Explicitly cloning over the "file://" protocol in t1092 in preparation
for merging a security release which will change the default value of
this configuration to be "user".
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Explicitly cloning over the "file://" protocol in t1092 in preparation
for merging a security release which will change the default value of
this configuration to be "user".
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Explicitly cloning over the "file://" protocol in t1092 in preparation
for merging a security release which will change the default value of
this configuration to be "user".
Signed-off-by: Taylor Blau <me@ttaylorr.com>
This function improperly uses an int to represent the number of entries
in the resulting argument array. This allows a malicious actor to
intentionally overflow the return value, leading to arbitrary heap
writes.
Because the resulting argv array is typically passed to execv(), it may
be possible to leverage this attack to gain remote code execution on a
victim machine. This was almost certainly the case for certain
configurations of git-shell until the previous commit limited the size
of input it would accept. Other calls to split_cmdline() are typically
limited by the size of argv the OS is willing to hand us, so are
similarly protected.
So this is not strictly fixing a known vulnerability, but is a hardening
of the function that is worth doing to protect against possible unknown
vulnerabilities.
One approach to fixing this would be modifying the signature of
`split_cmdline()` to look something like:
int split_cmdline(char *cmdline, const char ***argv, size_t *argc);
Where the return value of `split_cmdline()` is negative for errors, and
zero otherwise. If non-NULL, the `*argc` pointer is modified to contain
the size of the `**argv` array.
But this implies an absurdly large `argv` array, which more than likely
larger than the system's argument limit. So even if split_cmdline()
allowed this, it would fail immediately afterwards when we called
execv(). So instead of converting all of `split_cmdline()`'s callers to
work with `size_t` types in this patch, instead pursue the minimal fix
here to prevent ever returning an array with more than INT_MAX entries
in it.
Signed-off-by: Kevin Backhouse <kevinbackhouse@github.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
When git-shell is run in interactive mode (which must be enabled by
creating $HOME/git-shell-commands), it reads commands from stdin, one
per line, and executes them.
We read the commands with git_read_line_interactively(), which uses a
strbuf under the hood. That means we'll accept an input of arbitrary
size (limited only by how much heap we can allocate). That creates two
problems:
- the rest of the code is not prepared to handle large inputs. The
most serious issue here is that split_cmdline() uses "int" for most
of its types, which can lead to integer overflow and out-of-bounds
array reads and writes. But even with that fixed, we assume that we
can feed the command name to snprintf() (via xstrfmt()), which is
stuck for historical reasons using "int", and causes it to fail (and
even trigger a BUG() call).
- since the point of git-shell is to take input from untrusted or
semi-trusted clients, it's a mild denial-of-service. We'll allocate
as many bytes as the client sends us (actually twice as many, since
we immediately duplicate the buffer).
We can fix both by just limiting the amount of per-command input we're
willing to receive.
We should also fix split_cmdline(), of course, which is an accident
waiting to happen, but that can come on top. Most calls to
split_cmdline(), including the other one in git-shell, are OK because
they are reading from an OS-provided argv, which is limited in practice.
This patch should eliminate the immediate vulnerabilities.
I picked 4MB as an arbitrary limit. It's big enough that nobody should
ever run into it in practice (since the point is to run the commands via
exec, we're subject to OS limits which are typically much lower). But
it's small enough that allocating it isn't that big a deal.
The code is mostly just swapping out fgets() for the strbuf call, but we
have to add a few niceties like flushing and trimming line endings. We
could simplify things further by putting the buffer on the stack, but
4MB is probably a bit much there. Note that we'll _always_ allocate 4MB,
which for normal, non-malicious requests is more than we would before
this patch. But on the other hand, other git programs are happy to use
96MB for a delta cache. And since we'd never touch most of those pages,
on a lazy-allocating OS like Linux they won't even get allocated to
actual RAM.
