Message-ID: <53E7B39D.2060900@gmail.com>
Date:	Sun, 10 Aug 2014 20:02:05 +0200
From:	Mario Kleiner <mario.kleiner.de@...il.com>
To:	Thomas Hellstrom <thellstrom@...are.com>
CC:	Konrad Rzeszutek Wilk <konrad.wilk@...cle.com>,
	kamal@...onical.com, ben@...adent.org.uk,
	LKML <linux-kernel@...r.kernel.org>,
	"dri-devel@...ts.freedesktop.org" <dri-devel@...ts.freedesktop.org>,
	m.szyprowski@...sung.com
Subject: Re: CONFIG_DMA_CMA causes ttm performance problems/hangs.

On 08/10/2014 01:03 PM, Thomas Hellstrom wrote:
> On 08/10/2014 05:11 AM, Mario Kleiner wrote:
>> Resent this time without HTML formatting which lkml doesn't like. Sorry.
>>
>> On 08/09/2014 03:58 PM, Thomas Hellstrom wrote:
>>> On 08/09/2014 03:33 PM, Konrad Rzeszutek Wilk wrote:
>>>> On August 9, 2014 1:39:39 AM EDT, Thomas
>>>> Hellstrom<thellstrom@...are.com>  wrote:
>>>>> Hi.
>>>>>
>>>> Hey Thomas!
>>>>
>>>>> IIRC I don't think the TTM DMA pool allocates coherent pages more than
>>>>> one page at a time, and _if that's true_ it's pretty unnecessary for
>>>>> the
>>>>> dma subsystem to route those allocations to CMA. Maybe Konrad could
>>>>> shed
>>>>> some light over this?
>>>> It should allocate in batches and keep them in the TTM DMA pool for
>>>> some time to be reused.
>>>>
>>>> The pages that it gets are in 4kb granularity though.
>>> Then I feel inclined to say this is a DMA subsystem bug. Single page
>>> allocations shouldn't get routed to CMA.
>>>
>>> /Thomas
>> Yes, seems you're both right. I read through the code a bit more and
>> indeed the TTM DMA pool allocates only one page during each
>> dma_alloc_coherent() call, so it doesn't need CMA memory. The current
>> allocators don't check for single page CMA allocations and therefore
>> try to get it from the CMA area anyway, instead of skipping to the
>> much cheaper fallback.
>>
>> So the callers of dma_alloc_from_contiguous() could need that little
>> optimization of skipping it if only one page is requested. For
>>
>> dma_generic_alloc_coherent
>> <http://lxr.free-electrons.com/ident?i=dma_generic_alloc_coherent>
>> andintel_alloc_coherent
>> <http://lxr.free-electrons.com/ident?i=intel_alloc_coherent>  this
>> seems easy to do. Looking at the arm arch variants, e.g.,
>>
>> http://lxr.free-electrons.com/source/arch/arm/mm/dma-mapping.c#L1194
>>
>> and
>>
>> http://lxr.free-electrons.com/source/arch/arm64/mm/dma-mapping.c#L44
>>
>> i'm not sure if it is that easily done, as there aren't any fallbacks
>> for such a case and the code looks to me as if that's at least
>> somewhat intentional.
>>
>> As far as TTM goes, one quick one-line fix to prevent it from using
>> the CMA at least on SWIOTLB, NOMMU and Intel IOMMU (when using the
>> above methods) would be to clear the __GFP_WAIT
>> <http://lxr.free-electrons.com/ident?i=__GFP_WAIT> flag from the
>> passed gfp_t flags. That would trigger the well working fallback. So, is
>>
>> __GFP_WAIT  <http://lxr.free-electrons.com/ident?i=__GFP_WAIT>  needed
>> for those single page allocations that go through__ttm_dma_alloc_page
>> <http://lxr.free-electrons.com/ident?i=__ttm_dma_alloc_page>?
>>
>> It would be nice to have such a simple, non-intrusive one-line patch
>> that we still could get into 3.17 and then backported to older stable
>> kernels to avoid the same desktop hangs there if CMA is enabled. It
>> would be also nice for actual users of CMA to not use up lots of CMA
>> space for gpu's which don't need it. I think DMA_CMA was introduced
>> around 3.12.
>>
> I don't think that's a good idea. Omitting __GFP_WAIT would cause
> unnecessary memory allocation errors on systems under stress.
> I think this should be filed as a DMA subsystem kernel bug / regression
> and an appropriate solution should be worked out together with the DMA
> subsystem maintainers and then backported.

