ref: c947228f7cf50bbbeb3f1e1b1808280e9ae9247f
dir: /iobuf.c/
#include "all.h"
u32 nbuckets = 0; /* nbuckets derived from -m or Nbuckets */
Hiob *hiob = nil; /* array of nbuckets */
Extents frees = {0};
/*
add an Iobuf to the collisions lru linked list
hp must be locked
*/
Iobuf *
newbuf(Hiob *hp, u16 len)
{
Iobuf *p, *q;
if(len == 0){
panic("newbuf len == 0 caller %#p\n",
getcallerpc(&hp));
}
p = emalloc9p(sizeof(Iobuf));
q = hp->link;
if(q != nil){
p->fore = q;
p->back = q->back;
q->back = p;
p->back->fore = p;
}else{
hp->link = p;
p->fore = p;
p->back = p;
}
p->blkno = 0;
p->len = len;
p->xiobuf = emalloc9p(len*Rawblocksize);
return p;
}
/*
Get the Iobuf of the disk block at addr from the buffer cache
for my use.
All disk accesses go through the buffer cache. getbuf() selects
the Iobuf for our use from the buffer cache. putbuf() returns the
Iobuf back to the buffer cache.
Any Iobuf access happens only between the getbuf() and putbuf() calls.
The Iobuf's are grouped into a least-recently-used circular list
of buffers. The most recently used Iobuf is pointed to by Hiob.link.
Iobuf.fore is the next recently used buffer.
Iobuf.back is the oldest recently used buffer.
Hiob.link->back is the oldest buffer that will be reused first.
*/
Iobuf *
getbuf(u64 blkno, u16 len, u8 readonly, u8 freshalloc)
{
Hiob *hp;
Iobuf *s, *p;
s8 ncollisions;
if(len == 0){
panic("getbuf len == 0 blkno %llud readonly %d caller %#p\n",
blkno, readonly, getcallerpc(&blkno));
}
hp = &hiob[blkno%nbuckets];
if(chatty9p > 4)
dprint("getbuf blkno %llud len %d blkno%%nbuckets %llud pc 0x%p"
" hiob 0x%p hp 0x%p readonly %d\n",
blkno, len, blkno%nbuckets, getcallerpc(&blkno),
hiob, hp, readonly);
qlock(hp);
s = hp->link;
if(s == nil)
goto new;
for(p=s, ncollisions = 0;;){
ncollisions++;
if(p->blkno == blkno){
if(p != s){
/* remove p from its current position in the lru circular buffer */
p->back->fore = p->fore;
p->fore->back = p->back;
/* make p the hb->link and put it at the back of existing link */
p->fore = s;
p->back = s->back;
s->back = p;
p->back->fore = p;
hp->link = p;
}
incref(p);
qunlock(hp);
if(chatty9p > 4)
dprint(" in cache, after qunlock(hp) hp 0x%p blkno %llud\n",
hp, blkno);
if(p->len != len){
if(chatty9p > 4)
dprint("getbuf refresh used blkno %llud, size in memory is %d"
" and not %d, caller %#p\n",
blkno, p->len, len, getcallerpc(&blkno));
if(p->len == 0 || len == 0)
panic("getbuf: p->len == 0 || len == 0 p->len %d len %d",
p->len, len);
wlock(p);
if(chatty9p > 4)
dprint(" after wlock() blkno %llud\n", blkno);
if(p->dirty)
panic("p->dirty with p->len != len\n");
free(p->xiobuf);
p->xiobuf = emalloc9p(len*Rawblocksize);
p->len = len;
if(freshalloc == 0)
devread(blkno, p->xiobuf, len);
if(readonly){
if(chkwunlock(p) == 0){
showbuf(p);
panic("getbuf chkwunlock(p) == 0 called by %#p\n", getcallerpc(&blkno));
}
rlock(p);
}
}else if(readonly){
if(chatty9p > 4)
dprint(" in cache iobuf 0x%p has len %llud blkno %llud len %llud .."
" rlock()\n", p, p->len, blkno, len);
rlock(p);
}else{
wlock(p);
if(chatty9p > 4)
dprint(" after wlock() blkno %llud\n", blkno);
/* as we lock from top down, this should ensure that the
lower blocks in the hierarchy are removed from the write
queue only after the top blocks have been removed or erased.
*/
if(p->dirty)
rmwrite(p); /* remove the pending write as it will be
written by putbuf() anyway */
}
decref(p);
return p;
}
p = p->fore;
if(p == s)
break;
}
/* maxed out our allowed number of collisions,
try to steal an older Iobuf without any ref's.
Ncollisions is a soft limit.
