[PATCH] NFS: Use GFP_NOFS in nfs_direct_req_alloc

[PATCH] NFS: Use GFP_NOFS in nfs_direct_req_alloc

[Chuck Lever]

Don't dive into memory reclaim in the NFS direct I/O paths, otherwise
we can deadlock.

Reported by: Wengang Wang <wen.gang.wang@oracle.com>
Fix-suggested-by: Zach Brown <zach.brown@oracle.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
---

Trond-

Thoughts?

 fs/nfs/direct.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
index e4e089a..00f06b8 100644
--- a/fs/nfs/direct.c
+++ b/fs/nfs/direct.c
@@ -149,7 +149,7 @@ static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 {
 	struct nfs_direct_req *dreq;
 
-	dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
+	dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_NOFS);
 	if (!dreq)
 		return NULL;
 

Re: [PATCH] NFS: Use GFP_NOFS in nfs_direct_req_alloc

[Trond Myklebust]

On Tue, 2009-09-08 at 18:05 -0400, Chuck Lever wrote:
> Don't dive into memory reclaim in the NFS direct I/O paths, otherwise
> we can deadlock.
> 
> Reported by: Wengang Wang <wen.gang.wang@oracle.com>
> Fix-suggested-by: Zach Brown <zach.brown@oracle.com>
> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>

Wait... What??? How does an O_DIRECT read or write allocation deadlock
with memory reclaim? Both the read and the write path call
nfs_direct_req_alloc() before they pin any user pages in memory.

Trond

> ---
> 
> Trond-
> 
> Thoughts?
> 
>  fs/nfs/direct.c |    2 +-
>  1 files changed, 1 insertions(+), 1 deletions(-)
> 
> diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
> index e4e089a..00f06b8 100644
> --- a/fs/nfs/direct.c
> +++ b/fs/nfs/direct.c
> @@ -149,7 +149,7 @@ static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
>  {
>  	struct nfs_direct_req *dreq;
>  
> -	dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
> +	dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_NOFS);
>  	if (!dreq)
>  		return NULL;
>  
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH] NFS: Use GFP_NOFS in nfs_direct_req_alloc

[Chuck Lever]

On Sep 8, 2009, at 6:32 PM, Trond Myklebust wrote:
> On Tue, 2009-09-08 at 18:05 -0400, Chuck Lever wrote:
>> Don't dive into memory reclaim in the NFS direct I/O paths, otherwise
>> we can deadlock.
>>
>> Reported by: Wengang Wang <wen.gang.wang@oracle.com>
>> Fix-suggested-by: Zach Brown <zach.brown@oracle.com>
>> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
>
> Wait... What??? How does an O_DIRECT read or write allocation deadlock
> with memory reclaim? Both the read and the write path call
> nfs_direct_req_alloc() before they pin any user pages in memory.

This may be an issue only for loopback mounts where the backing device  
is an NFS O_DIRECT file.  This type of deadlock may not be able to  
happen in upstream kernels at this point.

Even so, it makes sense for this allocation to be consistent with  
similar allocations in the other NFS I/O paths.

> Trond
>
>> ---
>>
>> Trond-
>>
>> Thoughts?
>>
>> fs/nfs/direct.c |    2 +-
>> 1 files changed, 1 insertions(+), 1 deletions(-)
>>
>> diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
>> index e4e089a..00f06b8 100644
>> --- a/fs/nfs/direct.c
>> +++ b/fs/nfs/direct.c
>> @@ -149,7 +149,7 @@ static inline struct nfs_direct_req  
>> *nfs_direct_req_alloc(void)
>> {
>> 	struct nfs_direct_req *dreq;
>>
>> -	dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_KERNEL);
>> +	dreq = kmem_cache_alloc(nfs_direct_cachep, GFP_NOFS);
>> 	if (!dreq)
>> 		return NULL;
>>
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux- 
>> nfs" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>
> -- 
> Trond Myklebust
> Linux NFS client maintainer
>
> NetApp
> Trond.Myklebust@netapp.com
> www.netapp.com
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs"  
> in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

