Re: [PATCH v2 01/40] mm/vma: fix mmap_prepare file handling, remove file_doesnt_need_get

[Mike Rapoport]

On Mon, Sep 14, 2026 at 03:57:21PM +0100, Lorenzo Stoakes (ARM) wrote:
> The map->file_doesnt_need_get flag is confusing and the existing
> implementation has holes.
> 
> Drivers are permitted to change the owning file of a mapping. If they do
> so, they are required to take a reference on that file.
> 
> The mmap() operation which ultimately invokes __mmap_region() is guaranteed
> to drop the refcount for the original file the mapping was made under, but
> this is not true for the replaced file.
> 
> This has been addressed so far by tracking map->file_doesnt_need_get, which
> is rather poorly named and unfortunately fails to correctly track whether
> or not an additional put were needed in a number of cases.
> 
> Make life easier by removing this flag, and instead drop the reference for
> both mmap_prepare and the deprecated mmap callback in a new function
> put_map().
> 
> Track whether this needs to be done by aligning mmap_state with
> vm_area_desc and store the original file in the map->file field, keeping
> the updated file in map->vm_file.
> 
> In order to have the same behaviour for both types of hooks, only drop the
> reference __mmap_new_file_vma() itself took in its error path, deferring
> the replaced file's reference to put_map().
> 
> To make this work correctly, map->vm_file has to be updated before any
> error handling, so update __mmap_new_file_vma() and call_mmap_prepare() to
> set this field first.
> 
> Also when mmap_prepare() changes the file and is then merged, the reference
> count also must be decremented, so update the logic to call put_map() in
> this case too.
> 
> Also update __compat_vma_mmap() to manually perform this step for stacked
> file systems using the compatibility layer, and update
> compat_set_vma_from_desc() to replace vma_set_file() with a correct
> refcount/file update.
> 
> No in-tree driver is impacted by the incorrect implementation of this
> currently (no driver that does this is mergeable for one), so this does not
> need to be a fix.

But the patch iteslf needs to be fixed :)

It seems to be the change that broke the CI.

The rest is from LLM, take it with a grain of salt :)

mm-ci mm-unstable red build - bisect analysis
==============================================

Bad commit: 2a937a04babf1 "mm/vma: fix mmap_prepare file handling, remove
file_doesnt_need_get" - first patch of Lorenzo Stoakes' 40-patch series
"mm: make VMA flag semantics explicit, eliminate VM_SPECIAL" (v2).
https://lore.kernel.org/all/20260914-b4-mmap-prepare-vma-flag-sanify-v2-0-7d9781ed5361@xxxxxxxxxx

Symptom: tools/testing/selftests/mm/pfnmap.c triggers
"BUG: Bad page map in process pfnmap" during __zap_vma_range()/vm_normal_page(),
seen on process exit/munmap. CI's run-mm-selftests.sh greps guest dmesg for
BUG|WARNING and fails the job regardless of the test's own exit code.

Root cause:
In call_mmap_prepare() (mm/vma.c), the commit reordered:

    map->vma_flags = desc->vma_flags;

to run BEFORE call_action_prepare(). For PFN-remap mmap_prepare drivers
(e.g. /dev/mem), VMA_PFNMAP_BIT/VMA_IO_BIT are only added to
desc->vma_flags inside remap_pfn_range_prepare(), which runs as part of
call_action_prepare(). Snapshotting vma_flags too early means VM_PFNMAP/
VM_IO get silently dropped from the resulting VMA, while the PTEs are
still installed as pte_special (real PFN remap).

vm_normal_page() then sees a pte_special PTE in a VMA lacking
VM_PFNMAP/VM_MIXEDMAP and calls print_bad_page_map(), producing the
splat. Confirmed: crash log shows vm_flags:080000d1, missing VM_PFNMAP
(0x400).

This is a plain ordering bug in this one commit - not a driver-conversion
or vma_can_gup()/predicate-logic issue.

Bisect method: built/booted via virtme-ng, ran pfnmap selftest directly,
restricted to range 2a937a04babf1~1..0a80aa0fb08f1 (the 40-commit series;
tip = mm-unstable-2026-09-17). Pre-series commit: clean. First commit of
series: already reproduces the bug. All scratch worktrees/branches and
bisect state removed after; mm-unstable-2026-09-17/mm-new-2026-09-17 and
mmotm remote untouched.

