Date:   Fri, 21 Oct 2022 11:11:05 +0200
From:   David Hildenbrand <david@...hat.com>
To:     Peter Xu <peterx@...hat.com>
Cc:     linux-kernel@...r.kernel.org, linux-mm@...ck.org,
        linux-kselftest@...r.kernel.org,
        Andrew Morton <akpm@...ux-foundation.org>,
        Shuah Khan <shuah@...nel.org>, Hugh Dickins <hughd@...gle.com>,
        Vlastimil Babka <vbabka@...e.cz>,
        Andrea Arcangeli <aarcange@...hat.com>,
        "Matthew Wilcox (Oracle)" <willy@...radead.org>,
        Jason Gunthorpe <jgg@...dia.com>,
        John Hubbard <jhubbard@...dia.com>
Subject: Re: [PATCH v1 6/7] mm/ksm: convert break_ksm() to use
 walk_page_range_vma()

On 06.10.22 21:28, Peter Xu wrote:
> On Thu, Oct 06, 2022 at 11:20:42AM +0200, David Hildenbrand wrote:
>>>> +int break_ksm_pud_entry(pud_t *pud, unsigned long addr, unsigned long next,
>>>> +			struct mm_walk *walk)
>>>> +{
>>>> +	/* We only care about page tables to walk to a single base page. */
>>>> +	if (pud_leaf(*pud) || !pud_present(*pud))
>>>> +		return 1;
>>>> +	return 0;
>>>> +}
>>>
>>> Is this needed?  I thought the pgtable walker handlers this already.
>>>
>>> [...]
>>>
>>
>> Most probably yes. I was trying to avoid about PUD splits, but I guess we
>> simply should not care in VMAs that are considered by KSM (MERGABLE). Most
>> probably never ever happens.
> 
> I was surprised the split is the default approach; didn't really notice
> that before. Yeah maybe better to keep it.

Interestingly, one callback reduces the benchmark result by 100-200 MiB.

With only break_ksm_pmd_entry(), I get ~4900 MiB/s instead of ~5010 MiB/s (~2%).

I came to the conclusion that we really shouldn't have to worry about pud
THPs here: it could only be a file PUD and splitting that only zaps the
entry, but doesn't PMD- or PTE-map it.

Also, I think we will barely see large pud THP in a mergable mapping ... :)

[...]

>> My main motivation is to remove most of that GUP hackery here, which is
>> 1) Getting a reference on a page and waiting for migration to finish
>>     even though both is unnecessary.
>> 2) As we don't have sufficient control, we added FOLL_MIGRATION hacks to
>>     MM core to work around limitations in the GUP-based approacj.
> 
> I saw one thing of adding FOLL_MIGRATION from Hugh was to have a hint for
> follow page users:
> 
>    I'd have preferred to avoid another flag, and do it every time, in case
>    someone else makes the same easy mistake..
> 
> Though..

The important thing I think is that FOLL_MIGRATION really only applies to
follow_page(). In case of "modern" GUP we will just wait for migration
entries, handle swap entries ... when triggering a page fault.

> 
>> 3) We rely on legacy follow_page() interface that we should really get
>>     rid of in the long term.
> 
> ..this is part of effort to remove follow_page()?  More context will be
> helpful in that case.

The comment from Hugh is another example why follow_page() adds complexity.
One might wonder, how pages in the swapcache, device coherent pages, ...
would have to be handled.

Short-term, I want to cleanup GUP. Long-term we might want to consider
removing follow_page() completely.

[...]

>>
>> Yes, we have to extend page walking code, but it's just the natural,
>> non-hacky way of doing it.
>>
>> Regarding the 4% performance degradation (if I wouldn't have added the
>> benchmarks, nobody would know and probably care ;) ), I am not quite sure
>> why that is the case. We're just walking page tables after all in both
>> cases. Maybe the callback-based implementation of pagewalk code is less
>> efficient, but we might be able to improve that implementation if we really
>> care about performance here. Maybe removing break_ksm_pud_entry() already
>> improves the numbers slightly.
> 
> Yeah it could be the walker is just slower.  And for !ksm walking your code
> should be faster when hit migration entries, but that should really be rare
> anyway.


