Skip to contentSkip to conversation

fs/btrfs: Make extent item iteration to handle gaps

PATCH Superseded Qu Wenruo 5 participants 6 messages Oct 16, 2021 – Nov 2, 2021 Btrfs file system

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Qu Wenruo

author
#0 lore raw

PATCH 1/1 fs/btrfs: Make extent item iteration to handle gaps

[BUG]
Grub btrfs implementation can't handle two very basic btrfs file layouts:

1. Mixed inline/regualr extents

   # mkfs.btrfs -f test.img
   # mount test.img /mnt/btrfs
   # xfs_io -f -c "pwrite 0 1k" -c "sync" -c "falloc 0 4k" \
	       -c "pwrite 4k 4k" /mnt/btrfs/file
   # umount /mnt/btrfs
   # ./grub-fstest ./grub-fstest --debug=btrfs ~/test.img hex "/file"

Such mixed inline/regular extents case is not recommended layout, but all existing tools and kernel can handle it without problem

2. NO_HOLES feature

   # mkfs.btrfs -f test.img -O no_holes
   # mount test.img /mnt/btrfs
   # xfs_io -f -c "pwrite 0 4k" -c "pwrite 8k 4k" /mnt/btrfs/file
   # umount /mnt/btrfs
   # ./grub-fstest ./grub-fstest --debug=btrfs ~/test.img hex "/file"

NO_HOLES feature is going to be the default mkfs feature in the incoming v5.15 release, and kernel has support for it since v4.0.

[CAUSE]
The way GRUB btrfs code iterates through file extents relies on no gap between extents.

If any gap is hit, then grub btrfs will error out, without any proper reason to help debug the bug.

This is a bad assumption, since a long long time ago btrfs has a new feature called NO_HOLES to allow btrfs to skip the padding hole extent to reduce metadata usage.

The NO_HOLES feature is already stable since kernel v4.0 and is going to be the default mkfs feature in the incoming v5.15 btrfs-progs release.

[FIX]
When there is a extent gap, instead of error out, just try next item.

This is still not ideal, as kernel/progs/U-boot all do the iteration item by item, not relying on the file offset continuity.

But it will be way more time consuming to correct the whole behavior than starting from scratch to build a proper designed btrfs module for GRUB.

---
 grub-core/fs/btrfs.c | 33 ++++++++++++++++++++++++++++++---
 1 file changed, 30 insertions(+), 3 deletions(-)
grub-core/fs/btrfs.c +30 -3
index 63203034dfc6..4fbcbec7524a 100644
@@ -1443,6 +1443,7 @@ grub_btrfs_extent_read (struct grub_btrfs_data *data,
       grub_size_t csize;
       grub_err_t err;
       grub_off_t extoff;
+      struct grub_btrfs_leaf_descriptor desc;
       if (!data->extent || data->extstart > pos || data->extino != ino
 	  || data->exttree != tree || data->extend <= pos)
 	{
@@ -1455,7 +1456,7 @@ grub_btrfs_extent_read (struct grub_btrfs_data *data,
 	  key_in.type = GRUB_BTRFS_ITEM_TYPE_EXTENT_ITEM;
 	  key_in.offset = grub_cpu_to_le64 (pos);
 	  err = lower_bound (data, &key_in, &key_out, tree,
-			     &elemaddr, &elemsize, NULL, 0);
+			     &elemaddr, &elemsize, &desc, 0);
 	  if (err)
 	    return -1;
 	  if (key_out.object_id != ino
@@ -1494,10 +1495,36 @@ grub_btrfs_extent_read (struct grub_btrfs_data *data,
 			PRIxGRUB_UINT64_T "\n",
 			grub_le_to_cpu64 (key_out.offset),
 			grub_le_to_cpu64 (data->extent->size));
+	  /*
+	   * The way of extent item iteration is pretty bad, it completely
+	   * requires all extents are contiguous, which is not ensured.
+	   *
+	   * Features like NO_HOLE and mixed inline/regular extents can cause
+	   * gaps between file extent items.
+	   *
+	   * The correct way is to follow kernel/U-boot to iterate item by
+	   * item, without any assumption on the file offset continuity.
+	   *
+	   * Here we just manually skip to next item and re-do the verification.
+	   *
+	   * TODO: Rework the whole extent item iteration code, if not the
+	   * whole btrfs implementation.
+	   */
 	  if (data->extend <= pos)
 	    {
-	      grub_error (GRUB_ERR_BAD_FS, "extent not found");
-	      return -1;
+	      err = next(data, &desc, &elemaddr, &elemsize, &key_out);
+	      if (err < 0)
+		return -1;
+	      /* No next item for the inode, we hit the end */
+	      if (err == 0 || key_out.object_id != ino ||
+		  key_out.type != GRUB_BTRFS_ITEM_TYPE_EXTENT_ITEM)
+		      return pos - pos0;
+
+	      csize = grub_le_to_cpu64(key_out.offset) - pos;
+	      buf += csize;
+	      pos += csize;
+	      len -= csize;
+	      continue;
 	    }
 	}
       csize = data->extend - pos;
Signature · 1 line
2.33.0
author
#1 lore raw

