[PATCH] nfsd: more robust allocation failure handling in nfsd_reply_cache_init

[PATCH] nfsd: more robust allocation failure handling in nfsd_reply_cache_init

[Jeff Layton]

Currently, we try to allocate the cache as a single, large chunk, which
can fail if no big chunks of memory are available. We _do_ try to size
it according to the amount of memory in the box, but if the server is
started well after boot time, then the allocation can fail due to memory
fragmentation.

Try to handle this more gracefully by cutting the max_drc_entries in
half and then retrying if the allocation fails. Only give up if the
failed allocation is smaller than a page.

Reported-by: Olaf Hering <olaf@aepfle.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Jeff Layton <jlayton@redhat.com>
---
 fs/nfsd/nfscache.c | 27 +++++++++++++++++++++++----
 1 file changed, 23 insertions(+), 4 deletions(-)

While this would be good to get in, I don't see any particular urgency
here. This seems like it'd be reasonable for v4.9.

diff --git a/fs/nfsd/nfscache.c b/fs/nfsd/nfscache.c
index 54cde9a5864e..b8aaa7a71412 100644
--- a/fs/nfsd/nfscache.c
+++ b/fs/nfsd/nfscache.c
@@ -155,14 +155,12 @@ nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
 
 int nfsd_reply_cache_init(void)
 {
-	unsigned int hashsize;
+	unsigned int hashsize, target_hashsize;
 	unsigned int i;
 	int status = 0;
 
 	max_drc_entries = nfsd_cache_size_limit();
 	atomic_set(&num_drc_entries, 0);
-	hashsize = nfsd_hashsize(max_drc_entries);
-	maskbits = ilog2(hashsize);
 
 	status = register_shrinker(&nfsd_reply_cache_shrinker);
 	if (status)
@@ -173,9 +171,30 @@ int nfsd_reply_cache_init(void)
 	if (!drc_slab)
 		goto out_nomem;
 
-	drc_hashtbl = kcalloc(hashsize, sizeof(*drc_hashtbl), GFP_KERNEL);
+	/*
+	 * Attempt to allocate the hashtable, and progressively shrink the
+	 * size as the allocations fail. If the allocation size ends up being
+	 * smaller than a page however, then just give up.
+	 */
+	target_hashsize = nfsd_hashsize(max_drc_entries);
+	hashsize = target_hashsize;
+	do {
+		maskbits = ilog2(hashsize);
+		drc_hashtbl = kcalloc(hashsize, sizeof(*drc_hashtbl),
+					GFP_KERNEL|__GFP_NOWARN);
+		if (drc_hashtbl)
+			break;
+		max_drc_entries /= 2;
+		hashsize = nfsd_hashsize(max_drc_entries);
+	} while ((hashsize * sizeof(*drc_hashtbl)) >= PAGE_SIZE);
+
 	if (!drc_hashtbl)
 		goto out_nomem;
+
+	if (hashsize != target_hashsize)
+		pr_warn("NFSD: had to shrink reply cache hashtable (wanted %u, got %u)\n",
+			target_hashsize, hashsize);
+
 	for (i = 0; i < hashsize; i++) {
 		INIT_LIST_HEAD(&drc_hashtbl[i].lru_head);
 		spin_lock_init(&drc_hashtbl[i].cache_lock);
-- 
2.7.4

Re: [PATCH] nfsd: more robust allocation failure handling in nfsd_reply_cache_init

[Linus Torvalds]

On Tue, Aug 30, 2016 at 4:48 AM, Jeff Layton <jlayton@redhat.com> wrote:
>
> While this would be good to get in, I don't see any particular urgency
> here. This seems like it'd be reasonable for v4.9.

Agreed, looks ok to me. It certainly does not look like a new
regression or like a serious problem issue in practice. So 4.9 sounds
appropriate.

            Linus

Re: [PATCH] nfsd: more robust allocation failure handling in nfsd_reply_cache_init

[Bruce Fields]

On Tue, Aug 30, 2016 at 11:23:36AM -0700, Linus Torvalds wrote:
> On Tue, Aug 30, 2016 at 4:48 AM, Jeff Layton <jlayton@redhat.com> wrote:
> >
> > While this would be good to get in, I don't see any particular urgency
> > here. This seems like it'd be reasonable for v4.9.
> 
> Agreed, looks ok to me. It certainly does not look like a new
> regression or like a serious problem issue in practice. So 4.9 sounds
> appropriate.

Gah, Jeff points out I forgot to merge this.

Jeff was also wondering whether we could instead just allocate this with
vmalloc--is there any drawback?  We only allocate this on nfsd startup,
so if the only drawback is the allocation itself being expensive then
that's no big deal.

--b.

Re: [PATCH] nfsd: more robust allocation failure handling in nfsd_reply_cache_init

[Linus Torvalds]

On Thu, Oct 20, 2016 at 8:52 AM, Bruce Fields <bfields@fieldses.org> wrote:
>
> Jeff was also wondering whether we could instead just allocate this with
> vmalloc--is there any drawback?  We only allocate this on nfsd startup,
> so if the only drawback is the allocation itself being expensive then
> that's no big deal.

vmalloc is ok. Generally if it's *usually* a small allocation, the
best pattern tends to be to first try to kmalloc (of get_free_pages())
using __GFP_NORETRY | __GFP_NOWARN, and then fall back on vmalloc().
That way you don't end up doing vmalloc's for things that really don't
need it.

If you do that, we have a "kvfree()" helper that is "free either
kmalloc or vmalloc area", so you don't have to track after-the-fact
which one you did.

               Linus