On Wed, Feb 26, 2025 at 06:05:55PM +0000, Matthew Auld wrote:
> On 26/02/2025 15:12, Matthew Brost wrote:
> > On Wed, Feb 26, 2025 at 10:38:40AM +0000, Matthew Auld wrote:
> > > On 26/02/2025 04:18, Matthew Brost wrote:
> > > > On Tue, Feb 25, 2025 at 09:13:09PM -0600, Lucas De Marchi wrote:
> > > > > On Wed, Feb 26, 2025 at 10:00:18AM +0800, Mingcong Bai via B4 Relay wrote:
> > > > > > From: Mingcong Bai <jeffbai@xxxxxxx>
> > > > > >
> > > > > > The bo/ttm interfaces with kernel memory mapping from dedicated GPU
> > > > > > memory. It is not correct to assume that SZ_4K would suffice for page
> > > > > > alignment as there are a few hardware platforms that commonly uses non-4K
> > > > > > pages - for instance, currently, Loongson 3A5000/6000 devices (of the
> > > > > > LoongArch architecture) commonly uses 16K kernel pages.
> > > > > >
> > > > > > Per my testing Intel Xe/Arc families of GPUs works on at least
> > > > > > Loongson 3A6000 platforms so long as "Above 4G Decoding" and "Resizable
> > > > > > BAR" were enabled in the EFI firmware settings. I tested this patch series
> > > > > > on my Loongson XA61200 (3A6000) motherboard with an Intel Arc A750 GPU.
> > > > > >
> > > > > > Without this fix, the kernel will hang at a kernel BUG():
> > > > > >
> > > > > > [ 7.425445] ------------[ cut here ]------------
> > > > > > [ 7.430032] kernel BUG at drivers/gpu/drm/drm_gem.c:181!
> > > > > > [ 7.435330] Oops - BUG[#1]:
> > > > > > [ 7.438099] CPU: 0 UID: 0 PID: 102 Comm: kworker/0:4 Tainted: G E 6.13.3-aosc-main-00336-g60829239b300-dirty #3
> > > > > > [ 7.449511] Tainted: [E]=UNSIGNED_MODULE
> > > > > > [ 7.453402] Hardware name: Loongson Loongson-3A6000-HV-7A2000-1w-V0.1-EVB/Loongson-3A6000-HV-7A2000-1w-EVB-V1.21, BIOS Loongson-UDK2018-V4.0.05756-prestab
> > > > > > [ 7.467144] Workqueue: events work_for_cpu_fn
> > > > > > [ 7.471472] pc 9000000001045fa4 ra ffff8000025331dc tp 90000001010c8000 sp 90000001010cb960
> > > > > > [ 7.479770] a0 900000012a3e8000 a1 900000010028c000 a2 000000000005d000 a3 0000000000000000
> > > > > > [ 7.488069] a4 0000000000000000 a5 0000000000000000 a6 0000000000000000 a7 0000000000000001
> > > > > > [ 7.496367] t0 0000000000001000 t1 9000000001045000 t2 0000000000000000 t3 0000000000000000
> > > > > > [ 7.504665] t4 0000000000000000 t5 0000000000000000 t6 0000000000000000 t7 0000000000000000
> > > > > > [ 7.504667] t8 0000000000000000 u0 90000000029ea7d8 s9 900000012a3e9360 s0 900000010028c000
> > > > > > [ 7.504668] s1 ffff800002744000 s2 0000000000000000 s3 0000000000000000 s4 0000000000000001
> > > > > > [ 7.504669] s5 900000012a3e8000 s6 0000000000000001 s7 0000000000022022 s8 0000000000000000
> > > > > > [ 7.537855] ra: ffff8000025331dc ___xe_bo_create_locked+0x158/0x3b0 [xe]
> > > > > > [ 7.544893] ERA: 9000000001045fa4 drm_gem_private_object_init+0xcc/0xd0
> > > > > > [ 7.551639] CRMD: 000000b0 (PLV0 -IE -DA +PG DACF=CC DACM=CC -WE)
> > > > > > [ 7.557785] PRMD: 00000004 (PPLV0 +PIE -PWE)
> > > > > > [ 7.562111] EUEN: 00000000 (-FPE -SXE -ASXE -BTE)
> > > > > > [ 7.566870] ECFG: 00071c1d (LIE=0,2-4,10-12 VS=7)
> > > > > > [ 7.571628] ESTAT: 000c0000 [BRK] (IS= ECode=12 EsubCode=0)
> > > > > > [ 7.577163] PRID: 0014d000 (Loongson-64bit, Loongson-3A6000-HV)
