linux-mm.kvack.org archive mirror
 help / color / mirror / Atom feed
* [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
       [not found] <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcas1p4.samsung.com>
@ 2023-04-10  7:32 ` Jaewon Kim
  2023-04-12  8:22   ` Michal Hocko
       [not found]   ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p7>
  0 siblings, 2 replies; 15+ messages in thread
From: Jaewon Kim @ 2023-04-10  7:32 UTC (permalink / raw)
  To: jstultz, tjmercier, sumit.semwal, daniel.vetter, akpm, hannes, mhocko
  Cc: linux-mm, linux-kernel, jaewon31.kim, Jaewon Kim

Normal free:212600kB min:7664kB low:57100kB high:106536kB
  reserved_highatomic:4096KB active_anon:276kB inactive_anon:180kB
  active_file:1200kB inactive_file:0kB unevictable:2932kB
  writepending:0kB present:4109312kB managed:3689488kB mlocked:2932kB
  pagetables:13600kB bounce:0kB free_pcp:0kB local_pcp:0kB
  free_cma:200844kB
Out of memory and no killable processes...
Kernel panic - not syncing: System is deadlocked on memory

An OoM panic was reported. The log shows there were only native
processes which are non-killable as OOM_SCORE_ADJ_MIN. After looking
into the dump, I've found the dma-buf system heap was trying to allocate
a huge size. It seems to be a signed negative value.

dma_heap_ioctl_allocate(inline)
    |  heap_allocation = 0xFFFFFFC02247BD38 -> (
    |    len = 0xFFFFFFFFE7225100,

To avoid this invalid request, check if the requested size is bigger
than system total memory. Actually the old ion system heap had similar
policy with commit c9e8440eca61 ("staging: ion: Fix overflow and list
bugs in system heap").

Even with this sanity check, there is still risk of too much allocations
from the system_heap. Allocating multiple big size buffers may cause
oom. Add __GFP_RETRY_MAYFAIL. With this gfp, the allocation may fail,
but we can avoid oom panic.

Signed-off-by: Jaewon Kim <jaewon31.kim@samsung.com>
Acked-by: John Stultz <jstultz@google.com>
Reviewed-by: T.J. Mercier <tjmercier@google.com>
---
 drivers/dma-buf/heaps/system_heap.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/dma-buf/heaps/system_heap.c b/drivers/dma-buf/heaps/system_heap.c
index 920db302a273..583da8948679 100644
--- a/drivers/dma-buf/heaps/system_heap.c
+++ b/drivers/dma-buf/heaps/system_heap.c
@@ -41,7 +41,7 @@ struct dma_heap_attachment {
 	bool mapped;
 };
 
-#define LOW_ORDER_GFP (GFP_HIGHUSER | __GFP_ZERO)
+#define LOW_ORDER_GFP (GFP_HIGHUSER | __GFP_ZERO | __GFP_RETRY_MAYFAIL)
 #define HIGH_ORDER_GFP  (((GFP_HIGHUSER | __GFP_ZERO | __GFP_NOWARN \
 				| __GFP_NORETRY) & ~__GFP_RECLAIM) \
 				| __GFP_COMP)
@@ -350,6 +350,9 @@ static struct dma_buf *system_heap_allocate(struct dma_heap *heap,
 	struct page *page, *tmp_page;
 	int i, ret = -ENOMEM;
 
+	if (len / PAGE_SIZE > totalram_pages())
+		return ERR_PTR(-ENOMEM);
+
 	buffer = kzalloc(sizeof(*buffer), GFP_KERNEL);
 	if (!buffer)
 		return ERR_PTR(-ENOMEM);
-- 
2.17.1



^ permalink raw reply related	[flat|nested] 15+ messages in thread

* Re: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
  2023-04-10  7:32 ` [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation Jaewon Kim
@ 2023-04-12  8:22   ` Michal Hocko
       [not found]   ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p7>
  1 sibling, 0 replies; 15+ messages in thread
From: Michal Hocko @ 2023-04-12  8:22 UTC (permalink / raw)
  To: Jaewon Kim
  Cc: jstultz, tjmercier, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

Sorry for being late. I know there was some pre-existing discussion
around that but I didn't have time to participate.

On Mon 10-04-23 16:32:28, Jaewon Kim wrote:
> @@ -350,6 +350,9 @@ static struct dma_buf *system_heap_allocate(struct dma_heap *heap,
>  	struct page *page, *tmp_page;
>  	int i, ret = -ENOMEM;
>  
> +	if (len / PAGE_SIZE > totalram_pages())
> +		return ERR_PTR(-ENOMEM);
> +

This is an antipattern imho. Check 7661809d493b ("mm: don't allow
oversized kvmalloc() calls") how kvmalloc has dealt with a similar
issue. totalram_pages doesn't really tell you anything about incorrect
users. You might be on a low memory system where the request size is
sane normally, it just doesn't fit into memory on that particular
machine.


