Tim Ansell discovered his Debian server didn't permit git-daemon to
use as much memory as it needed to handle cloning a project with
a 128 MiB packfile. Filtering the strace provided by Tim of the
rev-list child showed this gem of a sequence:
open("./objects/pack/pack-*.pack", O_RDONLY|O_LARGEFILE <unfinished ...>
<... open resumed> ) = 5
OK, so the packfile is fd 5...
mmap2(NULL, 33554432, PROT_READ, MAP_PRIVATE, 5, 0 <unfinished ...>
<... mmap2 resumed> ) = 0xb5e2d000
and we mapped one 32 MiB window from it at position 0...
mmap2(NULL, 31020635, PROT_READ, MAP_PRIVATE, 5, 0x6000 <unfinished ...>
<... mmap2 resumed> ) = -1 ENOMEM (Cannot allocate memory)
And we asked for another window further into the file. But got
denied. In Tim's case this was due to a resource limit on the
git-daemon process, and its children.
Now where are we in the code? We're down inside use_pack(),
after we have called unuse_one_window() enough times to make sure
we stay within our allowed maximum window size. However since we
didn't unmap the prior window at 0xb5e2d000 we aren't exceeding
the current limit (which probably was just the defaults).
But we're actually down inside xmmap()...
So we release the window we do have (by calling release_pack_memory),
assuming there is some memory pressure...
munmap(0xb5e2d000, 33554432 <unfinished ...>
<... munmap resumed> ) = 0
close(5 <unfinished ...>
<... close resumed> ) = 0
And that was the last window in this packfile. So we closed it.
Way to go us. Our xmmap did not expect release_pack_memory to
close the fd its about to map...
mmap2(NULL, 31020635, PROT_READ, MAP_PRIVATE, 5, 0x6000 <unfinished ...>
<... mmap2 resumed> ) = -1 EBADF (Bad file descriptor)
And so the Linux kernel happily tells us f' off.
write(2, "fatal: ", 7 <unfinished ...>
<... write resumed> ) = 7
write(2, "Out of memory? mmap failed: Bad "..., 47 <unfinished ...>
<... write resumed> ) = 47
And we report the bad file descriptor error, and not the ENOMEM,
and die, claiming we are out of memory. But actually that mmap
should have succeeded, as we had enough memory for that window,
seeing as how we released the prior one.
Originally when I developed the sliding window mmap feature I had
this exact same bug in fast-import, and I dealt with it by handing
in the struct packed_git* we want to open the new window for, as the
caller wasn't prepared to reopen the packfile if unuse_one_window
closed it. The same is true here from xmmap, but the caller doesn't
have the struct packed_git* handy. So I'm using the file descriptor
instead to perform the same test.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
I stumbled across this in the context of the fchmod 0444 patch.
At first, I was going to unlink and call error like the two subsequent
tests do, but a failed write (above) provokes a "die", so I made
this do the same. This is testing for a write failure, after all.
Signed-off-by: Jim Meyering <jim@meyering.net>
Signed-off-by: Junio C Hamano <junkio@cox.net>
When some operations are interrupted (or "die()'d" or crashed) then the
partial object/pack/index file may remain around. Make it more obvious
in their name that those files are temporary stuff and can be cleaned up
if no operation is in progress.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
When creating a new object, we use "deflate(stream, Z_FINISH)" in a loop
until it no longer returns Z_OK, and then we do "deflateEnd()" to finish
up business.
That should all work, but the fact is, it's not how you're _supposed_ to
use the zlib return values properly:
- deflate() should never return Z_OK in the first place, except if we
need to increase the output buffer size (which we're not doing, and
should never need to do, since we pre-allocated a buffer that is
supposed to be able to hold the output in full). So the "while()" loop
was incorrect: Z_OK doesn't actually mean "ok, continue", it means "ok,
allocate more memory for me and continue"!
- if we got an error return, we would consider it to be end-of-stream,
but it could be some internal zlib error. In short, we should check
for Z_STREAM_END explicitly, since that's the only valid return value
anyway for the Z_FINISH case.
- we never checked deflateEnd() return codes at all.
Now, admittedly, none of these issues should ever happen, unless there is
some internal bug in zlib. So this patch should make zero difference, but
it seems to be the right thing to do.
We should probablybe anal and check the return value of "deflateInit()"
too!
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Use hash_sha1_file() instead of duplicating code to compute object SHA1.
While at it make it accept a const pointer.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
The thing is, if the output buffer is empty, we should *still* actually
use the zlib routines to *unpack* that empty output buffer.
But we had a test that said "only unpack if we still expect more output".
So we wouldn't use up all the zlib stream, because we felt that we didn't
need it, because we already had all the bytes we wanted. And it was
"true": we did have all the output data. We just needed to also eat all
the input data!