> > > > > > [ 7.583128] Modules linked in: xe(E+) drm_gpuvm(E) drm_exec(E) drm_buddy(E) gpu_sched(E) drm_suballoc_helper(E) drm_display_helper(E) loongson(E) r8169(E) cec(E) rc_core(E) realtek(E) i2c_algo_bit(E) tpm_tis_spi(E) led_class(E) hid_generic(E) drm_ttm_helper(E) ttm(E) drm_client_lib(E) drm_kms_helper(E) sunrpc(E) la_ow_syscall(E) i2c_dev(E)
> > > > > > [ 7.613049] Process kworker/0:4 (pid: 102, threadinfo=00000000bc26ebd1, task=0000000055480707)
> > > > > > [ 7.621606] Stack : 0000000000000000 3030303a6963702b 000000000005d000 0000000000000000
> > > > > > [ 7.629563] 0000000000000001 0000000000000000 0000000000000000 8e1bfae42b2f7877
> > > > > > [ 7.637519] 000000000005d000 900000012a3e8000 900000012a3e9360 0000000000000000
> > > > > > [ 7.645475] ffffffffffffffff 0000000000000000 0000000000022022 0000000000000000
> > > > > > [ 7.653431] 0000000000000001 ffff800002533660 0000000000022022 9000000000234470
> > > > > > [ 7.661386] 90000001010cba28 0000000000001000 0000000000000000 000000000005c300
> > > > > > [ 7.669342] 900000012a3e8000 0000000000000000 0000000000000001 900000012a3e8000
> > > > > > [ 7.677298] ffffffffffffffff 0000000000022022 900000012a3e9498 ffff800002533a14
> > > > > > [ 7.685254] 0000000000022022 0000000000000000 900000000209c000 90000000010589e0
> > > > > > [ 7.693209] 90000001010cbab8 ffff8000027c78c0 fffffffffffff000 900000012a3e8000
> > > > > > [ 7.701165] ...
> > > > > > [ 7.703588] Call Trace:
> > > > > > [ 7.703590] [<9000000001045fa4>] drm_gem_private_object_init+0xcc/0xd0
> > > > > > [ 7.712496] [<ffff8000025331d8>] ___xe_bo_create_locked+0x154/0x3b0 [xe]
> > > > > > [ 7.719268] [<ffff80000253365c>] __xe_bo_create_locked+0x228/0x304 [xe]
> > > > > > [ 7.725951] [<ffff800002533a10>] xe_bo_create_pin_map_at_aligned+0x70/0x1b0 [xe]
> > > > > > [ 7.733410] [<ffff800002533c7c>] xe_managed_bo_create_pin_map+0x34/0xcc [xe]
> > > > > > [ 7.740522] [<ffff800002533d58>] xe_managed_bo_create_from_data+0x44/0xb0 [xe]
> > > > > > [ 7.747807] [<ffff80000258d19c>] xe_uc_fw_init+0x3ec/0x904 [xe]
> > > > > > [ 7.753814] [<ffff80000254a478>] xe_guc_init+0x30/0x3dc [xe]
> > > > > > [ 7.759553] [<ffff80000258bc04>] xe_uc_init+0x20/0xf0 [xe]
> > > > > > [ 7.765121] [<ffff800002542abc>] xe_gt_init_hwconfig+0x5c/0xd0 [xe]
> > > > > > [ 7.771461] [<ffff800002537204>] xe_device_probe+0x240/0x588 [xe]
> > > > > > [ 7.777627] [<ffff800002575448>] xe_pci_probe+0x6c0/0xa6c [xe]
> > > > > > [ 7.783540] [<9000000000e9828c>] local_pci_probe+0x4c/0xb4
> > > > > > [ 7.788989] [<90000000002aa578>] work_for_cpu_fn+0x20/0x40
> > > > > > [ 7.794436] [<90000000002aeb50>] process_one_work+0x1a4/0x458
> > > > > > [ 7.800143] [<90000000002af5a0>] worker_thread+0x304/0x3fc
> > > > > > [ 7.805591] [<90000000002bacac>] kthread+0x114/0x138
> > > > > > [ 7.810520] [<9000000000241f64>] ret_from_kernel_thread+0x8/0xa4
> > > > > > [ 7.816489]
> > > > > > [ 7.817961] Code: 4c000020 29c3e2f9 53ff93ff <002a0001> 0015002c 03400000 02ff8063 29c04077 001500f7
> > > > > > [ 7.827651]
> > > > > > [ 7.829140] ---[ end trace 0000000000000000 ]---
> > > > > >
> > > > > > Revise all instances of `SZ_4K' with `PAGE_SIZE' and revise the call to
> > > > > > `drm_gem_private_object_init()' in `*___xe_bo_create_locked()' (last call
> > > > > > before BUG()) to use `size_t aligned_size' calculated from `PAGE_SIZE' to
> > > > > > fix the above error.
