Re: [EXTERNAL] [PATCH 2/2] drm/amdkfd: Add PCIe Hotplug Support for AMDKFD

Shuotao Xu <shuotaoxu@xxxxxxxxxxxxx> · Fri, 15 Apr 2022 10:12:18 +0000

Hi Andrey,



First I really appreciate the discussion! It helped me understand the driver code greatly. Thank you so much:)
Please see my inline comments. 




On Apr 14, 2022, at 11:13 PM, Andrey Grodzovsky <andrey.grodzovsky@xxxxxxx> wrote:







On 2022-04-14 10:00, Shuotao Xu wrote:








On Apr 14, 2022, at 1:31 AM, Andrey Grodzovsky <andrey.grodzovsky@xxxxxxx> wrote:







On 2022-04-13 12:03, Shuotao Xu wrote:








On Apr 11, 2022, at 11:52 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 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.









Actually I tried kill -7 $pid, and the process exists. The dmesg has some warning though.




[  711.769977] WARNING: CPU: 23 PID: 344 at .../amdgpu-rocm5.0.2/src/amd/amdgpu/amdgpu_object.c:1336 amdgpu_bo_release_notify+0x150/0x160 [amdgpu]
[  711.770528] Modules linked in: amdgpu(OE) amdttm(OE) amd_sched(OE) amdkcl(OE) iommu_v2 nf_conntrack_netlink nfnetlink xfrm_user xfrm_algo xt_addrtype br_netfilter xt_CHECKSUM iptable_mangle xt_MASQUERADE iptable_nat nf_nat xt_conntrack nf_conntrack
 nf_defrag_ipv6 nf_defrag_ipv4 ipt_REJECT nf_reject_ipv4 xt_tcpudp bridge stp llc ebtable_filter ebtables ip6table_filter ip6_tables iptable_filter overlay binfmt_misc intel_rapl_msr i40iw intel_rapl_common skx_edac nfit x86_pkg_temp_thermal intel_powerclamp
 coretemp kvm_intel rpcrdma kvm sunrpc ipmi_ssif ib_umad ib_ipoib rdma_ucm irqbypass rapl joydev acpi_ipmi input_leds intel_cstate ipmi_si ipmi_devintf mei_me mei intel_pch_thermal ipmi_msghandler ioatdma mac_hid lpc_ich dca acpi_power_meter acpi_pad sch_fq_codel
 ib_iser rdma_cm iw_cm ib_cm iscsi_tcp libiscsi_tcp libiscsi scsi_transport_iscsi pci_stub ip_tables x_tables autofs4 btrfs blake2b_generic zstd_compress raid10 raid456 async_raid6_recov async_memcpy async_pq async_xor async_tx xor
[  711.779359]  raid6_pq libcrc32c raid1 raid0 multipath linear mlx5_ib ib_uverbs ib_core ast drm_vram_helper i2c_algo_bit drm_ttm_helper ttm drm_kms_helper syscopyarea crct10dif_pclmul crc32_pclmul ghash_clmulni_intel sysfillrect uas hid_generic
 sysimgblt aesni_intel mlx5_core fb_sys_fops crypto_simd usbhid cryptd drm i40e pci_hyperv_intf usb_storage glue_helper mlxfw hid ahci libahci wmi
[  711.779752] CPU: 23 PID: 344 Comm: kworker/23:1 Tainted: G        W  OE     5.11.0+ #1
[  711.779755] Hardware name: Supermicro SYS-4029GP-TRT2/X11DPG-OT-CPU, BIOS 2.1 08/14/2018
[  711.779756] Workqueue: kfd_process_wq kfd_process_wq_release [amdgpu]
[  711.779955] RIP: 0010:amdgpu_bo_release_notify+0x150/0x160 [amdgpu]
[  711.780141] Code: e8 b5 af 34 f4 e9 1f ff ff ff 48 39 c2 74 07 0f 0b e9 69 ff ff ff 4c 89 e7 e8 3c b4 16 00 e9 5c ff ff ff e8 a2 ce fd f3 eb cf <0f> 0b eb cb e8 d7 02 34 f4 0f 1f 80 00 00 00 00 0f 1f 44 00 00 55
[  711.780143] RSP: 0018:ffffa8100dd67c30 EFLAGS: 00010282
[  711.780145] RAX: 00000000ffffffea RBX: ffff89980e792058 RCX: 0000000000000000
[  711.780147] RDX: 0000000000000000 RSI: ffff89a8f9ad8870 RDI: ffff89a8f9ad8870
[  711.780148] RBP: ffffa8100dd67c50 R08: 0000000000000000 R09: fffffffffff99b18
[  711.780149] R10: ffffa8100dd67bd0 R11: ffffa8100dd67908 R12: ffff89980e792000
[  711.780151] R13: ffff89980e792058 R14: ffff89980e7921bc R15: dead000000000100
[  711.780152] FS:  0000000000000000(0000) GS:ffff89a8f9ac0000(0000) knlGS:0000000000000000
[  711.780154] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[  711.780156] CR2: 00007ffddac6f71f CR3: 00000030bb80a003 CR4: 00000000007706e0
[  711.780157] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[  711.780159] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[  711.780160] PKRU: 55555554
[  711.780161] Call Trace:
[  711.780163]  ttm_bo_release+0x2ae/0x320 [amdttm]
[  711.780169]  amdttm_bo_put+0x30/0x40 [amdttm]
[  711.780357]  amdgpu_bo_unref+0x1e/0x30 [amdgpu]
[  711.780543]  amdgpu_gem_object_free+0x8b/0x160 [amdgpu]
[  711.781119]  drm_gem_object_free+0x1d/0x30 [drm]
[  711.781489]  amdgpu_amdkfd_gpuvm_free_memory_of_gpu+0x34a/0x380 [amdgpu]
[  711.782044]  kfd_process_device_free_bos+0xe0/0x130 [amdgpu]
[  711.782611]  kfd_process_wq_release+0x286/0x380 [amdgpu]
[  711.783172]  process_one_work+0x236/0x420
[  711.783543]  worker_thread+0x34/0x400
[  711.783911]  ? process_one_work+0x420/0x420
[  711.784279]  kthread+0x126/0x140
[  711.784653]  ? kthread_park+0x90/0x90
[  711.785018]  ret_from_fork+0x22/0x30
[  711.785387] ---[ end trace d8f50f6594817c84 ]---
[  711.798716] [drm] amdgpu: ttm finalized





