ref: 3878b725987e7429d21fb916672523ac81edb0ca
parent: 3482fd60b47db33b1873fe1912799360ff38c493
author: Ori Bernstein <ori@eigenstate.org>
date: Sat Sep 30 18:43:00 EDT 2023
gefs: remove deferred frees, use proper snapshots
--- a/blk.c
+++ b/blk.c
@@ -15,7 +15,7 @@
static vlong blkalloc_lk(Arena*, int);
static vlong blkalloc(int, uint);
-static int blkdealloc_lk(vlong);
+static int blkdealloc_lk(Arena*, vlong);
static Blk* initblk(Blk*, vlong, vlong, int);
int
@@ -286,7 +286,7 @@
* recursion.
*/
static int
-logappend(Arena *a, vlong off, vlong len, int op, int defer)
+logappend(Arena *a, vlong off, vlong len, int op)
{vlong o, ao;
Blk *nl, *lb;
@@ -293,7 +293,7 @@
char *p, *name;
nl = nil;
- lb = defer ? a->defertl : a->logtl;
+ lb = a->logtl;
assert((off & 0xff) == 0);
assert(op == LogAlloc || op == LogFree || op == LogSync);
if(op != LogSync){@@ -326,10 +326,7 @@
if(chainlog(lb, nl->bp.addr) == -1)
return -1;
lb = nl;
- if(defer)
- a->defertl = lb;
- else
- a->logtl = lb;
+ a->logtl = lb;
a->nlog++;
}
@@ -355,7 +352,7 @@
* current allocation. so that we don't
* reorder allocs and frees.
*/
- if(nl != nil && !defer){+ if(nl != nil){p = lb->data + lb->logsz;
ao = nl->bp.addr|LogAlloc1;
PACK64(p, ao);
@@ -599,7 +596,7 @@
}
lock(a);
cachedel(b->bp.addr);
- if(blkdealloc_lk(ba) == -1){+ if(blkdealloc_lk(a, ba) == -1){unlock(a);
return -1;
}
@@ -619,7 +616,7 @@
int
syncbarrier(Arena *a, vlong gen)
{- if(logappend(a, gen<<8, 0, LogSync, 0) == -1)
+ if(logappend(a, gen<<8, 0, LogSync) == -1)
return -1;
if(a->loghd.addr == -1)
a->loghd = a->logtl->bp;
@@ -667,14 +664,12 @@
}
static int
-blkdealloc_lk(vlong b)
+blkdealloc_lk(Arena *a, vlong b)
{- Arena *a;
int r;
r = -1;
- a = getarena(b);
- if(logappend(a, b, Blksz, LogFree, 0) == -1)
+ if(logappend(a, b, Blksz, LogFree) == -1)
return -1;
if(a->loghd.addr == -1)
a->loghd = a->logtl->bp;
@@ -687,19 +682,14 @@
}
int
-blkrelease(vlong b)
+blkdealloc(vlong b)
{Arena *a;
int r;
a = getarena(b);
- lock(a);
- r = -1;
- a = getarena(b);
- if(logappend(a, b, Blksz, LogFree, 1) == -1)
- return -1;
- if(a->deferhd.addr == -1)
- a->deferhd = a->defertl->bp;
+ lock(a);
+ r = blkdealloc_lk(a, b);
unlock(a);
return r;
}
@@ -736,7 +726,7 @@
unlock(a);
goto Again;
}
- if(logappend(a, b, Blksz, LogAlloc, 0) == -1){+ if(logappend(a, b, Blksz, LogAlloc) == -1){unlock(a);
return -1;
}
@@ -1010,6 +1000,7 @@
{ulong e, ge;
Bfree *p, *n;
+ Arena *a;
int i;
ge = agetl(&fs->epoch);
@@ -1024,9 +1015,12 @@
while(p != nil){n = p->next;
cachedel(p->bp.addr);
- blkrelease(p->bp.addr);
if(p->b != nil)
dropblk(p->b);
+ a = getarena(p->bp.addr);
+ lock(a);
+ blkdealloc_lk(a, p->bp.addr);
+ unlock(a);
free(p);
p = n;
}
@@ -1179,14 +1173,6 @@
for(i = 0; i < fs->narena; i++){a = &fs->arenas[i];
lock(a);
- if(a->deferhd.addr != -1){- finalize(a->defertl);
- chainlog(a->logtl, a->deferhd.addr);
- loadlog(a, a->deferhd);
- a->logtl = a->defertl;
- a->deferhd = (Bptr){-1, -1, -1};- a->defertl = nil;
- }
syncbarrier(a, gen);
finalize(a->logtl);
if(syncblk(a->logtl) == -1)
