<html><head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> </head> <body> <div class="moz-cite-prefix">Am 2020-05-05 um 1:29 p.m. schrieb Felix Kuehling: </div> <blockquote type="cite" cite="mid:388b628f-cfd3-b8e0-d43e-ed4cd280fee7@amd.com"> <div class="moz-cite-prefix">Am 2020-05-05 um 11:19 a.m. schrieb Christian König: </div> <blockquote type="cite" cite="mid:ca18ec07-a6f4-292c-c800-48dbf9c8854f@amd.com"> <div class="moz-cite-prefix">Am 05.05.20 um 16:58 schrieb Felix Kuehling: </div> <blockquote type="cite" cite="mid:170627d3-b5f8-edf5-d809-dd197c4e2018@amd.com"> <div class="moz-cite-prefix">Am 2020-05-05 um 3:47 a.m. schrieb Christian König: </div> <blockquote type="cite" cite="mid:6c4b3eb0-40b5-8c3c-dc5e-38b622864e12@gmail.com"> <div class="moz-cite-prefix">Just to reply here once more, this patch is a clear NAK. </div> </blockquote> Agreed. But see below. I don't think all is well. <blockquote type="cite" cite="mid:6c4b3eb0-40b5-8c3c-dc5e-38b622864e12@gmail.com"> <div class="moz-cite-prefix"> The references are grabbed in the call path of drm_gem_prime_export() and dropped again in drm_gem_dmabuf_release(). So they are perfectly balanced as far as I can see. </div> </blockquote> That is true for the GEM object references. But I believe there is still a problem with the TTM BO references. As far as I can tell amdgpu_bo_unref can free the TTM BO while there are still references to the GEM object from DMA buf exports. I think that's a fundamental problem with how we have two reference counts for the same physical object (TTM BO and the embedded GEM BO). </blockquote> Completely agree, I also mentioned that problem during my talk on FOSDEM. But calling amdgpu_bo_unref() to often is a bug in itself. </blockquote> That's not the problem here. </blockquote> I think we may have a fundamental issue with how we release memory in KFD using amdgpu_bo_unref. That worked OK before the GEM object was embedded in the TTM BO, and it continued working for memory that's not shared via DMABufs. Alejandro is testing a patch that changes KFD to release its BOs with drm_gem_object_put_unlocked instead of amdgpu_bo_unref. Ignore my musings below. I think they were based on incorrect assumptions that it's OK to release GEM BOs with amdgpu_bo_unref. Regards, Felix <blockquote type="cite" cite="mid:388b628f-cfd3-b8e0-d43e-ed4cd280fee7@amd.com"> <blockquote type="cite" cite="mid:ca18ec07-a6f4-292c-c800-48dbf9c8854f@amd.com"> What we could probably do to detect this is adding a BUG_ON() in TTMs release function and checking if the GEM reference count is really dead. </blockquote> The problem is, that we have to guess whether there are still any dmabuf references to the GEM BO. There is no way amdgpu can know that. You can't make amdgpu responsible for keeping a reference to the TTM BO while the GEM BO is still referenced by entities completely out of the control of amdgpu. Another weird thing I see is that amdgpu_gem_free_object calls amdgpu_bo_unref. That implies that the GEM object conceptually holds a reference to the amdpu/TTM BO. But that is not really the case. Amdgpu never takes that reference that GEM is supposed to own. If it did, we would leak all our memory because nobody would ever drop that reference. </blockquote> <blockquote type="cite" cite="mid:388b628f-cfd3-b8e0-d43e-ed4cd280fee7@amd.com"> <blockquote type="cite" cite="mid:ca18ec07-a6f4-292c-c800-48dbf9c8854f@amd.com"> <blockquote type="cite" cite="mid:170627d3-b5f8-edf5-d809-dd197c4e2018@amd.com">I think the correct solution is for amdgpu_bo_ref/unref to delegate its reference counting to drm_gem_object_get/put instead of ttm_bo_get/put. The amdgpu BO would hold one token reference to the TTM BO, which it can drop when the GEM BO refcount drops to 0. Finally, the amdgpu BO should only be freed once the TTM BO refcount also becomes 0. </blockquote> Just the other way around, but yes the long term plan should probably be to merge the two. </blockquote> I need a short term solution. Because I have a bug that causes a kernel oops with applications that are valid and correct, as far as I can tell. I'm thinking a solution that doesn't require major changes to the way TTM and GEM interact would put amdgpu in charge of coordinating the two. Unfortunately that would mean adding a third reference count in amdgpu_bo, in addition to the ones in TTM and GEM. The amdgpu BO would hold one token reference to each of the GEM and TTM BO. When amdgpu refcount goes to 0 it releases that GEM BO token reference. When the GEM BO refcount goes to 0, we get a callback amdgpu_gem_object_free. There we can drop the token reference to the TTM BO. Once the TTM BO reference goes to 0 we free the memory. Does this sound feasible? Regards, Felix <blockquote type="cite" cite="mid:ca18ec07-a6f4-292c-c800-48dbf9c8854f@amd.com"> The difficult is currently we have a mismatch what locks could be taken when we drop the references. Regards, Christian. <blockquote