The vast majority of files including object-store.h did not need dir.h
nor khash.h. Split the header into two files, and let most just depend
upon object-store-ll.h, while letting the two callers that need it
depend on the full object-store.h.
After this patch:
$ git grep -h include..object-store | sort | uniq -c
2 #include "object-store.h"
129 #include "object-store-ll.h"
Diff best viewed with `--color-moved`.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This also made it clear that several .c files that depended upon path.h
were missing a #include for it; add the missing includes while at it.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When "gc" needs to retain unreachable objects, packing them into
cruft packs (instead of exploding them into loose object files) has
been offered as a more efficient option for some time. Now the use
of cruft packs has been made the default and no longer considered
an experimental feature.
* tb/enable-cruft-packs-by-default:
repository.h: drop unused `gc_cruft_packs`
builtin/gc.c: make `gc.cruftPacks` enabled by default
t/t9300-fast-import.sh: prepare for `gc --cruft` by default
t/t6500-gc.sh: add additional test cases
t/t6500-gc.sh: refactor cruft pack tests
t/t6501-freshen-objects.sh: prepare for `gc --cruft` by default
t/t5304-prune.sh: prepare for `gc --cruft` by default
builtin/gc.c: ignore cruft packs with `--keep-largest-pack`
builtin/repack.c: fix incorrect reference to '-C'
pack-write.c: plug a leak in stage_tmp_packfiles()
The on-disk reverse index that allows mapping from the pack offset
to the object name for the object stored at the offset has been
enabled by default.
* tb/pack-revindex-on-disk:
t: invert `GIT_TEST_WRITE_REV_INDEX`
config: enable `pack.writeReverseIndex` by default
pack-revindex: introduce `pack.readReverseIndex`
pack-revindex: introduce GIT_TEST_REV_INDEX_DIE_ON_DISK
pack-revindex: make `load_pack_revindex` take a repository
t5325: mark as leak-free
pack-write.c: plug a leak in stage_tmp_packfiles()
The function `stage_tmp_packfiles()` generates a filename to use for
staging the contents of what will become the pack's ".mtimes" file.
The name is generated in `write_mtimes_file()` and the result is
returned back to `stage_tmp_packfiles()` which uses it to rename the
temporary file into place via `rename_tmp_packfiles()`.
`write_mtimes_file()` returns a `const char *`, indicating that callers
are not expected to free its result (similar to, e.g., `oid_to_hex()`).
But callers are expected to free its result, so this return type is
incorrect.
Change the function's signature to return a non-const `char *`, and free
it at the end of `stage_tmp_packfiles()`.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The function `stage_tmp_packfiles()` generates a filename to use for
staging the contents of what will become the pack's ".rev" file.
The name is generated in `write_rev_file_order()` (via its caller
`write_rev_file()`) in a string buffer, and the result is returned back
to `stage_tmp_packfiles()` which uses it to rename the temporary file
into place via `rename_tmp_packfiles()`.
That name is not visible outside of `stage_tmp_packfiles()`, so it can
(and should) be `free()`'d at the end of that function. We can't free it
in `rename_tmp_packfile()` since not all of its `source` arguments are
unreachable after calling it.
Instead, simply free() `rev_tmp_name` at the end of
`stage_tmp_packfiles()`.
(Note that the same leak exists for `mtimes_tmp_name`, but we do not
address it in this commit).
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Acked-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
cache.h's nature of a dumping ground of includes prevented it from
being included in some compat/ files, forcing us into a workaround
of having a double forward declaration of the read_in_full() function
(see commit 14086b0a13 ("compat/pread.c: Add a forward declaration to
fix a warning", 2007-11-17)). Now that we have moved functions like
read_in_full() from cache.h to wrapper.h, and wrapper.h isn't littered
with unrelated and scary #defines, get rid of the extra forward
declaration and just have compat/pread.c include wrapper.h.
Signed-off-by: Elijah Newren <newren@gmail.com>
Acked-by: Calvin Wan <calvinwan@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Several files were including cache.h solely to get other headers, such
as trace.h and trace2.h. Since the last few commits have modified
files to make these dependencies more explicit, the inclusion of cache.h
is no longer needed in several cases. Remove it.
