On Tue, 23 Jun 2026 20:55:32 +0000
Pranjal Shrivastava <[email protected]> wrote:

> On Tue, Jun 23, 2026 at 09:44:46AM +0100, David Laight wrote:
> 
> Hi David,
> 
> > On Tue, 23 Jun 2026 01:54:59 +0000
> > David Hu <[email protected]> wrote:
> >   
> > > Currently, `fill_sg_entry()` splits the scatterlist using `UINT_MAX`.
> > > This creates a non-page-aligned DMA length (`0xFFFFFFFF`) for the
> > > first entry, resulting in non-page-aligned DMA addresses for all
> > > subsequent entries.  
> > 
> > There is a separate issue of whether this code is even needed at all.
> > Where can transfers over 2G (never mind 4G) actually come from.
> > 
> > The read, write and similar system calls limit transfers to INT_MAX
> > (even on 64bit) and a lot of driver code will need fixing it longer
> > lengths are allowed though.
> > io_uring better enforce the same limits.
> > So the transfers can come directly from userspace.
> > 
> > Not only that but you also need a single physically contiguous buffer.
> > Good luck allocating that!
> > 
> > Now maybe there are some peer-to-peer places where the large buffer
> > is device memory, but they will be unusual and probably need
> > special treatment anyway.
> >   
> 
> I agree that traditional VFS read/write face the MAX_RW_COUNT limit 
> (~2GB), and io_uring has its limits, but I'm a little confused by the
> push to enforce these limits here in the SGL code?
> 
> File I/O seems to be only one side of the picture. In my view, this fix
> is necessary and certainly has a use-case:
> 
> For example, the RDMA subsystem has the capability to import dmabufs [1],
> which gives rise to use cases for dmabuf beyond standard file ops 
> (via VFS/io_uring). 
> 
> In these scenarios, GPU HBM can be exported as dmabufs. With recent GPUs,
> HBM capacity can be in the order of hundreds of GBs [2]. RDMA can employ
> infrastructure like the vfio-dmabuf-exporter [3] or similar dmabuf 
> exporters to frequently move huge blocks of data via P2PDMA.

Ok, that explains where big buffers can come from.
I just wasn't sure.

> If we restrict incoming dmabuf transfers to fit within VFS-centric 
> limits (2GB), we impose unnecessary overhead on the RDMA stack, forcing
> it to manage a significantly higher number of memory registrations. By 
> cleanly splitting these massive contiguous device buffers into 
> page-aligned SGL entries, we directly improve the efficiency of P2P 
> transfers and memory registration.

But a divide by '4G - PAGE_SIZE' is also non-trivial and (I think affects
a lot of io) when the quotient is always 1.
Splitting into 2G chunks is a lot cheaper.

> Since this change doesn't seem to have a negative impact on standard file
> I/O or break existing VFS constraints, I'm curious why we shouldn't 
> support splitting these >4GB P2P transfers? Am I missing something?

I was only wondering whether it was needed...
It does bring up the question of why the >4GB transfers even need splitting.
But that is another question.

If you want to split large transfers into 4G-PAGE_SIZE blocks
it is probably worth having a quick test that returns 1 for 'small' buffers.

        David

> 
> Thanks,
> Praan
> 
> [1] 
> https://elixir.bootlin.com/linux/v7.1.1/source/drivers/infiniband/core/umem_dmabuf.c#L174
>  
> [2] https://nvdam.widen.net/s/fdvdqvfvj2/hopper-h200-nvl-product-brief (Table 
> 2-2)
> [3] 
> https://elixir.bootlin.com/linux/v7.1.1/source/drivers/vfio/pci/vfio_pci_dmabuf.c#L297
> 

Reply via email to