Gentle ping?

Without this patch, the new mkfs.btrfs NO_HOLES feature would break any kernel/initramfs with hole in it.

And considering the modification is already small, I believe this patch is definitely worthy as a bug fix.

Thanks,
Qu

Quoted 114 lines · Qu Wenruo

[BUG]
Grub btrfs implementation can't handle two very basic btrfs file layouts:

1. Mixed inline/regualr extents

    # mkfs.btrfs -f test.img
    # mount test.img /mnt/btrfs
    # xfs_io -f -c "pwrite 0 1k" -c "sync" -c "falloc 0 4k" \
	       -c "pwrite 4k 4k" /mnt/btrfs/file
    # umount /mnt/btrfs
    # ./grub-fstest ./grub-fstest --debug=btrfs ~/test.img hex "/file"

Such mixed inline/regular extents case is not recommended layout, but all existing tools and kernel can handle it without problem

2. NO_HOLES feature

    # mkfs.btrfs -f test.img -O no_holes
    # mount test.img /mnt/btrfs
    # xfs_io -f -c "pwrite 0 4k" -c "pwrite 8k 4k" /mnt/btrfs/file
    # umount /mnt/btrfs
    # ./grub-fstest ./grub-fstest --debug=btrfs ~/test.img hex "/file"

NO_HOLES feature is going to be the default mkfs feature in the incoming v5.15 release, and kernel has support for it since v4.0.

[CAUSE]
The way GRUB btrfs code iterates through file extents relies on no gap between extents.

If any gap is hit, then grub btrfs will error out, without any proper reason to help debug the bug.

This is a bad assumption, since a long long time ago btrfs has a new feature called NO_HOLES to allow btrfs to skip the padding hole extent to reduce metadata usage.

The NO_HOLES feature is already stable since kernel v4.0 and is going to be the default mkfs feature in the incoming v5.15 btrfs-progs release.

[FIX]
When there is a extent gap, instead of error out, just try next item.

This is still not ideal, as kernel/progs/U-boot all do the iteration item by item, not relying on the file offset continuity.

But it will be way more time consuming to correct the whole behavior than starting from scratch to build a proper designed btrfs module for GRUB.