So it means the process was stuck in some wait_event_killable (maybe here drm_sched_entity_flush) - you can try 'cat/proc/$process_pid/stack' maybe before

you kill it to see where it was stuck so we can go from there.















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









The full dmesg will just the repetition of those two messages,

[186885.764079] amdgpu 0000:43:00.0: amdgpu: amdgpu: finishing device.
[186885.766916] [drm] free PSP TMR buffer
[186893.868173] amdgpu: amdgpu_vm_bo_update failed
[186893.868235] amdgpu: validate_invalid_user_pages: update PTE failed
[186893.876154] amdgpu: amdgpu_vm_bo_update failed
[186893.876190] amdgpu: validate_invalid_user_pages: update PTE failed
[186893.884150] amdgpu: amdgpu_vm_bo_update failed
[186893.884185] 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.





I tried something like *kfd_process_ref_release* which I think did what you described, but it did not work.




I don't see how kfd_process_ref_release is the same as I mentioned above, what i meant is calling the code above within kgd2kfd_suspend (where you tried to call kfd_kill_all_user_processes bellow)






Yes, you are right. It was not called by it.  











Instead I tried to kill the process from the kernel, but the amdgpu could **only** be hot-plugged in back successfully only if there was no rocm kernel running when it was plugged out. If not, amdgpu_probe will just hang later. (Maybe because
 amdgpu was plugged out while running state, it leaves a bad HW state which causes probe to hang).




We usually do asic_reset during probe to reset all HW state (checlk if amdgpu_device_init->amdgpu_asic_reset is running when you  plug back).






OK




  






I don’t know if this is a viable solution worth pursuing, but I attached the diff anyway.



Another solution could be let compute stack user mode detect a topology change via generation_count change, and abort gracefully there.