>  	buffer = kzalloc(sizeof(*buffer), GFP_KERNEL);
>  	if (!buffer)
>  		return ERR_PTR(-ENOMEM);
> -- 
> 2.17.1

-- 
Michal Hocko
SUSE Labs


^ permalink raw reply	[flat|nested] 15+ messages in thread

* RE: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
       [not found]   ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p7>
@ 2023-04-12  8:57     ` Jaewon Kim
  2023-04-12  9:23       ` Michal Hocko
       [not found]       ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p4>
  0 siblings, 2 replies; 15+ messages in thread
From: Jaewon Kim @ 2023-04-12  8:57 UTC (permalink / raw)
  To: Michal Hocko
  Cc: jstultz, tjmercier, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

>Sorry for being late. I know there was some pre-existing discussion
>around that but I didn't have time to participate.
>
>On Mon 10-04-23 16:32:28, Jaewon Kim wrote:
>> @@ -350,6 +350,9 @@ static struct dma_buf *system_heap_allocate(struct dma_heap *heap,
>>  	struct page *page, *tmp_page;
>>  	int i, ret = -ENOMEM;
>>  
>> +	if (len / PAGE_SIZE > totalram_pages())
>> +		return ERR_PTR(-ENOMEM);
>> +
>
>This is an antipattern imho. Check 7661809d493b ("mm: don't allow
>oversized kvmalloc() calls") how kvmalloc has dealt with a similar

Hello Thank you for the information.

I tried to search the macro of INT_MAX.

include/vdso/limits.h
#define INT_MAX         ((int)(~0U >> 1))

AFAIK the dma-buf system heap user can request that huge size more than 2GB. So
I think totalram_pages() is better than INT_MAX in this case.

>issue. totalram_pages doesn't really tell you anything about incorrect
>users. You might be on a low memory system where the request size is
>sane normally, it just doesn't fit into memory on that particular
>machine.

Sorry maybe I'm not fully understand what you meant. User may requested
a huge size like 3GB on 2GB ram device. But I think that should be rejected
because it is bigger than the device ram size.

Jaewon Kim

>
>
>>  	buffer = kzalloc(sizeof(*buffer), GFP_KERNEL);
>>  	if (!buffer)
>>  		return ERR_PTR(-ENOMEM);
>> -- 
>> 2.17.1
>
>-- 
>Michal Hocko
>SUSE Labs


^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
  2023-04-12  8:57     ` Jaewon Kim
@ 2023-04-12  9:23       ` Michal Hocko
       [not found]       ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p4>
  1 sibling, 0 replies; 15+ messages in thread
From: Michal Hocko @ 2023-04-12  9:23 UTC (permalink / raw)
  To: Jaewon Kim
  Cc: jstultz, tjmercier, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

On Wed 12-04-23 17:57:26, Jaewon Kim wrote:
> >Sorry for being late. I know there was some pre-existing discussion
> >around that but I didn't have time to participate.
> >
> >On Mon 10-04-23 16:32:28, Jaewon Kim wrote:
> >> @@ -350,6 +350,9 @@ static struct dma_buf *system_heap_allocate(struct dma_heap *heap,
> >>  	struct page *page, *tmp_page;
> >>  	int i, ret = -ENOMEM;
> >>  
> >> +	if (len / PAGE_SIZE > totalram_pages())
> >> +		return ERR_PTR(-ENOMEM);
> >> +
> >
> >This is an antipattern imho. Check 7661809d493b ("mm: don't allow
> >oversized kvmalloc() calls") how kvmalloc has dealt with a similar
> 
> Hello Thank you for the information.
> 
> I tried to search the macro of INT_MAX.
> 
> include/vdso/limits.h
> #define INT_MAX         ((int)(~0U >> 1))
> 
> AFAIK the dma-buf system heap user can request that huge size more than 2GB.

Do you have any pointers? This all is unreclaimable memory, right? How
are those users constrained to not go overboard?

> So
> I think totalram_pages() is better than INT_MAX in this case.
> 
> >issue. totalram_pages doesn't really tell you anything about incorrect
> >users. You might be on a low memory system where the request size is
> >sane normally, it just doesn't fit into memory on that particular
> >machine.
> 
> Sorry maybe I'm not fully understand what you meant. User may requested
> a huge size like 3GB on 2GB ram device. But I think that should be rejected
> because it is bigger than the device ram size.