Signed-off-by: Elijah Newren <newren@gmail.com>
Acked-by: Calvin Wan <calvinwan@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Dozens of files made use of gettext functions, without explicitly
including gettext.h. This made it more difficult to find which files
could remove a dependence on cache.h. Make C files explicitly include
gettext.h if they are using it.
However, while compat/fsmonitor/fsm-ipc-darwin.c should also gain an
include of gettext.h, it was left out to avoid conflicting with an
in-flight topic.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
These files are already included; we do not need to include them again
Signed-off-by: Seija Kijin <doremylover123@gmail.com>
Acked-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
write_mtimes_file() takes an mtimes parameter as its first option, but
the only caller passes a NULL constant. Drop this parameter to simplify
logic. This can be reverted if that parameter is needed in the future.
Signed-off-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Now that the `.mtimes` format is defined, supplement the pack-write API
to be able to conditionally write an `.mtimes` file along with a pack by
setting an additional flag and passing an oidmap that contains the
timestamps corresponding to each object in the pack.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are three definitions of an identical function which converts
`the_hash_algo` into either 1 (for SHA-1) or 2 (for SHA-256). There is a
copy of this function for writing both the commit-graph and
multi-pack-index file, and another inline definition used to write the
.rev header.
Consolidate these into a single definition in chunk-format.h. It's not
clear that this is the best header to define this function in, but it
should do for now.
(Worth noting, the .rev caller expects a 4-byte unsigned, but the other
two callers work with a single unsigned byte. The consolidated version
uses the latter type, and lets the compiler widen it when required).
Another caller will be added in a subsequent patch.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This structure will be used to communicate the per-object mtimes when
writing a cruft pack. Here, we need the full packing_data structure
because the mtime information is stored in an array there, not on the
individual object_entry's themselves (to avoid paying the overhead in
structure width for operations which do not generate a cruft pack).
We haven't passed this information down before because one of the two
callers (in bulk-checkin.c) does not have a packing_data structure at
all. In that case (where no cruft pack will be generated), NULL is
passed instead.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This commit introduces the infrastructure for the core.fsync
configuration knob. The repository components we want to sync
are identified by flags so that we can turn on or off syncing
for specific components.
If core.fsyncObjectFiles is set and the core.fsync configuration
also includes FSYNC_COMPONENT_LOOSE_OBJECT, we will fsync any
loose objects. This picks the strictest data integrity behavior
if core.fsync and core.fsyncObjectFiles are set to conflicting values.
This change introduces the currently unused fsync_component
helper, which will be used by a later patch that adds fsyncing to
the refs backend.
Actual configuration and documentation of the fsync components
list are in other patches in the series to separate review of
the underlying mechanism from the policy of how it's configured.
Helped-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The order in which various files that make up a single (conceptual)
packfile has been reevaluated and straightened up. This matters in
correctness, as an incomplete set of files must not be shown to a
running Git.
* tb/pack-finalize-ordering:
pack-objects: rename .idx files into place after .bitmap files
pack-write: split up finish_tmp_packfile() function
builtin/index-pack.c: move `.idx` files into place last
index-pack: refactor renaming in final()
builtin/repack.c: move `.idx` files into place last
pack-write.c: rename `.idx` files after `*.rev`
pack-write: refactor renaming in finish_tmp_packfile()
bulk-checkin.c: store checksum directly
pack.h: line-wrap the definition of finish_tmp_packfile()
The code that optionally creates the *.rev reverse index file has
been optimized to avoid needless computation when it is not writing
the file out.
* ab/reverse-midx-optim:
pack-write: skip *.rev work when not writing *.rev
Split up the finish_tmp_packfile() function and use the split-up version
in pack-objects.c in preparation for moving the step of renaming the
*.idx file later as part of a function change.
Since the only other caller of finish_tmp_packfile() was in
bulk-checkin.c, and it won't be needing a change to its *.idx renaming,
provide a thin wrapper for the old function as a static function in that
file. If other callers end up needing the simpler version it could be
moved back to "pack-write.c" and "pack.h".