*/
if(ncollisions >= Ncollisions){
Another:
do{
p = s->back;
if(p->ref == 0 && p->dirty == 0 && canwlock(p)){
if(p->dirty){
wunlock(p);
goto Another;
}
if(p->len != len){
free(p->xiobuf);
p->xiobuf = emalloc9p(len*Rawblocksize);
p->len = len;
}
hp->link = p;
if(chatty9p > 4)
dprint(" stealing iobuf 0x%p of %llud blocks"
" for blkno %llud len %llud\n",
p, p->len, blkno, len);
goto found; /* p is wlock() */
}
s = p;
}while(p != hp->link);
}
/* no unlocked blocks available; add a new one */
new:
if(chatty9p > 4)
dprint(" adding new Iobuf for blkno %llud\n", blkno);
p = newbuf(hp, len);
if(chatty9p > 4)
dprint(" .. wlock() blkno %llud\n", blkno);
wlock(p);
found:
p->blkno = blkno;
qunlock(hp);
if(chatty9p > 4)
dprint(" after qunlock(hp) hp 0x%p blkno %llud\n",
hp, blkno);
if(freshalloc == 0)
devread(blkno, p->xiobuf, len);
if(readonly){
if(chatty9p > 4)
dprint("new buffer: switching from wlock() to rlock() blkno %llud\n", blkno);
incref(p);
wunlock(p);
rlock(p);
decref(p);
}
return p;
}
Iobuf *
getbufchk(u64 blkno, u16 len, u8 readonly, int tag, u64 qpath)
{
Iobuf *b;
if(chatty9p > 4)
dprint("getbufchk caller pc 0x%p\n", getcallerpc(&blkno));
if(len == 0){
panic("getbufchk len == 0 blkno %llud readonly %d tag %d"
" qpath %llud caller %#p\n",
blkno, readonly, tag, qpath, getcallerpc(&blkno));
}
b = getbuf(blkno, len, readonly, Bused);
if(b != nil)
if(checktag(b, tag, qpath) == 0){
putbuf(b);
panic("checktag on %llud failed %s\n", blkno, errstring[Ephase]);
}
if(b->io == nil)
panic("b->io == nil blkno %llud readonly %d tag %d"
" qpath %llud b->blkno %llud %d caller %#p\n",
blkno, readonly, tag, qpath, b->blkno, b->len,
getcallerpc(&blkno));
return b;
}
/*
put the Iobuf of the disk block at addr back into
the buffer cache for others to use.
writes to disk if changed.
if(chatty9p > 4)
dprint("putbuf p->blkno 0x%d t->c->type %d devtab[t->c->type]->dc %c\n"
" p 0x%p p->readonly %d\n"
" p->xiobuf 0x%p",
p->blkno, t->c->type, devtab[t->c->type]->dc,
p, p->readonly, p->xiobuf);
*/
void
bkp(u64 srcbno, u8 *contents, u64 bno, u64 qpath)
{
Iobuf *buf;
if(bno == 0){
dprint("bkp %llud: invalid backup location %llud, qpath %llud\n",
srcbno, bno, qpath);
return;
}
buf = getbufchk(bno, 1, Bwritable, Tdata, qpath);
if(buf == nil){
panic("bkp: buf == nil\n");
}
memcpy(buf->io->buf, contents, Blocksize);
if(qpath == Qproot0 || qpath == Qproot1){
buf->io->d.mode &= ~DMDIR; /* to avoid recursive du -a */
}
putbuf(buf);
}
void
putbuf(Iobuf *p)
{
s8 i;
u8 buf[Blocksize];
u64 srcbno;
if(p == nil){
panic("putbuf p == nil called by %#p\n", getcallerpc(&p));
dprint("%s\n", errstring[Ephase]);
return;
}
if(p->io == nil){
showbuf(p);
panic("putbuf p->io == nil by %#p\n", getcallerpc(&p));
dprint("%s\n", errstring[Ephase]);
return;
}
if(p->len == 0){
showbuf(p);
panic("putbuf p->len == 0 by %#p\n", getcallerpc(&p));
dprint("%s\n", errstring[Ephase]);
return;
}
if(chatty9p > 4)
dprint("putbuf p->blkno %llud\n", p->blkno);
if(p->readers){
chkrunlock(p);
if(chatty9p > 4)
dprint(" .. runlock()'ed\n");
}else{
if(canwlock(p)){
panic("putbuf: buffer not locked %llud\n", p->blkno);
}
srcbno = p->blkno;
if(p->blkno == config.config.srcbno ||
p->blkno == config.super.srcbno ||
p->blkno == config.root.srcbno){
memcpy(buf, p->io->buf, Blocksize);
}
putwrite(p);
if(srcbno == config.config.srcbno){
for(i=0; i<Nbkp; i++)
bkp(srcbno, buf, config.config.dest[i], Qpconfig0+i*3);
}else if(srcbno == config.super.srcbno){
for(i=0; i<Nbkp; i++)
bkp(srcbno, buf, config.super.dest[i], Qpsuper0+i*3);
}else if(srcbno == config.root.srcbno){
for(i=0; i<Nbkp; i++)
bkp(srcbno, buf, config.root.dest[i], Qproot0+i*3);
}
}
}
/* only caller is freeblockbuf().
