Skip to contentSkip to conversation

maple_tree: maple_node slab object corruption via out-of-bounds write during VMA operations

BUG psg 2 participants 2 messages Feb 7 Memory management

lore
Share & tools
lore

psg

author
#0 lore raw

Hi,

We are hitting a reproducible maple_node slab corruption on 6.18.0-rc6 (ARM64) during early boot. The corruption manifests as
a left redzone overwrite detected by slub_debug, followed by a kernel panic due to panic_on_taint. We have captured two independent crash dumps showing the exact same corruption pattern.

We have verified that the CVE-2024-50200
fix (commit bea07fd63192 "maple_tree: correct tree corruption on spanning store") IS present in our kernel -- the r_mas.max > r_mas.last comparison uses 64-bit registers as expected. We have also ruled out CVE-2025-38364 (MA_STATE_PREALLOC flag), which causes NULL pointer dereference rather than an out-of-bounds write.

Environment
-----------
  Kernel: 6.18.0-rc6 (mainline, ARM64)
  Config: SMP PREEMPT, slub_debug=FZP enabled for maple_node Cmdline includes: panic_on_taint=0x20 slub_debug=FZP,zs_handle,...

Crash #1 (mmap path)
---------------------
  BUG maple_node (Tainted: G W OE): Object corrupt
  [Left Redzone overwritten] 0xffffff8828948300-0xffffff88289483ff @offset=768. First byte 0x1 instead of 0xcc

  Allocated in kmem_cache_prefill_sheaf+0x308/0x33c age=291 cpu=2 pid=87171

    kmem_cache_prefill_sheaf+0x308/0x33c
    mas_alloc_nodes+0x98/0xf0
    mas_preallocate+0x234/0x33c
    __split_vma+0x11c/0x364
    vms_gather_munmap_vmas+0x118/0x310
    mmap_region+0x2a8/0xae4
    do_mmap+0x470/0x578
    vm_mmap_pgoff+0x1e8/0x264
    ksys_mmap_pgoff+0xa4/0xf0
    __arm64_sys_mmap+0x34/0x44

  Freed in mt_destroy_walk+0x16c/0x344 age=391 cpu=6 pid=83838

    kmem_cache_free_bulk+0x3c4/0x9f8
    mt_destroy_walk+0x16c/0x344
    __mt_destroy+0x40/0x80
    exit_mmap+0x2ac/0x4b0
    __mmput+0x38/0x16c
    mmput+0x44/0x7c
    exec_mmap+0x208/0x2ac
    begin_new_exec+0x188/0x46c
    load_elf_binary+0x434/0xc68

  Slab 0xfffffffee0a25200 objects=21 used=16 fp=0xffffff882894a200 flags=0x4000000000000240(workingset|head|zone=1) Object 0xffffff8828948400 @offset=1024 fp=0xffffff8828949c00

  Panic call trace (detected during RCU free):

    check_bytes_and_report+0x104/0x31c
    check_object+0x98/0x3c8
    free_to_partial_list+0x174/0x638
    __slab_free+0x204/0x248
    kmem_cache_free_bulk+0x3c4/0x9f8
    kvfree_rcu_bulk+0x17c/0x320
    kfree_rcu_work+0xb8/0x144

Crash #2 (mprotect path)
-------------------------
  BUG maple_node (Tainted: G W OE): Object corrupt
  [Left Redzone overwritten] 0xffffff88184b8300-0xffffff88184b83ff @offset=768. First byte 0x1 instead of 0xbb

  Allocated in mas_alloc_nodes+0xcc/0xf0 age=343 cpu=3 pid=89696

    kmem_cache_alloc_noprof+0x3fc/0x55c
    mas_alloc_nodes+0xcc/0xf0
    mas_preallocate+0x234/0x33c
    __split_vma+0x11c/0x364
    vma_modify+0x424/0x4dc
    vma_modify_flags+0x74/0xa0
    mprotect_fixup+0x154/0x28c
    do_mprotect_pkey+0x410/0x5b0
    __arm64_sys_mprotect+0x20/0x34

  Freed in kvfree_rcu_bulk+0x17c/0x320 age=335 cpu=7 pid=9090

    kmem_cache_free_bulk+0x3c4/0x9f8
    kvfree_rcu_bulk+0x17c/0x320
    kfree_rcu_work+0xb8/0x144

