[Linux] Read vDSO range from /proc/PID/task/PID/maps instead of /proc/PID/maps

... as it's _much_ faster.

Hacking the gdb.threads/attach-many-short-lived-threads.exp test to
spawn thousands of threads instead of dozens to stress and debug
timeout problems with gdb.threads/attach-many-short-lived-threads.exp,
I saw that GDB would spend several seconds just reading the
/proc/PID/smaps file, to determine the vDSO mapping range.  GDB opens
and reads the whole file just once, and caches the result, but even
that is too slow.  For example, with almost 8000 threads:

 $ ls /proc/3518/task/ | wc -l
 7906

reading the /proc/PID/smaps file grepping for "vdso" takes over 15
seconds :

 $ time cat /proc/3518/smaps | grep vdso
 7ffdbafee000-7ffdbaff0000 r-xp 00000000 00:00 0                          [vdso]

 real    0m15.371s
 user    0m0.008s
 sys     0m15.017s

Looking around the web for hints, I found a nice description of the
issue here:

 http://backtrace.io/blog/blog/2014/11/12/large-thread-counts-and-slow-process-maps/

The problem is that /proc/PID/smaps wants to show the mappings as
being thread stack, and that has the kernel iterating over all threads
in the thread group, for each mapping.

The fix is to use the "map" file under /proc/PID/task/PID/ instead of
the /proc/PID/ one, as the former doesn't mark thread stacks for all
threads.

That alone drops the timing to the millisecond range on my machine:

 $ time cat /proc/3518/task/3518/smaps | grep vdso
 7ffdbafee000-7ffdbaff0000 r-xp 00000000 00:00 0                          [vdso]

 real    0m0.150s
 user    0m0.009s
 sys     0m0.084s

And since we only need the vdso mapping's address range, we can use
"maps" file instead of "smaps", and it's even cheaper:

/proc/PID/task/PID/maps :

 $ time cat /proc/3518/task/3518/maps | grep vdso
 7ffdbafee000-7ffdbaff0000 r-xp 00000000 00:00 0                          [vdso]

 real    0m0.027s
 user    0m0.000s
 sys     0m0.017s

gdb/ChangeLog:
2016-05-24  Pedro Alves  <palves@redhat.com>

	PR gdb/19828
	* linux-tdep.c (find_mapping_size): Delete.
	(linux_vsyscall_range_raw): Rewrite reading from
	/proc/PID/task/PID/maps directly instead of using
	gdbarch_find_memory_regions.
This commit is contained in:
Pedro Alves 2016-05-24 14:47:56 +01:00
parent aa01bd3689
commit 95e94c3f18
2 changed files with 62 additions and 23 deletions

View File

@ -1,3 +1,11 @@
2016-05-24 Pedro Alves <palves@redhat.com>
PR gdb/19828
* linux-tdep.c (find_mapping_size): Delete.
(linux_vsyscall_range_raw): Rewrite reading from
/proc/PID/task/PID/maps directly instead of using
gdbarch_find_memory_regions.
2016-05-24 Pedro Alves <palves@redhat.com>
PR gdb/19828

View File

@ -2277,39 +2277,70 @@ linux_gdb_signal_to_target (struct gdbarch *gdbarch,
return -1;
}
/* Rummage through mappings to find a mapping's size. */
static int
find_mapping_size (CORE_ADDR vaddr, unsigned long size,
int read, int write, int exec, int modified,
void *data)
{
struct mem_range *range = (struct mem_range *) data;
if (vaddr == range->start)
{
range->length = size;
return 1;
}
return 0;
}
/* Helper for linux_vsyscall_range that does the real work of finding
the vsyscall's address range. */
static int
linux_vsyscall_range_raw (struct gdbarch *gdbarch, struct mem_range *range)
{
char filename[100];
long pid;
char *data;
/* Can't access /proc if debugging a core file. */
if (!target_has_execution)
return 0;
/* We need to know the real target PID to access /proc. */
if (current_inferior ()->fake_pid_p)
return 0;
if (target_auxv_search (&current_target, AT_SYSINFO_EHDR, &range->start) <= 0)
return 0;
/* This is installed by linux_init_abi below, so should always be
available. */
gdb_assert (gdbarch_find_memory_regions_p (target_gdbarch ()));
pid = current_inferior ()->pid;
range->length = 0;
gdbarch_find_memory_regions (gdbarch, find_mapping_size, range);
return 1;
/* Note that reading /proc/PID/task/PID/maps (1) is much faster than
reading /proc/PID/maps (2). The later identifies thread stacks
in the output, which requires scanning every thread in the thread
group to check whether a VMA is actually a thread's stack. With
Linux 4.4 on an Intel i7-4810MQ @ 2.80GHz, with an inferior with
a few thousand threads, (1) takes a few miliseconds, while (2)
takes several seconds. Also note that "smaps", what we read for
determining core dump mappings, is even slower than "maps". */
xsnprintf (filename, sizeof filename, "/proc/%ld/task/%ld/maps", pid, pid);
data = target_fileio_read_stralloc (NULL, filename);
if (data != NULL)
{
struct cleanup *cleanup = make_cleanup (xfree, data);
char *line;
char *saveptr = NULL;
for (line = strtok_r (data, "\n", &saveptr);
line != NULL;
line = strtok_r (NULL, "\n", &saveptr))
{
ULONGEST addr, endaddr;
const char *p = line;
addr = strtoulst (p, &p, 16);
if (addr == range->start)
{
if (*p == '-')
p++;
endaddr = strtoulst (p, &p, 16);
range->length = endaddr - addr;
do_cleanups (cleanup);
return 1;
}
}
do_cleanups (cleanup);
}
else
warning (_("unable to open /proc file '%s'"), filename);
return 0;
}
/* Implementation of the "vsyscall_range" gdbarch hook. Handles