Update error handling for codepath that deals with corrupt loose
objects.
* jk/loose-object-info-report-error:
index-pack: detect local corruption in collision check
sha1_loose_object_info: return error for corrupted objects
When we notice that we have a local copy of an incoming
object, we compare the two objects to make sure we haven't
found a collision. Before we get to the actual object
bytes, though, we compare the type and size from
sha1_object_info().
If our local object is corrupted, then the type will be
OBJ_BAD, which obviously will not match the incoming type,
and we'll report "SHA1 COLLISION FOUND" (with capital
letters and everything). This is confusing, as the problem
is not a collision but rather local corruption. We should
report that instead (just like we do if reading the rest of
the object content fails a few lines later).
Note that we _could_ just ignore the error and mark it as a
non-collision. That would let you "git fetch" to replace a
corrupted object. But it's not a very reliable method for
repairing a repository. The earlier want/have negotiation
tries to get the other side to omit objects we already have,
and it would not realize that we are "missing" this
corrupted object. So we're better off complaining loudly
when we see corruption, and letting the user take more
drastic measures to repair (like making a full clone
elsewhere and copying the pack into place).
Note that the test sets transfer.unpackLimit in the
receiving repository so that we use index-pack (which is
what does the collision check). Normally for such a small
push we'd use unpack-objects, which would simply try to
write the loose object, and discard the new one when we see
that there's already an old one.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We'd prefer to avoid unchecked snprintf calls because
truncation can lead to unexpected results.
These are all cases where truncation shouldn't ever happen,
because the input to snprintf is fixed in size. That makes
them candidates for xsnprintf(), but it's simpler still to
just use the heap, and then nobody has to wonder if "100" is
big enough.
We'll use xstrfmt() where possible, and a strbuf when we need
the resulting size or to reuse the same buffer in a loop.
Signed-off-by: Jeff King <peff@peff.net>
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>
The final() function accepts a NULL value for certain
parameters, and falls back to writing into a reusable "name"
buffer, and then either:
1. For "keep_name", requiring all uses to do "keep_name ?
keep_name : name.buf". This is awkward, and it's easy
to accidentally look at the maybe-NULL keep_name.
2. For "final_index_name" and "final_pack_name", aliasing
those pointers to the "name" buffer. This is easier to
use, but the aliased pointers become invalid after the
buffer is reused (this isn't a bug now, but it's a
potential pitfall).
One way to make this safer would be to introduce an extra
pointer to do the aliasing, and have its lifetime match the
validity of the "name" buffer. But it's still easy to
accidentally use the wrong name (i.e., to use
"final_pack_name" instead of the aliased pointer).
Instead, let's use three separate buffers that will remain
valid through the function. That makes it safe to alias the
pointers and use them consistently. The extra allocations
shouldn't matter, as this function is not performance
sensitive.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In several places we write the name of the pack filename
into a fixed-size buffer using snprintf(), but do not check
the return value. As a result, a very long object directory
could cause us to quietly truncate the pack filename
(potentially leading to a corrupted repository, as a newly
written packfile could be missing its .pack extension).
We can use odb_pack_name() to do this with a strbuf (and
shorten the code, as well).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The odb_pack_keep() function generates the name of a .keep
file and opens it. This has two problems:
1. It requires a fixed-size buffer to create the filename
and doesn't notice when the result is truncated.
2. Of the two callers, one sometimes wants to open a
filename it already has, which makes things awkward (it
has to do so manually, and skips the leading-directory
creation).
Instead, let's have odb_pack_keep() just open the file.
Generating the name isn't hard, and a future patch will
switch callers over to odb_pack_name() anyway.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
You can run "git index-pack path/to/foo.pack" outside of a
repository to generate an index file, or just to verify the
contents. There's no point in doing a collision check, since
we obviously do not have any objects to collide with.
The current code will blindly look in .git/objects based on
the result of setup_git_env(). That effectively gives us the
right answer (since we won't find any objects), but it's a
waste of time, and it conflicts with our desire to
eventually get rid of the "fallback to .git" behavior of
setup_git_env().
