Re: [PATCH] cxl/acpi: Verify CHBS length for CXL2.0

[Dan Williams]

Zhijian Li (Fujitsu) wrote:
> 
> 
> On 27/03/2025 21:36, Dan Williams wrote:
> > Zhijian Li (Fujitsu) wrote:
> >>
> >>
> >> On 27/03/2025 11:44, Ira Weiny wrote:
> >>> Li Zhijian wrote:
> >>>> Per CXL Spec r3.1 Table 9-21, both CXL1.1 and CXL2.0 have defined their
> >>>> own length, verify it to avoid an invalid CHBS
> >>>
> >>>
> >>> I think this looks fine.  But did a platform have issues with this?
> >>
> >> Not really, actually, I discovered it while reviewing the code and
> >> CXL specification.
> >>
> >> Currently, this issue arises only when I inject an incorrect length
> >> via QEMU environment. Our hardware does not experience this problem.
> >>
> >>
> >>> Does this need to be backported?
> >> I remain neutral :)
> > 
> > What does the kernel do with this invalid CHBS from QEMU? I would be
> > happy to let whatever bad effect from injecting a corrupted CHBS just
> > happen because there are plenty of ways for QEMU to confuse the kernel
> > even if the table lengths are correct.
> > 
> > Unless it has real impact I would rather not touch the kernel for every
> > possible way that QEMU can make a mistake.
> 
> 
> 
> Thank you for the feedback.
> 
> If your earlier comments were specifically about ***backporting*** this patch,
> I agree there might not be an urgent need for that.
> 
> However, regarding the discussion on whether this patch should be accepted
> upstream, TBH, I believe it is necessary.
> 
> 1. The **CXL Specification (r3.1, Table 9-21)** explicitly defines `length`
> requirements for CHBS in both CXL 1.1 and CXL 2.0 cases. Failing to
> validate this field against the spec risks misinterpretation of invalid
> configurations.

The point is that the kernel has gotten by without this check and does
not need to be aggressive. Anything more than this hunk below is
overkill:

diff --git a/drivers/cxl/acpi.c b/drivers/cxl/acpi.c
index cb14829bb9be..fbcb93e5beb5 100644
--- a/drivers/cxl/acpi.c
+++ b/drivers/cxl/acpi.c
@@ -482,6 +482,10 @@ static int cxl_get_chbs_iter(union acpi_subtable_headers *header, void *arg,
            chbs->length != CXL_RCRB_SIZE)
                return 0;
 
+       if (chbs->cxl_version == ACPI_CEDT_CHBS_VERSION_CXL20 &&
+           chbs->length != ACPI_CEDT_CHBS_LENGTH_CXL20)
+               return 0;
+
        if (!chbs->base)
                return 0;
 
> 2. As mentioned in section **2.13.8** of the *CXL Memory Device Software Guide (Rev 1.0)*,
> It's recommended to verify the CHBS length.
> 
> While the immediate impact might be limited to edge cases (e.g., incorrect QEMU configurations),
> upstreaming this aligns the kernel with spec-mandated checks and improves
> robustness for future use cases.

What set me off was that this patch was:

 1 file changed, 29 insertions(+), 16 deletions(-)

...motivated by a buggy QEMU configuration, and that the kernel has been
fine to not carry self-defense against for years. So the check has
literally not mattered in practice for a long time.

I think it is ok to do that minimal validation I suggest above to pair
with the v1.1 length check, but in general there are more ways than the
length to produce a broken CHBS and I do not want to encourage a
cxl_chbs_verify() approach to gather more and more theoretical checks
unless and until we start seeing these quirks impacting the kernel in
production use cases. Buggy QEMU is not a suitable justification for
code refactoring.