On 2022-04-08 21:28, Shuotao Xu wrote:
On Apr 8, 2022, at 11:28 PM, Andrey Grodzovsky <andrey.grodzovsky@xxxxxxx> wrote:
[Some people who received this message don't often get email from andrey.grodzovsky@xxxxxxx. Learn why this is important at http://aka.ms/LearnAboutSenderIdentification.]
On 2022-04-08 04:45, Shuotao Xu wrote:
Adding PCIe Hotplug Support for AMDKFD: the support of hot-plug of GPU
devices can open doors for many advanced applications in data center
in the next few years, such as for GPU resource
disaggregation. Current AMDKFD does not support hotplug out b/o the
following reasons:
1. During PCIe removal, decrement KFD lock which was incremented at
the beginning of hw fini; otherwise kfd_open later is going to
fail.
I assumed you read my comment last time, still you do same approach.
More in details bellow
Aha, I like your fix:) I was not familiar with drm APIs so just only half understood your comment last time.
BTW, I tried hot-plugging out a GPU when rocm application is still running.
From dmesg, application is still trying to access the removed kfd device, and are met with some errors.
Application us supposed to keep running, it holds the drm_device
reference as long as it has an open
FD to the device and final cleanup will come only after the app will die
thus releasing the FD and the last
drm_device reference.
Application would hang and not exiting in this case.
For graphic apps what i usually see is a crash because of sigsev when
the app tries to access
an unmapped MMIO region on the device. I haven't tested for compute
stack and so there might
be something I haven't covered. Hang could mean for example waiting on a
fence which is not being
signaled - please provide full dmesg from this case.
Do you have any good suggestions on how to fix it down the line? (HIP runtime/libhsakmt or driver)
[64036.631333] amdgpu: amdgpu_vm_bo_update failed
[64036.631702] amdgpu: validate_invalid_user_pages: update PTE failed
[64036.640754] amdgpu: amdgpu_vm_bo_update failed
[64036.641120] amdgpu: validate_invalid_user_pages: update PTE failed
[64036.650394] amdgpu: amdgpu_vm_bo_update failed
[64036.650765] amdgpu: validate_invalid_user_pages: update PTE failed
This just probably means trying to update PTEs after the physical device
is gone - we usually avoid this by
first trying to do all HW shutdowns early before PCI remove completion
but when it's really tricky by
protecting HW access sections with drm_dev_enter/exit scope.
For this particular error it would be the best to flush
info->restore_userptr_work before the end of
amdgpu_pci_remove (rejecting new process creation and calling
cancel_delayed_work_sync(&process_info->restore_userptr_work) for all
running processes)
somewhere in amdgpu_pci_remove.
Andrey
Really appreciate your help!
Best,
Shuotao
2. Remove redudant p2p/io links in sysfs when device is hotplugged
out.
3. New kfd node_id is not properly assigned after a new device is
added after a gpu is hotplugged out in a system. libhsakmt will
find this anomaly, (i.e. node_from != <dev node id> in iolinks),
when taking a topology_snapshot, thus returns fault to the rocm
stack.
-- This patch fixes issue 1; another patch by Mukul fixes issues 2&3.
-- Tested on a 4-GPU MI100 gpu nodes with kernel 5.13.0-kfd; kernel
5.16.0-kfd is unstable out of box for MI100.
---
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 5 +++++
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 7 +++++++
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 1 +
drivers/gpu/drm/amd/amdkfd/kfd_device.c | 13 +++++++++++++
4 files changed, 26 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
index c18c4be1e4ac..d50011bdb5c4 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
@@ -213,6 +213,11 @@ int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm)
return r;
}
+int amdgpu_amdkfd_resume_processes(void)
+{
+ return kgd2kfd_resume_processes();
+}
+
int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev)
{
int r = 0;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index f8b9f27adcf5..803306e011c3 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -140,6 +140,7 @@ void amdgpu_amdkfd_fini(void);
void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm);
int amdgpu_amdkfd_resume_iommu(struct amdgpu_device *adev);
int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm);
+int amdgpu_amdkfd_resume_processes(void);
void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
const void *ih_ring_entry);
void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev);
@@ -347,6 +348,7 @@ void kgd2kfd_device_exit(struct kfd_dev *kfd);
void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm);
int kgd2kfd_resume_iommu(struct kfd_dev *kfd);
int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm);
+int kgd2kfd_resume_processes(void);
int kgd2kfd_pre_reset(struct kfd_dev *kfd);
int kgd2kfd_post_reset(struct kfd_dev *kfd);
void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
@@ -393,6 +395,11 @@ static inline int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm)
return 0;
}
+static inline int kgd2kfd_resume_processes(void)
+{
+ return 0;
+}
+
static inline int kgd2kfd_pre_reset(struct kfd_dev *kfd)
{
return 0;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index fa4a9f13c922..5827b65b7489 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -4004,6 +4004,7 @@ void amdgpu_device_fini_hw(struct amdgpu_device *adev)
if (drm_dev_is_unplugged(adev_to_drm(adev)))
amdgpu_device_unmap_mmio(adev);
+ amdgpu_amdkfd_resume_processes();
}
void amdgpu_device_fini_sw(struct amdgpu_device *adev)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index 62aa6c9d5123..ef05aae9255e 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -714,6 +714,19 @@ int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm)
return ret;
}
+/* for non-runtime resume only */
+int kgd2kfd_resume_processes(void)
+{
+ int count;
+
+ count = atomic_dec_return(&kfd_locked);
+ WARN_ONCE(count < 0, "KFD suspend / resume ref. error");
+ if (count == 0)
+ return kfd_resume_all_processes();
+
+ return 0;
+}
It doesn't make sense to me to just increment kfd_locked in
kgd2kfd_suspend to only decrement it again a few functions down the
road.
I suggest this instead - you only incrmemnt if not during PCI remove
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index 1c2cf3a33c1f..7754f77248a4 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -952,11 +952,12 @@ bool kfd_is_locked(void)
void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
{
+
if (!kfd->init_complete)
return;
/* for runtime suspend, skip locking kfd */
- if (!run_pm) {
+ if (!run_pm && !drm_dev_is_unplugged(kfd->ddev)) {
/* For first KFD device suspend all the KFD processes */
if (atomic_inc_return(&kfd_locked) == 1)
kfd_suspend_all_processes();
Andrey
+
int kgd2kfd_resume_iommu(struct kfd_dev *kfd)
{
int err = 0;