Diagnosis only - no fix proposed/applied.
 
> Signed-off-by: Lorenzo Stoakes (ARM) <ljs@xxxxxxxxxx>
> ---
>  mm/internal.h |  1 +
>  mm/util.c     |  5 +++-
>  mm/vma.c      | 95 ++++++++++++++++++++++++++++++++++-------------------------
>  mm/vma.h      |  6 ++--
>  4 files changed, 64 insertions(+), 43 deletions(-)
> 
> diff --git a/mm/internal.h b/mm/internal.h
> index 0dca33db068f..fe576d468af4 100644
> --- a/mm/internal.h
> +++ b/mm/internal.h
> @@ -7,6 +7,7 @@
>  #ifndef __MM_INTERNAL_H
>  #define __MM_INTERNAL_H
>  
> +#include <linux/file.h>
>  #include <linux/fs.h>
>  #include <linux/khugepaged.h>
>  #include <linux/mm.h>
> diff --git a/mm/util.c b/mm/util.c
> index bf0513d1d3d0..016932780925 100644
> --- a/mm/util.c
> +++ b/mm/util.c
> @@ -1228,8 +1228,11 @@ int __compat_vma_mmap(struct vm_area_desc *desc,
>  
>  	/* Perform any preparatory tasks for mmap action. */
>  	err = mmap_action_prepare(desc);
> -	if (err)
> +	if (err) {
> +		if (desc->vm_file != vma->vm_file)
> +			fput(desc->vm_file);
>  		return err;
> +	}
>  	/* Update the VMA from the descriptor. */
>  	compat_set_vma_from_desc(vma, desc);
>  	/* Complete any specified mmap actions. */
> diff --git a/mm/vma.c b/mm/vma.c
> index 55917d097933..a319a9fc2f29 100644
> --- a/mm/vma.c
> +++ b/mm/vma.c
> @@ -24,7 +24,8 @@ struct mmap_state {
>  		vm_flags_t vm_flags;
>  		vma_flags_t vma_flags;
>  	};
> -	struct file *file;
> +	struct file *file;	/* mmap()-specified file. */
> +	struct file *vm_file;	/* May be updated by mmap_prepare. */
>  	pgprot_t page_prot;
>  
>  	/* User-defined fields, perhaps updated by .mmap_prepare(). */
> @@ -43,8 +44,6 @@ struct mmap_state {
>  
>  	/* Determine if we can check KSM flags early in mmap() logic. */
>  	bool check_ksm_early :1;
> -	/* If .mmap_prepare changed the file, we don't need to pin. */
> -	bool file_doesnt_need_get :1;
>  };
>  
>  #define MMAP_STATE(name, mm_, vmi_, addr_, len_, pgoff_, anon_pgoff_, vma_flags_, file_) \
> @@ -58,6 +57,7 @@ struct mmap_state {
>  		.pglen = PHYS_PFN(len_),				\
>  		.vma_flags = vma_flags_,				\
>  		.file = file_,						\
> +		.vm_file = file_,					\
>  		.page_prot = vma_flags_to_page_prot(vma_flags_),	\
>  	}
>  
> @@ -70,7 +70,7 @@ struct mmap_state {
>  		.vma_flags = (map_)->vma_flags,				\
>  		.pgoff = (map_)->pgoff,					\
>  		.anon_pgoff = (map_)->anon_pgoff,			\
> -		.file = (map_)->file,					\
> +		.file = (map_)->vm_file,				\
>  		.prev = (map_)->prev,					\
>  		.middle = vma_,						\
>  		.next = (vma_) ? NULL : (map_)->next,			\
> @@ -2447,7 +2447,7 @@ void mm_drop_all_locks(struct mm_struct *mm)
>   */
>  static bool accountable_mapping(struct mmap_state *map)
>  {
> -	const struct file *file = map->file;
> +	const struct file *file = map->vm_file;
>  
>  	/*
>  	 * hugetlb has its own accounting separate from the core VM
> @@ -2496,7 +2496,7 @@ static void vms_abort_munmap_vmas(struct vma_munmap_struct *vms,
>  
>  static void update_ksm_flags(struct mmap_state *map)
>  {
> -	map->vma_flags = ksm_vma_flags(map->mm, map->file, map->vma_flags);
> +	map->vma_flags = ksm_vma_flags(map->mm, map->vm_file, map->vma_flags);
>  }
>  
>  static void set_desc_from_map(struct vm_area_desc *desc,
> @@ -2506,7 +2506,7 @@ static void set_desc_from_map(struct vm_area_desc *desc,
