On 2022-04-18 09:22, Shuotao Xu wrote:
On 2022-04-15 06:12, Shuotao Xu wrote:
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 2022-04-14 10:00, Shuotao Xu wrote:
On 2022-04-13 12:03, Shuotao Xu 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.
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.
Still I am not clear if ASIC reset happens on plug back or no, can you trace this please ?
I tried add pr_info into asic_reset functions, but cannot trace any upon plug-back.
This could possibly explain the hang on plug back. Can you see why we don't get there ?
Is amdgpu supposed to asic_reset each time when it is probed? I right now it seems to probe ok (it did not hang). I will trace back further
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.
Why you think it must exit ?
Because rocm will need to release the drm descriptor to get amdgpu_amdkfd_device_fini_sw called, which would eventually call kgd2kfd_device_exit called. This would clean up kfd_topology at least. Otherwise I don’t see how it would be added back without messing
up kfd topology to say the least.
However, those are all based my own observations. Please explain why it does not need exit if you believe so?
Note that when you add back a new device, pci device and drm device are created, I am not an expert on KFD code but i believe also a new KFD device is created independent of the old one and so the topology should see just 2 device instances (one
old zombie and one real new). I know at least this wasn't an issue for the graphic stack in exact same scenario and the libdrm tests i pointed to test exact this scenario.
Yes, regardless of the OOPS in ttm_bo_cleanup_refs, I plugged back the gpu, and I think it got probed all right, however the old kfd node is still there.
I can passed libdrm basic test suite on the plugged back. The bo test hangs out-of-box even without hotplug (see dmesg below).
kernel:[ 1609.029125] watchdog: BUG: soft lockup - CPU#39 stuck for 89s! [amdgpu_test:36407]
[ +0.000407] Code: 48 89 47 18 48 89 47 20 48 89 47 28 48 89 47 30 48 89 47 38 48 8d 7f 40 75 d9 90 c3 0f 1f 80 00 00 00 00 b9 00 10 00 00 31 c0 <f3> aa c3 cc cc cc cc cc cc 0f 1f 44 00 00 48 85 ff 0f 84 f2 00 00
[ +0.000856] RSP: 0018:ffffb996b57b3c40 EFLAGS: 00010246
[ +0.000434] RAX: 0000000000000000 RBX: ffff9cc7f8706e88 RCX: 0000000000000980
[ +0.000436] RDX: fffff935b17d9140 RSI: fffff935b17e0000 RDI: ffff9c831f645680
[ +0.000439] RBP: 0000000000000400 R08: fffff935b17d0000 R09: 0000000000000000
[ +0.000447] R10: 0000000000000000 R11: 0000000000000000 R12: 000000000000000a
[ +0.000437] R13: ffff9cc783980a20 R14: 000000000b5dbc00 R15: ffff9cc7f8706078
[ +0.000438] FS: 00007ff1ef611300(0000) GS:ffff9d453efc0000(0000) knlGS:0000000000000000
[ +0.000445] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ +0.000462] CR2: 00007f418bbb9320 CR3: 000000819fa84006 CR4: 0000000000770ee0
[ +0.000466] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[ +0.000451] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[ +0.000455] PKRU: 55555554
[ +0.000451] Call Trace:
[ +0.000448] ttm_pool_free+0x110/0x230 [ttm]
[ +0.000451] ttm_tt_unpopulate+0x5e/0xb0 [ttm]
[ +0.000454] ttm_tt_destroy_common+0xe/0x30 [ttm]
[ +0.000453] amdgpu_ttm_backend_destroy+0x1e/0x70 [amdgpu]
[ +0.000569] ttm_bo_cleanup_memtype_use+0x37/0x60 [ttm]
[ +0.000458] ttm_bo_release+0x286/0x500 [ttm]
[ +0.000450] amdgpu_bo_unref+0x1a/0x30 [amdgpu]
[ +0.000544] amdgpu_gem_object_free+0xad/0x160 [amdgpu]
[ +0.000534] drm_gem_object_release_handle+0x6a/0x80 [drm]
[ +0.000476] drm_gem_handle_delete+0x5b/0xa0 [drm]
[ +0.000465] ? drm_gem_handle_create+0x40/0x40 [drm]
[ +0.000469] drm_ioctl_kernel+0xab/0xf0 [drm]
[ +0.000458] drm_ioctl+0x1ec/0x390 [drm]
[ +0.000440] ? drm_gem_handle_create+0x40/0x40 [drm]
[ +0.000438] ? selinux_file_ioctl+0x17d/0x220
