[PATCH] drm/amdgpu: fix stall on CPU when allocate large system memory
James Zhu
jamesz at amd.com
Fri Nov 18 12:22:51 UTC 2022
On 2022-11-17 17:03, Felix Kuehling wrote:
> Am 2022-11-17 um 16:38 schrieb James Zhu:
>> When applications try to allocate large system (more than > 128GB),
>> "stall cpu" is reported.
>>
>> for such large system memory, walk_page_range takes more than 20s
>> usually.
>> The warning message can be removed when splitting hmm range into smaller
>> ones which is not more 64GB for each walk_page_range.
>>
>> [ 164.437617] amdgpu:amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu:1753:
>> amdgpu: create BO VA 0x7f63c7a00000 size 0x2f16000000 domain CPU
>> [ 164.488847] amdgpu:amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu:1785:
>> amdgpu: creating userptr BO for user_addr = 7f63c7a00000
>> [ 185.439116] rcu: INFO: rcu_sched self-detected stall on CPU
>> [ 185.439125] rcu: 8-....: (20999 ticks this GP)
>> idle=e22/1/0x4000000000000000 softirq=2242/2242 fqs=5249
>> [ 185.439137] (t=21000 jiffies g=6325 q=1215)
>> [ 185.439141] NMI backtrace for cpu 8
>> [ 185.439143] CPU: 8 PID: 3470 Comm: kfdtest Kdump: loaded Tainted:
>> G O 5.12.0-0_fbk5_zion_rc1_5697_g2c723fb88626 #1
>> [ 185.439147] Hardware name: HPE ProLiant XL675d Gen10 Plus/ProLiant
>> XL675d Gen10 Plus, BIOS A47 11/06/2020
>> [ 185.439150] Call Trace:
>> [ 185.439153] <IRQ>
>> [ 185.439157] dump_stack+0x64/0x7c
>> [ 185.439163] nmi_cpu_backtrace.cold.7+0x30/0x65
>> [ 185.439165] ? lapic_can_unplug_cpu+0x70/0x70
>> [ 185.439170] nmi_trigger_cpumask_backtrace+0xf9/0x100
>> [ 185.439174] rcu_dump_cpu_stacks+0xc5/0xf5
>> [ 185.439178] rcu_sched_clock_irq.cold.97+0x112/0x38c
>> [ 185.439182] ? tick_sched_handle.isra.21+0x50/0x50
>> [ 185.439185] update_process_times+0x8c/0xc0
>> [ 185.439189] tick_sched_timer+0x63/0x70
>> [ 185.439192] __hrtimer_run_queues+0xff/0x250
>> [ 185.439195] hrtimer_interrupt+0xf4/0x200
>> [ 185.439199] __sysvec_apic_timer_interrupt+0x51/0xd0
>> [ 185.439201] sysvec_apic_timer_interrupt+0x69/0x90
>> [ 185.439206] </IRQ>
>> [ 185.439207] asm_sysvec_apic_timer_interrupt+0x12/0x20
>> [ 185.439211] RIP: 0010:clear_page_rep+0x7/0x10
>> [ 185.439214] Code: e8 fe 7c 51 00 44 89 e2 48 89 ee 48 89 df e8 60
>> ff ff ff c6 03 00 5b 5d 41 5c c3 cc cc cc cc cc cc cc cc b9 00 02 00
>> 00 31 c0 <f3> 48 ab c3 0f 1f 44 00 00 31 c0 b9 40 00 00 00 66 0f 1f
>> 84 00 00
>> [ 185.439218] RSP: 0018:ffffc9000f58f818 EFLAGS: 00000246
>> [ 185.439220] RAX: 0000000000000000 RBX: 0000000000000881 RCX:
>> 000000000000005c
>> [ 185.439223] RDX: 0000000000100dca RSI: 0000000000000000 RDI:
>> ffff88a59e0e5d20
>> [ 185.439225] RBP: ffffea0096783940 R08: ffff888118c35280 R09:
>> ffffea0096783940
>> [ 185.439227] R10: ffff888000000000 R11: 0000160000000000 R12:
>> ffffea0096783980
>> [ 185.439228] R13: ffffea0096783940 R14: ffff88b07fdfdd00 R15:
>> 0000000000000000
>> [ 185.439232] prep_new_page+0x81/0xc0
>> [ 185.439236] get_page_from_freelist+0x13be/0x16f0
>> [ 185.439240] ? release_pages+0x16a/0x4a0
>> [ 185.439244] __alloc_pages_nodemask+0x1ae/0x340
>> [ 185.439247] alloc_pages_vma+0x74/0x1e0
>> [ 185.439251] __handle_mm_fault+0xafe/0x1360
>> [ 185.439255] handle_mm_fault+0xc3/0x280
>> [ 185.439257] hmm_vma_fault.isra.22+0x49/0x90
>> [ 185.439261] __walk_page_range+0x692/0x9b0
>> [ 185.439265] walk_page_range+0x9b/0x120
>> [ 185.439269] hmm_range_fault+0x4f/0x90
