Re: [PATCH v4 1/9] crash_dump: Fix potential double free and UAF of keys_header

From: Coiby Xu

Date: Wed Sep 02 2026 - 09:27:39 EST


On Tue, Sep 01, 2026 at 09:36:10AM +0800, Jinjie Ruan wrote:


在 2026/8/31 21:44, Coiby Xu 写道:
On Mon, Aug 31, 2026 at 03:39:35PM +0800, Jinjie Ruan wrote:


在 2026/8/28 16:48, Coiby Xu 写道:
If kexec_add_buffer somehow fails, keys_header will be freed. Depending
on /sys/kernel/config/crash_dm_crypt_key/reuse, it will lead to the
following two problems if the kexec_file_load syscall is called again,
  1. Double free of keys_header if reuse=false
  2. UAF of keys_header if reuse=true

To address these problems and also make it easier to reason about the
code, keep two invariants,
  1. keys_header will always be freed at the end of kexec_file_load
     syscall except during kdump image unloading for CPU/memory
     hot-plugging support
  2. There will always be valid keys_header if reuse=true

Fixes: 479e58549b0f ("crash_dump: store dm crypt keys in kdump
reserved memory")
Fixes: 9ebfa8dcaea7 ("crash_dump: reuse saved dm crypt keys for CPU/
memory hot-plugging")
Reported-by: Sourabh Jain <sourabhjain@xxxxxxxxxxxxx>
Signed-off-by: Coiby Xu <coiby.xu@xxxxxxxxx>
---
 include/linux/kexec.h        |  6 ++++
 kernel/crash_dump_dm_crypt.c | 66 ++++++++++++++++++++++++++----------
 kernel/kexec_file.c          |  2 ++
 3 files changed, 56 insertions(+), 18 deletions(-)


[...]

@@ -369,9 +387,6 @@ static int build_keys_header(void)
     struct config_key *key;
     int i, r;

-    if (keys_header != NULL)
-        kvfree(keys_header);
-
     keys_header = kzalloc(get_keys_header_size(key_count), GFP_KERNEL);
     if (!keys_header)
         return -ENOMEM;
@@ -415,8 +430,7 @@ int crash_load_dm_crypt_keys(struct kimage *image)
         .top_down = false,
         .random = true,
     };
-    int r;
-
+    int r = 0;

     if (key_count <= 0) {
         kexec_dprintk("No dm-crypt keys\n");
@@ -424,14 +438,15 @@ int crash_load_dm_crypt_keys(struct kimage *image)
     }

     if (!is_dm_key_reused) {
-        image->dm_crypt_keys_addr = 0;
         r = build_keys_header();
-        if (r) {
-            pr_err("Failed to build dm-crypt keys header, ret=%d\n",
r);
-            return r;
-        }
+        if (r)
+            goto out;
     }

+    /*
+     * keys_header will be copied to reserver memory later and then be
+     * cleaned up at the end of kexec_file_load syscall
+     */
     kbuf.buffer = keys_header;
     kbuf.bufsz = get_keys_header_size(key_count);

@@ -441,18 +456,33 @@ int crash_load_dm_crypt_keys(struct kimage *image)
     r = kexec_add_buffer(&kbuf);
     if (r) {
         pr_err("Failed to call kexec_add_buffer, ret=%d\n", r);
-        kvfree((void *)kbuf.buffer);
-        return r;
+        goto out;
     }
+
     image->dm_crypt_keys_addr = kbuf.mem;
     image->dm_crypt_keys_sz = kbuf.bufsz;
     kexec_dprintk(
         "Loaded dm crypt keys to kexec_buffer bufsz=0x%lx
memsz=0x%lx\n",
         kbuf.bufsz, kbuf.memsz);

+out:
+    is_dm_key_reused = false;

I think there's still a bug for arm64 here.

When is_dm_key_reused set to true, the outer while loop prematurely
clears the is_dm_key_reused flag at the end of
crash_load_dm_crypt_keys() in the very first iteration.

Consequently, subsequent re-entrant calls to crash_load_dm_crypt_keys()
falsely perceive the keys as not reused, forcing a redundant execution
of build_keys_header() which overwrites the global keys_header pointer;

This induces a silent kernel memory leak.

Thanks for reviewing this patch and raising the concern! Unless I miss
something, I don't think the global keys_header will be overwritten
because kexec_file_post_load_cleanup_dm_crypt will called to free the
memory during the end of kexec_file_load syscall.

Hi Coiby,

Hi Jinjie,


Not really on arm64, as below:

kexec_file_load syscall
-> kimage_file_alloc_init()
-> kimage_file_prepare_segments()
-> kexec_image_load_default()
-> arm64 image_load()
-> load_other_segments() while retry loop
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-> crash_load_dm_crypt_keys() -> May be called
repeatedly!!! Memory leak bug here as I described above,

Thanks for providing the stack trace to help me understand the problem!
Now I see what I missed. I also notice your "[PATCH v3 04/17] arm64: kexec_file: Fix elf_headers memory leak in retry loop" is addressing the same problem for the elf_headers. I'll have a careful
look at related patches and re-think the solution.


-> kimage_file_post_load_cleanup()



92 >-------/*
93 >------- * The location of the kernel segment may make it impossible
to satisfy
94 >------- * the other segment requirements, so we try repeatedly to
find a
95 >------- * location that will work.
96 >------- */
97 >-------while ((ret = kexec_add_buffer(&kbuf)) == 0) {
98 >------->-------/* Try to load additional data */
99 >------->-------kernel_segment = &image-
>segment[kernel_segment_number];
100 >------->-------ret = load_other_segments(image, kernel_segment->mem,
101 >------->------->------->------->-------  kernel_segment->memsz,
initrd,
102 >------->------->------->------->-------  initrd_len, cmdline);
103 >------->-------if (!ret)
104 >------->------->-------break;
105
106 >------->-------/*
107 >------->------- * We couldn't find space for the other segments;
erase the
108 >------->------- * kernel segment and try the next available hole.
109 >------->------- */
110 >------->-------image->nr_segments -= 1;
111 >------->-------kbuf.buf_min = kernel_segment->mem +
kernel_segment->memsz;
112 >------->-------kbuf.mem = KEXEC_BUF_MEM_UNKNOWN;
113 >-------}

Btw, I assume the above text is not related to the discussion, right?

The code closely related to my description above — the arm64
crash_load_dm_crypt_keys() — is exactly called within the
load_other_segments() in the while loop mentioned earlier. Therefore,
crash_load_dm_crypt_keys() may be called repeatedly before
kexec_file_post_load_cleanup_dm_crypt().

We can refer to the arm64 commit 108aa503657e ("arm64: kexec_file: try
more regions if loading segments fails") , which introduced the retry loop.

Thanks for the explanation and also pointing me to the reference commit!


Best regards,
Jinjie

[...]



--
Best regards,
Coiby