Re: [PATCH v2 03/11] mm/damon/paddr: support addr_unit for DAMOS_PAGEOUT
From: Quanmin Yan
Date: Wed Aug 27 2025 - 07:37:54 EST
Hi SJ,
在 2025/8/27 10:42, SeongJae Park 写道:
On Wed, 27 Aug 2025 10:21:41 +0800 Quanmin Yan <yanquanmin1@xxxxxxxxxx> wrote:
在 2025/8/26 22:21, SeongJae Park 写道:
On Tue, 26 Aug 2025 12:51:17 +0800 Quanmin Yan <yanquanmin1@xxxxxxxxxx> wrote:
Hi SJ,
在 2025/8/26 11:21, SeongJae Park 写道:
[...]
==== Attachment 0 (0001-mm-damon-paddr-use-do_div-on-i386-for-damon_pa_pageo.patch) ====
From hackermail Thu Jan 1 00:00:00 1970
From: SeongJae Park <sj@xxxxxxxxxx>
To: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
Cc: SeongJae Park <sj@xxxxxxxxxx>
Cc: damon@xxxxxxxxxxxxxxx
Cc: kernel-team@xxxxxxxx
Cc: linux-kernel@xxxxxxxxxxxxxxx
Cc: linux-mm@xxxxxxxxx
Date: Mon, 25 Aug 2025 07:41:33 -0700
Subject: [PATCH 1/3] mm/damon/paddr: use do_div() on i386 for damon_pa_pageout()
return value
Otherwise, __udidi3 linking problem happens on certain configs.
Reported-by: kernel test robot <lkp@xxxxxxxxx>
Closes: https://lore.kernel.org/oe-kbuild-all/202508241831.EKwdwXZL-lkp@xxxxxxxxx/
Signed-off-by: SeongJae Park <sj@xxxxxxxxxx>
---
mm/damon/paddr.c | 14 +++++++++++++-
1 file changed, 13 insertions(+), 1 deletion(-)
diff --git a/mm/damon/paddr.c b/mm/damon/paddr.c
index 5fad2f9a99a0..09c87583af6c 100644
--- a/mm/damon/paddr.c
+++ b/mm/damon/paddr.c
@@ -135,6 +135,18 @@ static bool damon_pa_invalid_damos_folio(struct folio *folio, struct damos *s)
return false;
}
+/* convert physical address to core-layer address */
+static unsigned long damon_pa_core_addr(phys_addr_t pa,
+ unsigned long addr_unit)
+{
+#ifdef __i386__
Can we use the following condition instead?
#if !defined(CONFIG_64BIT) && defined(CONFIG_PHYS_ADDR_T_64BIT)
To my understanding, this issue happens only on i386, not every 32bit
architectures. So I think i386 specific condition is better.
I understand. However, the aforementioned general condition is essential,
and we should propose a new patch to address this. After introducing addr_unit,
any 32-bit architecture should support monitoring of 64-bit phys_addr_t.
The issue is that we cannot divide 64bit values with plain '/' operator on
"i386", and hence this patch makes it to use do_div() instead of '/' on "i386".
No such or other problems at supporting 64-bit phys_addr_t is found on other
32bit architectures, to my understanding. My understanding is that at least
you confirmed that on your arm-based test environment. So we don't need a new
patch to my understanding.
Am I missing somthing?
This is because I seem to have made a mistake earlier: I adjusted the local
compilation toolchain. When the __udivdi3 issue mentioned above occurred, it
reminded me of a potential problem. After switching to a completely new environment
for testing, I discovered the __aeabi_uldivmod linking error in arm, which is similar
to the __udivdi3 issue.🙁
Thank you for sharing this. Then I agree the current fixup is insufficient.
Andrew, could you please drop this patch series from mm tree for now? I will
further discuss with Quanmin about the proper fix and post next version of this
series with the proper fixup.
To prevent similar environment-related problems in the
future, I suggest adjusting the condition to the following:
#if !defined(CONFIG_64BIT) && defined(CONFIG_PHYS_ADDR_T_64BIT)
Please consider approving this fix.
I'm yet in travel, and I'd like to take more time on thinking about the proper
fix. Quanmin, could you share more details about your test setups, both for
the compiling success case and failing case?
Apologies for disturbing you during your travels. Please allow me to explain:
When CONFIG_PHYS_ADDR_T_64BIT is enabled on "i386" [1], the phys_addr_t type
becomes 64-bit, requiring the use of the do_div function. We are in agreement
on this point.
On arm32, if LPAE (which we intend to support in this series) is enabled,
CONFIG_PHYS_ADDR_T_64BIT will also be enabled. In this case, pa / addr_unit
will involve 64-bit division and similarly require the do_div function.
Obviously, such link errors should normally occur under these circumstances.
Unfortunately, the expected anomalies did not manifest in my previous tests.
This may be related to some incorrect adjustments I had made to my local build
environment quite some time ago — though I cannot be entirely certain. That
said, I have since cleaned up the old configurations and ensured the current
environment is clean and normal. For now, we have confirmed the actual problem
and its root cause, shall we focus on fixing it?
In summary, after introducing addr_unit, we expect that any 32-bit architecture
should support monitoring of 64-bit phys_addr_t. Therefore, we can consider the
following adjustment:
#if !defined(CONFIG_64BIT) && defined(CONFIG_PHYS_ADDR_T_64BIT)
Or at least adjust it to:
#if defined(__i386__) || (defined(__arm__) && defined(CONFIG_PHYS_ADDR_T_64BIT))
I have thoroughly re-validated the feature functionality today and confirmed the
correctness of the aforementioned modifications. Therefore, could I kindly ask
you to consider the aforementioned modifications when you have some free time?
[1]https://download.01.org/0day-ci/archive/20250824/202508241831.EKwdwXZL-lkp@xxxxxxxxx/config
Thanks,
Quanmin Yan
[...]