I have the following right now:


 From 7f767f9e9e673a29793cd35f1c3d66ed593b67cb Mon Sep 17 00:00:00 2001
From: David Hildenbrand <david@...hat.com>
Date: Mon, 25 Jul 2022 10:36:20 +0200
Subject: [PATCH] mm/ksm: convert break_ksm() to use walk_page_range_vma()

FOLL_MIGRATION exists only for the purpose of break_ksm(), and
actually, there is not even the need to wait for the migration to
finish, we only want to know if we're dealing with a KSM page.

Using follow_page() just to identify a KSM page overcomplicates GUP
code. Let's use walk_page_range_vma() instead, because we don't actually
care about the page itself, we only need to know a single property --
no need to even grab a reference.

So, get rid of follow_page() usage such that we can get rid of
FOLL_MIGRATION now and eventually be able to get rid of follow_page() in
the future.

In my setup (AMD Ryzen 9 3900X), running the KSM selftest to test unmerge
performance on 2 GiB (taskset 0x8 ./ksm_tests -D -s 2048), this results in
a performance degradation of ~2% (old: ~5010 MiB/s, new: ~4900 MiB/s).
I don't think we particularly care for now.

Interestingly, the benchmark reduction is due to the single callback.
Adding a second callback (e.g., pud_entry()) reduces the benchmark by
another 100-200 MiB/s.

Signed-off-by: David Hildenbrand <david@...hat.com>
---
  mm/ksm.c | 49 +++++++++++++++++++++++++++++++++++++++----------
  1 file changed, 39 insertions(+), 10 deletions(-)

diff --git a/mm/ksm.c b/mm/ksm.c
index c6f58aa6e731..5cdb852ff132 100644
--- a/mm/ksm.c
+++ b/mm/ksm.c
@@ -39,6 +39,7 @@
  #include <linux/freezer.h>
  #include <linux/oom.h>
  #include <linux/numa.h>
+#include <linux/pagewalk.h>
  
  #include <asm/tlbflush.h>
  #include "internal.h"
@@ -419,6 +420,39 @@ static inline bool ksm_test_exit(struct mm_struct *mm)
  	return atomic_read(&mm->mm_users) == 0;
  }
  
+static int break_ksm_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long next,
+			struct mm_walk *walk)
+{
+	struct page *page = NULL;
+	spinlock_t *ptl;
+	pte_t *pte;
+	int ret;
+
+	if (pmd_leaf(*pmd) || !pmd_present(*pmd))
+		return 0;
+
+	pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
+	if (pte_present(*pte)) {
+		page = vm_normal_page(walk->vma, addr, *pte);
+	} else if (!pte_none(*pte)) {
+		swp_entry_t entry = pte_to_swp_entry(*pte);
+
+		/*
+		 * As KSM pages remain KSM pages until freed, no need to wait
+		 * here for migration to end.
+		 */
+		if (is_migration_entry(entry))
+			page = pfn_swap_entry_to_page(entry);
+	}
+	ret = page && PageKsm(page);
+	pte_unmap_unlock(pte, ptl);
+	return ret;
+}
+
+static const struct mm_walk_ops break_ksm_ops = {
+	.pmd_entry = break_ksm_pmd_entry,
+};
+
  /*
   * We use break_ksm to break COW on a ksm page by triggering unsharing,
   * such that the ksm page will get replaced by an exclusive anonymous page.
@@ -434,21 +468,16 @@ static inline bool ksm_test_exit(struct mm_struct *mm)
   */
  static int break_ksm(struct vm_area_struct *vma, unsigned long addr)
  {
-	struct page *page;
  	vm_fault_t ret = 0;
  
  	do {
-		bool ksm_page = false;
+		int ksm_page;
  
  		cond_resched();
-		page = follow_page(vma, addr,
-				FOLL_GET | FOLL_MIGRATION | FOLL_REMOTE);
-		if (IS_ERR_OR_NULL(page))
-			break;
-		if (PageKsm(page))
-			ksm_page = true;
-		put_page(page);
-
+		ksm_page = walk_page_range_vma(vma, addr, addr + 1,
+					       &break_ksm_ops, NULL);
+		if (WARN_ON_ONCE(ksm_page < 0))
+			return ksm_page;
  		if (!ksm_page)
  			return 0;
  		ret = handle_mm_fault(vma, addr,
-- 
2.37.3


-- 
Thanks,

David / dhildenb

Powered by blists - more mailing lists