drm/i915: Lock llist_del_first() vs llist_del_all()
authorChris Wilson <chris@chris-wilson.co.uk>
Mon, 6 Nov 2017 11:15:08 +0000 (11:15 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Mon, 6 Nov 2017 12:05:02 +0000 (12:05 +0000)
commitf991c492aa55fb1c6834882c5d786d5bb3b25f07
tree7f5cf6e21571b1b51cff8c94531c50df73651516
parent69ea47a5a98b198c9d36fe157a3986748a9e2554
drm/i915: Lock llist_del_first() vs llist_del_all()

An oversight in commit 87701b4b5593 ("drm/i915: Only free the oldest
stale object before a fresh allocation") was that not only do we have to
serialise concurrent users of llist_del_first(), but we also have to
lock llist_del_first() vs llist_del_all().

From llist.h,

 * This can be summarized as follows:
 *
 *           |   add    | del_first |  del_all
 * add       |    -     |     -     |     -
 * del_first |          |     L     |     L
 * del_all   |          |           |     -
 *
 * Where, a particular row's operation can happen concurrently with a column's
 * operation, with "-" being no lock needed, while "L" being lock is needed.

This should hopefully explain:

<4>[   89.287106] general protection fault: 0000 [#1] PREEMPT SMP
<4>[   89.287126] Modules linked in: snd_hda_codec_hdmi snd_hda_codec_realtek snd_hda_codec_generic x86_pkg_temp_thermal intel_powerclamp coretemp i915 crct10dif_pclmul crc32_pclmul ghash_clmulni_intel snd_hda_intel snd_hda_codec snd_hwdep snd_hda_core r8169 mii mei_me mei snd_pcm prime_numbers i2c_hid pinctrl_geminilake pinctrl_intel
<4>[   89.287226] CPU: 2 PID: 23 Comm: ksoftirqd/2 Tainted: G     U          4.14.0-rc8-CI-CI_DRM_3315+ #1
<4>[   89.287247] Hardware name: Intel Corp. Geminilake/GLK RVP2 LP4SD (07), BIOS GELKRVPA.X64.0062.B30.1708222146 08/22/2017
<4>[   89.287270] task: ffff88017ab34ec0 task.stack: ffffc90000128000
<4>[   89.287290] RIP: 0010:llist_add_batch+0x4/0x20
<4>[   89.287301] RSP: 0018:ffffc9000012bdb8 EFLAGS: 00010296
<4>[   89.287314] RAX: ffffffff811017ad RBX: 6e468801a1560000 RCX: ef3e53fceecdeb81
<4>[   89.287330] RDX: 6e468801a1566130 RSI: ffff880103d73d98 RDI: ffff880103d73d98
<4>[   89.287346] RBP: ffffc9000012bdb8 R08: ffff88017ab35780 R09: 0000000000000000
<4>[   89.287361] R10: ffffc9000012bd68 R11: 00000000abb18c3d R12: ffffffffa01369e0
<4>[   89.287377] R13: ffff88017fd1b8f8 R14: ffff88017ab34ec0 R15: 000000000000000a
<4>[   89.287393] FS:  0000000000000000(0000) GS:ffff88017fd00000(0000) knlGS:0000000000000000
<4>[   89.287411] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
<4>[   89.287424] CR2: 00007ff0c0755018 CR3: 000000016df9b000 CR4: 00000000003406e0
<4>[   89.287440] Call Trace:
<4>[   89.287511]  __i915_gem_free_object_rcu+0x20/0x40 [i915]
<4>[   89.287527]  rcu_process_callbacks+0x27a/0x730
<4>[   89.287544]  __do_softirq+0xc0/0x4ae
<4>[   89.287559]  ? smpboot_thread_fn+0x2d/0x280
<4>[   89.287571]  run_ksoftirqd+0x1f/0x70
<4>[   89.287582]  smpboot_thread_fn+0x18a/0x280
<4>[   89.287595]  kthread+0x114/0x150
<4>[   89.287605]  ? sort_range+0x30/0x30
<4>[   89.287615]  ? kthread_create_on_node+0x40/0x40
<4>[   89.287628]  ret_from_fork+0x27/0x40
<4>[   89.287641] Code: 0d 48 83 ea 01 4c 89 c1 48 83 fa ff 74 12 48 23 0c d7 74 ed 48 c1 e2 06 48 0f bd c9 48 8d 04 0a 5d c3 90 90 90 90 90 55 48 89 e5 <48> 8b 0a 48 89 0e 48 89 c8 f0 48 0f b1 3a 48 39 c1 75 ed 48 85
<1>[   89.287774] RIP: llist_add_batch+0x4/0x20 RSP: ffffc9000012bdb8
<4>[   89.287826] ---[ end trace e775d15174d8ae02 ]---

(Lockless lists are only easy (and lockless) when only using
llist_add/llist_del_all!)

Fixes: 87701b4b5593 ("drm/i915: Only free the oldest stale object before
a fresh allocation")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171106111508.11941-1-chris@chris-wilson.co.uk
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
drivers/gpu/drm/i915/i915_gem.c