> > > > > >
> > > > > > Cc: <stable@xxxxxxxxxxxxxxx>
> > > > > > Fixes: 4e03b584143e ("drm/xe/uapi: Reject bo creation of unaligned size")
> > > > > > Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
> > > > > > Tested-by: Mingcong Bai <jeffbai@xxxxxxx>
> > > > > > Tested-by: Haien Liang <27873200@xxxxxx>
> > > > > > Tested-by: Shirong Liu <lsr1024@xxxxxx>
> > > > > > Tested-by: Haofeng Wu <s2600cw2@xxxxxxx>
> > > > > > Link: https://github.com/FanFansfan/loongson-linux/commit/22c55ab3931c32410a077b3ddb6dca3f28223360
> > > > > > Co-developed-by: Shang Yatsen <429839446@xxxxxx>
> > > > > > Signed-off-by: Shang Yatsen <429839446@xxxxxx>
> > > > > > Signed-off-by: Mingcong Bai <jeffbai@xxxxxxx>
> > > > > > ---
> > > > > > drivers/gpu/drm/xe/xe_bo.c | 8 ++++----
> > > > > > 1 file changed, 4 insertions(+), 4 deletions(-)
> > > > > >
> > > > > > diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
> > > > > > index 3f5391d416d469c636d951dd6f0a2b3b5ae95dab..dd03c581441f352eff51d0eafe1298fca7d9653d 100644
> > > > > > --- a/drivers/gpu/drm/xe/xe_bo.c
> > > > > > +++ b/drivers/gpu/drm/xe/xe_bo.c
> > > > > > @@ -1441,9 +1441,9 @@ struct xe_bo *___xe_bo_create_locked(struct xe_device *xe, struct xe_bo *bo,
> > > > > > flags |= XE_BO_FLAG_INTERNAL_64K;
> > > > > > alignment = align >> PAGE_SHIFT;
> > > > > > } else {
> > > >
> > > > } else if (type == ttm_bo_type_device) {
> > > > new code /w PAGE_SIZE
> > > > } else {
> > > > old code /w SZ_4K (or maybe XE_PAGE_SIZE now)?
> > > > }
> > > >
> > > > See below for further explaination.
> > > >
> > > > > > - aligned_size = ALIGN(size, SZ_4K);
> > > > > > + aligned_size = ALIGN(size, PAGE_SIZE);
> > > > >
> > > > > in the very beginning of the driver we were set to use XE_PAGE_SIZE
> > > > > for things like this. It seems thing went side ways though.
> > > > >
> > > > > Thanks for fixing these. XE_PAGE_SIZE is always 4k, but I think we should
> > > > > uxe XE_PAGE_SIZE for clarity. For others in Cc... any thoughts?
> > > > >
> > > >
> > > > It looks like you have a typo here, Lucas. Could you please clarify?
> > > >
> > > > However, XE_PAGE_SIZE should always be 4k, as it refers to the GPU page
> > > > size, which is fixed.
> > > >
> > > > I think using PAGE_SIZE makes sense in some cases. See my other
> > > > comments.
> > > >
> > > > > > flags &= ~XE_BO_FLAG_INTERNAL_64K;
> > > > > > - alignment = SZ_4K >> PAGE_SHIFT;
> > > > > > + alignment = PAGE_SIZE >> PAGE_SHIFT;
> > > > > > }
> > > > > >
> > > > > > if (type == ttm_bo_type_device && aligned_size != size)
> > > > > > @@ -1457,7 +1457,7 @@ struct xe_bo *___xe_bo_create_locked(struct xe_device *xe, struct xe_bo *bo,
> > > > > >
> > > > > > bo->ccs_cleared = false;
> > > > > > bo->tile = tile;
> > > > > > - bo->size = size;
> > > > > > + bo->size = aligned_size;
> > > > >
> > > > > the interface of this function is that the caller needs to pass the
> > > > > correct size, it's not really expected the function will adjust it and
> > > > > the check is there to gurantee to return the appropriate error. There
> > > >
> > > > Let me expand further on Lucas's comment. We reject user BOs that are
> > > > unaligned here in ___xe_bo_create_locked.
