Re: [PATCH v2 4/6] cxl/pci: Make cxl_dvsec_ranges() failure not fatal to cxl_pci

[Ben Widawsky <ben.widawsky@intel.com>]

On 22-03-17 11:20:48, Dan Williams wrote:
> On Thu, Mar 17, 2022 at 10:52 AM Ben Widawsky <ben.widawsky@intel.com> wrote:
> >
> > On 22-03-14 18:22:38, Dan Williams wrote:
> > > cxl_dvsec_ranges(), the helper for enumerating the presence of an active
> > > legacy CXL.mem configuration on a CXL 2.0 Memory Expander, is not fatal
> > > for cxl_pci because there is still value to enable mailbox operations
> > > even if CXL.mem operation is disabled. Recall that the reason cxl_pci
> > > does this initialization and not cxl_mem is to preserve the useful
> > > property (for unit testing) that cxl_mem is cxl_memdev + mmio generic,
> > > and does not require access to a 'struct pci_dev' to issue config
> > > cycles.
> > >
> > > Update 'struct cxl_endpoint_dvsec_info' to carry either a positive
> > > number of non-zero size legacy CXL DVSEC ranges, or the negative error
> > > code from __cxl_dvsec_ranges() in its @ranges member.
> > >
> > > Reported-by: Krzysztof Zach <krzysztof.zach@intel.com>
> > > Fixes: 560f78559006 ("cxl/pci: Retrieve CXL DVSEC memory info")
> > > Signed-off-by: Dan Williams <dan.j.williams@intel.com>
> > > ---
> > >  drivers/cxl/pci.c |   27 ++++++++++++++++++---------
> > >  1 file changed, 18 insertions(+), 9 deletions(-)
> > >
> > > diff --git a/drivers/cxl/pci.c b/drivers/cxl/pci.c
> > > index 257cf735505d..994c79bf6afd 100644
> > > --- a/drivers/cxl/pci.c
> > > +++ b/drivers/cxl/pci.c
> > > @@ -463,13 +463,18 @@ static int wait_for_media_ready(struct cxl_dev_state *cxlds)
> > >       return 0;
> > >  }
> > >
> > > -static int cxl_dvsec_ranges(struct cxl_dev_state *cxlds)
> > > +/*
> > > + * Return positive number of non-zero ranges on success and a negative
> > > + * error code on failure. The cxl_mem driver depends on ranges == 0 to
> > > + * init HDM operation.
> > > + */
> >
> > It shouldn't depend on 0 ranges, it depends on ranges matching existing HDM
> > decoder ranges. I took a shortcut by just checking global enable originally
> > because we didn't yet have code to enumerate decoders (anything else would be a
> > crazy BIOS bug that's probably not worth working around).
> 
> Hmm, I don't see that as a shortcut. If global enable is set then
> there is no requirement that CXL DVSEC matches the HDM decoder ranges,
> if global enable is not set then the HDM decoder ranges are not in
> effect.

It's true that there is no requirement that the DVSEC matches HDM decoder
ranges, except as you point out before it's going against spec recommendation.

> 
> This is what the new comment in cxl_hdm_decode_init() is trying to
> clarify. My proposal is that Linux ignores the recommendation which I
> think is trying to accommodate legacy CXL 1.1 software stacks which
> Linux never had.
> 

Okay, that's fine.

>         /*
>          * Per CXL 2.0 Section 8.1.3.8.3 and 8.1.3.8.4 DVSEC CXL Range 1 Base
>          * [High,Low] when HDM operation is enabled the range register values
>          * are ignored by the device, but the spec also recommends matching the
>          * DVSEC Range 1,2 to HDM Decoder Range 0,1 so, non-zero info->ranges
>          * are expected.
>          */

Maybe explicitly say the driver does not attempt to check this recommendation.