We've had this bug before - thinking that we don't need to inflate()
anything because we already had it all..
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This provides a smoother degradation in performance when the cache
gets trashed due to the delta_base_cache_limit being reached. Limited
testing with really small delta_base_cache_limit values appears to confirm
this.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Currently there are 3 different ways to deal with the cache size.
Let's stick to only one. The compiler is smart enough to produce the exact
same code in those cases anyway.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
The new configuration variable core.deltaBaseCacheLimit allows the
user to control how much memory they are willing to give to Git for
caching base objects of deltas. This is not normally meant to be
a user tweakable knob; the "out of the box" settings are meant to
be suitable for almost all workloads.
We default to 16 MiB under the assumption that the cache is not
meant to consume all of the user's available memory, and that the
cache's main purpose was to cache trees, for faster path limiters
during revision traversal. Since trees tend to be relatively small
objects, this relatively small limit should still allow a large
number of objects.
On the other hand we don't want the cache to start storing 200
different versions of a 200 MiB blob, as this could easily blow
the entire address space of a 32 bit process.
We evict OBJ_BLOB from the cache first (credit goes to Junio) as
we want to favor OBJ_TREE within the cache. These are the objects
that have the highest inflate() startup penalty, as they tend to
be small and thus don't have that much of a chance to ammortize
that penalty over the entire data.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
A malloc() + memcpy() will always be faster than mmap() +
malloc() + inflate(). If the data is already there it is
certainly better to copy it straight away.
With this patch below I can do 'git log drivers/scsi/ >
/dev/null' about 7% faster. I bet it might be even more on
those platforms with bad mmap() support.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This trivial 256-entry delta_base cache improves performance for some
loads by a factor of 2.5 or so.
Instead of always re-generating the delta bases (possibly over and over
and over again), just cache the last few ones. They often can get re-used.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This doesn't change any code, it just creates a point for where we'd
actually do the caching of delta bases that have been generated.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Especially with the new index format to come, it is more appropriate
to encapsulate more into check_packed_git_idx() and assume less of the
index format in struct packed_git.
To that effect, the index_base is renamed to index_data with void * type
so it is not used directly but other pointers initialized with it. This
allows for a couple pointer cast removal, as well as providing a better
generic name to grep for when adding support for new index versions or
formats.
And index_data is declared const too while at it.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
When accessing objects, we first look for them in packs that
are linked together in the reverse order of discovery.
Since younger packs tend to contain more recent objects, which
are more likely to be accessed often, and local packs tend to
contain objects more relevant to our specific projects, sort the
list of packs before starting to access them. In addition,
favoring local packs over the ones borrowed from alternates can
be a win when alternates are mounted on network file systems.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Some systems have sizeof(off_t) == 8 while sizeof(size_t) == 4.
This implies that we are able to access and work on files whose
maximum length is around 2^63-1 bytes, but we can only malloc or
mmap somewhat less than 2^32-1 bytes of memory.
On such a system an implicit conversion of off_t to size_t can cause
the size_t to wrap, resulting in unexpected and exciting behavior.
Right now we are working around all gcc warnings generated by the
-Wshorten-64-to-32 option by passing the off_t through xsize_t().
In the future we should make xsize_t on such problematic platforms
detect the wrapping and die if such a file is accessed.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Not all platforms have declared 'unsigned long' to be a 64 bit value,
but we want to support a 64 bit packfile (or close enough anyway)
in the near future as some projects are getting large enough that
their packed size exceeds 4 GiB.
By using off_t, the POSIX type that is declared to mean an offset
within a file, we support whatever maximum file size the underlying
operating system will handle. For most modern systems this is up
around 2^60 or higher.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
As we permit up to 2^32-1 objects in a single packfile we cannot
use a signed int to represent the object offset within a packfile,
after 2^31-1 objects we will start seeing negative indexes and
error out or compute bad addresses within the mmap'd index.
This is a minor cleanup that does not introduce any significant
logic changes. It is roach free.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
We shouldn't attempt to assign constant strings into char*, as the
string is not writable at runtime. Likewise we should always be
treating unsigned values as unsigned values, not as signed values.
Most of these are very straightforward. The only exception is the
(unnecessary) xstrdup/free in builtin-branch.c for the detached
head case. Since this is a user-level interactive type program
and that particular code path is executed no more than once, I feel
that the extra xstrdup call is well worth the easy elimination of
this warning.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
If an index is corrupt, or is simply too new for us to understand,
we were leaking the mmap that held the entire content of the index.
This could be a considerable size on large projects, given that
the index is at least 24 bytes * nr_objects.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Because we are currently cheating and never supplying the delta base
for an OBJ_OFS_DELTA we get a random SHA-1 in the delta base field.
