: buddy allocator: ext3 failed to alloc with __GFP_NOFAIL

Hello Andrew,

we had the following ext3 problems once during stress testing (on 2.6.8):
-----------------------------------------------------------------
journal_get_undo_access: No memory for committed data
ext3_free_blocks: aborting transaction: Out of memory in __ext3_journal_get_undo_access<2>EXT3-fs error (device hda7) in ext3_free_blocks: Out of memory Aborting journal on device hda7.
EXT3-fs error (device hda7) in ext3_ordered_commit_write: IO failure
ext3_abort called.
EXT3-fs abort (device hda7): ext3_journal_start: Detected aborted journal
Remounting filesystem read-only
ext3_free_blocks: aborting transaction: Journal has aborted in __ext3_journal_get_undo_access<2>EXT3-fs error (device hda7) in ext3_free_blocks: Journal has aborted
ext3_reserve_inode_write: aborting transaction: Journal has aborted in __ext3_journal_get_write_access<2>EXT3-fs error (device hda7) in ext3_reserve_inode_write: Journal has aborted EXT3-fs error (device hda7) in ext3_truncate: Out of memory
ext3_reserve_inode_write: aborting transaction: Journal has aborted in __ext3_journal_get_write_access<2>EXT3-fs error (device hda7) in ext3_reserve_inode_write: Journal has aborted EXT3-fs error (device hda7) in ext3_orphan_del: Journal has aborted
ext3_reserve_inode_write: aborting transaction: Journal has aborted in __ext3_journal_get_write_access<2>EXT3-fs error (device hda7) in ext3_reserve_inode_write: Journal has aborted EXT3-fs error (device hda7) in ext3_delete_inode: Out of memory
__journal_remove_journal_head: freeing b_committed_data
..................
------------------------------------------------------------------

As it is seen from the messages journal_get_undo_access() failed to allocate some memory with jbd_kmalloc(), which in turn called kmalloc() with __GFP_NOFAIL flag.

How could it happen?
The only possible reason for this we suppose is a piece of code in __alloc_pages():

if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
        /* go through the zonelist yet again, ignoring mins */
        for (i = 0; zones[i] != NULL; i++) {
                struct zone *z = zones[i];

                page = buffered_rmqueue(z, order, gfp_mask);
                if (page) {
                        zone_statistics(zonelist, z);
                        goto got_pg;
                }
        }
        goto nopage;				<<<< HERE!!! FAIL...
}

So kswapd (which has PF_MEMALLOC flag) can fail to allocate memory even when it allocates it with __GFP_NOFAIL flag.

The attached patch should fix the problem (against 2.6.14).

Kirill

@@ -867,6 +867,7 @@ zone_reclaim_retry:
 
 	/* This allocation should allow future memory freeing. */
 
+rebalance:
 	if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
 			&& !in_interrupt()) {
 		if (!(gfp_mask & __GFP_NOMEMALLOC)) {

@@ -879,14 +880,13 @@ zone_reclaim_retry:
 					goto got_pg;
 			}
 		}
-		goto nopage;
+		goto empty;
 	}
 
 	/* Atomic allocations - we can't balance anything */
 	if (!wait)
 		goto nopage;
 
-rebalance:
 	cond_resched();
 
 	/* We now go into synchronous reclaim */

@@ -946,6 +946,7 @@ rebalance:
 	 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
 	 * <= 3, but that may not be true in other implementations.
 	 */
+empty:
 	do_retry = 0;
 	if (!(gfp_mask & __GFP_NORETRY)) {
 		if ((order <= 3) || (gfp_mask & __GFP_REPEAT))

Hello,

as Andrey Savochkin pointed to me, the previous patch is incorrect since makes allocations with PF_MEMALLOC to be always success for order <= 3 which is not what we usually want. I remade the patch to be more explicit about the issue - now it retries allocation _only_ if __GFP_NOFAIL is set.

