[PATCH v29 3/9] arm64: kdump: reserve memory for crash dump kernel

Mark Rutland mark.rutland at arm.com
Mon Jan 23 02:23:15 PST 2017


On Mon, Jan 23, 2017 at 06:51:46PM +0900, AKASHI Takahiro wrote:
> Mark,
> 
> On Thu, Jan 19, 2017 at 11:28:50AM +0000, Mark Rutland wrote:
> > On Thu, Jan 19, 2017 at 06:49:42PM +0900, AKASHI Takahiro wrote:
> > > On Tue, Jan 17, 2017 at 11:54:42AM +0000, Mark Rutland wrote:
> > > > On Tue, Jan 17, 2017 at 05:20:44PM +0900, AKASHI Takahiro wrote:
> > > > > On Fri, Jan 13, 2017 at 11:39:15AM +0000, Mark Rutland wrote:

> > > > I don't think we have much code useful for unmapping. We could re-use 
> > > > create_mapping_late for this, passing a set of prot bits that means the
> > > > entries are invalid (e.g. have a PAGE_KERNEL_INVALID).
> > > 
> > > Do you really think that we should totally invalidate mmu entries?
> > > I guess that, given proper cache & TLB flush operations, RO attribute is
> > > good enough for memory consistency, no?
> > > (None accesses the region, as I said, except in the case of re-loading
> > > crash dump kernel though.)
> > 
> > My worry is that the first kernel and kdump kernel may map (portions of)
> > the region with potentially confliciting memory attributes. So it would
> > be necessary to completely unmap the region.
> 
> I think that this can happen only if the second kernel boots up,
> leaving non-crashed cpus still running for some reason.

Yes. I was considering a kdump case where a secondary was stuck in the
first kernel.

> > You raise a good point that this would also mean we need to perform some
> > cache maintenance, which makes that a little more painful. We'd need a
> > sequence like:
> > 
> > * Unmap the region
> > * TLB invalidation
> > * Remap the region with non-cacheable attributes
> > * Cache maintenance
> > * Unmap the region
> > * TLB invalidation
> 
> I don't get why we need to remap the region and do cache
> maintenance here. Please elaborate a bit more?

I think I was wrong, and we don't need to. Sorry about that.

My thought was that to ensure that there aren't stale lines with
differing attributes, we'd need to do a clean+invalidate while the
caches are guaranteed to not allocate anything furthe. Hence, we'd need
to use a non-cacheable mapping to perform the clean+invalidate.

However, I now think that so long as we unmap the range, this shouldn't
matter. The new kernel can perform the maintenance if it wishes to use
different attributes, similarly to what the first kernel must do per the
boot protocol.

> My current implementation of arch_kexec_protect_crashkres() is:
> 
>         kexec_segment_flush(kexec_crash_image);
>         create_mapping_late(crashk_res.start, ..., __pgprot(0));
>                                                 or PAGE_KERNEL_INVALID
>         flush_tlb_all();
> 
> kexec_segment_flush() will eventually do dcache-flush for all the modified
> data in crash dump kernel memory.

I now think this should be fine, per the above.

Thanks,
Mark.



More information about the kexec mailing list