The ideal would be a version of strbuf_getline() that accepted a maximum
value. But for a minimal vulnerability fix, let's keep things localized
and simple. We can always refactor further on top.
The included test fails in an obvious way with ASan or UBSan (which
notice the integer overflow and out-of-bounds reads). Without them, it
fails in a less obvious way: we may segfault, or we may try to xstrfmt()
a long string, leading to a BUG(). Either way, it fails reliably before
this patch, and passes with it. Note that we don't need an EXPENSIVE
prereq on it. It does take 10-15s to fail before this patch, but with
the new limit, we fail almost immediately (and the perl process
generating 2GB of data exits via SIGPIPE).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
We have no tests of even basic functionality of git-shell. Let's add a
couple of obvious ones. This will serve as a framework for adding tests
for new things we fix, as well as making sure we don't screw anything up
too badly while doing so.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
An earlier patch discussed and fixed a scenario where Git could be used
as a vector to exfiltrate sensitive data through a Docker container when
a potential victim clones a suspicious repository with local submodules
that contain symlinks.
That security hole has since been plugged, but a similar one still
exists. Instead of convincing a would-be victim to clone an embedded
submodule via the "file" protocol, an attacker could convince an
individual to clone a repository that has a submodule pointing to a
valid path on the victim's filesystem.
For example, if an individual (with username "foo") has their home
directory ("/home/foo") stored as a Git repository, then an attacker
could exfiltrate data by convincing a victim to clone a malicious
repository containing a submodule pointing at "/home/foo/.git" with
`--recurse-submodules`. Doing so would expose any sensitive contents in
stored in "/home/foo" tracked in Git.
For systems (such as Docker) that consider everything outside of the
immediate top-level working directory containing a Dockerfile as
inaccessible to the container (with the exception of volume mounts, and
so on), this is a violation of trust by exposing unexpected contents in
the working copy.
To mitigate the likelihood of this kind of attack, adjust the "file://"
protocol's default policy to be "user" to prevent commands that execute
without user input (including recursive submodule initialization) from
taking place by default.
Suggested-by: Jeff King <peff@peff.net>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that interact with submodules a handful of times use
`test_config_global`.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`. Tests that interact with submodules a handful of
times use `test_config_global` instead. Test scripts that rely on
submodules throughout use a `git config --global` during a setup test
towards the beginning of the script.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`. Tests that interact with submodules a handful of
times use `test_config_global` instead. Test scripts that rely on
submodules throughout use a `git config --global` during a setup test
towards the beginning of the script.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`. Tests that interact with submodules a handful of
times use `test_config_global` instead. Test scripts that rely on
submodules throughout use a `git config --global` during a setup test
towards the beginning of the script.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`. Tests that interact with submodules a handful of
times use `test_config_global` instead. Test scripts that rely on
submodules throughout use a `git config --global` during a setup test
towards the beginning of the script.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`. Tests that interact with submodules a handful of
times use `test_config_global` instead. Test scripts that rely on
submodules throughout use a `git config --global` during a setup test
towards the beginning of the script.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
To prepare for the default value of `protocol.file.allow` to change to
"user", ensure tests that rely on local submodules can initialize them
over the file protocol.
Tests that only need to interact with submodules in a limited capacity
have individual Git commands annotated with the appropriate
configuration via `-c`. Tests that interact with submodules a handful of
times use `test_config_global` instead.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
To prepare for changing the default value of `protocol.file.allow` to
"user", update the `prolog()` function in lib-submodule-update to allow
submodules to be cloned over the file protocol.
This is used by a handful of submodule-related test scripts, which
themselves will have to tweak the value of `protocol.file.allow` in
certain locations. Those will be done in subsequent commits.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
When cloning a repository with `--local`, Git relies on either making a
hardlink or copy to every file in the "objects" directory of the source
repository. This is done through the callpath `cmd_clone()` ->
`clone_local()` -> `copy_or_link_directory()`.
