Hi Marek,
On Thu, Mar 28, 2019 at 4:19 AM Marek Vasut <marek.vasut@xxxxxxxxx> wrote:
> On 3/27/19 1:22 PM, Geert Uytterhoeven wrote:
> > On Wed, Mar 27, 2019 at 12:30 PM Simon Horman <horms@xxxxxxxxxxxx> wrote:
> >> On Mon, Mar 25, 2019 at 12:41:01PM +0100, marek.vasut@xxxxxxxxx wrote:
> >>> From: Marek Vasut <marek.vasut+renesas@xxxxxxxxx>
> >>> The MSI message address in the RC address space can be 64 bit. The
> >>> R-Car PCIe RC supports such a 64bit MSI message address as well.
> >>> The code currently uses virt_to_phys(__get_free_pages()) to obtain
> >>> a reserved page for the MSI message address, and the return value
> >>> of which can be a 64 bit physical address on 64 bit system.
> >>>
> >>> However, the driver only programs PCIEMSIALR register with the bottom
> >>> 32 bits of the virt_to_phys(__get_free_pages()) return value and does
> >>> not program the top 32 bits into PCIEMSIAUR, but rather programs the
> >>> PCIEMSIAUR register with 0x0. This worked fine on older 32 bit R-Car
> >>> SoCs, however may fail on new 64 bit R-Car SoCs.
> >>>
> >>> Since from a PCIe controller perspective, an inbound MSI is a memory
> >>> write to a special address (in case of this controller, defined by
> >>> the value in PCIEMSIAUR:PCIEMSIALR), which triggers an interrupt, but
> >>> never hits the DRAM _and_ because allocation of an MSI by a PCIe card
> >>> driver obtains the MSI message address by reading PCIEMSIAUR:PCIEMSIALR
> >>> in rcar_msi_setup_irqs(), incorrectly programmed PCIEMSIAUR cannot
> >>> cause memory corruption or other issues.
> >>>
> >>> There is however the possibility that if virt_to_phys(__get_free_pages())
> >>> returned address above the 32bit boundary _and_ PCIEMSIAUR was programmed
> >>> to 0x0 _and_ if the system had physical RAM at the address matching the
> >>> value of PCIEMSIALR, a PCIe card driver could allocate a buffer with a
> >>> physical address matching the value of PCIEMSIALR and a remote write to
> >>> such a buffer by a PCIe card would trigger a spurious MSI.
> >>>
> >>> Signed-off-by: Marek Vasut <marek.vasut+renesas@xxxxxxxxx>
> >>> Cc: Geert Uytterhoeven <geert+renesas@xxxxxxxxx>
> >>> Cc: Phil Edworthy <phil.edworthy@xxxxxxxxxxx>
> >>> Cc: Simon Horman <horms+renesas@xxxxxxxxxxxx>
> >>> Cc: Wolfram Sang <wsa@xxxxxxxxxxxxx>
> >>> Cc: linux-renesas-soc@xxxxxxxxxxxxxxx
> >>> To: linux-pci@xxxxxxxxxxxxxxx
> >>> Reviewed-by: Geert Uytterhoeven <geert+renesas@xxxxxxxxx>
> >>
> >> Does this warrant a Fixes tag?
> >
> > (digging in old sent email)
> > Fixes: 290c1fb358605402 ("PCI: rcar: Add MSI support for PCIe")
>
> But does it really fix that commit, given that on Gen2 and earlier, it
> was not broken as those were 32bit platforms ?
It does not fix the bug on that commit, as the bug cannot happen on arm32.
It does fix that commit, in that that commit used "unsigned long" for a
physical address, which is wrong, even on arm32 (esp. with LPAE).
If you insist on having a Fixes tag for a commit where the bug could be
seen:
Fixes: e015f88c368da1e6 ("PCI: rcar: Add support for R-Car H3 to pcie-rcar")
Apart from that, drivers should use the DMA API instead of virt_to_phys().
However, now we have a better understanding of how MSI interrupts work,
we don't even need to allocate that page. All we need is the physical
address of a page that is guaranteed not to be backed by RAM
(i.e. not to be a valid target for a legitimate PCI bus mastering transaction).
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