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index 4e7d9cb09a69..79b4c9b84cd0 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -697,12 +697,15 @@ void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm, bool force)
                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(force);
        }



+       if (drm_dev_is_unplugged(kfd->ddev))
+               kfd_kill_all_user_processes();
+
        kfd->dqm->ops.stop(kfd->dqm);
        kfd_iommu_suspend(kfd);
 }
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 55c9e1922714..84cbcd857856 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -1065,6 +1065,7 @@ void kfd_unref_process(struct kfd_process *p);
 int kfd_process_evict_queues(struct kfd_process *p, bool force);
 int kfd_process_restore_queues(struct kfd_process *p);
 void kfd_suspend_all_processes(bool force);
+void kfd_kill_all_user_processes(void);
 /*
  * kfd_resume_all_processes:
  *     bool sync: If kfd_resume_all_processes() should wait for the
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index 6cdc855abb6d..fb0c753b682c 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -2206,6 +2206,24 @@ void kfd_suspend_all_processes(bool force)
        srcu_read_unlock(&kfd_processes_srcu, idx);
 }



+
+void kfd_kill_all_user_processes(void)
+{
+       struct kfd_process *p;
+       struct amdkfd_process_info *p_info;
+       unsigned int temp;
+       int idx = srcu_read_lock(&kfd_processes_srcu);
+
+       pr_info("Killing all processes\n");
+       hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
+               p_info = p->kgd_process_info;
+               pr_info("Killing  processes, pid = %d", pid_nr(p_info->pid));
+               kill_pid(p_info->pid, SIGBUS, 1);




From looking into kill_pid I see it only sends a signal but doesn't wait for completion, it would make sense to wait for completion here. In any case I would actually try to put here





I have made a version which does that with some atomic counters. Please read later in the diff.







hash_for_each_rcu(p_info)
    

    cancel_delayed_work_sync(&p_info->restore_userptr_work)



instead 
 at least that what i meant in the previous mail. 



I actually tried that earlier, and it did not work. Application still keeps running, and you have to send a kill to the user process.



I have made the following version. It waits for processes to terminate synchronously after sending SIGBUS. After that it does the real work of amdgpu_pci_remove.
However, it hangs at amdgpu_device_ip_fini_early when it is trying to deinit ip_block 6 <sdma_v4_0> (https://gitlab.freedesktop.org/agd5f/linux/-/blob/amd-staging-drm-next/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c#L2818).
 I assume that there are still some inflight dma, therefore fini of this ip block thus hangs? 



The following is an excerpt of the dmesg: please excuse for putting my own pr_info, but I hope you get my point of where it hangs.




[  392.344735] amdgpu: all processes has been fully released
[  392.346557] amdgpu: amdgpu_acpi_fini done
[  392.346568] amdgpu 0000:b3:00.0: amdgpu: amdgpu: finishing device.
[  392.349238] amdgpu: amdgpu_device_ip_fini_early enter ip_blocks = 9
[  392.349248] amdgpu: Free mem_obj = 000000007bf54275, range_start = 14, range_end = 14
[  392.350299] amdgpu: Free mem_obj = 00000000a85bc878, range_start = 12, range_end = 12
[  392.350304] amdgpu: Free mem_obj = 00000000b8019e32, range_start = 13, range_end = 13
[  392.350308] amdgpu: Free mem_obj = 000000002d296168, range_start = 4, range_end = 11
[  392.350313] amdgpu: Free mem_obj = 000000001fc4f934, range_start = 0, range_end = 3
[  392.350322] amdgpu: amdgpu_amdkfd_suspend(adev, false) done
[  392.350672] amdgpu: hw_fini of IP block[8] <jpeg_v2_5> done 0
[  392.350679] amdgpu: hw_fini of IP block[7] <vcn_v2_5> done 0






I just remembered that the idea to actively kill and wait for running user processes during unplug was rejected

as a bad idea in the first iteration of unplug work I did (don't remember why now, need to look) so this is a no go.





Maybe an application has kfd open, but was not accessing the dev. So kill it at unplug could kill the process unnecessarily.
However, the latest version I had with the sleep function got rid of the IP block fini hang.



