fs/ntfs3: fix slab-out-of-bounds write in ni_create_attr_list()

From: HE WEI (ギカク) <[email protected]>

ni_create_attr_list() allocates a fixed buffer of al_aligned(record_size) (== record_size) bytes and then walks every attribute of the primary MFT record, writing one ATTR_LIST_ENTRY per attribute and advancing the cursor by le_size(name_len), with no check against the end of the buffer; the total size is only computed after the loop.

A minimum-size resident attribute occupies SIZEOF_RESIDENT (0x18 = 24) bytes on disk, but an unnamed attribute expands to le_size(0) (0x20 = 32) bytes in the list. Because the number of attributes in a record is not bounded (mi_enum_attr() accepts arbitrarily many equal-type, nameless minimum-size attributes), a crafted record packed with such attributes produces a list larger than record_size and overflows the heap buffer.

This is reachable from a crafted, loop-mounted NTFS image: opening the file and adding an attribute (e.g. via setxattr) drives ntfs_set_ea() -> ni_insert_resident() -> ni_insert_attr() -> ni_ins_attr_ext() -> ni_create_attr_list().

  BUG: KASAN: slab-out-of-bounds in ni_create_attr_list+0xc48/0x1058
  Write of size 4 at addr ffff000008984c00 by task setfattr/345
   ni_create_attr_list+0xc48/0x1058
   ni_ins_attr_ext+0x510/0x7c0
   ni_insert_attr+0x3f8/0x70c
   ni_insert_resident+0xc8/0x3b0
   ntfs_set_ea+0x66c/0xd28
   ntfs_setxattr+0x4d8/0x5b0
   __arm64_sys_setxattr+0xa4/0x124
  Allocated by task 345:
   ni_create_attr_list+0x188/0x1058
  The buggy address belongs to the cache kmalloc-1k of size 1024
  (the write lands at object+1024).

Size the buffer from the actual attributes instead of assuming a single record_size is always enough.

Fixes 4342306f0f0d ("fs/ntfs3: Add file operations and implementation")

Cc [email protected]

Signed-off-by HE WEI (ギカク) <[email protected]>

---
v2:
- Add Cc: [email protected]: this is an attacker-controlled on-disk image heap out-of-bounds write and should be backported.
- No functional change from v1; widening Cc (linux-fsdevel, VFS) for review, as the v1 posting received no response.
- Drop a redundant self Reported-by.

v1: https://lore.kernel.org/all/[email protected]/

---
 fs/ntfs3/frecord.c | 20 ++++++++++++++++----
 1 file changed, 16 insertions(+), 4 deletions(-)

fs/ntfs3/frecord.c +16 -5

index 2e901d073fe9..6488d7a415c0 100644
@@ -768,10 +768,23 @@ int ni_create_attr_list(struct ntfs_inode *ni)
        rs = sbi->record_size;
 
        /*
-        * Skip estimating exact memory requirement.
-        * Looks like one record_size is always enough.
+        * Compute the exact size of the attribute list.  Each attribute in the
+        * record yields one ATTR_LIST_ENTRY of le_size(name_len) bytes.  The
+        * minimum on-disk attribute is SIZEOF_RESIDENT (0x18) bytes, but an
+        * unnamed one expands to le_size(0) (0x20) here, so a record crafted
+        * with many such attributes needs more than a single record_size; the
+        * previous fixed kzalloc(record_size) could therefore be overflowed by
+        * an attacker-controlled record.
         */
-       le = kzalloc(al_aligned(rs), GFP_NOFS);
+       lsize = 0;
+       attr = NULL;
+       while ((attr = mi_enum_attr(ni, &ni->mi, attr)))
+               lsize += le_size(attr->name_len);
+
+       if (!lsize)
+               return -EINVAL;
+
+       le = kzalloc(al_aligned(lsize), GFP_NOFS);
        if (!le)
                return -ENOMEM;
 

@@ -781,7 +794,6 @@ int ni_create_attr_list(struct ntfs_inode *ni)
        attr = NULL;
        nb = 0;
        free_b = 0;
-       attr = NULL;
 
        for (; (attr = mi_enum_attr(ni, &ni->mi, attr)); le = Add2Ptr(le, sz)) {
                sz = le_size(attr->name_len);
--

2.43.0

Hi Konstantin,

Gentle ping on this v2. It's an attacker-controlled on-disk image heap out-of-bounds write in ni_create_attr_list(),
reachable via setxattr on a crafted, loop-mounted NTFS image (KASAN trace is in the commit message), which is why it's Cc'd to stable.

