Re: [PATCH] ipc/mqueue: use correct gfp flags in msg_insert
From: Doug Ledford
Date: Mon May 14 2012 - 22:46:19 EST
On 5/14/2012 7:54 PM, Andrew Morton wrote:
> On Mon, 14 May 2012 23:05:25 +0200
> Sasha Levin <levinsasha928@xxxxxxxxx> wrote:
>> msg_insert() tries to allocate using GFP_KERNEL, while in both cases when it's called,
>> it's coming from an atomic context. Introduced by 7dd7edf ("ipc/mqueue: improve
>> performance of send/recv").
>> Use GFP_ATOMIC instead.
>> Also, fix up coding style in the kzalloc while we're there.
>> Signed-off-by: Sasha Levin <levinsasha928@xxxxxxxxx>
>> ipc/mqueue.c | 2 +-
>> 1 files changed, 1 insertions(+), 1 deletions(-)
>> diff --git a/ipc/mqueue.c b/ipc/mqueue.c
>> index 30f6f8f..9ec6896 100644
>> --- a/ipc/mqueue.c
>> +++ b/ipc/mqueue.c
>> @@ -133,7 +133,7 @@ static int msg_insert(struct msg_msg *msg, struct mqueue_inode_info *info)
>> p = &(*p)->rb_right;
>> - leaf = kzalloc(sizeof(struct posix_msg_tree_node), GFP_KERNEL);
>> + leaf = kzalloc(sizeof(*leaf), GFP_ATOMIC);
>> if (!leaf)
>> return -ENOMEM;
> hm, that should have spewed warnings everywhere the first time anyone
> tested it. Doug, is a re-read of Documentation/SubmitChecklist needed?
Re-read? I never it read it a first time, so hard for me to re-read it.
But thanks for pointing it out. Now I've read it.
> Switching to GFP_ATOMIC is a bit regrettable. Can we avoid this by
> speculatively allocating the memory before taking the lock, then free
> it again if we ended up not using it?
Not really, we take the lock in a different function than this and would
have to pass around a node struct and then free it if we didn't use it.
I mean, it could be done, but it would fugly the calls around this up.
The msg_insert() routine is called in two places. In one place, the
lock is taken right there so you could allocate before and then call.
In the other, it is another function called with the lock held so now
you would have to pass the possible mem allocation around two functions.
Doable, but ugly. On the other hand, this is a small struct that
should be coming off one of the small size kmem cache pools (4 pointers
total, a long, and an int, so kmalloc-32 or kmalloc-64 depending on
arch). That doesn't seem like a likely candidate to fail if there is
memory pressure, especially considering that immediately prior to taking
the lock we call kmalloc with GFP_KERNEL (as part of load_msg()) and so
we should either not be under serious memory pressure or we would have
slept waiting for it to ease up.
I think I can imagine a better way to do this though as part of the
whole request to cache at least one rbnode entry so we get the 0 message
performance of the queue back. I'll send that patch through once I've
verified it does what I think it will.
Doug Ledford <dledford@xxxxxxxxxx>
GPG KeyID: 0E572FDD
Infiniband specific RPMs available at
Description: OpenPGP digital signature