cvs commit: src/sys/ufs/ffs ffs_softdep.c softdep.h

[Doug Ambrisko]

Alexandr Kovalenko writes:
| Hello, Stephan Uphoff!
| 
| On Wed, Aug 10, 2005 at 02:09:26PM +0000, you wrote:
| 
| > ups         2005-08-10 14:09:26 UTC
| > 
| >   FreeBSD src repository
| > 
| >   Modified files:        (Branch: RELENG_6)
| >     sys/ufs/ffs          ffs_softdep.c softdep.h 
| >   Log:
| >   MFC ffs_softdep.c 1.182, softdep.h 1.18
| >   
| >   Delay freeing disk space for file system blocks until all
| >   dirty buffers are safely released. This fixes softdep
| >   problems on truncation (deletion) of files with dirty
| >   buffers.
| 
| Could this be the fix for the problem when unpacking large archives on
| soft-updates-enabled volumes? (I experience complete lockup of
| filesystem operations at some point of time during extracting files, for
| example - cd /usr/ports/editors/openoffice-1.1 && make extract)

Don't think so.  Different bug that I found:

Index: sys/kern/vfs_bio.c
===================================================================
RCS file: /usr/local/cvsroot/freebsd/src/sys/kern/vfs_bio.c,v
retrieving revision 1.493
diff -u -p -r1.493 vfs_bio.c
--- sys/kern/vfs_bio.c	3 Aug 2005 05:02:08 -0000	1.493
+++ sys/kern/vfs_bio.c	12 Aug 2005 05:35:16 -0000
@@ -1646,6 +1646,9 @@ getnewbuf(int slpflag, int slptimeo, int
 	 * async I/O rather then sync I/O.
 	 */
 
+	/* XXX DJA hack prevent fragmentation problems for now */
+	maxsize = MAXBSIZE;
+
 	atomic_add_int(&getnewbufcalls, 1);
 	atomic_subtract_int(&getnewbufrestarts, 1);
 restart:
@@ -1690,6 +1693,14 @@ restart:
 			nqindex = QUEUE_EMPTY;
 			nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
 		}
+		if (nbp == NULL && curthread->td_proc == bufdaemonproc) {
+			if (defrag == 0 && bufspace + maxsize < maxbufspace){
+				printf("buf daemon has some potential space %d\n",maxbufspace-(bufspace + maxsize));
+				nqindex = QUEUE_EMPTY;
+				nbp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]);
+			}
+		}
+
 	}
 
 	/*
@@ -1882,6 +1893,8 @@ restart:
 		mtx_lock(&nblock);
 		needsbuffer |= flags;
 		while (needsbuffer & flags) {
+			if (curthread->td_proc == bufdaemonproc)
+				panic("Buffer Daemon lockup %s\n", waitmsg);
 			if (msleep(&needsbuffer, &nblock,
 			    (PRIBIO + 4) | slpflag, waitmsg, slptimeo)) {
 				mtx_unlock(&nblock);

The problem is that the buffer pool memory space has extra space for
the buffer daemon to use in case of emergency.  It may need to use
to space to clear some space.  If all of the space is used up it
ends up waiting on itself which never completes.  So that is the easy
thing to fix.  If it is the buffer daemon then let it use the memory that
been reserved for it for emergencies.

Now the tougher problem is that later on when the buffer memory gets
fragmented and then the buffer daemon needs to go into the emergency
area it fails due to fragmentation and waits on itself again :-(
Since the buffer pool points to blocks of memory it is difficult to
defrag the space via garbage collection.  Remember we can't flush it
which the buffer daemon is trying to do since it needs space to do it.
So we've have to walk the buffer pool and find adjacent free slots
and then combine them into a bigger one or trickle things down to
make a large contigous free space again.  Since the buffer daemon
tends to do larger then average size requests things get easily fragemented.
So I just force everything to MAXBSIZE so effectively there can
be no fragmentation with the trade off less then maximum number of 
active buffer pools :-( but atleast the buffer daemon stops
waiting on itself and doesn't lock up.  You can see this via DDB and ps.
Over time it gets worse.  Most people start to reboot before that
happens.  Advanced readers can figure out the defrag problem!  It doesn't 
appear easy to me.

Yes I have some debug printf's in it to confirm it is surviving versus
locking up.

We found this problem untar'ing large files with lots of little files
with a smaller nbuf.  A good example was untar of a tar'ed up image
of FreeBSD including src, ports and all CVS metadata.

Please try this patch and let me know if it helps.  Watch for:
	buf daemon has some potential space
in dmesg etc.  I haven't tested this version much in -current but
the same basic code saves our 4.X systems.

Also I'll be gone until Mon.

Thanks,

Doug A.