For the record, the fix was first posted as v1 on 2026-06-10:
https://lore.kernel.org/ntfs3/[email protected]/
This v2 (2026-06-25) only adds Cc: stable and widens review to linux-fsdevel; the fix itself is unchanged from v1.

Could you let me know if you'd like any changes, or whether it can be queued for a future bugfix pull? I'm happy to rebase or adjust as needed.

Thanks,
HE WEI (ギカク)

Quoted 90 lines · hewei-gikaku <[email protected]> 于2026年6月25日周四 12:19写道：

From: HE WEI (ギカク) <[email protected]>

ni_create_attr_list() allocates a fixed buffer of al_aligned(record_size) (== record_size) bytes and then walks every attribute of the primary MFT record, writing one ATTR_LIST_ENTRY per attribute and advancing the cursor by le_size(name_len), with no check against the end of the buffer; the total size is only computed after the loop.

A minimum-size resident attribute occupies SIZEOF_RESIDENT (0x18 = 24) bytes on disk, but an unnamed attribute expands to le_size(0) (0x20 = 32) bytes in the list. Because the number of attributes in a record is not bounded (mi_enum_attr() accepts arbitrarily many equal-type, nameless minimum-size attributes), a crafted record packed with such attributes produces a list larger than record_size and overflows the heap buffer.

This is reachable from a crafted, loop-mounted NTFS image: opening the file and adding an attribute (e.g. via setxattr) drives ntfs_set_ea() -> ni_insert_resident() -> ni_insert_attr() -> ni_ins_attr_ext() -> ni_create_attr_list().

   BUG: KASAN: slab-out-of-bounds in ni_create_attr_list+0xc48/0x1058
   Write of size 4 at addr ffff000008984c00 by task setfattr/345
    ni_create_attr_list+0xc48/0x1058
    ni_ins_attr_ext+0x510/0x7c0
    ni_insert_attr+0x3f8/0x70c
    ni_insert_resident+0xc8/0x3b0
    ntfs_set_ea+0x66c/0xd28
    ntfs_setxattr+0x4d8/0x5b0
    __arm64_sys_setxattr+0xa4/0x124
   Allocated by task 345:
    ni_create_attr_list+0x188/0x1058
   The buggy address belongs to the cache kmalloc-1k of size 1024
   (the write lands at object+1024).

Size the buffer from the actual attributes instead of assuming a single record_size is always enough.

Fixes 4342306f0f0d ("fs/ntfs3: Add file operations and implementation")

Cc [email protected]

Signed-off-by HE WEI (ギカク) <[email protected]>

---
v2:
- Add Cc: [email protected]: this is an attacker-controlled on-disk image heap out-of-bounds write and should be backported.
- No functional change from v1; widening Cc (linux-fsdevel, VFS) for review, as the v1 posting received no response.
- Drop a redundant self Reported-by.

v1: https://lore.kernel.org/all/[email protected]/

---
  fs/ntfs3/frecord.c | 20 ++++++++++++++++----
  1 file changed, 16 insertions(+), 4 deletions(-)

fs/ntfs3/frecord.c +16 -4

index 2e901d073fe9..6488d7a415c0 100644
@@ -768,10 +768,23 @@ int ni_create_attr_list(struct ntfs_inode *ni)
         rs = sbi->record_size;
 
