Regression due to "mm: fix-up zone present pages"

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi all,

Our QA noticed a regression in one of our i915/GEM testcases in 3.7:

https://bugs.freedesktop.org/show_bug.cgi?id=56859

Direct link to dmesg of the machine:
https://bugs.freedesktop.org/attachment.cgi?id=70052 Note that the
machine is 32bit, which seems to be important since Chris Wilson
confirmed the bug on his 32bit Sandybridge machine, whereas mine here
with a 64bit kernel works flawlessly.

The testcase is gem_tiled_swapping:

http://cgit.freedesktop.org/xorg/app/intel-gpu-tools/tree/tests/gem_tiled_swapping.c

Quick high-level description of the workload:

It allocates a working set larger than available memory, then fills it
by writing it through the gpu gart (required to get a linear view of
tiled buffers) and afterwards reads it to check whether anything got
corrupted. Since the working set is too large to fit into ram, this
will force all buffers through swap. We've written this testcase to
exercise the reswizzle swapin path since some platforms have a tiling
layout depending upon physical pfn (awesome feature btw), but not snb.
So within the kernel this workload simply grabs the backing storage
from shmemfs with shmem_read_mapping_page_gfp and then binds them into
the gpu pagetables (the GTT). This happens in the i915_gem_fault
fucntion. Unbinding in this workload happens either directly (if the
gem code can't get enough memory) or through our shrinker
(i915_gem_inactive_shrink). Swapout is then left to shmemfs to handle.
All the above stuff is in drivers/gpu/drm/i915_gem.c

Testcase fails because it detects a mismatch between what has been
written and what has been read back.

Our qa people bisected the regression to

commit 7f1290f2f2a4d2c3f1b7ce8e87256e052ca23125
Author: Jianguo Wu <wujianguo@xxxxxxxxxx>
Date:   Mon Oct 8 16:33:06 2012 -0700

    mm: fix-up zone present pages

and confirmed the revert on top of the latest drm-intel-nightly branch
(which is based on top of 3.7-rc2 and contains the -next stuff for
3.8). They've also tested the for-QA branch which had latest Linus
upstream merged in, which did not fix the problem. For reference the
intel trees are at (but I don't think it matters really that it's not
plain upstream, nothing really changed in the relevant i915/gem paths
compared to upstream):

http://cgit.freedesktop.org/~danvet/drm-intel

I have no idea how that early boot zone init fix could even corrupt
swapping in such a fashion, so ideas highly welcome. QA people are
cc'ed, and hopefully I haven't missed anyone else on the cc list.

Yours, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
dri-devel mailing list
dri-devel@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/dri-devel


[Index of Archives]     [Linux DRI Users]     [Linux Intel Graphics]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]
  Powered by Linux