--- a/dat.h
+++ b/dat.h
@@ -454,6 +454,8 @@
/* superblocks */
Blk *sb0; /* primary */
Blk *sb1; /* backup */
+ long *statemap; /* for debugging */
+
/* arena allocation */
Arena *arenas;
long roundrobin;
@@ -536,8 +538,6 @@
vlong nlog; /* logged since last copression */
Bptr loghd; /* allocation log */
Blk *logtl; /* end of the log, open for writing */
- Bptr deferhd; /* allocation log */
- Blk *defertl; /* end of the log, open for writing */
Syncq *sync;
};
--- a/fns.h
+++ b/fns.h
@@ -57,7 +57,7 @@
void freebp(Tree*, Bptr);
int dlappend(Dlist *dl, Bptr);
int killblk(Tree*, Bptr);
-int blkrelease(vlong);
+int blkdealloc(vlong);
ushort blkfill(Blk*);
uvlong blkhash(Blk*);
uvlong bufhash(void*, usize);
@@ -66,8 +66,9 @@
char* updatesnap(Tree**, Tree*, char*);
char* labelsnap(Tree*, char*);
-char* delsnap(Tree*, char*);
+char* delsnap(Tree*, vlong, char*);
Tree* opensnap(char*);
+vlong successor(vlong);
void closesnap(Tree*);
void reamfs(char*);
--- a/fs.c
+++ b/fs.c
@@ -40,6 +40,7 @@
snapfs(int fd, char *old, char *new)
{Mount *mnt;
+ vlong succ;
Tree *t, *s;
char *e;
@@ -62,7 +63,8 @@
return;
}
if(strlen(new) == 0){- if((e = delsnap(t, old)) != nil){+ succ = successor(t->gen);
+ if((e = delsnap(t, succ, old)) != nil){fprint(fd, "snap: error deleting '%s': %s\n", new, e);
unlock(&fs->mountlk);
return;
--- a/load.c
+++ b/load.c
@@ -60,6 +60,7 @@
char *e;
Tree *t;
int i, k;
+ Dir *d;
if((mnt = mallocz(sizeof(*mnt), 1)) == nil)
sysfatal("malloc: %r");@@ -87,6 +88,10 @@
sysfatal("superblock: %r");if((fs->arenas = calloc(fs->narena, sizeof(Arena))) == nil)
sysfatal("malloc: %r");+ if((d = dirfstat(fs->fd)) == nil)
+ sysfatal("wut");+ if((fs->statemap = malloc((d->length/Blksz+1)*sizeof(long))) == nil)
+ sysfatal("wut"); for(i = 0; i < fs->narena; i++){a = &fs->arenas[i];
memset(a, 0, sizeof(Arena));
--- a/pack.c
+++ b/pack.c
@@ -551,10 +551,6 @@
a->used = UNPACK64(p); p += 8;
a->logtl = nil;
- a->deferhd.addr = -1;
- a->deferhd.hash = -1;
- a->deferhd.gen = -1;
- a->defertl = nil;
assert(p <= e);
return p;
}
--- a/ream.c
+++ b/ream.c
@@ -177,6 +177,8 @@
b->type = Tlog;
b->bp.addr = addr;
b->data = b->buf + Loghdsz;
+ if(b->bp.addr == 512*MiB)
+ b->bp.addr += Blksz;
setflag(b, Bdirty);
p = b->data + Loghashsz;
--- a/snap.c
+++ b/snap.c
@@ -154,7 +154,7 @@
return Efs;
if(docontents){ for(p = b->data; p != b->data+b->deadsz; p += 8){- if(blkrelease(UNPACK64(p)) == -1){+ if(blkdealloc(UNPACK64(p)) == -1){dropblk(b);
return Efs;
}
@@ -224,7 +224,7 @@
return err;
}
-static vlong
+vlong
successor(vlong gen)
{char *e, pfx[9];
@@ -254,21 +254,15 @@
}
static char*
-reclaimblocks(vlong gen, vlong prev)
+reclaimblocks(vlong gen, vlong succ, vlong prev)
{char *e, pfx[9];
- vlong succ;
Dlist dl;
Scan s;
Msg m;
- succ = successor(gen);
-
pfx[0] = Kdlist;
- if(succ == -1)
- PACK64(pfx+1, gen);
- else
- PACK64(pfx+1, succ);
+ PACK64(pfx+1, gen);
btnewscan(&s, pfx, sizeof(pfx));
if((e = btenter(&fs->snap, &s)) != nil)
return e;
@@ -309,26 +303,30 @@
* it will be merged with that successor.