>  	desc->end = map->end;
>  
>  	desc->pgoff = map->pgoff;
> -	desc->vm_file = map->file;
> +	desc->vm_file = map->vm_file;
>  	desc->vma_flags = map->vma_flags;
>  	desc->page_prot = map->page_prot;
>  }
> @@ -2586,6 +2586,10 @@ static int __mmap_setup(struct mmap_state *map, struct vm_area_desc *desc,
>  	return 0;
>  }
>  
> +static bool map_same_file(struct mmap_state *map)
> +{
> +	return map->vm_file == map->file;
> +}
>  
>  static int __mmap_new_file_vma(struct mmap_state *map,
>  			       struct vm_area_struct *vma)
> @@ -2593,20 +2597,23 @@ static int __mmap_new_file_vma(struct mmap_state *map,
>  	struct vma_iterator *vmi = map->vmi;
>  	int error;
>  
> -	vma->vm_file = map->file;
> -	if (!map->file_doesnt_need_get)
> -		get_file(map->file);
> +	vma->vm_file = map->vm_file;
> +	if (map_same_file(map))
> +		get_file(map->vm_file);
>  
> -	if (!map->file->f_op->mmap)
> +	if (!map->vm_file->f_op->mmap)
>  		return 0;
>  
>  	error = mmap_file(vma->vm_file, vma);
> +	map->vm_file = vma->vm_file;
> +
>  	if (error) {
>  		UNMAP_STATE(unmap, vmi, vma, vma->vm_start, vma->vm_end,
>  			    map->prev, map->next);
> -		fput(vma->vm_file);
> -		vma->vm_file = NULL;
> +		if (map_same_file(map))
> +			fput(map->vm_file);
>  
> +		vma->vm_file = NULL;
>  		vma_iter_set(vmi, vma->vm_end);
>  		/* Undo any partial mapping done by a device driver. */
>  		unmap_region(&unmap);
> @@ -2623,7 +2630,6 @@ static int __mmap_new_file_vma(struct mmap_state *map,
>  			!vma_flags_test(&map->vma_flags, VMA_MAYWRITE_BIT) &&
>  			vma_test(vma, VMA_MAYWRITE_BIT));
>  
> -	map->file = vma->vm_file;
>  	map->vma_flags = vma->flags;
>  
>  	return 0;
> @@ -2631,7 +2637,7 @@ static int __mmap_new_file_vma(struct mmap_state *map,
>  
>  static void map_set_anon(struct mmap_state *map)
>  {
> -	map->file = NULL;
> +	map->vm_file = NULL;
>  	map->vm_ops = NULL;
>  	map->pgoff = map->addr >> PAGE_SHIFT;
>  }
> @@ -2643,7 +2649,7 @@ static bool map_is_private(const struct mmap_state *map)
>  
>  static bool map_is_anon(const struct mmap_state *map)
>  {
> -	return map_is_private(map) && !map->file;
> +	return map_is_private(map) && !map->vm_file;
>  }
>  
>  /*
> @@ -2688,7 +2694,7 @@ static int __mmap_new_vma(struct mmap_state *map, struct vm_area_struct **vmap,
>  	}
>  
>  	/* Invoke callbacks. */
> -	if (map->file)
> +	if (map->vm_file)
>  		error = __mmap_new_file_vma(map, vma);
>  	else if (!is_anon)
>  		error = shmem_zero_setup(vma);
> @@ -2797,36 +2803,34 @@ static int call_mmap_prepare(struct mmap_state *map,
>  	int err;
>  
>  	/* Invoke the hook. */
> -	err = vfs_mmap_prepare(map->file, desc);
> -	if (err)
> -		return err;
> -
> -	/* It's invalid for mmap_preprare hooks to clear vm_ops. */
> -	if (!desc->vm_ops)
> -		return -EINVAL;
> -
> -	err = call_action_prepare(map, desc);
> +	err = vfs_mmap_prepare(map->vm_file, desc);
>  	if (err)
>  		return err;
>  
>  	/* Update fields permitted to be changed. */
>  	map->pgoff = desc->pgoff;
> -	if (desc->vm_file != map->file) {
> -		map->file_doesnt_need_get = true;
> -		map->file = desc->vm_file;
> -	}
> +	if (desc->vm_file != map->vm_file)
> +		map->vm_file = desc->vm_file;
>  	map->vma_flags = desc->vma_flags;
>  	map->page_prot = desc->page_prot;
>  	/* User-defined fields. */
