[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20180317005513.GB22414@axis.com>
Date: Sat, 17 Mar 2018 01:55:13 +0100
From: Niklas Cassel <niklas.cassel@...s.com>
To: Lorenzo Pieralisi <lorenzo.pieralisi@....com>
Cc: kishon@...com, Bjorn Helgaas <bhelgaas@...gle.com>,
Sekhar Nori <nsekhar@...com>, John Keeping <john@...anate.com>,
Shawn Lin <shawn.lin@...k-chips.com>,
Cyrille Pitchen <cyrille.pitchen@...e-electrons.com>,
linux-pci@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH v4 1/5] PCI: endpoint: BAR width should not depend on
sizeof dma_addr_t
On Fri, Mar 16, 2018 at 06:02:20PM +0000, Lorenzo Pieralisi wrote:
> On Thu, Mar 08, 2018 at 02:33:26PM +0100, Niklas Cassel wrote:
> > If a BAR supports 64-bit width or not depends on the hardware,
> > and should thus not depend on sizeof(dma_addr_t).
> >
> > Since this driver is generic, default to always using BAR width
> > of 32-bits. 64-bit BARs can easily be tested by replacing
> > PCI_BASE_ADDRESS_MEM_TYPE_32 with PCI_BASE_ADDRESS_MEM_TYPE_64
> > in bar_flags.
> >
> > Signed-off-by: Niklas Cassel <niklas.cassel@...s.com>
> > ---
> > Note to Lorenzo/Bjorn:
> > It is not trivial to convert the bar_size + bar_flags +
> > struct pci_epf->bar member array to an array of struct resources,
> > since we need to be able to store the addresses returned
> > by dma_alloc_coherent(), which is of type dma_addr_t.
> > struct resource uses resource_size_t, which is defined as phys_addr_t.
> > E.g. ARTPEC-7 uses 64-bit dma_addr_t, but only 32-bit phys_addr_t.
> >
> > drivers/pci/endpoint/functions/pci-epf-test.c | 15 +++++++++------
> > 1 file changed, 9 insertions(+), 6 deletions(-)
> >
> > diff --git a/drivers/pci/endpoint/functions/pci-epf-test.c b/drivers/pci/endpoint/functions/pci-epf-test.c
> > index 800da09d9005..7c70433b11a7 100644
> > --- a/drivers/pci/endpoint/functions/pci-epf-test.c
> > +++ b/drivers/pci/endpoint/functions/pci-epf-test.c
> > @@ -71,6 +71,14 @@ struct pci_epf_test_data {
> > };
> >
> > static int bar_size[] = { 512, 512, 1024, 16384, 131072, 1048576 };
> > +static int bar_flags[] = {
> > + PCI_BASE_ADDRESS_SPACE_MEMORY | PCI_BASE_ADDRESS_MEM_TYPE_32,
> > + PCI_BASE_ADDRESS_SPACE_MEMORY | PCI_BASE_ADDRESS_MEM_TYPE_32,
> > + PCI_BASE_ADDRESS_SPACE_MEMORY | PCI_BASE_ADDRESS_MEM_TYPE_32,
> > + PCI_BASE_ADDRESS_SPACE_MEMORY | PCI_BASE_ADDRESS_MEM_TYPE_32,
> > + PCI_BASE_ADDRESS_SPACE_MEMORY | PCI_BASE_ADDRESS_MEM_TYPE_32,
> > + PCI_BASE_ADDRESS_SPACE_MEMORY | PCI_BASE_ADDRESS_MEM_TYPE_32
> > +};
>
> Niklas,
>
> I think you are almost there, I have one question though to address
> that can even simplify the patchset.
>
> If, according to your own commit logs (and my reading of the code), the
> Cadence driver makes a decision on the BAR size just by checking the
> corresponding region size (I would be happy to hear the reason
> underpinning that choice, BTW), why can't we do the same for DWC (ie to
> let the DWC driver decides whether a BAR should be 64 or 32 bits ?)
We could, but I think that would be a mistake.
The API that the user/"endpoint function" has available to configure
the BARs:
pci_epc_set_bar()
If the user, for some reason, wants to configure a BAR with a
64-bit width, even though the BAR size is less than 4 GB,
I think that the API should allow that.
This would not be possible if pci_epc_set_bar() would start to
ignore PCI_BASE_ADDRESS_MEM_TYPE_64 from the flags parameter
(and BAR width thus only being controlled by the size parameter).
int pci_epc_set_bar(struct pci_epc *epc, u8 func_no, enum pci_barno bar,
dma_addr_t bar_phys, size_t size, int flags);
Best regards,
Niklas
Powered by blists - more mailing lists