The way this optimization works is by enumerating every file and
directory recursively in the source repository's `$GIT_DIR/objects`
directory, and then either making a copy or hardlink of each file. The
only exception to this rule is when copying the "alternates" file, in
which case paths are rewritten to be absolute before writing a new
"alternates" file in the destination repo.
One quirk of this implementation is that it dereferences symlinks when
cloning. This behavior was most recently modified in 36596fd2df (clone:
better handle symlinked files at .git/objects/, 2019-07-10), which
attempted to support `--local` clones of repositories with symlinks in
their objects directory in a platform-independent way.
Unfortunately, this behavior of dereferencing symlinks (that is,
creating a hardlink or copy of the source's link target in the
destination repository) can be used as a component in attacking a
victim by inadvertently exposing the contents of file stored outside of
the repository.
Take, for example, a repository that stores a Dockerfile and is used to
build Docker images. When building an image, Docker copies the directory
contents into the VM, and then instructs the VM to execute the
Dockerfile at the root of the copied directory. This protects against
directory traversal attacks by copying symbolic links as-is without
dereferencing them.
That is, if a user has a symlink pointing at their private key material
(where the symlink is present in the same directory as the Dockerfile,
but the key itself is present outside of that directory), the key is
unreadable to a Docker image, since the link will appear broken from the
container's point of view.
This behavior enables an attack whereby a victim is convinced to clone a
repository containing an embedded submodule (with a URL like
"file:///proc/self/cwd/path/to/submodule") which has a symlink pointing
at a path containing sensitive information on the victim's machine. If a
user is tricked into doing this, the contents at the destination of
those symbolic links are exposed to the Docker image at runtime.
One approach to preventing this behavior is to recreate symlinks in the
destination repository. But this is problematic, since symlinking the
objects directory are not well-supported. (One potential problem is that
when sharing, e.g. a "pack" directory via symlinks, different writers
performing garbage collection may consider different sets of objects to
be reachable, enabling a situation whereby garbage collecting one
repository may remove reachable objects in another repository).
Instead, prohibit the local clone optimization when any symlinks are
present in the `$GIT_DIR/objects` directory of the source repository.
Users may clone the repository again by prepending the "file://" scheme
to their clone URL, or by adding the `--no-local` option to their `git
clone` invocation.
The directory iterator used by `copy_or_link_directory()` must no longer
dereference symlinks (i.e., it *must* call `lstat()` instead of `stat()`
in order to discover whether or not there are symlinks present). This has
no bearing on the overall behavior, since we will immediately `die()` on
encounter a symlink.
Note that t5604.33 suggests that we do support local clones with
symbolic links in the source repository's objects directory, but this
was likely unintentional, or at least did not take into consideration
the problem with sharing parts of the objects directory with symbolic
links at the time. Update this test to reflect which options are and
aren't supported.
Helped-by: Johannes Schindelin <Johannes.Schindelin@gmx.de>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
As the 'master' front will soon tag a preview and then release
candidates for 2.38, it is unknown if we are going to issue another
maintenance release on the 2.37.x track, but as we have accumulated
enough material there, let's prepare a draft for it.
Even if we end up not tagging 2.37.4, it would help motivated distro
packagers to maintain their slightly older and "more stable" versions.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The auto-stashed local changes created by "git merge --autostash"
was mixed into a conflicted state left in the working tree, which
has been corrected.
* en/merge-unstash-only-on-clean-merge:
merge: only apply autostash when appropriate
Update the version of Ubuntu used for GitHub Actions CI from 18.04
to 22.04.
* ds/github-actions-use-newer-ubuntu:
ci: update 'static-analysis' to Ubuntu 22.04
The preload-index codepath made copies of pathspec to give to
multiple threads, which were left leaked.
* ad/preload-plug-memleak:
preload-index: fix memleak
xcalloc(), imitating calloc(), takes "number of elements of the
array", and "size of a single element", in this order. A call that
does not follow this ordering has been corrected.