[ +0.000423] ? lock_release+0x1ce/0x270
[ +0.000416] ? trace_hardirqs_on+0x1b/0xd0
[ +0.000418] ? _raw_spin_unlock_irqrestore+0x2d/0x40
[ +0.000419] amdgpu_drm_ioctl+0x49/0x80 [amdgpu]
[ +0.000499] __x64_sys_ioctl+0x80/0xb0
[ +0.000414] do_syscall_64+0x3a/0x70
[ +0.000400] entry_SYSCALL_64_after_hwframe+0x44/0xae
[ +0.000387] RIP: 0033:0x7ff1ef7263db
[ +0.000371] Code: 0f 1e fa 48 8b 05 b5 7a 0d 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d 85 7a 0d 00 f7 d8 64 89 01 48
[ +0.000763] RSP: 002b:00007ffdf1cd0278 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
[ +0.000386] RAX: ffffffffffffffda RBX: 00007ffdf1cd02b0 RCX: 00007ff1ef7263db
[ +0.000383] RDX: 00007ffdf1cd02b0 RSI: 0000000040086409 RDI: 0000000000000007
[ +0.000396] RBP: 0000000040086409 R08: 00005574eefd5c60 R09: 00005574eefdd360
[ +0.000391] R10: 00005574eefd4010 R11: 0000000000000246 R12: 00005574eefd66d8
[ +0.000386] R13: 0000000000000007 R14: 0000000000000000 R15: 00007ff1ef830143
I also tried to run tf benchmark to the newly plugged nodes (one of the node is dummy), but failed.
Can we have some confirmation from kfd team that they have considered a zombie kfd node?
Also note that even with running grpahic stack there is always a KFD device and KFD topology present but of course probably not the same as when u run a KFD facing process so there could be some issues there.
Also note that because of this patch
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=267d51d77fdae8708b94e1a24b8e5d961297edb7 all MMIO accesses from such zombie/orphan user processes will be remapped to zero page and so will not necessarily experience a segfault
when device removal happnes but rather maybe some crash due to NULL data read from MMIO by the process and used in some manner.
It depends on where the application is when the device is plugged out.
For example, in one case, the application keeps saying out-of-memory, but won’t exit.
For one of the cases. The other case is that it would wait for a signal.
2022-04-18 12:42:38.939303: E tensorflow/stream_executor/rocm/ rocm_driver.cc:692] failed to alloc 2304 bytes on host: HIP_ERROR_OutOfMemory 2022-04-18 12:42:38.939322: W ./tensorflow/core/common_runtime/gpu/gpu_host_allocator.h:44] could not allocate pinned host memory of size: 2304
2022-04-18 12:42:38.940772: E tensorflow/stream_executor/rocm/ rocm_driver.cc:692] failed to alloc 2304 bytes on host: HIP_ERROR_OutOfMemory 2022-04-18 12:42:38.940791: W ./tensorflow/core/common_runtime/gpu/gpu_host_allocator.h:44] could not allocate pinned host memory of size: 2304
2022-04-18 12:42:38.942379: E tensorflow/stream_executor/rocm/ rocm_driver.cc:692] failed to alloc 2304 bytes on host: HIP_ERROR_OutOfMemory 2022-04-18 12:42:38.942399: W ./tensorflow/core/common_runtime/gpu/gpu_host_allocator.h:44] could not allocate pinned host memory of size: 2304
2022-04-18 12:42:38.943829: E tensorflow/stream_executor/rocm/ rocm_driver.cc:692] failed to alloc 2304 bytes on host: HIP_ERROR_OutOfMemory 2022-04-18 12:42:38.943849: W ./tensorflow/core/common_runtime/gpu/gpu_host_allocator.h:44] could not allocate pinned host memory of size: 2304
2022-04-18 12:42:38.945272: E tensorflow/stream_executor/rocm/ rocm_driver.cc:692] failed to alloc 2304 bytes on host: HIP_ERROR_OutOfMemory 2022-04-18 12:42:38.945292: W ./tensorflow/core/common_runtime/gpu/gpu_host_allocator.h:44] could not allocate pinned host memory of size: 2304
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]
This is a known regression, all SYSFS components must be removed before pci_remove code runs otherwise you get either warnings for single file removals or
OOPSEs for sysfs gorup removals like here. Please try to move amdgpu_ras_sysfs_remove from amdgpu_ras_late_fini to the end of amdgpu_ras_pre_fini (which happens before pci remove)
I fixed it in the newer patch, please see it below.
