Re: [PATCH v3 bpf-next 01/10] lib/buildid: harden build ID parsing logic

On Wed, Aug 7, 2024 at 8:12 AM Jann Horn [off-list ref] wrote:

+Matthew and fsdevel list for pagecache question

On Tue, Jul 30, 2024 at 10:39 PM Andrii Nakryiko [off-list ref] wrote:

quoted

Harden build ID parsing logic, adding explicit READ_ONCE() where it's
important to have a consistent value read and validated just once.

Fixes tag below points to the code that moved this code into
lib/buildid.c, and then subsequently was used in perf subsystem, making
this code exposed to perf_event_open() users in v5.12+.

One thing that still seems dodgy to me with this patch applied is the
call from build_id_parse() to find_get_page(), followed by reading the
page contents. My understanding of the page cache (which might be
incorrect) is that find_get_page() can return a page whose contents
have not been initialized yet, and you're supposed to check for
PageUptodate() or something like that before reading from it.

Maybe Matthew can check if I understood that right?


Also, it might be a good idea to liberally spray READ_ONCE() around
all the remaining unannotated shared memory accesses in
build_id_parse(), get_build_id_32(), get_build_id_64() and
parse_build_id_buf().

Andi was against that, so I kept READ_ONCE() only where strictly
necessary, AFAICT.

quoted

Cc: stable@vger.kernel.org
Cc: Jann Horn <jannh@google.com>
Suggested-by: Andi Kleen <redacted>
Fixes: bd7525dacd7e ("bpf: Move stack_map_get_build_id into lib")
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
---
 lib/buildid.c | 51 +++++++++++++++++++++++++++------------------------
 1 file changed, 27 insertions(+), 24 deletions(-)

diff --git a/lib/buildid.c b/lib/buildid.c
index e02b5507418b..d21d86f6c19a 100644
--- a/lib/buildid.c
+++ b/lib/buildid.c

@@ -18,28 +18,29 @@ static int parse_build_id_buf(unsigned char *build_id,
                              const void *note_start,
                              Elf32_Word note_size)
 {
+       const char note_name[] = "GNU";
+       const size_t note_name_sz = sizeof(note_name);
        Elf32_Word note_offs = 0, new_offs;
+       u32 name_sz, desc_sz;
+       const char *data;

        while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
                Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);

+               name_sz = READ_ONCE(nhdr->n_namesz);
+               desc_sz = READ_ONCE(nhdr->n_descsz);
                if (nhdr->n_type == BUILD_ID &&
-                   nhdr->n_namesz == sizeof("GNU") &&
-                   !strcmp((char *)(nhdr + 1), "GNU") &&
-                   nhdr->n_descsz > 0 &&
-                   nhdr->n_descsz <= BUILD_ID_SIZE_MAX) {
-                       memcpy(build_id,
-                              note_start + note_offs +
-                              ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
-                              nhdr->n_descsz);
-                       memset(build_id + nhdr->n_descsz, 0,
-                              BUILD_ID_SIZE_MAX - nhdr->n_descsz);
+                   name_sz == note_name_sz &&
+                   strcmp((char *)(nhdr + 1), note_name) == 0 &&
+                   desc_sz > 0 && desc_sz <= BUILD_ID_SIZE_MAX) {
+                       data = note_start + note_offs + ALIGN(note_name_sz, 4);

I don't think we have any guarantee here that this addition won't
result in an OOB pointer?

quoted

+                       memcpy(build_id, data, desc_sz);

I think this can access OOB data (because "data" can already be OOB
and because "desc_sz" hasn't been checked against the amount of
remaining space in the page).

Andi already pointed this out and I fixed it locally, thanks.

quoted

+                       memset(build_id + desc_sz, 0, BUILD_ID_SIZE_MAX - desc_sz);
                        if (size)
-                               *size = nhdr->n_descsz;
+                               *size = desc_sz;
                        return 0;
                }
-               new_offs = note_offs + sizeof(Elf32_Nhdr) +
-                       ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
+               new_offs = note_offs + sizeof(Elf32_Nhdr) + ALIGN(name_sz, 4) + ALIGN(desc_sz, 4);
                if (new_offs <= note_offs)  /* overflow */
                        break;

You check whether "new_offs" has wrapped here, but then on the next
loop iteration, you check for "note_offs + sizeof(Elf32_Nhdr) <
note_size". So if new_offs is 0xffffffff at this point, then I think
the overflow check here will be passed, the loop condition will be
true on 32-bit kernels (on 64-bit kernels it won't be because the
addition happens on 64-bit numbers thanks to sizeof()), and "nhdr"
will point in front of the note?

Correct, and so I moved this new_offs calculation and overflow check
to the beginning of the loop, which I think should capture this issue.

For the while() condition itself I have:

        if (check_add_overflow(note_offs, note_size, &note_end))
                return -EINVAL;

        while (note_offs < note_end - sizeof(Elf32_Nhdr) - note_name_sz) {
            ...
        }

I'll try to post an updated version soon-ish, have been waiting for mm
folks feedback before posting a new version.

quoted

                note_offs = new_offs;