Re: [PATCH] drm/shmem-helper: Fix Map huge page mapping in fault handler

[Boris Brezillon <boris.brezillon@collabora.com>]

On Thu, 12 Mar 2026 12:50:27 -0300
Pedro Demarchi Gomes <pedrodemargomes@gmail.com> wrote:

> When running ./tools/testing/selftests/mm/split_huge_page_test multiple
> times with /sys/kernel/mm/transparent_hugepage/shmem_enabled and
> /sys/kernel/mm/transparent_hugepage/enabled set as always the following BUG
> occurs:
> 
> [  232.728858] ------------[ cut here ]------------
> [  232.729458] kernel BUG at mm/memory.c:2276!
> [  232.729726] Oops: invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC KASAN PTI
> [  232.730217] CPU: 19 UID: 60578 PID: 1497 Comm: llvmpipe-9 Not tainted 7.0.0-rc1mm-new+ #19 PREEMPT(lazy)
> [  232.730855] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.17.0-9.fc43 06/10/2025
> [  232.731360] RIP: 0010:walk_to_pmd+0x29e/0x3c0
> [  232.731569] Code: d8 5b 5d 41 5c 41 5d 41 5e 41 5f c3 cc cc cc cc 48 89 ea 48 89 de 4c 89 f7 e8 ae 85 ff ff 85 c0 0f 84 1f fe ff ff 31 db eb d0 <0f> 0b 48 89 ea 48 89 de 4c 89 f7 e8 92 8b ff ff 85 c0 75 e8 48 b8
> [  232.732614] RSP: 0000:ffff8881aa6ff9a8 EFLAGS: 00010282
> [  232.732991] RAX: 8000000142e002e7 RBX: ffff8881433cae10 RCX: dffffc0000000000
> [  232.733362] RDX: 0000000000000000 RSI: 00007fb47840b000 RDI: 8000000142e002e7
> [  232.733801] RBP: 00007fb47840b000 R08: 0000000000000000 R09: 1ffff110354dff46
> [  232.734168] R10: fffffbfff0cb921d R11: 00000000910da5ce R12: 1ffffffff0c1fcdd
> [  232.734459] R13: 1ffffffff0c23f36 R14: ffff888171628040 R15: 0000000000000000
> [  232.734861] FS:  00007fb4907f86c0(0000) GS:ffff888791f2c000(0000) knlGS:0000000000000000
> [  232.735265] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [  232.735548] CR2: 00007fb47840be00 CR3: 000000015e6dc000 CR4: 00000000000006f0
> [  232.736031] Call Trace:
> [  232.736273]  <TASK>
> [  232.736500]  get_locked_pte+0x1f/0xa0
> [  232.736878]  insert_pfn+0x9f/0x350
> [  232.737190]  ? __pfx_pat_pagerange_is_ram+0x10/0x10
> [  232.737614]  ? __pfx_insert_pfn+0x10/0x10
> [  232.737990]  ? __pfx_css_rstat_updated+0x10/0x10
> [  232.738281]  ? __pfx_pfn_modify_allowed+0x10/0x10
> [  232.738552]  ? lookup_memtype+0x62/0x180
> [  232.738761]  vmf_insert_pfn_prot+0x14b/0x340
> [  232.739012]  ? __pfx_vmf_insert_pfn_prot+0x10/0x10
> [  232.739247]  ? __pfx___might_resched+0x10/0x10
> [  232.739475]  drm_gem_shmem_fault.cold+0x18/0x39
> [  232.739677]  ? rcu_read_unlock+0x20/0x70
> [  232.739882]  __do_fault+0x251/0x7b0
> [  232.740028]  do_fault+0x6e1/0xc00
> [  232.740167]  ? __lock_acquire+0x590/0xc40
> [  232.740335]  handle_pte_fault+0x439/0x760
> [  232.740498]  ? mtree_range_walk+0x252/0xae0
> [  232.740669]  ? __pfx_handle_pte_fault+0x10/0x10
> [  232.740899]  __handle_mm_fault+0xa02/0xf30
> [  232.741066]  ? __pfx___handle_mm_fault+0x10/0x10
> [  232.741255]  ? find_vma+0xa1/0x120
> [  232.741403]  handle_mm_fault+0x2bf/0x8f0
> [  232.741564]  do_user_addr_fault+0x2d3/0xed0
> [  232.741736]  ? trace_page_fault_user+0x1bf/0x240
> [  232.741969]  exc_page_fault+0x87/0x120
> [  232.742124]  asm_exc_page_fault+0x26/0x30
> [  232.742288] RIP: 0033:0x7fb4d73ed546