I first plugout the device, then kill the rocm user process. Then it has other OOPSES related to ttm_bo_cleanup_refs.
[ +0.000006] BUG: kernel NULL pointer dereference, address: 0000000000000010
[ +0.000349] #PF: supervisor read access in kernel mode
[ +0.000340] #PF: error_code(0x0000) - not-present page
[ +0.000341] PGD 0 P4D 0
[ +0.000336] Oops: 0000 [#1] SMP NOPTI
[ +0.000345] CPU: 9 PID: 95 Comm: kworker/9:1 Tainted: G W E 5.13.0-kfd #1
[ +0.000367] Hardware name: INGRASYS TURING /MB , BIOS K71FQ28A 10/05/2021
[ +0.000376] Workqueue: events delayed_fput
[ +0.000422] RIP: 0010:ttm_resource_free+0x24/0x40 [ttm]
[ +0.000464] Code: 00 00 0f 1f 40 00 0f 1f 44 00 00 53 48 89 f3 48 8b 36 48 85 f6 74 21 48 8b 87 28 02 00 00 48 63 56 10 48 8b bc d0 b8 00 00 00 <48> 8b 47 10 ff 50 08 48 c7 03 00 00 00 00 5b c3 66 66 2e 0f 1f 84
[ +0.001009] RSP: 0018:ffffb21c59413c98 EFLAGS: 00010282
[ +0.000515] RAX: ffff8b1aa4285f68 RBX: ffff8b1a823b5ea0 RCX: 00000000002a000c
[ +0.000536] RDX: 0000000000000000 RSI: ffff8b1acb84db80 RDI: 0000000000000000
[ +0.000539] RBP: 0000000000000001 R08: 0000000000000000 R09: ffffffffc03c3e00
[ +0.000543] R10: 0000000000000000 R11: 0000000000000000 R12: ffff8b1a823b5ec8
[ +0.000542] R13: 0000000000000000 R14: ffff8b1a823b5d90 R15: ffff8b1a823b5ec8
[ +0.000544] FS: 0000000000000000(0000) GS:ffff8b187f440000(0000) knlGS:0000000000000000
[ +0.000559] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ +0.000575] CR2: 0000000000000010 CR3: 00000076e6812004 CR4: 0000000000770ee0
[ +0.000575] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[ +0.000579] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[ +0.000575] PKRU: 55555554
[ +0.000568] Call Trace:
[ +0.000567] ttm_bo_cleanup_refs+0xe4/0x290 [ttm]
[ +0.000588] ttm_bo_delayed_delete+0x147/0x250 [ttm]
[ +0.000589] ttm_device_fini+0xad/0x1b0 [ttm]
[ +0.000590] amdgpu_ttm_fini+0x2a7/0x310 [amdgpu]
[ +0.000730] gmc_v9_0_sw_fini+0x3a/0x40 [amdgpu]
[ +0.000753] amdgpu_device_fini_sw+0xae/0x260 [amdgpu]
[ +0.000734] amdgpu_driver_release_kms+0x12/0x30 [amdgpu]
[ +0.000737] drm_dev_release+0x20/0x40 [drm]
[ +0.000626] drm_release+0xa8/0xf0 [drm]
[ +0.000625] __fput+0xa5/0x250
[ +0.000606] delayed_fput+0x1f/0x30
[ +0.000607] process_one_work+0x26e/0x580
[ +0.000608] ? process_one_work+0x580/0x580
[ +0.000616] worker_thread+0x4d/0x3d0
[ +0.000614] ? process_one_work+0x580/0x580
[ +0.000617] kthread+0x117/0x150
[ +0.000615] ? kthread_park+0x90/0x90
[ +0.000621] ret_from_fork+0x1f/0x30
[ +0.000603] 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
drm_ttm_helper iommu_v2 ttm gpu_sched drm_kms_helper cfbfillrect syscopyarea cfbimgblt sysfillrect sysimgblt fb_sys_fops cfbcopyarea drm drm_panel_orientation_quirks [last unloaded: amdgpu]
[ +0.002840] CR2: 0000000000000010