type="cite" cite="mid:170627d3-b5f8-edf5-d809-dd197c4e2018@amd.com"> Regards, Felix <blockquote type="cite" cite="mid:6c4b3eb0-40b5-8c3c-dc5e-38b622864e12@gmail.com"> <div class="moz-cite-prefix"> Regards, Christian. Am 01.05.20 um 16:44 schrieb Felix Kuehling: </div> </blockquote> <blockquote type="cite" cite="mid:6c4b3eb0-40b5-8c3c-dc5e-38b622864e12@gmail.com"> <blockquote type="cite" cite="mid:551849ec-bc90-0bd2-d46c-f6d8e5c0fee0@amd.com"> [dropping my gmail address] We saw this backtrace showing the call chain while investigating a kernel oops caused by this issue on the DKMS branch with the KFD IPC API. It happens after a dma-buf file is released with fput: <pre>[ 1255.049330] BUG: kernel NULL pointer dereference, address: 000000000000051e [ 1255.049727] #PF: supervisor read access in kernel mode [ 1255.050092] #PF: error_code(0x0000) - not-present page [ 1255.050416] PGD 0 P4D 0 [ 1255.050736] Oops: 0000 [#1] SMP PTI [ 1255.051060] CPU: 27 PID: 2292 Comm: kworker/27:2 Tainted: G OE 5.3.0-46-generic #38~18.04.1-Ubuntu [ 1255.051400] Hardware name: Supermicro SYS-4029GP-TRT2/X11DPG-OT-CPU, BIOS 3.0a 02/26/2019 [ 1255.051752] Workqueue: events delayed_fput [ 1255.052111] RIP: 0010:drm_gem_object_put_unlocked+0x1c/0x70 [drm] [ 1255.052465] Code: 4d 80 c8 ee 0f 0b eb d8 66 0f 1f 44 00 00 0f 1f 44 00 00 48 85 ff 74 34 55 48 89 e5 41 54 53 48 89 fb 48 8b 7f 08 48 8b 47 20 <48> 83 b8 a0 00 00 00 00 74 1a 4c 8d 67 68 48 89 df 4c 89 e6 e8 9b [ 1255.053224] RSP: 0018:ffffb4b62035fdc8 EFLAGS: 00010286 [ 1255.053613] RAX: 000000000000047e RBX: ffff9f2add197850 RCX: 0000000000000000 [ 1255.054032] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff9f2aa2548aa0 [ 1255.054440] RBP: ffffb4b62035fdd8 R08: 0000000000000000 R09: 0000000000000000 [ 1255.054860] R10: 0000000000000010 R11: ffff9f2a4b1cc310 R12: 0000000000080005 [ 1255.055268] R13: ffff9f2a4b1cc310 R14: ffff9f4e369161e0 R15: ffff9f2a1b2f9080 [ 1255.055674] FS: 0000000000000000(0000) GS:ffff9f4e3f740000(0000) knlGS:0000000000000000 [ 1255.056087] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 1255.056501] CR2: 000000000000051e CR3: 00000002df00a004 CR4: 00000000007606e0 [ 1255.056923] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 1255.057345] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 1255.057763] PKRU: 55555554 [ 1255.058179] Call Trace: [ 1255.058603] drm_gem_dmabuf_release+0x1a/0x30 [drm] [ 1255.059025] dma_buf_release+0x56/0x130 [ 1255.059443] __fput+0xc6/0x260 [ 1255.059856] delayed_fput+0x20/0x30 [ 1255.060272] process_one_work+0x1fd/0x3f0 [ 1255.060686] worker_thread+0x34/0x410 [ 1255.061099] kthread+0x121/0x140 [ 1255.061510] ? process_one_work+0x3f0/0x3f0 [ 1255.061923] ? kthread_park+0xb0/0xb0 [ 1255.062336] ret_from_fork+0x35/0x40 </pre> drm_gem_object_put_unlocked calls drm_gem_object_free when the obj->refcount reaches 0. From there it calls dev->driver->gem_free_object_unlocked, which is amdgpu_gem_object_free in amdgpu. Regards, Felix <div class="moz-cite-prefix">Am 2020-05-01 um 10:29 a.m. schrieb Christian König: </div> <blockquote type="cite" cite="mid:70e5d202-34ed-532f-e6b6-c195a3effad3@gmail.com">Am 01.05.20 um 16:21 schrieb Felix Kuehling: <blockquote type="cite">From: Felix Kuehling <a class="moz-txt-link-rfc2396E" href="mailto:felix.kuehling@gmail.com" moz-do-not-send="true"><felix.kuehling@gmail.com></a> That reference gets dropped when the the dma-buf is freed. Not incrementing the refcount can lead to use-after-free errors. Signed-off-by: Felix Kuehling <a class="moz-txt-link-rfc2396E" href="mailto:felix.kuehling@gmail.com" moz-do-not-send="true"><felix.kuehling@gmail.com></a> --- drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c index ffeb20f11c07..a0f9b3ef4aad 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c @@ -398,8 +398,15 @@ struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj, return ERR_PTR(-EPERM); buf = drm_gem_prime_export(gobj, flags); - if (!IS_ERR(buf)) + if (!IS_ERR(buf)) { buf->ops = &amdgpu_dmabuf_ops; + /* GEM needs a reference to the underlying object + * that gets dropped when the dma-buf is released, + * through the amdgpu_gem_object_free callback + * from drm_gem_object_put_unlocked. + */ + amdgpu_bo_ref(bo); + } </blockquote> Of hand that doesn't sounds correct to me. Why should the exported bo be closed through amdgpu_gem_object_free()? Regards, Christian. <blockquote type="cite"> return buf; } </blockquote> </blockquote> <fieldset class="mimeAttachmentHeader"></fieldset> <pre class="moz-quote-pre" wrap="">_______________________________________________ amd-gfx mailing list <a class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org">amd-gfx@lists.freedesktop.org</a> <a class="moz-txt-link-freetext" href="https://lists.freedesktop.org/mailman/listinfo/amd-gfx">https://lists.freedesktop.org/mailman/listinfo/amd-gfx</a> </pre> </blockquote> </blockquote> </blockquote> </blockquote> </blockquote> </body> </html>