Instead lets clear the hash out so its at least all 0's. This is
somewhat more obvious that something fishy is going on, like we
don't actually have the SHA-1 of the base handy. :)
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
We did not detect broken loose object files, either when
underlying inflate() signalled the breakage, nor inflate()
finished and we had garbage trailing at the end. We do better
now.
We also make unpack_sha1_file() a static function to
sha1_file.c, since it is not used by anybody outside.
Signed-off-by: Junio C Hamano <junkio@cox.net>
We currently have two parallel notation for dealing with object types
in the code: a string and a numerical value. One of them is obviously
redundent, and the most used one requires more stack space and a bunch
of strcmp() all over the place.
This is an initial step for the removal of the version using a char array
found in object reading code paths. The patch is unfortunately large but
there is no sane way to split it in smaller parts without breaking the
system.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Sometime typename() is used, sometimes type_names[] is accessed directly.
Let's enforce typename() all the time which allows for validating the
type.
Also let's add a function to go from a name to a type and use it instead
of manual memcpy() when appropriate.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
First there are too many offsets there and it is getting confusing.
So 'offset' is now 'curpos' to distinguish from other offsets like
'obj_offset'.
Then structures like x = foo(x, &y) are now done as y = foo(&x).
It looks more natural that the result y be returned directly and
x be passed as reference to be updated in place. This has the effect
of reducing some line length and removing a few, needing a bit less
stack space, and it even reduces the compiled code size.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Let's have hdr be a simple char pointer/array when possible, and let's
reduce its storage to 32 bytes. Especially for sha1_loose_object_info()
where 128 bytes is way excessive and wastes extra CPU cycles inflating.
The object type is already restricted to 10 bytes in parse_sha1_header()
and the size, even if it is 64 bits, will fit in 20 decimal numbers. So
32 bytes is plenty.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
* lt/crlf:
Teach core.autocrlf to 'git apply'
t0020: add test for auto-crlf
Make AutoCRLF ternary variable.
Lazy man's auto-CRLF
* jc/apply-config:
t4119: test autocomputing -p<n> for traditional diff input.
git-apply: guess correct -p<n> value for non-git patches.
git-apply: notice "diff --git" patch again
Fix botched "leak fix"
t4119: add test for traditional patch and different p_value
apply: fix memory leak in prefix_one()
git-apply: require -p<n> when working in a subdirectory.
git-apply: do not lose cwd when run from a subdirectory.
Teach 'git apply' to look at $HOME/.gitconfig even outside of a repository
Teach 'git apply' to look at $GIT_DIR/config
This ensures that a given area is mapped at most twice, and greatly
reduces the virtual address space usage.
Signed-off-by: Alexandre Julliard <julliard@winehq.org>
Acked-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
It currently does NOT know about file attributes, so it does its
conversion purely based on content. Maybe that is more in the "git
philosophy" anyway, since content is king, but I think we should try to do
the file attributes to turn it off on demand.
Anyway, BY DEFAULT it is off regardless, because it requires a
[core]
AutoCRLF = true
in your config file to be enabled. We could make that the default for
Windows, of course, the same way we do some other things (filemode etc).
But you can actually enable it on UNIX, and it will cause:
- "git update-index" will write blobs without CRLF
- "git diff" will diff working tree files without CRLF
- "git checkout" will write files to the working tree _with_ CRLF
and things work fine.
Funnily, it actually shows an odd file in git itself:
git clone -n git test-crlf
cd test-crlf
git config core.autocrlf true
git checkout
git diff
shows a diff for "Documentation/docbook-xsl.css". Why? Because we have
actually checked in that file *with* CRLF! So when "core.autocrlf" is
true, we'll always generate a *different* hash for it in the index,
because the index hash will be for the content _without_ CRLF.
Is this complete? I dunno. It seems to work for me. It doesn't use the
filename at all right now, and that's probably a deficiency (we could
certainly make the "is_binary()" heuristics also take standard filename
heuristics into account).
I don't pass in the filename at all for the "index_fd()" case
(git-update-index), so that would need to be passed around, but this
actually works fine.
NOTE NOTE NOTE! The "is_binary()" heuristics are totally made-up by yours
truly. I will not guarantee that they work at all reasonable. Caveat
emptor. But it _is_ simple, and it _is_ safe, since it's all off by
default.
The patch is pretty simple - the biggest part is the new "convert.c" file,
but even that is really just basic stuff that anybody can write in
"Teaching C 101" as a final project for their first class in programming.
Not to say that it's bug-free, of course - but at least we're not talking
about rocket surgery here.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Here's a patch that I think we can merge right now. There may be
other places that need this, but this at least points out the
three places that read/write working tree files for git
update-index, checkout and diff respectively. That should cover
a lot of it [jc: git-apply uses an entirely different codepath
both for reading and writing].
