object-name: avoid use-after-free in get_oid_with_context_1()
PATCH 1/1 object-name: avoid use-after-free in get_oid_with_context_1()
When a ":<path>" argument names a relative path, resolve_relative_path() returns a newly allocated string and "cp" is pointed at it:
new_path = resolve_relative_path(repo, cp);
if (!new_path) {
namelen = namelen - (cp - name);
} else {
cp = new_path;
namelen = strlen(cp);
}From there on "cp" and "new_path" name the same allocation. Later the memory location that "new_path" points to is freed.
free(new_path); if (reject_tree_in_index(repo, only_to_die, ce, stage, prefix, cp))
But here the reject_tree_in_index() passes "cp" to diagnose_invalid_index_path(), which calls strlen() on it, looks it up in the index, and formats it into its messages, allocating as it goes. All of this reads memory that has already been freed.
Collapse the two exits into one to ensure a single free() that happens after the last use.
Three things have to coincide to reach this:
1. The path has to be relative, or nothing is allocated and "cp" still points into the argument.
2. The entry found has to be a sparse
directory, which needs a sparse index.
3. The argument has to get past the check in die_verify_filename() that skips a leading ':' followed by a non-alphanumeric, so ":0:./dir/" arrives here where ":./dir/" does not.
Add a test to t1092 that covers the combination. It fails under SANITIZE=address without the change to object-name.c.
This was reported in [1], and the shape used here was suggested in review [2], but that series was not rerolled and the fix never landed.
[1] https://lore.kernel.org/git/cf6bcdb43e5b4abab464c30a914d64dc8e7a9925.1655336146.git.gitgitgadget@gmail.com/
[2] https://lore.kernel.org/git/[email protected]/
---
The three conditions make this awkward to reach by hand, so here is the recipe:
git init sparse && cd sparse && mkdir folder1 folder2 && echo a >folder1/a && echo b >folder2/b && git add -A && git commit -m init && git sparse-checkout init --cone --sparse-index && git sparse-checkout set folder1 && git show :0:./folder2/
Without the change below, no sanitizer is needed to see it. On 2.52.0 the buffer has already been reused by the time the message is formatted, so the path printed is whatever now sits in that memory, and it differs from run to run:
fatal: path '' does not exist (neither on disk nor in the index)
fatal: path 'M-6?:xM-@M-:M-L??X' does not exist (neither on disk nor in the index)
fatal: path '?M-*JM-^M->M-YM-tn?H' does not exist (neither on disk nor in the index)
Still without the change, built with SANITIZE=address, the same command reports
ERROR: AddressSanitizer: heap-use-after-free READ of size 3 at 0x607000002a20
#1 diagnose_invalid_index_path object-name.c:1653
#2 get_oid_with_context_1 object-name.c:1807
#3 maybe_die_on_misspelt_object_name
#4 die_verify_filename setup.c:216
#6 setup_revisions revision.c:3103
#8 cmd_show log.c:694
freed by thread T0 here:
#1 get_oid_with_context_1 object-name.c:1806
previously allocated by thread T0 here:
#5 prefix_path setup.c:149
#6 get_oid_with_context_1 object-name.c:1784
With the change, the message reads "folder2/" every time and the sanitizer stays quiet.
"git diff" and "git rev-parse" reach it the same way, and so does "../" from a subdirectory.
The new t1092 test fails without the object-name.c hunk and passes with it, under SANITIZE=address.
