Re: [PATCH] drm/ttm: only use DMA32 if needed for dummy_read_page

[Christian König]

Am 16.01.24 um 12:27 schrieb Yangyu Chen:
> On 1/16/24 17:45, Christian König wrote:
> > Am 16.01.24 um 08:24 schrieb Yangyu Chen:
> > > Some platforms may not have any memory in ZONE_DMA32 and use IOMMU 
> > > to allow
> > > 32-bit-DMA-only device to work. Forcing GFP_DMA32 on dummy_read_page 
> > > will
> > > fail in such platforms. Only use DMA32 when it must to get the bug
> > > resolved.
> >
> > Well that makes no sense.
> >
> > If a platform doesn't have a ZONE_DMA32 then GFP_DMA32 is just 
> > ignored as far as I know.
>
> Well, but it only applies when such ISA has no ZONE_DMA32 such as x86-32
> without PAE. It does not apply to some ISAs which has ZONE_DMA32 defined
> in the kernel but there is no memory below the 4GB address range. It
> happens on some RISC-V SoC chips and I know at least two chips have this
> issue.
>
>
> Below is some dmesg from my platform, using such RISC-V without 
> ZONE_DMA32 with AMD RX550 GPU:
>
> [    0.000000] Linux version 6.7.0-dirty (cyy@cyy-pc) 
> (riscv64-linux-gnu-gcc (Debian 13.2.0-7) 13.2.0, GNU ld (GNU Binutils 
> for Debian) 2.41.50.20231227) #12 SMP Tue Jan 16 18:55:13 CST 2024
> [    0.000000] Machine model:
> [    0.000000] SBI specification v2.0 detected
> [    0.000000] SBI implementation ID=0x1 Version=0x10004
> [    0.000000] SBI TIME extension detected
> [    0.000000] SBI IPI extension detected
> [    0.000000] SBI RFENCE extension detected
> [    0.000000] efi: UEFI not found.
> [    0.000000] OF: reserved mem: 
> 0x0000002000000000..0x000000200003ffff (256 KiB) nomap non-reusable 
> mmode_resv1@20,0
> [    0.000000] OF: reserved mem: 
> 0x0000002000040000..0x000000200005ffff (128 KiB) nomap non-reusable 
> mmode_resv0@20,40000
> [    0.000000] Zone ranges:
> [    0.000000]   DMA32    empty
> [    0.000000]   Normal   [mem 0x0000002000000000-0x00000021ffffffff]
>
> [   35.738619] [drm] amdgpu kernel modesetting enabled.
> [   35.743916] [drm] initializing kernel modesetting (POLARIS12 
> 0x1002:0x699F 0x1043:0x0513 0xC7).
> [   35.752671] [drm] register mmio base: 0xA8100000
> [   35.763558] [drm] register mmio size: 262144
> [   35.768491] [drm] add ip block number 0 <vi_common>
> [   35.773403] [drm] add ip block number 1 <gmc_v8_0>
> [   35.773407] [drm] add ip block number 2 <tonga_ih>
> [   35.773410] [drm] add ip block number 3 <gfx_v8_0>
> [   35.773412] [drm] add ip block number 4 <sdma_v3_0>
> [   35.773415] [drm] add ip block number 5 <powerplay>
> [   35.773418] [drm] add ip block number 6 <dm>
> [   35.806668] [drm] add ip block number 7 <uvd_v6_0>
> [   35.806671] [drm] add ip block number 8 <vce_v3_0>
> [   36.014036] amdgpu 0000:05:00.0: amdgpu: Fetched VBIOS from ROM BAR
> [   36.020322] amdgpu: ATOM BIOS: 115-C994PI2-100
> [   36.043586] [drm] UVD is enabled in VM mode
> [   36.047813] [drm] UVD ENC is enabled in VM mode
> [   36.052370] [drm] VCE enabled in VM mode
> [   36.052377] amdgpu 0000:05:00.0: amdgpu: Trusted Memory Zone (TMZ) 
> feature not supported
> [   36.064404] amdgpu 0000:05:00.0: amdgpu: PCIE atomic ops is not 
> supported
> [   36.071209] [drm] GPU posting now...
> [   36.190029] [drm] vm size is 64 GB, 2 levels, block size is 10-bit, 
> fragment size is 9-bit
> [   36.199385] amdgpu 0000:05:00.0: amdgpu: VRAM: 4096M 
> 0x000000F400000000 - 0x000000F4FFFFFFFF (4096M used)
> [   36.208978] amdgpu 0000:05:00.0: amdgpu: GART: 256M 
> 0x000000FF00000000 - 0x000000FF0FFFFFFF
> [   36.217345] [drm] Detected VRAM RAM=4096M, BAR=256M
> [   36.222226] [drm] RAM width 128bits GDDR5
> [   36.226288] warn_alloc: 8 callbacks suppressed
> [   36.226290] (udev-worker): page allocation failure: order:0, 