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The index-pack builtin is marked as RUN_SETUP_GENTLY,
because it's perfectly fine to index a pack in the
filesystem outside of any repository. However, --stdin mode
will write the result to the object database, which does not
make sense outside of a repository. Doing so creates a bogus
".git" directory with nothing in it except the newly-created
pack and its index.
Instead, let's flag this as an error and abort.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Convert two more qsort(3) calls to QSORT to reduce code size and for
better safety and consistency.
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Apply the semantic patch contrib/coccinelle/qsort.cocci to the code
base, replacing calls of qsort(3) with QSORT. The resulting code is
shorter and supports empty arrays with NULL pointers.
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When receiving a pack-file, it can be useful to abort the
`git index-pack`, if the pack-file is too big.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git push" and "git clone" learned to give better progress meters
to the end user who is waiting on the terminal.
* jk/push-progress:
receive-pack: send keepalives during quiet periods
receive-pack: turn on connectivity progress
receive-pack: relay connectivity errors to sideband
receive-pack: turn on index-pack resolving progress
index-pack: add flag for showing delta-resolution progress
clone: use a real progress meter for connectivity check
check_connected: add progress flag
check_connected: relay errors to alternate descriptor
check_everything_connected: use a struct with named options
check_everything_connected: convert to argv_array
rev-list: add optional progress reporting
check_everything_connected: always pass --quiet to rev-list
After a client has sent us the complete pack, we may spend
some time processing the data and running hooks. If the
client asked us to be quiet, receive-pack won't send any
progress data during the index-pack or connectivity-check
steps. And hooks may or may not produce their own progress
output. In these cases, the network connection is totally
silent from both ends.
Git itself doesn't care about this (it will wait forever),
but other parts of the system (e.g., firewalls,
load-balancers, etc) might hang up the connection. So we'd
like to send some sort of keepalive to let the network and
the client side know that we're still alive and processing.
We can use the same trick we did in 05e9515 (upload-pack:
send keepalive packets during pack computation, 2013-09-08).
Namely, we will send an empty sideband data packet every `N`
seconds that we do not relay any stderr data over the
sideband channel. As with 05e9515, this means that we won't
bother sending keepalives when there's actual progress data,
but will kick in when it has been disabled (or if there is a
lull in the progress data).
The concept is simple, but the details are subtle enough
that they need discussing here.
Before the client sends us the pack, we don't want to do any
keepalives. We'll have sent our ref advertisement, and we're
waiting for them to send us the pack (and tell us that they
support sidebands at all).
While we're receiving the pack from the client (or waiting
for it to start), there's no need for keepalives; it's up to
them to keep the connection active by sending data.
Moreover, it would be wrong for us to do so. When we are the
server in the smart-http protocol, we must treat our
connection as half-duplex. So any keepalives we send while
receiving the pack would potentially be buffered by the
webserver. Not only does this make them useless (since they
would not be delivered in a timely manner), but it could
actually cause a deadlock if we fill up the buffer with
keepalives. (It wouldn't be wrong to send keepalives in this
phase for a full-duplex connection like ssh; it's simply
pointless, as it is the client's responsibility to speak).
As soon as we've gotten all of the pack data, then the
client is waiting for us to speak, and we should start
keepalives immediately. From here until the end of the
connection, we send one any time we are not otherwise
sending data.
But there's a catch. Receive-pack doesn't know the moment
we've gotten all the data. It passes the descriptor to
index-pack, who reads all of the data, and then starts
resolving the deltas. We have to communicate that back.
To make this work, we instruct the sideband muxer to enable
keepalives in three phases:
1. In the beginning, not at all.
2. While reading from index-pack, wait for a signal
indicating end-of-input, and then start them.
3. Afterwards, always.
The signal from index-pack in phase 2 has to come over the
stderr channel which the muxer is reading. We can't use an
extra pipe because the portable run-command interface only
gives us stderr and stdout.