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We treat the presence of an `.idx` file as the indicator of whether or
not it's safe to use a packfile. But `finish_tmp_packfile()` (which is
responsible for writing the pack and moving the temporary versions of
all of its auxiliary files into place) is inconsistent about the write
order.
Specifically, it moves the `.rev` file into place after the `.idx`,
leaving open the possibility to open a pack which looks "ready" (because
the `.idx` file exists and is readable) but appears momentarily to not
have a `.rev` file. This causes Git to fall back to generating the
pack's reverse index in memory.
Though racy, this amounts to an unnecessary slow-down at worst, and
doesn't affect the correctness of the resulting reverse index.
Close this race by moving the .rev file into place before moving the
.idx file into place.
This still leaves the issue of `.idx` files being renamed into place
before the auxiliary `.bitmap` file is renamed when in pack-object.c's
write_pack_file() "write_bitmap_index" is true. That race will be
addressed in subsequent commits.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Refactor the renaming in finish_tmp_packfile() into a helper function.
The callers are now expected to pass a "name_buffer" ending in
"pack-OID." instead of the previous "pack-", we then append "pack",
"idx" or "rev" to it.
By doing the strbuf_setlen() in rename_tmp_packfile() we reuse the
buffer and avoid the repeated allocations we'd get if that function had
its own temporary "struct strbuf".
This approach of reusing the buffer does make the last user in
pack-object.c's write_pack_file() slightly awkward, since we needlessly
do a strbuf_setlen() before calling strbuf_release() for consistency. In
subsequent changes we'll move that bitmap writing code around, so let's
not skip the strbuf_setlen() now.
The previous strbuf_reset() idiom originated with 5889271114
(finish_tmp_packfile():use strbuf for pathname construction,
2014-03-03), which in turn was a minimal adjustment of pre-strbuf code
added in 0e990530ae (finish_tmp_packfile(): a helper function,
2011-10-28).
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fix a performance regression introduced in a587b5a786 (pack-write.c:
extract 'write_rev_file_order', 2021-03-30) and stop needlessly
allocating the "pack_order" array and sorting it with
"pack_order_cmp()", only to throw that work away when we discover that
we're not writing *.rev files after all.
This redundant work was not present in the original version of this
code added in 8ef50d9958 (pack-write.c: prepare to write 'pack-*.rev'
files, 2021-01-25). There we'd call write_rev_file() from
e.g. finish_tmp_packfile(), but we'd "return NULL" early in
write_rev_file() if not doing a "WRITE_REV" or "WRITE_REV_VERIFY".
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add and apply a semantic patch for using xopen() instead of calling
open(2) and die() or die_errno() explicitly. This makes the error
messages more consistent and shortens the code.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
An on-disk reverse-index to map the in-pack location of an object
back to its object name across multiple packfiles is introduced.
* tb/reverse-midx:
midx.c: improve cache locality in midx_pack_order_cmp()
pack-revindex: write multi-pack reverse indexes
pack-write.c: extract 'write_rev_file_order'
pack-revindex: read multi-pack reverse indexes
Documentation/technical: describe multi-pack reverse indexes
midx: make some functions non-static
midx: keep track of the checksum
midx: don't free midx_name early
midx: allow marking a pack as preferred
t/helper/test-read-midx.c: add '--show-objects'
builtin/multi-pack-index.c: display usage on unrecognized command
builtin/multi-pack-index.c: don't enter bogus cmd_mode
builtin/multi-pack-index.c: split sub-commands
builtin/multi-pack-index.c: define common usage with a macro
builtin/multi-pack-index.c: don't handle 'progress' separately
builtin/multi-pack-index.c: inline 'flags' with options
Existing callers provide the reverse index code with an array of 'struct
pack_idx_entry *'s, which is then sorted by pack order (comparing the
offsets of each object within the pack).
Prepare for the multi-pack index to write a .rev file by providing a way
to write the reverse index without an array of pack_idx_entry (which the
MIDX code does not have).
Instead, callers can invoke 'write_rev_index_positions()', which takes
an array of uint32_t's. The ith entry in this array specifies the ith
object's (in index order) position within the pack (in pack order).
