nfs: don't assume fscache_acquire_cookie will always succeed

We seem to be assuming fscache_acquire_cookie will always succeed and will obtain a valid cookie. But, under low memory conditions kmem_cache_alloc could fail and that would leave nfsi->fscache being set to NULL.

This could lead to a possibility of a page with PG_fscache flag set but cookie not set(i.e. NULL). This inconsistency could trigger a kernel BUG in nfs_fscache_release_page() and other places as we do BUG_ON(!cookie), I think.

I'm aware that there are other places where we would need such checks, but wanted to validate this patch before getting on to them.

---

 fs/nfs/client.c |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)

index ee77713..7776e8a 100644
@@ -155,6 +155,8 @@ static struct nfs_client *nfs_alloc_client(const struct nfs_client_initdata *cl_
 		clp->cl_machine_cred = cred;
 
 	nfs_fscache_get_client_cookie(clp);
+	if (!clp->fscache)
+		goto error_cleanup;
 
 	return clp;
 

Suresh Jayaraman

Did you mean __fscache_acquire_cookie() function or any other post operations after the cookie has been acquired?

Because, nfs_fscache_get_client_cookie() calls fscache_acquire_cookie() which inturn calls __fscache_acquire_cookie() that returns NULL if kmem_cache_alloc fails and hence clp->fscache gets a negative cookie in nfs_fscache_get_client_cookie.

I happened to stumble upon this code when I tried to investigate a bug that reported a kerbel BUG in 2.6.32 while running LTP testsuite.

--snip--
kernel BUG at fs/nfs/fscache.c:360!

Here's the backtrace:

#0 [ffff88001cc573e0] machine_kexec at ffffffff81021532 #1 [ffff88001cc57430] crash_kexec at ffffffff81094280 #2 [ffff88001cc57500] oops_end at ffffffff814b0290
#3 [ffff88001cc57520] do_invalid_op at ffffffff81004af4 #4 [ffff88001cc575c0] invalid_op at ffffffff81003cd5
#5 [ffff88001cc57648] nfs_fscache_release_page at ffffffffa028a2cd #6 [ffff88001cc57690] shrink_page_list at ffffffff810e525e #7 [ffff88001cc57710] isolate_lru_pages at ffffffff810e47f3 #8 [ffff88001cc577a0] shrink_inactive_list at ffffffff810e5669 #9 [ffff88001cc57930] shrink_zone at ffffffff810e62c4
#10 [ffff88001cc579d0] shrink_zones at ffffffff810e63c3 #11 [ffff88001cc57a10] do_try_to_free_pages at ffffffff810e747d #12 [ffff88001cc57a70] try_to_free_pages at ffffffff810e7832 #13 [ffff88001cc57ad0] __alloc_pages_slowpath at ffffffff810ddef3 #14 [ffff88001cc57b90] __alloc_pages_nodemask at ffffffff810de35c #15 [ffff88001cc57bb0] load_balance_newidle at ffffffff81048c69 #16 [ffff88001cc57c80] ondemand_readahead at ffffffff810e1d43 #17 [ffff88001cc57cc0] do_generic_file_read at ffffffff810d8cfe #18 [ffff88001cc57d40] generic_file_aio_read at ffffffff810d9536 #19 [ffff88001cc57dd0] do_sync_read at ffffffff81122f53 #20 [ffff88001cc57f00] vfs_read at ffffffff811237a7
#21 [ffff88001cc57f30] sys_read at ffffffff81123980
#22 [ffff88001cc57f80] system_call_fastpath at ffffffff81002f3b

    RIP: 00002b6e30d4b2b0  RSP: 00007fff30463528  RFLAGS: 00010206

--snip--

My interpretation of the BUG is:

Under low (seemingly very low) memory conditions, when we try to do a NFS read and hence ondemand_readahead(), while trying to preallocate memory, we find that get_page_from_freelist() didn't manage to find pages from the free list and hence __alloc_pages_slowpath() is being invoked.

Direct reclaim kicks in as our attempt to get free list fails even after the background reclaim. Direct reclaim attempts to shrink caches. And in shrink_page_list() we check if either page->PG_private or page->PG_private1 (i.e. PG_fscache) flag is set before calling try_to_release_page().

try_to_release_page() calls the nfs release_page method where the check for PG_private fails and nfs_fscache_release_page() gets called and the kernel BUGs when it finds nfsi->fscache is NULL.

The fact PG_private is not set in nfs_release_page implies that PG_fscache is set for the page because of check in shrink_page_list() which in turn implies it's a cache object and cookies should not be NULL. It appears that the chances of a race condition is also unlikely because we seem to be holding page lock in try_to_release_page().

This lead to the question of whether we are setting the cookie properly in all cases and the allocation failure in fscache_acquire_cookie() seemed like a potential place where it could still get set to NULL and remain as it is.

Perhaps, my hypothesis has a hole or two or I'm miss something.

Thanks,

Suresh Jayaraman

Also, while poking around the code, noticed that nfs_clear_inode() calls nfs_fscache_release_inode_cookie() which doesn't seem to check whether "fsc" mount option (i.e. NFS_FSCACHE(inode)) was used or not and goes ahead to call fscache_relinquish_cookie(). As I understand we should not be calling any fscache related functions when the client doesn't use fsc mount option (though CONFIG_NFS_FSCACHE is set), right?

Thanks,

Suresh Jayaraman

Not necessarily. It can be done that way, but if you just set the top-level cookie to NULL, fscache.h will just od the right thing all the way down. It adds a test and jump to each path.

Anyway, I can't really look at it now, I'm on holiday in NZ for a couple more weeks.

David