---
  grub-core/fs/btrfs.c | 33 ++++++++++++++++++++++++++++++---
  1 file changed, 30 insertions(+), 3 deletions(-)
grub-core/fs/btrfs.c +30 -3
index 63203034dfc6..4fbcbec7524a 100644
@@ -1443,6 +1443,7 @@ grub_btrfs_extent_read (struct grub_btrfs_data *data,
        grub_size_t csize;
        grub_err_t err;
        grub_off_t extoff;
+      struct grub_btrfs_leaf_descriptor desc;
        if (!data->extent || data->extstart > pos || data->extino != ino
  	  || data->exttree != tree || data->extend <= pos)
  	{
@@ -1455,7 +1456,7 @@ grub_btrfs_extent_read (struct grub_btrfs_data *data,
  	  key_in.type = GRUB_BTRFS_ITEM_TYPE_EXTENT_ITEM;
  	  key_in.offset = grub_cpu_to_le64 (pos);
  	  err = lower_bound (data, &key_in, &key_out, tree,
-			     &elemaddr, &elemsize, NULL, 0);
+			     &elemaddr, &elemsize, &desc, 0);
  	  if (err)
  	    return -1;
  	  if (key_out.object_id != ino
@@ -1494,10 +1495,36 @@ grub_btrfs_extent_read (struct grub_btrfs_data *data,
  			PRIxGRUB_UINT64_T "\n",
  			grub_le_to_cpu64 (key_out.offset),
  			grub_le_to_cpu64 (data->extent->size));
+	  /*
+	   * The way of extent item iteration is pretty bad, it completely
+	   * requires all extents are contiguous, which is not ensured.
+	   *
+	   * Features like NO_HOLE and mixed inline/regular extents can cause
+	   * gaps between file extent items.
+	   *
+	   * The correct way is to follow kernel/U-boot to iterate item by
+	   * item, without any assumption on the file offset continuity.
+	   *
+	   * Here we just manually skip to next item and re-do the verification.
+	   *
+	   * TODO: Rework the whole extent item iteration code, if not the
+	   * whole btrfs implementation.
+	   */
  	  if (data->extend <= pos)
  	    {
-	      grub_error (GRUB_ERR_BAD_FS, "extent not found");
-	      return -1;
+	      err = next(data, &desc, &elemaddr, &elemsize, &key_out);
+	      if (err < 0)
+		return -1;
+	      /* No next item for the inode, we hit the end */
+	      if (err == 0 || key_out.object_id != ino ||
+		  key_out.type != GRUB_BTRFS_ITEM_TYPE_EXTENT_ITEM)
+		      return pos - pos0;
+
+	      csize = grub_le_to_cpu64(key_out.offset) - pos;
+	      buf += csize;
+	      pos += csize;
+	      len -= csize;
+	      continue;
  	    }
  	}
        csize = data->extend - pos;
#2 lore raw
Quoted 7 lines · The development of GNU GRUB

Gentle ping?

Without this patch, the new mkfs.btrfs NO_HOLES feature would break any kernel/initramfs with hole in it.

And considering the modification is already small, I believe this patch is definitely worthy as a bug fix.

+1

This does worth more attention as booting would have been blocked.

Quoted 106 lines

Thanks,
Qu

Quoted 101 lines · Qu Wenruo

[BUG]
Grub btrfs implementation can't handle two very basic btrfs file layouts:

1. Mixed inline/regualr extents

    # mkfs.btrfs -f test.img
    # mount test.img /mnt/btrfs
    # xfs_io -f -c "pwrite 0 1k" -c "sync" -c "falloc 0 4k" \
	       -c "pwrite 4k 4k" /mnt/btrfs/file
    # umount /mnt/btrfs
    # ./grub-fstest ./grub-fstest --debug=btrfs ~/test.img hex "/file"

Such mixed inline/regular extents case is not recommended layout, but all existing tools and kernel can handle it without problem

2. NO_HOLES feature

    # mkfs.btrfs -f test.img -O no_holes
    # mount test.img /mnt/btrfs
    # xfs_io -f -c "pwrite 0 4k" -c "pwrite 8k 4k" /mnt/btrfs/file
    # umount /mnt/btrfs
    # ./grub-fstest ./grub-fstest --debug=btrfs ~/test.img hex "/file"

NO_HOLES feature is going to be the default mkfs feature in the incoming v5.15 release, and kernel has support for it since v4.0.