Stdout is already used to pass the .keep filename back to
receive-pack. We could also send a signal there, but then we
would find out about it in the main thread. And the
keepalive needs to be done by the async muxer thread (since
it's the one writing sideband data back to the client). And
we can't reliably signal the async thread from the main
thread, because the async code sometimes uses threads and
sometimes uses forked processes.
Therefore the signal must come over the stderr channel,
where it may be interspersed with other random
human-readable messages from index-pack. This patch makes
the signal a single NUL byte. This is easy to parse, should
not appear in any normal stderr output, and we don't have to
worry about any timing issues (like seeing half the signal
bytes in one read(), and half in a subsequent one).
This is a bit ugly, but it's simple to code and should work
reliably.
Another option would be to stop using an async thread for
muxing entirely, and just poll() both stderr and stdout of
index-pack from the main thread. This would work for
index-pack (because we aren't doing anything useful in the
main thread while it runs anyway). But it would make the
connectivity check and the hook muxers much more
complicated, as they need to simultaneously feed the
sub-programs while reading their stderr.
The index-pack phase is the only one that needs this
signaling, so it could simply behave differently than the
other two. That would mean having two separate
implementations of copy_to_sideband (and the keepalive
code), though. And it still doesn't get rid of the
signaling; it just means we can write a nicer message like
"END_OF_INPUT" or something on stdout, since we don't have
to worry about separating it from the stderr cruft.
One final note: this signaling trick is only done with
index-pack, not with unpack-objects. There's no point in
doing it for the latter, because by definition it only kicks
in for a small number of objects, where keepalives are not
as useful (and this conveniently lets us avoid duplicating
the implementation).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The index-pack command has two progress meters: one for
"receiving objects", and one for "resolving deltas". You get
neither by default, or both with "-v".
But for a push through receive-pack, we would want only the
"resolving deltas" phase, _not_ the "receiving objects"
progress. There are two reasons for this.
One is simply that existing clients are already printing
"writing objects" progress at the same time. Arguably
"receiving" from the far end is more useful, because it
tells you what has actually gotten there, as opposed to what
might be stuck in a buffer somewhere between the client and
server. But that would require a protocol extension to tell
clients not to print their progress. Possible, but
complexity for little gain.
The second reason is much more important. In a full-duplex
connection like git-over-ssh, we can print progress while
the pack is incoming, and it will immediately get to the
client. But for a half-duplex connection like git-over-http,
we should not say anything until we have received the full
request. Anything we write is subject to being stuck in a
buffer by the webserver. Worse, we can end up in a deadlock
if that buffer fills up.
So our best bet is to avoid writing anything that isn't a
small fixed size until we've received the full pack.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
unpack_entry_data() receives an off_t value from unpack_raw_entry(),
which could be larger than unsigned long on 32-bit systems with large
file support. Correct the type so truncation does not happen. This
only affects bad object reporting though.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Use the right type for offsets in this case, off_t, which makes a
difference on 32-bit systems with large file support, and change
formatting code accordingly.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
On 32-bit systems with large file support, one entry could be larger
than 4GB and overflow "len". Correct it so we can unpack a full entry.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Mark several messages for translation.
* va/i18n-misc-updates:
i18n: unpack-trees: avoid substituting only a verb in sentences
i18n: builtin/pull.c: split strings marked for translation
i18n: builtin/pull.c: mark placeholders for translation
i18n: git-parse-remote.sh: mark strings for translation
i18n: branch: move comment for translators
i18n: branch: unmark string for translation
i18n: builtin/rm.c: remove a comma ',' from string
i18n: unpack-trees: mark strings for translation
i18n: builtin/branch.c: mark option for translation
i18n: index-pack: use plural string instead of normal one
Mark several messages for translation.
* va/i18n-misc-updates:
i18n: unpack-trees: avoid substituting only a verb in sentences
i18n: builtin/pull.c: split strings marked for translation
i18n: builtin/pull.c: mark placeholders for translation
i18n: git-parse-remote.sh: mark strings for translation
i18n: branch: move comment for translators
i18n: branch: unmark string for translation
i18n: builtin/rm.c: remove a comma ',' from string
i18n: unpack-trees: mark strings for translation
i18n: builtin/branch.c: mark option for translation
i18n: index-pack: use plural string instead of normal one
Git could output "completed with 1 local objects", but in this
case using "object" instead of "objects" is the correct form.