*/
char*
-delsnap(Tree *t, char *name)
+delsnap(Tree *t, vlong succ, char *name)
{char buf[2][Kvmax], *e;
Msg m[2];
int nm;
- if(strcmp(name, "dump") == 0 || strcmp(name, "empty") == 0 || strcmp(name, "adm") == 0)
- return Ename;
-
nm = 0;
- m[nm].op = Odelete;
- m[nm].k = buf[nm];
- if((e = packlabel(buf[nm], sizeof(buf[nm]), name)) == nil)
- return Ename;
- m[nm].nk = e - m[nm].k;
- m[nm].v = nil;
- m[nm].nv = 0;
- nm++;
+ if(name != nil){+ if(strcmp(name, "dump") == 0
+ || strcmp(name, "empty") == 0
+ || strcmp(name, "adm") == 0)
+ return Ename;
+
+ m[nm].op = Odelete;
+ m[nm].k = buf[nm];
+ if((e = packlabel(buf[nm], sizeof(buf[nm]), name)) == nil)
+ return Ename;
+ m[nm].nk = e - m[nm].k;
+ m[nm].v = nil;
+ m[nm].nv = 0;
+ t->nlbl--;
+ nm++;
+ }
- t->nlbl--;
if(t->nlbl == 0 && t->nsucc <= 1){m[nm].op = Odelete;
m[nm].k = buf[nm];
@@ -347,7 +345,7 @@
return e;
if((e = btupsert(&fs->snap, m, nm)) != nil)
return e;
- if((e = reclaimblocks(t->gen, t->prev)) != nil)
+ if((e = reclaimblocks(t->gen, succ, t->prev)) != nil)
return e;
return nil;
}
@@ -363,7 +361,9 @@
char buf[2][Kvmax];
Msg m[2];
- if(strcmp(name, "dump") == 0 || strcmp(name, "empty") == 0 || strcmp(name, "adm") == 0)
+ if(strcmp(name, "dump") == 0
+ || strcmp(name, "empty") == 0
+ || strcmp(name, "adm") == 0)
return Ename;
t->nlbl++;
m[0].op = Oinsert;
@@ -390,15 +390,6 @@
if(!o->dirty)
return nil;
- /* this snap can be modified in place, so do that */
- if(o->nlbl == 1 && o->nsucc == 0){- m[0].op = Oinsert;
- tree2kv(o, &m[0], buf[0], sizeof(buf[0]));
- if((e = btupsert(&fs->snap, &m[0], 1)) != nil)
- return e;
- o->dirty = 0;
- return nil;
- }
/* update the old kvp */
o->nsucc++;
@@ -416,7 +407,10 @@
t->nsucc = 0;
t->ht = o->ht;
t->bp = o->bp;
- t->prev = o->gen;
+ if(o->nlbl == 0 && o->nsucc == 1)
+ t->prev = o->prev;
+ else
+ t->prev = o->gen;
t->gen = aincv(&fs->nextgen, 1);
m[1].op = Oinsert;
@@ -432,6 +426,10 @@
/* only update the dirty status after we sync */
o->dirty = 0;
+
+ /* this was the last ref to the snap */
+ if(o->nlbl == 0 && o->nsucc == 1)
+ delsnap(o, t->gen, nil);
closesnap(o);
*r = t;
return nil;