Re: [PATCH v2 00/10] ublk: add shared memory zero-copy support

[Jens Axboe <axboe@kernel.dk>]

On 4/7/26 9:03 PM, Ming Lei wrote:
> On Tue, Apr 07, 2026 at 12:29:39PM -0700, Caleb Sander Mateos wrote:
>> On Mon, Apr 6, 2026 at 7:39 PM Ming Lei <ming.lei@redhat.com> wrote:
>>>
>>> On Tue, Mar 31, 2026 at 11:31:51PM +0800, Ming Lei wrote:
>>>> Hello,
>>>>
>>>> Add shared memory based zero-copy (UBLK_F_SHMEM_ZC) support for ublk.
>>>>
>>>> The ublk server and its client share a memory region (e.g. memfd or
>>>> hugetlbfs file) via MAP_SHARED mmap. The server registers this region
>>>> with the kernel via UBLK_U_CMD_REG_BUF, which pins the pages and
>>>> builds a PFN maple tree. When I/O arrives, the driver looks up bio
>>>> pages in the maple tree — if they match registered buffer pages, the
>>>> data is used directly without copying.
>>>>
>>>> Please see details on document added in patch 3.
>>>>
>>>> Patches 1-4 implement the kernel side:
>>>>  - buffer register/unregister control commands with PFN coalescing,
>>>>    including read-only buffer support (UBLK_SHMEM_BUF_READ_ONLY)
>>>>  - PFN-based matching in the I/O path, with enforcement that read-only
>>>>    buffers reject non-WRITE requests
>>>>  - UBLK_F_SHMEM_ZC feature flag
>>>>  - eliminate permanent pages[] array from struct ublk_buf; the maple
>>>>    tree already stores PFN ranges, so pages[] becomes temporary
>>>>
>>>> Patches 5-10 add kublk (selftest server) support and tests:
>>>>  - hugetlbfs buffer sharing (both kublk and fio mmap the same file)
>>>>  - null target and loop target tests with fio verify
>>>>  - filesystem-level test (ext4 on ublk, fio verify on a file)
>>>>  - read-only buffer registration test (--rdonly_shmem_buf)
>>>>
>>>> Changes since V1:
>>>>  - rename struct ublk_buf_reg to struct ublk_shmem_buf_reg, add __u32
>>>>    flags field for extensibility, narrow __u64 len to __u32 (max 4GB
>>>>    per UBLK_SHMEM_ZC_OFF_MASK), remove __u32 reserved (patch 1)
>>>>  - add UBLK_SHMEM_BUF_READ_ONLY flag: pin pages without FOLL_WRITE,
>>>>    enabling registration of write-sealed memfd buffers (patch 1)
>>>>  - use backward-compatible struct reading: memset zero + copy
>>>>    min(header->len, sizeof(struct)) (patch 1)
>>>>  - reorder struct ublk_buf_range fields for better packing (16 bytes
>>>>    vs 24 bytes), change buf_index to unsigned short, add unsigned short
>>>>    flags to store per-range read-only state (patch 1)
>>>>  - enforce read-only buffer semantics in ublk_try_buf_match(): reject
>>>>    non-WRITE requests on read-only buffers since READ I/O needs to
>>>>    write data into the buffer (patch 2)
>>>>  - narrow struct ublk_buf::nr_pages to unsigned int, narrow struct
>>>>    ublk_buf_range::base_offset to unsigned int (patch 1)
>>>>  - add new patch 4: eliminate permanent pages[] array from struct
>>>>    ublk_buf — recover struct page pointers via pfn_to_page() from the
>>>>    maple tree during unregistration, saving 2MB per 1GB buffer
>>>>  - add UBLK_F_SHMEM_ZC to feat_map in kublk (patch 5)
>>>>  - add new patch 10: read-only buffer registration selftest with
>>>>    --rdonly_shmem_buf option on null target + hugetlbfs
>>>
>>> Hello,
>>>
>>> Ping...
>>
>> Sorry I didn't get a chance to look at this earlier. It looks really
>> nice, thank you for implementing it! I have just a few comments. One
>> is about the UAPI (can we just use virtual addresses instead of buffer
>> index + offset), so I might wait on landing this patchset until we've
>> finalized that.
> 
> Hi Jens,
> 
> Can you drop V2 from for-7.1/block so that we can polish up it in V3?
> 
> Or I am fine to cook up a followup for misc clean & fix?

Please just do a fixup series. If I drop it now, it's not going into
7.1, and the issues aren't that big of a deal to fix up.

-- 
Jens Axboe