[CAUSE]
The way GRUB btrfs code iterates through file extents relies on no gap between extents.

If any gap is hit, then grub btrfs will error out, without any proper reason to help debug the bug.

This is a bad assumption, since a long long time ago btrfs has a new feature called NO_HOLES to allow btrfs to skip the padding hole extent to reduce metadata usage.

The NO_HOLES feature is already stable since kernel v4.0 and is going to be the default mkfs feature in the incoming v5.15 btrfs-progs release.

[FIX]
When there is a extent gap, instead of error out, just try next item.

This is still not ideal, as kernel/progs/U-boot all do the iteration item by item, not relying on the file offset continuity.

But it will be way more time consuming to correct the whole behavior than starting from scratch to build a proper designed btrfs module for GRUB.

---
  grub-core/fs/btrfs.c | 33 ++++++++++++++++++++++++++++++---
  1 file changed, 30 insertions(+), 3 deletions(-)
grub-core/fs/btrfs.c +21 -3
index 63203034dfc6..4fbcbec7524a 100644
@@ -1443,6 +1443,7 @@ grub_btrfs_extent_read (struct grub_btrfs_data *data,
        grub_size_t csize;
        grub_err_t err;
        grub_off_t extoff;
+      struct grub_btrfs_leaf_descriptor desc;
        if (!data->extent || data->extstart > pos || data->extino != ino
  	  || data->exttree != tree || data->extend <= pos)
  	{
@@ -1455,7 +1456,7 @@ grub_btrfs_extent_read (struct grub_btrfs_data *data,
  	  key_in.type = GRUB_BTRFS_ITEM_TYPE_EXTENT_ITEM;
  	  key_in.offset = grub_cpu_to_le64 (pos);
  	  err = lower_bound (data, &key_in, &key_out, tree,
-			     &elemaddr, &elemsize, NULL, 0);
+			     &elemaddr, &elemsize, &desc, 0);
  	  if (err)
  	    return -1;
  	  if (key_out.object_id != ino
@@ -1494,10 +1495,36 @@ grub_btrfs_extent_read (struct grub_btrfs_data *data,
  			PRIxGRUB_UINT64_T "\n",
  			grub_le_to_cpu64 (key_out.offset),
  			grub_le_to_cpu64 (data->extent->size));
+	  /*
+	   * The way of extent item iteration is pretty bad, it completely
+	   * requires all extents are contiguous, which is not ensured.
+	   *
+	   * Features like NO_HOLE and mixed inline/regular extents can cause
+	   * gaps between file extent items.
+	   *
+	   * The correct way is to follow kernel/U-boot to iterate item by
+	   * item, without any assumption on the file offset continuity.
+	   *
+	   * Here we just manually skip to next item and re-do the verification.
+	   *
+	   * TODO: Rework the whole extent item iteration code, if not the
+	   * whole btrfs implementation.
+	   */
  	  if (data->extend <= pos)
  	    {
-	      grub_error (GRUB_ERR_BAD_FS, "extent not found");
-	      return -1;
+	      err = next(data, &desc, &elemaddr, &elemsize, &key_out);
+	      if (err < 0)
+		return -1;
+	      /* No next item for the inode, we hit the end */
Quoted 8 lines

grub-core/fs/btrfs.c grub_btrfs_extent_read (struct grub_btrfs_data *data, · line 1526

Quoted 8 lines
grub-core/fs/btrfs.c +8 -0
+	      if (err == 0 || key_out.object_id != ino ||
+		  key_out.type != GRUB_BTRFS_ITEM_TYPE_EXTENT_ITEM)
+		      return pos - pos0;
+
+	      csize = grub_le_to_cpu64(key_out.offset) - pos;
+	      buf += csize;
+	      pos += csize;
+	      len -= csize;

Does it make sense to add some sort of range check here to safegard len from being underflow ? My justfication is that csize value is read out from disk so can be anything wild presumably.