Use Q_() instead of _().
Signed-off-by: Vasco Almeida <vascomalmeida@sapo.pt>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The code to read the pack data using the offsets stored in the pack
idx file has been made more carefully check the validity of the
data in the idx.
* jk/pack-idx-corruption-safety:
sha1_file.c: mark strings for translation
use_pack: handle signed off_t overflow
nth_packed_object_offset: bounds-check extended offset
t5313: test bounds-checks of corrupted/malicious pack/idx files
These are automatically named by replacing .pack suffix in the
name of the packfile. Add a small helper to do so, as I'll be
adding another one soonish.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When 592ce208 (index-pack: use strip_suffix to avoid magic numbers,
2014-06-30) refactored the code to derive names of .idx and .keep
files from the name of .pack file, a copy-and-paste typo crept in,
mistakingly attempting to create and store the keep message file in
the .idx file we just created, instead of .keep file.
As we create the .keep file with O_CREAT|O_EXCL, and we do so after
we write the .idx file, we luckily do not clobber the .idx file, but
because we deliberately ignored EEXIST when creating .keep file
(which is justifiable because only the existence of .keep file
matters), nobody noticed this mistake so far.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If a pack .idx file has a corrupted offset for an object, we
may try to access an offset in the .idx or .pack file that
is larger than the file's size. For the .pack case, we have
use_pack() to protect us, which realizes the access is out
of bounds. But if the corrupted value asks us to look in the
.idx file's secondary 64-bit offset table, we blindly add it
to the mmap'd index data and access arbitrary memory.
We can fix this with a simple bounds-check compared to the
size we found when we opened the .idx file.
Note that there's similar code in index-pack that is
triggered only during "index-pack --verify". To support
both, we pull the bounds-check into a separate function,
which dies when it sees a corrupted file.
It would be nice if we could return an error, so that the
pack code could try to find a good copy of the object
elsewhere. Currently nth_packed_object_offset doesn't have
any way to return an error, but it could probably use "0" as
a sentinel value (since no object can start there). This is
the minimal fix, and we can improve the resilience later on
top.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If our size computation overflows size_t, we may allocate a
much smaller buffer than we expected and overflow it. It's
probably impossible to trigger an overflow in most of these
sites in practice, but it is easy enough convert their
additions and multiplications into overflow-checking
variants. This may be fixing real bugs, and it makes
auditing the code easier.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Each of these cases can be converted to use ALLOC_ARRAY or
REALLOC_ARRAY, which has two advantages:
1. It automatically checks the array-size multiplication
for overflow.
2. It always uses sizeof(*array) for the element-size,
so that it can never go out of sync with the declared
type of the array.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Convert all instances of get_object_hash to use an appropriate reference
to the hash member of the oid member of struct object. This provides no
functional change, as it is essentially a macro substitution.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Jeff King <peff@peff.net>
struct object is one of the major data structures dealing with object
IDs. Convert it to use struct object_id instead of an unsigned char
array. Convert get_object_hash to refer to the new member as well.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Jeff King <peff@peff.net>
Convert most instances where the sha1 member of struct object is
dereferenced to use get_object_hash. Most instances that are passed to
functions that have versions taking struct object_id, such as
get_sha1_hex/get_oid_hex, or instances that can be trivially converted
to use struct object_id instead, are not converted.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Jeff King <peff@peff.net>
We generally use 32-byte buffers to format git's "type size"
header fields. These should not generally overflow unless
you can produce some truly gigantic objects (and our types
come from our internal array of constant strings). But it is
a good idea to use xsnprintf to make sure this is the case.
