Linux-Doc Archive on lore.kernel.org
 help / color / Atom feed
From: Catalin Marinas <catalin.marinas@arm.com>
To: chenzhou <chenzhou10@huawei.com>
Cc: tglx@linutronix.de, mingo@redhat.com, dyoung@redhat.com,
	bhe@redhat.com, will@kernel.org, james.morse@arm.com,
	robh+dt@kernel.org, arnd@arndb.de, John.P.donnelly@oracle.com,
	prabhakar.pkin@gmail.com, nsaenzjulienne@suse.de, corbet@lwn.net,
	bhsharma@redhat.com, horms@verge.net.au, guohanjun@huawei.com,
	xiexiuqi@huawei.com, huawei.libin@huawei.com,
	linux-kernel@vger.kernel.org,
	linux-arm-kernel@lists.infradead.org, kexec@lists.infradead.org,
	linux-doc@vger.kernel.org
Subject: Re: [PATCH v10 4/5] arm64: kdump: fix kdump broken with ZONE_DMA reintroduced
Date: Wed, 29 Jul 2020 16:20:29 +0100
Message-ID: <20200729152028.GE5524@gaia> (raw)
In-Reply-To: <217004f5-dd8e-d04c-038b-c88b132d5495@huawei.com>

On Wed, Jul 29, 2020 at 10:14:32PM +0800, chenzhou wrote:
> On 2020/7/29 19:58, Catalin Marinas wrote:
> > On Wed, Jul 29, 2020 at 11:52:39AM +0800, chenzhou wrote:
> >> How about like this:
> >> 1. For ZONE_DMA issue, use Bhupesh's solution, keep the crashkernel=
> >>    behaviour to ZONE_DMA allocations.
> >> 2. For this patch series, make the reserve_crashkernel_low() to
> >>    ZONE_DMA allocations.
> > 
> > So you mean rebasing your series on top of Bhupesh's? I guess you can
> > combine the two, I really don't care which way as long as we fix both
> > issues and agree on the crashkernel= semantics. I think with some tweaks
> > we can go with your series alone.
> >
> > IIUC from the x86 code (especially the part you #ifdef'ed out for
> > arm64), if ",low" is not passed (so just standard crashkernel=X), it
> > still allocates sufficient low memory for the swiotlb in ZONE_DMA. The
> > rest can go in a high region. Why can't we do something similar on
> > arm64? Of course, you can keep the ",low" argument for explicit
> > allocation but I don't want to mandate it.
> 
> It is a good idea to combine the two.
> 
> For parameter crashkernel=X, we do like this:
> 1. allocate some low memory in ZONE_DMA(or ZONE_DMA32 if CONFIG_ZONE_DMA=n)
> 2. allocate X size memory in a high region
> 
> ",low" argument can be used to specify the low memory.
> 
> Do i understand correctly?

Yes, although we could follow the x86 approach:

1. Try low (ZONE_DMA for arm64) allocation, fallback to high allocation
   if it fails.

2. If crash_base is outside ZONE_DMA, call reserve_crashkernel_low()
   which either honours the ,low option or allocates some small amount
   in ZONE_DMA.

If at some point we have platforms failing step 2, we'll look at
changing ZONE_DMA to the full 4GB on non-RPi4 platforms.

It looks to me like x86 ignores the ,low option if the first step
managed to get some low memory. Shall we do the same on arm64?

> > So with an implicit ZONE_DMA allocation similar to the x86 one, we
> > probably don't need Bhupesh's series at all. In addition, we can limit
> > crashkernel= to the first 4G with a fall-back to high like x86 (not sure
> > if memblock_find_in_range() is guaranteed to search in ascending order).
> > I don't think we need an explicit ",high" annotation.
> >
> > So with the above, just a crashkernel=1G gives you at least 256MB in
> > ZONE_DMA followed by the rest anywhere, with a preference for
> > ZONE_DMA32. This way we can also keep the reserve_crashkernel_low()
> > mostly intact from x86 (less #ifdef's).
> 
> Yes. We can let crashkernel=X  try to reserve low memory and fall back to use high memory
> if failing to find a low range.

