Re: [PATCH v2 1/2] drm/ttm: Only allocate huge pages with new flag TTM_PAGE_FLAG_TRANSHUGE

From: Christian KÃnig
Date: Sun Apr 29 2018 - 03:03:18 EST

Am 29.04.2018 um 01:02 schrieb Michel DÃnzer:
On 2018-04-28 06:30 PM, Ilia Mirkin wrote:
On Fri, Apr 27, 2018 at 9:08 AM, Michel DÃnzer <michel@xxxxxxxxxxx> wrote:
From: Michel DÃnzer <michel.daenzer@xxxxxxx>

Previously, TTM would always (with CONFIG_TRANSPARENT_HUGEPAGE enabled)
try to allocate huge pages. However, not all drivers can take advantage
of huge pages, but they would incur the overhead for allocating and
freeing them anyway.

Now, drivers which can take advantage of huge pages need to set the new
flag TTM_PAGE_FLAG_TRANSHUGE to get them. Drivers not setting this flag
no longer incur any overhead for allocating or freeing huge pages.

* Also guard swapping of consecutive pages in ttm_get_pages
* Reword commit log, hopefully clearer now

Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Michel DÃnzer <michel.daenzer@xxxxxxx>
Both I and lots of other people, based on reports, are still seeing
plenty of issues with this as late as 4.16.4.
"lots of other people", "plenty of issues" sounds a bit exaggerated from
what I've seen. FWIW, while I did see the original messages myself, I
haven't seen any since Christian's original fix (see below), neither
with amdgpu nor radeon, even before this patch you followed up to.

Admittedly I'm on nouveau, but others have reported issues with
radeon/amdgpu as well. It's been going on since the feature was merged
in v4.15, with what seems like little investigation from the authors
introducing the feature.
That's not a fair assessment. See and following

Christian fixed the original issue in
d0bc0c2a31c95002d37c3cc511ffdcab851b3256 "swiotlb: suppress warning when
__GFP_NOWARN is set". Christian did his best to try and get the fix in
before 4.15 final, but for reasons beyond his control, it was delayed
until 4.16-rc1 and then backported to 4.15.5.

Unfortunately, there was an swiotlb regression (not directly related to
Christian's work) shortly after this fix, also in 4.16-rc1, which is now
fixed in 4.17-rc1 and will be backported to 4.16.y.

And that's exactly the reason why I intentionally kept this enabled for all users of the TTM DMA page pool and not put it behind a flag.

This change has surfaced quite a number of bugs in the swiotlb code which could have caused issues before. It's just that those code path where never exercised massively before.

Additional to that using huge pages is beneficial for the MM and CPU TLB (not implemented yet) even when the GPU driver can't make much use of it.

It looks like there's at least one more bug left, but it's not clear yet
when that was introduced, whether it's directly related to Christian's
work, or indeed what the impact is. Let's not get ahead of ourselves.

Well my patches surfaced the problems, but the underlying issues where present even before those changes and I'm very well involved in fixing the underlying issues.

I even considered to just revert the huge page path for the DMA pool allocator, but it's just that the TTM patches seem to work exactly as they are intended. So that doesn't feel like doing the right thing here.

We now have *two* broken releases, v4.15 and v4.16 (anything that
spews error messages and stack traces ad-infinitum in dmesg is, by
definition, broken).
I haven't seen any evidence that there's still an issue in 4.15, is
there any?

Not that I know of, the fix was backported as far as I know.

You're putting this behind a flag now (finally),
I wrote this patch because I realized due to some remark I happened to
see you make this week on IRC that the huge page support in TTM was
enabled for all drivers. Instead of making that kind of remark on IRC,
it would have been more constructive, and more conducive to quick
implementation, to suggest making the feature not active for drivers
which don't need it in a mailing list post.

I have to admit that I'm lacking behind taking care of the amdgpu/radeon user space issues just because of more important stuff to do, but the issues affecting other drivers should be fixed by now.

BTW: The user space problems for amdgpu/radeon seems to come from either the DDX or Glamour.

For example try playing a video user firefox with Glamour enabled and take a look at how much memory we free/allocate.

It's multiple gigabytes for just a few seconds playback, that strongly indicates that we allocate/free a texture for each displayed frame which is quite far from optimal.


At least, please do more research before making this kind of negative

P.S. You might also want to look into whether nouveau really should be
hitting swiotlb in these cases.