libbpf: harden parse_vma_segs() path parsing
PATCH v4 1/1 libbpf: harden parse_vma_segs() path parsing
parse_vma_segs() in tools/lib/bpf/usdt.c parses /proc/<pid>/maps with two widthless scansets, "%s" into mode[16] and "%[^\n]" into line[4096]. A VMA name in maps is not limited to that local buffer; a deeply nested backing path can produce a maps record long enough to overflow the stack buffer.
Bound both scansets to the declared buffer sizes ("%15s" for mode[16] and "%4095[^\n]" for line[4096]) and drain any residue past line[4094] with "%*[^\n]" before the trailing "\n". Without the drain, the residue of an over-long record would stay in the stream and break the next "%zx-%zx" parse, so the loop would exit early and silently skip later maps records.
Also stop using sscanf(..., "%s") to peel the /proc/<pid>/root prefix from lib_path. Parse the pid and prefix length with "%n", check for the following slash, and copy the remainder with libbpf_strlcpy(). That removes a second unbounded stack write and preserves paths containing spaces.
---
v4:
- Carry Emil's Reviewed-by.
- Simplify the /proc/<pid>/root prefix handling with sscanf() %n, removing the unreachable snprintf() length check.
- Initialize and check the %n output before using it, so partial literal matches after the pid cannot use an unassigned offset.
- Add a short comment for the %n return-value rule.
- Declare the maps-line buffer as line[4096] to match the %4095 scanset width.
- Reword the maps-line comment without seq_file implementation detail.
v3:
- Correct Fixes tag to the initial USDT implementation commit, per BPF CI review after adding second site.
v2:
- Replace the unbounded /proc/<pid>/root sscanf() path peeling with bounded prefix handling, addressing review feedback on v1 and preserving paths containing spaces.
- Keep the v1 maps parser fix using bounded fscanf() scansets and a suppressed scanset drain for over-long records.
- Re-ran real parse_vma_segs() ASAN harnesses for the original maps overflow, the proc-root overflow, proc-root paths with spaces, and adjacent successful parses after an over-long maps record.
Reproduced with Debian 12 on rootless podman: an unprivileged container process mkdirs 50 nested 200-char directories and mmaps a file at the bottom, producing a >10KB /proc/<host-pid>/maps line. A harness on the host then calls the real parse_vma_segs() against the container's PID; libbpf is built with -fsanitize=address and the only local source change is dropping the "static" keyword on parse_vma_segs so the symbol is linkable from the harness.
Stock libbpf reports:
==ERROR: AddressSanitizer: stack-buffer-overflow WRITE of size >10KB at <line> thread T0
#0 scanf_common -> #1 __isoc99_fscanf
#3 parse_vma_segs tools/lib/bpf/usdt.c:509 Address ... in frame parse_vma_segs at offset 8512, just past line[PATH_MAX].
Patched libbpf parses the same maps cleanly. Follow-up calls return 0 with seg_cnt > 0 for libc.so.6 and for ld-linux-x86-64.so.2 (format drain), which appears in maps after the over-long entry.
On normal hardened builds the stack canary aborts the consumer; on builds without stack protector the bytes past line[] are attacker-influenced path bytes.
Selftest gate
=============
tools/testing/selftests/bpf/test_progs -t usdt under QEMU x86_64 (KVM) on the patched kernel: all 6 subtests pass (usdt/basic, basic_optimized, optimized_attach, multispec, urand_auto_attach, urand_pid_attach) on both stock and patched libbpf, diff-clean. The in-tree selftest does not itself exercise long maps records.