The only question is whether we need to preserve some more ZONE_DMA on
the current system. If for example we pass a crashkernel=512M and some
cma=, we may end up with very little free memory in ZONE_DMA. That's
mostly an issue for RPi4 since other platforms would work with
ZONE_DMA32. We could add a threshold and go for high allocation directly
if the required size is too large.

> About the function reserve_crashkernel_low(), if we put it in arch/arm64, there is some common
> code with x86_64. Some suggestions about this?

If we can use this function almost intact, just move it in a common
place. But if it gets sprinkled with #ifdef CONFIG_ARM64, I'd rather
duplicate it. I'd still prefer to move it to a common place if possible.

You can go a step further and also move the x86 reserve_crashkernel() to
common code. I don't think there a significant difference between arm64
and x86 here. You'd have to define arch-specific specific
CRASH_ADDR_LOW_MAX etc.

Also patches moving code should not have any functional change. The
CRASH_ALIGN change from 16M to 2M on x86 should be a separate patch as
it needs to be acked by the x86 maintainers (IIRC, Ingo only acked the
function move if there was no functional change; CRASH_ALIGN is used for
the start address, not just alignment, on x86).

-- 
Catalin

  reply index

Thread overview: 18+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-07-03  3:58 [PATCH v10 0/5] support reserving crashkernel above 4G on arm64 kdump Chen Zhou
2020-07-03  3:58 ` [PATCH v10 1/5] x86: kdump: move reserve_crashkernel_low() into crash_core.c Chen Zhou
2020-07-03  3:58 ` [PATCH v10 2/5] arm64: kdump: reserve crashkenel above 4G for crash dump kernel Chen Zhou
2020-07-03  3:58 ` [PATCH v10 3/5] arm64: kdump: add memory for devices by DT property linux,usable-memory-range Chen Zhou
2020-07-03  3:58 ` [PATCH v10 4/5] arm64: kdump: fix kdump broken with ZONE_DMA reintroduced Chen Zhou
2020-07-27 17:30   ` Catalin Marinas
2020-07-29  3:52     ` chenzhou
2020-07-29 11:58       ` Catalin Marinas
2020-07-29 14:14         ` chenzhou
2020-07-29 15:20           ` Catalin Marinas [this message]
2020-07-30  8:22             ` chenzhou
2020-07-03  3:58 ` [PATCH v10 5/5] kdump: update Documentation about crashkernel on arm64 Chen Zhou
2020-07-03  4:46   ` Dave Young
2020-07-03  4:50     ` Dave Young
2020-07-03  9:11       ` Dave Young
2020-07-03  7:26 ` [PATCH v10 0/5] support reserving crashkernel above 4G on arm64 kdump Bhupesh Sharma
2020-07-03  8:38   ` chenzhou
2020-07-27 12:38     ` John Donnelly

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20200729152028.GE5524@gaia \
    --to=catalin.marinas@arm.com \
    --cc=John.P.donnelly@oracle.com \
    --cc=arnd@arndb.de \
    --cc=bhe@redhat.com \
    --cc=bhsharma@redhat.com \
    --cc=chenzhou10@huawei.com \
    --cc=corbet@lwn.net \
    --cc=dyoung@redhat.com \
    --cc=guohanjun@huawei.com \
    --cc=horms@verge.net.au \
    --cc=huawei.libin@huawei.com \
    --cc=james.morse@arm.com \
    --cc=kexec@lists.infradead.org \
    --cc=linux-arm-kernel@lists.infradead.org \
    --cc=linux-doc@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mingo@redhat.com \
    --cc=nsaenzjulienne@suse.de \
    --cc=prabhakar.pkin@gmail.com \
    --cc=robh+dt@kernel.org \
    --cc=tglx@linutronix.de \
    --cc=will@kernel.org \
    --cc=xiexiuqi@huawei.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

Linux-Doc Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/linux-doc/0 linux-doc/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 linux-doc linux-doc/ https://lore.kernel.org/linux-doc \
		linux-doc@vger.kernel.org
	public-inbox-index linux-doc

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.kernel.vger.linux-doc


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git