Ok, so it is needed. I'll file a bug report.

>> The other problem is that probably TTM does not reuse pages from the
>> DMA pool. If i trace the __ttm_dma_alloc_page
>> <http://lxr.free-electrons.com/ident?i=__ttm_dma_alloc_page> and
>> __ttm_dma_free_page
>> <http://lxr.free-electrons.com/ident?i=__ttm_dma_alloc_page> calls for
>> those single page allocs/frees, then over a 20 second interval of
>> tracing and switching tabs in firefox, scrolling things around etc. i
>> find about as many alloc's as i find free's, e.g., 1607 allocs vs.
>> 1648 frees.
> This is because historically the pools have been designed to keep only
> pages with nonstandard caching attributes since changing page caching
> attributes have been very slow but the kernel page allocators have been
> reasonably fast.
>
> /Thomas

Ok. A bit more ftraceing showed my hang problem case goes through the 
"if (is_cached)" paths, so the pool doesn't recycle anything and i see 
it bouncing up and down by 4 pages all the time.

But for the non-cached case, which i don't hit with my problem, could 
one of you look at line 954...

http://lxr.free-electrons.com/source/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c#L954

... and tell me why that unconditional npages = count; assignment makes sense? It seems to essentially disable all recycling for the dma pool whenever the pool isn't filled up to/beyond its maximum with free pages? When the pool is filled up, lots of stuff is recycled, but when it is already somewhat below capacity, it gets "punished" by not getting refilled? I'd just like to understand the logic behind that line.

