Stack overflow at write_one()

I have found a stack overflow at builtin/pack-objects.c:write_one(), where it calls itself endlessly. This is caused by the object_entry e and e->delta->delta being the same. But I have no idea how that happened.

First, the full story:

I used Google's repo tool to mirror AOSP to my machine. This mirrors several kernel trees (six last time I counted), without sharing objects one with another. To save space, I decided to point their objects/info/alternates to my mirror of the Linus kernel tree (which should be safe, since Linus makes it always fast-forward), and run "git gc" on them to create a smaller pack. This worked for all trees except one, where it core dumped (abrt report at https://bugzilla.redhat.com/show_bug.cgi?id=755132).

I compiled the latest git (v1.7.8-rc3-17-gf56ef11) to see if it still happened, and here is what I could get from gdb. I attached to the pack-objects process before it crashed (full command line "git pack-objects --keep-true-parents --honor-pack-keep --non-empty --all --reflog --unpack-unreachable --local --delta-base-offset /home/cesarb/src/bug755132/omap.git/objects/pack/.tmp-5171-pack"), continued, and let it crash:

(gdb) cont Continuing. [New Thread 0x7f3f2bad3700 (LWP 5205)] [New Thread 0x7f3f2b2d2700 (LWP 5206)] [New Thread 0x7f3f2aad1700 (LWP 5207)] [New Thread 0x7f3f2a2d0700 (LWP 5208)] [Thread 0x7f3f2b2d2700 (LWP 5206) exited] [Thread 0x7f3f2bad3700 (LWP 5205) exited] [Thread 0x7f3f2aad1700 (LWP 5207) exited] [Thread 0x7f3f2a2d0700 (LWP 5208) exited]

Program received signal SIGSEGV, Segmentation fault.
0x00000000004472b9 in write_one (f=0x6a97db0, e=0x7f3f30233490,
     offset=0x7fff79b53908) at builtin/pack-objects.c:415
415	{

Unlike on Fedora's git binary, where it happened on a call instruction, this time it happened on a push instruction:

(gdb) disassemble
Dump of assembler code for function write_one:
    0x00000000004472b0 <+0>:	push   %r15
    0x00000000004472b2 <+2>:	push   %r14
    0x00000000004472b4 <+4>:	push   %r13
    0x00000000004472b6 <+6>:	mov    %rdx,%r13
=> 0x00000000004472b9 <+9>:	push   %r12
    0x00000000004472bb <+11>:	mov    %rdi,%r12

The last few frames on the stack show the endless recursion:

(gdb) where
#0  0x00000000004472b9 in write_one (f=0x6a97db0, e=0x7f3f30233490,
     offset=0x7fff79b53908) at builtin/pack-objects.c:415
#1  0x00000000004472ed in write_one (f=0x6a97db0, e=0x7f3f30277390,
     offset=0x7fff79b53908) at builtin/pack-objects.c:423
#2  0x00000000004472ed in write_one (f=0x6a97db0, e=0x7f3f30233490,
     offset=0x7fff79b53908) at builtin/pack-objects.c:423
#3  0x00000000004472ed in write_one (f=0x6a97db0, e=0x7f3f30277390,
     offset=0x7fff79b53908) at builtin/pack-objects.c:423
#4  0x00000000004472ed in write_one (f=0x6a97db0, e=0x7f3f30233490,
     offset=0x7fff79b53908) at builtin/pack-objects.c:423

And here is the loop in the data structures:

(gdb) p e $1 = (struct object_entry *) 0x7f3f30233490 (gdb) p e->delta $2 = (struct object_entry *) 0x7f3f30277390 (gdb) p e->delta->delta $3 = (struct object_entry *) 0x7f3f30233490

Unfortunately, I do not know enough of git's internals to debug further. In case it helps, here is the contents of a few of these structures:

(gdb) p *e
$4 = {idx = {
     sha1 = "\257>J\241)\266\023\064\a\342J\320\375ӆ\262M\245", 
<incomplete sequence \356>, crc32 = 0, offset = 0}, size = 20, in_pack = 
0x259b610,
   in_pack_offset = 231061238, delta = 0x7f3f30277390,
   delta_child = 0x7f3f30277390, delta_sibling = 0x7f3f30413b10,
   delta_data = 0x0, delta_size = 20, z_delta_size = 0, hash = 2099915708,
   type = OBJ_OFS_DELTA, in_pack_type = OBJ_OFS_DELTA,
   in_pack_header_size = 5 '\005', preferred_base = 0 '\000',
   no_try_delta = 0 '\000', tagged = 0 '\000', filled = 1 '\001'}
(gdb) p *(e->delta)
$5 = {idx = {
     sha1 = 
"\372\307\035\372\017\350\307\f\310R\t\236\006\034\063N*T\216\253",
     crc32 = 0, offset = 0}, size = 14, in_pack = 0x259b610,
   in_pack_offset = 39990, delta = 0x7f3f30233490,
   delta_child = 0x7f3f30233490, delta_sibling = 0x0, delta_data = 0x0,
   delta_size = 14, z_delta_size = 0, hash = 2099915708, type = 
OBJ_REF_DELTA,
   in_pack_type = OBJ_REF_DELTA, in_pack_header_size = 21 '\025',
   preferred_base = 0 '\000', no_try_delta = 0 '\000', tagged = 0 '\000',
   filled = 1 '\001'}
(gdb) p *(e->in_pack)
$6 = {next = 0x25a53c0, windows = 0x259bc40, pack_size = 449155894,
   index_data = 0x7f3f4f0a9000, index_size = 58351420, num_objects = 
2083941,
   num_bad_objects = 0, bad_object_sha1 = 0x0, index_version = 2,
   mtime = 1321387261, pack_fd = -1, pack_local = 1, pack_keep = 0,
   do_not_close = 0, sha1 = 
"\371Q4\177.ȳv\364\246\332Z\234\025?\352ݠP\210",
   pack_name = 0x259b671 
"/home/cesarb/src/bug755132/omap.git/objects/pack/pack-f951347f2ec8b376f4a6da5a9c153feadda05088.pack"}

I tried using "git fsck" to see if it could find anything strange, but it seems to get stuck (using 100% CPU) after these lines:

[...] Checking commit fb630b9fc902e24209166b1659a8b375bf38099c Checking tree fc32c012c750084eb1d82782cee7c80a45a78289 Checking blob fc7bbba585cee2c2b0d5282c42fb986bfb032a0a Checking commit fdcb23634c9b6649bb02c681033d4973491b0e35 Checking tree fe773cf73ff553249be2f24ddf770f5dc43a41f1 Checking blob fe67b5c79f0ff33d92ebe7469a89c5a5d044fc0a Checking blob fe73276e026bf263f494a917c84c6a3fcaeaaeda Checking tree fe30eda9d92d074816f9c3a47fd3ffb9b89ca835 Checking tree fe9c75396e6d433b289d0e40c7e47921b91cad3a Checking blob ff3ed6086ce1c6b6b4b5111c034d14a208c0d045 Checking blob ff66638ff54d5ad7067e4f246d392059eef1a7bf Checking tree ff126d2bc67017199049ddba761979f3bda57eb9

Unfortunately, the reproducer I have (a copy of both trees with objects/info/alternates modified) is 1.8G in size, and I do not know how to create a smaller reproducer. If you know of a command which would get more relevant information from them, just ask; I plan on keeping them around for a while.

Already found the real cause (jGit bug) and workaround posted, I think.

See $gmane/185573

Em 19-11-2011 19:08, Junio C Hamano escreveu:

I presume the cause then is what was fixed by http://egit.eclipse.org/w/?p=jgit.git;a=commit;h=2fbf296fda205446eac11a13abd4fcdb182f28d9 ?

That did it, thanks! The patch had an offset, a fuzz, and a reject, but it was easy to fix by hand.

$ ../git/git gc Counting objects: 30254, done. Delta compression using up to 4 threads. Compressing objects: 100% (6614/6614), done. warning: recursive delta detected for object fac71dfa0fe8c70cc852099e061c334e2a548eab warning: recursive delta detected for object 1b730f5b2e0bdb2a2206af8ed30170509e75a2f5 warning: recursive delta detected for object 2f25a87e67fa3a226e367b9e080f11aa90c9f953 warning: recursive delta detected for object d5e5eefac91788da9a94efe9a15e0b928a77489e Writing objects: 100% (30254/30254), done. Total 30254 (delta 24008), reused 28803 (delta 23266)

And after that the repack does not break anymore:

$ ../git/git gc Counting objects: 30254, done. Delta compression using up to 4 threads. Compressing objects: 100% (5876/5876), done. Writing objects: 100% (30254/30254), done. Total 30254 (delta 24008), reused 30254 (delta 24008)

On Sat, Nov 19, 2011 at 13:46, Cesar Eduardo Barros <cesarb@cesarb.net> wrote:

Yes. The AOSP servers were all updated with the above JGit patch, so the servers are no longer sending duplicate objects. But yes, for a period of time there were duplicates in the kernel repositories, particularly kernal/omap.

Em 19-11-2011 21:30, Shawn Pearce escreveu:

So, would an alternative workaround in my situation be to delete kernel/omap.git and let repo sync recreate it? It seems repo does not have extra metadata anywhere else, so just removing the directory should be enough for it to clone again from scratch, hopefully getting a corrected pack from the server.

On Sat, Nov 19, 2011 at 16:02, Cesar Eduardo Barros <cesarb@cesarb.net> wrote:

Yes. repo does not have extra state, so just removing the directory and running `repo sync` again to clone the repository would correct the local repository.