* Re: [bug report] PCI: Check BAR index for validity
2025-03-10 7:54 ` Philipp Stanner
@ 2025-03-10 8:57 ` Philipp Stanner
2025-03-10 9:18 ` Dan Carpenter
2025-03-10 19:53 ` Bjorn Helgaas
2 siblings, 0 replies; 6+ messages in thread
From: Philipp Stanner @ 2025-03-10 8:57 UTC (permalink / raw)
To: phasta, Bjorn Helgaas, Dan Carpenter; +Cc: linux-pci
On Mon, 2025-03-10 at 08:54 +0100, Philipp Stanner wrote:
> On Sat, 2025-03-08 at 15:07 -0600, Bjorn Helgaas wrote:
> > On Sat, Mar 08, 2025 at 01:23:28PM +0300, Dan Carpenter wrote:
> > > Hello Philipp Stanner,
> > >
> > > Commit ba10e5011d05 ("PCI: Check BAR index for validity") from
> > > Mar
> > > 4,
> > > 2025 (linux-next), leads to the following Smatch static checker
> > > warning:
> > >
> > > drivers/pci/devres.c:632
> > > pcim_remove_bar_from_legacy_table()
> > > error: buffer overflow 'legacy_iomap_table' 6 <= 15
> >
> > Thanks, I dropped this patch for now.
> >
> > > drivers/pci/devres.c
> > > 621 static void pcim_remove_bar_from_legacy_table(struct
> > > pci_dev *pdev, int bar)
> > > 622 {
> > > 623 void __iomem **legacy_iomap_table;
> > > 624
> > > 625 if (!pci_bar_index_is_valid(bar))
> > >
> > > This line used to check PCI_STD_NUM_BARS (6) but now it's
> > > checking
> > > PCI_NUM_RESOURCES (15).
>
> What is even going on here. Why are thos different values? Does a PCI
> device now have at most 6, or 15 BARs?
>
> Or is a BAR different from a "resource"?
>
> And why would it be 15? I haven't read the standard, but I would
> suspect it should be 16.
>
> And which of those two here should be used?
> https://elixir.bootlin.com/linux/v6.14-rc4/source/include/linux/pci.h#L133
>
> The comment doesn't say *which one* is "preserved for backwards
> compatibility".
Furthermore, I just saw that the old pcim_ code would then also be
half-broken, because it also uses PCI_STD_NUM_BARS, whereas the pci_
functions use PCI_NUM_RESOURCES:
https://elixir.bootlin.com/linux/v6.8.9/source/drivers/pci/pci.c#L6555
P.
>
> So many questions…
>
> But granted, the check is wrong for the devres resource array, and I
> suppose it should be made the same size as pci_dev.resource.
>
>
> > >
> > > 626 return;
> > > 627
> > > 628 legacy_iomap_table = (void __iomem
> > > **)pcim_iomap_table(pdev);
> > > 629 if (!legacy_iomap_table)
> > > 630 return;
> > > 631
> > > --> 632 legacy_iomap_table[bar] = NULL;
> > > ^^^^^^^^^^^^^^^^^^^^^^^
> > > Leading to a buffer overflow.
>
> Leading to a *potential* buffer overflow.
>
> Anyways, thanks for reporting.
>
> P.
>
>
> > >
> > > 633 }
> > >
> > > regards,
> > > dan carpenter
>
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: [bug report] PCI: Check BAR index for validity
2025-03-10 7:54 ` Philipp Stanner
2025-03-10 8:57 ` Philipp Stanner
@ 2025-03-10 9:18 ` Dan Carpenter
2025-03-10 19:53 ` Bjorn Helgaas
2 siblings, 0 replies; 6+ messages in thread
From: Dan Carpenter @ 2025-03-10 9:18 UTC (permalink / raw)
To: phasta; +Cc: Bjorn Helgaas, linux-pci
On Mon, Mar 10, 2025 at 08:54:54AM +0100, Philipp Stanner wrote:
> > >
> > > 626 return;
> > > 627
> > > 628 legacy_iomap_table = (void __iomem
> > > **)pcim_iomap_table(pdev);
> > > 629 if (!legacy_iomap_table)
> > > 630 return;
> > > 631
> > > --> 632 legacy_iomap_table[bar] = NULL;
> > > ^^^^^^^^^^^^^^^^^^^^^^^
> > > Leading to a buffer overflow.