object-name.c | 15 +++++++++------ t/t1092-sparse-checkout-compatibility.sh | 11 +++++++++++ 2 files changed, 20 insertions(+), 6 deletions(-)
object-name.c +9 -6
@@ -1803,13 +1803,16 @@ static enum get_oid_result get_oid_with_context_1(struct repository *repo, memcmp(ce->name, cp, namelen))break ;if (ce_stage (ce) == stage) { +int ret =-1 ; + +if (!reject_tree_in_index (repo, only_to_die, ce, + stage, prefix, cp)) { +oidcpy (oid, &ce->oid); + oc->mode = ce->ce_mode; + ret =0 ; + }free (new_path); -if (reject_tree_in_index (repo, only_to_die, ce, - stage, prefix, cp)) -return -1 ; -oidcpy (oid, &ce->oid); - oc->mode = ce->ce_mode; -return 0 ; +return ret; } pos++; }
t/t1092-sparse-checkout-compatibility.sh +11 -0
@@ -1357,6 +1357,17 @@ do " done +test_expect_success 'relative path to a sparse directory' ' + init_repos && + + # A ":<stage>:<path>" argument whose path is relative is resolved + # into a heap-allocated buffer, and a sparse directory found at that + # path is reported through it. Cover that combination, so that the + # reporting does not read the buffer after it has been released. + test_sparse_match test_must_fail git show :0:./folder1/ && + test_sparse_match test_must_fail git rev-parse :0:./folder1/ +' + test_expect_success 'submodule handling' ' init_repos &&
Signature · 1 line
2.52.0
Quoted 37 lines · Shlok Kulshreshtha
When a ":<path>" argument names a relative path,
resolve_relative_path()returns a newly allocated string and "cp" is pointed at it:new_path = resolve_relative_path(repo, cp); if (!new_path) { namelen = namelen - (cp - name); } else { cp = new_path; namelen = strlen(cp); }From there on "cp" and "
new_path" name the same allocation. Later the memory location that "new_path" points to is freed.free(new_path); if (reject_tree_in_index(repo, only_to_die, ce, stage, prefix, cp))But here the
reject_tree_in_index()passes "cp" todiagnose_invalid_index_path(), which callsstrlen()on it, looks it up in the index, and formats it into its messages, allocating as it goes. All of this reads memory that has already been freed.Collapse the two exits into one to ensure a single
free()that happens after the last use.Three things have to coincide to reach this:
1. The path has to be relative, or nothing is allocated and "cp" still points into the argument.
2. The entry found has to be a sparse
directory, which needs a sparse index.3. The argument has to get past the check in
die_verify_filename()that skips a leading ':' followed by a non-alphanumeric, so ":0:./dir/" arrives here where ":./dir/" does not.
Quoted 8 lines
Add a test to t1092 that covers the combination. It fails under SANITIZE=address without the change to
object-name.c.This was reported in [1], and the shape used here was suggested in review [2], but that series was not rerolled and the fix never landed.
[1] https://lore.kernel.org/git/cf6bcdb43e5b4abab464c30a914d64dc8e7a9925.1655336146.git.gitgitgadget@gmail.com/
[2] https://lore.kernel.org/git/[email protected]/
Oh, from 2022, good find.
Quoted 1 line
Reported-by Johannes Schindelin <[email protected]>
Original-patch-by even, no?
Quoted 66 lines
Helped-by Junio C Hamano <[email protected]>Signed-off-by Shlok Kulshreshtha <[email protected]>---
The three conditions make this awkward to reach by hand, so here is the recipe:git init sparse && cd sparse && mkdir folder1 folder2 && echo a >folder1/a && echo b >folder2/b && git add -A && git commit -m init && git sparse-checkout init --cone --sparse-index && git sparse-checkout set folder1 && git show :0:./folder2/Without the change below, no sanitizer is needed to see it. On 2.52.0 the buffer has already been reused by the time the message is formatted, so the path printed is whatever now sits in that memory, and it differs from run to run:
fatal: path '' does not exist (neither on disk nor in the index)
fatal: path 'M-6?:xM-@M-:M-L??X' does not exist (neither on disk nor in the index)
fatal: path '?M-*JM-^M->M-YM-tn?H' does not exist (neither on disk nor in the index)Still without the change, built with SANITIZE=address, the same command reports
ERROR: AddressSanitizer: heap-use-after-free READ of size 3 at
0x607000002a20
#1diagnose_invalid_index_pathobject-name.c:1653
#2get_oid_with_context_1object-name.c:1807
#3maybe_die_on_misspelt_object_name
#4die_verify_filenamesetup.c:216
#6setup_revisionsrevision.c:3103
#8cmd_showlog.c:694
freed by thread T0 here:
#1get_oid_with_context_1object-name.c:1806
previously allocated by thread T0 here:
#5prefix_pathsetup.c:149
#6get_oid_with_context_1object-name.c:1784With the change, the message reads "folder2/" every time and the sanitizer stays quiet.
