On 5/3/23 8:53 AM, David Hildenbrand wrote: > On 03.05.23 13:25, Matthew Rosato wrote: >> On 5/3/23 3:08 AM, David Hildenbrand wrote: >>> On 03.05.23 02:31, Matthew Rosato wrote: >>>> On 5/2/23 6:51 PM, Lorenzo Stoakes wrote: >>>>> Writing to file-backed mappings which require folio dirty tracking using >>>>> GUP is a fundamentally broken operation, as kernel write access to GUP >>>>> mappings do not adhere to the semantics expected by a file system. >>>>> >>>>> A GUP caller uses the direct mapping to access the folio, which does not >>>>> cause write notify to trigger, nor does it enforce that the caller marks >>>>> the folio dirty. >>>>> >>>>> The problem arises when, after an initial write to the folio, writeback >>>>> results in the folio being cleaned and then the caller, via the GUP >>>>> interface, writes to the folio again. >>>>> >>>>> As a result of the use of this secondary, direct, mapping to the folio no >>>>> write notify will occur, and if the caller does mark the folio dirty, this >>>>> will be done so unexpectedly. >>>>> >>>>> For example, consider the following scenario:- >>>>> >>>>> 1. A folio is written to via GUP which write-faults the memory, notifying >>>>> the file system and dirtying the folio. >>>>> 2. Later, writeback is triggered, resulting in the folio being cleaned and >>>>> the PTE being marked read-only. >>>>> 3. The GUP caller writes to the folio, as it is mapped read/write via the >>>>> direct mapping. >>>>> 4. The GUP caller, now done with the page, unpins it and sets it dirty >>>>> (though it does not have to). >>>>> >>>>> This change updates both the PUP FOLL_LONGTERM slow and fast APIs. As >>>>> pin_user_pages_fast_only() does not exist, we can rely on a slightly >>>>> imperfect whitelisting in the PUP-fast case and fall back to the slow case >>>>> should this fail. >>>>> >>>>> v8: >>>>> - Fixed typo writeable -> writable. >>>>> - Fixed bug in writable_file_mapping_allowed() - must check combination of >>>>> FOLL_PIN AND FOLL_LONGTERM not either/or. >>>>> - Updated vma_needs_dirty_tracking() to include write/shared to account for >>>>> MAP_PRIVATE mappings. >>>>> - Move to open-coding the checks in folio_pin_allowed() so we can >>>>> READ_ONCE() the mapping and avoid unexpected compiler loads. Rename to >>>>> account for fact we now check flags here. >>>>> - Disallow mapping == NULL or mapping & PAGE_MAPPING_FLAGS other than >>>>> anon. Defer to slow path. >>>>> - Perform GUP-fast check _after_ the lowest page table level is confirmed to >>>>> be stable. >>>>> - Updated comments and commit message for final patch as per Jason's >>>>> suggestions. >>>> >>>> Tested again on s390 using QEMU with a memory backend file (on ext4) and vfio-pci -- This time both vfio_pin_pages_remote (which will call pin_user_pages_remote(flags | FOLL_LONGTERM)) and the pin_user_pages_fast(FOLL_WRITE | FOLL_LONGTERM) in kvm_s390_pci_aif_enable are being allowed (e.g. returning positive pin count) >>> >>> At least it's consistent now ;) And it might be working as expected ... >>> >>> In v7: >>> * pin_user_pages_fast() succeeded >>> * vfio_pin_pages_remote() failed >>> >>> But also in v7: >>> * GUP-fast allows pinning (anonymous) pages in MAP_PRIVATE file >>> mappings >>> * Ordinary GUP allows pinning pages in MAP_PRIVATE file mappings >>> >>> In v8: >>> * pin_user_pages_fast() succeeds >>> * vfio_pin_pages_remote() succeeds >>> >>> But also in v8: >>> * GUP-fast allows pinning (anonymous) pages in MAP_PRIVATE file >>> mappings >>> * Ordinary GUP allows pinning pages in MAP_PRIVATE file mappings >>> >>> >>> I have to speculate, but ... could it be that you are using a private mapping? >>> >>> In QEMU, unfortunately, the default for memory-backend-file is "share=off" (private) ... for memory-backend-memfd it is "share=on" (shared). The default is stupid ... >>> >>> If you invoke QEMU manually, can you specify "share=on" for the memory-backend-file? I thought libvirt would always default to "share=on" for file mappings (everything else doesn't make much sense) ... but you might have to specify >>> <access mode="shared"/> >>> in addition to >>> <source type="file"/> >>> >> >> Ah, there we go. Yes, I was using the default of share=off. When I instead specify share=on, now the pins will fail in both cases. >> > > Out of curiosity, how does that manifest? > > I assume the VM is successfully created and as Linux tries initializing and using the device, we get a bunch of errors inside the VM, correct? > Yes, that's correct. Which error comes first (an attempt at mapping something via type1 iommu or an attempt to register AEN) depends on the device type and the order of operations of the associated driver. But in either case, you're going to see guest errors associated with that action. mlx5 and ism give up rather quickly and just fail their probe. nvme in the guest is persistent and its actions keep re-attempting to setup AEN by issuing the associated instruction; but the associated blockdev will never show up.