* sg/xcalloc-cocci-fix:
promisor-remote: fix xcalloc() argument order
Fix deadlocks between main Git process and subprocess spawned via
the pipe_command() API, that can kill "git add -p" that was
reimplemented in C recently.
* jk/pipe-command-nonblock:
pipe_command(): mark stdin descriptor as non-blocking
pipe_command(): handle ENOSPC when writing to a pipe
pipe_command(): avoid xwrite() for writing to pipe
git-compat-util: make MAX_IO_SIZE define globally available
nonblock: support Windows
compat: add function to enable nonblocking pipes
An earlier optimization discarded a tree-object buffer that is
still in use, which has been corrected.
* jk/is-promisor-object-keep-tree-in-use:
is_promisor_object(): fix use-after-free of tree buffer
Documentation for "git add --renormalize" has been improved.
source: <20220810144450.470-2-philipoakley@iee.email>
* po/doc-add-renormalize:
doc add: renormalize is not idempotent for CRCRLF
Fixes to sparse index compatibility work for "reset" and "checkout"
commands.
source: <pull.1312.v3.git.1659985672.gitgitgadget@gmail.com>
* vd/sparse-reset-checkout-fixes:
unpack-trees: unpack new trees as sparse directories
cache.h: create 'index_name_pos_sparse()'
oneway_diff: handle removed sparse directories
checkout: fix nested sparse directory diff in sparse index
"git fsck" reads mode from tree objects but canonicalizes the mode
before passing it to the logic to check object sanity, which has
hid broken tree objects from the checking logic. This has been
corrected, but to help exiting projects with broken tree objects
that they cannot fix retroactively, the severity of anomalies this
code detects has been demoted to "info" for now.
source: <YvQcNpizy9uOZiAz@coredump.intra.peff.net>
* jk/fsck-tree-mode-bits-fix:
fsck: downgrade tree badFilemode to "info"
fsck: actually detect bad file modes in trees
tree-walk: add a mechanism for getting non-canonicalized modes
Platform-specific code that determines if a directory is OK to use
as a repository has been taught to report more details, especially
on Windows.
source: <pull.1286.v2.git.1659965270.gitgitgadget@gmail.com>
* js/safe-directory-plus:
mingw: handle a file owned by the Administrators group correctly
mingw: be more informative when ownership check fails on FAT32
mingw: provide details about unsafe directories' ownership
setup: prepare for more detailed "dubious ownership" messages
setup: fix some formatting
Avoid repeatedly running getconf to ask libc version in the test
suite, and instead just as it once per script.
source: <pull.1311.git.1659620305757.gitgitgadget@gmail.com>
* pw/use-glibc-tunable-for-malloc-optim:
tests: cache glibc version check
A follow-up fix to a fix for a regression in 2.36.
source: <patch-1.1-2450e3e65cf-20220805T141402Z-avarab@gmail.com>
* ab/hooks-regression-fix:
hook API: don't segfault on strbuf_addf() to NULL "out"
Plug memory leaks in the failure code path in the "merge-ort" merge
strategy backend.
source: <pull.1307.v2.git.1659114727.gitgitgadget@gmail.com>
* js/ort-clean-up-after-failed-merge:
merge-ort: do leave trace2 region even if checkout fails
merge-ort: clean up after failed merge
Older gcc with -Wall complains about the universal zero initializer
"struct s = { 0 };" idiom, which makes developers' lives
inconvenient (as -Werror is enabled by DEVELOPER=YesPlease). The
build procedure has been tweaked to help these compilers.
source: <YuQ60ZUPBHAVETD7@coredump.intra.peff.net>
* jk/struct-zero-init-with-older-gcc:
config.mak.dev: squelch -Wno-missing-braces for older gcc
Conditionally allow building Python interpreter on Windows
source: <pull.1306.v2.git.1659109272.gitgitgadget@gmail.com>
* js/mingw-with-python:
mingw: remove unneeded `NO_CURL` directive
mingw: remove unneeded `NO_GETTEXT` directive
windows: include the Python bits when building Git for Windows
Fix build procedure for Windows that uses CMake so that it can pick
up the shell interpreter from local installation location.
source: <pull.1304.git.1658912756815.gitgitgadget@gmail.com>
* ca/unignore-local-installation-on-windows:
cmake: support local installations of git
GitHub Actions scheduled a brownout of Ubuntu 18.04, which canceled all
runs of the 'static-analysis' job in our CI runs. Update to 22.04 to
avoid this as the brownout later turns into a complete deprecation.
