revert-must-not-fail-allocation-loops-back-to-GFP_NOFAIL

   1 jbd2: revert must-not-fail allocation loops back to GFP_NOFAIL
   2
   3 From: Michal Hocko <mhocko@suse.cz>
   4
   5 This basically reverts 47def82672b3 (jbd2: Remove __GFP_NOFAIL from jbd2
   6 layer). The deprecation of __GFP_NOFAIL was a bad choice because it led
   7 to open coding the endless loop around the allocator rather than
   8 removing the dependency on the non failing allocation. So the
   9 deprecation was a clear failure and the reality tells us that
  10 __GFP_NOFAIL is not even close to go away.
  11
  12 It is still true that __GFP_NOFAIL allocations are generally discouraged
  13 and new uses should be evaluated and an alternative (pre-allocations or
  14 reservations) should be considered but it doesn't make any sense to lie
  15 the allocator about the requirements. Allocator can take steps to help
  16 making a progress if it knows the requirements.
  17
  18 Signed-off-by: Michal Hocko <mhocko@suse.cz>
  19 Signed-off-by: Theodore Ts'o <tytso@mit.edu>
  20 Acked-by: David Rientjes <rientjes@google.com>
  21 ---
  22
  23 Hi,
  24 this has been posted few months ago
  25 (http://marc.info/?l=linux-mm&m=142530454419654&w=2) but it hasn't
  26 gone anywhere so I am reposting. I've just rebased it on top of the
  27 ext4/for-linus tree. It wasn't clear to me which branch should I use so
  28 I've just picked this one as it was one of the most recently updated.
  29
  30  fs/jbd2/journal.c     | 11 +----------
  31  fs/jbd2/transaction.c | 20 +++++++-------------
  32  2 files changed, 8 insertions(+), 23 deletions(-)
  33
  34 diff --git a/fs/jbd2/journal.c b/fs/jbd2/journal.c
  35 index b96bd8076b70..0bc333b4a594 100644
  36 --- a/fs/jbd2/journal.c
  37 +++ b/fs/jbd2/journal.c
  38 @@ -371,16 +371,7 @@ int jbd2_journal_write_metadata_buffer(transaction_t *transaction,
  39          */
  40         J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));
  41
  42 -retry_alloc:
  43 -       new_bh = alloc_buffer_head(GFP_NOFS);
  44 -       if (!new_bh) {
  45 -               /*
  46 -                * Failure is not an option, but __GFP_NOFAIL is going
  47 -                * away; so we retry ourselves here.
  48 -                */
  49 -               congestion_wait(BLK_RW_ASYNC, HZ/50);
  50 -               goto retry_alloc;
  51 -       }
  52 +       new_bh = alloc_buffer_head(GFP_NOFS|__GFP_NOFAIL);
  53
  54         /* keep subsequent assertions sane */
  55         atomic_set(&new_bh->b_count, 1);
  56 diff --git a/fs/jbd2/transaction.c b/fs/jbd2/transaction.c
  57 index ff2f2e6ad311..799242cecffb 100644
  58 --- a/fs/jbd2/transaction.c
  59 +++ b/fs/jbd2/transaction.c
  60 @@ -278,22 +278,16 @@ static int start_this_handle(journal_t *journal, handle_t *handle,
  61
  62  alloc_transaction:
  63         if (!journal->j_running_transaction) {
  64 +               /*
  65 +                * If __GFP_FS is not present, then we may be being called from
  66 +                * inside the fs writeback layer, so we MUST NOT fail.
  67 +                */
  68 +               if ((gfp_mask & __GFP_FS) == 0)
  69 +                       gfp_mask |= __GFP_NOFAIL;
  70                 new_transaction = kmem_cache_zalloc(transaction_cache,
  71                                                     gfp_mask);
  72 -               if (!new_transaction) {
  73 -                       /*
  74 -                        * If __GFP_FS is not present, then we may be
  75 -                        * being called from inside the fs writeback
  76 -                        * layer, so we MUST NOT fail.  Since
  77 -                        * __GFP_NOFAIL is going away, we will arrange
  78 -                        * to retry the allocation ourselves.
  79 -                        */
  80 -                       if ((gfp_mask & __GFP_FS) == 0) {
  81 -                               congestion_wait(BLK_RW_ASYNC, HZ/50);
  82 -                               goto alloc_transaction;
  83 -                       }
  84 +               if (!new_transaction)
  85                         return -ENOMEM;
  86 -               }
  87         }
  88
  89         jbd_debug(3, "New handle %p going live.\n", handle);
  90 --
  91 2.1.4
  92
  93 --
  94 To unsubscribe, send a message with 'unsubscribe linux-mm' in
  95 the body to majordomo@kvack.org.  For more info on Linux MM,
  96 see: http://www.linux-mm.org/ .
  97 Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
  98