[ +0.000755] ---[ end trace 9737737402551e39 ]--
This looks like another regression - try seeing where is the NULL reference and then we can see how to avoid this.
Those are the line of code.
(gdb) l *(ttm_bo_cleanup_refs+0xe4)
0x19c4 is in ttm_bo_cleanup_refs (drivers/gpu/drm/ttm/ttm_bo.c:360).
355 ttm_bo_move_to_pinned(bo);
356 list_del_init(&bo->ddestroy);
357 spin_unlock(&bo->bdev->lru_lock);
358 ttm_bo_cleanup_memtype_use(bo);
359
360 if (unlock_resv)
361 dma_resv_unlock(amdkcl_ttm_resvp(bo));
362
363 ttm_bo_put(bo);
364
(gdb) l *(ttm_resource_free+0x24)
0x57f4 is in ttm_resource_free (drivers/gpu/drm/ttm/ttm_resource.c:65).
60
61 if (!*res)
62 return;
63
64 man = ttm_manager_type(bo->bdev, (*res)->mem_type);
65 man->func->free(man, *res);
66 *res = NULL;
67 }
68 EXPORT_SYMBOL(ttm_resource_free);
69
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.
Still - it's better to do it this way even for those failures to go awaya
Cancel_delayed_work is insufficient, you will need to make sure the work won’t be processed after plugout. Please see my patch
Saw, see my comment.
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?
Let's try to fix more as I advised above.
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!
I am currently not sure why it helps. I will need to setup my own ROCm setup and retest hot plug to check this in more depth but currently i have higher priorities. Please try to confirm ASIC reset always takes place on plug back
and fix the sysfs OOPs as I advised above to clear up at least some of the issues. Also please describe to me exactly what you steps to reproduce this scenario so later I might be able to do it myself.
I can still try to help to fix the bug in my spare time. My setup is as follows
- I have a server with 4 AMD MI100 GPUs. I think 1 GPU would also work.
- I used the https://github.com/RadeonOpenCompute/ROCK-Kernel-Driver/tree/roc-5.0.x as
the starting point, and apply Mukul’s patch and my patch.
- Then I run a tensorflow benchmark from a docker.
- docker run -it --device=/dev/kfd --device=/dev/dri --group-add video rocm/tensorflow:rocm4.5.2-tf1.15-dev
- And run the following benchmark in the docker: python benchmarks/scripts/tf_cnn_benchmarks/tf_cnn_benchmarks.py --num_gpus=4 --batch_size=32 --model=resnet50 --variable_update=parameter_server
- Might to need to adjust num_gpus parameter based on your setup
- Remove a GPU at random time.
- Do whatever is needed to before plugback and reverify the benchmark can still run.
Also, we have hotplug test suite in libdrm (graphic stack), so maybe u can install libdrm and run that test suite to see if it exposes more issues.
OK I could try it some time.
I tried suite 13, the hotplugout test, but it says it got killed? There are a some oops from dmesg during ttm_pool_free_page.
