On Tue, Jun 27, 2023 at 8:28 AM Peter Xu <peterx@xxxxxxxxxx> wrote:
>
> On Mon, Jun 26, 2023 at 09:23:16PM -0700, Suren Baghdasaryan wrote:
> > handle_mm_fault returning VM_FAULT_RETRY or VM_FAULT_COMPLETED means
> > mmap_lock has been released. However with per-VMA locks behavior is
> > different and the caller should still release it. To make the
> > rules consistent for the caller, drop the per-VMA lock before returning
> > from handle_mm_fault when page fault should be retried or is completed.
> >
> > Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
> > ---
> > arch/arm64/mm/fault.c | 3 ++-
> > arch/powerpc/mm/fault.c | 3 ++-
> > arch/s390/mm/fault.c | 3 ++-
> > arch/x86/mm/fault.c | 3 ++-
> > mm/memory.c | 12 +++++++++++-
> > 5 files changed, 19 insertions(+), 5 deletions(-)
> >
> > diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
> > index 6045a5117ac1..89f84e9ea1ff 100644
> > --- a/arch/arm64/mm/fault.c
> > +++ b/arch/arm64/mm/fault.c
> > @@ -601,7 +601,8 @@ static int __kprobes do_page_fault(unsigned long far, unsigned long esr,
> > goto lock_mmap;
> > }
> > fault = handle_mm_fault(vma, addr, mm_flags | FAULT_FLAG_VMA_LOCK, regs);
> > - vma_end_read(vma);
> > + if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
> > + vma_end_read(vma);
> >
> > if (!(fault & VM_FAULT_RETRY)) {
> > count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
> > diff --git a/arch/powerpc/mm/fault.c b/arch/powerpc/mm/fault.c
> > index 531177a4ee08..4697c5dca31c 100644
> > --- a/arch/powerpc/mm/fault.c
> > +++ b/arch/powerpc/mm/fault.c
> > @@ -494,7 +494,8 @@ static int ___do_page_fault(struct pt_regs *regs, unsigned long address,
> > }
> >
> > fault = handle_mm_fault(vma, address, flags | FAULT_FLAG_VMA_LOCK, regs);
> > - vma_end_read(vma);
> > + if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
> > + vma_end_read(vma);
> >
> > if (!(fault & VM_FAULT_RETRY)) {
> > count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
> > diff --git a/arch/s390/mm/fault.c b/arch/s390/mm/fault.c
> > index b65144c392b0..cccefe41038b 100644
> > --- a/arch/s390/mm/fault.c
> > +++ b/arch/s390/mm/fault.c
> > @@ -418,7 +418,8 @@ static inline vm_fault_t do_exception(struct pt_regs *regs, int access)
> > goto lock_mmap;
> > }
> > fault = handle_mm_fault(vma, address, flags | FAULT_FLAG_VMA_LOCK, regs);
> > - vma_end_read(vma);
> > + if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
> > + vma_end_read(vma);
> > if (!(fault & VM_FAULT_RETRY)) {
> > count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
> > goto out;
> > diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c
> > index e4399983c50c..d69c85c1c04e 100644
> > --- a/arch/x86/mm/fault.c
> > +++ b/arch/x86/mm/fault.c
> > @@ -1347,7 +1347,8 @@ void do_user_addr_fault(struct pt_regs *regs,
> > goto lock_mmap;
> > }
> > fault = handle_mm_fault(vma, address, flags | FAULT_FLAG_VMA_LOCK, regs);
> > - vma_end_read(vma);
> > + if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
> > + vma_end_read(vma);
> >
> > if (!(fault & VM_FAULT_RETRY)) {
> > count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
> > diff --git a/mm/memory.c b/mm/memory.c
> > index f69fbc251198..9011ad63c41b 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -5086,7 +5086,17 @@ static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
> > }
> > }
> >
> > - return handle_pte_fault(&vmf);
> > + ret = handle_pte_fault(&vmf);
> > + if (ret & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)) {
> > + /*
> > + * In case of VM_FAULT_RETRY or VM_FAULT_COMPLETED we might
> > + * be still holding per-VMA lock to keep the vma stable as long
> > + * as possible. Drop it before returning.
> > + */
> > + if (vmf.flags & FAULT_FLAG_VMA_LOCK)
> > + vma_end_read(vma);
> > + }
>
> This smells hackish.. I'd think better we just release the lock at the
> place where we'll return RETRY, and AFAIU swap is the only place vma lock
> returns a RETRY with current code base?
>
> do_swap_page():
> if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
> + vma_end_read(vma);
> ret = VM_FAULT_RETRY;
> goto out;
> }
>
> I.e., I don't think VM_FAULT_COMPLETED can even be returned with vma lock
> paths yet as it doesn't yet support VM_SHARED.
Ack.
>
> --
> Peter Xu
>
