Skip to contentSkip to conversation

mm/migrate_device: fix double unlock and remove dead code

PATCH In review Sunny Patel 4 participants 8 messages Apr 13 – Apr 14

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 mm/migrate_device: fix double unlock and remove dead code

Fix two bugs in device migration paths:

1) migrate_vma_collect_huge_pmd() calls spin_unlock after softleaf_entry_wait_on_locked(), which already releases the ptl.

2) migrate_vma_insert_huge_pmd_page() has a dead else-if branch and this branch is always unreachable.

---
 mm/migrate_device.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
mm/migrate_device.c +1 -3
index 8079676c8f1f..0e005c26ee88 100644
@@ -177,7 +177,6 @@ static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start,
 
 		if (softleaf_is_migration(entry)) {
 			softleaf_entry_wait_on_locked(entry, ptl);
-			spin_unlock(ptl);
 			return -EAGAIN;
 		}
 
@@ -869,8 +868,7 @@ static int migrate_vma_insert_huge_pmd_page(struct migrate_vma *migrate,
 		if (!is_huge_zero_pmd(*pmdp))
 			goto unlock_abort;
 		flush = true;
-	} else if (!pmd_none(*pmdp))
-		goto unlock_abort;
+	}
 
 	add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 	folio_add_new_anon_rmap(folio, vma, addr, RMAP_EXCLUSIVE);
Signature · 1 line
2.43.0
#1 lore raw
Quoted 8 lines · Sunny Patel

Fix two bugs in device migration paths:

1) migrate_vma_collect_huge_pmd() calls spin_unlock after softleaf_entry_wait_on_locked(), which already releases the ptl.

2) migrate_vma_insert_huge_pmd_page() has a dead else-if branch and this branch is always unreachable.

Can you move 1) into a separate patch, add a Fixes: tag an CC stable?

I think it is

Fixes a30b48bf1b24 ("mm/migrate_device: implement THP migration of zone

device pages")

2) will then be a pure cleanup patch.

Thanks!

Quoted 10 lines
---
 mm/migrate_device.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
mm/migrate_device.c +0 -0
index 8079676c8f1f..0e005c26ee88 100644

mm/migrate_device.c static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start, · line 180

Quoted 4 lines
mm/migrate_device.c +0 -1
@@ -177,7 +177,6 @@ static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start,
 
 		if (softleaf_is_migration(entry)) {
 			softleaf_entry_wait_on_locked(entry, ptl);
-			spin_unlock(ptl);

Yes, that looks correct to me.

mm/migrate_device.c static int migrate_vma_insert_huge_pmd_page(struct migrate_vma *migrate, · line 871

Quoted 10 lines
mm/migrate_device.c +1 -2
 			return -EAGAIN;
 		}
 
@@ -869,8 +868,7 @@ static int migrate_vma_insert_huge_pmd_page(struct migrate_vma *migrate,
 		if (!is_huge_zero_pmd(*pmdp))
 			goto unlock_abort;
 		flush = true;
-	} else if (!pmd_none(*pmdp))
-		goto unlock_abort;
+	}

Huh, how did that happen. I hope that it's not a typo and we wanted to check for something else.

Signature · 3 lines
Cheers,

David
reviewer
#2 lore raw
Quoted 42 lines · David Hildenbrand (Arm)
Quoted 8 lines · Sunny Patel

Fix two bugs in device migration paths:

1) migrate_vma_collect_huge_pmd() calls spin_unlock after softleaf_entry_wait_on_locked(), which already releases the ptl.

2) migrate_vma_insert_huge_pmd_page() has a dead else-if branch and this branch is always unreachable.

Can you move 1) into a separate patch, add a Fixes: tag an CC stable?

I think it is

Fixes a30b48bf1b24 ("mm/migrate_device: implement THP migration of zone

device pages")

2) will then be a pure cleanup patch.

Thanks!

Quoted 10 lines
---
 mm/migrate_device.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
mm/migrate_device.c +0 -0
index 8079676c8f1f..0e005c26ee88 100644

mm/migrate_device.c static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start, · line 180

Quoted 4 lines
mm/migrate_device.c +0 -1
@@ -177,7 +177,6 @@ static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start,
 
 		if (softleaf_is_migration(entry)) {
 			softleaf_entry_wait_on_locked(entry, ptl);
-			spin_unlock(ptl);

Yes, that looks correct to me.

Quoted 3 lines
mm/migrate_device.c +0 -0
 			return -EAGAIN;
 		}
 
Quoted 11 lines

mm/migrate_device.c static int migrate_vma_insert_huge_pmd_page(struct migrate_vma *migrate, · line 871

Quoted 7 lines
mm/migrate_device.c +1 -2
@@ -869,8 +868,7 @@ static int migrate_vma_insert_huge_pmd_page(struct migrate_vma *migrate,
 		if (!is_huge_zero_pmd(*pmdp))
 			goto unlock_abort;
 		flush = true;
-	} else if (!pmd_none(*pmdp))
-		goto unlock_abort;
+	}

Huh, how did that happen. I hope that it's not a typo and we wanted to check for something else.

Looking at the function and trying to figure this out, but find VM_WARN_ON_FOLIO(!folio, folio) at the top, where folio is from page_folio(page). It is either a nop or a zero dereferencing. It should be removed.

VM_WARN_ON_ONCE(!pmd_none(*pmdp) && !is_huge_zero_pmd(*pmdp)) can be removed too, since the above ifs takes !pmd_none(*pmdp) && !is_huge_zero_pmd(*pmdp) to unlock_abort.

Back to the above code:

        if (!pmd_none(*pmdp)) {
                if (!is_huge_zero_pmd(*pmdp))
                        goto unlock_abort;
                flush = true;
        } else if (!pmd_none(*pmdp))
                goto unlock_abort;

It seems to me that the first if should be removed, since if pmdp is not pmd_none(), others filled pmd entry before us, so no further action should be taken, otherwise, the function will overwrite some valid pmd entry.

OK, look at migrate_vma_insert_page(), which does PTE level work, the above code might be intended to do:

        if (pmd_present(*pmdp)) {
                if (!is_huge_zero_pmd(*pmdp))
                        goto unlock_abort;
                flush = true;
        } else if (!pmd_none(*pmdp))
                goto unlock_abort;

Best Regards,
Yan, Zi

#3 lore raw
Quoted 57 lines · Zi Yan
Quoted 42 lines · David Hildenbrand (Arm)
Quoted 8 lines · Sunny Patel

Fix two bugs in device migration paths:

1) migrate_vma_collect_huge_pmd() calls spin_unlock after softleaf_entry_wait_on_locked(), which already releases the ptl.

2) migrate_vma_insert_huge_pmd_page() has a dead else-if branch and this branch is always unreachable.

Can you move 1) into a separate patch, add a Fixes: tag an CC stable?

I think it is

Fixes a30b48bf1b24 ("mm/migrate_device: implement THP migration of zone

device pages")

2) will then be a pure cleanup patch.

Thanks!

Quoted 10 lines
---
 mm/migrate_device.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
mm/migrate_device.c +0 -0
index 8079676c8f1f..0e005c26ee88 100644

mm/migrate_device.c static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start, · line 180

Quoted 4 lines
mm/migrate_device.c +0 -1
@@ -177,7 +177,6 @@ static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start,
 
 		if (softleaf_is_migration(entry)) {
 			softleaf_entry_wait_on_locked(entry, ptl);
-			spin_unlock(ptl);

Yes, that looks correct to me.

Quoted 3 lines
mm/migrate_device.c +0 -0
 			return -EAGAIN;
 		}
 
Quoted 11 lines

mm/migrate_device.c static int migrate_vma_insert_huge_pmd_page(struct migrate_vma *migrate, · line 871

Quoted 7 lines
mm/migrate_device.c +1 -2
@@ -869,8 +868,7 @@ static int migrate_vma_insert_huge_pmd_page(struct migrate_vma *migrate,
 		if (!is_huge_zero_pmd(*pmdp))
 			goto unlock_abort;
 		flush = true;
-	} else if (!pmd_none(*pmdp))
-		goto unlock_abort;
+	}

Huh, how did that happen. I hope that it's not a typo and we wanted to check for something else.

Quoted 3 lines

Looking at the function and trying to figure this out, but find VM_WARN_ON_FOLIO(!folio, folio) at the top, where folio is from page_folio(page). It is either a nop or a zero dereferencing. It should be removed.

Heh, VM_WARN_ON_FOLIO(!folio, folio) itself is completely odd. Dump NULL if NULL.

Agreed that this should be removed. Or replaced by a

VM_WARN_ON_ONCE(page);

Quoted 4 lines

VM_WARN_ON_ONCE(!pmd_none(*pmdp) && !is_huge_zero_pmd(*pmdp)) can be removed too, since the above ifs takes !pmd_none(*pmdp) && !is_huge_zero_pmd(*pmdp) to unlock_abort.