Even totalram_pages/10 can be just unfeasible amount of data to be
allocated without a major disruption. totalram_pages is no measure of
the memory availability.
If you want to have a ballpark estimation then si_mem_available might be
something you are looking for. But I thought the sole purpose of this
patch is to catch obviously buggy callers (like sign overflow lenght
etc) rather than any memory consumption sanity check.

-- 
Michal Hocko
SUSE Labs


^ permalink raw reply	[flat|nested] 15+ messages in thread

* RE: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
       [not found]       ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p4>
@ 2023-04-12  9:44         ` Jaewon Kim
  2023-04-12 11:02           ` Michal Hocko
       [not found]           ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p8>
  0 siblings, 2 replies; 15+ messages in thread
From: Jaewon Kim @ 2023-04-12  9:44 UTC (permalink / raw)
  To: Michal Hocko
  Cc: jstultz, tjmercier, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

>On Wed 12-04-23 17:57:26, Jaewon Kim wrote:
>> >Sorry for being late. I know there was some pre-existing discussion
>> >around that but I didn't have time to participate.
>> >
>> >On Mon 10-04-23 16:32:28, Jaewon Kim wrote:
>> >> @@ -350,6 +350,9 @@ static struct dma_buf *system_heap_allocate(struct dma_heap *heap,
>> >>  	struct page *page, *tmp_page;
>> >>  	int i, ret = -ENOMEM;
>> >>  
>> >> +	if (len / PAGE_SIZE > totalram_pages())
>> >> +		return ERR_PTR(-ENOMEM);
>> >> +
>> >
>> >This is an antipattern imho. Check 7661809d493b ("mm: don't allow
>> >oversized kvmalloc() calls") how kvmalloc has dealt with a similar
>> 
>> Hello Thank you for the information.
>> 
>> I tried to search the macro of INT_MAX.
>> 
>> include/vdso/limits.h
>> #define INT_MAX         ((int)(~0U >> 1))
>> 
>> AFAIK the dma-buf system heap user can request that huge size more than 2GB.
>
>Do you have any pointers? This all is unreclaimable memory, right? How
>are those users constrained to not go overboard?

Correct dma-buf system heap memory is unreclaimable. To avoid that huge request,
this patch includes __GFP_RETRY_MAYFAIL.

 #define LOW_ORDER_GFP (GFP_HIGHUSER | __GFP_ZERO | __GFP_RETRY_MAYFAIL)

>
>> So
>> I think totalram_pages() is better than INT_MAX in this case.
>> 
>> >issue. totalram_pages doesn't really tell you anything about incorrect
>> >users. You might be on a low memory system where the request size is
>> >sane normally, it just doesn't fit into memory on that particular
>> >machine.
>> 
>> Sorry maybe I'm not fully understand what you meant. User may requested
>> a huge size like 3GB on 2GB ram device. But I think that should be rejected
>> because it is bigger than the device ram size.
>
>Even totalram_pages/10 can be just unfeasible amount of data to be
>allocated without a major disruption. totalram_pages is no measure of
>the memory availability.
>If you want to have a ballpark estimation then si_mem_available might be
>something you are looking for. But I thought the sole purpose of this
>patch is to catch obviously buggy callers (like sign overflow lenght
>etc) rather than any memory consumption sanity check.

Yes if we want to avoid some big size, si_mem_available could be one option.
Actually I tried to do totalram_pages() / 2 like the old ion system heap in
the previous patch version. Anyway totalram_pages in this patch is used to
avoid the buggy size.

And as we discussed in v2 patch, __GFP_RETRY_MAYFAIL was added. And I think
the gfp makes us feel better in memory perspective.

>
>-- 
>Michal Hocko
>SUSE Labs


^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
  2023-04-12  9:44         ` Jaewon Kim
@ 2023-04-12 11:02           ` Michal Hocko
       [not found]           ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p8>
  1 sibling, 0 replies; 15+ messages in thread
From: Michal Hocko @ 2023-04-12 11:02 UTC (permalink / raw)
  To: Jaewon Kim
  Cc: jstultz, tjmercier, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

