ref: 40d71baf7f44b17fcf323ebed8b835d86f340e84
parent: d7378c10d846ce21ef43271ae4daa8a5e5250d46
author: cinap_lenrek <cinap_lenrek@felloff.net>
date: Fri Feb 21 00:23:21 EST 2014
exportfs: bring back the changes, bug was due to rendezvous group being shared by listen exportfs used pid of slave proc as rendezvous tag. when we changed it to use Proc* memory address, there where tag collisions because listen didnt fork the rendezvous group (bug!). for now, just do rfork(RFREND) in main just in case. will fix aux/listen in a follow up changeset. -- cinap
--- a/sys/src/cmd/exportfs/exportfs.c
+++ b/sys/src/cmd/exportfs/exportfs.c
@@ -230,7 +230,7 @@
DEBUG(DFD, "exportfs: started\n");
- rfork(RFNOTEG);
+ rfork(RFNOTEG|RFREND);
if(messagesize == 0){
messagesize = iounit(netfd);
@@ -460,8 +460,8 @@
fatal(Enomem);
n = convS2M(t, data, messagesize);
if(write(netfd, data, n)!=n){
- syslog(0, "exportfs", "short write: %r");
- fatal("mount write");
+ /* not fatal, might have got a note due to flush */
+ fprint(2, "exportfs: short write in reply: %r\n");
}
free(data);
}
@@ -570,8 +570,6 @@
unlock(&sbufalloc);
w = emallocz(sizeof(*w) + messagesize);
}
- w->pid = 0;
- w->canint = 0;
w->flushtag = NOTAG;
return w;
}
@@ -579,8 +577,6 @@
void
putsbuf(Fsrpc *w)
{
- w->pid = 0;
- w->canint = 0;
w->flushtag = NOTAG;
lock(&sbufalloc);
w->next = sbufalloc.free;
--- a/sys/src/cmd/exportfs/exportfs.h
+++ b/sys/src/cmd/exportfs/exportfs.h
@@ -15,8 +15,6 @@
struct Fsrpc
{
Fsrpc *next; /* freelist */
- uintptr pid; /* Pid of slave process executing the rpc */
- int canint; /* Interrupt gate */
int flushtag; /* Tag on which to reply to flush */
Fcall work; /* Plan 9 incoming Fcall */
uchar buf[]; /* Data buffer */
@@ -53,9 +51,10 @@
struct Proc
{
- uintptr pid;
+ Lock;
Fsrpc *busy;
Proc *next;
+ int pid;
};
struct Qidtab
@@ -70,7 +69,6 @@
enum
{
- MAXPROC = 50,
FHASHSIZE = 64,
Fidchunk = 1000,
Npsmpt = 32,
@@ -128,7 +126,7 @@
void slaveopen(Fsrpc*);
void slaveread(Fsrpc*);
void slavewrite(Fsrpc*);
-void blockingslave(void);
+void blockingslave(Proc*);
void reopen(Fid *f);
void noteproc(int, char*);
void flushaction(void*, char*);
--- a/sys/src/cmd/exportfs/exportsrv.c
+++ b/sys/src/cmd/exportfs/exportsrv.c
@@ -64,14 +64,20 @@
for(m = Proclist; m; m = m->next){
w = m->busy;
- if(w != 0 && w->pid == m->pid && w->work.tag == t->work.oldtag) {
+ if(w == nil || w->work.tag != t->work.oldtag)
+ continue;
+
+ lock(m);
+ w = m->busy;
+ if(w != nil && w->work.tag == t->work.oldtag) {
w->flushtag = t->work.tag;
DEBUG(DFD, "\tset flushtag %d\n", t->work.tag);
- if(w->canint)
- postnote(PNPROC, w->pid, "flush");
+ postnote(PNPROC, m->pid, "flush");
+ unlock(m);
putsbuf(t);
return;
}
+ unlock(m);
}
reply(&t->work, &rhdr, 0);
@@ -459,10 +465,10 @@
void
slave(Fsrpc *f)
{
- Proc *p;
- uintptr pid;
- Fcall rhdr;
static int nproc;
+ Proc *m, **l;
+ Fcall rhdr;
+ int pid;
if(readonly){
switch(f->work.type){
@@ -479,30 +485,41 @@
}
}
for(;;) {
- for(p = Proclist; p; p = p->next) {
- if(p->busy == 0) {
- f->pid = p->pid;
- p->busy = f;
- do {
- pid = (uintptr)rendezvous((void*)p->pid, f);
- }
- while(pid == ~0); /* Interrupted */
- if(pid != p->pid)
- fatal("rendezvous sync fail");
- return;
- }
+ for(l = &Proclist; (m = *l) != nil; l = &m->next) {
+ if(m->busy != nil)
+ continue;
+
+ m->busy = f;
+ while(rendezvous(m, f) == (void*)~0)
+ ;
+
+ /* swept a slave proc */
+ if(f == nil){
+ *l = m->next;
+ free(m);
+ nproc--;
+ break;
+ }
+ f = nil;
+
+ /*
+ * as long as the number of slave procs
+ * is small, dont bother sweeping.