Our policy is to let zombie processes (zombie in a sense that the underlying device is gone) live as long as they want


(as long as you able to terminate them - which you do, so that ok)

and the system should finish PCI remove gracefully and be able to hot plug back the device.  Hence, i suggest dropping

this direction of forcing all user processes to be killed, confirm you have graceful shutdown and remove of device

from PCI topology and then concentrate on why when you plug back it hangs. 



So I basically revert back to the original solution which you suggested.




diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index 4e7d9cb09a69..5504a18b5a45 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -697,7 +697,7 @@ void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm, bool force)
                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(force);







First confirm if ASIC reset happens on

next init. 



This patch works great at planned plugout, where all the rocm processes are killed before plugout. And device can be added back without a problem.
However unplanned plugout when there is rocm processes are running just don’t work.



Second please confirm if the timing you kill manually the user process has impact on whether you have a hang

on next plug back (if you kill before 


Scenario 0: Kill before plug back




1. echo 1 > /sys/bus/pci/…/remove, would finish. 
But the application won’t exit until there is a kill signal.



2. kill the the process. The application does several things and seems trigger drm_release in the kernel, which are met with kernel NULL pointer deference related to sysfs_remove. Then the whole fs just freeze.




[  +0.002498] BUG: kernel NULL pointer dereference, address: 0000000000000098
[  +0.000486] #PF: supervisor read access in kernel mode
[  +0.000545] #PF: error_code(0x0000) - not-present page
[  +0.000551] PGD 0 P4D 0
[  +0.000553] Oops: 0000 [#1] SMP NOPTI
[  +0.000540] CPU: 75 PID: 4911 Comm: kworker/75:2 Tainted: G        W   E     5.13.0-kfd #1
[  +0.000559] Hardware name: INGRASYS         TURING  /MB      , BIOS K71FQ28A 10/05/2021
[  +0.000567] Workqueue: events delayed_fput
[  +0.000563] RIP: 0010:kernfs_find_ns+0x1b/0x100
[  +0.000569] Code: ff ff e8 88 59 9f 00 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 41 57 8b 05 df f0 7b 01 41 56 41 55 49 89 f5 41 54 55 48 89 fd 53 <44> 0f b7 b7 98 00 00 00 48 89 d3 4c 8b 67 70 66 41 83 e6 20 41 0f
[  +0.001193] RSP: 0018:ffffb9875db5fc98 EFLAGS: 00010246
[  +0.000602] RAX: 0000000000000000 RBX: ffffa101f79507d8 RCX: 0000000000000000
[  +0.000612] RDX: 0000000000000000 RSI: ffffffffc09a9417 RDI: 0000000000000000
[  +0.000604] RBP: 0000000000000000 R08: 0000000000000001 R09: 0000000000000000
[  +0.000760] R10: ffffb9875db5fcd0 R11: 0000000000000000 R12: 0000000000000000
[  +0.000597] R13: ffffffffc09a9417 R14: ffffa08363fb2d18 R15: 0000000000000000
[  +0.000702] FS:  0000000000000000(0000) GS:ffffa0ffbfcc0000(0000) knlGS:0000000000000000
[  +0.000666] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[  +0.000658] CR2: 0000000000000098 CR3: 0000005747812005 CR4: 0000000000770ee0
[  +0.000715] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[  +0.000655] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[  +0.000592] PKRU: 55555554
[  +0.000580] Call Trace:
[  +0.000591]  kernfs_find_and_get_ns+0x2f/0x50
[  +0.000584]  sysfs_remove_file_from_group+0x20/0x50
[  +0.000580]  amdgpu_ras_sysfs_remove+0x3d/0xd0 [amdgpu]
[  +0.000737]  amdgpu_ras_late_fini+0x1d/0x40 [amdgpu]
[  +0.000750]  amdgpu_sdma_ras_fini+0x96/0xb0 [amdgpu]
[  +0.000742]  ? gfx_v10_0_resume+0x10/0x10 [amdgpu]
[  +0.000738]  sdma_v4_0_sw_fini+0x23/0x90 [amdgpu]
[  +0.000717]  amdgpu_device_fini_sw+0xae/0x260 [amdgpu]
[  +0.000704]  amdgpu_driver_release_kms+0x12/0x30 [amdgpu]
[  +0.000687]  drm_dev_release+0x20/0x40 [drm]
[  +0.000583]  drm_release+0xa8/0xf0 [drm]
[  +0.000584]  __fput+0xa5/0x250
[  +0.000621]  delayed_fput+0x1f/0x30
[  +0.000726]  process_one_work+0x26e/0x580
[  +0.000581]  ? process_one_work+0x580/0x580
[  +0.000611]  worker_thread+0x4d/0x3d0
[  +0.000611]  ? process_one_work+0x580/0x580
[  +0.000605]  kthread+0x117/0x150
[  +0.000611]  ? kthread_park+0x90/0x90
[  +0.000619]  ret_from_fork+0x1f/0x30
[  +0.000625] Modules linked in: amdgpu(E) xt_conntrack xt_MASQUERADE nfnetlink xt_addrtype iptable_filter iptable_nat nf_nat nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 br_netfilter x86_pkg_temp_thermal cdc_ether usbnet acpi_pad msr ip_tables x_tables
 ast drm_vram_helper iommu_v2 drm_ttm_helper gpu_sched ttm drm_kms_helper cfbfillrect syscopyarea cfbimgblt sysfillrect sysimgblt fb_sys_fops cfbcopyarea drm drm_panel_orientati
on_quirks [last unloaded: amdgpu]



