On Tuesday, 9 March 2021 6:44:41 AM AEDT Ralph Campbell wrote:
On 3/3/21 10:16 PM, Alistair Popple wrote:
Some devices require exclusive write access to shared virtual memory (SVM) ranges to perform atomic operations on that memory. This requires CPU page tables to be updated to deny access whilst atomic operations are occurring.
In order to do this introduce a new swap entry type (SWP_DEVICE_EXCLUSIVE). When a SVM range needs to be marked for exclusive access by a device all page table mappings for the particular range are replaced with device exclusive swap entries. This causes any CPU access to the page to result in a fault.
Faults are resovled by replacing the faulting entry with the original mapping. This results in MMU notifiers being called which a driver uses to update access permissions such as revoking atomic access. After notifiers have been called the device will no longer have exclusive access to the region.
Signed-off-by: Alistair Popple apopple@nvidia.com
I see in the next two patches how make_device_exclusive_entry() and check_device_exclusive_range() are used. This points out a similar problem that migrate_vma_setup() had before I added the mmu_notifier_range_init_migrate() helper to pass a cookie from migrate_vma_setup() to the invalidation callback so the device driver could ignore an invalidation callback triggered by the caller and thus resulting in a deadlock or having to invalidate device PTEs that wouldn't be migrating.
I think you can eliminate the need for check_device_exclusive_range() in the same way by adding a "void *" pointer to make_device_exclusive_entry() and passing that through to try_to_protect(), setting rmap_walk_control
rwc.arg
and then passing arg to mmu_notifier_range_init_migrate().
Thanks for the idea, I had missed there was already a way of passing a "void *" as part of mmu_notifier_range_init_migrate(). Agree that should allow a single pass without needing check_device_exclusive_range().
As Jason points out still need to check the GUP page is mapped at the expected address but that can be done as part of installing the exclusive swap entry in try_to_protect_one().
Although, maybe it would be better to define a new mmu_notifier_range_init_exclusive() and event type MMU_NOTIFY_EXCLUSIVE so that a device driver can revoke atomic/exclusive access but keep read/write access to other parts of the page.
Agree, I don't think overloading mmu_notifier_range_init_migrate() with the exclusive usage is correct. Better to define a new helper.
I thought about how make_device_exclusive_entry() is similar to
hmm_range_fault()
and whether it would be possible to add a new HMM_PFN_REQ_EXCLUSIVE flag but
I
see that make_device_exclusive_entry() returns the pages locked and with an additional get_page() reference. This doesn't fit well with the other hmm_range_fault() entries being returned as a "snapshot" so having a
different
API makes sense. I think it would be useful to add a HMM_PFN_EXCLUSIVE flag
so
that snapshots of the page tables can at least report that a page is
exclusively
being accessed by *some* device. Unfortunately, there is no pgmap pointer to
be
able to tell which device has exclusive access (since any struct page could
be
exclusively accessed, not just device private ones).
I have also experimented with integrating this with HMM but it just didn't end up being a good fit for the reasons you mention.
I also don't think adding HMM_PFN_EXCLUSIVE to read page table snapshots is that useful because there is no way to tell *which* device has exclusive access. So unless I've missed some particular usage for it now I think it can probably be added as a future improvement to HMM if/when it is needed.
- Alistair