Skip to contentSkip to conversation

zram: Prevent page allocation failure during zcomp_strm_alloc

PATCH Minchan Kim 2 participants 3 messages Nov 23, 2015

lore
Share & tools
lore
#0 lore raw

New thread

Quoted 20 lines · Sergey Senozhatsky

Hello,

On (11/23/15 11:15), Minchan Kim wrote:
[..]

Quoted 6 lines
Quoted 6 lines

static void *zcomp_lz4_create(void)
{

(quoted diff) +3 -1
-	return kzalloc(LZ4_MEM_COMPRESS, GFP_KERNEL);
+	void *ret;
+
+	ret = kzalloc(LZ4_MEM_COMPRESS,
Quoted 9 lines

Review

Quoted 3 lines
(quoted diff) +3 -0
+			__GFP_NORETRY|__GFP_NOWARN|__GFP_NOMEMALLOC);
+	if (!ret)
+		ret = vzalloc(LZ4_MEM_COMPRESS);

One thing I feel bad smell is that call vzalloc with GFP_KERNEL. This function can be called in direct reclaim path with holding fs lock and GFP_KERNEL can enter recursive reclaim path so lockdep would complain theoretically if I don't miss something.

Quoted 3 lines

yes, GFP_KERNEL looks a bit fragile to me too. And may be zcomp_strm_alloc() and comp->backend->create() deserve GFP_NOFS. I believe I sent a patch doing this a while ago: https://lkml.org/lkml/2015/6/16/465

Sorry, I have missed that.
It's worth to fix that you proved it that could happen. But when I read your patch, GFP_NOIO instead GFP_NOFS would better way. Could you resend it?

Quoted 11 lines
Quoted 5 lines

If it is true, we should fix several allocation flags in zcomp_strm_alloc. I just want to record this warning for the future in this thread so someone who is finding for the contribution material will prove and fix it. :)

I can re-send the patch.

And, in case if you missed it, what's your opinion on the idea of reducing ->max_strm if we can't allocate new streams. Here:
http://marc.info/?l=linux-kernel&m=144798049429861

Hmm, auto backoff max_comp_streams with warn to user, I'm not sure we really need it. Alloc failure is depenent on the workload and timing so although there are a fail in t1, it could be successful in t2 so if we *really* want to introduce auto backoff, the logic should include advance routine as well as backoff. With that, I want to handle it traparently without notice to user.

*HOWEVER*, I think the problem Kyeongdon said came from warning and memset by __GFP_ZERO flood. Other overheads for alloc/free would be not heavy because they are higly optimized path in the MM part. About reclaiming part, it should be almost no problem because reclaimer cannot make forward progress by deadlock so it doesn't scan any LRU.

So, Kyeongdon's patch will remove warning overhead and likely to make zcomp_stram_alloc successful with vmalloc so I want to roll it out first. And let's add a WARN_ON_ONCE to detect of failure and rethink it when we receive such report.

Thanks.

#1 lore raw

On (11/23/15 13:18), Minchan Kim wrote:
[..]

Quoted 6 lines
Quoted 1 line

https://lkml.org/lkml/2015/6/16/465

Sorry, I have missed that.
It's worth to fix that you proved it that could happen. But when I read your patch, GFP_NOIO instead GFP_NOFS would better way. Could you resend it?

no problem.

agree. we also would want to switch from vzalloc() to

	__vmalloc_node_flags(size, NUMA_NO_NODE,
			GFP_NOIO | __GFP_HIGHMEM |  __GFP_ZERO)

in fallbacks. I'll send the patch later today.

Quoted 12 lines
Quoted 11 lines
Quoted 5 lines

If it is true, we should fix several allocation flags in zcomp_strm_alloc. I just want to record this warning for the future in this thread so someone who is finding for the contribution material will prove and fix it. :)

I can re-send the patch.

And, in case if you missed it, what's your opinion on the idea of reducing ->max_strm if we can't allocate new streams. Here:
http://marc.info/?l=linux-kernel&m=144798049429861

Quoted 6 lines

Hmm, auto backoff max_comp_streams with warn to user, I'm not sure we really need it. Alloc failure is depenent on the workload and timing so although there are a fail in t1, it could be successful in t2 so if we *really* want to introduce auto backoff, the logic should include advance routine as well as backoff. With that, I want to handle it traparently without notice to user.

yes. auto roll-back is important (that's why I mentioned it). the idea is to avoid stealing of pages for streams. for example, in case of low memory decrease ->max_strm to

	MAX(->avail_strm, ->max_strm, online cpus() / 2)

and even probably release some idle streams (um... as a reaction to shrinker call???). or at least prevent setting of ->max_sgtrm to some unreasonably huge values... "> 4 * NR_CPUS", for instance. but this is not so critical.

Quoted 4 lines

So, Kyeongdon's patch will remove warning overhead and likely to make zcomp_stram_alloc successful with vmalloc so I want to roll it out first. And let's add a WARN_ON_ONCE to detect of failure and rethink it when we receive such report.

	-ss
#2 lore raw

On (11/23/15 16:43), Sergey Senozhatsky wrote:
[..]

Quoted 3 lines

agree. we also would want to switch from vzalloc() to

	__vmalloc_node_flags(size, NUMA_NO_NODE,
			GFP_NOIO | __GFP_HIGHMEM |  __GFP_ZERO)

[..]

Quoted 4 lines
Quoted 4 lines

So, Kyeongdon's patch will remove warning overhead and likely to make zcomp_stram_alloc successful with vmalloc so I want to roll it out first. And let's add a WARN_ON_ONCE to detect of failure and rethink it when we receive such report.

hm... for k{z,m}alloc() it does reduce the warning overhead, but not for vmalloc() -> warn_alloc_failed() [in theory]. So I guess, I'll change vmalloc() to

	__vmalloc(XXX,
		GFP_NOIO | __GFP_NOWARN | __GFP_HIGHMEM | __GFP_ZERO,
		PAGE_KERNEL);

/* passing __GFP_NOWARN */.

	-ss