3.  echo 1 > /sys/bus/pci/rescan. This would just hang. I assume the sysfs is broken.



Based on 1 & 2, it seems that 1 won’t let the amdgpu exit gracefully, because 2 will do some cleanup maybe should have happened before 1.



or you kill after plug back does it makes a difference). 





Scenario 2: Kill after plug back



If I perform rescan before kill, then the driver seemed probed fine. But kill will have the same issue which messed up the sysfs the same way as in Scenario 2.






Final Comments:



0. cancel_delayed_work_sync(&p_info->restore_userptr_work) would make the repletion of
amdgpu_vm_bo_update failure go away, but it does not solve the issues in those scenarios.



1. For planned hotplug, this patch should work as long as you follow some protocol, i.e. kill before plugout. Is this patch an acceptable one since it provides some added feature than before?



2. For unplanned hotplug when there is rocm app running, the patch that kill all processes and wait for 5 sec would work consistently. But it seems that it is an unacceptable solution for official release. I can hold it for our own internal usage.  It
 seems that kill after removal would cause problems, and I don’t know if there is a quick fix by me because of my limited understanding of the amdgpu driver. Maybe AMD could have a quick fix; Or it is really a difficult one. This feature may or may not be a
 blocking issue in our GPU disaggregation research down the way. Please let us know for either cases, and we would like to learn and help as much as we could!



Thank you so much!



Best regards,
Shuotao




Andrey













diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
index 8fa9b86ac9d2..c0b27f722281 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
@@ -188,6 +188,12 @@ void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
  kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry);
 }
 
+void amdgpu_amdkfd_kill_all_processes(struct amdgpu_device *adev)
+{
+ if (adev->kfd.dev)
+ kgd2kfd_kill_all_user_processes(adev->kfd.dev);
+}
+
 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm)
 {
  if (adev->kfd.dev)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index 27c74fcec455..f4e485d60442 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -141,6 +141,7 @@ struct amdkfd_process_info {
 int amdgpu_amdkfd_init(void);
 void amdgpu_amdkfd_fini(void);
 
+void amdgpu_amdkfd_kill_all_processes(struct amdgpu_device *adev);
 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, bool sync);
@@ -405,6 +406,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
  const struct kgd2kfd_shared_resources *gpu_resources);
 void kgd2kfd_device_exit(struct kfd_dev *kfd);
 void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm, bool force);
+void kgd2kfd_kill_all_user_processes(struct kfd_dev *kfd);
 int kgd2kfd_resume_iommu(struct kfd_dev *kfd);
 int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm, bool sync);
 int kgd2kfd_pre_reset(struct kfd_dev *kfd);
@@ -443,6 +445,9 @@ static inline void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm, bool force)
 {
 }
 
+void kgd2kfd_kill_all_user_processes(struct kfd_dev *kfd){
+}
+
 static int __maybe_unused kgd2kfd_resume_iommu(struct kfd_dev *kfd)
 {
  return 0;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index 3d5fc0751829..af6fe5080cfa 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
@@ -2101,6 +2101,9 @@ amdgpu_pci_remove(struct pci_dev *pdev)
 {
  struct drm_device *dev = pci_get_drvdata(pdev);
 
