Kernel 5.15 returns empty content for topology/thread_siblings on
aarch64 platform, which in conjunction of uninitialized `buf` memory
buffer results in the garbage:
(gdb) p buf
$14 = " @\377\367\177\000\000\000\275\000\347j\032\236"
This garbage is then being later consumed by underlying helper functions
like for example cpumask_parse() and this leads to the following crash
later:
in __libc_free (p=0x7ff7f67c00) at src/malloc/mallocng/free.c:105
in free (p=<optimized out>) at src/malloc/free.c:5
in add_cpuset_to_array (setsize=<optimized out>, set=<optimized out>, items=<optimized out>, ary=<optimized out>) at ../sys-utils/lscpu-topology.c:29
in cputype_read_topology (cxt=cxt@entry=0x7ff7fffe70, ct=0x4298a0) at ../sys-utils/lscpu-topology.c:153
in lscpu_read_topology (cxt=cxt@entry=0x7ff7fffe70) at ../sys-utils/lscpu-topology.c:629
in main (argc=1, argv=0x7ffffffdb8) at ../sys-utils/lscpu.c:1341
It looks like the problem is that current logic expects fgets() to set
errno on failure, but fgets() is not documented to do so and and neither
glibc nor musl set errno. So if errno was set to 0 before fgets() call,
the failure from fgets() is ignored and then invalid buffer is being
parsed.
Fixes: #1810
Suggested-by: Thomas Weißschuh <thomas@t-8ch.de>
Signed-off-by: Petr Štetiar <ynezz@true.cz>
According to POSIX, readlink() makes no effort to null-terminate buffer
with the result. It seems better to hide this disadvantage in the
ul_path_...() API rather than assume buf[sz] = '\0' everywhere.
Reported-by: Reported-by: Jan Pazdziora <jpazdziora@redhat.com>
Signed-off-by: Karel Zak <kzak@redhat.com>
* add ul_path_stat()
* make sure all paths for ul_path_..() functions are always
interpreted relatively to the context directory and prefix. This is
difference between ul_path_ API and standard libc "at" functions. We
do not use any exception for absolute paths. The reason is that we
need to read from prefixed paths although application assume absolute
path (/dev/sda1 means /prefix/dev/sda1 if a /prefix is defined).
Signed-off-by: Karel Zak <kzak@redhat.com>
The goal is to avoid duplicate code in path.c and sysfs.c and make it
possible to define prefix for paths for all sysfs and procfs based
utils. Now we have /proc snapshots (for tests) for lscpu only. It
would be nice to have the same (for sysfs) for lsblk and another tools.
* very simple API to read numbers, strings and symlinks
* based on openat()
pc = ul_new_path("/sys/block/sda");
ul_path_read_u64(pc, &size, "size");
ul_path_read_u64(pc, &lsz, "queue/logical_block_size");
* printf-like API to generate paths, for example:
ul_path_readf_u64(pc, &num, "sda%d/size", partno)
* allow to define prefix to redirect hardcoded paths to another
location, for example:
pc = ul_new_path("/sys/block/sda");
ul_path_set_prefix(pc, "/my/regression/dump");
ul_path_read_u64(pc, &num, "size");
to read /my/regression/dump/sys/block/sda/size
* allow to extend the API by "dialects", for example for sysfs:
pc = ul_new_path(NULL);
sysfs_blkdev_init_path(pc, devno, NULL);
and use ul_path_* functions to read from @pc initialized by
sysfs_blkdev_init_path()
* add test_path binary
Signed-off-by: Karel Zak <kzak@redhat.com>
Simply avoiding strdup(). Error handling improved.
This was the Clang Analyzer warning:
Memory Error, Use-after-free
sys-utils/lsmem.c:259:3: warning: Use of memory after it is freed
err(EXIT_FAILURE, _("Failed to open %s"), path);
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Signed-off-by: Ruediger Meier <ruediger.meier@ga-group.nl>
Do not operate on truncated/random paths. Note, path_strdup()
can now really return NULL, to be handled in next commit.
Signed-off-by: Ruediger Meier <ruediger.meier@ga-group.nl>