> mode:0x104(GFP_DMA32|__GFP_ZERO), nodemask=(null),cpuset=/,mems_allowed=0
> [   36.242837] CPU: 3 PID: 165 Comm: (udev-worker) Not tainted 
> 6.7.0-dirty #12
> [   36.249796] Hardware name: (DT)
> [   36.254670] Call Trace:
> [   36.257113] [<ffffffff80006056>] dump_backtrace+0x1c/0x24
> [   36.262513] [<ffffffff8096b860>] show_stack+0x2c/0x38
> [   36.267562] [<ffffffff80977f14>] dump_stack_lvl+0x3c/0x54
> [   36.272958] [<ffffffff80977f40>] dump_stack+0x14/0x1c
> [   36.278006] [<ffffffff8018e66c>] warn_alloc+0xee/0x16c
> [   36.283143] [<ffffffff8018f17a>] __alloc_pages+0xa90/0xb40
> [   36.288627] [<ffffffff0174f5fe>] ttm_global_init+0x12a/0x1ac [ttm]
> [   36.294830] [<ffffffff0174f6b2>] ttm_device_init+0x32/0x158 [ttm]
> [   36.300935] [<ffffffff018448bc>] amdgpu_ttm_init+0x7a/0x638 [amdgpu]
> [   36.310260] [<ffffffff01b4c8b8>] amdgpu_bo_init+0x76/0x82 [amdgpu]
> [   36.319323] [<ffffffff018c2b6e>] gmc_v8_0_sw_init+0x36c/0x652 [amdgpu]
> [   36.328710] [<ffffffff018328fa>] amdgpu_device_init+0x1648/0x20fa 
> [amdgpu]
> [   36.338395] [<ffffffff01834c1a>] amdgpu_driver_load_kms+0x1e/0x158 
> [amdgpu]
> [   36.348164] [<ffffffff0182b9d4>] amdgpu_pci_probe+0x124/0x46c [amdgpu]
> [   36.357543] [<ffffffff804ae3ae>] pci_device_probe+0x7a/0xf0
> [   36.363115] [<ffffffff8058cfb6>] really_probe+0x86/0x242
> [   36.368427] [<ffffffff8058d1ce>] __driver_probe_device+0x5c/0xda
> [   36.374431] [<ffffffff8058d278>] driver_probe_device+0x2c/0xb2
> [   36.380260] [<ffffffff8058d3f8>] __driver_attach+0x6c/0x11a
> [   36.385829] [<ffffffff8058b14c>] bus_for_each_dev+0x60/0xae
> [   36.391398] [<ffffffff8058ca08>] driver_attach+0x1a/0x22
> [   36.396707] [<ffffffff8058c368>] bus_add_driver+0xd0/0x1ba
> [   36.402189] [<ffffffff8058e046>] driver_register+0x3e/0xd8
> [   36.407671] [<ffffffff804ad04a>] __pci_register_driver+0x58/0x62
> [   36.413679] [<ffffffff0207e078>] amdgpu_init+0x78/0x1000 [amdgpu]
> [   36.422616] [<ffffffff8000212c>] do_one_initcall+0x58/0x19c
> [   36.428188] [<ffffffff80086d22>] do_init_module+0x4e/0x1b0
> [   36.433672] [<ffffffff80088456>] load_module+0x1374/0x1768
> [   36.439155] [<ffffffff80088a10>] init_module_from_file+0x76/0xaa
> [   36.445158] [<ffffffff80088c2c>] __riscv_sys_finit_module+0x1cc/0x2ba
> [   36.451597] [<ffffffff80978952>] do_trap_ecall_u+0xba/0x12e
> [   36.457167] [<ffffffff809819f4>] ret_from_exception+0x0/0x64
> [   36.462847] Mem-Info:
> [   36.465124] active_anon:41 inactive_anon:4533 isolated_anon:0
>                 active_file:12928 inactive_file:24854 isolated_file:0
>                 unevictable:0 dirty:176 writeback:0
>                 slab_reclaimable:3390 slab_unreclaimable:3922
>                 mapped:4609 shmem:333 pagetables:564
>                 sec_pagetables:0 bounce:0
>                 kernel_misc_reclaimable:0
>                 free:1983561 free_pcp:0 free_cma:0
> [   36.490913] Node 0 hugepages_total=0 hugepages_free=0 
> hugepages_surp=0 hugepages_size=2048kB
> [   36.508722] 38077 total pagecache pages
> [   36.512560] 0 pages in swap cache
> [   36.515875] Free swap  = 0kB
> [   36.518757] Total swap = 0kB
> [   36.521633] 2097152 pages RAM
> [   36.524601] 0 pages HighMem/MovableOnly
> [   36.528438] 45269 pages reserved
> [   36.531699] [drm:amdgpu_ttm_init [amdgpu]] *ERROR* failed 
> initializing buffer object driver(-12).
> [   36.543451] [drm:amdgpu_device_init [amdgpu]] *ERROR* sw_init of IP 
> block <gmc_v8_0> failed -12
> [   36.555011] amdgpu 0000:05:00.0: amdgpu: amdgpu_device_ip_init failed
> [   36.561458] amdgpu 0000:05:00.0: amdgpu: Fatal error during GPU init
> [   36.567816] amdgpu 0000:05:00.0: amdgpu: amdgpu: finishing device.

