Hi Michael,
On 11/11/16 at 09:37am, Michael Chan wrote:
> On Fri, Nov 11, 2016 at 6:02 AM, Baoquan He <b...@redhat.com> wrote:
> > On 11/11/16 at 09:46pm, Baoquan He wrote:
> >> Hi bnx2 experts,
> >>
> >> In commit 3e1be7a ("bnx2: Reset device during driver initialization"),
> >> firmware requesting code was moved from open stage to probe stage.
> >> The reason is in kdump kernel hardware iommu need device be reset in
> >> driver probe stage, otherwise those in-flight DMA from 1st kernel
> >> will continue going and look up into the newly created io-page tables.
> >> So we need reset device to stop in-flight DMA as early as possibe.
> >>
> >> But with commit 3e1be7a merged, people reported their bnx2 driver init
> >> failed because of failed firmware loading. After discussion, it's found
> >> that they built bnx2 driver into kernel, and that makes probe function
> >> bnx2_init_one be called in do_initcalls(). But at this time the initramfs
> >> has not been uncompressed yet and mounted, kernel can't detect firmware.
> >>
> >> So there's only one way to cover both. Try to hard reset the bnx2 device
> >> at probe stage, without involving firmware issues. I tried to add function
> >> bnx2_hard_reset_chip() to do this and it's only called in kdump kernel.
> >> The thing is I am not quite familiar with bnx2 chip spec, just abstract
> >> code from bnx2_reset_chip, the testing result is good.
> >
> > Here I changed to send BNX2_MISC_COMMAND_HD_RESET in BNX2_CHIP_5709
> > case.
> >
>
> From my old 5709 Documentation:
>
> Bit 6 HD_RESET: Writing this bit as 1 will cause the chip to do a
> hard reset like bit 5 except the sticky bits in the PCI function are
> not reset.
>
> Bit 5 POR_RESET: Writing this bit as 1 will cause the chip to do an
> internal reset exactly like a power-up reset. There is no protection
> for this request and it may cause any current PCI cycle to lock up.
> This reset is intended for use under manufacturing conditions only.
>
> So it sounds like doing HD_RESET can potentially cause a PCI bus lock up.
>
> Why not just disable DMA gracefully as done at the beginning of
> bnx2_reset_chip()?
Thanks for your suggestion.
If what I undertand is correct, you meant waiting for the current PCI
transaction to complete, right? I tried and it also works. I like this
idea, will repost v2. Please help check if it meets your thoughts.
Thanks
Baoquan