tools/lib/bpf/usdt.c | 16 ++++++++++++---- 1 file changed, 12 insertions(+), 4 deletions(-)
tools/lib/bpf/usdt.c +12 -4
@@ -468,10 +468,10 @@ static int parse_elf_segs(Elf *elf, const char *path, struct elf_seg **segs, sizstatic int parse_vma_segs (int pid,const char *lib_path,struct elf_seg **segs,size_t *seg_cnt) { -char path[PATH_MAX ], line[PATH_MAX ], mode[16 ]; +char path[PATH_MAX ], line[4096 ], mode[16 ];size_t seg_start, seg_end, seg_off;struct elf_seg *seg; -int tmp_pid, i, err; +int tmp_pid, n, i, err;FILE *f; *seg_cnt =0 ;
@@ -480,8 +480,13 @@ static int parse_vma_segs(int pid, const char *lib_path, struct elf_seg **segs,* /proc/<pid>/root/<path>. They will be reported as just /<path> in * /proc/<pid>/maps. */ -if (sscanf (lib_path,"/proc/%d/root%s" , &tmp_pid, path) ==2 && pid == tmp_pid) +/* %n is not counted in sscanf() return value, so initialize it. */ + n =0 ; +if (sscanf (lib_path,"/proc/%d/root%n" , &tmp_pid, &n) ==1 && + n >0 && pid == tmp_pid && lib_path[n] =='/' ) { +libbpf_strlcpy (path, lib_path + n,sizeof (path)); goto proceed; + }if (!realpath (lib_path, path)) { pr_warn("usdt: failed to get absolute path of '%s' (err %s), using path as is...\n",
@@ -504,8 +509,11 @@ static int parse_vma_segs(int pid, const char *lib_path, struct elf_seg **segs,* 7f5c6f5d1000-7f5c6f5d3000 rw-p 001c7000 08:04 21238613 /usr/lib64/libc-2.17.so * 7f5c6f5d3000-7f5c6f5d8000 rw-p 00000000 00:00 0 * 7f5c6f5d8000-7f5c6f5d9000 r-xp 00000000 103:01 362990598 /data/users/andriin/linux/tools/bpf/usdt/libhello_usdt.so +* +* Some VMA names can be longer than the local buffer. Bound the +* writes, but still consume the rest of the line. */ -while (fscanf (f,"%zx-%zx %s %zx %*s %*d%[^\n]\n" , +while (fscanf (f,"%zx-%zx %15s %zx %*s %*d%4095[^\n]%*[^\n]\n" , &seg_start, &seg_end, mode, &seg_off, line) ==5 ) {void *tmp;
Signature · 1 line
2.53.0
Andrii Nakryiko
AppliedQuoted 14 lines · Michael Bommarito
parse_vma_segs()intools/lib/bpf/usdt.cparses /proc/<pid>/maps with two widthless scansets, "%s" into mode[16] and "%[^\n]" into line[4096]. A VMA name in maps is not limited to that local buffer; a deeply nested backing path can produce a maps record long enough to overflow the stack buffer.Bound both scansets to the declared buffer sizes ("%15s" for mode[16] and "%4095[^\n]" for line[4096]) and drain any residue past line[4094] with "%*[^\n]" before the trailing "\n". Without the drain, the residue of an over-long record would stay in the stream and break the next "%zx-%zx" parse, so the loop would exit early and silently skip later maps records.
Quoted 8 lines
Also stop using
sscanf(..., "%s")to peel the /proc/<pid>/root prefix fromlib_path. Parse the pid and prefix length with "%n", check for the following slash, and copy the remainder withlibbpf_strlcpy(). That removes a second unbounded stack write and preserves paths containing spaces.Fixes 74cc6311cec9 ("libbpf: Add USDT notes parsing and resolution logic")
dropped cc: stable, libbpf is packaged through github repo and in-kernel libbpf usage doesn't use USDTs for anything, there is nothing to backport
applied to bpf-next, thanks!
Quoted 19 lines
Reviewed-by Emil Tsalapatis <[email protected]>Assisted-by Claude:claude-opus-4-7Signed-off-by Michael Bommarito <[email protected]>---
v4:
- Carry Emil's Reviewed-by.
- Simplify the /proc/<pid>/root prefix handling withsscanf()%n, removing the unreachablesnprintf()length check.
- Initialize and check the %n output before using it, so partial literal matches after the pid cannot use an unassigned offset.
- Add a short comment for the %n return-value rule.
- Declare the maps-line buffer as line[4096] to match the %4095 scanset width.
- Reword the maps-line comment withoutseq_fileimplementation detail.v3:
- Correct Fixes tag to the initial USDT implementation commit, per BPF CI review after adding second site.
[...]