Thread (8 messages) 8 messages, 2 authors, 11h ago

Re: [PATCH v6 2/2] ring-buffer: Improve nr_pages type

From: Vincent Donnefort <hidden>
Date: 2026-08-17 10:14:52
Also in: sashiko-reviews

On Fri, Aug 14, 2026 at 03:59:44PM +0000, sashiko-bot@kernel.org wrote:
quoted hunk ↗ jump to hunk
Thank you for your contribution! Sashiko AI review found 4 potential issue(s) to consider:

New issues:
- [High] When `nr_pages` is exactly `U32_MAX - 1`, the expression `meta->nr_subbufs + 1` overflows to 0, silently skipping persistent ring buffer validation and wiping data.

Pre-existing issues:
- [Critical] Loop counters iterating over `meta->nr_subbufs` are typed as `signed int`, which wrap to negative values for extremely large buffers, leading to out-of-bounds memory accesses.
- [Critical] The `meta->nr_subbufs` field is read from potentially corrupted persistent memory and used without validation against `nr_pages`, causing a heap out-of-bounds write in `bitmap_clear()`.
- [High] Multiplication of `subbuf_size` and `idx` overflows 32 bits for ring buffers > 4GB, causing severe memory aliasing.
--
--- Patch [2]: [PATCH v6 2/2] ring-buffer: Improve nr_pages type ---
commit 030ec6a7ceec6e71123077bd1fc64e0cf5c27cc8
Author: Vincent Donnefort [off-list ref]

ring-buffer: Improve nr_pages type

If ring_buffer_per_cpu::nr_pages is defined as unsigned long, it is
capped to 32-bits in a few places, limiting the operations possible on a
very large buffer.

[ ... ]
quoted
diff --git a/kernel/trace/ring_buffer.c b/kernel/trace/ring_buffer.c
index ec13779922ffb..ec127e2ad0527 100644
--- a/kernel/trace/ring_buffer.c
+++ b/kernel/trace/ring_buffer.c
[ ... ]
quoted
@@ -1746,8 +1745,8 @@ static void *rb_subbufs_from_meta(struct ring_buffer_cpu_meta *meta)
 static void *rb_range_buffer(struct ring_buffer_per_cpu *cpu_buffer, int idx)
 {
 	struct ring_buffer_cpu_meta *meta;
+	unsigned int subbuf_size;
 	unsigned long ptr;
-	int subbuf_size;
 
 	meta = rb_range_meta(cpu_buffer->buffer, 0, cpu_buffer->cpu);
 	if (!meta)
[Severity: High]
This is a pre-existing issue, but does the multiplication of subbuf_size and
idx overflow 32 bits for ring buffers larger than 4GB?

In rb_range_buffer(), we have:

    subbuf_size = meta->subbuf_size;
    idx = meta->buffers[idx];
    ptr = (unsigned long)rb_subbufs_from_meta(meta);
    ptr += subbuf_size * idx;

Since subbuf_size is an unsigned int and idx is an int, the multiplication
evaluates as a 32-bit unsigned integer. For ring buffers over 4GB, this can
overflow and wrap around before being zero-extended and added to the 64-bit
ptr, causing memory aliasing by overwriting earlier subbuffers.
I can add this to this patch/
[ ... ]
quoted
@@ -1840,10 +1839,10 @@ static bool rb_meta_init(struct trace_buffer *buffer, int scratch_size)
  * must be the same.
  */
 static bool rb_cpu_meta_valid(struct ring_buffer_cpu_meta *meta, int cpu,
-			      struct trace_buffer *buffer, int nr_pages,
+			      struct trace_buffer *buffer, unsigned long nr_pages,
 			      unsigned long *subbuf_mask)
 {
-	int subbuf_size = PAGE_SIZE;
+	unsigned int subbuf_size = PAGE_SIZE;
 	unsigned long buffers_start;
 	unsigned long buffers_end;
 	int i;
[Severity: Critical]
This is a pre-existing issue, but will loop counters iterating over
meta->nr_subbufs wrap to negative values for extremely large buffers?

In rb_cpu_meta_valid():

    for (i = 0; i < meta->nr_subbufs; i++) {
        if (meta->buffers[i] < 0 || ...

And in rb_range_meta_init():

    for (i = 0; i < meta->nr_subbufs; i++) {
        meta->buffers[i] = i;

The loop counter i is a signed int. Since nr_pages can now go up to
U32_MAX - 1, if the buffer size pushes meta->nr_subbufs past INT_MAX, i
will wrap to negative. i < meta->nr_subbufs will promote i to unsigned,
evaluating to true, and the negative index will be sign-extended on 64-bit
architectures, causing out-of-bounds reads and writes.
Yeah sure, I'll make it unsigned int... doesn't sound too critical though!
[Severity: Critical]
This is a pre-existing issue, but could meta->nr_subbufs cause a heap
out-of-bounds write if it doesn't match nr_pages?

In rb_cpu_meta_valid():

    bitmap_clear(subbuf_mask, 0, meta->nr_subbufs);

    for (i = 0; i < meta->nr_subbufs; i++) {
        ...
        set_bit(meta->buffers[i], subbuf_mask);

meta->nr_subbufs is read directly from persistent memory and used without
being validated against the dynamically allocated subbuf_mask size, which is
based on nr_pages + 1. If meta->nr_subbufs is maliciously or accidentally set
larger than nr_pages + 1, bitmap_clear() and set_bit() will write zeroes and
ones past the end of the bitmap allocation.
Sure, but that'll be another patch.
[ ... ]
quoted
@@ -2837,6 +2837,10 @@ static struct trace_buffer *alloc_buffer(unsigned long size, unsigned flags,
 		 */
 		nr_pages = (size - sizeof(struct ring_buffer_cpu_meta)) /
 			(subbuf_size + sizeof(int));
+
+		/* limited by ring_buffer_cpu_meta::nr_subbufs */
+		if (nr_pages > U32_MAX - 1)
+			goto fail_free_buffers;
 		/* Need at least two pages plus the reader page */
 		if (nr_pages < 3)
 			goto fail_free_buffers;
[Severity: High]
Does meta->nr_subbufs + 1 overflow when nr_pages is exactly U32_MAX - 1?

If nr_pages is exactly U32_MAX - 1, meta->nr_subbufs becomes U32_MAX.
In rb_meta_validate_events(), there are loops with this condition:

    for (i = 0; i < meta->nr_subbufs + 1; i++, rb_dec_page(&head_page)) {

U32_MAX + 1 overflows to 0 in 32-bit unsigned arithmetic. This makes the loop
condition i < 0, causing the loops to terminate immediately without rewinding
and validating the persistent ring buffer from a previous boot.
Sure, this can be  unsigned it as well...

And there are plenty of other functions that need fixing then.
rb_meta_subbuf_idx, rb_cpu_meta_valid, __rb_get_reader_page, rb_iter_peek... 

I need a bit more time to audit what must be modified.

-- 
Vincent
-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260814154823.755406-1-vdonnefort@google.com?part=2
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help