The use of 18.04 was set in d051ed77ee (.github/workflows/main.yml: run
static-analysis on bionic, 2021-02-08) due to the lack of Coccinelle
being available on 20.04 (which continues today).
Signed-off-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If a merge failed and we are leaving conflicts in the working directory
for the user to resolve, we should not attempt to apply any autostash.
Further, if we fail to apply the autostash (because either the merge
failed, or the user requested --no-commit), then we should instruct the
user how to apply it later.
Add a testcase verifying we have corrected this behavior.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Pass the number of elements first and their size second, as expected
by xcalloc().
Patch generated with:
make SPATCH_FLAGS=--recursive-includes contrib/coccinelle/xcalloc.cocci.patch
Our default SPATCH_FLAGS ('--all-includes') doesn't catch this
transformation by default, unless used in combination with a large-ish
SPATCH_BATCH_SIZE which happens to put 'promisor-remote.c' with a file
that includes 'repository.h' directly in the same batch.
Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fix a memory leak occuring in case of pathspec copy in preload_index.
Direct leak of 8 byte(s) in 8 object(s) allocated from:
#0 0x7f0a353ead47 in __interceptor_malloc (/usr/lib/gcc/x86_64-pc-linux-gnu/11.3.0/libasan.so.6+0xb5d47)
#1 0x55750995e840 in do_xmalloc /home/anthony/src/c/git/wrapper.c:51
#2 0x55750995e840 in xmalloc /home/anthony/src/c/git/wrapper.c:72
#3 0x55750970f824 in copy_pathspec /home/anthony/src/c/git/pathspec.c:684
#4 0x557509717278 in preload_index /home/anthony/src/c/git/preload-index.c:135
#5 0x55750975f21e in refresh_index /home/anthony/src/c/git/read-cache.c:1633
#6 0x55750915b926 in cmd_status builtin/commit.c:1547
#7 0x5575090e1680 in run_builtin /home/anthony/src/c/git/git.c:466
#8 0x5575090e1680 in handle_builtin /home/anthony/src/c/git/git.c:720
#9 0x5575090e284a in run_argv /home/anthony/src/c/git/git.c:787
#10 0x5575090e284a in cmd_main /home/anthony/src/c/git/git.c:920
#11 0x5575090dbf82 in main /home/anthony/src/c/git/common-main.c:56
#12 0x7f0a348230ab (/lib64/libc.so.6+0x290ab)
Signed-off-by: Anthony Delannoy <anthony.2lannoy@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Our pipe_command() helper lets you both write to and read from a child
process on its stdin/stdout. It's supposed to work without deadlocks
because we use poll() to check when descriptors are ready for reading or
writing. But there's a bug: if both the data to be written and the data
to be read back exceed the pipe buffer, we'll deadlock.
The issue is that the code assumes that if you have, say, a 2MB buffer
to write and poll() tells you that the pipe descriptor is ready for
writing, that calling:
write(cmd->in, buf, 2*1024*1024);
will do a partial write, filling the pipe buffer and then returning what
it did write. And that is what it would do on a socket, but not for a
pipe. When writing to a pipe, at least on Linux, it will block waiting
for the child process to read() more. And now we have a potential
deadlock, because the child may be writing back to us, waiting for us to
read() ourselves.