> > > >
> > > > 1490 if (type == ttm_bo_type_device && aligned_size != size)
> > > > 1491 return ERR_PTR(-EINVAL);
> > > >
> > > > What we allow are kernel BOs (!= ttm_bo_type_device), which are never
> > > > mapped to the CPU or the PPGTT (user GPU mappings), to be a smaller
> > > > size. Examples of this include memory for GPU page tables, LRC state,
> > > > etc. Memory for GPU page tables is always allocated in 4k blocks, so
> > > > changing the allocation to the CPU page size of 16k or 64k would be
> > > > wasteful.
> > > >
> > > > AFAIK, kernel memory is always a VRAM allocation, so we don't have any
> > > > CPU page size requirements. If this is not true (I haven't checked), or
> > > > perhaps just to future-proof, change the snippet in my first comment to:
> > > >
> > > > } else if (type == ttm_bo_type_device || flags & XE_BO_FLAG_SYSTEM)) {
> > > > new code /w PAGE_SIZE
> > > > } else {
> > > > old code /w SZ_4K
> > > > }
> > > >
> > > > Then change BO assignment size too:
> > > >
> > > > bo->size = flags & XE_BO_FLAG_SYSTEM ? aligned_size : size;
> > > >
> > > > This should enable kernel VRAM allocations to be smaller than the CPU
> > > > page size (I think). Can you try out this suggestion and see if the Xe
> > > > boots with non-4k pages?
> > >
> > > The vram allocator chunk size is PAGE_SIZE so that would also need some
> > > attention, I think.
> > >
> >
> > Agree. So I think __xe_ttm_vram_mgr_init should be called with
> > s/PAGE_SIZE/SZ_4K?
>
> Should be fine from allocator pov. But also need to update the upper layers
> in the VRAM manager itself, I think.
>
Right. A lot of pfn logic is based on PAGE_SIZE/SHIFT and that would
likely need to be updated too.
> >
> > > But I think we also then need to deal with the assert in: https://elixir.bootlin.com/linux/v6.14-rc4/source/drivers/gpu/drm/drm_gem.c#L181.
> > >
> >
> > Yep. I think that would need to be adjusted as well to be bypassed if we
> > are never going to CPU map the BO—specifically, CPU map it to user space
> > or if the BO is not in VRAM. For kernel VRAM mapping, this resolves to
> > an offset within an existing large PCIe BAR mapping, so allocations
> > unaligned to PAGE_SIZE should work.
>
> Yeah, agree. I thinks it's possible.
>
Yes, might be a little more difficult than I originally thought though.
> >
> > Maybe export __drm_gem_private_object_init, which skips the BUG_ON, and
> > call this in Xe to avoid interfering with other drivers' expectations?
>
> Some other places I spotted are the VRAM manager, and stuff like
> xe_bo_vmap() and then into TTM itself. So it might be quite widespread.
>
Ok, so maybe for initial merge we drop this idea and circle back given
the complexity as we'd just be wasting memory for things like PTEs?
Matt
> >
> > Matt
> >
> > > >
> > > > Also others in Cc... thoughts / double check my input?
> > > >
> > > > > are other places that would need some additional fixes leading to this
> > > > > function. Example:
> > > > >
> > > > > xe_gem_create_ioctl()
> > > > > {
> > > > > ...
> > > > > if (XE_IOCTL_DBG(xe, args->size & ~PAGE_MASK))
> > > > > return -EINVAL;
> > > >
> > > > This actually looks right, the minimum allocation size for user BOs
> > > > should be PAGE_SIZE aligned. The last patch in the series fixes the
> > > > query for this.
> > > >
> > > > Matt
> > > >
> > > > > ...
> > > > > }
> > > > >
> > > > >
> > > > > Lucas De Marchi
> > > > >
> > > > > > bo->flags = flags;
> > > > > > bo->cpu_caching = cpu_caching;
> > > > > > bo->ttm.base.funcs = &xe_gem_object_funcs;
> > > > > > @@ -1468,7 +1468,7 @@ struct xe_bo *___xe_bo_create_locked(struct xe_device *xe, struct xe_bo *bo,
> > > > > > #endif
> > > > > > INIT_LIST_HEAD(&bo->vram_userfault_link);
> > > > > >
> > > > > > - drm_gem_private_object_init(&xe->drm, &bo->ttm.base, size);
> > > > > > + drm_gem_private_object_init(&xe->drm, &bo->ttm.base, aligned_size);
> > > > > >
> > > > > > if (resv) {
> > > > > > ctx.allow_res_evict = !(flags & XE_BO_FLAG_NO_RESV_EVICT);
> > > > > >
> > > > > > --
> > > > > > 2.48.1
> > > > > >
> > > > > >
> > >
>
