Re: [PATCH 1/2] PCI: endpoint: strictly apply bar fixed size to allocate space

[Niklas Cassel <cassel@kernel.org>]

On Mon, Mar 31, 2025 at 04:39:33PM +0200, Jerome Brunet wrote:
> On Mon 31 Mar 2025 at 10:14, Niklas Cassel <cassel@kernel.org> wrote:
> >
> > Perhaps the solution is to add another struct member to struct pci_epf_bar,
> > size (meaning actual BAR size, which will be written to the BAR mask register)
> > and backing_mem_size.
> >
> > Or.. we modify pci_epf_alloc_space() to allocate an aligned size, but the
> > size that we store in (struct pci_epf_bar).size is the unaligned size.
> 
> I tried this and it works. As pointed above, as long as pci_epc_set_bar() is
> happy, it will work for me since the dwc-ep driver does not really care for
> the size given with fixed BARs.
> 
> However, when doing so, it gets a bit trick to properly call
> dma_free_coherent() as we don't have the size actually allocated
> anymore. It is possible to compute it again but it is rather ugly.

You are right.


> 
> It would probably be best to add a parameter indeed, to track the size
> allocated with dma_alloc_coherent(). What about .aligned_size ? Keeping
> .size to track the actual bar size requires less modification I think.

Your problem should be able to be reproducible also for BAR type
BAR_PROGRAMMABLE, when sending in a size smaller than epc_features.align
to pci_epf_alloc_space(), which will then modify epf_bar.size to be aligned.

The call to set_bar() will then use the aligned size instead of the
"BAR size" when writing the BAR mask register, which is wrong.
(This means that the BAR size seen by the host is the aligned size and not
"BAR size".)

So yes, I think having both size and aligned_size (or whatever name) is the
way to go.


Kind regards,
Niklas