On Tue, Jun 02, 2020 at 07:34:47PM +0530, Prabhakar Kushwaha wrote:

On Mon, Jun 1, 2020 at 1:10 PM Will Deacon [off-list ref] wrote:

quoted

On Thu, May 21, 2020 at 04:52:02PM +0530, Prabhakar Kushwaha wrote:

quoted

On Thu, May 21, 2020 at 2:53 PM Will Deacon [off-list ref] wrote:

quoted

On Tue, May 19, 2020 at 08:24:21AM +0530, Prabhakar Kushwaha wrote:

quoted

What kind of issue you are foreseeing in using memcpy(). May be we can
try to find a solution.

Well the thing might not be cache-coherent to start with...

Thanks for telling possible issue area.  Let me try to explain why
this should not be an issue.

kdump kernel runs from reserved memory space defined during the boot
of first kernel. kdump does not touch memory of the previous kernel.
So no page has been created in kdump kernel  and  there should not be
any data/attribute/coherency issue from MMU point of view .

Then how does this work?:

        rdcfg.strtab = memremap(rdcfg.strtab_dma, size, MEMREMAP_WB);

You're explicitly asking for a write-back mapping.

As i mentioned earlier, I will replace it with MEMREMAP_WT to make
sure data is written into the memory.

Please note, this memmap is temporary for copying older SMMU table to
cfg->strtab.
Here, cfg->strtab & cfg->strtab_dma allocated via dmam_alloc_coherent
during SMMU probe.

quoted

quoted

During SMMU probe functions,  dmem_alloc_coherent() will be used
allocate new memory (part of existing flow).
This patch copy STE or first level descriptor to *this* memory, after
mapping physical address using memremap().
It just copy everything  so there should not be any issue related to
attribute/content.

Yes, copying  done after mapping it as MEMREMAP_WB. if you want I can
use it as MEMREMAP_WT

You need to take into account whether or not the device is coherent, and the
DMA API is designed to handle that for you. But even then, this is fragile
as hell because you end up having to infer the hardware configuration
from the device to understand the size and format of the data structures.
If the crashkernel isn't identical to the host kernel (in terms of kconfig,
driver version, firmware tables, cmdline etc) then this is very likely to
go wrong.

There are two possible scenarios for mismatched kdump kernel
1.  kdump kernel does not have the devices' driver
2.  kdump kernel have the different variation/configuration of driver

This patch create temporary SMMU table entries which are overwritten
by driver-probe.

What exactly does this achieve, given that you don't copy the context
descriptors or the page tables?

Driver's probe will overwrite SMMU entries based on its new
requirement (size, format, data structures etc).

for "1",  As  no device driver,  SMMU entry will remain there.
Means no-one looking for the copied content (even if device continued
to perform DMA).

About coherency between Cores and Memory(DMA).
At the time of crash:  Only one CPU is allowed to remain continue,
rest are stopped.
__crash_kexec --> machine_crash_shutdown --> crash_smp_send_stop()

The active CPU is used to boot kdump kernel. hence none of the CPUs is
looking for data copied by DMA.
Coherency issue should not be there.

I'm talking about coherency between the SMMU and the CPU, so I don't think
the number of CPUs is relevant.

please let me know your view.

It still seems extremely fragile to me, so I continue to think that this
is the wrong approach.

Will

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel