Attachment 202890 Details for Bug 236480 – Fix dn->dn_type != DMU_OT_NONE in dbuf_create()

[patch] Fix dn->dn_type != DMU_OT_NONE in dbuf_create()

zfs-dnode-race (text/plain), 2.83 KB, created by ncrogers on 2019-03-15 21:20:29 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: ncrogers

Created: 2019-03-15 21:20:29 UTC

Size: 2.83 KB

patch

obsolete

>Attempt to fix a ZFS race bug that causes the fs and thereby the
>system to "hang" under an unknown and rare deadlock condition. The
>hang/deadlock behavior APPEARS to transition into a complete kernel
>panic when INVARIANTS is enabled in the kernel, but the causality
>between the below bug reports is still debatable.
>
>https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=236220
>https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=236480
>
>******************************************************************************
>Andriy Gapon  freebsd_committer 2019-03-14 23:04:59 UTC
>
>    I believe that the problem could be a race condition between dnode_hold
>    and dnode_sync_free possibly introduced in base r337669 which is the
>    large dnode feature from ZoL.
>    Perhaps, the problem has already been fixed in ZoL...
>
>    This looks like the fix that I had in mind:
>    https://github.com/behlendorf/zfs/commit/50f32ed74e42aa28522e9681fb8ae55239fa33a7
>    but note that it is from a personal branch, not ZoL.
>
>    Here is a relevant bug report:
>    https://github.com/zfsonlinux/zfs/issues/5396
>    and another one:
>    https://github.com/zfsonlinux/zfs/issues/6522
>
>    And this is what ultimately went into ZoL:
>    https://github.com/zfsonlinux/zfs/commit/4c5b89f59e4e5c8f5b4680040118ebde09598bbe
>    Much larger and much harder to understand.
>
>******************************************************************************
>https://github.com/behlendorf/zfs/commit/50f32ed74e42aa28522e9681fb8ae55239fa33a7.patch
>
>From 50f32ed74e42aa28522e9681fb8ae55239fa33a7 Mon Sep 17 00:00:00 2001
>From: Brian Behlendorf <behlendorf1@llnl.gov>
>Date: Thu, 24 Aug 2017 17:29:11 -0700
>Subject: [PATCH] Fix dn->dn_type != DMU_OT_NONE in dbuf_create()
>
>As part of commit 50c957f7 this check was pulled up before the
>call to dnode_create().  This is racy since the dnode_phys_t
>in the dbuf could be updated after the check passed but before
>it's created by dnode_create().  Close the race by adding the
>original check back to detect this unlikely case.
>
>Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
>Closes #5396
>Closes #6522
>
>******************************************************************************
>
>Index: sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dnode.c
>===================================================================
>--- sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dnode.c	(revision 345185)
>+++ sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dnode.c	(working copy)
>@@ -1324,7 +1324,9 @@
> 	mutex_enter(&dn->dn_mtx);
> 	type = dn->dn_type;
> 	if (dn->dn_free_txg ||
>-	    ((flag & DNODE_MUST_BE_FREE) && !refcount_is_zero(&dn->dn_holds))) {
>+	    ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
>+	    ((flag & DNODE_MUST_BE_FREE) &&
>+	    (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
> 		mutex_exit(&dn->dn_mtx);
> 		zrl_remove(&dnh->dnh_zrlock);
> 		dbuf_rele(db, FTAG);
>

Actions: View | Diff

Attachments on bug 236480: 202858 | 202859 | 202890