Note that we slightly modify the interface to
write_sha1_file_prepare, which nows uses "hdrlen" as an "in"
parameter as well as an "out" (on the way in it stores the
allocated size of the header, and on the way out it returns
the ultimate size of the header).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since 5a688fe4 ("core.sharedrepository = 0mode" should set, not
loosen, 2009-03-25), we kept reminding ourselves:
NEEDSWORK: this should be renamed to finalize_temp_file() as
"moving" is only a part of what it does, when no patch between
master to pu changes the call sites of this function.
without doing anything about it. Let's do so.
The purpose of this function was not to move but to finalize. The
detail of the primarily implementation of finalizing was to link the
temporary file to its final name and then to unlink, which wasn't
even "moving". The alternative implementation did "move" by calling
rename(2), which is a fun tangent.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Allow ignoring fsck errors on specific set of known-to-be-bad
objects, and also tweaking warning level of various kinds of non
critical breakages reported.
* js/fsck-opt:
fsck: support ignoring objects in `git fsck` via fsck.skiplist
fsck: git receive-pack: support excluding objects from fsck'ing
fsck: introduce `git fsck --connectivity-only`
fsck: support demoting errors to warnings
fsck: document the new receive.fsck.<msg-id> options
fsck: allow upgrading fsck warnings to errors
fsck: optionally ignore specific fsck issues completely
fsck: disallow demoting grave fsck errors to warnings
fsck: add a simple test for receive.fsck.<msg-id>
fsck: make fsck_tag() warn-friendly
fsck: handle multiple authors in commits specially
fsck: make fsck_commit() warn-friendly
fsck: make fsck_ident() warn-friendly
fsck: report the ID of the error/warning
fsck (receive-pack): allow demoting errors to warnings
fsck: offer a function to demote fsck errors to warnings
fsck: provide a function to parse fsck message IDs
fsck: introduce identifiers for fsck messages
fsck: introduce fsck options
Disable "have we lost a race with competing repack?" check while
receiving a huge object transfer that runs index-pack.
* jk/index-pack-reduce-recheck:
index-pack: avoid excessive re-reading of pack directory
When c6458e60 (index-pack: kill union delta_base to save memory,
2015-04-18) attempted to reduce the memory footprint of index-pack,
one of the key thing it did was to keep track of ref-deltas and
ofs-deltas separately.
In fix_unresolved_deltas(), however it forgot that it now wants to
look only at ref deltas in one place. The code allocated an array
for nr_unresolved, which is sum of number of ref- and ofs-deltas
minus nr_resolved, which may be larger or smaller than the number
ref-deltas. Depending on nr_resolved, this was either under or over
allocating.
Also, the old code before this change had to use 'i' and 'n' because
some of the things we see in the (old) deltas[] array we scanned
with 'i' would not make it into the sorted_by_pos[] array in the old
world order, but now because you have only ref delta in a separate
ref_deltas[] array, they increment lock&step. We no longer need
separate variables. And most importantly, we shouldn't pass the
nr_unresolved parameter, as this number does not play a role in the
working of this helper function.
Helped-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Disable "have we lost a race with competing repack?" check while
receiving a huge object transfer that runs index-pack.
* jk/index-pack-reduce-recheck:
index-pack: avoid excessive re-reading of pack directory
For example, missing emails in commit and tag objects can be demoted to
mere warnings with
git config receive.fsck.missingemail=warn
The value is actually a comma-separated list.
In case that the same key is listed in multiple receive.fsck.<msg-id>
lines in the config, the latter configuration wins (this can happen for
example when both $HOME/.gitconfig and .git/config contain message type
settings).
As git receive-pack does not actually perform the checks, it hands off
the setting to index-pack or unpack-objects in the form of an optional
argument to the --strict option.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Just like the diff machinery, we are about to introduce more settings,
therefore it makes sense to carry them around as a (pointer to a) struct
containing all of them.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
An earlier optimization broke index-pack for a large object
transfer; this fixes it before the breakage hits any released
version.
* nd/slim-index-pack-memory-usage:
index-pack: fix truncation of off_t in comparison
Since 45e8a74 (has_sha1_file: re-check pack directory before
giving up, 2013-08-30), we spend extra effort for
has_sha1_file to give the right answer when somebody else is
repacking. Usually this effort does not matter, because
after finding that the object does not exist, the next step
is usually to die().