--
Chuck Lever
chuck[dot]lever[at]oracle[dot]com

Re: [PATCH] NFS: Use GFP_NOFS in nfs_direct_req_alloc

[Trond Myklebust]

On Tue, 2009-09-08 at 18:43 -0400, Chuck Lever wrote:
> On Sep 8, 2009, at 6:32 PM, Trond Myklebust wrote:
> > On Tue, 2009-09-08 at 18:05 -0400, Chuck Lever wrote:
> >> Don't dive into memory reclaim in the NFS direct I/O paths, otherwise
> >> we can deadlock.
> >>
> >> Reported by: Wengang Wang <wen.gang.wang@oracle.com>
> >> Fix-suggested-by: Zach Brown <zach.brown@oracle.com>
> >> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
> >
> > Wait... What??? How does an O_DIRECT read or write allocation deadlock
> > with memory reclaim? Both the read and the write path call
> > nfs_direct_req_alloc() before they pin any user pages in memory.
> 
> This may be an issue only for loopback mounts where the backing device  
> is an NFS O_DIRECT file.  This type of deadlock may not be able to  
> happen in upstream kernels at this point.

I don't see how that makes any difference whatsoever. If the backing
device is a non-O_DIRECT file, then you have GFP_KERNEL allocation of
the pages.

Anything that calls down into a filesystem on a read() or write() path
had better not assume that it won't block.

> Even so, it makes sense for this allocation to be consistent with  
> similar allocations in the other NFS I/O paths.

I don't buy the 'symmetry' argument. The reason for the GFP_NOFS in the
nfs_writedata_alloc() is that you have a deadlock when the VM calls
->writepages() in order to reclaim memory.
That is not the case here, and so this is not a symmetrical case.

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH] NFS: Use GFP_NOFS in nfs_direct_req_alloc

[Chuck Lever]

On Sep 8, 2009, at 7:05 PM, Trond Myklebust wrote:
> On Tue, 2009-09-08 at 18:43 -0400, Chuck Lever wrote:
>> On Sep 8, 2009, at 6:32 PM, Trond Myklebust wrote:
>>> On Tue, 2009-09-08 at 18:05 -0400, Chuck Lever wrote:
>>>> Don't dive into memory reclaim in the NFS direct I/O paths,  
>>>> otherwise
>>>> we can deadlock.
>>>>
>>>> Reported by: Wengang Wang <wen.gang.wang@oracle.com>
>>>> Fix-suggested-by: Zach Brown <zach.brown@oracle.com>
>>>> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
>>>
>>> Wait... What??? How does an O_DIRECT read or write allocation  
>>> deadlock
>>> with memory reclaim? Both the read and the write path call
>>> nfs_direct_req_alloc() before they pin any user pages in memory.
>>
>> This may be an issue only for loopback mounts where the backing  
>> device
>> is an NFS O_DIRECT file.  This type of deadlock may not be able to
>> happen in upstream kernels at this point.
>
> I don't see how that makes any difference whatsoever. If the backing
> device is a non-O_DIRECT file, then you have GFP_KERNEL allocation of
> the pages.
>
> Anything that calls down into a filesystem on a read() or write() path
> had better not assume that it won't block.

Basically we're treating an O_DIRECT file just like a block device.   
If the block I/O path blocks when a kernel file system calls in to do  
a memory reclaim, we're in dutch.

>> Even so, it makes sense for this allocation to be consistent with
>> similar allocations in the other NFS I/O paths.
>
> I don't buy the 'symmetry' argument. The reason for the GFP_NOFS in  
> the
> nfs_writedata_alloc() is that you have a deadlock when the VM calls
> ->writepages() in order to reclaim memory.
> That is not the case here, and so this is not a symmetrical case.