"git diff" and "git rev-parse" reach it the same way, and so does "../" from a subdirectory.
The new t1092 test fails without the
object-name.chunk and passes with it, under SANITIZE=address.object-name.c | 15 +++++++++------ t/t1092-sparse-checkout-compatibility.sh | 11 +++++++++++ 2 files changed, 20 insertions(+), 6 deletions(-)object-name.c +6
-0@@ -1803,13 +1803,16 @@ static enum get_oid_result get_oid_with_context_1(struct repository *repo, memcmp(ce->name, cp, namelen))break ;if (ce_stage (ce) == stage) { +int ret =-1 ; + +if (!reject_tree_in_index (repo, only_to_die, ce, + stage, prefix, cp)) { +oidcpy (oid, &ce->oid); + oc->mode = ce->ce_mode;
object-name.c static enum get_oid_result get_oid_with_context_1(struct repository *repo, · line 1815
Quoted 10 lines
object-name.c +3
-6+ ret =0 ; + }free (new_path); -if (reject_tree_in_index (repo, only_to_die, ce, - stage, prefix, cp)) -return -1 ; -oidcpy (oid, &ce->oid); - oc->mode = ce->ce_mode; -return 0 ; +return ret;
OK
Quoted 15 lines
object-name.c +0
-0} pos++; }t/t1092-sparse-checkout-compatibility.sh +4
-0@@ -1357,6 +1357,17 @@ do " done +test_expect_success 'relative path to a sparse directory' ' + init_repos && + + # A ":<stage>:<path>" argument whose path is relative is resolved
t/t1092-sparse-checkout-compatibility.sh do · line 1373
Quoted 10 lines
t/t1092-sparse-checkout-compatibility.sh +7
-0+ # into a heap-allocated buffer, and a sparse directory found at that + # path is reported through it. Cover that combination, so that the + # reporting does not read the buffer after it has been released. + test_sparse_match test_must_fail git show :0:./folder1/ && + test_sparse_match test_must_fail git rev-parse :0:./folder1/ +' + test_expect_success 'submodule handling' ' init_repos &&
Good idea to add a test.
René
Quoted 1 line · René Scharfe
Original-patch-by even, no?
Yes, that's more accurate I guess. I'll update this in a v2.
Thanks for the review.
Shlok
PATCH v2 1/1 object-name: avoid use-after-free in get_oid_with_context_1()
When a ":<path>" argument names a relative path, resolve_relative_path() returns a newly allocated string and "cp" is pointed at it:
new_path = resolve_relative_path(repo, cp);
if (!new_path) {
namelen = namelen - (cp - name);
} else {
cp = new_path;
namelen = strlen(cp);
}From there on "cp" and "new_path" name the same allocation. Later the memory location that "new_path" points to is freed.
free(new_path); if (reject_tree_in_index(repo, only_to_die, ce, stage, prefix, cp))
But here the reject_tree_in_index() passes "cp" to diagnose_invalid_index_path(), which calls strlen() on it, looks it up in the index, and formats it into its messages, allocating as it goes. All of this reads memory that has already been freed.
Collapse the two exits into one to ensure a single free() that happens after the last use.
Three things have to coincide to reach this:
1. The path has to be relative, or nothing is allocated and "cp" still points into the argument.
2. The entry found has to be a sparse
directory, which needs a sparse index.
3. The argument has to get past the check in die_verify_filename() that skips a leading ':' followed by a non-alphanumeric, so ":0:./dir/" arrives here where ":./dir/" does not.
Add a test to t1092 that covers the combination. It fails under SANITIZE=address without the change to object-name.c.
This was reported in [1], and the shape used here was suggested in review [2], but that series was not rerolled and the fix never landed.
[1] https://lore.kernel.org/git/cf6bcdb43e5b4abab464c30a914d64dc8e7a9925.1655336146.git.gitgitgadget@gmail.com/
[2] https://lore.kernel.org/git/[email protected]/
---
Changes since v1: credit Johannes as both, Reported-by and Original-patch-by, as pointed out by René. No change to the code or the test.