Interesting, I think I remember seeing something like this on ARM as 
well but we didn't run into any problems with GFP_DMA32 there.


> > This patch here won't work since the use_dma32 flag is a per device 
> > flag which can't be used for the global initialization.
> >
> > Otherwise this can randomly fail depending on if a DMA32 device 
> > initializes first or after some device with larger addressing 
> > capabilities. This configuration is quite common on older 
> > motherboards with both integrated and dedicated graphics.
> >
> > Regards,
> > Christian.
>
> Thanks for your review. However, the bug I mentioned above still needs
> to be resolved. I want to submit patch v2 with a different solution 
> that tries to allocate with GFP_DMA32 flag, if it fails, then try to 
> allocate without GFP_DMA32 flag. Would this solution be acceptable?

Yeah, that's probably a good idea.

I mean in general we should avoid GFP_DMA32 altogether here, but for 
this we would need a function to get memory with the best possible 
accessibility for all devices.

Regards,
Christian.

> Thanks,
> Yangyu Chen
>
> > > Signed-off-by: Yangyu Chen <cyy@xxxxxxxxxxxx>
> > > ---
> > >   drivers/gpu/drm/ttm/ttm_device.c | 7 ++++---
> > >   1 file changed, 4 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/ttm/ttm_device.c 
> > > b/drivers/gpu/drm/ttm/ttm_device.c
> > > index d48b39132b32..62f16fb72428 100644
> > > --- a/drivers/gpu/drm/ttm/ttm_device.c
> > > +++ b/drivers/gpu/drm/ttm/ttm_device.c
> > > @@ -63,7 +63,7 @@ static void ttm_global_release(void)
> > >       mutex_unlock(&ttm_global_mutex);
> > >   }
> > > -static int ttm_global_init(void)
> > > +static int ttm_global_init(bool use_dma32)
> > >   {
> > >       struct ttm_global *glob = &ttm_glob;
> > >       unsigned long num_pages, num_dma32;
> > > @@ -95,7 +95,8 @@ static int ttm_global_init(void)
> > >       ttm_pool_mgr_init(num_pages);
> > >       ttm_tt_mgr_init(num_pages, num_dma32);
> > > -    glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
> > > +    glob->dummy_read_page = use_dma32 ? alloc_page(__GFP_ZERO | 
> > > GFP_DMA32) :
> > > +                        alloc_page(__GFP_ZERO);
> > >       if (unlikely(glob->dummy_read_page == NULL)) {
> > >           ret = -ENOMEM;
> > > @@ -200,7 +201,7 @@ int ttm_device_init(struct ttm_device *bdev, 
> > > const struct ttm_device_funcs *func
> > >       if (WARN_ON(vma_manager == NULL))
> > >           return -EINVAL;
> > > -    ret = ttm_global_init();
> > > +    ret = ttm_global_init(use_dma32);
> > >       if (ret)
> > >           return ret;