回复: 回复: [PATCH 1/2] drm/amdgpu: race issue when jobs on 2 ring timeout
Chen, Horace
Horace.Chen at amd.com
Tue Jan 19 17:09:16 UTC 2021
[AMD Official Use Only - Internal Distribution Only]
OK, I understand. You mean one device in the hive may be locked up independently without locking up the whole hive.
It could happen, I'll change my code.
Thanks & Regards,
Horace.
________________________________
发件人: Grodzovsky, Andrey <Andrey.Grodzovsky at amd.com>
发送时间: 2021年1月20日 0:58
收件人: Chen, Horace <Horace.Chen at amd.com>; amd-gfx at lists.freedesktop.org <amd-gfx at lists.freedesktop.org>
抄送: Quan, Evan <Evan.Quan at amd.com>; Tuikov, Luben <Luben.Tuikov at amd.com>; Koenig, Christian <Christian.Koenig at amd.com>; Deucher, Alexander <Alexander.Deucher at amd.com>; Xiao, Jack <Jack.Xiao at amd.com>; Zhang, Hawking <Hawking.Zhang at amd.com>; Liu, Monk <Monk.Liu at amd.com>; Xu, Feifei <Feifei.Xu at amd.com>; Wang, Kevin(Yang) <Kevin1.Wang at amd.com>; Xiaojie Yuan <xiaojie.yuan at amd.com>
主题: Re: 回复: [PATCH 1/2] drm/amdgpu: race issue when jobs on 2 ring timeout
On 1/19/21 11:39 AM, Chen, Horace wrote:
[AMD Official Use Only - Internal Distribution Only]
Hi Andrey,
I think the list in the XGMI hive won't be break in the middle if we lock the device before we change the list. Because if 2 devices in 1 hive went into the function, it will follow the same sequence to lock the devices. So one of them will definately break at the first device. I add iterate devices here is just to lock all device in the hive since we will change the device sequence in the hive soon after.
I didn't mean break in a sense of breaking the list itself, I just meant the literal 'break' instruction
to terminate the iteration once you failed to lock a particular device.
The reason to break the interation in the middle is that the list is changed during the iteration without taking any lock. It is quite bad since I'm fixing one of this issue. And for XGMI hive, there are 2 locks protecting the list, one is the device lock I changed here, the other one is in front of my change, there is a hive->lock to protect the hive.
Even the bad thing really happened, I think moving back through the list is also very dengerous since we don't know what the list finally be, Unless we stack the devices we have iterated through a mirrored list. That can be a big change.
Not sure we are on the same page, my concern is let's sat your XGMI hive consists of 2 devices, you manged to call successfully do
amdgpu_device_lock_adev for dev1 but then failed for dev2, in this case you will bail out without releasing dev1, no ?
Andrey
I'm ok to seperate the locking in amdgpu_device_lock_adev here, I'll do some test and update the code later.
Thanks & Regards,
Horace.
________________________________
发件人: Grodzovsky, Andrey <Andrey.Grodzovsky at amd.com><mailto:Andrey.Grodzovsky at amd.com>
发送时间: 2021年1月19日 22:33
收件人: Chen, Horace <Horace.Chen at amd.com><mailto:Horace.Chen at amd.com>; amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org> <amd-gfx at lists.freedesktop.org><mailto:amd-gfx at lists.freedesktop.org>
抄送: Quan, Evan <Evan.Quan at amd.com><mailto:Evan.Quan at amd.com>; Tuikov, Luben <Luben.Tuikov at amd.com><mailto:Luben.Tuikov at amd.com>; Koenig, Christian <Christian.Koenig at amd.com><mailto:Christian.Koenig at amd.com>; Deucher, Alexander <Alexander.Deucher at amd.com><mailto:Alexander.Deucher at amd.com>; Xiao, Jack <Jack.Xiao at amd.com><mailto:Jack.Xiao at amd.com>; Zhang, Hawking <Hawking.Zhang at amd.com><mailto:Hawking.Zhang at amd.com>; Liu, Monk <Monk.Liu at amd.com><mailto:Monk.Liu at amd.com>; Xu, Feifei <Feifei.Xu at amd.com><mailto:Feifei.Xu at amd.com>; Wang, Kevin(Yang) <Kevin1.Wang at amd.com><mailto:Kevin1.Wang at amd.com>; Xiaojie Yuan <xiaojie.yuan at amd.com><mailto:xiaojie.yuan at amd.com>
主题: Re: [PATCH 1/2] drm/amdgpu: race issue when jobs on 2 ring timeout
On 1/19/21 7:22 AM, Horace Chen wrote:
> Fix a racing issue when jobs on 2 rings timeout simultaneously.