  Slab 0xfffffffee0612e00 objects=21 used=8 fp=0xffffff88184b8100 flags=0x4000000000000240(workingset|head|zone=1) Object 0xffffff88184b8400 @offset=1024 fp=0xffffff88184b8100

  Panic call trace (detected during sheaf prefill alloc):

    check_bytes_and_report+0x104/0x31c
    check_object+0x98/0x3c8
    alloc_debug_processing+0x104/0x1b8
    ___slab_alloc+0xb10/0x1314
    __kmem_cache_alloc_bulk+0x1d0/0x460
    kmem_cache_prefill_sheaf+0x308/0x33c
    mas_alloc_nodes+0x98/0xf0
    mas_preallocate+0x234/0x33c
    mmap_region+0x548/0xae4
    do_mmap+0x470/0x578

Redzone corruption pattern analysis
------------------------------------
Both crashes show IDENTICAL structured data in the left redzone of the object at slot 1 (offset 1024). The redzone occupies bytes 768-1023 (256 bytes). The corruption originates from the PREVIOUS maple_node (slot 0, offset 0-255) writing past its 256-byte boundary.

Corrupted left redzone dump (crash #1, 0xcc = SLUB_RED_ACTIVE):

  Redzone  ffffff8828948300: 01 00 00 00 cc cc cc cc cc cc cc cc 78 59 ef ff
  Redzone  ffffff8828948310: 08 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00
  Redzone  ffffff8828948320: cc cc cc cc cc cc cc cc 80 59 ef ff 04 00 00 00
  Redzone  ffffff8828948330: 00 00 00 00 00 00 00 00 01 00 00 00 cc cc cc cc
  Redzone  ffffff8828948340: cc cc cc cc 88 59 ef ff 08 00 00 00 00 00 00 00
  Redzone  ffffff8828948350: 00 00 00 00 01 00 00 00 cc cc cc cc cc cc cc cc
  Redzone  ffffff8828948360: 90 59 ef ff 04 00 00 00 00 00 00 00 00 00 00 00
  Redzone  ffffff8828948370: 01 00 00 00 cc cc cc cc cc cc cc cc 98 59 ef ff
  Redzone  ffffff8828948380: 08 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00
  Redzone  ffffff8828948390: cc cc cc cc cc cc cc cc a0 59 ef ff 18 00 00 00
  Redzone  ffffff88289483a0: 00 00 00 00 00 00 00 00 01 00 00 00 cc cc cc cc
  Redzone  ffffff88289483b0: cc cc cc cc b8 59 ef ff 18 00 00 00 00 00 00 00
  Redzone  ffffff88289483c0: 00 00 00 00 01 00 00 00 cc cc cc cc cc cc cc cc
  Redzone  ffffff88289483d0: d0 59 ef ff 18 00 00 00 00 00 00 00 00 00 00 00
  Redzone  ffffff88289483e0: 01 00 00 00 cc cc cc cc cc cc cc cc e8 59 ef ff
  Redzone  ffffff88289483f0: 18 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00

Corrupted left redzone dump (crash #2, 0xbb = SLUB_RED_INACTIVE):

  Redzone  ffffff88184b8300: 01 00 00 00 bb bb bb bb bb bb bb bb 78 59 ef ff
  Redzone  ffffff88184b8310: 08 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00
  Redzone  ffffff88184b8320: bb bb bb bb bb bb bb bb 80 59 ef ff 04 00 00 00
  Redzone  ffffff88184b8330: 00 00 00 00 00 00 00 00 01 00 00 00 bb bb bb bb
  Redzone  ffffff88184b8340: bb bb bb bb 88 59 ef ff 08 00 00 00 00 00 00 00
  Redzone  ffffff88184b8350: 00 00 00 00 01 00 00 00 bb bb bb bb bb bb bb bb
  Redzone  ffffff88184b8360: 90 59 ef ff 04 00 00 00 00 00 00 00 00 00 00 00
  Redzone  ffffff88184b8370: 01 00 00 00 bb bb bb bb bb bb bb bb 98 59 ef ff
  Redzone  ffffff88184b8380: 08 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00
  Redzone  ffffff88184b8390: bb bb bb bb bb bb bb bb a0 59 ef ff 18 00 00 00
  Redzone  ffffff88184b83a0: 00 00 00 00 00 00 00 00 01 00 00 00 bb bb bb bb
  Redzone  ffffff88184b83b0: bb bb bb bb b8 59 ef ff 18 00 00 00 00 00 00 00
  Redzone  ffffff88184b83c0: 00 00 00 00 01 00 00 00 bb bb bb bb bb bb bb bb
  Redzone  ffffff88184b83d0: d0 59 ef ff 18 00 00 00 00 00 00 00 00 00 00 00
  Redzone  ffffff88184b83e0: 01 00 00 00 bb bb bb bb bb bb bb bb e8 59 ef ff
  Redzone  ffffff88184b83f0: 18 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00