+ */
+ if(nproc < 16)
+ break;
}
+ if(f == nil)
+ return;
- if(nproc >= MAXPROC){
+ m = emallocz(sizeof(Proc));
+ pid = rfork(RFPROC|RFMEM|RFNOWAIT);
+ switch(pid) {
+ case -1:
reply(&f->work, &rhdr, Enoprocs);
putsbuf(f);
+ free(m);
return;
- }
- nproc++;
- pid = rfork(RFPROC|RFMEM);
- switch(pid) {
- case -1:
- fatal("rfork");
case 0:
if (local[0] != '\0')
@@ -511,44 +528,34 @@
local, remote);
else
procsetname("%s -> %s", local, remote);
- blockingslave();
- fatal("slave");
+ blockingslave(m);
+ _exits(0);
default:
- p = emallocz(sizeof(Proc));
- p->busy = 0;
- p->pid = pid;
- p->next = Proclist;
- Proclist = p;
- while(rendezvous((void*)pid, p) == (void*)~0)
- ;
+ m->pid = pid;
+ m->next = Proclist;
+ Proclist = m;
+ nproc++;
}
}
}
void
-blockingslave(void)
+blockingslave(Proc *m)
{
Fsrpc *p;
Fcall rhdr;
- Proc *m;
- uintptr pid;
notify(flushaction);
- pid = getpid();
-
- do {
- m = rendezvous((void*)pid, 0);
- }
- while(m == (void*)~0); /* Interrupted */
-
for(;;) {
- p = rendezvous((void*)pid, (void*)pid);
- if(p == (void*)~0) /* Interrupted */
+ p = rendezvous(m, nil);
+ if(p == (void*)~0) /* Interrupted */
continue;
+ if(p == nil) /* Swept */
+ break;
- DEBUG(DFD, "\tslave: %p %F p %p\n", pid, &p->work, p->pid);
+ DEBUG(DFD, "\tslave: %d %F\n", m->pid, &p->work);
if(p->flushtag != NOTAG)
goto flushme;
@@ -568,13 +575,17 @@
default:
reply(&p->work, &rhdr, "exportfs: slave type error");
}
- if(p->flushtag != NOTAG) {
flushme:
+ lock(m);
+ m->busy = nil;
+ unlock(m);
+
+ /* no more flushes can come in now */
+ if(p->flushtag != NOTAG) {
p->work.type = Tflush;
p->work.tag = p->flushtag;
reply(&p->work, &rhdr, 0);
}
- m->busy = 0;
putsbuf(p);
}
}
@@ -654,16 +665,8 @@
path = makepath(f->f, "");
DEBUG(DFD, "\topen: %s %d\n", path, work->mode);
-
- p->canint = 1;
- if(p->flushtag != NOTAG){
- free(path);
- return;
- }
- /* There is a race here I ignore because there are no locks */
f->fid = open(path, work->mode);
free(path);
- p->canint = 0;
if(f->fid < 0 || (d = dirfstat(f->fid)) == nil) {
Error:
errstr(err, sizeof err);
@@ -703,9 +706,6 @@
}
n = (work->count > messagesize-IOHDRSZ) ? messagesize-IOHDRSZ : work->count;
- p->canint = 1;
- if(p->flushtag != NOTAG)
- return;
data = malloc(n);
if(data == 0) {
reply(work, &rhdr, Enomem);
@@ -717,7 +717,6 @@
r = preaddir(f, (uchar*)data, n, work->offset);
else
r = pread(f->fid, data, n, work->offset);
- p->canint = 0;
if(r < 0) {
free(data);
errstr(err, sizeof err);
@@ -724,7 +723,6 @@
reply(work, &rhdr, err);
return;
}
-
DEBUG(DFD, "\tread: fd=%d %d bytes\n", f->fid, r);
rhdr.data = data;
@@ -750,11 +748,7 @@
}
n = (work->count > messagesize-IOHDRSZ) ? messagesize-IOHDRSZ : work->count;
- p->canint = 1;
- if(p->flushtag != NOTAG)
- return;
n = pwrite(f->fid, work->data, n, work->offset);
- p->canint = 0;
if(n < 0) {
errstr(err, sizeof err);
reply(work, &rhdr, err);