The three conditions make this awkward to reach by hand, so here is the recipe:
git init sparse && cd sparse && mkdir folder1 folder2 && echo a >folder1/a && echo b >folder2/b && git add -A && git commit -m init && git sparse-checkout init --cone --sparse-index && git sparse-checkout set folder1 && git show :0:./folder2/
Without the change below, no sanitizer is needed to see it. On 2.52.0 the buffer has already been reused by the time the message is formatted, so the path printed is whatever now sits in that memory, and it differs from run to run:
fatal: path '' does not exist (neither on disk nor in the index)
fatal: path 'M-6?:xM-@M-:M-L??X' does not exist (neither on disk nor in the index)
fatal: path '?M-*JM-^M->M-YM-tn?H' does not exist (neither on disk nor in the index)
Still without the change, built with SANITIZE=address, the same command reports
ERROR: AddressSanitizer: heap-use-after-free READ of size 3 at 0x607000002a20
#1 diagnose_invalid_index_path object-name.c:1653
#2 get_oid_with_context_1 object-name.c:1807
#3 maybe_die_on_misspelt_object_name
#4 die_verify_filename setup.c:216
#6 setup_revisions revision.c:3103
#8 cmd_show log.c:694
freed by thread T0 here:
#1 get_oid_with_context_1 object-name.c:1806
previously allocated by thread T0 here:
#5 prefix_path setup.c:149
#6 get_oid_with_context_1 object-name.c:1784
With the change, the message reads "folder2/" every time and the sanitizer stays quiet.
"git diff" and "git rev-parse" reach it the same way, and so does "../" from a subdirectory.
The new t1092 test fails without the object-name.c hunk and passes with it, under SANITIZE=address.
object-name.c | 15 +++++++++------ t/t1092-sparse-checkout-compatibility.sh | 11 +++++++++++ 2 files changed, 20 insertions(+), 6 deletions(-)
object-name.c +9 -6
@@ -1803,13 +1803,16 @@ static enum get_oid_result get_oid_with_context_1(struct repository *repo, memcmp(ce->name, cp, namelen))break ;if (ce_stage (ce) == stage) { +int ret =-1 ; + +if (!reject_tree_in_index (repo, only_to_die, ce, + stage, prefix, cp)) { +oidcpy (oid, &ce->oid); + oc->mode = ce->ce_mode; + ret =0 ; + }free (new_path); -if (reject_tree_in_index (repo, only_to_die, ce, - stage, prefix, cp)) -return -1 ; -oidcpy (oid, &ce->oid); - oc->mode = ce->ce_mode; -return 0 ; +return ret; } pos++; }
t/t1092-sparse-checkout-compatibility.sh +11 -0
@@ -1357,6 +1357,17 @@ do " done +test_expect_success 'relative path to a sparse directory' ' + init_repos && + + # A ":<stage>:<path>" argument whose path is relative is resolved + # into a heap-allocated buffer, and a sparse directory found at that + # path is reported through it. Cover that combination, so that the + # reporting does not read the buffer after it has been released. + test_sparse_match test_must_fail git show :0:./folder1/ && + test_sparse_match test_must_fail git rev-parse :0:./folder1/ +' + test_expect_success 'submodule handling' ' init_repos &&
Signature · 1 line
2.52.0
Quoted 6 lines · Shlok Kulshreshtha
object-name.c +0
-0@@ -1803,13 +1803,16 @@ static enum get_oid_result get_oid_with_context_1(struct repository *repo, memcmp(ce->name, cp, namelen))
object-name.c static enum get_oid_result get_oid_with_context_1(struct repository *repo, · line 1813
Quoted 10 lines
object-name.c +8
-0break ;if (ce_stage (ce) == stage) { +int ret =-1 ; + +if (!reject_tree_in_index (repo, only_to_die, ce, + stage, prefix, cp)) { +oidcpy (oid, &ce->oid); + oc->mode = ce->ce_mode; + ret =0 ; + }
The function only ever returns -1 or 0 itself, so we could've written it this way:
int ret = reject_tree_in_index(repo, only_to_die, ce,
stage, prefix, cp);
if (!ret) {
oidcpy(oid, &ce->oid);
oc->mode = ce->ce_mode;
}
free(new_path);
return ret;But I won't insist on that change, this is already a clear improvement.
