_lk(Arena *a, vlong b) { int r; r = -1; assert(b >= a->hd->bp.addr + Blksz); assert(b < a->tl->bp.addr); if(logappend(a, b, Blksz, LogFree) == -1) return -1; if(a->loghd.addr == -1) a->loghd = a->logtl->bp; if(freerange(a->free, b, Blksz) == -1) goto out; a->used -= Blksz; r = 0; out: return r; } int blkdealloc(vlong b) { Arena *a; int r; a = getarena(b); qlock(a); r = blkdealloc_lk(a, b); qunlock(a); return r; } static vlong blkalloc(int ty, uint hint) { Arena *a; vlong b; int tries; tries = 0; Again: a = pickarena(ty, hint, tries); if(a == nil || tries == fs->narena){ werrstr("no empty arenas"); return -1; } /* * TODO: there's an extreme edge case * here. * * If the file system has room to alloc * a data block but no log block, then * we end up with it in a stuck state. * The fix is to reserve alloc blocks, * so that we're guaranteed to be able * to log an alloc if the disk is working * correctly. */ tries++; if(canqlock(a) == 0) goto Again; if((b = blkalloc_lk(a, 0)) == -1){ qunlock(a); goto Again; } if(logappend(a, b, Blksz, LogAlloc) == -1){ qunlock(a); return -1; } if(a->loghd.addr == -1) a->loghd = a->logtl->bp; qunlock(a); return b; } static Blk* initblk(Blk *b, vlong bp, vlong gen, int ty) { Blk *ob; ob = cacheget(bp); if(ob != nil){ fprint(2, "dup block: %#p %B (alloced %#llx freed %#llx lasthold: %#llx, lastdrop: %#llx)\n", ob, ob->bp, ob->alloced, ob->freed, ob->lasthold, ob->lastdrop); abort(); } b->type = ty; b->bp.addr = bp; b->bp.hash = -1; b->bp.gen = gen; switch(ty){ case Tdat: case Tarena: b->data = b->buf; break; case Tdlist: case Tlog: b->logsz = 0; b->logp = (Bptr){-1, -1, -1}; b->data = b->buf + Loghdsz; break; case Tpivot: b->data = b->buf + Pivhdsz; break; case Tleaf: b->data = b->buf + Leafhdsz; break; } b->fnext = nil; setflag(b, Bdirty); b->nval = 0; b->valsz = 0; b->nbuf = 0; b->bufsz = 0; b->logsz = 0; b->alloced = getcallerpc(&b); return b; } Blk* newdblk(Tree *t, int ty, vlong hint) { vlong bp; Blk *b; if((bp = blkalloc(ty, hint)) == -1) return nil; if((b = cachepluck()) == nil) return nil; initblk(b, bp, t->memgen, ty); b->alloced = getcallerpc(&t); return b; } Blk* newblk(Tree *t, int ty) { return newdblk(t, ty, 0); } Blk* dupblk(Tree *t, Blk *b) { Blk *r; if((r = newblk(t, b->type)) == nil) return nil; setflag(r, Bdirty); r->bp.hash = -1; r->nval = b->nval; r->valsz = b->valsz; r->nbuf = b->nbuf; r->bufsz = b->bufsz; r->logsz = b->logsz; r->alloced = getcallerpc(&t); memcpy(r->buf, b->buf, sizeof(r->buf)); return r; } void finalize(Blk *b) { if(b->type != Tdat) PACK16(b->buf, b->type); switch(b->type){ default: abort(); break; case Tpivot: PACK16(b->buf+2, b->nval); PACK16(b->buf+4, b->valsz); PACK16(b->buf+6, b->nbuf); PACK16(b->buf+8, b->bufsz); break; case Tleaf: PACK16(b->buf+2, b->nval); PACK16(b->buf+4, b->valsz); break; case Tdlist: case Tlog: b->logh = bufhash(b->data, b->logsz); PACK16(b->buf+2, b->logsz); PACK64(b->buf+4, b->logh); packbp(b->buf+12, Ptrsz, &b->logp); break; case Tdat: case Tarena: case Tsuper: break; } b->bp.hash = blkhash(b); setflag(b, Bfinal); cacheins(b); b->cached = getcallerpc(&b); } Blk* getblk(Bptr bp, int flg) { uvlong xh, ck; char *t; Blk *b; int i; i = ihash(bp.addr) % nelem(fs->blklk); qlock(&fs->blklk[i]); if((b = cacheget(bp.addr)) != nil){ qunlock(&fs->blklk[i]); return b; } if((b = readblk(bp.addr, flg)) == nil){ qunlock(&fs->blklk[i]); return nil; } b->alloced = getcallerpc(&bp); b->bp.hash = blkhash(b); if((flg&GBnochk) == 0){ if(b->type == Tlog || b->type == Tdlist){ t = "log"; xh = b->logh; ck = bufhash(b->data, b->logsz); }else{ t = "block"; xh = bp.hash; ck = b->bp.hash; } if(ck != xh){ fprint(2, "corrupt %s %p %B: %.16llux != %.16llux\n", t, b, bp, xh, ck); qunlock(&fs->blklk[i]); return nil; } } b->bp.gen = bp.gen; cacheins(b); qunlock(&fs->blklk[i]); return b; } Blk* holdblk(Blk *b) { ainc(&b->ref); b->lasthold = getcallerpc(&b); return b; } void dropblk(Blk *b) { assert(b == nil || b->ref > 0); if(b == nil || adec(&b->ref) != 0) return; b->lastdrop = getcallerpc(&b); /* * freed blocks go to the LRU bottom * for early reuse. */ if(checkflag(b, Bfreed)) lrubot(b); else lrutop(b); } ushort blkfill(Blk *b) { switch(b->type){ case Tpivot: return 2*b->nbuf + b->bufsz + 2*b->nval + b->valsz; case Tleaf: return 2*b->nval + b->valsz; default: fprint(2, "invalid block @%lld\n", b->bp.addr); abort(); } return 0; // shut up kencc } void freeblk(Tree *t, Blk *b, Bptr bp) { Bfree *f; ulong ge; if(t == &fs->snap || (t != nil && bp.gen <= t->gen)){ killblk(t, bp); return; } if((f = malloc(sizeof(Bfree))) == nil) abort(); f->op = DFblk; f->bp = bp; if(b != nil) f->b = holdblk(b); else f->b = nil; ge = agetl(&fs->epoch); f->next = fs->limbo[ge]; fs->limbo[ge] = f; } void epochstart(int tid) { ulong ge; ge = agetl(&fs->epoch); asetl(&fs->lepoch[tid], ge | Eactive); } void epochend(int tid) { ulong le; le = agetl(&fs->lepoch[tid]); asetl(&fs->lepoch[tid], le &~ Eactive); } void epochclean(void) { ulong e, ge; Bfree *p, *n; Arena *a; Qent qe; int i; ge = agetl(&fs->epoch); for(i = 0; i < fs->nworker; i++){ e = agetl(&fs->lepoch[i]); if((e & Eactive) && e != (ge | Eactive)) return; } p = asetp(&fs->limbo[(ge+1)%3], nil); asetl(&fs->epoch, (ge+1)%3); for(; p != nil; p = n){ n = p->next; switch(p->op){ case DFtree: free(p->t); break; case DFblk: a = getarena(p->bp.addr); qe.op = Qfree; qe.bp = p->bp; qe.b = nil; qput(a->sync, qe); if(p->b != nil){ setflag(p->b, Bfreed); dropblk(p->b); } break; default: abort(); } free(p); } } void enqueue(Blk *b) { Arena *a; Qent qe; b->enqueued = getcallerpc(&b); a = getarena(b->bp.addr); assert(checkflag(b, Bdirty)); assert(b->bp.addr >= 0); holdblk(b); finalize(b); qe.op = Qwrite; qe.bp = b->bp; qe.b = b; qput(a->sync, qe); } void qinit(Syncq *q) { q->fullrz.l = &q->lk; q->emptyrz.l = &q->lk; q->nheap = 0; q->heapsz = fs->cmax; if((q->heap = malloc(q->heapsz*sizeof(Qent))) == nil) sysfatal("alloc queue: %r"); } int qcmp(Qent *a, Qent *b) { if(a->qgen != b->qgen) return (a->qgen < b->qgen) ? -1 : 1; if(a->bp.addr != b->bp.addr) return (a->bp.addr < b->bp.addr) ? -1 : 1; if(a->op != b->op){ if(a->op == Qflush) return -1; if(a->op == Qfree) return 1; } return 0; } void qput(Syncq *q, Qent qe) { int i; qlock(&q->lk); qe.qgen = agetv(&fs->qgen); while(q->nheap == q->heapsz) rsleep(&q->fullrz); for(i = q->nheap; i > 0; i = (i-1)/2){ if(qcmp(&qe, &q->heap[(i-1)/2]) == 1) break; q->heap[i] = q->heap[(i-1)/2]; } q->heap[i] = qe; q->nheap++; rwakeup(&q->emptyrz); qunlock(&q->lk); } static Qent qpop(Syncq *q) { int i, l, r, m; Qent e, t; qlock(&q->lk); while(q->nheap == 0) rsleep(&q->emptyrz); e = q->heap[0]; if(--q->nheap == 0) goto Out; i = 0; q->heap[0] = q->heap[q->nheap]; while(1){ m = i; l = 2*i+1; r = 2*i+2; if(l < q->nheap && qcmp(&q->heap[m], &q->heap[l]) == 1) m = l; if(r < q->nheap && qcmp(&q->heap[m], &q->heap[r]) == 1) m = r; if(m == i) break; t = q->heap[m]; q->heap[m] = q->heap[i]; q->heap[i] = t; i = m; } Out: rwakeup(&q->fullrz); qunlock(&q->lk); return e; } void runsync(int, void *p) { Arena *a; Syncq *q; Qent qe; q = p; while(1){ qe = qpop(q); if(qe.op == Qfree){ a = getarena(qe.bp.addr); qlock(a); cachedel(qe.bp.addr); blkdealloc_lk(a, qe.bp.addr); if(qe.b != nil) dropblk(qe.b); qunlock(a); }else if(qe.op == Qflush){ qlock(&fs->synclk); if(--fs->syncing == 0) rwakeupall(&fs->syncrz); qunlock(&fs->synclk); }else{ if(checkflag(qe.b, Bfreed) == 0) if(syncblk(qe.b) == -1){ ainc(&fs->broken); fprint(2, "write: %r\n"); abort(); } dropblk(qe.b); } } } void sync(void) { uvlong gen; Arena *a; Qent qe; int i; if(fs->rdonly) return; qlock(&fs->synclk); if(!fs->snap.dirty){ qunlock(&fs->synclk); return; } flushdlcache(0); gen = aincv(&fs->qgen, 1); fs->syncing = fs->nsyncers; for(i = 0; i < fs->nsyncers; i++){ qe.op = Qflush; qe.bp.addr = 0; qe.bp.hash = -1; qe.bp.gen = -1; qe.b = nil; qput(&fs->syncq[i], qe); } while(fs->syncing != 0) rsleep(&fs->syncrz); /* pass 0: sync arena contents */ for(i = 0; i < fs->narena; i++){ a = &fs->arenas[i]; qlock(a); syncbarrier(a, gen); finalize(a->logtl); if(syncblk(a->logtl) == -1) sysfatal("sync arena: %r"); qunlock(a); } /* * pass 1: sync block headers; if we crash here, * the block footers are consistent, and we can * use them. */ for(i = 0; i < fs->narena; i++){ a = &fs->arenas[i]; qlock(a); packarena(a->hd->data, Blksz, a); finalize(a->hd); if(syncblk(a->hd) == -1) sysfatal("sync arena: %r"); qunlock(a); } /* * pass 2: sync superblock; we have a consistent * set of block headers, so if we crash, we can * use the loaded block headers; the footers will * get synced after so that we can use them next * time around. */ for(i = 0; i < fs->narena; i++) fs->arenabp[i] = fs->arenas[i].hd->bp; packsb(fs->sb0->buf, Blksz, fs); packsb(fs->sb1->buf, Blksz, fs); finalize(fs->sb0); finalize(fs->sb1); if(syncblk(fs->sb0) == -1) sysfatal("sync sb: %r"); if(syncblk(fs->sb1) == -1) sysfatal("sync sb: %r"); fs->snap.dirty = 0; /* * pass 3: sync block footers; if we crash here, * the block headers are consistent, and we can * use them. */ for(i = 0; i < fs->narena; i++){ a = &fs->arenas[i]; qlock(a); packarena(a->tl->data, Blksz, a); finalize(a->tl); if(syncblk(a->tl) == -1) sysfatal("sync arena: %r"); qunlock(a); } /* * Pass 4: clean up the old snap tree's deadlist */ freedl(&fs->snapdl, 1); fs->snapdl.hd.addr = -1; fs->snapdl.hd.hash = -1; fs->snapdl.hd.gen = -1; fs->snapdl.tl.addr = -1; fs->snapdl.tl.hash = -1; fs->snapdl.tl.gen = -1; fs->snapdl.ins = nil; qunlock(&fs->synclk); }