That is precisely the case here, in fact.  The upper file system is  
attempting to reclaim memory in the same kernel where the NFS client  
is trying to allocate with GFP_KERNEL.

>
> -- 
> Trond Myklebust
> Linux NFS client maintainer
>
> NetApp
> Trond.Myklebust@netapp.com
> www.netapp.com

--
Chuck Lever
chuck[dot]lever[at]oracle[dot]com

Re: [PATCH] NFS: Use GFP_NOFS in nfs_direct_req_alloc

[Trond Myklebust]

On Tue, 2009-09-08 at 21:01 -0400, Chuck Lever wrote:
> On Sep 8, 2009, at 7:05 PM, Trond Myklebust wrote:
> > On Tue, 2009-09-08 at 18:43 -0400, Chuck Lever wrote:
> >> On Sep 8, 2009, at 6:32 PM, Trond Myklebust wrote:
> >>> On Tue, 2009-09-08 at 18:05 -0400, Chuck Lever wrote:
> >>>> Don't dive into memory reclaim in the NFS direct I/O paths,  
> >>>> otherwise
> >>>> we can deadlock.
> >>>>
> >>>> Reported by: Wengang Wang <wen.gang.wang@oracle.com>
> >>>> Fix-suggested-by: Zach Brown <zach.brown@oracle.com>
> >>>> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
> >>>
> >>> Wait... What??? How does an O_DIRECT read or write allocation  
> >>> deadlock
> >>> with memory reclaim? Both the read and the write path call
> >>> nfs_direct_req_alloc() before they pin any user pages in memory.
> >>
> >> This may be an issue only for loopback mounts where the backing  
> >> device
> >> is an NFS O_DIRECT file.  This type of deadlock may not be able to
> >> happen in upstream kernels at this point.
> >
> > I don't see how that makes any difference whatsoever. If the backing
> > device is a non-O_DIRECT file, then you have GFP_KERNEL allocation of
> > the pages.
> >
> > Anything that calls down into a filesystem on a read() or write() path
> > had better not assume that it won't block.
> 
> Basically we're treating an O_DIRECT file just like a block device.   
> If the block I/O path blocks when a kernel file system calls in to do  
> a memory reclaim, we're in dutch.

Without a lot more changelog context that explains what you are wanting
to do, why it is relevant to NFS (and O_DIRECT in particular), and why
you can't work around it in other ways (PF_MEMALLOC comes to mind), I'm
not at all interested in applying this patch.

> >> Even so, it makes sense for this allocation to be consistent with
> >> similar allocations in the other NFS I/O paths.
> >
> > I don't buy the 'symmetry' argument. The reason for the GFP_NOFS in  
> > the
> > nfs_writedata_alloc() is that you have a deadlock when the VM calls
> > ->writepages() in order to reclaim memory.
> > That is not the case here, and so this is not a symmetrical case.
> 
> That is precisely the case here, in fact.  The upper file system is  
> attempting to reclaim memory in the same kernel where the NFS client  
> is trying to allocate with GFP_KERNEL.

That's the "upper file system"'s problem, not ours... Stacking
filesystems causes issues. Screwing over the existing users of the
underlying filesystem is not a fix for those issues...

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com

Re: [PATCH] NFS: Use GFP_NOFS in nfs_direct_req_alloc

[Chuck Lever]