The corruption data is interleaved with original redzone poison bytes (0xcc or 0xbb), which are preserved at 8-byte intervals. The corrupted bytes form a repeating 28-byte structure that resembles maple_range_64 pivot entries containing VMA page-boundary addresses:

  Bytes 0-3: flags/refcount (0x00000001)
  Bytes 4-11: [original redzone poison - NOT overwritten] Bytes 12-15: VMA address fragment (e.g., 0xffef5978, incrementing) Bytes 16-19: size/length field (0x04, 0x08, or 0x18 pages) Bytes 20-27: zero padding

The VMA addresses form a sequential series:
  0x????ffef5978, 0x????ffef5980, 0x????ffef5988, 0x????ffef5990, 0x????ffef5998, 0x????ffef59a0, 0x????ffef59b8, 0x????ffef59d0, 0x????ffef59e8

This pattern is consistent with a maple_range_64 node's pivot/slot data being written beyond the 256-byte maple_node allocation boundary, overflowing into the right redzone of slot 0 and the left redzone of slot 1.

Our analysis of slot 0 data from crash #1 (via physical memory reconstruction from the DDR dump) revealed DUPLICATE PIVOT entries in the previous maple_node -- a pattern reminiscent of CVE-2024-50200, but occurring despite the fix being present. This suggests there may be another code path in the maple tree that can produce similar spanning store corruption.

Thanks

#1 lore raw
Quoted 20 lines · psg

Hi,

We are hitting a reproducible maple_node slab corruption on 6.18.0-rc6 (ARM64) during early boot. The corruption manifests as
a left redzone overwrite detected by slub_debug, followed by a kernel panic due to panic_on_taint. We have captured two independent crash dumps showing the exact same corruption pattern.

We have verified that the CVE-2024-50200
fix (commit bea07fd63192 "maple_tree: correct tree corruption on spanning store") IS present in our kernel -- the r_mas.max > r_mas.last comparison uses 64-bit registers as expected. We have also ruled out CVE-2025-38364 (MA_STATE_PREALLOC flag), which causes NULL pointer dereference rather than an out-of-bounds write.

Quoted 3 lines

Environment
-----------
  Kernel: 6.18.0-rc6 (mainline, ARM64)

Did you mean 6.19?

Quoted 2 lines

  Config: SMP PREEMPT, slub_debug=FZP enabled for maple_node Cmdline includes: panic_on_taint=0x20 slub_debug=FZP,zs_handle,...

Can we see the full command line and config?

Have you tried 6.18 or any other released kernel?

Did you try enabling CONFIG_DEBUG_VM_MAPLE_TREE ? Please set no_hash_pointers on the command line. This way we can see the full tree dump and where it is happening.

Quoted 23 lines

Crash #1 (mmap path)
---------------------
  BUG maple_node (Tainted: G W OE): Object corrupt
  [Left Redzone overwritten] 0xffffff8828948300-0xffffff88289483ff @offset=768. First byte 0x1 instead of 0xcc

  Allocated in kmem_cache_prefill_sheaf+0x308/0x33c age=291 cpu=2 pid=87171

    kmem_cache_prefill_sheaf+0x308/0x33c
    mas_alloc_nodes+0x98/0xf0
    mas_preallocate+0x234/0x33c
    __split_vma+0x11c/0x364
    vms_gather_munmap_vmas+0x118/0x310
    mmap_region+0x2a8/0xae4
    do_mmap+0x470/0x578
    vm_mmap_pgoff+0x1e8/0x264
    ksys_mmap_pgoff+0xa4/0xf0
    __arm64_sys_mmap+0x34/0x44

  Freed in mt_destroy_walk+0x16c/0x344 age=391 cpu=6 pid=83838

Quoted 9 lines
    kmem_cache_free_bulk+0x3c4/0x9f8
    mt_destroy_walk+0x16c/0x344
    __mt_destroy+0x40/0x80
    exit_mmap+0x2ac/0x4b0
    __mmput+0x38/0x16c
    mmput+0x44/0x7c
    exec_mmap+0x208/0x2ac
    begin_new_exec+0x188/0x46c
    load_elf_binary+0x434/0xc68

This is on exit of the process, so it's not very early in the boot process. Hopefully you can reproduce it with the debug flag without waiting too long.