An easy way to trigger this is:
git -c add.interactive.useBuiltin=true \
-c interactive.diffFilter=cat \
checkout -p HEAD~200
The diff against HEAD~200 will be big, and the filter wants to write all
of it back to us (obviously this is a dummy filter, but in the real
world something like diff-highlight would similarly stream back a big
output).
If you set add.interactive.useBuiltin to false, the problem goes away,
because now we're not using pipe_command() anymore (instead, that part
happens in perl). But this isn't a bug in the interactive code at all.
It's the underlying pipe_command() code which is broken, and has been
all along.
We presumably didn't notice because most calls only do input _or_
output, not both. And the few that do both, like gpg calls, may have
large inputs or outputs, but never both at the same time (e.g., consider
signing, which has a large payload but a small signature comes back).
The obvious fix is to put the descriptor into non-blocking mode, and
indeed, that makes the problem go away. Callers shouldn't need to
care, because they never see the descriptor (they hand us a buffer to
feed into it).
The included test fails reliably on Linux without this patch. Curiously,
it doesn't fail in our Windows CI environment, but has been reported to
do so for individual developers. It should pass in any environment after
this patch (courtesy of the compat/ layers added in the last few
commits).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When write() to a non-blocking pipe fails because the buffer is full,
POSIX says we should see EAGAIN. But our mingw_write() compat layer on
Windows actually returns ENOSPC for this case. This is probably
something we want to correct, but given that we don't plan to use
non-blocking descriptors in a lot of places, we can work around it by
just catching ENOSPC alongside EAGAIN. If we ever do fix mingw_write(),
then this patch can be reverted.
We don't actually use a non-blocking pipe yet, so this is still just
preparation.
Helped-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If xwrite() sees an EAGAIN response, it will loop forever until the
write succeeds (or encounters a real error). This is due to ef1cf0167a
(xwrite: poll on non-blocking FDs, 2016-06-26), with the idea that we
won't be surprised by a descriptor unexpectedly set as non-blocking.
But that will make things awkward when we do want a non-blocking
descriptor, and a future patch will switch pipe_command() to using one.
In that case, looping on EAGAIN is bad, because the process on the other
end of the pipe may be waiting on us before doing another read() on the
pipe, which would mean we deadlock.
In practice we're not supposed to ever see EAGAIN here, since poll()
will have just told us the descriptor is ready for writing. But our
Windows emulation of poll() will always return "ready" for writing to a
pipe descriptor! This is due to 94f4d01932 (mingw: workaround for hangs
when sending STDIN, 2020-02-17).
Our best bet in that case is to keep handling other descriptors, as any
read() we do may allow the child command to make forward progress (i.e.,
its write() finishes, and then it read()s from its stdin, freeing up
space in the pipe buffer). This means we might busy-loop between poll()
and write() on Windows if the child command is slow to read our input,
but it's much better than the alternative of deadlocking.
In practice, this busy-looping should be rare:
- for small inputs, we'll just write the whole thing in a single
write() anyway, non-blocking or not
- for larger inputs where the child reads input and then processes it
before writing (e.g., gpg verifying a signature), we may make a few
extra write() calls that get EAGAIN during the initial write, but
once it has taken in the whole input, we'll correctly block waiting
to read back the data.
- for larger inputs where the child process is streaming output back
(like a diff filter), we'll likewise see some extra EAGAINs, but
most of them will be followed immediately by a read(), which will
let the child command make forward progress.
Of course it won't happen at all for now, since we don't yet use a
non-blocking pipe. This is just preparation for when we do.
Helped-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We define MAX_IO_SIZE within wrapper.c, but it's useful for any code
that wants to do a raw write() for whatever reason (say, because they
want different EAGAIN handling). Let's make it available everywhere.