>  	map->vm_ops = desc->vm_ops;
>  	map->vm_private_data = desc->private_data;
>  
> +	/* It's invalid for mmap_prepare hooks to clear vm_ops. */
> +	if (!desc->vm_ops)
> +		return -EINVAL;
> +
> +	err = call_action_prepare(map, desc);
> +	if (err)
> +		return err;
> +
>  	/*
>  	 * MAP_PRIVATE-/dev/zero mappings are an ancient way of getting
>  	 * anonymous mappings. Rather than allowing these mappings to be odd
>  	 * outliers, simply make them truly anonymous.
>  	 */
> -	if (map_is_private(map) && file_is_dev_zero(map->file))
> +	if (map_is_private(map) && file_is_dev_zero(map->vm_file))
>  		map_set_anon(map);
>  
>  	return 0;
> @@ -2845,7 +2849,7 @@ static void set_vma_user_defined_fields(struct vm_area_struct *vma,
>   */
>  static bool can_set_ksm_flags_early(struct mmap_state *map)
>  {
> -	struct file *file = map->file;
> +	struct file *file = map->vm_file;
>  
>  	/* Anonymous mappings have no driver which can change them. */
>  	if (!file)
> @@ -2868,6 +2872,20 @@ static bool can_set_ksm_flags_early(struct mmap_state *map)
>  	return false;
>  }
>  
> +static void put_map(struct mmap_state *map)
> +{
> +	/*
> +	 * An error occurred or the VMA was merged.
> +	 *
> +	 * If the file was changed by the driver (which is required to increment
> +	 * the replacement file's reference count), drop its reference count.
> +	 *
> +	 * On error, the caller always drops the original file regardless.
> +	 */
> +	if (map->vm_file && !map_same_file(map))
> +		fput(map->vm_file);
> +}
> +
>  static unsigned long __mmap_region(struct file *file, unsigned long addr,
>  		unsigned long len, vma_flags_t vma_flags,
>  		unsigned long pgoff, struct list_head *uf)
> @@ -2922,7 +2940,10 @@ static unsigned long __mmap_region(struct file *file, unsigned long addr,
>  
>  	__mmap_complete(&map, vma);
>  
> -	if (have_mmap_prepare && allocated_new) {
> +	if (!allocated_new) {
> +		/* Merged, so need to drop refcount. */
> +		put_map(&map);
> +	} else if (have_mmap_prepare) {
>  		error = mmap_action_complete(vma, &desc.action,
>  					     /*is_compat=*/false);
>  		if (error)
> @@ -2936,13 +2957,7 @@ static unsigned long __mmap_region(struct file *file, unsigned long addr,
>  	if (map.charged)
>  		vm_unacct_memory(map.charged);
>  abort_munmap:
> -	/*
> -	 * This indicates that .mmap_prepare has set a new file, differing from
> -	 * desc->vm_file. But since we're aborting the operation, only the
> -	 * original file will be cleaned up. Ensure we clean up both.
> -	 */
> -	if (map.file_doesnt_need_get)
> -		fput(map.file);
> +	put_map(&map);
>  	vms_abort_munmap_vmas(&map.vms, &map.mas_detach);
>  	return error;
>  }
> diff --git a/mm/vma.h b/mm/vma.h
> index e97bd2dfa786..f15faa83f3d6 100644
> --- a/mm/vma.h
> +++ b/mm/vma.h
> @@ -394,8 +394,10 @@ static inline void compat_set_vma_from_desc(struct vm_area_struct *vma,
>  
>  	/* Mutable fields. Populated with initial state. */
>  	vma_set_pgoff(vma, desc->pgoff);
> -	if (desc->vm_file != vma->vm_file)
> -		vma_set_file(vma, desc->vm_file);
> +	if (desc->vm_file != vma->vm_file) {
> +		fput(vma->vm_file);
> +		vma->vm_file = desc->vm_file;
> +	}
>  	vma->flags = desc->vma_flags;
>  	vma->vm_page_prot = desc->page_prot;
>  
> 
> -- 
> 2.55.0
> 

-- 
Sincerely yours,
Mike.