Quoted 135 lines

  Slab 0xfffffffee0a25200 objects=21 used=16 fp=0xffffff882894a200 flags=0x4000000000000240(workingset|head|zone=1) Object 0xffffff8828948400 @offset=1024 fp=0xffffff8828949c00

  Panic call trace (detected during RCU free):

    check_bytes_and_report+0x104/0x31c
    check_object+0x98/0x3c8
    free_to_partial_list+0x174/0x638
    __slab_free+0x204/0x248
    kmem_cache_free_bulk+0x3c4/0x9f8
    kvfree_rcu_bulk+0x17c/0x320
    kfree_rcu_work+0xb8/0x144

Crash #2 (mprotect path)
-------------------------
  BUG maple_node (Tainted: G W OE): Object corrupt
  [Left Redzone overwritten] 0xffffff88184b8300-0xffffff88184b83ff @offset=768. First byte 0x1 instead of 0xbb

  Allocated in mas_alloc_nodes+0xcc/0xf0 age=343 cpu=3 pid=89696

    kmem_cache_alloc_noprof+0x3fc/0x55c
    mas_alloc_nodes+0xcc/0xf0
    mas_preallocate+0x234/0x33c
    __split_vma+0x11c/0x364
    vma_modify+0x424/0x4dc
    vma_modify_flags+0x74/0xa0
    mprotect_fixup+0x154/0x28c
    do_mprotect_pkey+0x410/0x5b0
    __arm64_sys_mprotect+0x20/0x34

  Freed in kvfree_rcu_bulk+0x17c/0x320 age=335 cpu=7 pid=9090

    kmem_cache_free_bulk+0x3c4/0x9f8
    kvfree_rcu_bulk+0x17c/0x320
    kfree_rcu_work+0xb8/0x144

  Slab 0xfffffffee0612e00 objects=21 used=8 fp=0xffffff88184b8100 flags=0x4000000000000240(workingset|head|zone=1) Object 0xffffff88184b8400 @offset=1024 fp=0xffffff88184b8100

  Panic call trace (detected during sheaf prefill alloc):

    check_bytes_and_report+0x104/0x31c
    check_object+0x98/0x3c8
    alloc_debug_processing+0x104/0x1b8
    ___slab_alloc+0xb10/0x1314
    __kmem_cache_alloc_bulk+0x1d0/0x460
    kmem_cache_prefill_sheaf+0x308/0x33c
    mas_alloc_nodes+0x98/0xf0
    mas_preallocate+0x234/0x33c
    mmap_region+0x548/0xae4
    do_mmap+0x470/0x578

Redzone corruption pattern analysis
------------------------------------
Both crashes show IDENTICAL structured data in the left redzone of the object at slot 1 (offset 1024). The redzone occupies bytes 768-1023 (256 bytes). The corruption originates from the PREVIOUS maple_node (slot 0, offset 0-255) writing past its 256-byte boundary.

Corrupted left redzone dump (crash #1, 0xcc = SLUB_RED_ACTIVE):

  Redzone  ffffff8828948300: 01 00 00 00 cc cc cc cc cc cc cc cc 78 59 ef ff
  Redzone  ffffff8828948310: 08 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00
  Redzone  ffffff8828948320: cc cc cc cc cc cc cc cc 80 59 ef ff 04 00 00 00
  Redzone  ffffff8828948330: 00 00 00 00 00 00 00 00 01 00 00 00 cc cc cc cc
  Redzone  ffffff8828948340: cc cc cc cc 88 59 ef ff 08 00 00 00 00 00 00 00
  Redzone  ffffff8828948350: 00 00 00 00 01 00 00 00 cc cc cc cc cc cc cc cc
  Redzone  ffffff8828948360: 90 59 ef ff 04 00 00 00 00 00 00 00 00 00 00 00
  Redzone  ffffff8828948370: 01 00 00 00 cc cc cc cc cc cc cc cc 98 59 ef ff
  Redzone  ffffff8828948380: 08 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00
  Redzone  ffffff8828948390: cc cc cc cc cc cc cc cc a0 59 ef ff 18 00 00 00
  Redzone  ffffff88289483a0: 00 00 00 00 00 00 00 00 01 00 00 00 cc cc cc cc
  Redzone  ffffff88289483b0: cc cc cc cc b8 59 ef ff 18 00 00 00 00 00 00 00
  Redzone  ffffff88289483c0: 00 00 00 00 01 00 00 00 cc cc cc cc cc cc cc cc
  Redzone  ffffff88289483d0: d0 59 ef ff 18 00 00 00 00 00 00 00 00 00 00 00
  Redzone  ffffff88289483e0: 01 00 00 00 cc cc cc cc cc cc cc cc e8 59 ef ff
  Redzone  ffffff88289483f0: 18 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00

