mm: introduce new .mmap_prepare() file callback

mm: introduce new .mmap_prepare() file callback

Patch series "eliminate mmap() retry merge, add .mmap_prepare hook", v2.

During the mmap() of a file-backed mapping, we invoke the underlying
driver file's mmap() callback in order to perform driver/file system
initialisation of the underlying VMA.

This has been a source of issues in the past, including a significant
security concern relating to unwinding of error state discovered by Jann
Horn, as fixed in commit 5de195060b2e ("mm: resolve faulty mmap_region()
error path behaviour") which performed the recent, significant, rework of
mmap() as a whole.

However, we have had a fly in the ointment remain - drivers have a great
deal of freedom in the .mmap() hook to manipulate VMA state (as well as
page table state).

This can be problematic, as we can no longer reason sensibly about VMA
state once the call is complete (the ability to do - anything - here does
rather interfere with that).

In addition, callers may choose to do odd or unusual things which might
interfere with subsequent steps in the mmap() process, and it may do so
and then raise an error, requiring very careful unwinding of state about
which we can make no assumptions.

Rather than providing such an open-ended interface, this series provides
an alternative, far more restrictive one - we expose a whitelist of fields
which can be adjusted by the driver, along with immutable state upon which
the driver can make such decisions:

struct vm_area_desc {
	/* Immutable state. */
	struct mm_struct *mm;
	unsigned long start;
	unsigned long end;

	/* Mutable fields. Populated with initial state. */
	pgoff_t pgoff;
	struct file *file;
	vm_flags_t vm_flags;
	pgprot_t page_prot;

	/* Write-only fields. */
	const struct vm_operations_struct *vm_ops;
	void *private_data;
};

The mmap logic then updates the state used to either merge with a VMA or
establish a new VMA based upon this logic.

This is achieved via new file hook .mmap_prepare(), which is, importantly,
invoked very early on in the mmap() process.

If an error arises, we can very simply abort the operation with very
little unwinding of state required.

The existing logic contains another, related, peccadillo - since the
.mmap() callback might do anything, it may also cause a previously
unmergeable VMA to become mergeable with adjacent VMAs.

Right now the logic will retry a merge like this only if the driver
changes VMA flags, and changes them in such a way that a merge might
succeed (that is, the flags are not 'special', that is do not contain any
of the flags specified in VM_SPECIAL).

This has also been the source of a great deal of pain - it's hard to
reason about an .mmap() callback that might do - anything - but it's also
hard to reason about setting up a VMA and writing to the maple tree, only
to do it again utilising a great deal of shared state.

Since .mmap_prepare() sets fields before the first merge is even
attempted, the use of this callback obviates the need for this retry merge
logic.

A driver may only specify .mmap_prepare() or the deprecated .mmap()
callback.  In future we may add futher callbacks beyond .mmap_prepare() to
faciliate all use cass as we convert drivers.

In researching this change, I examined every .mmap() callback, and
discovered only a very few that set VMA state in such a way that a.  the
VMA flags changed and b.  this would be mergeable.

In the majority of cases, it turns out that drivers are mapping kernel
memory and thus ultimately set VM_PFNMAP, VM_MIXEDMAP, or other
unmergeable VM_SPECIAL flags.

Of those that remain I identified a number of cases which are only
applicable in DAX, setting the VM_HUGEPAGE flag:

* dax_mmap()
* erofs_file_mmap()
* ext4_file_mmap()
* xfs_file_mmap()

For this remerge to not occur and to impact users, each of these cases
would require a user to mmap() files using DAX, in parts, immediately
adjacent to one another.

This is a very unlikely usecase and so it does not appear to be worthwhile
to adjust this functionality accordingly.

We can, however, very quickly do so if needed by simply adding an
.mmap_prepare() callback to these as required.

There are two further non-DAX cases I idenitfied:

* orangefs_file_mmap() - Clears VM_RAND_READ if set, replacing with
  VM_SEQ_READ.
* usb_stream_hwdep_mmap() - Sets VM_DONTDUMP.

Both of these cases again seem very unlikely to be mmap()'d immediately
adjacent to one another in a fashion that would result in a merge.

Finally, we are left with a viable case:

* secretmem_mmap() - Set VM_LOCKED, VM_DONTDUMP.

This is viable enough that the mm selftests trigger the logic as a matter
of course.  Therefore, this series replace the .secretmem_mmap() hook with
.secret_mmap_prepare().

This patch (of 3):

Provide a means by which drivers can specify which fields of those
permitted to be changed should be altered to prior to mmap()'ing a range
(which may either result from a merge or from mapping an entirely new
VMA).

Doing so is substantially safer than the existing .mmap() calback which
provides unrestricted access to the part-constructed VMA and permits
drivers and file systems to do 'creative' things which makes it hard to
reason about the state of the VMA after the function returns.

The existing .mmap() callback's freedom has caused a great deal of issues,
especially in error handling, as unwinding the mmap() state has proven to
be non-trivial and caused significant issues in the past, for instance
those addressed in commit 5de195060b2e ("mm: resolve faulty mmap_region()
error path behaviour").

It also necessitates a second attempt at merge once the .mmap() callback
has completed, which has caused issues in the past, is awkward, adds
overhead and is difficult to reason about.

The .mmap_prepare() callback eliminates this requirement, as we can update
fields prior to even attempting the first merge.  It is safer, as we
heavily restrict what can actually be modified, and being invoked very
early in the mmap() process, error handling can be performed safely with
very little unwinding of state required.

The .mmap_prepare() and deprecated .mmap() callbacks are mutually
exclusive, so we permit only one to be invoked at a time.

Update vma userland test stubs to account for changes.

Link: https://lkml.kernel.org/r/cover.1746792520.git.lorenzo.stoakes@oracle.com
Link: https://lkml.kernel.org/r/adb36a7c4affd7393b2fc4b54cc5cfe211e41f71.1746792520.git.lorenzo.stoakes@oracle.com
Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christian Brauner <brauner@kernel.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jann Horn <jannh@google.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>