------------- original comment below ------------------

we had the following ext3 problems once during stress testing (on 2.6.8):
-----------------------------------------------------------------
journal_get_undo_access: No memory for committed data
ext3_free_blocks: aborting transaction: Out of memory in __ext3_journal_get_undo_access<2>EXT3-fs error (device hda7) in ext3_free_blocks: Out of memory Aborting journal on device hda7.
EXT3-fs error (device hda7) in ext3_ordered_commit_write: IO failure
ext3_abort called.
EXT3-fs abort (device hda7): ext3_journal_start: Detected aborted journal
Remounting filesystem read-only
ext3_free_blocks: aborting transaction: Journal has aborted in __ext3_journal_get_undo_access<2>EXT3-fs error (device hda7) in ext3_free_blocks: Journal has aborted
ext3_reserve_inode_write: aborting transaction: Journal has aborted in __ext3_journal_get_write_access<2>EXT3-fs error (device hda7) in ext3_reserve_inode_write: Journal has aborted EXT3-fs error (device hda7) in ext3_truncate: Out of memory
ext3_reserve_inode_write: aborting transaction: Journal has aborted in __ext3_journal_get_write_access<2>EXT3-fs error (device hda7) in ext3_reserve_inode_write: Journal has aborted EXT3-fs error (device hda7) in ext3_orphan_del: Journal has aborted
ext3_reserve_inode_write: aborting transaction: Journal has aborted in __ext3_journal_get_write_access<2>EXT3-fs error (device hda7) in ext3_reserve_inode_write: Journal has aborted EXT3-fs error (device hda7) in ext3_delete_inode: Out of memory
__journal_remove_journal_head: freeing b_committed_data
..................
------------------------------------------------------------------

As it is seen from the messages journal_get_undo_access() failed to allocate some memory with jbd_kmalloc(), which in turn called kmalloc() with __GFP_NOFAIL flag.

How could it happen?
The only possible reason for this we suppose is a piece of code in __alloc_pages():

if ((p->flags & (PF_MEMALLOC | PF_MEMDIE)) && !in_interrupt()) {
        /* go through the zonelist yet again, ignoring mins */
        for (i = 0; zones[i] != NULL; i++) {
                struct zone *z = zones[i];

                page = buffered_rmqueue(z, order, gfp_mask);
                if (page) {
                        zone_statistics(zonelist, z);
                        goto got_pg;
                }
        }
        goto nopage;                <<<< HERE!!! FAIL...
}

So kswapd (which has PF_MEMALLOC flag) can fail to allocate memory even when it allocates it with __GFP_NOFAIL flag.

The attached patch should fix the problem (against 2.6.14).

Kirill

@@ -870,6 +870,7 @@ zone_reclaim_retry:
 	if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
 			&& !in_interrupt()) {
 		if (!(gfp_mask & __GFP_NOMEMALLOC)) {
+nofail_alloc:
 			/* go through the zonelist yet again, ignoring mins */
 			for (i = 0; (z = zones[i]) != NULL; i++) {
 				if (!cpuset_zone_allowed(z, gfp_mask))

@@ -878,6 +879,10 @@ zone_reclaim_retry:
 				if (page)
 					goto got_pg;
 			}
+			if (gfp_mask & __GFP_NOFAIL) {
+				blk_congestion_wait(WRITE, HZ/50);
+				goto nofail_alloc;
+			}
 		}
 		goto nopage;
 	}

OK, thanks. This patch was still in my things-i-need-to-spend-half-an-hour-thinking-about queue anyway.

The problem here is that we'll loop if TIF_MEMDIE is set.

But given that the caller has specified __GFP_NOFAIL, I think that's correct behaviour - __GFP_NOFAIL means "I am lame, and will oops if you
cannot allocate memory". So we just ignore TIF_MEMDIE..

That being said, why do we need another loop here? Would it not be sufficient to do:

@@ -907,7 +907,8 @@ zone_reclaim_retry:
 					goto got_pg;
 			}
 		}
-		goto nopage;
+		if (!(gfp_mask & __GFP_NOFAIL))
+			goto nopage;
 	}
 
 	/* Atomic allocations - we can't balance anything */

_

Answer: because that way we'll go recursive if PF_MEMALLOC is set. Ho-hum.