Corrupted left redzone dump (crash #2, 0xbb = SLUB_RED_INACTIVE):

  Redzone  ffffff88184b8300: 01 00 00 00 bb bb bb bb bb bb bb bb 78 59 ef ff
  Redzone  ffffff88184b8310: 08 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00
  Redzone  ffffff88184b8320: bb bb bb bb bb bb bb bb 80 59 ef ff 04 00 00 00
  Redzone  ffffff88184b8330: 00 00 00 00 00 00 00 00 01 00 00 00 bb bb bb bb
  Redzone  ffffff88184b8340: bb bb bb bb 88 59 ef ff 08 00 00 00 00 00 00 00
  Redzone  ffffff88184b8350: 00 00 00 00 01 00 00 00 bb bb bb bb bb bb bb bb
  Redzone  ffffff88184b8360: 90 59 ef ff 04 00 00 00 00 00 00 00 00 00 00 00
  Redzone  ffffff88184b8370: 01 00 00 00 bb bb bb bb bb bb bb bb 98 59 ef ff
  Redzone  ffffff88184b8380: 08 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00
  Redzone  ffffff88184b8390: bb bb bb bb bb bb bb bb a0 59 ef ff 18 00 00 00
  Redzone  ffffff88184b83a0: 00 00 00 00 00 00 00 00 01 00 00 00 bb bb bb bb
  Redzone  ffffff88184b83b0: bb bb bb bb b8 59 ef ff 18 00 00 00 00 00 00 00
  Redzone  ffffff88184b83c0: 00 00 00 00 01 00 00 00 bb bb bb bb bb bb bb bb
  Redzone  ffffff88184b83d0: d0 59 ef ff 18 00 00 00 00 00 00 00 00 00 00 00
  Redzone  ffffff88184b83e0: 01 00 00 00 bb bb bb bb bb bb bb bb e8 59 ef ff
  Redzone  ffffff88184b83f0: 18 00 00 00 00 00 00 00 00 00 00 00 01 00 00 00

The corruption data is interleaved with original redzone poison bytes (0xcc or 0xbb), which are preserved at 8-byte intervals. The corrupted bytes form a repeating 28-byte structure that resembles maple_range_64 pivot entries containing VMA page-boundary addresses:

  Bytes 0-3: flags/refcount (0x00000001)
  Bytes 4-11: [original redzone poison - NOT overwritten] Bytes 12-15: VMA address fragment (e.g., 0xffef5978, incrementing) Bytes 16-19: size/length field (0x04, 0x08, or 0x18 pages) Bytes 20-27: zero padding

The VMA addresses form a sequential series:
  0x????ffef5978, 0x????ffef5980, 0x????ffef5988, 0x????ffef5990, 0x????ffef5998, 0x????ffef59a0, 0x????ffef59b8, 0x????ffef59d0, 0x????ffef59e8

This pattern is consistent with a maple_range_64 node's pivot/slot data being written beyond the 256-byte maple_node allocation boundary, overflowing into the right redzone of slot 0 and the left redzone of slot 1.

Quoted 6 lines

Our analysis of slot 0 data from crash #1 (via physical memory reconstruction from the DDR dump) revealed DUPLICATE PIVOT entries in the previous maple_node -- a pattern reminiscent of CVE-2024-50200, but occurring despite the fix being present. This suggests there may be another code path in the maple tree that can produce similar spanning store corruption.

What are the pivots?

At rcu free time, the data in the nodes may not be reliable so it would be good to try and use the debug validation code in the conf option mentioned above.

Thanks,
Liam