Eliminate lock c->dirtylk which is the cause of fossil hanging while
taking a snapshot, as mentioned in BUGS of fossilcons(8).

Dirtylk prevents any thread from dirtying new blocks while flushThread
is writing old dirty blocks to disk.  Deadlock occurs when a thread T
tries to dirty a block B while holding a lock on an already-dirty block
A. T can't proceed until c->dirtylk is unlocked, which won't happen
until flushThread has written out block A; but block A can't be written
out until T unlocks it.

Dirtylk serves no useful purpose: whenever cacheFlush is called with
wait=1, a write lock is already held on the "epoch lock" fs->elk, which
prevents any new 9p requests from touching the file system.  Any
dirtying of blocks is limited to internal fossil activities for requests
which were already in progress when cacheFlush was called, and thus
strictly bounded even without c->dirtylk.

Reference: /n/sources/patch/applied/fossil-snap-deadlock
Date: Thu Mar 22 12:55:40 CET 2012
Signed-off-by: miller@hamnavoe.com

--- /sys/src/cmd/fossil/cache.c	Thu Mar 22 12:55:24 2012
+++ /sys/src/cmd/fossil/cache.c	Thu Mar 22 13:09:24 2012
@@ -25,7 +25,6 @@
 struct Cache
 {
 	VtLock	*lk;
-	VtLock	*dirtylk;
 	int 	ref;
 	int	mode;
 
@@ -163,7 +162,6 @@
 	nbl = nblocks * 4;
 
 	c->lk = vtLockAlloc();
-	c->dirtylk = vtLockAlloc();	/* allowed to dirty blocks */
 	c->ref = 1;
 	c->disk = disk;
 	c->z = z;
@@ -1100,14 +1098,12 @@
 		return 1;
 	assert(b->iostate == BioClean);
 
-	vtLock(c->dirtylk);
 	vtLock(c->lk);
 	b->iostate = BioDirty;
 	c->ndirty++;
 	if(c->ndirty > (c->maxdirty>>1))
 		vtWakeup(c->flush);
 	vtUnlock(c->lk);
-	vtUnlock(c->dirtylk);
 
 	return 1;
 }
@@ -2088,12 +2084,6 @@
 void
 cacheFlush(Cache *c, int wait)
 {
-	/*
-	 * Lock c->dirtylk so that more blocks aren't being dirtied
-	 * while we try to write out what's already here.
-	 * Otherwise we might not ever finish!
-	 */
-	vtLock(c->dirtylk);
 	vtLock(c->lk);
 	if(wait){
 		while(c->ndirty){
@@ -2106,7 +2096,6 @@
 	}else if(c->ndirty)
 		vtWakeup(c->flush);
 	vtUnlock(c->lk);
-	vtUnlock(c->dirtylk);
 }
 
 /*
--- /sys/man/8/fossilcons	Thu Mar 22 12:55:29 2012
+++ /sys/man/8/fossilcons	Thu Mar 22 12:55:26 2012
@@ -1205,13 +1205,3 @@
 .EX
 snap -a -s /snapshot/2003/1220/0700 -d /archive/2003/1220
 .EE
-.SH BUGS
-It is prudent to avoid taking a snapshot at the same time as an
-archival dump.
-.I Fossil
-has been seen to sometimes hang when they collide.
-Snapshots are taken when
-.BI time(0)/60% interval
-is zero, so
-an interval of 60 will take snapshots on the hour.
-It's easiest to schedule the archival dumps to happen not exactly on the hour.