Thanks,
Michael

Quoted 11 lines

grub-core/fs/btrfs.c grub_btrfs_extent_read (struct grub_btrfs_data *data, · line 1530

Quoted 5 lines
grub-core/fs/btrfs.c +1 -0
+	      continue;
  	    }
  	}
        csize = data->extend - pos;

_______________________________________________
Grub-devel mailing list
[email protected]
https://lists.gnu.org/mailman/listinfo/grub-devel

#3 lore raw
Quoted 122 lines · Michael Chang
Quoted 7 lines · The development of GNU GRUB

Gentle ping?

Without this patch, the new mkfs.btrfs NO_HOLES feature would break any kernel/initramfs with hole in it.

And considering the modification is already small, I believe this patch is definitely worthy as a bug fix.

+1

This does worth more attention as booting would have been blocked.

Quoted 109 lines

Thanks,
Qu

Quoted 104 lines · Qu Wenruo

[BUG]
Grub btrfs implementation can't handle two very basic btrfs file layouts:

1. Mixed inline/regualr extents

     # mkfs.btrfs -f test.img
     # mount test.img /mnt/btrfs
     # xfs_io -f -c "pwrite 0 1k" -c "sync" -c "falloc 0 4k" \
	       -c "pwrite 4k 4k" /mnt/btrfs/file
     # umount /mnt/btrfs
     # ./grub-fstest ./grub-fstest --debug=btrfs ~/test.img hex "/file"

Such mixed inline/regular extents case is not recommended layout, but all existing tools and kernel can handle it without problem

2. NO_HOLES feature

     # mkfs.btrfs -f test.img -O no_holes
     # mount test.img /mnt/btrfs
     # xfs_io -f -c "pwrite 0 4k" -c "pwrite 8k 4k" /mnt/btrfs/file
     # umount /mnt/btrfs
     # ./grub-fstest ./grub-fstest --debug=btrfs ~/test.img hex "/file"

NO_HOLES feature is going to be the default mkfs feature in the incoming v5.15 release, and kernel has support for it since v4.0.

[CAUSE]
The way GRUB btrfs code iterates through file extents relies on no gap between extents.

If any gap is hit, then grub btrfs will error out, without any proper reason to help debug the bug.

This is a bad assumption, since a long long time ago btrfs has a new feature called NO_HOLES to allow btrfs to skip the padding hole extent to reduce metadata usage.

The NO_HOLES feature is already stable since kernel v4.0 and is going to be the default mkfs feature in the incoming v5.15 btrfs-progs release.

[FIX]
When there is a extent gap, instead of error out, just try next item.

This is still not ideal, as kernel/progs/U-boot all do the iteration item by item, not relying on the file offset continuity.

But it will be way more time consuming to correct the whole behavior than starting from scratch to build a proper designed btrfs module for GRUB.