Some day we can actually implement it. In the meantime, this
points out a place for people to start. We *can* even start with
a really simple "we do CRLF conversion automatically, regardless
of filename" kind of approach, that just look at the data (all
three cases have the _full_ file data already in memory) and
says "ok, this is text, so let's convert to/from DOS format
directly".
THAT somebody can write in ten minutes, and it would already
make git much nicer on a DOS/Windows platform, I suspect.
And it would be totally zero-cost if you just make it a config
option (but please make it dynamic with the _default_ just being
0/1 depending on whether it's UNIX/Windows, just so that UNIX
people can _test_ it easily).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
The new interface allows an application to temporarily hash a
small number of objects and pretend that they are available in
the object store without actually writing them.
Signed-off-by: Junio C Hamano <junkio@cox.net>
If open_packed_git failed it may have been because the packfile
actually exists and is readable, but some sort of verification
did not pass. In this case open_packed_git left pack_fd filled
in, as the file descriptor is valid. We don't want to leak the
file descriptor, nor do we want to allow someone in the future
to use this packed_git.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Matthias Lederhofer identified a race condition where a Git reader
process was able to locate an object in a packed_git index, but
was then preempted while a `git repack -a -d` ran and completed.
By the time the reader was able to seek in the packfile to get the
object data, the packfile no longer existed on disk.
In this particular case the reader process did not attempt to
open the packfile before it was deleted, so it did not already
have the pack_fd field popuplated. With the packfile itself gone,
there was no way for the reader to open it and fetch the data.
I'm fixing the race condition by teaching find_pack_entry to ignore
a packed_git whose packfile is not currently open and which cannot
be opened. If none of the currently known packs can supply the
object, we will return 0 and the caller will decide the object is
not available. If this is the first attempt at finding an object,
the caller will reprepare_packed_git and try again. If it was
the second attempt, the caller will typically return NULL back,
and an error message about a missing object will be reported.
This patch does not address the situation of a reader which is
being starved out by a tight sequence of `git repack -a -d` runs.
In this particular case the reader will try twice, probably fail
both times, and declare the object in question cannot be found.
As it is highly unlikely that a real world `git repack -a -d` can
complete faster than a reader can open a packfile, so I don't think
this is a huge concern.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Because I want to reuse open_packed_git in a context where I don't
want the process to die if the packfile in question is bogus, I'm
changing its behavior to return error("...") rather than die("...")
when it detects something is wrong with the packfile it was given.
Right now we still must die out of use_pack should open_packed_git
fail, as none of use_pack's callers are prepared to handle a failure
from that function.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
After staring at the comment and the associated for loop, I
realized the comment was completely bogus. The section of
code its talking about is trying to avoid duplicate mapping
of the same packfile.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
There is little point in having the linked list insertion code
appearing in install_packed_git, and then again just 30 lines
further down in the same file.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
We used to call find_pack_entry() twice from read_sha1_file() in order
to avoid printing an error message, when the object did not exist. This
is fixed by moving the call to error() to the only place it really
could be called.
Signed-off-by: Peter Eriksen <s022018@student.dtu.dk>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Way back when Junio developed the 64 bit index topic he came up
with a means of changing the .idx file format so that older Git
clients would recognize that they don't understand the file and
refuse to read it, while newer clients could tell the difference
between the old-style and new-style .idx files. Unfortunately
this wasn't recorded anywhere.
This change documents how we might go about changing the .idx
file format by using a special signature in the first four bytes.
Credit (and possible blame) goes completely to Junio for thinking
up this technique.
The change also modifies the error message of the current Git code
so that users get a recommendation to upgrade their Git software
should this version or later encounter a new-style .idx which it
cannot process. We already do this for the .pack files, but since
we usually process the .idx files first its important that these
files are recognized and encourage an upgrade.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Originally I introduced read_or_die for the purpose of reading
the pack header and trailer, and I was too lazy to print proper
error messages.
Linus Torvalds <torvalds@osdl.org>:
> For a read error, at the very least you have to say WHICH FILE
> couldn't be read, because it's usually a matter of some file just
> being too short, not some system-wide problem.
and of course Linus is right. Make it so.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
With the new-and-improved write_in_full() semantics, where a partial write
simply always returns a real error (and always sets 'errno' when that
happens, including for the disk full case), a lot of the callers of
write_in_full() were just unnecessarily complex.
In particular, there's no reason to ever check for a zero length or
return: if the length was zero, we'll return zero, otherwise, if a disk
full resulted in the actual write() system call returning zero the
write_in_full() logic would have correctly turned that into a negative
return value, with 'errno' set to ENOSPC.
I really wish every "write_in_full()" user would just check against "<0"
now, but this fixes the nasty and stupid ones.
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