However, some code paths make a large number of
has_sha1_file checks which are _not_ expected to return 1.
The collision test in index-pack.c is such a case. On a
local system, this can cause a performance slowdown of
around 5%. But on a system with high-latency system calls
(like NFS), it can be much worse.
This patch introduces a "quick" flag to has_sha1_file which
callers can use when they would prefer high performance at
the cost of false negatives during repacks. There may be
other code paths that can use this, but the index-pack one
is the most obviously critical, so we'll start with
switching that one.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit c6458e6 (index-pack: kill union delta_base to save
memory, 2015-04-18) refactored the comparison functions used
in sorting and binary searching our delta list. The
resulting code does something like:
int cmp_offsets(off_t a, off_t b)
{
return a - b;
}
This works most of the time, but produces nonsensical
results when the difference between the two offsets is
larger than what can be stored in an "int". This can lead to
unresolved deltas if the packsize is larger than 2G (even on
64-bit systems, an int is still typically 32 bits):
$ git clone git://github.com/mozilla/gecko-dev
Cloning into 'gecko-dev'...
remote: Counting objects: 4800161, done.
remote: Compressing objects: 100% (178/178), done.
remote: Total 4800161 (delta 88), reused 0 (delta 0), pack-reused 4799978
Receiving objects: 100% (4800161/4800161), 2.21 GiB | 3.26 MiB/s, done.
Resolving deltas: 99% (3808820/3811944), completed with 0 local objects.
fatal: pack has 3124 unresolved deltas
fatal: index-pack failed
We can fix it by doing direct comparisons between the
offsets and returning constants; the callers only care about
the sign of the comparison, not the magnitude.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Memory usage of "git index-pack" has been trimmed by tens of
per-cent.
* nd/slim-index-pack-memory-usage:
index-pack: kill union delta_base to save memory
index-pack: reduce object_entry size to save memory
Once we know the number of objects in the input pack, we allocate an
array of nr_objects of struct delta_entry. On x86-64, this struct is
32 bytes long. The union delta_base, which is part of struct
delta_entry, provides enough space to store either ofs-delta (8 bytes)
or ref-delta (20 bytes).
Because ofs-delta encoding is more efficient space-wise and more
performant at runtime than ref-delta encoding, Git packers try to use
ofs-delta whenever possible, and it is expected that objects encoded
as ref-delta are minority.
In the best clone case where no ref-delta object is present, we waste
(20-8) * nr_objects bytes because of this union. That's about 38MB out
of 100MB for deltas[] with 3.4M objects, or 38%. deltas[] would be
around 62MB without the waste.
This patch attempts to eliminate that. deltas[] array is split into
two: one for ofs-delta and one for ref-delta. Many functions are also
duplicated because of this split. With this patch, ofs_deltas[] array
takes 51MB. ref_deltas[] should remain unallocated in clone case (0
bytes). This array grows as we see ref-delta. We save about half in
this case, or 25% of total bookkeeping.
The saving is more than the calculation above because some padding in
the old delta_entry struct is removed. ofs_delta_entry is 16 bytes,
including the 4 bytes padding. That's 13MB for padding, but packing
the struct could break platforms that do not support unaligned
access. If someone on 32-bit is really low on memory and only deals
with packs smaller than 2G, using 32-bit off_t would eliminate the
padding and save 27MB on top.
A note about ofs_deltas allocation. We could use ref_deltas memory
allocation strategy for ofs_deltas. But that probably just adds more
overhead on top. ofs-deltas are generally the majority (1/2 to 2/3) in
any pack. Incremental realloc may lead to too many memcpy. And if we
preallocate, say 1/2 or 2/3 of nr_objects initially, the growth rate
of ALLOC_GROW() could make this array larger than nr_objects, wasting
more memory.
Brought-up-by: Matthew Sporleder <msporleder@gmail.com>
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