Userspace log:
$ sudo ./tests/amdgpu/amdgpu_test -f -s 13
The ASIC NOT support UVD, suite disabled
The ASIC NOT support VCE, suite disabled
The ASIC NOT support UVD ENC, suite disabled.
Don't support TMZ (trust memory zone), security suite disabled
CUnit - A unit testing framework for C - Version 2.1-3
Suite: Hotunplug Tests
Test: Unplug card and rescan the bus to plug it back …Killed
Dmesg log:
[ +0.000479] BUG: unable to handle page fault for address: ffffc01343fc81b4
[ +0.000054] #PF: supervisor write access in kernel mode
[ +0.000033] #PF: error_code(0x0002) - not-present page
[ +0.000032] PGD 807ffc1067 P4D 807ffc1067 PUD 807ffc0067 PMD 0
[ +0.000038] Oops: 0002 [#1] SMP NOPTI
[ +0.000025] CPU: 92 PID: 7534 Comm: amdgpu_test Tainted: G W E 5.13.0-kfd #1
[ +0.000048] Hardware name: INGRASYS TURING /MB , BIOS K71FQ28A 10/05/2021
[ +0.000045] RIP: 0010:__free_pages+0xc/0x80
[ +0.000031] Code: 01 00 74 0f 0f b6 77 51 85 f6 74 07 31 d2 e9 3b dc ff ff e9 66 ff ff ff 66 0f 1f 44 00 00 0f 1f 44 00 00 41 54 55 48 89 fd 53 <f0> ff 4f 34 74 46 48 8b 07 a9 00 00 01 00 75 54 44 8d 66 ff 85 f6
[ +0.000103] RSP: 0018:ffff96f71ba6fd60 EFLAGS: 00010246
[ +0.000032] RAX: 00000000ffffffff RBX: ffff89f1ccf86078 RCX: 0000000003fc8180
[ +0.000041] RDX: ffff89f1b4746000 RSI: 0000000000000000 RDI: ffffc01343fc8180
[ +0.000042] RBP: ffffc01343fc8180 R08: 0000000000000000 R09: 0000000000000246
[ +0.000040] R10: 00000080b4746fff R11: 0000000000000003 R12: 0000000000000000
[ +0.000041] R13: ffff89f1ccf85f80 R14: ffff89f1ccf86ef8 R15: ffff8972293b0000
[ +0.000042] FS: 00007fcfb843a300(0000) GS:ffff89ef80100000(0000) knlGS:0000000000000000
[ +0.000046] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ +0.000033] CR2: ffffc01343fc81b4 CR3: 0000000178154006 CR4: 0000000000770ee0
[ +0.000041] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[ +0.000041] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[ +0.000041] PKRU: 55555554
[ +0.000017] Call Trace:
[ +0.000018] ttm_pool_free_page+0x69/0x90 [ttm]
[ +0.000038] ttm_pool_type_fini+0x58/0x70 [ttm]
[ +0.000034] ttm_pool_fini+0x30/0x50 [ttm]
[ +0.000031] ttm_device_fini+0xf3/0x1b0 [ttm]
[ +0.000032] amdgpu_ttm_fini+0x2a7/0x310 [amdgpu]
[ +0.000265] gmc_v9_0_sw_fini+0x3a/0x40 [amdgpu]
[ +0.000246] amdgpu_device_fini_sw+0xae/0x260 [amdgpu]
[ +0.000219] amdgpu_driver_release_kms+0x12/0x30 [amdgpu]
[ +0.000219] drm_dev_release+0x20/0x40 [drm]
[ +0.000059] drm_release+0xa8/0xf0 [drm]
[ +0.000053] __fput+0xa5/0x250
[ +0.000023] task_work_run+0x5c/0xa0
[ +0.000026] exit_to_user_mode_prepare+0x1db/0x1e0
[ +0.000033] syscall_exit_to_user_mode+0x19/0x50
[ +0.000030] do_syscall_64+0x47/0x70
[ +0.000018] entry_SYSCALL_64_after_hwframe+0x44/0xae
[ +0.000025] RIP: 0033:0x7fcfb86403d7
[ +0.000869] Code: 64 89 02 48 c7 c0 ff ff ff ff eb bb 0f 1f 80 00 00 00 00 f3 0f 1e fa 64 8b 04 25 18 00 00 00 85 c0 75 10 b8 03 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 41 c3 48 83 ec 18 89 7c 24 0c e8 f3 fb ff ff