On Wed 12-04-23 18:44:40, Jaewon Kim wrote:
> >On Wed 12-04-23 17:57:26, Jaewon Kim wrote:
> >> >Sorry for being late. I know there was some pre-existing discussion
> >> >around that but I didn't have time to participate.
> >> >
> >> >On Mon 10-04-23 16:32:28, Jaewon Kim wrote:
> >> >> @@ -350,6 +350,9 @@ static struct dma_buf *system_heap_allocate(struct dma_heap *heap,
> >> >>  	struct page *page, *tmp_page;
> >> >>  	int i, ret = -ENOMEM;
> >> >>  
> >> >> +	if (len / PAGE_SIZE > totalram_pages())
> >> >> +		return ERR_PTR(-ENOMEM);
> >> >> +
> >> >
> >> >This is an antipattern imho. Check 7661809d493b ("mm: don't allow
> >> >oversized kvmalloc() calls") how kvmalloc has dealt with a similar
> >> 
> >> Hello Thank you for the information.
> >> 
> >> I tried to search the macro of INT_MAX.
> >> 
> >> include/vdso/limits.h
> >> #define INT_MAX         ((int)(~0U >> 1))
> >> 
> >> AFAIK the dma-buf system heap user can request that huge size more than 2GB.
> >
> >Do you have any pointers? This all is unreclaimable memory, right? How
> >are those users constrained to not go overboard?
> 
> Correct dma-buf system heap memory is unreclaimable. To avoid that huge request,
> this patch includes __GFP_RETRY_MAYFAIL.

__GFP_RETRY_MAYFAIL doesn't avoud huge requests. It will drain the free
available memory to the edge of OOM (especially for low order requests)
so effectively anybody else requesting any memory (GFP_KERNEL like req.)
will hit the oom killer very likely).

>  #define LOW_ORDER_GFP (GFP_HIGHUSER | __GFP_ZERO | __GFP_RETRY_MAYFAIL)
> 
> >
> >> So
> >> I think totalram_pages() is better than INT_MAX in this case.
> >> 
> >> >issue. totalram_pages doesn't really tell you anything about incorrect
> >> >users. You might be on a low memory system where the request size is
> >> >sane normally, it just doesn't fit into memory on that particular
> >> >machine.
> >> 
> >> Sorry maybe I'm not fully understand what you meant. User may requested
> >> a huge size like 3GB on 2GB ram device. But I think that should be rejected
> >> because it is bigger than the device ram size.
> >
> >Even totalram_pages/10 can be just unfeasible amount of data to be
> >allocated without a major disruption. totalram_pages is no measure of
> >the memory availability.
> >If you want to have a ballpark estimation then si_mem_available might be
> >something you are looking for. But I thought the sole purpose of this
> >patch is to catch obviously buggy callers (like sign overflow lenght
> >etc) rather than any memory consumption sanity check.
> 
> Yes if we want to avoid some big size, si_mem_available could be one option.
> Actually I tried to do totalram_pages() / 2 like the old ion system heap in
> the previous patch version. Anyway totalram_pages in this patch is used to
> avoid the buggy size.

So let me repeat that totalram_pages is a wrong thing to do(tm).

This is not a subsystem I would feel like nacking a patch, but consider
this feedback as strong of a rejection as somebody external can give
you. A mm internal allocator would get an outright nack.

What you are doing is just wrong and an antipattern to what other
allocators do. Either use something like INT_MAX to catch overflows or
do not try to catch buggy code but pretend a better memory consumer
citizen by using something like si_mem_available (ideally think of
other potential memory users so do not allow any request to use all
of it). The later might require much more involved interface and I do
rememeber some attempts to account and limit dmabuf memory better.

> And as we discussed in v2 patch, __GFP_RETRY_MAYFAIL was added. And I think
> the gfp makes us feel better in memory perspective.

wishful thinking that is.
-- 
Michal Hocko
SUSE Labs


^ permalink raw reply	[flat|nested] 15+ messages in thread

* RE: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
       [not found]           ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p8>
@ 2023-04-12 11:37             ` Jaewon Kim
  2023-04-12 11:51               ` Michal Hocko
                                 ` (3 more replies)
  0 siblings, 4 replies; 15+ messages in thread
From: Jaewon Kim @ 2023-04-12 11:37 UTC (permalink / raw)
  To: Michal Hocko
  Cc: jstultz, tjmercier, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

