Janos Haar of First NetCenter Bt. reported numerous crashes involving the
NBD driver. With his help, this was tracked down to bogus bio vectors
which in turn was the result of a race condition between the
receive/transmit routines in the NBD driver.
The bug manifests itself like this:
CPU0 CPU1
do_nbd_request
add req to queuelist
nbd_send_request
send req head
for each bio
kmap
send
nbd_read_stat
nbd_find_request
nbd_end_request
kunmap
When CPU1 finishes nbd_end_request, the request and all its associated
bio's are freed. So when CPU0 calls kunmap whose argument is derived from
the last bio, it may crash.
Under normal circumstances, the race occurs only on the last bio. However,
if an error is encountered on the remote NBD server (such as an incorrect
magic number in the request), or if there were a bug in the server, it is
possible for the nbd_end_request to occur any time after the request's
addition to the queuelist.
The following patch fixes this problem by making sure that requests are not
added to the queuelist until after they have been completed transmission.
In order for the receiving side to be ready for responses involving
requests still being transmitted, the patch introduces the concept of the
active request.
When a response matches the current active request, its processing is
delayed until after the tranmission has come to a stop.
This has been tested by Janos and it has been successful in curing this
race condition.
From: Herbert Xu <herbert@gondor.apana.org.au>
Here is an updated patch which removes the active_req wait in
nbd_clear_queue and the associated memory barrier.
I've also clarified this in the comment.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Cc: <djani22@dynamicweb.hu>
Cc: Paul Clements <Paul.Clements@SteelEye.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
Herbert Xu