Attachment 897576 Details for Bug 1099511 – Fix for bdb issue #3514381; el6/db4

[patch] Fix for bdb issue #3514381; el6/db4

0001-Fix-a-CDB-race.patch (text/plain), 2.67 KB, created by Lubomir Rintel on 2014-05-20 13:37:45 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Lubomir Rintel

Created: 2014-05-20 13:37:45 UTC

Size: 2.67 KB

patch

obsolete

>From ae71c395ccd7652eeb5c158adbd8b5217e4e7f7e Mon Sep 17 00:00:00 2001
>From: michael brey <michael.brey@oracle.com>
>Date: Tue, 20 May 2014 15:20:47 +0200
>Subject: [PATCH] Fix a CDB race
>
>Report and reproducer here:
>https://community.oracle.com/thread/3514381
>
>From: michael brey <michael.brey@oracle.com>
>To: Lubomir Rintel <lkundrak@v3.sk>
>Subject: Re: BDB crash
>Date: Tue, 13 May 2014 09:07:45 -0600 (05/13/2014 05:07:45 PM)
>Message-id: <53723541.7040203@oracle.com>
>
>  attached are patches for each release.  the 5.3.28 patch will apply on
>top of 5.3.21.
>
>thanks
>mike
>---
> env/env_failchk.c |   24 ++++++++++++++++++++++++
> 1 files changed, 24 insertions(+), 0 deletions(-)
>
>diff --git a/env/env_failchk.c b/env/env_failchk.c
>index 16d0bcb..b3f2738 100644
>--- a/env/env_failchk.c
>+++ b/env/env_failchk.c
>@@ -206,6 +206,7 @@ __env_in_api(env)
> 	REGINFO *infop;
> 	THREAD_INFO *thread;
> 	u_int32_t i;
>+	pid_t pid;
> 	int unpin, ret;
> 
> 	if ((htab = env->thr_hashtab) == NULL)
>@@ -219,6 +220,7 @@ __env_in_api(env)
> 
> 	for (i = 0; i < env->thr_nbucket; i++)
> 		SH_TAILQ_FOREACH(ip, &htab[i], dbth_links, __db_thread_info) {
>+			pid = ip->dbth_pid;
> 			if (ip->dbth_state == THREAD_SLOT_NOT_IN_USE ||
> 			    (ip->dbth_state == THREAD_OUT &&
> 			    thread->thr_count <  thread->thr_max))
>@@ -235,6 +237,28 @@ __env_in_api(env)
> 				ip->dbth_state = THREAD_SLOT_NOT_IN_USE;
> 				continue;
> 			}
>+			/*
>+			 * The above tests are not atomic, so it is possible that
>+			 * the process pointed by ip has changed during the tests.
>+			 * In particular, if the process pointed by ip when is_alive
>+			 * was executed terminated normally, a new process may reuse
>+			 * the same ip structure and change its dbth_state before the
>+			 * next two tests were performed. Therefore, we need to test
>+			 * here that all four tests above are done on the same process.
>+			 * If the process pointed by ip changed, all tests are invalid
>+			 * and can be ignored.
>+			 * Similarly, it's also possible for two processes racing to
>+			 * change the dbth_state of the same ip structure. For example,
>+			 * both process A and B reach the above test for the same
>+			 * terminated process C where C's dbth_state is THREAD_OUT.
>+			 * If A goes into the 'if' block and changes C's dbth_state to
>+			 * THREAD_SLOT_NOT_IN_USE before B checks the condition, B
>+			 * would incorrectly fail the test and run into this line.
>+			 * Therefore, we need to check C's dbth_state again and fail
>+			 * the db only if C's dbth_state is indeed THREAD_ACTIVE.
>+			 */
>+			if (ip->dbth_state != THREAD_ACTIVE || ip->dbth_pid != pid)
>+				continue;
> 			return (__db_failed(env,
> 			     "Thread died in Berkeley DB library",
> 			     ip->dbth_pid, ip->dbth_tid));
>-- 
>1.7.1
>

Actions: View | Diff

Attachments on bug 1099511: 897576