Re: [PATCH] mm/gup: batch PTE-mapped large folios in gup_fast_pte_range()

[David Hildenbrand (Arm)]

On 9/17/26 10:37, Yuan-Hao Hsu wrote:
> GUP-fast grabs a PTE-mapped large folio one page at a time.  Every PTE
> costs a try_grab_folio_fast() (a refcount cmpxchg, plus the pincount
> atomic, a full barrier and a node stat update for FOLL_PIN), a
> gup_fast_folio_allowed() and a folio_set_referenced(), so a 64 kB mTHP
> pays sixteen of each and a PTE-mapped 2 MB THP pays 512.  The PMD and
> PUD leaf paths already take the whole range with one
> try_grab_folio_fast() call, and the slow path is getting the same
> treatment for PTEs in Rik's follow_page_mask() series.
> 
> The users are the hot ones: iov_iter_extract_user_pages() pins every
> O_DIRECT buffer through pin_user_pages_fast(), io_uring registers its
> buffers through it, and so does RDMA memory registration.  With mTHP
> those buffers are PTE-mapped large folios.
> 
> After the head page has been grabbed and verified as before, extend the
> run to the following PTEs that map the next pages of the same folio
> with the same protection bits, take the remaining references for the
> run with try_grab_folio(), and fill pages[] from the run.  The PTEs of
> the run are read while the reference on the folio is held, so the folio
> can neither be freed nor split underneath the scan; a THP collapse can
> still detach the page table, so the pmd is checked again after the run
> is grabbed and the run is dropped if it changed, leaving the next
> iteration to hit the existing pmd check and bail out.  A read-only run
> of an anonymous folio stops at the first page that gup_must_unshare()
> rejects, so the per-page exclusivity guarantee for FOLL_PIN is kept.
> 
> folio_pte_batch_flags() does the scan.  It reads the entries with
> ptep_get(), which is not safe without the page table lock on
> CONFIG_GUP_GET_PXX_LOW_HIGH (x86 PAE, mips32, sh) or with arm64 contpte,
> and GUP-fast reads every PTE with ptep_get_lockless() for that reason.
> Add FPB_LOCKLESS to make the helper do the same; it is inlined per call
> site, so the other callers do not change.
> 
> The order-0 path keeps its instruction sequence.  The folio flags are
> tested before folio_set_referenced() on purpose: a load of the flags
> right after that locked instruction cost about 25% on order-0 pages in
> the gup_test benchmark below, and reading them first brings that path
> back to the baseline.
> 
> gup_test ioctl over a 256 MB anonymous region, FOLL_WRITE, 65536 pages
> per call, median of 15 runs, two baseline boots / one patched boot, on
> an x86-64 VM:
> 
>                        get_user_pages_fast()   pin_user_pages_fast()
>                        before       after      before       after
>   4 kB pages           849/789 us   783 us     1174/1138 us 1094 us
>   64 kB mTHP           812/811 us   126 us     1215/1113 us  188 us
>   1 MB mTHP            799/774 us    59 us     1185/1077 us   64 us
>   2 MB THP, PTE-mapped 803/786 us    57 us     1159/1120 us   58 us
>   2 MB THP, PMD-mapped  42/40 us     40 us       46/41 us     42 us
> 
> fio O_DIRECT randread of a null_blk device (CPU bound) with the I/O
> buffer in 64 kB mTHP, median of 15 five-second runs:
> 
>   bs=1M psync                 126.2/125.6 GB/s  ->  220.6 GB/s
>   bs=1M io_uring, iodepth 16   71.2/71.6 GB/s   ->  112.1 GB/s
>   bs=64k io_uring, iodepth 16  50.2/50.7 GB/s   ->   60.1 GB/s
> 
> IORING_REGISTER_BUFFERS of 1 GB, median of 15:
> 
>   64 kB mTHP           6179/5998 us  ->  1599 us
>   1 MB mTHP            5790/5801 us  ->   893 us
>   4 kB pages           7173/6741 us  ->  6763 us
> 

I think Rik was already working on this and sent some patches. Anyhow, there is
quite some GUP review backlog I have t go through, so this will have to wait.

-- 
Cheers,

David