Re: [PATCH 00/13] Enable compound page for p2pdma memory

[Leon Romanovsky <leon@kernel.org>]

On Wed, Dec 24, 2025 at 09:37:39AM +0800, Hou Tao wrote:
> 
> 
> On 12/24/2025 9:18 AM, Hou Tao wrote:
> > Hi,
> >
> > On 12/21/2025 8:19 PM, Leon Romanovsky wrote:
> >> On Sat, Dec 20, 2025 at 12:04:33PM +0800, Hou Tao wrote:
> >>> From: Hou Tao <houtao1@huawei.com>
> >>>
> >>> Hi,
> >>>
> >>> device-dax has already supported compound page. It not only reduces the
> >>> cost of struct page significantly, it also improve the performance of
> >>> get_user_pages when 2MB or 1GB page size is used. We are experimenting
> >>> to use p2p dma to directly transfer the content of NVMe SSD into NPU.
> >> I’ll admit my understanding here is limited, and lately everything tends  
> >> to look like a DMABUF problem to me. Could you explain why DMABUF support 
> >> is not being used for this use case?
> > I have limited knowledge of dma-buf, so correct me if I am wrong. It
> > seems that as for now there is no available way to use the dma-buf to
> > read/write files. For the userspace vaddr backended by  the dma-buf, it
> > is a PFN mapping, get_user_pages() will reject such address.
> 
> Hit the send button too soon :) So In my understanding, the advantage of
> dma-buf is that it doesn't need struct page.

The primary advantage of dma-buf is that it provides a safe mechanism for
sharing a DMA region between devices or subsystems. This allows reliable
p2p communication between two devices. For example, a GPU and an RDMA NIC
can share a memory region for data transfer.

The ability to operate without a struct page is an important part of this
design.

> and it also means that it needs special handling to support IO
> from/to dma-buf (e.g.,  [RFC v2 00/11] Add dmabuf read/write via io_uring [1])

It looks like that read/write support is needed for IO data transfer,
but you talked about CMB.

I would imagine that NVMe exported CMB through dmabuf and your NPU will
import it without need to do any read/write at all.

Thanks

> 
> [1]
> https://lore.kernel.org/io-uring/cover.1763725387.git.asml.silence@gmail.com/
> >> Thanks
> >>
> >>> The size of NPU HBM is 32GB or larger and there are at most 8 NPUs in
> >>> the host. When using the base page, the memory overhead is about 4GB for
> >>> 128GB HBM, and the mapping of 32GB HBM into userspace takes about 0.8
> >>> second. Considering ZONE_DEVICE memory type has already supported the
> >>> compound page, enabling the compound page support for p2pdma memory as
> >>> well. After applying the patch set, when using the 1GB page, the memory
> >>> overhead is about 2MB and the mmap costs about 0.04 ms.
> >>>
> >>> The main difference between the compound page support of device-dax and
> >>> p2pdma is that p2pdma inserts the page into user vma during mmap instead
> >>> of page fault. The main reason is simplicity. The patch set is
> >>> structured as shown below:
> >>>
> >>> Patch #1~#2: tiny bug fixes for p2pdma
> >>> Patch #3~#5: add callbacks support in kernfs and sysfs, include
> >>> pagesize, may_split and get_unmapped_area. These callbacks are necessary
> >>> for the support of compound page when mmaping sysfs binary file.
> >>> Patch #6~#7: create compound page for p2pdma memory in the kernel. 
> >>> Patch #8~#10: support the mapping of compound page in userspace. 
> >>> Patch #11~#12: support the compound page for NVMe CMB.
> >>> Patch #13: enable the support for compound page for p2pdma memory.
> >>>
> >>> Please see individual patches for more details. Comments and
> >>> suggestions are always welcome.
> >>>
> >>> Hou Tao (13):
> >>>   PCI/P2PDMA: Release the per-cpu ref of pgmap when vm_insert_page()
> >>>     fails
> >>>   PCI/P2PDMA: Fix the warning condition in p2pmem_alloc_mmap()
> >>>   kernfs: add support for get_unmapped_area callback
> >>>   kernfs: add support for may_split and pagesize callbacks
> >>>   sysfs: support get_unmapped_area callback for binary file
> >>>   PCI/P2PDMA: add align parameter for pci_p2pdma_add_resource()
> >>>   PCI/P2PDMA: create compound page for aligned p2pdma memory
> >>>   mm/huge_memory: add helpers to insert huge page during mmap
> >>>   PCI/P2PDMA: support get_unmapped_area to return aligned vaddr
> >>>   PCI/P2PDMA: support compound page in p2pmem_alloc_mmap()
> >>>   PCI/P2PDMA: add helper pci_p2pdma_max_pagemap_align()
> >>>   nvme-pci: introduce cmb_devmap_align module parameter
> >>>   PCI/P2PDMA: enable compound page support for p2pdma memory
> >>>
> >>>  drivers/accel/habanalabs/common/hldio.c |   3 +-
> >>>  drivers/nvme/host/pci.c                 |  10 +-
> >>>  drivers/pci/p2pdma.c                    | 140 ++++++++++++++++++++++--
> >>>  fs/kernfs/file.c                        |  79 +++++++++++++
> >>>  fs/sysfs/file.c                         |  15 +++
> >>>  include/linux/huge_mm.h                 |   4 +
> >>>  include/linux/kernfs.h                  |   3 +
> >>>  include/linux/pci-p2pdma.h              |  30 ++++-
> >>>  include/linux/sysfs.h                   |   4 +
> >>>  mm/huge_memory.c                        |  66 +++++++++++
> >>>  10 files changed, 339 insertions(+), 15 deletions(-)
> >>>
> >>> -- 
> >>> 2.29.2
> >>>
> >>>
> 
>