---
   grub-core/fs/btrfs.c | 33 ++++++++++++++++++++++++++++++---
   1 file changed, 30 insertions(+), 3 deletions(-)
grub-core/fs/btrfs.c +24 -3
index 63203034dfc6..4fbcbec7524a 100644
@@ -1443,6 +1443,7 @@ grub_btrfs_extent_read (struct grub_btrfs_data *data,
         grub_size_t csize;
         grub_err_t err;
         grub_off_t extoff;
+      struct grub_btrfs_leaf_descriptor desc;
         if (!data->extent || data->extstart > pos || data->extino != ino
   	  || data->exttree != tree || data->extend <= pos)
   	{
@@ -1455,7 +1456,7 @@ grub_btrfs_extent_read (struct grub_btrfs_data *data,
   	  key_in.type = GRUB_BTRFS_ITEM_TYPE_EXTENT_ITEM;
   	  key_in.offset = grub_cpu_to_le64 (pos);
   	  err = lower_bound (data, &key_in, &key_out, tree,
-			     &elemaddr, &elemsize, NULL, 0);
+			     &elemaddr, &elemsize, &desc, 0);
   	  if (err)
   	    return -1;
   	  if (key_out.object_id != ino
@@ -1494,10 +1495,36 @@ grub_btrfs_extent_read (struct grub_btrfs_data *data,
   			PRIxGRUB_UINT64_T "\n",
   			grub_le_to_cpu64 (key_out.offset),
   			grub_le_to_cpu64 (data->extent->size));
+	  /*
+	   * The way of extent item iteration is pretty bad, it completely
+	   * requires all extents are contiguous, which is not ensured.
+	   *
+	   * Features like NO_HOLE and mixed inline/regular extents can cause
+	   * gaps between file extent items.
+	   *
+	   * The correct way is to follow kernel/U-boot to iterate item by
+	   * item, without any assumption on the file offset continuity.
+	   *
+	   * Here we just manually skip to next item and re-do the verification.
+	   *
+	   * TODO: Rework the whole extent item iteration code, if not the
+	   * whole btrfs implementation.
+	   */
   	  if (data->extend <= pos)
   	    {
-	      grub_error (GRUB_ERR_BAD_FS, "extent not found");
-	      return -1;
+	      err = next(data, &desc, &elemaddr, &elemsize, &key_out);
+	      if (err < 0)
+		return -1;
+	      /* No next item for the inode, we hit the end */
+	      if (err == 0 || key_out.object_id != ino ||
+		  key_out.type != GRUB_BTRFS_ITEM_TYPE_EXTENT_ITEM)
+		      return pos - pos0;
Quoted 8 lines
Quoted 5 lines

grub-core/fs/btrfs.c grub_btrfs_extent_read (struct grub_btrfs_data *data, · line 1526

Quoted 5 lines
grub-core/fs/btrfs.c +5 -0
+
+	      csize = grub_le_to_cpu64(key_out.offset) - pos;
+	      buf += csize;
+	      pos += csize;
+	      len -= csize;

Does it make sense to add some sort of range check here to safegard len from being underflow ?

Yes, thanks for catching this.

Quoted 2 lines

My justfication is that csize value is read out
from disk so can be anything wild presumably.

It can be even more easier to trigger the underflow.

If the read range ends inside a hole, and NO_HOLES feature is enabled, then csize will be larger than len, and underflow.

Thankfully for GRUB it always read the whole file, not part of it. So it's fine for GRUB, but definitely not for a generic fs implementation.

So I'll update the patch to do a clamp on the size.

Thanks,
Qu

Quoted 17 lines

Thanks,
Michael

Quoted 11 lines

grub-core/fs/btrfs.c grub_btrfs_extent_read (struct grub_btrfs_data *data, · line 1530

Quoted 5 lines
grub-core/fs/btrfs.c +1 -0
+	      continue;
   	    }
   	}
         csize = data->extend - pos;

_______________________________________________
Grub-devel mailing list
[email protected]
https://lists.gnu.org/mailman/listinfo/grub-devel

#4 lore raw
In reply to Qu Wenruo
Quoted 4 lines · Qu Wenruo

Gentle ping?

Without this patch, the new mkfs.btrfs NO_HOLES feature would break any kernel/initramfs with hole in it.

Just to clarify, it's not a new feature, it's been out for a long time (since 3.14). The new thing is that it's going to be enabled by default so the potential impact would be higher.

#5 lore raw
Quoted 9 lines · David Sterba
Quoted 4 lines · Qu Wenruo

Gentle ping?

Without this patch, the new mkfs.btrfs NO_HOLES feature would break any kernel/initramfs with hole in it.

Just to clarify, it's not a new feature, it's been out for a long time (since 3.14). The new thing is that it's going to be enabled by default so the potential impact would be higher.

I will take a look at this patch probably next week.

Daniel