Re: [PATCH v4 1/3] mm/mprotect: Fix soft-dirty check in can_change_pte_writable()

From: David Hildenbrand
Date: Mon Jan 02 2023 - 07:30:36 EST


On 28.12.22 15:14, Muhammad Usama Anjum wrote:
On 12/19/22 5:19 PM, Muhammad Usama Anjum wrote:
Addition of vma_soft_dirty_enabled() has tinkered with the soft-dirty PTE
bit status setting. The internal behavior has changed. The test case was
shared by David
(https://lore.kernel.org/all/bfcae708-db21-04b4-0bbe-712badd03071@xxxxxxxxxx/).
The explanation is as following:

_Before_ addition of this patch(76aefad628aae),
m = mmap(2 pages)
clear_softdirty()
mremap(m + pag_size)
mprotect(READ)
mprotect(READ | WRITE);
memset(m)
After memset(),
PAGE-1 PAGE-2
VM_SOFTDIRTY set set
PTE softdirty flag set set
/proc//pagemap view set set


_After_ addition of this patch(76aefad628aae)
m = mmap(2 pages)
clear_softdirty()
mremap(m + page_size)
mprotect(READ)
mprotect(READ | WRITE);
memset(m)
After memset(),
PAGE-1 PAGE-2
VM_SOFTDIRTY set set
PTE softdirty flag *not set* set
/proc//pagemap view set set

The user's point of view hasn't changed. But internally after this patch,
the soft-dirty tracking in PTEs gets turn off if VM_SOFTDIRTY is set. The
soft-dirty tracking in the PTEs shouldn't be just turned off when mprotect
is used. Why? Because soft-dirty tracking in the PTEs is always enabled
regardless of VM_SOFTDIRTY is set or not. Example:

m = mem(2 pages)
At this point:
PAGE-1 PAGE-2
VM_SOFTDIRTY set set
PTE softdirty flag not set not set
/proc//pagemap view set set
memset(m)
At this point:
PAGE-1 PAGE-2
VM_SOFTDIRTY set set
PTE softdirty flag set set
/proc//pagemap view set set

This example proves that soft-dirty flag on the PTE is set regardless of
the VM_SOFTDIRTY.

Hi Andrew and Cyrill,

Peter doesn't agree with me here that this change in behavior should be
reverted etc. Please comment.

For the records, I agree with Peter: As 76aefad628aa ("mm/mprotect: fix soft-dirty check in can_change_pte_writable()") documents, this patch fixed real problems.

/proc/pagemap works as expected right now such that we don't have an under-indication. Internal representation is an implementation detail.

Whatever we do, there must not be an under-indication of softdirty. That is the ABI guaranteed (especially for anonymous memory). "No over-indication" was never the ABI guarantee.

For your use case, you want to reduce over-indication. I suggested looked into alternatives.

--
Thanks,

David / dhildenb