>The problem is following:
>
>FS driver recieves VOP_FSYNC request,
>then it scan the queue and pick up dirty buffers to
>bwrite(bp) them. bwrite calls VOP_STRATEGY,
>FS's strategy routine is trying to VOP_BMAP buf, and
>fails (suppose it fails), then xxfs_strategy do:
>
> bp->b_error = error;
> bp->b_flags |= B_ERROR;
> biodone(bp);
> return(bp);
>
>(Looks right, at least IMO)
>But:
>biodone will not rel*se(bp), as it is not B_ASYNC,
>then it returns to bwrite, where it will brelse(bp) after
>biowait(bp). Then bp will come to brelse with B_ERROR set.
>in the begining, brelse:
>
> ...
>
> if ((bp->b_flags & (B_READ | B_ERROR)) == B_ERROR) {
> bp->b_flags &= ~B_ERROR;
> bdirty(bp);
> } ...
>
>then buffer returns to dirty queue, and is dirty, all
>repeates infinite.
Also, for async writes, the file may be closed before the write completes
(the write never completes...). Then vinvalbuf() calls VOP_FSYNC()
to write all dirty buffers and panics when VOP_FSYNC() fails to do this.
>How this is solved?
This isn't solved. It was less serious before rev.1.196 of vfs_bio.c
when B_ERROR buffers were discarded insead of re-dirtied in the above
code fragment. See also PR 11697, and about 20 PRs reporting problems
with i/o errors and EOF "errors" (ENOSPC/EINVAL) for (mis)using buffered
devices (especially fd0).
I use the following variant of the patch in PR 11697:
diff -c2 vfs_bio.c~ vfs_bio.c
*** vfs_bio.c~ Thu May 13 14:42:10 1999
--- vfs_bio.c Thu May 13 16:36:09 1999
***************
*** 744,752 ****
bp->b_flags &= ~B_ERROR;
! if ((bp->b_flags & (B_READ | B_ERROR)) == B_ERROR) {
/*
* Failed write, redirty. Must clear B_ERROR to prevent
! * pages from being scrapped. Note: B_INVAL is ignored
! * here but will presumably be dealt with later.
*/
bp->b_flags &= ~B_ERROR;
--- 800,808 ----
bp->b_flags &= ~B_ERROR;
! if ((bp->b_flags & (B_READ | B_ERROR | B_INVAL)) == B_ERROR &&
! (bp->b_error == EIO || bp->b_error == 0)) {
/*
* Failed write, redirty. Must clear B_ERROR to prevent
! * pages from being scrapped.
*/
bp->b_flags &= ~B_ERROR;
This fixes the primary problem in all cases except the most interesting one:
for real i/o errors. It doesn't touch the secondary problem that most
VOP_FSYNC() routines don't try hard enough to write all dirty buffers in
the MNT_WAIT case, so vinvalbuf() may panic.
Bruce
To Unsubscribe: send mail to majord...@freebsd.org
with "unsubscribe freebsd-current" in the body of the message
