On 01/03/2024 18:47, Catalin Marinas wrote: > On Mon, Feb 26, 2024 at 12:03:21PM +0000, Ryan Roberts wrote: >> Make clear the atmicity/consistency requirements of the API and how we >> achieve them. >> >> Link: https://lore.kernel.org/linux-mm/Zc-Tqqfksho3BHmU@xxxxxxx/ >> Signed-off-by: Ryan Roberts <ryan.roberts@xxxxxxx> >> --- >> arch/arm64/mm/contpte.c | 24 ++++++++++++++---------- >> 1 file changed, 14 insertions(+), 10 deletions(-) >> >> diff --git a/arch/arm64/mm/contpte.c b/arch/arm64/mm/contpte.c >> index be0a226c4ff9..1b64b4c3f8bf 100644 >> --- a/arch/arm64/mm/contpte.c >> +++ b/arch/arm64/mm/contpte.c >> @@ -183,16 +183,20 @@ EXPORT_SYMBOL_GPL(contpte_ptep_get); >> pte_t contpte_ptep_get_lockless(pte_t *orig_ptep) >> { >> /* >> - * Gather access/dirty bits, which may be populated in any of the ptes >> - * of the contig range. We may not be holding the PTL, so any contiguous >> - * range may be unfolded/modified/refolded under our feet. Therefore we >> - * ensure we read a _consistent_ contpte range by checking that all ptes >> - * in the range are valid and have CONT_PTE set, that all pfns are >> - * contiguous and that all pgprots are the same (ignoring access/dirty). >> - * If we find a pte that is not consistent, then we must be racing with >> - * an update so start again. If the target pte does not have CONT_PTE >> - * set then that is considered consistent on its own because it is not >> - * part of a contpte range. >> + * The ptep_get_lockless() API requires us to read and return *orig_ptep >> + * so that it is self-consistent, without the PTL held, so we may be >> + * racing with other threads modifying the pte. Usually a READ_ONCE() >> + * would suffice, but for the contpte case, we also need to gather the >> + * access and dirty bits from across all ptes in the contiguous block, >> + * and we can't read all of those neighbouring ptes atomically, so any >> + * contiguous range may be unfolded/modified/refolded under our feet. >> + * Therefore we ensure we read a _consistent_ contpte range by checking >> + * that all ptes in the range are valid and have CONT_PTE set, that all >> + * pfns are contiguous and that all pgprots are the same (ignoring >> + * access/dirty). If we find a pte that is not consistent, then we must >> + * be racing with an update so start again. If the target pte does not >> + * have CONT_PTE set then that is considered consistent on its own >> + * because it is not part of a contpte range. >> */ > > I haven't had the time to properly think about this function but, > depending on what its semantics are, we might not guarantee that, at the > time of reading a pte, we have the correct dirty state from the other > ptes in the range. > > Theoretical: let's say we read the first pte in the contig range and > it's clean but further down there's a dirty one. Another (v)CPU breaks > the contig range, sets the dirty bit everywhere, there's some > pte_mkclean for all of them and they are collapsed into a contig range > again. The function above on the first (v)CPU returns a clean pte when > it should have actually been dirty at the time of read. But I think that still conforms to semantics of the function. If you had the same situation with a non-contpte mapping, the first thread may read the PTE at any time; when it's dirty, or after its been cleaned by the other thread. It's inherrently racy. All that matters is that what is returned is _consistent_; you don't want to be in a position where the first read of the block is mapping one folio, then by the time all the access/dirty bits are read, the mapping is actually to a different folio - that's an example of being inconsistent. > > Throughout the callers of this function, I couldn't find one where it > matters. So I concluded that they don't need the dirty state. Normally > the dirty state is passed to the page flags, so not lost after the pte > has been cleaned. I agree we can simplify the semantics. But I think its better done in a separate series (which I previously linked). What's the bottom line here? Are you ok with this comment as a short term solution for now, or do you want something more radical (i.e. push to get the series that does these simplifications reviewed and in time for v6.9). I still believe the current ptep_get_lockless() implementation is correct. So given I have a plan to simplify in the long run, I hope we can still get this series into v6.9 as planned. Thanks, Ryan