>
> Leading to a *potential* buffer overflow.
>
Smatch is doing cross function analysis in this case. Smatch knows
that pcim_iounmap_regions() is fine but the bug is when this is
called from pcim_iomap_regions().
drivers/pci/devres.c | pcim_iounmap_regions | pcim_remove_bar_from_legacy_table | PARAM_VALUE | 1 | bar | 0-5
drivers/pci/devres.c | pcim_iomap_regions | pcim_remove_bar_from_legacy_table | PARAM_VALUE | 1 | bar | 0-15
But, that raises a different question because you would expect
the map and unmap functions to loop over the same bars.
regards,
dan carpenter
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [bug report] PCI: Check BAR index for validity
2025-03-10 7:54 ` Philipp Stanner
2025-03-10 8:57 ` Philipp Stanner
2025-03-10 9:18 ` Dan Carpenter
@ 2025-03-10 19:53 ` Bjorn Helgaas
2 siblings, 0 replies; 6+ messages in thread
From: Bjorn Helgaas @ 2025-03-10 19:53 UTC (permalink / raw)
To: phasta; +Cc: Dan Carpenter, linux-pci
On Mon, Mar 10, 2025 at 08:54:54AM +0100, Philipp Stanner wrote:
> On Sat, 2025-03-08 at 15:07 -0600, Bjorn Helgaas wrote:
> > On Sat, Mar 08, 2025 at 01:23:28PM +0300, Dan Carpenter wrote:
> > > Hello Philipp Stanner,
> > >
> > > Commit ba10e5011d05 ("PCI: Check BAR index for validity") from Mar
> > > 4,
> > > 2025 (linux-next), leads to the following Smatch static checker
> > > warning:
> > >
> > > drivers/pci/devres.c:632
> > > pcim_remove_bar_from_legacy_table()
> > > error: buffer overflow 'legacy_iomap_table' 6 <= 15
> >
> > Thanks, I dropped this patch for now.
> >
> > > drivers/pci/devres.c
> > > 621 static void pcim_remove_bar_from_legacy_table(struct
> > > pci_dev *pdev, int bar)
> > > 622 {
> > > 623 void __iomem **legacy_iomap_table;
> > > 624
> > > 625 if (!pci_bar_index_is_valid(bar))
> > >
> > > This line used to check PCI_STD_NUM_BARS (6) but now it's checking
> > > PCI_NUM_RESOURCES (15).
>
> What is even going on here. Why are thos different values? Does a PCI
> device now have at most 6, or 15 BARs?
You included the link below, so I guess you found the answer to this,
but PCI devices can have:
- up to six normal BARs (pci_dev.resource[0-5])
- a ROM BAR (pci_dev.resource[6])
- up to six SR-IOV BARs (pci_dev.resource[7-12])
- four windows (for bridges) (pci_dev.resource[13-16])
> Or is a BAR different from a "resource"?
Yes, as above. Not all can be used at once, e.g., bridges can only
have two BARs and may not implement all four windows, and SR-IOV VFs
can't use the normal BARs at all.
> And why would it be 15? I haven't read the standard, but I would
> suspect it should be 16.
This is an implementation thing, not a PCI spec thing, but I count a
maximum of 17 resources if CONFIG_PCI_IOV is defined.
> And which of those two here should be used?
> https://elixir.bootlin.com/linux/v6.14-rc4/source/include/linux/pci.h#L133
^ permalink raw reply [flat|nested] 6+ messages in thread