>On Wed 12-04-23 18:44:40, Jaewon Kim wrote:
>> >On Wed 12-04-23 17:57:26, Jaewon Kim wrote:
>> >> >Sorry for being late. I know there was some pre-existing discussion
>> >> >around that but I didn't have time to participate.
>> >> >
>> >> >On Mon 10-04-23 16:32:28, Jaewon Kim wrote:
>> >> >> @@ -350,6 +350,9 @@ static struct dma_buf *system_heap_allocate(struct dma_heap *heap,
>> >> >>  	struct page *page, *tmp_page;
>> >> >>  	int i, ret = -ENOMEM;
>> >> >>  
>> >> >> +	if (len / PAGE_SIZE > totalram_pages())
>> >> >> +		return ERR_PTR(-ENOMEM);
>> >> >> +
>> >> >
>> >> >This is an antipattern imho. Check 7661809d493b ("mm: don't allow
>> >> >oversized kvmalloc() calls") how kvmalloc has dealt with a similar
>> >> 
>> >> Hello Thank you for the information.
>> >> 
>> >> I tried to search the macro of INT_MAX.
>> >> 
>> >> include/vdso/limits.h
>> >> #define INT_MAX         ((int)(~0U >> 1))
>> >> 
>> >> AFAIK the dma-buf system heap user can request that huge size more than 2GB.
>> >
>> >Do you have any pointers? This all is unreclaimable memory, right? How
>> >are those users constrained to not go overboard?
>> 
>> Correct dma-buf system heap memory is unreclaimable. To avoid that huge request,
>> this patch includes __GFP_RETRY_MAYFAIL.
>
>__GFP_RETRY_MAYFAIL doesn't avoud huge requests. It will drain the free
>available memory to the edge of OOM (especially for low order requests)
>so effectively anybody else requesting any memory (GFP_KERNEL like req.)
>will hit the oom killer very likely).
>
>>  #define LOW_ORDER_GFP (GFP_HIGHUSER | __GFP_ZERO | __GFP_RETRY_MAYFAIL)
>> 
>> >
>> >> So
>> >> I think totalram_pages() is better than INT_MAX in this case.
>> >> 
>> >> >issue. totalram_pages doesn't really tell you anything about incorrect
>> >> >users. You might be on a low memory system where the request size is
>> >> >sane normally, it just doesn't fit into memory on that particular
>> >> >machine.
>> >> 
>> >> Sorry maybe I'm not fully understand what you meant. User may requested
>> >> a huge size like 3GB on 2GB ram device. But I think that should be rejected
>> >> because it is bigger than the device ram size.
>> >
>> >Even totalram_pages/10 can be just unfeasible amount of data to be
>> >allocated without a major disruption. totalram_pages is no measure of
>> >the memory availability.
>> >If you want to have a ballpark estimation then si_mem_available might be
>> >something you are looking for. But I thought the sole purpose of this
>> >patch is to catch obviously buggy callers (like sign overflow lenght
>> >etc) rather than any memory consumption sanity check.
>> 
>> Yes if we want to avoid some big size, si_mem_available could be one option.
>> Actually I tried to do totalram_pages() / 2 like the old ion system heap in
>> the previous patch version. Anyway totalram_pages in this patch is used to
>> avoid the buggy size.
>
>So let me repeat that totalram_pages is a wrong thing to do(tm).
>
>This is not a subsystem I would feel like nacking a patch, but consider
>this feedback as strong of a rejection as somebody external can give
>you. A mm internal allocator would get an outright nack.
>
>What you are doing is just wrong and an antipattern to what other
>allocators do. Either use something like INT_MAX to catch overflows or
>do not try to catch buggy code but pretend a better memory consumer
>citizen by using something like si_mem_available (ideally think of
>other potential memory users so do not allow any request to use all
>of it). The later might require much more involved interface and I do
>rememeber some attempts to account and limit dmabuf memory better.
>
>> And as we discussed in v2 patch, __GFP_RETRY_MAYFAIL was added. And I think
>> the gfp makes us feel better in memory perspective.
>
>wishful thinking that is.
>-- 
>Michal Hocko
>SUSE Labs

Yes I think you're right. As a allocator, dma-buf system heap looks to be loose
in memory allocation. Limiting dmabuf memory may be required. But I think there
is no nice and reasonable way so far. And the dma-buf system heap is being
widely used in Android mobile system. AFAIK the camera consumes huge memory
through this dma-buf system heap. I actually even looked a huge size request
over 2GB in one dma-buf request.

Jaewon Kim


^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
  2023-04-12 11:37             ` Jaewon Kim
@ 2023-04-12 11:51               ` Michal Hocko
       [not found]               ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p2>
                                 ` (2 subsequent siblings)
  3 siblings, 0 replies; 15+ messages in thread
From: Michal Hocko @ 2023-04-12 11:51 UTC (permalink / raw)
  To: Jaewon Kim
  Cc: jstultz, tjmercier, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

On Wed 12-04-23 20:37:59, Jaewon Kim wrote:
> Limiting dmabuf memory may be required. But I think there
> is no nice and reasonable way so far.

If that is really the way then the patch doesn't really add a big
benefit. It doesn't really prevent OOMs (or panics due to OOM) as the
allocator still allows to consume arbitrary amount of memory. The
provided check is not able to tell between buggy and legit calls.
-- 
Michal Hocko
SUSE Labs


^ permalink raw reply	[flat|nested] 15+ messages in thread

* RE: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
       [not found]               ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p2>
@ 2023-04-12 12:35                 ` Jaewon Kim
  2023-04-12 13:01                   ` Michal Hocko
  0 siblings, 1 reply; 15+ messages in thread
From: Jaewon Kim @ 2023-04-12 12:35 UTC (permalink / raw)
  To: Michal Hocko
  Cc: jstultz, tjmercier, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

>On Wed 12-04-23 20:37:59, Jaewon Kim wrote:
>> Limiting dmabuf memory may be required. But I think there
>> is no nice and reasonable way so far.
>
>If that is really the way then the patch doesn't really add a big
>benefit. It doesn't really prevent OOMs (or panics due to OOM) as the
>allocator still allows to consume arbitrary amount of memory. The
>provided check is not able to tell between buggy and legit calls.
>-- 
>Michal Hocko
>SUSE Labs

Yes it could be. Though the buggy call is blocked by totalram_pages check,
mm may suffer memory shortage due to the huge memory consumption through
dma-buf system heap. We just hope Android LMKD or oomk kills the memory
hoggers prior to oom panic.

IMO if possible mm should be able to track the dma-buf size as stat in
mm_rss_stat for each process. Then oomk is able to know size through
get_mm_rss. Actually in mobile device environment, there is one more
this kind of allocator. That is graphics drivers. They also consumes
quite much memory and they may allocate as userspace wants.

Jaewon Kim


^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
  2023-04-12 12:35                 ` Jaewon Kim
@ 2023-04-12 13:01                   ` Michal Hocko
  0 siblings, 0 replies; 15+ messages in thread
From: Michal Hocko @ 2023-04-12 13:01 UTC (permalink / raw)
  To: Jaewon Kim
  Cc: jstultz, tjmercier, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

On Wed 12-04-23 21:35:32, Jaewon Kim wrote:
> >On Wed 12-04-23 20:37:59, Jaewon Kim wrote:
> >> Limiting dmabuf memory may be required. But I think there
> >> is no nice and reasonable way so far.
> >
> >If that is really the way then the patch doesn't really add a big
> >benefit. It doesn't really prevent OOMs (or panics due to OOM) as the
> >allocator still allows to consume arbitrary amount of memory. The
> >provided check is not able to tell between buggy and legit calls.
> >-- 
> >Michal Hocko
> >SUSE Labs
> 
> Yes it could be. Though the buggy call is blocked by totalram_pages check,

It seems our definitions of buggy differ here. I do not see much
difference between totalram_pages +- PAGE_SIZE (or any epsilon for that
matter). Both would put the system down to its knees without a way out
other than panic.

> mm may suffer memory shortage due to the huge memory consumption through
> dma-buf system heap. We just hope Android LMKD or oomk kills the memory
> hoggers prior to oom panic.

You seem to be missing an important point. If the global OOM killer is
not able to find a victim the LMKD or oomk are highly unlikely as well
(unless they ignore OOM_SCORE_ADJ_MIN).

> IMO if possible mm should be able to track the dma-buf size as stat in
> mm_rss_stat for each process.

I do remember some proposals from the past and IIRC the main problem was
how to attribute those buffers to the actual owner.

I believe I have give you some arguments to consider. The rest is up to
you. As I've said I do not have any stakes in dmabuf. The patch itself
is not actively harmful, it is just adding an illusion of a fix while it
doesn't give much.
-- 
Michal Hocko
SUSE Labs


^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
  2023-04-12 11:37             ` Jaewon Kim
  2023-04-12 11:51               ` Michal Hocko
       [not found]               ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p2>
@ 2023-04-12 16:49               ` T.J. Mercier
  2023-04-12 22:10                 ` Jaewon Kim
       [not found]               ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p6>
  3 siblings, 1 reply; 15+ messages in thread
From: T.J. Mercier @ 2023-04-12 16:49 UTC (permalink / raw)
  To: jaewon31.kim
  Cc: Michal Hocko, jstultz, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

On Wed, Apr 12, 2023 at 4:38 AM Jaewon Kim <jaewon31.kim@samsung.com> wrote:

> Yes I think you're right. As a allocator, dma-buf system heap looks to be loose
> in memory allocation. Limiting dmabuf memory may be required. But I think there
> is no nice and reasonable way so far. And the dma-buf system heap is being
> widely used in Android mobile system. AFAIK the camera consumes huge memory
> through this dma-buf system heap. I actually even looked a huge size request
> over 2GB in one dma-buf request.
>
Hey can you point me to where you saw a request that big? That's a
non-buggy request?!


^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
  2023-04-12 16:49               ` T.J. Mercier
@ 2023-04-12 22:10                 ` Jaewon Kim
  0 siblings, 0 replies; 15+ messages in thread
From: Jaewon Kim @ 2023-04-12 22:10 UTC (permalink / raw)
  To: T.J. Mercier
  Cc: Jaewon Kim, Michal Hocko, jstultz, Sumit Semwal, daniel.vetter,
	Andrew Morton, hannes, linux-mm, linux-kernel

[-- Attachment #1: Type: text/plain, Size: 987 bytes --]

It was one of camera scenarios. I internally asked and heard that was not a
bug but normal. I think 2GB looks too big for one graphics buffer but it
could be for other purposes like camera. I think the system heap should
support that.

2023년 4월 13일 (목) 오전 1:49, T.J. Mercier <tjmercier@google.com>님이 작성:

> On Wed, Apr 12, 2023 at 4:38 AM Jaewon Kim <jaewon31.kim@samsung.com>
> wrote:
>
> > Yes I think you're right. As a allocator, dma-buf system heap looks to
> be loose
> > in memory allocation. Limiting dmabuf memory may be required. But I
> think there
> > is no nice and reasonable way so far. And the dma-buf system heap is
> being
> > widely used in Android mobile system. AFAIK the camera consumes huge
> memory
> > through this dma-buf system heap. I actually even looked a huge size
> request
> > over 2GB in one dma-buf request.
> >
> Hey can you point me to where you saw a request that big? That's a
> non-buggy request?!
>

[-- Attachment #2: Type: text/html, Size: 1402 bytes --]

^ permalink raw reply	[flat|nested] 15+ messages in thread

* RE: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
       [not found]               ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p6>
@ 2023-04-13  0:16                 ` Jaewon Kim
  2023-04-13  6:55                   ` Michal Hocko
       [not found]                   ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p1>
  0 siblings, 2 replies; 15+ messages in thread
From: Jaewon Kim @ 2023-04-13  0:16 UTC (permalink / raw)
  To: T.J. Mercier
  Cc: Michal Hocko, jstultz, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

>On Wed, Apr 12, 2023 at 4:38?AM Jaewon Kim <jaewon31.kim@samsung.com> wrote:
>
>> Yes I think you're right. As a allocator, dma-buf system heap looks to be loose
>> in memory allocation. Limiting dmabuf memory may be required. But I think there
>> is no nice and reasonable way so far. And the dma-buf system heap is being
>> widely used in Android mobile system. AFAIK the camera consumes huge memory
>> through this dma-buf system heap. I actually even looked a huge size request
>> over 2GB in one dma-buf request.
>>
>Hey can you point me to where you saw a request that big? That's a
>non-buggy request?!

(let me resend as plain text)
It was one of camera scenarios. I internally asked and heard that was not a bug
but normal. I think 2GB looks too big for one graphics buffer but it could be
for other purposes like camera. I think the system heap should support that.

Regarding __GFP_RETRY_MAYFAIL, we may need to say dma-buf system heap was
designed to gather many pages up to a requested size. If mm returns NULL due to
__GFP_RETRY_MAYFAIL, dma-buf system heap will release other already allocated
pages, so that it may help to avoid oom.


^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
  2023-04-13  0:16                 ` Jaewon Kim
@ 2023-04-13  6:55                   ` Michal Hocko
       [not found]                   ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p1>
  1 sibling, 0 replies; 15+ messages in thread
From: Michal Hocko @ 2023-04-13  6:55 UTC (permalink / raw)
  To: Jaewon Kim
  Cc: T.J. Mercier, jstultz, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

On Thu 13-04-23 09:16:58, Jaewon Kim wrote:
> >On Wed, Apr 12, 2023 at 4:38?AM Jaewon Kim <jaewon31.kim@samsung.com> wrote:
> >
> >> Yes I think you're right. As a allocator, dma-buf system heap looks to be loose
> >> in memory allocation. Limiting dmabuf memory may be required. But I think there
> >> is no nice and reasonable way so far. And the dma-buf system heap is being
> >> widely used in Android mobile system. AFAIK the camera consumes huge memory
> >> through this dma-buf system heap. I actually even looked a huge size request
> >> over 2GB in one dma-buf request.
> >>
> >Hey can you point me to where you saw a request that big? That's a
> >non-buggy request?!
> 
> (let me resend as plain text)
> It was one of camera scenarios. I internally asked and heard that was not a bug
> but normal. I think 2GB looks too big for one graphics buffer but it could be
> for other purposes like camera. I think the system heap should support that.