[ +0.001788] RSP: 002b:00007ffc8fc26c28 EFLAGS: 00000246 ORIG_RAX: 0000000000000003
[ +0.000888] RAX: 0000000000000000 RBX: 000055d67a05b6a0 RCX: 00007fcfb86403d7
[ +0.000867] RDX: 00007fcfb8627be0 RSI: 0000000000000000 RDI: 0000000000000003
[ +0.000846] RBP: 000055d67a05b8a0 R08: 0000000000000007 R09: 0000000000000000
[ +0.000816] R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000000
[ +0.000791] R13: 0000000000000000 R14: 0000000000000000 R15: 00007fcfb8659143
[ +0.000770] 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 drm_ttm_helper ttm iommu_v2 gpu_sched drm_kms_helper cfbfillrect syscopyarea cfbimgblt sysfillrect sysimgblt fb_sys_fops cfbcopyarea drm drm_panel_orientation_quirks [last unloaded: amdgpu]
[ +0.003303] CR2: ffffc01343fc81b4
[ +0.000799] ---[ end trace 2360927435b19009 ]—
The following is the new diff.
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 182b7eae598a..48c3cd4054de 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -1327,7 +1327,7 @@ int emu_soc_asic_init(struct amdgpu_device *adev);
* ASICs macro.
*/
#define amdgpu_asic_set_vga_state(adev, state) (adev)->asic_funcs->set_vga_state((adev), (state))
-#define amdgpu_asic_reset(adev) (adev)->asic_funcs->reset((adev))
+#define amdgpu_asic_reset(adev) ({int r; pr_info("performing amdgpu_asic_reset\n"); r = (adev)->asic_funcs->reset((adev));r;})
#define amdgpu_asic_reset_method(adev) (adev)->asic_funcs->reset_method((adev))
#define amdgpu_asic_get_xclk(adev) (adev)->asic_funcs->get_xclk((adev))
#define amdgpu_asic_set_uvd_clocks(adev, v, d) (adev)->asic_funcs->set_uvd_clocks((adev), (v), (d))
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index 27c74fcec455..842abd7150a6 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -134,6 +134,7 @@ struct amdkfd_process_info {
/* MMU-notifier related fields */
atomic_t evicted_bos;
+ atomic_t invalid;
struct delayed_work restore_userptr_work;
struct pid *pid;
};
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 99d2b15bcbf3..2a588eb9f456 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -1325,6 +1325,7 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
atomic_set(&info->evicted_bos, 0);
+ atomic_set(&info->invalid, 0);
INIT_DELAYED_WORK(&info->restore_userptr_work,
amdgpu_amdkfd_restore_userptr_worker);
@@ -2693,6 +2694,9 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
struct mm_struct *mm;
int evicted_bos;
+ if (atomic_read(&process_info->invalid))
+ return;
+
Probably better to again use drm_dev_enter/exit guard pair instead of this flag.
I don’t know if I could use drm_dev_enter/exit efficiently because a process can have multiple drm_dev open. And I don’t know how I can recover/refer drm_dev(s) efficiently in the worker function in order to use drm_dev_enter/exit.