         /*
-        * Skip estimating exact memory requirement.
-        * Looks like one record_size is always enough.
+        * Compute the exact size of the attribute list.  Each attribute in the
+        * record yields one ATTR_LIST_ENTRY of le_size(name_len) bytes.  The
+        * minimum on-disk attribute is SIZEOF_RESIDENT (0x18) bytes, but an
+        * unnamed one expands to le_size(0) (0x20) here, so a record crafted
+        * with many such attributes needs more than a single record_size; the
+        * previous fixed kzalloc(record_size) could therefore be overflowed by
+        * an attacker-controlled record.
          */
-       le = kzalloc(al_aligned(rs), GFP_NOFS);
+       lsize = 0;
+       attr = NULL;
+       while ((attr = mi_enum_attr(ni, &ni->mi, attr)))
+               lsize += le_size(attr->name_len);
+
+       if (!lsize)
+               return -EINVAL;
+
+       le = kzalloc(al_aligned(lsize), GFP_NOFS);
         if (!le)
                 return -ENOMEM;
 

@@ -781,7 +794,6 @@ int ni_create_attr_list(struct ntfs_inode *ni)
         attr = NULL;
         nb = 0;
         free_b = 0;
-       attr = NULL;
 

Quoted 4 lines

fs/ntfs3/frecord.c +0 -1

         for (; (attr = mi_enum_attr(ni, &ni->mi, attr)); le = Add2Ptr(le, sz)) {
                 sz = le_size(attr->name_len);
--

2.43.0

Hi Konstantin,

Gentle ping on this v2. It's an attacker-controlled on-disk image heap out-of-bounds write in ni_create_attr_list(),
reachable via setxattr on a crafted, loop-mounted NTFS image (KASAN trace is in the commit message), which is why it's Cc'd to stable.

For the record, the fix was first posted as v1 on 2026-06-10:
https://lore.kernel.org/ntfs3/[email protected]/
This v2 (2026-06-25) only adds Cc: stable and widens review to linux-fsdevel; the fix itself is unchanged from v1.

Could you let me know if you'd like any changes, or whether it can be queued for a future bugfix pull? I'm happy to rebase or adjust as needed.

Thanks,
HE WEI (ギカク)

Quoted 90 lines · hewei-gikaku <[email protected]> 于2026年6月25日周四 12:19写道：

From: HE WEI (ギカク) <[email protected]>

ni_create_attr_list() allocates a fixed buffer of al_aligned(record_size) (== record_size) bytes and then walks every attribute of the primary MFT record, writing one ATTR_LIST_ENTRY per attribute and advancing the cursor by le_size(name_len), with no check against the end of the buffer; the total size is only computed after the loop.

A minimum-size resident attribute occupies SIZEOF_RESIDENT (0x18 = 24) bytes on disk, but an unnamed attribute expands to le_size(0) (0x20 = 32) bytes in the list. Because the number of attributes in a record is not bounded (mi_enum_attr() accepts arbitrarily many equal-type, nameless minimum-size attributes), a crafted record packed with such attributes produces a list larger than record_size and overflows the heap buffer.

This is reachable from a crafted, loop-mounted NTFS image: opening the file and adding an attribute (e.g. via setxattr) drives ntfs_set_ea() -> ni_insert_resident() -> ni_insert_attr() -> ni_ins_attr_ext() -> ni_create_attr_list().

   BUG: KASAN: slab-out-of-bounds in ni_create_attr_list+0xc48/0x1058
   Write of size 4 at addr ffff000008984c00 by task setfattr/345
    ni_create_attr_list+0xc48/0x1058
    ni_ins_attr_ext+0x510/0x7c0
    ni_insert_attr+0x3f8/0x70c
    ni_insert_resident+0xc8/0x3b0
    ntfs_set_ea+0x66c/0xd28
    ntfs_setxattr+0x4d8/0x5b0
    __arm64_sys_setxattr+0xa4/0x124
   Allocated by task 345:
    ni_create_attr_list+0x188/0x1058
   The buggy address belongs to the cache kmalloc-1k of size 1024
   (the write lands at object+1024).

Size the buffer from the actual attributes instead of assuming a single record_size is always enough.