>> [ 185.439274] amdgpu_hmm_range_get_pages+0x24f/0x260 [amdgpu]
>> [ 185.439463] amdgpu_ttm_tt_get_user_pages+0xc2/0x190 [amdgpu]
>> [ 185.439603] amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu+0x49f/0x7a0
>> [amdgpu]
>> [ 185.439774] kfd_ioctl_alloc_memory_of_gpu+0xfb/0x410 [amdgpu]
>>
>> Signed-off-by: James Zhu <James.Zhu at amd.com>
>> ---
>> drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c | 47 +++++++++++++++++--------
>> 1 file changed, 32 insertions(+), 15 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c
>> index a48ea62b12b0..0425fc6a49aa 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c
>> @@ -163,6 +163,7 @@ int amdgpu_hmm_range_get_pages(struct
>> mmu_interval_notifier *notifier,
>> struct hmm_range **phmm_range)
>> {
>> struct hmm_range *hmm_range;
>> + unsigned long hmm_range_end;
>> unsigned long timeout;
>> unsigned long i;
>> unsigned long *pfns;
>> @@ -184,25 +185,41 @@ int amdgpu_hmm_range_get_pages(struct
>> mmu_interval_notifier *notifier,
>> hmm_range->default_flags |= HMM_PFN_REQ_WRITE;
>> hmm_range->hmm_pfns = pfns;
>> hmm_range->start = start;
>> - hmm_range->end = start + npages * PAGE_SIZE;
>> + hmm_range_end = start + npages * PAGE_SIZE;
>
> This variable name is too easy to confuse with hmm_range->end. I would
> suggest calling it "end", analogous to "start".
[JZ] Sure
>
>
>> hmm_range->dev_private_owner = owner;
>> - /* Assuming 512MB takes maxmium 1 second to fault page address */
>> - timeout = max(npages >> 17, 1ULL) * HMM_RANGE_DEFAULT_TIMEOUT;
>> - timeout = jiffies + msecs_to_jiffies(timeout);
>> +#define MAX_WALK_BYTE (64ULL<<30)
>> + do {
>> + hmm_range->end = min(hmm_range->start + MAX_WALK_BYTE,
>> hmm_range_end);
>> +
>> + pr_debug("hmm range: start = 0x%lx, end = 0x%lx",
>> + hmm_range->start, hmm_range->end);
>> +
>> + /* Assuming 512MB takes maxmium 1 second to fault page
>> address */
>> + timeout = max((hmm_range->end - hmm_range->start) >> 29,
>> 1ULL) *
>> + HMM_RANGE_DEFAULT_TIMEOUT;
>> + timeout = jiffies + msecs_to_jiffies(timeout);
>> retry:
>> - hmm_range->notifier_seq = mmu_interval_read_begin(notifier);
>> - r = hmm_range_fault(hmm_range);
>> - if (unlikely(r)) {
>> - /*
>> - * FIXME: This timeout should encompass the retry from
>> - * mmu_interval_read_retry() as well.
>> - */
>> - if (r == -EBUSY && !time_after(jiffies, timeout))
>> - goto retry;
>> - goto out_free_pfns;
>> - }
>> + hmm_range->notifier_seq = mmu_interval_read_begin(notifier);
>> + r = hmm_range_fault(hmm_range);
>> + if (unlikely(r)) {
>> + /*
>> + * FIXME: This timeout should encompass the retry from
>> + * mmu_interval_read_retry() as well.
>> + */
>> + if (r == -EBUSY && !time_after(jiffies, timeout))
>> + goto retry;
>> + goto out_free_pfns;
>> + }
>> +
>> + hmm_range->hmm_pfns += MAX_WALK_BYTE >> PAGE_SHIFT;
>> + hmm_range->start = hmm_range->end;
>> + schedule();
>
> We don't need to schedule for ranges < MAX_WALK_BYTES. I guess these
> three lines could be conditional:
>
> + if (hmm_range->end == end)
> + break;
>
> hmm_range->hmm_pfns += MAX_WALK_BYTE >> PAGE_SHIFT;
> hmm_range->start = hmm_range->end;
> schedule();
[JZ] Yes, it will be more accurate. I will update it. Thanks!
>
> Regards,
> Felix
>
>
>> + } while (hmm_range->end < hmm_range_end);
>> +
>> + hmm_range->start = start;
>> + hmm_range->hmm_pfns = pfns;
>> /*
>> * Due to default_flags, all pages are HMM_PFN_VALID or
More information about the amd-gfx
mailing list