Quoted 8 lines
t/t1092-sparse-checkout-compatibility.sh +0
-0@@ -1357,6 +1357,17 @@ do " done
t/t1092-sparse-checkout-compatibility.sh do · line 1369
Quoted 10 lines
t/t1092-sparse-checkout-compatibility.sh +10
-0+test_expect_success 'relative path to a sparse directory' ' + init_repos && + + # A ":<stage>:<path>" argument whose path is relative is resolved + # into a heap-allocated buffer, and a sparse directory found at that + # path is reported through it. Cover that combination, so that the + # reporting does not read the buffer after it has been released. + test_sparse_match test_must_fail git show :0:./folder1/ && + test_sparse_match test_must_fail git rev-parse :0:./folder1/ +'
Yup, this test indeed catches the bug:
--- sparse-checkout-err 2026-08-10 06:13:59.698011294 +0000
+++ sparse-index-err 2026-08-10 06:13:59.703981906 +0000
@@ -1 +1 @@
-fatal: path 'folder1/' does not exist (neither on disk nor in the index)
+fatal: path '�[UU?' does not exist (neither on disk nor in the index)Overall this looks good to me, thanks!
Patrick
PATCH v3 1/1 object-name: avoid use-after-free in get_oid_with_context_1()
When a ":<path>" argument names a relative path, resolve_relative_path() returns a newly allocated string and "cp" is pointed at it:
new_path = resolve_relative_path(repo, cp);
if (!new_path) {
namelen = namelen - (cp - name);
} else {
cp = new_path;
namelen = strlen(cp);
}From there on "cp" and "new_path" name the same allocation. Later the memory location that "new_path" points to is freed.
free(new_path); if (reject_tree_in_index(repo, only_to_die, ce, stage, prefix, cp))
But here the reject_tree_in_index() passes "cp" to diagnose_invalid_index_path(), which calls strlen() on it, looks it up in the index, and formats it into its messages, allocating as it goes. All of this reads memory that has already been freed.
Collapse the two exits into one to ensure a single free() that happens after the last use.
Three things have to coincide to reach this:
1. The path has to be relative, or nothing is allocated and "cp" still points into the argument.
2. The entry found has to be a sparse
directory, which needs a sparse index.
3. The argument has to get past the check in die_verify_filename() that skips a leading ':' followed by a non-alphanumeric, so ":0:./dir/" arrives here where ":./dir/" does not.
Add a test to t1092 that covers the combination. It fails under SANITIZE=address without the change to object-name.c.
This was reported in [1], and the shape used here was suggested in review [2], but that series was not rerolled and the fix never landed.
[1] https://lore.kernel.org/git/cf6bcdb43e5b4abab464c30a914d64dc8e7a9925.1655336146.git.gitgitgadget@gmail.com/
[2] https://lore.kernel.org/git/[email protected]/
---
Changes since v2: apply Patrick's suggestion to assign reject_tree_in_index()'s return value directly, since it only returns 0 or -1, which is exactly what this function needs to return too. No change to the test. This reduces redundancy.
The three conditions make this awkward to reach by hand, so here is the recipe:
git init sparse && cd sparse && mkdir folder1 folder2 && echo a >folder1/a && echo b >folder2/b && git add -A && git commit -m init && git sparse-checkout init --cone --sparse-index && git sparse-checkout set folder1 && git show :0:./folder2/
Without the change below, no sanitizer is needed to see it. On 2.52.0 the buffer has already been reused by the time the message is formatted, so the path printed is whatever now sits in that memory, and it differs from run to run:
fatal: path '' does not exist (neither on disk nor in the index)
fatal: path 'M-6?:xM-@M-:M-L??X' does not exist (neither on disk nor in the index)
fatal: path '?M-*JM-^M->M-YM-tn?H' does not exist (neither on disk nor in the index)
Still without the change, built with SANITIZE=address, the same command reports
ERROR: AddressSanitizer: heap-use-after-free READ of size 3 at 0x607000002a20
#1 diagnose_invalid_index_path object-name.c:1653
#2 get_oid_with_context_1 object-name.c:1807
#3 maybe_die_on_misspelt_object_name
#4 die_verify_filename setup.c:216
#6 setup_revisions revision.c:3103
#8 cmd_show log.c:694
freed by thread T0 here:
#1 get_oid_with_context_1 object-name.c:1806
previously allocated by thread T0 here:
#5 prefix_path setup.c:149
#6 get_oid_with_context_1 object-name.c:1784
With the change, the message reads "folder2/" every time and the sanitizer stays quiet.