Fixes 4342306f0f0d ("fs/ntfs3: Add file operations and implementation")

Cc [email protected]

Signed-off-by HE WEI (ギカク) <[email protected]>

---
v2:
- Add Cc: [email protected]: this is an attacker-controlled on-disk image heap out-of-bounds write and should be backported.
- No functional change from v1; widening Cc (linux-fsdevel, VFS) for review, as the v1 posting received no response.
- Drop a redundant self Reported-by.

v1: https://lore.kernel.org/all/[email protected]/

---
  fs/ntfs3/frecord.c | 20 ++++++++++++++++----
  1 file changed, 16 insertions(+), 4 deletions(-)

fs/ntfs3/frecord.c +16 -4

index 2e901d073fe9..6488d7a415c0 100644
@@ -768,10 +768,23 @@ int ni_create_attr_list(struct ntfs_inode *ni)
         rs = sbi->record_size;
 
         /*
-        * Skip estimating exact memory requirement.
-        * Looks like one record_size is always enough.
+        * Compute the exact size of the attribute list.  Each attribute in the
+        * record yields one ATTR_LIST_ENTRY of le_size(name_len) bytes.  The
+        * minimum on-disk attribute is SIZEOF_RESIDENT (0x18) bytes, but an
+        * unnamed one expands to le_size(0) (0x20) here, so a record crafted
+        * with many such attributes needs more than a single record_size; the
+        * previous fixed kzalloc(record_size) could therefore be overflowed by
+        * an attacker-controlled record.
          */
-       le = kzalloc(al_aligned(rs), GFP_NOFS);
+       lsize = 0;
+       attr = NULL;
+       while ((attr = mi_enum_attr(ni, &ni->mi, attr)))
+               lsize += le_size(attr->name_len);
+
+       if (!lsize)
+               return -EINVAL;
+
+       le = kzalloc(al_aligned(lsize), GFP_NOFS);
         if (!le)
                 return -ENOMEM;
 

@@ -781,7 +794,6 @@ int ni_create_attr_list(struct ntfs_inode *ni)
         attr = NULL;
         nb = 0;
         free_b = 0;
-       attr = NULL;
 

Quoted 4 lines

fs/ntfs3/frecord.c +0 -1

         for (; (attr = mi_enum_attr(ni, &ni->mi, attr)); le = Add2Ptr(le, sz)) {
                 sz = le_size(attr->name_len);
--

2.43.0

Quoted 110 lines · HE WEI（ギカク）

Hi Konstantin,

Gentle ping on this v2. It's an attacker-controlled on-disk image heap out-of-bounds write in ni_create_attr_list(),
reachable via setxattr on a crafted, loop-mounted NTFS image (KASAN trace is in the commit message), which is why it's Cc'd to stable.

For the record, the fix was first posted as v1 on 2026-06-10:
https://lore.kernel.org/ntfs3/[email protected]/
This v2 (2026-06-25) only adds Cc: stable and widens review to linux-fsdevel; the fix itself is unchanged from v1.

Could you let me know if you'd like any changes, or whether it can be queued for a future bugfix pull? I'm happy to rebase or adjust as needed.

Thanks,
HE WEI (ギカク)

Quoted 90 lines · hewei-gikaku <[email protected]> 于2026年6月25日周四 12:19写道：

From: HE WEI (ギカク) <[email protected]>

ni_create_attr_list() allocates a fixed buffer of al_aligned(record_size) (== record_size) bytes and then walks every attribute of the primary MFT record, writing one ATTR_LIST_ENTRY per attribute and advancing the cursor by le_size(name_len), with no check against the end of the buffer; the total size is only computed after the loop.

A minimum-size resident attribute occupies SIZEOF_RESIDENT (0x18 = 24) bytes on disk, but an unnamed attribute expands to le_size(0) (0x20 = 32) bytes in the list. Because the number of attributes in a record is not bounded (mi_enum_attr() accepts arbitrarily many equal-type, nameless minimum-size attributes), a crafted record packed with such attributes produces a list larger than record_size and overflows the heap buffer.

This is reachable from a crafted, loop-mounted NTFS image: opening the file and adding an attribute (e.g. via setxattr) drives ntfs_set_ea() -> ni_insert_resident() -> ni_insert_attr() -> ni_ins_attr_ext() -> ni_create_attr_list().