Expose this new function for use in a later patch, and rewrite the
existing write_rev_file() in terms of this new function.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The approach to "fsck" the incoming objects in "index-pack" is
attractive for performance reasons (we have them already in core,
inflated and ready to be inspected), but fundamentally cannot be
applied fully when we receive more than one pack stream, as a tree
object in one pack may refer to a blob object in another pack as
".gitmodules", when we want to inspect blobs that are used as
".gitmodules" file, for example. Teach "index-pack" to emit
objects that must be inspected later and check them in the calling
"fetch-pack" process.
* jt/transfer-fsck-across-packs:
fetch-pack: print and use dangling .gitmodules
fetch-pack: with packfile URIs, use index-pack arg
http-fetch: allow custom index-pack args
http: allow custom index-pack args
Teach index-pack to print dangling .gitmodules links after its "keep" or
"pack" line instead of declaring an error, and teach fetch-pack to check
such lines printed.
This allows the tree side of the .gitmodules link to be in one packfile
and the blob side to be in another without failing the fsck check,
because it is now fetch-pack which checks such objects after all
packfiles have been downloaded and indexed (and not index-pack on an
individual packfile, as it is before this commit).
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This patch prepares for callers to be able to write reverse index files
to disk.
It adds the necessary machinery to write a format-compliant .rev file
from within 'write_rev_file()', which is called from
'finish_tmp_packfile()'.
Similar to the process by which the reverse index is computed in memory,
these new paths also have to sort a list of objects by their offsets
within a packfile. These new paths use a qsort() (as opposed to a radix
sort), since our specialized radix sort requires a full revindex_entry
struct per object, which is more memory than we need to allocate.
The qsort is obviously slower, but the theoretical slowdown would
require a repository with a large amount of objects, likely implying
that the time spent in, say, pack-objects during a repack would dominate
the overall runtime.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
write_promisor_file() already uses xfopen(), so it would die
if the file cannot be opened for writing. To be consistent
with this behavior and not overlook issues, let's also die if
there are errors when we are actually writing to the file.
Suggested-by: Jeff King <peff@peff.net>
Suggested-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Let's replace the 2 different pieces of code that write a
promisor file in 'builtin/repack.c' and 'fetch-pack.c'
with a new function called 'write_promisor_file()' in
'pack-write.c' and 'pack.h'.
This might also help us in the future, if we want to put
back the ref names and associated hashes that were in
the promisor files we are repacking in 'builtin/repack.c'
as suggested by a NEEDSWORK comment just above the code
we are refactoring.
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Call hashwrite_be64() to write a 64-bit value instead of open-coding it
using htonl() and hashwrite(). This shortens the code, gets rid of a
buffer and several magic numbers, and makes the intent clearer.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
hashwrite() already buffers writes, so pass the fanout table entries
individually via hashwrite_be32(), which also does the endianess
conversion for us. This avoids a memory copy, shortens the code and
reduces the number of magic numbers.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Call hashwrite_be32() instead of open-coding it. This shortens the code
a bit and makes it easier to read.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The index-pack documentation explicitly states that the pack
name is derived from the sorted list of object names, but
since commit 1190a1acf8 ("pack-objects: name pack files
after trailer hash") that isn't true anymore.
Be less explicit in the docs as to what the exact output is,
and just say that it's whatever goes into the pack name.
Also update a comment on write_idx_file() since it no longer
modifies the sha1 variable (it's const now anyway), as noted
by Junio.
Fixes: 1190a1acf8 ("pack-objects: name pack files after trailer hash")
Signed-off-by: Johannes Berg <johannes@sipsolutions.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Pack checksums always use the current hash algorithm in use, so switch
from sha1_to_hex to hash_to_hex.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This rounds out the previous three patches, covering the
inequality logic for the "hash" variant of the functions.
As with the previous three, the accompanying code changes
are the mechanical result of applying the coccinelle patch;
see those patches for more discussion.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Using the more restrictive oideq() should, in the long run,
give the compiler more opportunities to optimize these
callsites. For now, this conversion should be a complete
noop with respect to the generated code.