On Sep 8, 2009, at 9:37 PM, Trond Myklebust wrote:
> On Tue, 2009-09-08 at 21:01 -0400, Chuck Lever wrote:
>> On Sep 8, 2009, at 7:05 PM, Trond Myklebust wrote:
>>> On Tue, 2009-09-08 at 18:43 -0400, Chuck Lever wrote:
>>>> On Sep 8, 2009, at 6:32 PM, Trond Myklebust wrote:
>>>>> On Tue, 2009-09-08 at 18:05 -0400, Chuck Lever wrote:
>>>>>> Don't dive into memory reclaim in the NFS direct I/O paths,
>>>>>> otherwise
>>>>>> we can deadlock.
>>>>>>
>>>>>> Reported by: Wengang Wang <wen.gang.wang@oracle.com>
>>>>>> Fix-suggested-by: Zach Brown <zach.brown@oracle.com>
>>>>>> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
>>>>>
>>>>> Wait... What??? How does an O_DIRECT read or write allocation
>>>>> deadlock
>>>>> with memory reclaim? Both the read and the write path call
>>>>> nfs_direct_req_alloc() before they pin any user pages in memory.
>>>>
>>>> This may be an issue only for loopback mounts where the backing
>>>> device
>>>> is an NFS O_DIRECT file.  This type of deadlock may not be able to
>>>> happen in upstream kernels at this point.
>>>
>>> I don't see how that makes any difference whatsoever. If the backing
>>> device is a non-O_DIRECT file, then you have GFP_KERNEL allocation  
>>> of
>>> the pages.
>>>
>>> Anything that calls down into a filesystem on a read() or write()  
>>> path
>>> had better not assume that it won't block.
>>
>> Basically we're treating an O_DIRECT file just like a block device.
>> If the block I/O path blocks when a kernel file system calls in to do
>> a memory reclaim, we're in dutch.
>
> Without a lot more changelog context that explains what you are  
> wanting
> to do, why it is relevant to NFS (and O_DIRECT in particular), and why
> you can't work around it in other ways (PF_MEMALLOC comes to mind),  
> I'm
> not at all interested in applying this patch.

I didn't ask you to apply it, I just asked for your thoughts.

We have a dm target that uses an NFS file as a backing device.  It  
converts bios to NFS read and write requests using direct and async I/ 
O.  It's a loopback block device with a local file system residing in  
it.

(OK, so "loopback mount" was probably not a clear way to explain what  
is going on).

>>>> Even so, it makes sense for this allocation to be consistent with
>>>> similar allocations in the other NFS I/O paths.
>>>
>>> I don't buy the 'symmetry' argument. The reason for the GFP_NOFS in
>>> the
>>> nfs_writedata_alloc() is that you have a deadlock when the VM calls
>>> ->writepages() in order to reclaim memory.
>>> That is not the case here, and so this is not a symmetrical case.
>>
>> That is precisely the case here, in fact.  The upper file system is
>> attempting to reclaim memory in the same kernel where the NFS client
>> is trying to allocate with GFP_KERNEL.
>
> That's the "upper file system"'s problem, not ours... Stacking
> filesystems causes issues. Screwing over the existing users of the
> underlying filesystem is not a fix for those issues...

How does this change "screw over" the existing users of NFS O_DIRECT?

--
Chuck Lever
chuck[dot]lever[at]oracle[dot]com

Re: [PATCH] NFS: Use GFP_NOFS in nfs_direct_req_alloc

[Trond Myklebust]

On Tue, 2009-09-08 at 22:16 -0400, Chuck Lever wrote:
> On Sep 8, 2009, at 9:37 PM, Trond Myklebust wrote:
> > On Tue, 2009-09-08 at 21:01 -0400, Chuck Lever wrote:
> >> That is precisely the case here, in fact.  The upper file system is
> >> attempting to reclaim memory in the same kernel where the NFS client
> >> is trying to allocate with GFP_KERNEL.
> >
> > That's the "upper file system"'s problem, not ours... Stacking
> > filesystems causes issues. Screwing over the existing users of the
> > underlying filesystem is not a fix for those issues...
> 
> How does this change "screw over" the existing users of NFS O_DIRECT?

If they are low on memory, and call read() or write() on an O_DIRECT
file, the kernel will fail to start the necessary memory reclaim. Given
that by far the most common users of NFS O_DIRECT these days tend to be
large databases with rather heavy memory requirements our code changes
should rather be moving in the opposite direction w.r.t. use of
GFP_NOFS.

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com