+ /* kill all kfd processes before drm_dev_unplug */
+ amdgpu_amdkfd_kill_all_processes(drm_to_adev(dev));
+
 #ifdef HAVE_DRM_DEV_UNPLUG
  drm_dev_unplug(dev);
 #else
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index 5504a18b5a45..480c23bef5e2 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -691,6 +691,12 @@ bool kfd_is_locked(void)
  return  (atomic_read(&kfd_locked) > 0);
 }
 
+inline void kgd2kfd_kill_all_user_processes(struct kfd_dev* dev)
+{
+ kfd_kill_all_user_processes();
+}
+
+
 void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm, bool force)
 {
  if (!kfd->init_complete)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 55c9e1922714..a35a2cb5bb9f 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -1064,6 +1064,7 @@ static inline struct kfd_process_device *kfd_process_device_from_gpuidx(
 void kfd_unref_process(struct kfd_process *p);
 int kfd_process_evict_queues(struct kfd_process *p, bool force);
 int kfd_process_restore_queues(struct kfd_process *p);
+void kfd_kill_all_user_processes(void);
 void kfd_suspend_all_processes(bool force);
 /*
  * kfd_resume_all_processes:
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index 6cdc855abb6d..17e769e6951d 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -46,6 +46,9 @@ struct mm_struct;
 #include "kfd_trace.h"
 #include "kfd_debug.h"
 
+static atomic_t kfd_process_locked = ATOMIC_INIT(0);
+static atomic_t kfd_inflight_kills = ATOMIC_INIT(0);
+
 /*
  * List of struct kfd_process (field kfd_process).
  * Unique/indexed by mm_struct*
@@ -802,6 +805,9 @@ struct kfd_process *kfd_create_process(struct task_struct *thread)
  struct kfd_process *process;
  int ret;
 
+ if ( atomic_read(&kfd_process_locked) > 0 )
+ return ERR_PTR(-EINVAL);
+
  if (!(thread->mm && mmget_not_zero(thread->mm)))
  return ERR_PTR(-EINVAL);
 
@@ -1126,6 +1132,10 @@ static void kfd_process_wq_release(struct work_struct *work)
  put_task_struct(p->lead_thread);
 
  kfree(p);
+
+ if ( atomic_read(&kfd_process_locked) > 0 ){
+ atomic_dec(&kfd_inflight_kills);
+ }
 }
 
 static void kfd_process_ref_release(struct kref *ref)
@@ -2186,6 +2196,35 @@ static void restore_process_worker(struct work_struct *work)
  pr_err("Failed to restore queues of pasid 0x%x\n", p->pasid);
 }
 
+void kfd_kill_all_user_processes(void)
+{
+ struct kfd_process *p;
+ /* struct amdkfd_process_info *p_info; */
+ unsigned int temp;
+ int idx;
+ atomic_inc(&kfd_process_locked);
+
+ idx = srcu_read_lock(&kfd_processes_srcu);
+ pr_info("Killing all processes\n");
+ hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
+ dev_warn(kfd_device,
+ "Sending SIGBUS to process %d (pasid 0x%x)",
+ p->lead_thread->pid, p->pasid);
+ send_sig(SIGBUS, p->lead_thread, 0);
+ atomic_inc(&kfd_inflight_kills);
+ }
+ srcu_read_unlock(&kfd_processes_srcu, idx);
+
+ while ( atomic_read(&kfd_inflight_kills) > 0 ){
+ dev_warn(kfd_device,
+ "kfd_processes_table is not empty, going to sleep for 10ms\n");
+ msleep(10);
+ }
+
+ atomic_dec(&kfd_process_locked);
+ pr_info("all processes has been fully released");
+}
+
 void kfd_suspend_all_processes(bool force)
 {
  struct kfd_process *p;










Regards,
Shuotao








Andrey





+       }
+       srcu_read_unlock(&kfd_processes_srcu, idx);
+}
+
+
 int kfd_resume_all_processes(bool sync)
 {
        struct kfd_process *p;






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;



























