Re: [PATCH v4] arm64: kdump: simplify the reservation behaviour of crashkernel=,high

[Baoquan He <bhe@redhat.com>]

On 03/23/23 at 05:25pm, Catalin Marinas wrote:
> On Mon, Mar 20, 2023 at 09:12:08PM +0800, Baoquan He wrote:
> > On 03/17/23 at 06:05pm, Catalin Marinas wrote:
> > > On Fri, Mar 17, 2023 at 11:09:13PM +0800, Baoquan He wrote:
> > > > In fact, what I want to achieve is we set CRASH_ADDR_LOW_MAX to 4G
> > > > fixedly on arm64, just like what we do on x86_64. As for RPi4 platform,
> > > > we leave it to crashkernel=size@offset syntax. Two reasons for this:
> > > > 1) crashkernel is needed on enterprise platform, such as workstation or
> > > > server. While RPi4 is obviously not the target. I contacted several RPi4
> > > > players in Redhat and my friends, none of them ever played kdump
> > > > testing. If they really have to, crashkernel=size@offset is enough for
> > > > them to set.
> > > 
> > > I'd like crashkernel=size (without @offset) on RPi4 to still do the
> > > right thing: a low allocation at least as we had until recently (or
> > > high+low where high here is maybe above 1GB). IOW, no regression for
> > > this crashkernel=size case. We can then change the explicit
> > > crashkernel=x,high to mean only above 4GB irrespective of the platform
> > > and crashkernel=x,low to be below arm64_dma_phys_limit.
> > 
> > Since crashkernel=,high and crashkernel=size fallback was added in arm64
> > recently, with my understanding, you are suggesting:
> > 
> > on arm64:
> > RPi4:
> > crashkernel=size
> > 0~1G: low memory (no regression introduced)
> 
> And, if not enough low memory, fall back to memory above 1GB (for RPi4;
> it would be above 4GB for any other system).
> 
> > crashkernel=size,high
> > 0~1G: low memory | 4G~top: high memory
> 
> Yes.
> 
> > Other normal system:
> > crashkernel=size|crashkernel=size,high
> > 0~4G: low memory | 4G~top: high memory
> 
> Yes.
> 
> IOW, specifying 'high' only forces the high allocation above 4GB instead
> of arm64_dma_phys_limit, irrespective of the platform. If no 'high'
> specified search_base remains CRASH_ADDR_LOW_MAX (1GB on RPi4, 4GB for
> the rest).
> 
> > > > 2) with the fixed CRASH_ADDR_LOW_MAX as 4G, we can easily fix the
> > > > problem of base page mapping for the whole linear mapping if crsahkernel=
> > > > is set in kernel parameter shown in [1] at bottom. 
> > > 
> > > That's a different problem ;). I should re-read that thread, forgot most
> > > of the details but I recall one of the counter arguments was that there
> > > isn't a strong case to unmap the crashkernel reservation. Now, if we
> > > place crashdump kernel image goes in the 'high' reservation, can we not
> > > leave the 'low' reservation mapped? We don't really care about it as it
> > > wouldn't have any meaningful code/data to be preserved. If the 'high'
> > > one goes above 4G always, we don't depend on the arm64_dma_phys_limit.
> > 
> > Yes, this looks ideal. While it only works when crashkernel=,high case and
> > it succeeds to reserve a memory region for the specified size of crashkernel
> > high memory. At below, we have 4 cases of crashkernel= syntax:
> > 
> > crashkernel=size
> > 1)first attempt:  low memory under arm64_dma_phys_limit
> > 2)fallback:       finding memory above 4G
> 
> (2) should be 'finding memory above arm64_dma_phys_limit' to keep the
> current behaviour for RPi4.

Then for RPi4, with case 2), it will find memory above
arm64_dma_phys_limit, namely 1G. Then it will get two memory regions,
one could be in [1G, 4G], another is below 4G. I am fine with this, as
long as it won't cause confusion that people may think two low memory
regions you mentioned earlier. Please help confirm if I understand your
suggestioin correctly. I will start making patch with this clarified.
Thanks.

> 
> > crashkernel=size,high
> > 3)first attempt:  finding memory above 4G
> > 4)fallback:       low memory under arm64_dma_phys_limit
> 
> Yes.
> 
> > case 3) works with your suggestion. However, 1), 2), 4) all need to
> > defer to bootmem_init(). With these cases and different handling,
> > reserve_crashkernel() could be too complicated.
> 
> Ah, because of the fallback below arm64_dma_phys_limit as in (4), we
> still can't move the full crashkernel reservation early. Well, we could
> do it in two steps: (a) early attempt at crashkernel reservation above
> 4G if 'high' was specified and we avoid mapping it if successful and (b)
> do the late crashkernel reservation below arm64_dma_phys_limit and skip
> unmapping as being too late. This way most server-like platforms would
> get a reservation above 4G, unmapped.

Yeah, this covers case 3), while other cases are still in pit.

> 
> > I am wondering if we can cancel the protection of crashkernel memory
> > region on arm64 for now. In earlier discussion, people questioned if the
> > protection is necessary on arm64. After comparison, I would rather take
> > away the protection method of crashkernel region since they try to
> > protect in a chance in one million , while the base page mapping for the
> > whole linear mapping is mitigating arm64 high end server always.
> 
> This works for me. We can add the protection later for addresses above
> 4GB only as mentioned above.

Thanks, I have posted a patchset to cancel the protection on crashkernel
memory region as per your confirmation here. This can give distros a
chance to back port them to fix the performance issue caused by the base
page mapping. I personally expect we can hold the crashkernel region
unprotected till we have a ideal solution since the code will be elegant
with comfortable simplicity.

Let's wait and see the code change if people interested want to keep the
protection methods.


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