Right, and if that's an unexpected case, we should warn on that path instead. Like

	if (WARN_ON_ONCE(!is_huge_zero_pmd(*pmdp)))
		goto unlock_abort;
Quoted 18 lines

Back to the above code:

        if (!pmd_none(*pmdp)) {
                if (!is_huge_zero_pmd(*pmdp))
                        goto unlock_abort;
                flush = true;
        } else if (!pmd_none(*pmdp))
                goto unlock_abort;

It seems to me that the first if should be removed, since if pmdp is not pmd_none(), others filled pmd entry before us, so no further action should be taken, otherwise, the function will overwrite some valid pmd entry.

OK, look at migrate_vma_insert_page(), which does PTE level work, the above code might be intended to do:

Quoted 6 lines
        if (pmd_present(*pmdp)) {
                if (!is_huge_zero_pmd(*pmdp))
                        goto unlock_abort;
                flush = true;
        } else if (!pmd_none(*pmdp))
                goto unlock_abort;

Agreed, that makes more sense. Thanks!

Signature · 3 lines
Cheers,

David
reviewer
#6 lore raw
In reply to Sunny Patel
Quoted 18 lines · Sunny Patel

Fix two bugs in device migration paths:

1) migrate_vma_collect_huge_pmd() calls spin_unlock after softleaf_entry_wait_on_locked(), which already releases the ptl.

2) migrate_vma_insert_huge_pmd_page() has a dead else-if branch and this branch is always unreachable.

---
 mm/migrate_device.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
mm/migrate_device.c +0 -0
index 8079676c8f1f..0e005c26ee88 100644

mm/migrate_device.c static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start, · line 181

Quoted 6 lines
mm/migrate_device.c +0 -1
@@ -177,7 +177,6 @@ static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start,
 
 		if (softleaf_is_migration(entry)) {
 			softleaf_entry_wait_on_locked(entry, ptl);
-			spin_unlock(ptl);
 			return -EAGAIN;
 		}

This entire if statement is dead code, since we bail out and collect a skip on !softleaf_is_device_private immediately above. I thought we already had a patch from Davidlohr Bueso 'mm/migrate_device: remove dead migration entry check in migrate_vma_collect_huge_pmd()' that removed this whole if statement. I even have an email from Andrew pulling this patch, but it looks like the code is still present in 7.0.

I found this link via a google search which contains the above patch too:

https://kernel.googlesource.com/pub/scm/linux/kernel/git/akpm/25-new/+/f405e7bc5d93a541cf11663b2424baf0a03b496f%5E%21/

Look like this change got lost somehow?

Matt

Quoted 14 lines
mm/migrate_device.c +1 -2
@@ -869,8 +868,7 @@ static int migrate_vma_insert_huge_pmd_page(struct migrate_vma *migrate,
 		if (!is_huge_zero_pmd(*pmdp))
 			goto unlock_abort;
 		flush = true;
-	} else if (!pmd_none(*pmdp))
-		goto unlock_abort;
+	}
 
 	add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 	folio_add_new_anon_rmap(folio, vma, addr, RMAP_EXCLUSIVE);
Signature · 1 line
2.43.0
#7 lore raw
Quoted 33 lines · Matthew Brost
Quoted 18 lines · Sunny Patel

Fix two bugs in device migration paths:

1) migrate_vma_collect_huge_pmd() calls spin_unlock after softleaf_entry_wait_on_locked(), which already releases the ptl.

2) migrate_vma_insert_huge_pmd_page() has a dead else-if branch and this branch is always unreachable.

---
 mm/migrate_device.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
mm/migrate_device.c +0 -0
index 8079676c8f1f..0e005c26ee88 100644

mm/migrate_device.c static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start, · line 181

Quoted 6 lines
mm/migrate_device.c +0 -1
@@ -177,7 +177,6 @@ static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start,
 
 		if (softleaf_is_migration(entry)) {
 			softleaf_entry_wait_on_locked(entry, ptl);
-			spin_unlock(ptl);
 			return -EAGAIN;
 		}

This entire if statement is dead code, since we bail out and collect a skip on !softleaf_is_device_private immediately above. I thought we already had a patch from Davidlohr Bueso 'mm/migrate_device: remove dead migration entry check in migrate_vma_collect_huge_pmd()' that removed this whole if statement. I even have an email from Andrew pulling this patch, but it looks like the code is still present in 7.0.

Quoted 6 lines

I found this link via a google search which contains the above patch too:

https://kernel.googlesource.com/pub/scm/linux/kernel/git/akpm/25-new/+/f405e7bc5d93a541cf11663b2424baf0a03b496f%5E%21/

Look like this change got lost somehow?

Oh, even better, then we don't even need a Fixes:

Signature · 3 lines
Cheers,

David