The result is also perhaps a little more readable, as it
avoids the "zero is equal" idiom. Since it's so prevalent in
C, I think seasoned programmers tend not to even notice it
anymore, but it can sometimes make for awkward double
negations (e.g., we can drop a few !!oidcmp() instances
here).
This patch was generated almost entirely by the included
coccinelle patch. This mechanical conversion should be
completely safe, because we check explicitly for cases where
oidcmp() is compared to 0, which is what oideq() is doing
under the hood. Note that we don't have to catch "!oidcmp()"
separately; coccinelle's standard isomorphisms make sure the
two are treated equivalently.
I say "almost" because I did hand-edit the coccinelle output
to fix up a few style violations (it mostly keeps the
original formatting, but sometimes unwraps long lines).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If we want to use a hashfile on the temporary file for a lockfile, then
we need finalize_hashfile() to fully write the trailing hash but also keep
the file descriptor open.
Do this by adding a new CSUM_HASH_IN_STREAM flag along with a functional
change that checks this flag before writing the checksum to the stream.
This differs from previous behavior since it would be written if either
CSUM_CLOSE or CSUM_FSYNC is provided.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The hashclose() method behaves very differently depending on the flags
parameter. In particular, the file descriptor is not always closed.
Perform a simple rename of "hashclose()" to "finalize_hashfile()" in
preparation for functional changes.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Rename struct sha1file to struct hashfile, along with all of its related
functions.
The transformation in this commit was made by global search-and-replace.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Convert various uses of hardcoded 20- and 40-based numbers to use
the_hash_algo, along with direct calls to SHA-1. Adjust the names of
variables to refer to "hash" instead of "sha1".
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When a caller tries to read a particular set of bytes via
read_in_full(), there are three possible outcomes:
1. An error, in which case -1 is returned and errno is
set.
2. A short read, in which fewer bytes are returned and
errno is unspecified (we never saw a read error, so we
may have some random value from whatever syscall failed
last).
3. The full read completed successfully.
Many callers handle cases 1 and 2 together by just checking
the result against the requested size. If their combined
error path looks at errno (e.g., by calling die_errno), they
may report a nonsense value.
Let's fix these sites by having them distinguish between the
two error cases. That avoids the random errno confusion, and
lets us give more detailed error messages.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Convert struct pack_idx_entry to use struct object_id by changing the
definition and applying the following semantic patch, plus the standard
object_id transforms:
@@
struct pack_idx_entry E1;
@@
- E1.sha1
+ E1.oid.hash
@@
struct pack_idx_entry *E1;
@@
- E1->sha1
+ E1->oid.hash
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The odb_mkstemp() function expects the caller to provide a
fixed buffer to write the resulting tempfile name into. But
it creates the template using snprintf without checking the
return value. This means we could silently truncate the
filename.
In practice, it's unlikely that the truncation would end in
the template-pattern that mkstemp needs to open the file. So
we'd probably end up failing either way, unless the path was
specially crafted.
The simplest fix would be to notice the truncation and die.
However, we can observe that most callers immediately
xstrdup() the result anyway. So instead, let's switch to
using a strbuf, which is easier for them (and isn't a big
deal for the other 2 callers, who can just strbuf_release
when they're done with it).
Note that many of the callers used static buffers, but this
was purely to avoid putting a large buffer on the stack. We
never passed the static buffers out of the function, so
there's no complicated memory handling we need to change.
Signed-off-by: Jeff King <peff@peff.net>
The odb_mkstemp function does not return an error; it dies
on failure instead. But many of its callers compare the
resulting descriptor against -1 and die themselves.
Mostly this is just pointless, but it does raise a question
when looking at the callers: if they show the results of the
"template" buffer after a failure, what's in it? The answer
is: it doesn't matter, because it cannot happen.
So let's make that clear by removing the bogus error checks.
In bitmap_writer_finish(), we can drop the error-handling
code entirely. In the other two cases, it's shared with the
open() in another code path; we can just move the
error-check next to that open() call.
And while we're at it, let's flesh out the function's
docstring a bit to make the error behavior clear.
Signed-off-by: Jeff King <peff@peff.net>