>
> If 2 rings timed out at the same time, the amdgpu_device_gpu_recover
> will be reentered. Then the adev->gmc.xgmi.head will be grabbed
> by 2 local linked list, which may cause wild pointer issue in
> iterating.
>
> lock the device earily to prevent the node be added to 2 different
> lists.
>
> Signed-off-by: Horace Chen <horace.chen at amd.com><mailto:horace.chen at amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 42 +++++++++++++++-------
> 1 file changed, 30 insertions(+), 12 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index 4d434803fb49..9574da3abc32 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -4540,6 +4540,7 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
> int i, r = 0;
> bool need_emergency_restart = false;
> bool audio_suspended = false;
> + bool get_dev_lock = false;
>
> /*
> * Special case: RAS triggered and full reset isn't supported
> @@ -4582,28 +4583,45 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
> * Build list of devices to reset.
> * In case we are in XGMI hive mode, resort the device list
> * to put adev in the 1st position.
> + *
> + * lock the device before we try to operate the linked list
> + * if didn't get the device lock, don't touch the linked list since
> + * others may iterating it.
> */
> INIT_LIST_HEAD(&device_list);
> if (adev->gmc.xgmi.num_physical_nodes > 1) {
> if (!hive)
> return -ENODEV;
> - if (!list_is_first(&adev->gmc.xgmi.head, &hive->device_list))
> - list_rotate_to_front(&adev->gmc.xgmi.head, &hive->device_list);
> - device_list_handle = &hive->device_list;
> +
> + list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
> + get_dev_lock = amdgpu_device_lock_adev(tmp_adev, hive);
> + if (!get_dev_lock)
> + break;
What about unlocking back all the devices you already locked if the break
happens in the middle of the iteration ?
Note that at skip_recovery: we don't do it. BTW, i see this issue is already in
the current code.
Also, maybe now it's better to separate the actual locking in
amdgpu_device_lock_adev
from the other stuff going on there since I don't think you would wont to toggle
stuff
like adev->mp1_state back and forth and also the function name is not
descriptive of
the other stuff going on there anyway.
Andrey
> + }
> + if (get_dev_lock) {
> + if (!list_is_first(&adev->gmc.xgmi.head, &hive->device_list))
> + list_rotate_to_front(&adev->gmc.xgmi.head, &hive->device_list);
> + device_list_handle = &hive->device_list;
> + }
> } else {
> - list_add_tail(&adev->gmc.xgmi.head, &device_list);
> - device_list_handle = &device_list;
> + get_dev_lock = amdgpu_device_lock_adev(adev, hive);
> + tmp_adev = adev;
> + if (get_dev_lock) {
> + list_add_tail(&adev->gmc.xgmi.head, &device_list);
> + device_list_handle = &device_list;
> + }
> + }
> +
> + if (!get_dev_lock) {
> + dev_info(tmp_adev->dev, "Bailing on TDR for s_job:%llx, as another already in progress",
> + job ? job->base.id : -1);
> + r = 0;
> + /* even we skipped this reset, still need to set the job to guilty */
> + goto skip_recovery;
> }
>
> /* block all schedulers and reset given job's ring */
> list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
> - if (!amdgpu_device_lock_adev(tmp_adev, hive)) {
> - dev_info(tmp_adev->dev, "Bailing on TDR for s_job:%llx, as another already in progress",
> - job ? job->base.id : -1);
> - r = 0;
> - goto skip_recovery;
> - }
> -
> /*
> * Try to put the audio codec into suspend state
> * before gpu reset started.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20210119/32fd3612/attachment-0001.htm>
More information about the amd-gfx
mailing list