thanks,
-mario


>> This bit of code fromttm_dma_unpopulate
>> <http://lxr.free-electrons.com/ident?i=ttm_dma_unpopulate>()  (line
>> 954 in 3.16) looks suspicious:
>>
>> http://lxr.free-electrons.com/source/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c#L954
>>
>>
>> Alloc's from a tt_cached cached pool ( if (is_cached)...) always get
>> freed and are not given back to the cached pool. But in the uncached
>> case, there's logic to make sure the pool doesn't grow forever (line
>> 955, checking against _manager->options.max_size), but before that
>> check in line 954 there's an uncoditional assignment of npages =
>> count; which seems to force freeing all pages as well, instead of
>> recycling? Is this some debug code left over, or intentional and just
>> me not understanding what happens there?
>>
>> thanks,
>> -mario
>>
>>
>>>>> /Thomas
>>>>>
>>>>>
>>>>> On 08/08/2014 07:42 PM, Mario Kleiner wrote:
>>>>>> Hi all,
>>>>>>
>>>>>> there is a rather severe performance problem i accidentally found
>>>>> when
>>>>>> trying to give Linux 3.16.0 a final test on a x86_64 MacBookPro under
>>>>>> Ubuntu 14.04 LTS with nouveau as graphics driver.
>>>>>>
>>>>>> I was lazy and just installed the Ubuntu precompiled mainline kernel.
>>>>>> That kernel happens to have CONFIG_DMA_CMA=y set, with a default CMA
>>>>>> (contiguous memory allocator) size of 64 MB. Older Ubuntu kernels
>>>>>> weren't compiled with CMA, so i only observed this on 3.16, but
>>>>>> previous kernels would likely be affected too.
>>>>>>
>>>>>> After a few minutes of regular desktop use like switching workspaces,
>>>>>> scrolling text in a terminal window, Firefox with multiple tabs open,
>>>>>> Thunderbird etc. (tested with KDE/Kwin, with/without desktop
>>>>>> composition), i get chunky desktop updates, then multi-second
>>>>> freezes,
>>>>>> after a few minutes the desktop hangs for over a minute on almost any
>>>>>> GUI action like switching windows etc. --> Unuseable.
>>>>>>
>>>>>> ftrace'ing shows the culprit being this callchain (typical good/bad
>>>>>> example ftrace snippets at the end of this mail):
>>>>>>
>>>>>> ...ttm dma coherent memory allocations, e.g., from
>>>>>> __ttm_dma_alloc_page() ... --> dma_alloc_coherent() --> platform
>>>>>> specific hooks ... -> dma_generic_alloc_coherent() [on x86_64] -->
>>>>>> dma_alloc_from_contiguous()
>>>>>>
>>>>>> dma_alloc_from_contiguous() is a no-op without CONFIG_DMA_CMA, or
>>>>> when
>>>>>> the machine is booted with kernel boot cmdline parameter "cma=0", so
>>>>>> it triggers the fast alloc_pages_node() fallback at least on x86_64.
>>>>>>
>>>>>> With CMA, this function becomes progressively more slow with every
>>>>>> minute of desktop use, e.g., runtimes going up from < 0.3 usecs to
>>>>>> hundreds or thousands of microseconds (before it gives up and
>>>>>> alloc_pages_node() fallback is used), so this causes the
>>>>>> multi-second/minute hangs of the desktop.
>>>>>>
>>>>>> So it seems ttm memory allocations quickly fragment and/or exhaust
>>>>> the
>>>>>> CMA memory area, and dma_alloc_from_contiguous() tries very hard to
>>>>>> find a fitting hole big enough to satisfy allocations with a retry
>>>>>> loop (see
>>>>>>
>>>>> https://urldefense.proofpoint.com/v1/url?u=http://lxr.free-electrons.com/source/drivers/base/dma-contiguous.c%23L339&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=6cy0madhpBCtEyOKu95ucqhzU%2FjAHPP7ODVTc47UYQs%3D%0A&s=42356aad2ff181236f4704283dc058fdd7b7e213cdea7378665094b35ee0dfdf)
>>>>>
>>>>>> that takes forever.
>>>> I am curious why it does not end up using the pool. As in use the
>>>> TTM DMA pool to pick pages instead of allocating (and freeing) new
>>>> ones?
>>>>
>>>>>> This is not good, also not for other devices which actually need a
>>>>>> non-fragmented CMA for DMA, so what to do? I doubt most current gpus
>>>>>> still need physically contiguous dma memory, maybe with exception of
>>>>>> some embedded gpus?
>>>> Oh. If I understood you correctly - the CMA ends up giving huge
>>>> chunks of contiguous area. But if the sizes are 4kb I wonder why it
>>>> would do that?
>>>>
>>>> The modern GPUs on x86 can deal with scatter gather and as you
>>>> surmise don't need contiguous physical contiguous areas.
>>>>>> My naive approach would be to add a new gfp_t flag a la
>>>>>> ___GFP_AVOIDCMA, and make callers of dma_alloc_from_contiguous()
>>>>>> refrain from doing so if they have some fallback for getting memory.
>>>>>> And then add that flag to ttm's ttm_dma_populate() gfp_flags, e.g.,
>>>>>> around here:
>>>>>>
>>>>> https://urldefense.proofpoint.com/v1/url?u=http://lxr.free-electrons.com/source/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c%23L884&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=6cy0madhpBCtEyOKu95ucqhzU%2FjAHPP7ODVTc47UYQs%3D%0A&s=0c2a37c8bac57e0ab7333a9580eb5114e09566d1d34ab43be7a80de8316bdcdd