Is that any of the upstream drivers or something sitting out of the
tree.

> Regarding __GFP_RETRY_MAYFAIL, we may need to say dma-buf system heap was
> designed to gather many pages up to a requested size. If mm returns NULL due to
> __GFP_RETRY_MAYFAIL, dma-buf system heap will release other already allocated
> pages, so that it may help to avoid oom.

This really depends on the other activity on the system. If you have a
more concurrent memory demand at the time then you might be just out of
the luck. Really, claiming huge portion of the memory shouldn't be done
nilly willy.
-- 
Michal Hocko
SUSE Labs


^ permalink raw reply	[flat|nested] 15+ messages in thread

* RE: [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation
       [not found]                   ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p1>
@ 2023-04-13  7:01                     ` Jaewon Kim
  0 siblings, 0 replies; 15+ messages in thread
From: Jaewon Kim @ 2023-04-13  7:01 UTC (permalink / raw)
  To: Michal Hocko, Jaewon Kim
  Cc: T.J. Mercier, jstultz, sumit.semwal, daniel.vetter, akpm, hannes,
	linux-mm, linux-kernel, jaewon31.kim

>On Thu 13-04-23 09:16:58, Jaewon Kim wrote:
>> >On Wed, Apr 12, 2023 at 4:38?AM Jaewon Kim <jaewon31.kim@samsung.com> wrote:
>> >
>> >> Yes I think you're right. As a allocator, dma-buf system heap looks to be loose
>> >> in memory allocation. Limiting dmabuf memory may be required. But I think there
>> >> is no nice and reasonable way so far. And the dma-buf system heap is being
>> >> widely used in Android mobile system. AFAIK the camera consumes huge memory
>> >> through this dma-buf system heap. I actually even looked a huge size request
>> >> over 2GB in one dma-buf request.
>> >>
>> >Hey can you point me to where you saw a request that big? That's a
>> >non-buggy request?!
>> 
>> (let me resend as plain text)
>> It was one of camera scenarios. I internally asked and heard that was not a bug
>> but normal. I think 2GB looks too big for one graphics buffer but it could be
>> for other purposes like camera. I think the system heap should support that.
>
>Is that any of the upstream drivers or something sitting out of the
>tree.

I don't think so. I guess that is userspace library rather than kernel driver.
The user library directly might request the size through dma-buf ioctl. Even
though that is kernel driver, I think the driver may not be upstreamed.

>
>> Regarding __GFP_RETRY_MAYFAIL, we may need to say dma-buf system heap was
>> designed to gather many pages up to a requested size. If mm returns NULL due to
>> __GFP_RETRY_MAYFAIL, dma-buf system heap will release other already allocated
>> pages, so that it may help to avoid oom.
>
>This really depends on the other activity on the system. If you have a
>more concurrent memory demand at the time then you might be just out of
>the luck. Really, claiming huge portion of the memory shouldn't be done
>nilly willy.

I agree on that.

>-- 
>Michal Hocko
>SUSE Labs


^ permalink raw reply	[flat|nested] 15+ messages in thread

end of thread, other threads:[~2023-04-13  7:02 UTC | newest]

Thread overview: 15+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
     [not found] <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcas1p4.samsung.com>
2023-04-10  7:32 ` [PATCH v3] dma-buf/heaps: system_heap: avoid too much allocation Jaewon Kim
2023-04-12  8:22   ` Michal Hocko
     [not found]   ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p7>
2023-04-12  8:57     ` Jaewon Kim
2023-04-12  9:23       ` Michal Hocko
     [not found]       ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p4>
2023-04-12  9:44         ` Jaewon Kim
2023-04-12 11:02           ` Michal Hocko
     [not found]           ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p8>
2023-04-12 11:37             ` Jaewon Kim
2023-04-12 11:51               ` Michal Hocko
     [not found]               ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p2>
2023-04-12 12:35                 ` Jaewon Kim
2023-04-12 13:01                   ` Michal Hocko
2023-04-12 16:49               ` T.J. Mercier
2023-04-12 22:10                 ` Jaewon Kim
     [not found]               ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p6>
2023-04-13  0:16                 ` Jaewon Kim
2023-04-13  6:55                   ` Michal Hocko
     [not found]                   ` <CGME20230410073304epcas1p4cf3079b096994d69472b7801bd530bc7@epcms1p1>
2023-04-13  7:01                     ` Jaewon Kim

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).