[v3] binfmt_elf: Dump smaller VMAs first in ELF cores

[PATCH v3] binfmt_elf: Dump smaller VMAs first in ELF cores

Posted by Brian Mak 1 year, 4 months ago

Large cores may be truncated in some scenarios, such as with daemons
with stop timeouts that are not large enough or lack of disk space. This
impacts debuggability with large core dumps since critical information
necessary to form a usable backtrace, such as stacks and shared library
information, are omitted.

We attempted to figure out which VMAs are needed to create a useful
backtrace, and it turned out to be a non-trivial problem. Instead, we
try simply sorting the VMAs by size, which has the intended effect.

By sorting VMAs by dump size and dumping in that order, we have a
simple, yet effective heuristic.

Signed-off-by: Brian Mak <makb@juniper.net>
---

Hi all,

Still need to run rr tests on this, per Kees Cook's suggestion, will
update back once done. GDB and readelf show that this patch works
without issue though.

Thanks,
Brian Mak

v3: Edited commit message to better convey alternative solution as
    non-trivial

    Moved sorting logic to fs/coredump.c to make it in place

    Above edits suggested by Eric Biederman <ebiederm@xmission.com>

v2: Edited commit message to include more reasoning for sorting VMAs
    
    Removed conditional VMA sorting with debugfs knob
    
    Above edits suggested by Eric Biederman <ebiederm@xmission.com>

 fs/coredump.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/fs/coredump.c b/fs/coredump.c
index 7f12ff6ad1d3..33c5ac53ab31 100644
--- a/fs/coredump.c
+++ b/fs/coredump.c
@@ -18,6 +18,7 @@
 #include <linux/personality.h>
 #include <linux/binfmts.h>
 #include <linux/coredump.h>
+#include <linux/sort.h>
 #include <linux/sched/coredump.h>
 #include <linux/sched/signal.h>
 #include <linux/sched/task_stack.h>
@@ -1191,6 +1192,18 @@ static void free_vma_snapshot(struct coredump_params *cprm)
 	}
 }
 
+static int cmp_vma_size(const void *vma_meta_lhs_ptr, const void *vma_meta_rhs_ptr)
+{
+	const struct core_vma_metadata *vma_meta_lhs = vma_meta_lhs_ptr;
+	const struct core_vma_metadata *vma_meta_rhs = vma_meta_rhs_ptr;
+
+	if (vma_meta_lhs->dump_size < vma_meta_rhs->dump_size)
+		return -1;
+	if (vma_meta_lhs->dump_size > vma_meta_rhs->dump_size)
+		return 1;
+	return 0;
+}
+
 /*
  * Under the mmap_lock, take a snapshot of relevant information about the task's
  * VMAs.
@@ -1253,5 +1266,8 @@ static bool dump_vma_snapshot(struct coredump_params *cprm)
 		cprm->vma_data_size += m->dump_size;
 	}
 
+	sort(cprm->vma_meta, cprm->vma_count, sizeof(*cprm->vma_meta),
+		cmp_vma_size, NULL);
+
 	return true;
 }

base-commit: eb5e56d1491297e0881c95824e2050b7c205f0d4
-- 
2.25.1

[PATCH v3] binfmt_elf: Dump smaller VMAs first in ELF cores

Posted by Michael Stapelberg 10 months ago

Hey Brian and folks

> […]
> backtrace, and it turned out to be a non-trivial problem. Instead, we
> try simply sorting the VMAs by size, which has the intended effect.
> […]
> Still need to run rr tests on this, per Kees Cook's suggestion, will
> update back once done. GDB and readelf show that this patch works
> without issue though.

I think in your testing, you probably did not try the eu-stack tool
from the elfutils package, because I think I found a bug:

Current elfutils cannot symbolize core dumps created by Linux 6.12+.
I noticed this because systemd-coredump(8) uses elfutils, and when
a program crashed on my machine, syslog did not show function names.

I reported this issue with elfutils at:
https://sourceware.org/bugzilla/show_bug.cgi?id=32713
…but figured it would be good to give a heads-up here, too.

Is this breakage sufficient reason to revert the commit?
Or are we saying userspace just needs to be updated to cope?

Thanks
Best regards
Michael