The alternative would be adding xwrite_foo() variants to give callers
more options. But there's really no reason MAX_IO_SIZE needs to be
abstracted away, so this give callers the most flexibility.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Implement enable_pipe_nonblock() using the Windows API. This works only
for pipes, but that is sufficient for this limited interface. Despite
the API calls used, it handles both "named" and anonymous pipes from our
pipe() emulation.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We'd like to be able to make some of our pipes nonblocking so that
poll() can be used effectively, but O_NONBLOCK isn't portable. Let's
introduce a compat wrapper so this can be abstracted for each platform.
The interface is as narrow as possible to let platforms do what's
natural there (rather than having to implement fcntl() and a fake
O_NONBLOCK for example, or having to handle other types of descriptors).
The next commit will add Windows support, at which point we should be
covering all platforms in practice. But if we do find some other
platform without O_NONBLOCK, we'll return ENOSYS. Arguably we could just
trigger a build-time #error in this case, which would catch the problem
earlier. But since we're not planning to use this compat wrapper in many
code paths, a seldom-seen runtime error may be friendlier for such a
platform than blocking compilation completely. Our test suite would
still notice it.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since commit fcc07e980b (is_promisor_object(): free tree buffer after
parsing, 2021-04-13), we'll always free the buffers attached to a
"struct tree" after searching them for promisor links. But there's an
important case where we don't want to do so: if somebody else is already
using the tree!
This can happen during a "rev-list --missing=allow-promisor" traversal
in a partial clone that is missing one or more trees or blobs. The
backtrace for the free looks like this:
#1 free_tree_buffer tree.c:147
#2 add_promisor_object packfile.c:2250
#3 for_each_object_in_pack packfile.c:2190
#4 for_each_packed_object packfile.c:2215
#5 is_promisor_object packfile.c:2272
#6 finish_object__ma builtin/rev-list.c:245
#7 finish_object builtin/rev-list.c:261
#8 show_object builtin/rev-list.c:274
#9 process_blob list-objects.c:63
#10 process_tree_contents list-objects.c:145
#11 process_tree list-objects.c:201
#12 traverse_trees_and_blobs list-objects.c:344
[...]
We're in the middle of walking through the entries of a tree object via
process_tree_contents(). We see a blob (or it could even be another tree
entry) that we don't have, so we call is_promisor_object() to check it.
That function loops over all of the objects in the promisor packfile,
including the tree we're currently walking. When we're done with it
there, we free the tree buffer. But as we return to the walk in
process_tree_contents(), it's still holding on to a pointer to that
buffer, via its tree_desc iterator, and it accesses the freed memory.
Even a trivial use of "--missing=allow-promisor" triggers this problem,
as the included test demonstrates (it's just a vanilla --blob:none
clone).
We can detect this case by only freeing the tree buffer if it was
allocated on our behalf. This is a little tricky since that happens
inside parse_object(), and it doesn't tell us whether the object was
already parsed, or whether it allocated the buffer itself. But by
checking for an already-parsed tree beforehand, we can distinguish the
two cases.
That feels a little hacky, and does incur an extra lookup in the
object-hash table. But that cost is fairly minimal compared to actually
loading objects (and since we're iterating the whole pack here, we're
likely to be loading most objects, rather than reusing cached results).
It may also be a good direction for this function in general, as there
are other possible optimizations that rely on doing some analysis before
parsing:
- we could detect blobs and avoid reading their contents; they can't
link to other objects, but parse_object() doesn't know that we don't
care about checking their hashes.
- we could avoid allocating object structs entirely for most objects
(since we really only need them in the oidset), which would save
some memory.
- promisor commits could use the commit-graph rather than loading the
object from disk
This commit doesn't do any of those optimizations, but I think it argues
that this direction is reasonable, rather than relying on parse_object()
and trying to teach it to give us more information about whether it
parsed.
The included test fails reliably under SANITIZE=address just when
running "rev-list --missing=allow-promisor". Checking the output isn't
strictly necessary to detect the bug, but it seems like a reasonable
addition given the general lack of coverage for "allow-promisor" in the
test suite.
Reported-by: Andrew Olsen <andrew.olsen@koordinates.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
