Re: [PATCH] drm/etnaviv: Remove GFP_HIGHUSER in systems with 32 address limits

[Lucas Stach <l.stach@pengutronix.de>]

Hi Xiaolei,

Am Dienstag, dem 06.08.2024 um 18:47 +0800 schrieb Xiaolei Wang:
> GFP_HIGHUSER is for userspace allocations that may be mapped
> to userspace,An example may be a hardware allocation that maps
> data directly into userspace but has no addressing limitations,
> this conflicts with GFP_DMA32,The kernel reports a BUG:
> 
GFP_HIGHUSER is a combination of GFP_USER | __GFP_HIGHMEM. Only the
highmem part is incompatible with DMA32. You don't want to clear the
GFP_USER bit here, as the driver allocated buffers might be mapped to
userspace.

Regards,
Lucas

> kernel BUG at include/linux/gfp.h:139!
> Internal error: Oops - BUG: 00000000f2000800 [#1] PREEMPT SMP
> Modules linked in:
> Hardware name: NXP i.MX8MPlus EVK board (DT)
> pstate: 40000005 (nZcv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
>  pc : __alloc_pages_noprof+0x5d8/0x72c
>  lr : alloc_pages_mpol_noprof+0x100/0x4e0
>  sp : ffffffc08c6a71c0
>  x29: ffffffc08c6a71c0 x28: ffffffc086e46000 x27: ffffffc086e46a68
>  x26: 1ffffff81122b260 x25: ffffffc089159304 x24: ffffff80da938000
>  x23: 0000000000000000 x22: 0000000000000000 x21: ffffff80da938000
>  x20: 1ffffff8118d4e46 x19: 0000000000146cc6 x18: 0000000000000000
>  x17: ffffffc081b00980 x16: ffffffc081b002a8 x15: 1ffffff8118d4e56
>  x14: 00000000f1f1f1f1 x13: 00000000f3f3f300 x12: 0000000000000000
>  x11: ffffff80da9384c8 x10: ffffff80da938000 x9 : 00000000f2f2f200
>  x8 : 0000000041b58ab3 x7 : 00000000f3000000 x6 : 00000000f3f3f3f3
>  x5 : 1ffffff01b527005 x4 : 000000000000000c x3 : 0000000000000006
>  x2 : 0000000000000000 x1 : 00000000000003a3 x0 : 0000000000000000
>  Call trace:
>   __alloc_pages_noprof+0x5d8/0x72c
>   alloc_pages_mpol_noprof+0x100/0x4e0
>   folio_alloc_mpol_noprof+0x18/0xb8
>   shmem_alloc_folio+0x154/0x1a8
>   shmem_alloc_and_add_folio+0x180/0xee8
>   shmem_get_folio_gfp+0x660/0x103c
>   shmem_read_folio_gfp+0x98/0x104
>   drm_gem_get_pages+0x174/0x5ac
>   etnaviv_gem_shmem_get_pages+0x18/0x5c
>   etnaviv_gem_get_pages+0x100/0x328
>   etnaviv_gem_cpu_prep+0x2e8/0x438
>   etnaviv_ioctl_gem_cpu_prep+0xb0/0x1ac
>   drm_ioctl_kernel+0x158/0x2c8
>   drm_ioctl+0x494/0xb48
>   __arm64_sys_ioctl+0x120/0x18c
>   invoke_syscall+0x6c/0x25c
>   el0_svc_common.constprop.0+0x174/0x278
>   do_el0_svc+0x40/0x58
>   el0_svc+0x50/0xc0
>   el0t_64_sync_handler+0xc0/0xc4
>   el0t_64_sync+0x190/0x194
>  Code: 52800021 39003c01 d4210000 17ffff57 (d4210000)
> 
> Fixes: b72af445cd38 ("drm/etnaviv: request pages from DMA32 zone when needed")
> Signed-off-by: Xiaolei Wang <xiaolei.wang@windriver.com>
> ---
>  drivers/gpu/drm/etnaviv/etnaviv_gpu.c | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> index 7c7f97793ddd..c3f329226bed 100644
> --- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> +++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> @@ -844,8 +844,10 @@ int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
>  	 * request pages for our SHM backend buffers from the DMA32 zone to
>  	 * hopefully avoid performance killing SWIOTLB bounce buffering.
>  	 */
> -	if (dma_addressing_limited(gpu->dev))
> -		priv->shm_gfp_mask |= GFP_DMA32;
> +	if (dma_addressing_limited(gpu->dev)) {
> +		priv->shm_gfp_mask |= GFP_DMA32 & GFP_USER;
> +		priv->shm_gfp_mask &= ~GFP_HIGHUSER;
> +	}
>  
>  	/* Create buffer: */
>  	ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer,