mm/gup: tolerate NULL unlocked in fixup_user_fault()
Subsystems (2)
PATCH 1/1 mm/gup: tolerate NULL unlocked in fixup_user_fault()
fixup_user_fault() takes a "bool *unlocked" output parameter that callers may set to NULL when they do not want the retry/unlock machinery. The function honours that contract on the way in:
if (unlocked) fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
so callers passing NULL never set FAULT_FLAG_ALLOW_RETRY. In return, handle_mm_fault() is not expected to produce VM_FAULT_RETRY or VM_FAULT_COMPLETED for them, which is why the dereferences of unlocked further down used to be considered unreachable.
That invariant is implicit, not enforced. At least one caller in arch/s390/pci/pci_mmio.c does pass NULL:
fixup_user_fault(current->mm, mmio_addr, FAULT_FLAG_WRITE, NULL);
If a future change in handle_mm_fault() ever returned VM_FAULT_COMPLETED or VM_FAULT_RETRY without ALLOW_RETRY having been requested, the unconditional "*unlocked = true" stores would NULL-deref and crash the kernel for this path.
smatch flags both stores:
mm/gup.c:1597 fixup_user_fault() error: we previously assumed
'unlocked' could be null (see line 1573)
mm/gup.c:1612 fixup_user_fault() error: we previously assumed
'unlocked' could be null (see line 1573)
Make the NULL handling consistent on both sides of the function:
guard the two stores with "if (unlocked)" so fixup_user_fault() tolerates a NULL output pointer regardless of which fault outcome handle_mm_fault() returns.
No functional change for callers that already pass a non-NULL pointer.
--- mm/gup.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
mm/gup.c +4 -2
@@ -1594,7 +1594,8 @@ int fixup_user_fault(struct mm_struct *mm,* could tell the callers so they do not need to unlock. */ mmap_read_lock (mm); - *unlocked = true; +if (unlocked) + *unlocked = true;return 0 ; }
@@ -1608,7 +1609,8 @@ int fixup_user_fault(struct mm_struct *mm,if (ret &VM_FAULT_RETRY ) {mmap_read_lock (mm); - *unlocked = true; +if (unlocked) + *unlocked = true; fault_flags |=FAULT_FLAG_TRIED ; goto retry; }
Signature · 1 line
2.43.0
David Hildenbrand (Arm)
maintainerQuoted 15 lines · Stepan Ionichev
fixup_user_fault()takes a "bool *unlocked" output parameter that callers may set to NULL when they do not want the retry/unlock machinery. The function honours that contract on the way in:if (unlocked) fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;so callers passing NULL never set
FAULT_FLAG_ALLOW_RETRY. In return,handle_mm_fault()is not expected to produceVM_FAULT_RETRYorVM_FAULT_COMPLETEDfor them, which is why the dereferences of unlocked further down used to be considered unreachable.That invariant is implicit, not enforced. At least one caller in
arch/s390/pci/pci_mmio.cdoes pass NULL:
Quoted 6 lines
fixup_user_fault(current->mm, mmio_addr, FAULT_FLAG_WRITE, NULL);If a future change in
handle_mm_fault()ever returnedVM_FAULT_COMPLETEDorVM_FAULT_RETRYwithoutALLOW_RETRYhaving been requested, the unconditional "*unlocked = true" stores would NULL-deref and crash the kernel for this path.
That would be completely broken. We must not drop the mmap lock unless FAULT_FLAG_ALLOW_RETRY was set. Returning VM_FAULT_COMPLETED/VM_FAULT_RETRY would mean that we did that. Broken.
And the function documents "If NULL, the caller must guarantee that fault_flags does not contain FAULT_FLAG_ALLOW_RETRY."
Signature · 3 lines
Cheers, David
Quoted 6 lines · David Hildenbrand
That would be completely broken. We must not drop the mmap lock unless
FAULT_FLAG_ALLOW_RETRYwas set. Returning VM_FAULT_COMPLETED/VM_FAULT_RETRY would mean that we did that. Broken.And the function documents "If NULL, the caller must guarantee that
fault_flagsdoes not containFAULT_FLAG_ALLOW_RETRY."
You're right -- the contract is already enforced by the surrounding code and documented at the function. Defensive guards against an impossible scenario would just hide future real bugs.
Please drop this patch.
Thanks for taking the time to explain.
Stepan