[Intel-gfx] [PATCH] drm/i915: apply PM_EARLY for non-GTT mappable objects
Thomas Hellström
thomas.hellstrom at linux.intel.com
Fri Feb 25 10:42:18 UTC 2022
On 2/25/22 11:34, Matthew Auld wrote:
> On DG2 we allow objects that are smaller than the min_page_size, under
> the premise that these are never mapped by the GTT, like with the paging
> structures. Currently the suspend-resume path will try to map such
> objects through the migration vm, which hits:
>
> [ 560.529217] kernel BUG at drivers/gpu/drm/i915/gt/intel_migrate.c:431!
> [ 560.536081] invalid opcode: 0000 [#1] PREEMPT SMP NOPTI
> [ 560.541629] CPU: 4 PID: 2062 Comm: rtcwake Tainted: G W 5.17.0-rc5-demarchi+ #175
> [ 560.550716] Hardware name: Intel Corporation CoffeeLake Client Platform/CoffeeLake S UDIMM RVP, BIOS CNLSFWR1.R00.X220.B00.2103302221 03/30/2021
> [ 560.563627] RIP: 0010:emit_pte+0x2e7/0x380 [i915]
> [ 560.568665] Code: ee 02 48 89 69 04 83 c6 05 83 c0 05 39 f0 0f 4f c6 48 8b 73 08 39 d0 0f 4f c2 44 89 f2 4c 8d 4a ff 49 85 f1 0f 84 62 fe ff ff <0f> 0b 48 c7 03 00 00 00 00 4d 89 c6 8b 01 48 29 ce 48 8d 57 0c 48
> [ 560.587691] RSP: 0018:ffffc9000104f8a0 EFLAGS: 00010206
> [ 560.592906] RAX: 0000000000000040 RBX: ffffc9000104f908 RCX: ffffc900025114d0
> [ 560.600024] RDX: 0000000000010000 RSI: 00000003f9fe2000 RDI: ffffc900025114dc
> [ 560.607458] RBP: 0000000001840000 R08: ffff88810f335540 R09: 000000000000ffff
> [ 560.614865] R10: 000000000000081b R11: 0000000000000001 R12: 000000000000081b
> [ 560.622300] R13: 0000000000000000 R14: 0000000000010000 R15: ffff888107c3e240
> [ 560.629716] FS: 00007f5b7c086580(0000) GS:ffff88846dc00000(0000) knlGS:0000000000000000
> [ 560.638090] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [ 560.644132] CR2: 00007f3ab0a133a8 CR3: 000000010a43e003 CR4: 00000000003706e0
> [ 560.651590] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> [ 560.659002] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> [ 560.666438] Call Trace:
> [ 560.668885] <TASK>
> [ 560.670983] intel_context_migrate_copy+0x1b1/0x4c0 [i915]
> [ 560.676794] __i915_ttm_move+0x628/0x790 [i915]
> [ 560.681704] ? dma_resv_iter_next+0x8f/0xb0
> [ 560.686223] ? dma_resv_iter_first+0xe5/0x140
> [ 560.690894] ? i915_deps_add_resv+0x4b/0x110 [i915]
> [ 560.696147] ? dma_resv_reserve_shared+0x161/0x310
> [ 560.701228] i915_gem_obj_copy_ttm+0x10f/0x220 [i915]
> [ 560.706650] i915_ttm_backup+0x191/0x2f0 [i915]
> [ 560.711558] i915_gem_process_region+0x266/0x3b0 [i915]
> [ 560.717153] ? verify_cpu+0xf0/0x100
> [ 560.721040] ? pci_pm_resume_early+0x20/0x20
> [ 560.725603] i915_ttm_backup_region+0x47/0x70 [i915]
> [ 560.730927] i915_gem_backup_suspend+0x141/0x170 [i91
>
> For now let's just force the memcpy path for such objects during
> suspend-resume.
>
> Fixes: 00e27ad85bc9 ("drm/i915/migrate: add acceleration support for DG2")
> Reported-by: Lucas De Marchi <lucas.demarchi at intel.com>
> Signed-off-by: Matthew Auld <matthew.auld at intel.com>
> Cc: Thomas Hellström <thomas.hellstrom at linux.intel.com>
> ---
> drivers/gpu/drm/i915/gem/i915_gem_region.c | 11 +++++++++++
> 1 file changed, 11 insertions(+)
>
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_region.c b/drivers/gpu/drm/i915/gem/i915_gem_region.c
> index a4350227e9ae..6d65c5849874 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_region.c
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_region.c
> @@ -67,6 +67,17 @@ i915_gem_object_create_region(struct intel_memory_region *mem,
> if (!obj)
> return ERR_PTR(-ENOMEM);
>
> + /*
> + * Anything smaller than the min_page_size can't be freely inserted into
> + * the GTT, due to alignemnt restrictions. For such special objects,
s/alignemnt/alignment/
Reviewed-by: Thomas Hellström <thomas.hellstrom at linux.intel.com>
> + * make sure we force memcpy based suspend-resume. In the future we can
> + * revisit this, either by allowing special mis-aligned objects in the
> + * migration path, or by mapping all of LMEM upfront using cheap 1G
> + * GTT entries.
> + */
> + if (default_page_size < mem->min_page_size)
> + flags |= I915_BO_ALLOC_PM_EARLY;
> +
> err = mem->ops->init_object(mem, obj, size, page_size, flags);
> if (err)
> goto err_object_free;
More information about the Intel-gfx
mailing list