> [  232.742441] Code: 66 41 0f 6f fb 66 44 0f 6d dc 66 44 0f 6f c6 66 41 0f 6d f1 66 0f 6c fc 66 45 0f 6c c1 66 44 0f 6f c9 66 0f 6d ca 66 0f db f0 <66> 0f df 04 08 66 44 0f 6c ca 66 45 0f db c2 66 44 0f df 10 66 44
> [  232.743193] RSP: 002b:00007fb4907f68a0 EFLAGS: 00010206
> [  232.743565] RAX: 00007fb47840aa00 RBX: 00007fb4d73ec070 RCX: 0000000000001400
> [  232.743871] RDX: 0000000000002800 RSI: 0000000000003c00 RDI: 0000000000000001
> [  232.744150] RBP: 0000000000000004 R08: 0000000000001400 R09: 00007fb4d73ec060
> [  232.744433] R10: 000055f0261a4288 R11: 00007fb4c013da40 R12: 0000000000000008
> [  232.744712] R13: 0000000000000000 R14: 4332322132212110 R15: 0000000000000004
> [  232.746616]  </TASK>
> [  232.746711] Modules linked in: nft_nat nft_masq veth bridge stp llc snd_seq_dummy snd_hrtimer snd_seq snd_seq_device snd_timer snd soundcore overlay rfkill nf_conntrack_netbios_ns nf_conntrack_broadcast nft_fib_inet nft_fib_ipv4 nft_fib_ipv6 nft_fib nft_reject_inet nf_reject_ipv4 nf_reject_ipv6 nft_reject nft_ct nft_chain_nat nf_nat nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 nf_tables qrtr ppdev 9pnet_virtio 9pnet parport_pc i2c_piix4 netfs pcspkr parport i2c_smbus joydev sunrpc vfat fat loop dm_multipath nfnetlink vsock_loopback vmw_vsock_virtio_transport_common vmw_vsock_vmci_transport zram lz4hc_compress vmw_vmci lz4_compress vsock e1000 bochs serio_raw ata_generic pata_acpi scsi_dh_rdac scsi_dh_emc scsi_dh_alua i2c_dev fuse qemu_fw_cfg
> [  232.749308] ---[ end trace 0000000000000000 ]---
> [  232.749507] RIP: 0010:walk_to_pmd+0x29e/0x3c0
> [  232.749692] Code: d8 5b 5d 41 5c 41 5d 41 5e 41 5f c3 cc cc cc cc 48 89 ea 48 89 de 4c 89 f7 e8 ae 85 ff ff 85 c0 0f 84 1f fe ff ff 31 db eb d0 <0f> 0b 48 89 ea 48 89 de 4c 89 f7 e8 92 8b ff ff 85 c0 75 e8 48 b8
> [  232.750428] RSP: 0000:ffff8881aa6ff9a8 EFLAGS: 00010282
> [  232.750645] RAX: 8000000142e002e7 RBX: ffff8881433cae10 RCX: dffffc0000000000
> [  232.750954] RDX: 0000000000000000 RSI: 00007fb47840b000 RDI: 8000000142e002e7
> [  232.751232] RBP: 00007fb47840b000 R08: 0000000000000000 R09: 1ffff110354dff46
> [  232.751514] R10: fffffbfff0cb921d R11: 00000000910da5ce R12: 1ffffffff0c1fcdd
> [  232.751837] R13: 1ffffffff0c23f36 R14: ffff888171628040 R15: 0000000000000000
> [  232.752124] FS:  00007fb4907f86c0(0000) GS:ffff888791f2c000(0000) knlGS:0000000000000000
> [  232.752441] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [  232.752674] CR2: 00007fb47840be00 CR3: 000000015e6dc000 CR4: 00000000000006f0
> [  232.752983] Kernel panic - not syncing: Fatal exception
> [  232.753510] Kernel Offset: disabled
> [  232.754643] ---[ end Kernel panic - not syncing: Fatal exception ]---
> 
> This happens when two concurrent page faults occur within the same PMD range.
> One fault installs a PMD mapping through vmf_insert_pfn_pmd(), while the other
> attempts to install a PTE mapping via vmf_insert_pfn(). The bug is
> triggered because a pmd_trans_huge is not expected when walking the page
> table inside vmf_insert_pfn.
> 
> Avoid this race by adding a huge_fault callback to drm_gem_shmem_vm_ops so that
> PMD-sized mappings are handled through the appropriate huge page fault path.