"git diff" and "git rev-parse" reach it the same way, and so does "../" from a subdirectory.
Re-verified against the current tree: the t1092 test still fails without the object-name.c change and passes with it, under SANITIZE=address.
object-name.c | 14 ++++++++------ t/t1092-sparse-checkout-compatibility.sh | 11 +++++++++++ 2 files changed, 19 insertions(+), 6 deletions(-)
object-name.c +8 -6
@@ -1803,13 +1803,15 @@ static enum get_oid_result get_oid_with_context_1(struct repository *repo, memcmp(ce->name, cp, namelen))break ;if (ce_stage (ce) == stage) { +int ret =reject_tree_in_index (repo, only_to_die, ce, + stage, prefix, cp); + +if (!ret) { +oidcpy (oid, &ce->oid); + oc->mode = ce->ce_mode; + }free (new_path); -if (reject_tree_in_index (repo, only_to_die, ce, - stage, prefix, cp)) -return -1 ; -oidcpy (oid, &ce->oid); - oc->mode = ce->ce_mode; -return 0 ; +return ret; } pos++; }
t/t1092-sparse-checkout-compatibility.sh +11 -0
@@ -1405,6 +1405,17 @@ do " done +test_expect_success 'relative path to a sparse directory' ' + init_repos && + + # A ":<stage>:<path>" argument whose path is relative is resolved + # into a heap-allocated buffer, and a sparse directory found at that + # path is reported through it. Cover that combination, so that the + # reporting does not read the buffer after it has been released. + test_sparse_match test_must_fail git show :0:./folder1/ && + test_sparse_match test_must_fail git rev-parse :0:./folder1/ +' + test_expect_success 'submodule handling' ' init_repos &&
Signature · 1 line
2.52.0
Quoted 11 lines · Shlok Kulshreshtha
When a ":<path>" argument names a relative path,
resolve_relative_path()returns a newly allocated string and "cp" is pointed at it:new_path = resolve_relative_path(repo, cp); if (!new_path) { namelen = namelen - (cp - name); } else { cp = new_path; namelen = strlen(cp); }
Quoted 5 lines
From there on "cp" and "
new_path" name the same allocation. Later the memory location that "new_path" points to is freed.free(new_path); if (reject_tree_in_index(repo, only_to_die, ce, stage, prefix, cp))
Nicely described and ...
Quoted 14 lines
object-name.c +6
-0@@ -1803,13 +1803,15 @@ static enum get_oid_result get_oid_with_context_1(struct repository *repo, memcmp(ce->name, cp, namelen))break ;if (ce_stage (ce) == stage) { +int ret =reject_tree_in_index (repo, only_to_die, ce, + stage, prefix, cp); + +if (!ret) { +oidcpy (oid, &ce->oid); + oc->mode = ce->ce_mode;
object-name.c static enum get_oid_result get_oid_with_context_1(struct repository *repo, · line 1815
Quoted 10 lines
object-name.c +2
-6+ }free (new_path); -if (reject_tree_in_index (repo, only_to_die, ce, - stage, prefix, cp)) -return -1 ; -oidcpy (oid, &ce->oid); - oc->mode = ce->ce_mode; -return 0 ; +return ret; }
... the fix matches exactly what anybody would expect from the problem description, i.e., "Do not free new_path before we are done with using cp".
Will queue. Thanks.
Quoted 1 line · Patrick Steinhardt
But I won't insist on that change, this is already a clear improvement.
Thanks for the review. That does read cleaner, so I've included it in v3.
Shlok