evicted_bos = atomic_read(&process_info->evicted_bos);
if (!evicted_bos)
return;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index ec38517ab33f..e7d85d8d282d 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -1054,6 +1054,7 @@ void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
*/
void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
{
+ pr_debug("%s called\n",__func__);
pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
}
@@ -1066,6 +1067,7 @@ void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
*/
int amdgpu_device_pci_reset(struct amdgpu_device *adev)
{
+ pr_debug("%s called\n",__func__);
return pci_reset_function(adev->pdev);
}
@@ -4702,6 +4704,8 @@ int amdgpu_do_asic_reset(struct list_head *device_list_handle,
bool need_full_reset, skip_hw_reset, vram_lost = false;
int r = 0;
+ pr_debug("%s called\n",__func__);
+
/* Try reset handler method first */
tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
reset_list);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c
index 49bdf9ff7350..b469acb65c1e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c
@@ -2518,7 +2518,6 @@ void amdgpu_ras_late_fini(struct amdgpu_device *adev,
if (!ras_block || !ih_info)
return;
- amdgpu_ras_sysfs_remove(adev, ras_block);
if (ih_info->cb)
amdgpu_ras_interrupt_remove_handler(adev, ih_info);
}
@@ -2577,6 +2576,7 @@ void amdgpu_ras_suspend(struct amdgpu_device *adev)
int amdgpu_ras_pre_fini(struct amdgpu_device *adev)
{
struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
+ struct ras_manager *obj, *tmp;
if (!adev->ras_enabled || !con)
return 0;
@@ -2585,6 +2585,13 @@ int amdgpu_ras_pre_fini(struct amdgpu_device *adev)
/* Need disable ras on all IPs here before ip [hw/sw]fini */
amdgpu_ras_disable_all_features(adev, 0);
amdgpu_ras_recovery_fini(adev);
+
+ /* remove sysfs before pci_remove to avoid OOPSES from sysfs_remove_groups */
+ list_for_each_entry_safe(obj, tmp, &con->head, node) {
+ amdgpu_ras_sysfs_remove(adev, &obj->head);
+ put_obj(obj);
+ }
+
return 0;
}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index 4e7d9cb09a69..0fa806a78e39 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -693,16 +693,35 @@ bool kfd_is_locked(void)
void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm, bool force)
{
+ struct kfd_process *p;
+ struct amdkfd_process_info *p_info;
+ unsigned int temp;
+
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(force);
}
+ if (drm_dev_is_unplugged(kfd->ddev)){
+ int idx = srcu_read_lock(&kfd_processes_srcu);
+ pr_debug("cancel restore_userptr_wor\n");
+ hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
+ if ( kfd_process_gpuidx_from_gpuid(p, kfd->id) >= 0 ) {
+ p_info = p->kgd_process_info;
+ pr_debug("cancel processes, pid = %d for gpu_id = %d", pid_nr(p_info->pid), kfd->id);
+ cancel_delayed_work_sync(&p_info->restore_userptr_work);
+ /* block all future restore_userptr_work */
+ atomic_inc(&p_info->invalid);
Same as i mentioned above with drm.dev_eneter/exit
Same as I mentioned as the process can have many drm_dev open.
Final comments:
I suspect that the my linux kernel version might not have all the fixes you did for hotplug. Can you give me a pointer to the lowest version of linux kernel (5.14 from linux kernel repo? amd-drm-staging-next does not work for MI100 out-of-box), which would
pass all libdrm tests including hotplug tests (some tests hang, some failed now) ?
Thank you so much!
Andrey
+ }
+ }
+ srcu_read_unlock(&kfd_processes_srcu, idx);
+ }
+
kfd->dqm->ops.stop(kfd->dqm);
kfd_iommu_suspend(kfd);
}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
index 600ba2a728ea..7e3d1848eccc 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
@@ -669,11 +669,12 @@ static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
#ifdef HAVE_AMD_IOMMU_PC_SUPPORTED
if (dev->kobj_perf) {
list_for_each_entry(perf, &dev->perf_props, list) {
+ sysfs_remove_group(dev->kobj_perf, perf->attr_group);
kfree(perf->attr_group);
perf->attr_group = NULL;
}
kobject_del(dev->kobj_perf);
- kobject_put(dev->kobj_perf);
+ /* kobject_put(dev->kobj_perf); */
dev->kobj_perf = NULL;
}
#endif
Thank you so much! Looking forward to your comments!
Regards,
Shuotao
Andrey
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;
|