>>>>>
>>>>>> However i'm not familiar enough with memory management, so likely
>>>>>> greater minds here have much better ideas on how to deal with this?
>>>>>>
>>>> That is a bit of hack to deal with CMA being slow.
>>>>
>>>> Hmm. Let's first figure out why TTM DMA pool is not reusing pages.
>>>>>> thanks,
>>>>>> -mario
>>>>>>
>>>>>> Typical snippet from an example trace of a badly stalling desktop
>>>>> with
>>>>>> CMA (alloc_pages_node() fallback may have been missing in this traces
>>>>>> ftrace_filter settings):
>>>>>>
>>>>>> 1)               |                          ttm_dma_pool_get_pages
>>>>>> [ttm]() {
>>>>>>    1)               | ttm_dma_page_pool_fill_locked [ttm]() {
>>>>>>    1)               | ttm_dma_pool_alloc_new_pages [ttm]() {
>>>>>>    1)               | __ttm_dma_alloc_page [ttm]() {
>>>>>>    1)               | dma_generic_alloc_coherent() {
>>>>>>    1) ! 1873.071 us | dma_alloc_from_contiguous();
>>>>>>    1) ! 1874.292 us |                                  }
>>>>>>    1) ! 1875.400 us |                                }
>>>>>>    1)               | __ttm_dma_alloc_page [ttm]() {
>>>>>>    1)               | dma_generic_alloc_coherent() {
>>>>>>    1) ! 1868.372 us | dma_alloc_from_contiguous();
>>>>>>    1) ! 1869.586 us |                                  }
>>>>>>    1) ! 1870.053 us |                                }
>>>>>>    1)               | __ttm_dma_alloc_page [ttm]() {
>>>>>>    1)               | dma_generic_alloc_coherent() {
>>>>>>    1) ! 1871.085 us | dma_alloc_from_contiguous();
>>>>>>    1) ! 1872.240 us |                                  }
>>>>>>    1) ! 1872.669 us |                                }
>>>>>>    1)               | __ttm_dma_alloc_page [ttm]() {
>>>>>>    1)               | dma_generic_alloc_coherent() {
>>>>>>    1) ! 1888.934 us | dma_alloc_from_contiguous();
>>>>>>    1) ! 1890.179 us |                                  }
>>>>>>    1) ! 1890.608 us |                                }
>>>>>>    1)   0.048 us    | ttm_set_pages_caching [ttm]();
>>>>>>    1) ! 7511.000 us |                              }
>>>>>>    1) ! 7511.306 us |                            }
>>>>>>    1) ! 7511.623 us |                          }
>>>>>>
>>>>>> The good case (with cma=0 kernel cmdline, so
>>>>>> dma_alloc_from_contiguous() no-ops,)
>>>>>>
>>>>>> 0)               |                          ttm_dma_pool_get_pages
>>>>>> [ttm]() {
>>>>>>    0)               | ttm_dma_page_pool_fill_locked [ttm]() {
>>>>>>    0)               | ttm_dma_pool_alloc_new_pages [ttm]() {
>>>>>>    0)               | __ttm_dma_alloc_page [ttm]() {
>>>>>>    0)               | dma_generic_alloc_coherent() {
>>>>>>    0)   0.171 us    | dma_alloc_from_contiguous();
>>>>>>    0)   0.849 us    | __alloc_pages_nodemask();
>>>>>>    0)   3.029 us    |                                  }
>>>>>>    0)   3.882 us    |                                }
>>>>>>    0)               | __ttm_dma_alloc_page [ttm]() {
>>>>>>    0)               | dma_generic_alloc_coherent() {
>>>>>>    0)   0.037 us    | dma_alloc_from_contiguous();
>>>>>>    0)   0.163 us    | __alloc_pages_nodemask();
>>>>>>    0)   1.408 us    |                                  }
>>>>>>    0)   1.719 us    |                                }
>>>>>>    0)               | __ttm_dma_alloc_page [ttm]() {
>>>>>>    0)               | dma_generic_alloc_coherent() {
>>>>>>    0)   0.035 us    | dma_alloc_from_contiguous();
>>>>>>    0)   0.153 us    | __alloc_pages_nodemask();
>>>>>>    0)   1.454 us    |                                  }
>>>>>>    0)   1.720 us    |                                }
>>>>>>    0)               | __ttm_dma_alloc_page [ttm]() {
>>>>>>    0)               | dma_generic_alloc_coherent() {
>>>>>>    0)   0.036 us    | dma_alloc_from_contiguous();
>>>>>>    0)   0.112 us    | __alloc_pages_nodemask();
>>>>>>    0)   1.211 us    |                                  }
>>>>>>    0)   1.541 us    |                                }
>>>>>>    0)   0.035 us    | ttm_set_pages_caching [ttm]();
>>>>>>    0) + 10.902 us   |                              }
>>>>>>    0) + 11.577 us   |                            }
>>>>>>    0) + 11.988 us   |                          }
>>>>>>
>>>>>> _______________________________________________
>>>>>> dri-devel mailing list
>>>>>> dri-devel@...ts.freedesktop.org
>>>>>> https://urldefense.proofpoint.com/v1/url?u=http://lists.freedesktop.org/mailman/listinfo/dri-devel&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=6cy0madhpBCtEyOKu95ucqhzU%2FjAHPP7ODVTc47UYQs%3D%0A&s=d2636419e1f7f56c0d270e29ffe6ab6c6e29249876a578d70d973058f9411831
>>>>>>
>> _______________________________________________
>> dri-devel mailing list
>> dri-devel@...ts.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/dri-devel

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists