Skip to contentSkip to conversation

mm: shmem: don't truncate page if memory failure happens

PATCH Superseded Yang Shi 2 participants 4 messages Nov 14, 2021 – Nov 15, 2021

Subsystems (4)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Yang Shi

author
#0 lore raw

PATCH 1/1 mm: shmem: don't truncate page if memory failure happens

The current behavior of memory failure is to truncate the page cache regardless of dirty or clean. If the page is dirty the later access will get the obsolete data from disk without any notification to the users. This may cause silent data loss. It is even worse for shmem since shmem is in-memory filesystem, truncating page cache means discarding data blocks. The later read would return all zero.

The right approach is to keep the corrupted page in page cache, any later access would return error for syscalls or SIGBUS for page fault, until the file is truncated, hole punched or removed. The regular storage backed filesystems would be more complicated so this patch is focused on shmem. This also unblock the support for soft offlining shmem THP.

[[email protected]: fix uninitialized variable use in me_pagecache_clean()]

[Fix invalid pointer dereference in shmem_read_mapping_page_gfp() with a
slight different implementation from what Ajay Garg <[email protected]> and Muchun Song <[email protected]> proposed and reworked the error handling of shmem_write_begin() suggested by Linus]

Cc Hugh Dickins <[email protected]>
Cc Kirill A. Shutemov <[email protected]>
Cc Matthew Wilcox <[email protected]>
Cc Naoya Horiguchi <[email protected]>
Cc Oscar Salvador <[email protected]>
Cc Peter Xu <[email protected]>
Cc Ajay Garg <[email protected]>
Cc Muchun Song <[email protected]>

---
Fixed pointer dereference oops reported by Ajay Garg and reworked the error handling of shmem_write_begin() suggested by Linus.

 mm/memory-failure.c | 14 ++++++++++---
 mm/shmem.c          | 48 +++++++++++++++++++++++++++++++++++++++------
 mm/userfaultfd.c    |  5 +++++
 3 files changed, 58 insertions(+), 9 deletions(-)
mm/memory-failure.c +11 -3
index 07c875fdeaf0..f64ebb6226cb 100644
@@ -58,6 +58,7 @@
 #include <linux/ratelimit.h>
 #include <linux/page-isolation.h>
 #include <linux/pagewalk.h>
+#include <linux/shmem_fs.h>
 #include "internal.h"
 #include "ras/ras_event.h"
 
@@ -867,6 +868,7 @@ static int me_pagecache_clean(struct page_state *ps, struct page *p)
 {
 	int ret;
 	struct address_space *mapping;
+	bool extra_pins;
 
 	delete_from_lru_cache(p);
 
@@ -895,18 +897,24 @@ static int me_pagecache_clean(struct page_state *ps, struct page *p)
 		goto out;
 	}
 
+	/*
+	 * The shmem page is kept in page cache instead of truncating
+	 * so is expected to have an extra refcount after error-handling.
+	 */
+	extra_pins = shmem_mapping(mapping);
+
 	/*
 	 * Truncation is a bit tricky. Enable it per file system for now.
 	 *
 	 * Open: to take i_rwsem or not for this? Right now we don't.
 	 */
 	ret = truncate_error_page(p, page_to_pfn(p), mapping);
+	if (has_extra_refcount(ps, p, extra_pins))
+		ret = MF_FAILED;
+
 out:
 	unlock_page(p);
 
-	if (has_extra_refcount(ps, p, false))
-		ret = MF_FAILED;
-
 	return ret;
 }
 
mm/shmem.c +42 -6

96 lines not on this page · raw message

mm/userfaultfd.c +5 -0
index ac6f036298cd..0780c2a57ff1 100644
@@ -232,6 +232,11 @@ static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
 		goto out;
 	}
 
+	if (PageHWPoison(page)) {
+		ret = -EIO;
+		goto out_release;
+	}
+
 	ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 				       page, false, wp_copy);
 	if (ret)
Signature · 1 line
2.26.2
#1 lore raw
Quoted 4 lines · Yang Shi
mm/shmem.c +0 -0
@@ -2466,7 +2467,18 @@ shmem_write_begin(struct file *file, struct address_space *mapping,
 			return -EPERM;
 	}
 

mm/shmem.c shmem_write_begin(struct file *file, struct address_space *mapping, · line 2478

Quoted 10 lines
mm/shmem.c +9 -1
-	return shmem_getpage(inode, index, pagep, SGP_WRITE);
+	ret = shmem_getpage(inode, index, pagep, SGP_WRITE);
+
+	if (ret)
+		return ret;
+
+	if (*pagep && PageHWPoison(*pagep)) {
+		unlock_page(*pagep);
+		put_page(*pagep);
+		ret = -EIO;

You definitely need to add:

		*pagep = NULL;

I'm not entirely convinced that you need the conditional on '*pagep'. If we returned 0, there had better be a page at pagep!

I also think this would be clearer if written as:

	if (PageHWPoison(*pagep)) {
		unlock_page(*pagep);
		put_page(*pagep);
		*pagep = NULL;
		return -EIO;
	}

	return 0;

instead of re-using ret. Sometimes that can make the code flow clearer, but here, I don't think it does.

mm/shmem.c struct page *shmem_read_mapping_page_gfp(struct address_space *mapping, · line 4208

Quoted 12 lines
mm/shmem.c +5 -3
@@ -4168,9 +4201,12 @@ struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
 	error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
 				  gfp, NULL, NULL, NULL);
 	if (error)
-		page = ERR_PTR(error);
-	else
-		unlock_page(page);
+		return ERR_PTR(error);
+
+	unlock_page(page);
+	if (PageHWPoison(page))
+		return ERR_PTR(-EIO);

Do we need to put_page() the page in this error case?

author
#2 lore raw
Quoted 19 lines · Matthew Wilcox
Quoted 4 lines · Yang Shi
mm/shmem.c +0 -0
@@ -2466,7 +2467,18 @@ shmem_write_begin(struct file *file, struct address_space *mapping,
                      return -EPERM;
      }
 

mm/shmem.c shmem_write_begin(struct file *file, struct address_space *mapping, · line 2478

Quoted 10 lines
mm/shmem.c +9 -1
-     return shmem_getpage(inode, index, pagep, SGP_WRITE);
+     ret = shmem_getpage(inode, index, pagep, SGP_WRITE);
+
+     if (ret)
+             return ret;
+
+     if (*pagep && PageHWPoison(*pagep)) {
+             unlock_page(*pagep);
+             put_page(*pagep);
+             ret = -EIO;

You definitely need to add:

Quoted 1 line
                *pagep = NULL;

Thanks, will do it.

Quoted 3 lines

I'm not entirely convinced that you need the conditional on '*pagep'. If we returned 0, there had better be a page at pagep!

For SGP_WRITE, yes, it has a page at pagep if ret is 0, but shmem_getpage() could return 0 with NULL page at pagep for SGP_READ.

In the other thread that Linus elaborated why this commit was reverted and needed some rework, the discussion about not relying on implementation detail for error handling taught me it may be not a robust implementation to assume it is never NULL.

We might refactor shmem_getpage() code in the future to make sure when it returns success there must be a valid pagep so that we just need to care about the return value for error handling.

Quoted 12 lines

I also think this would be clearer if written as:

        if (PageHWPoison(*pagep)) {
                unlock_page(*pagep);
                put_page(*pagep);
                *pagep = NULL;
                return -EIO;
        }

        return 0;
Quoted 2 lines

instead of re-using ret. Sometimes that can make the code flow clearer, but here, I don't think it does.

Sure.

Quoted 15 lines

mm/shmem.c struct page *shmem_read_mapping_page_gfp(struct address_space *mapping, · line 4208

Quoted 13 lines
mm/shmem.c +5 -3
@@ -4168,9 +4201,12 @@ struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
      error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
                                gfp, NULL, NULL, NULL);
      if (error)
-             page = ERR_PTR(error);
-     else
-             unlock_page(page);
+             return ERR_PTR(error);
+
+     unlock_page(page);
+     if (PageHWPoison(page))
+             return ERR_PTR(-EIO);

Do we need to put_page() the page in this error case?

Aha, yes. Sorry for missing this. I was fooled by shmem_pin_map() in i915 driver which does put page, but I realized it just puts the valid pages pinned *before* meeting error page by second look
.

Quoted 1 line
author
#3 lore raw
Quoted 32 lines · Yang Shi
Quoted 19 lines · Matthew Wilcox
Quoted 4 lines · Yang Shi
mm/shmem.c +0 -0
@@ -2466,7 +2467,18 @@ shmem_write_begin(struct file *file, struct address_space *mapping,
                      return -EPERM;
      }
 

mm/shmem.c shmem_write_begin(struct file *file, struct address_space *mapping, · line 2478

Quoted 10 lines
mm/shmem.c +9 -1
-     return shmem_getpage(inode, index, pagep, SGP_WRITE);
+     ret = shmem_getpage(inode, index, pagep, SGP_WRITE);
+
+     if (ret)
+             return ret;
+
+     if (*pagep && PageHWPoison(*pagep)) {
+             unlock_page(*pagep);
+             put_page(*pagep);
+             ret = -EIO;

You definitely need to add:

Quoted 1 line
                *pagep = NULL;

Thanks, will do it.

Quoted 3 lines

I'm not entirely convinced that you need the conditional on '*pagep'. If we returned 0, there had better be a page at pagep!

For SGP_WRITE, yes, it has a page at pagep if ret is 0, but shmem_getpage() could return 0 with NULL page at pagep for SGP_READ.

Quoted 20 lines

In the other thread that Linus elaborated why this commit was reverted and needed some rework, the discussion about not relying on implementation detail for error handling taught me it may be not a robust implementation to assume it is never NULL.

We might refactor shmem_getpage() code in the future to make sure when it returns success there must be a valid pagep so that we just need to care about the return value for error handling.

Quoted 11 lines

I also think this would be clearer if written as:

        if (PageHWPoison(*pagep)) {
                unlock_page(*pagep);
                put_page(*pagep);
                *pagep = NULL;
                return -EIO;
        }

        return 0;

By rethinking the code, I do prefer the above. "if (*pagep && PageHWPoison(*pagep))" does give extra protection from NULL pointer dereference for the future, but on the opposite side it seems confusing to *not* have error handling for NULL page. I bet it may incur more confusion than the protection for future, anyway it can't be NULL if ret == 0 and !SGP_READ.

shmem_read_mapping_page_gfp() doesn't check page pointer either. IIUC only SGP_READ case has NULL page pointer checked in shmem code, so I'd like to follow the convention in this patch. If we think it is not good, we could refactor the code (for example, guarantee page have valid page as long as ret value is not error) in a perarate patch.

Quoted 27 lines
Quoted 3 lines

instead of re-using ret. Sometimes that can make the code flow clearer, but here, I don't think it does.

Sure.

Quoted 15 lines

mm/shmem.c struct page *shmem_read_mapping_page_gfp(struct address_space *mapping, · line 4208

Quoted 13 lines
mm/shmem.c +5 -3
@@ -4168,9 +4201,12 @@ struct page *shmem_read_mapping_page_gfp(struct address_space *mapping,
      error = shmem_getpage_gfp(inode, index, &page, SGP_CACHE,
                                gfp, NULL, NULL, NULL);
      if (error)
-             page = ERR_PTR(error);
-     else
-             unlock_page(page);
+             return ERR_PTR(error);
+
+     unlock_page(page);
+     if (PageHWPoison(page))
+             return ERR_PTR(-EIO);

Do we need to put_page() the page in this error case?

Aha, yes. Sorry for missing this. I was fooled by shmem_pin_map() in i915 driver which does put page, but I realized it just puts the valid pages pinned *before* meeting error page by second look
.

Quoted 1 line