   BUG: KASAN: slab-out-of-bounds in ni_create_attr_list+0xc48/0x1058
   Write of size 4 at addr ffff000008984c00 by task setfattr/345
    ni_create_attr_list+0xc48/0x1058
    ni_ins_attr_ext+0x510/0x7c0
    ni_insert_attr+0x3f8/0x70c
    ni_insert_resident+0xc8/0x3b0
    ntfs_set_ea+0x66c/0xd28
    ntfs_setxattr+0x4d8/0x5b0
    __arm64_sys_setxattr+0xa4/0x124
   Allocated by task 345:
    ni_create_attr_list+0x188/0x1058
   The buggy address belongs to the cache kmalloc-1k of size 1024
   (the write lands at object+1024).

Size the buffer from the actual attributes instead of assuming a single record_size is always enough.

Fixes 4342306f0f0d ("fs/ntfs3: Add file operations and implementation")

Cc [email protected]

Signed-off-by HE WEI (ギカク) <[email protected]>

---
v2:
- Add Cc: [email protected]: this is an attacker-controlled on-disk image heap out-of-bounds write and should be backported.
- No functional change from v1; widening Cc (linux-fsdevel, VFS) for review, as the v1 posting received no response.
- Drop a redundant self Reported-by.

v1: https://lore.kernel.org/all/[email protected]/

---
  fs/ntfs3/frecord.c | 20 ++++++++++++++++----
  1 file changed, 16 insertions(+), 4 deletions(-)

fs/ntfs3/frecord.c +16 -4

index 2e901d073fe9..6488d7a415c0 100644
@@ -768,10 +768,23 @@ int ni_create_attr_list(struct ntfs_inode *ni)
         rs = sbi->record_size;
 
         /*
-        * Skip estimating exact memory requirement.
-        * Looks like one record_size is always enough.
+        * Compute the exact size of the attribute list.  Each attribute in the
+        * record yields one ATTR_LIST_ENTRY of le_size(name_len) bytes.  The
+        * minimum on-disk attribute is SIZEOF_RESIDENT (0x18) bytes, but an
+        * unnamed one expands to le_size(0) (0x20) here, so a record crafted
+        * with many such attributes needs more than a single record_size; the
+        * previous fixed kzalloc(record_size) could therefore be overflowed by
+        * an attacker-controlled record.
          */
-       le = kzalloc(al_aligned(rs), GFP_NOFS);
+       lsize = 0;
+       attr = NULL;
+       while ((attr = mi_enum_attr(ni, &ni->mi, attr)))
+               lsize += le_size(attr->name_len);
+
+       if (!lsize)
+               return -EINVAL;
+
+       le = kzalloc(al_aligned(lsize), GFP_NOFS);
         if (!le)
                 return -ENOMEM;
 

@@ -781,7 +794,6 @@ int ni_create_attr_list(struct ntfs_inode *ni)
         attr = NULL;
         nb = 0;
         free_b = 0;
-       attr = NULL;
 

Quoted 4 lines

Quoted 4 lines

fs/ntfs3/frecord.c +0 -1

         for (; (attr = mi_enum_attr(ni, &ni->mi, attr)); le = Add2Ptr(le, sz)) {
                 sz = le_size(attr->name_len);
--

2.43.0

Hello,

Sorry for the confusion. The fix is already upstream: v1 was applied as commit 7c4841e2a627 ("fs/ntfs3: fix slab-out-of-bounds write in ni_create_attr_list()"), in v7.3-rc1. v2 is functionally identical, so there is nothing further for me to apply.

The applied commit does not carry Cc: stable, though, so it has not been backported.

Thanks for the fix, and for following up on it.

Regards,
Konstantin

Thanks for clarifying. Since the upstream commit did not carry the stable tag, could the stable team please consider backporting commit 7c4841e2a627 to the applicable supported stable trees? This fixes an attacker-controlled on-disk slab-out-of-bounds write. My v2 carried
Cc: [email protected], but the functionally identical v1 was the version merged upstream.

Thanks for clarifying. Since the upstream commit did not carry the stable tag, could the stable team please consider backporting commit 7c4841e2a627 to the applicable supported stable trees? This fixes an attacker-controlled on-disk slab-out-of-bounds write. My v2 carried
Cc: [email protected], but the functionally identical v1 was the version merged upstream.