These blocks do not need to be written to the disk.
Hence, avoiding putwrite().
*/
void
putbuffree(Iobuf *p)
{
if(p == nil){
panic("putbuffree p == nil called by %#p\n", getcallerpc(&p));
dprint("%s\n", errstring[Ephase]);
return;
}
if(p->io == nil){
showbuf(p);
panic("putbuffree p->io == nil by %#p\n", getcallerpc(&p));
dprint("%s\n", errstring[Ephase]);
return;
}
if(p->len == 0){
showbuf(p);
panic("putbuffree p->len == 0 by %#p\n", getcallerpc(&p));
dprint("%s\n", errstring[Ephase]);
return;
}
if(chatty9p > 4)
dprint("putbuffree p->blkno %llud\n", p->blkno);
if(p->readers){
chkrunlock(p);
if(chatty9p > 4)
dprint(" .. runlock()'ed\n");
}else{
if(canwlock(p)){
panic("putbuffree: buffer not locked %llud\n", p->blkno);
}
p->dirty = 0;
if(chkwunlock(p) == 0){
showbuf(p);
panic("putbuffree chkwunlock(p) == 0 called by %#p\n", getcallerpc(&p));
}
}
}
int
checktag(Iobuf *p, u16 tag, u64 qpath)
{
uintptr pc;
if(tag != p->io->type || (qpath != Qpnone && qpath != p->io->path)){
pc = getcallerpc(&p);
dprint(" tag = %G; expected %G; blkno = %llud\n",
(uint)p->io->type, (uint)tag, p->blkno);
if(qpath == Qpnone){
dprint("checktag pc=%p disk %s(block %llud) tag/path=%s/%llud;"
" expected %s\n",
pc, devfile, (u64)p->blkno,
tagnames[p->io->type], (u64)p->io->path,
tagnames[tag]);
} else {
dprint(" tag/path = %G/%llux; expected %G/%llux\n",
(uint)p->io->type, p->io->path, tag, qpath);
dprint("checktag pc=%p disk %s(block %llud) tag/path=%s/%llud;"
" expected %s/%llud\n",
pc, devfile, (u64)p->blkno,
tagnames[p->io->type], (u64)p->io->path,
tagnames[tag], (u64)qpath);
}
return 0;
}
return 1;
}
void
settag(Iobuf *p, int tagtype, u64 qpath)
{
if(p->readers)
panic("settag %s(%llux) tag/path=%s/%llud: not Bwritable\n",
devfile, (u64)p->blkno, tagnames[tagtype], qpath);
if(p->io == nil)
panic("settag %s(%llux) tag/path=%s/%llud: p->io == nil\n",
devfile, (u64)p->blkno, tagnames[tagtype], qpath);
p->io->len = p->len;
p->io->type = tagtype;
p->io->path = qpath;
}
void *amalloc(u64 n){
void *p;
if(p = mallocalign(n, sizeof(u64), 0, 0))
memset(p, 0, n);
else
sysfatal("malloc: %r");
setmalloctag(p, getcallerpc(&n));
return p;
}
/*
* Prepare nbuckets of hash buckets. Each bucket will point to a
* linked list of collisions. The collisions are ordered into a
* least-recently-used (lru) linked list.
*/
void
iobufinit(void)
{
int i;
Hiob *hp;
while(prime(nbuckets) == 0)
nbuckets++;
if(chatty9p)
dprint("mafs: iobufinit %ud hash buckets\n", nbuckets);
hiob = amalloc(nbuckets*sizeof(Hiob));
hp = hiob;
if(chatty9p > 4)
dprint("iobufinit: hiob 0x%p\n", hiob);
for(i=0; i<nbuckets; i++){
qlock(hp);
qunlock(hp);
hp++;
}
}
void
showbuf(Iobuf *p)
{
if(p == nil){
dprint("showbuf p == nil called by %#p\n",
p, getcallerpc(&p));
return;
}
dprint("showbuf p 0x%p ref %d readers %d writer %d"
" blkno %llud len %d"
" fore 0x%p back 0x%p"
" xiobuf 0x%p"
" caller %#p\n",
p, p->ref, p->readers, p->writer,
p->blkno, p->len,
p->fore, p->back,
p->xiobuf,
getcallerpc(&p));
if(p->io != nil)
showblock((u8*)p->io);
}