Yikes, looks like we got it wrong between v4 and v6 :-/: v4 had the
.huge_fault implementation, v5 tried moved to .map_pages, and v6 went
back to .fault, but augmented to support huge pages, which I guess is
wrong for the reasons you mentioned.

> 
> Fixes: 211b9a39f261 ("drm/shmem-helper: Map huge pages in fault handler")
> Signed-off-by: Pedro Demarchi Gomes <pedrodemargomes@gmail.com>
> ---
>  drivers/gpu/drm/drm_gem_shmem_helper.c | 59 +++++++++++++++++++-------
>  1 file changed, 43 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/gpu/drm/drm_gem_shmem_helper.c b/drivers/gpu/drm/drm_gem_shmem_helper.c
> index 7b5a49935ae4..4cd9080c598b 100644
> --- a/drivers/gpu/drm/drm_gem_shmem_helper.c
> +++ b/drivers/gpu/drm/drm_gem_shmem_helper.c
> @@ -550,24 +550,53 @@ int drm_gem_shmem_dumb_create(struct drm_file *file, struct drm_device *dev,
>  }
>  EXPORT_SYMBOL_GPL(drm_gem_shmem_dumb_create);
>  
> -static bool drm_gem_shmem_try_map_pmd(struct vm_fault *vmf, unsigned long addr,
> -				      struct page *page)
> +
> +static vm_fault_t drm_gem_shmem_huge_fault(struct vm_fault *vmf,
> +					   unsigned int order)
>  {
> -#ifdef CONFIG_ARCH_SUPPORTS_PMD_PFNMAP
> -	unsigned long pfn = page_to_pfn(page);
> -	unsigned long paddr = pfn << PAGE_SHIFT;
> -	bool aligned = (addr & ~PMD_MASK) == (paddr & ~PMD_MASK);
> +	struct vm_area_struct *vma = vmf->vma;
> +	struct drm_gem_object *obj = vma->vm_private_data;
> +	struct drm_gem_shmem_object *shmem = to_drm_gem_shmem_obj(obj);
> +	loff_t num_pages = obj->size >> PAGE_SHIFT;
> +	vm_fault_t ret = VM_FAULT_FALLBACK;
> +	struct page **pages = shmem->pages;
> +	pgoff_t page_offset;
> +	unsigned long pfn;
> +	unsigned long paddr;
> +	bool aligned;
> +	struct page *page;
> +
> +	if (order != PMD_ORDER)
> +		goto out_unlocked;
> +
> +	/* Offset to faulty address in the VMA. */
> +	page_offset = vmf->pgoff - vma->vm_pgoff;
> +
> +	dma_resv_lock(shmem->base.resv, NULL);
> +
> +	if (page_offset >= num_pages ||
> +	    drm_WARN_ON_ONCE(obj->dev, !shmem->pages) ||
> +	    shmem->madv < 0) {
> +		ret = VM_FAULT_SIGBUS;
> +		goto out;
> +	}
> +
> +	page = pages[page_offset];
> +	pfn = page_to_pfn(page);
> +	paddr = pfn << PAGE_SHIFT;
> +	aligned = (vmf->address & ~PMD_MASK) == (paddr & ~PMD_MASK);
>  
>  	if (aligned &&
> -	    pmd_none(*vmf->pmd) &&
>  	    folio_test_pmd_mappable(page_folio(page))) {
>  		pfn &= PMD_MASK >> PAGE_SHIFT;
> -		if (vmf_insert_pfn_pmd(vmf, pfn, false) == VM_FAULT_NOPAGE)
> -			return true;
> +		ret = vmf_insert_pfn_pmd(vmf, pfn, false);
>  	}
> -#endif
>  
> -	return false;
> + out:
> +	dma_resv_unlock(shmem->base.resv);
> +
> + out_unlocked:
> +	return ret;
>  }

Do we really need to duplicate everything? Can't we rename
drm_gem_shmem_fault() into drm_gem_shmem_any_fault() with a
`bool try_pmd` passed as an argument, and then have
drm_gem_shmem_[huge_]fault() as wrappers?

static vm_fault_t drm_gem_shmem_huge_fault(struct vm_fault *vmf,
					   unsigned int order)
{
	if (order != PMD_ORDER)
		return VM_FAULT_FALLBACK;

	return drm_gem_shmem_any_fault(vmf, true);
}

static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf)
{
	return drm_gem_shmem_any_fault(vmf, false);
}

>  
>  static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf)
> @@ -593,11 +622,6 @@ static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf)
>  		goto out;
>  	}
>  
> -	if (drm_gem_shmem_try_map_pmd(vmf, vmf->address, pages[page_offset])) {
> -		ret = VM_FAULT_NOPAGE;
> -		goto out;
> -	}
> -
>  	pfn = page_to_pfn(pages[page_offset]);
>  	ret = vmf_insert_pfn(vma, vmf->address, pfn);
>  
> @@ -643,6 +667,9 @@ static void drm_gem_shmem_vm_close(struct vm_area_struct *vma)
>  
>  const struct vm_operations_struct drm_gem_shmem_vm_ops = {
>  	.fault = drm_gem_shmem_fault,
> +	#ifdef CONFIG_ARCH_SUPPORTS_HUGE_PFNMAP
> +	.huge_fault = drm_gem_shmem_huge_fault,
> +	#endif

Let's keep the #ifdef unindented:

#ifdef CONFIG_ARCH_SUPPORTS_HUGE_PFNMAP
	.huge_fault = drm_gem_shmem_huge_fault,
#endif

>  	.open = drm_gem_shmem_vm_open,
>  	.close = drm_gem_shmem_vm_close,
>  };