Get rid of sort_comparator() as it impose a run time double indirect
function call for little compile time type checking gain.
Also get rid of create_sorted_list() as it only has one user which would
as well be just fine doing its sorting locally. Eventually the list of
deltifiable objects might be shorter than the whole object list.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Objects that have delta "children" from pack data reuse must consider the
depth of their deepest child when they try to deltify themselves for those
children not to become too deep.
However, in the context of a "thin" pack, the delta children depth was
skipped entirely on the presumption that the pack was always going to be
exploded on the receiving end, hence the delta length wasn't an issue.
Now that we keep received packs as is and reuse pack data when repacking,
those packs do contain delta chains that are longer than expected. Worse,
those delta chain may even grow longer when the pack is further repacked
into another thin pack for a subsequent transmission.
So this patch restores strict delta length even for thin packs, and it
moves check_delta_limit() usage directly in the delta loop where it is
needed. This way the delta_limit can be removed from struct object_entry
as well. Oh and the initial value was wrong too.
The progress_interval() function was moved to a more logical location in
the process.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Before finding best delta combinations, we sort objects by name hash,
then by size, then by their position in memory. Then we walk the list
backwards to test delta candidates.
We hope that a bigger size usually means a newer objects. But a bigger
address in memory does not mean a newer object. So the last comparison
must be reversed.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Let's avoid some cycles when there is no base to test against, and avoid
unnecessary object lookups.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This function used to call locate_object_entry_hash() _twice_ per added
object while only once should suffice. Let's reorganize that code a bit.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This replaces the inflate validation with a CRC validation when reusing
data from a pack which uses index version 2. That makes repacking much
safer against corruptions, and it should be a bit faster too.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This is necessary for testing the new capabilities in some automated
way without having an actual 4GB+ pack.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Pack index version 2 goes as follows:
- 8 bytes of header with signature and version.
- 256 entries of 4-byte first-level fan-out table.
- Table of sorted 20-byte SHA1 records for each object in pack.
- Table of 4-byte CRC32 entries for raw pack object data.
- Table of 4-byte offset entries for objects in the pack if offset is
representable with 31 bits or less, otherwise it is an index in the next
table with top bit set.
- Table of 8-byte offset entries indexed from previous table for offsets
which are 32 bits or more (optional).
- 20-byte SHA1 checksum of sorted object names.
- 20-byte SHA1 checksum of the above.
The object SHA1 table is all contiguous so future pack format that would
contain this table directly won't require big changes to the code. It is
also tighter for slightly better cache locality when looking up entries.
Support for large packs exceeding 31 bits in size won't impose an index
size bloat for packs within that range that don't need a 64-bit offset.
And because newer objects which are likely to be the most frequently used
are located at the beginning of the pack, they won't pay the 64-bit offset
lookup at run time either even if the pack is large.
Right now an index version 2 is created only when the biggest offset in a
pack reaches 31 bits. It might be a good idea to always use index version
2 eventually to benefit from the CRC32 it contains when reusing pack data
while repacking.
[jc: with the "oops" fix to keep track of the last offset correctly]
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
The most important optimization for performance when repacking is the
ability to reuse data from a previous pack as is and bypass any delta
or even SHA1 computation by simply copying the raw data from one pack
to another directly.
The problem with this is that any data corruption within a copied object
would go unnoticed and the new (repacked) pack would be self-consistent
with its own checksum despite containing a corrupted object. This is a
real issue that already happened at least once in the past.
In some attempt to prevent this, we validate the copied data by inflating
it and making sure no error is signaled by zlib. But this is still not
perfect as a significant portion of a pack content is made of object
headers and references to delta base objects which are not deflated and
therefore not validated when repacking actually making the pack data reuse
still not as safe as it could be.
Of course a full SHA1 validation could be performed, but that implies
full data inflating and delta replaying which is extremely costly, which
cost the data reuse optimization was designed to avoid in the first place.
So the best solution to this is simply to store a CRC32 of the raw pack
data for each object in the pack index. This way any object in a pack can
be validated before being copied as is in another pack, including header
and any other non deflated data.
Why CRC32 instead of a faster checksum like Adler32? Quoting Wikipedia:
Jonathan Stone discovered in 2001 that Adler-32 has a weakness for very
short messages. He wrote "Briefly, the problem is that, for very short
packets, Adler32 is guaranteed to give poor coverage of the available
bits. Don't take my word for it, ask Mark Adler. :-)" The problem is
that sum A does not wrap for short messages. The maximum value of A for
a 128-byte message is 32640, which is below the value 65521 used by the
modulo operation. An extended explanation can be found in RFC 3309,
which mandates the use of CRC32 instead of Adler-32 for SCTP, the
Stream Control Transmission Protocol.
In the context of a GIT pack, we have lots of small objects, especially
deltas, which are likely to be quite small and in a size range for which
Adler32 is dimed not to be sufficient. Another advantage of CRC32 is the
possibility for recovery from certain types of small corruptions like
single bit errors which are the most probable type of corruptions.
OK what this patch does is to compute the CRC32 of each object written to
a pack within pack-objects. It is not written to the index yet and it is
obviously not validated when reusing pack data yet either.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Change a few size and offset variables to more appropriate type, then
add overflow tests on those offsets. This prevents any bad data to be
generated/processed if off_t happens to not be large enough to handle
some big packs.
Better be safe than sorry.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This patch introduces the MSB() macro to obtain the desired number of
most significant bits from a given variable independently of the variable
type.
It is then used to better implement the overflow test on the OBJ_OFS_DELTA
base offset variable with the property of always working correctly
regardless of the type/size of that variable.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
The coming index format change doesn't allow for the number of objects
to be determined from the size of the index file directly. Instead, Let's
initialize a field in the packed_git structure with the object count when
the index is validated since the count is always known at that point.
While at it let's reorder some struct packed_git fields to avoid padding
due to needed 64-bit alignment for some of them.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Let's avoid the open coded pack index reference in pack-object and use
nth_packed_object_sha1() instead. This will help encapsulating index
format differences in one place.
And while at it there is no reason to copy SHA1's over and over while a
direct pointer to it in the index will do just fine.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Acked-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This removes slightly more lines than it adds, but the real reason for
doing this is that future optimizations will require more setup of the
tree descriptor, and so we want to do it in one place.
Also renamed the "desc.buf" field to "desc.buffer" just to trigger
compiler errors for old-style manual initializations, making sure I
didn't miss anything.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Since we have the "tree_entry_len()" helper function these days, and
don't need to do a full strlen(), there's no point in saving the path
length - it's just redundant information.
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>
Always use an off_t value in pack-objects anytime we are dealing
with an offset to some data within a packfile.
Also fixed a minor uintmax_t that was incorrectly defined before.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
As we technically try to support up to a maximum of 2**32-1 objects
in a single packfile we should act like it and use unsigned 32 bit
integers for all of our object counts and progress output.
This change does not modify everything in pack-objects that probably
needs to change to fully support the maximum of 2**32-1 objects.
I'm intentionally breaking the improvements into slightly smaller
commits to make them easier to follow.
No logic change should be occuring here, with the exception that
some comparsions will now work properly when the number of objects
exceeds 2**31-1.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
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>
If a repository ever gets in a situation where there are too many
packs (more than 60 or so), perhaps because of frequent use of
git-fetch -k or incremental git-repack, then it becomes impossible to
fully repack the repository with git-repack -a. That command just
dies with the cryptic message
fatal: too many internal rev-list options
This message comes from git-pack-objects, which is passed one command
line option like --unpacked=pack-<SHA1>.pack for each pack file to be
repacked. However, the current code has a static limit of 64 command
line arguments and just aborts if more arguments are passed to it.
Fix this by dynamically allocating the array of command line
arguments, and doubling the size each time it overflows.
Signed-off-by: Roland Dreier <roland@digitalvampire.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Previous step converted use of strncmp() with literal string
mechanically even when the result is only used as a boolean:
if (!strncmp("foo", arg, 3)) ==> if (!(-prefixcmp(arg, "foo")))
This step manually cleans them up to read:
if (!prefixcmp(arg, "foo"))
Signed-off-by: Junio C Hamano <junkio@cox.net>
This mechanically converts strncmp() to use prefixcmp(), but only when
the parameters match specific patterns, so that they can be verified
easily. Leftover from this will be fixed in a separate step, including
idiotic conversions like
if (!strncmp("foo", arg, 3))
=>
if (!(-prefixcmp(arg, "foo")))
This was done by using this script in px.perl
#!/usr/bin/perl -i.bak -p
if (/strncmp\(([^,]+), "([^\\"]*)", (\d+)\)/ && (length($2) == $3)) {
s|strncmp\(([^,]+), "([^\\"]*)", (\d+)\)|prefixcmp($1, "$2")|;
}
if (/strncmp\("([^\\"]*)", ([^,]+), (\d+)\)/ && (length($1) == $3)) {
s|strncmp\("([^\\"]*)", ([^,]+), (\d+)\)|(-prefixcmp($2, "$1"))|;
}
and running:
$ git grep -l strncmp -- '*.c' | xargs perl px.perl
Signed-off-by: Junio C Hamano <junkio@cox.net>
The code calls use_pack() to make that the variably encoded
offset fits in the mmap'ed window, but it forgot that the
operation gives the pointer to the beginning of the asked
region.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Junio noticed that 'non-trivial' pushes were failing if executed
using the sliding window mmap changes. This was somewhat difficult
to track down as the failure was appearing randomly.
It turns out this was a failure caused by the delta base reference
(either ref or offset format) spanning over the end of a mmap window.
The error in pack-objects was we were not recalling use_pack
after the object header was unpacked, and therefore we did not
get the promise of at least 20 bytes in the buffer for the delta
base parsing. This would case later memcmp() calls to walk into
unassigned address space at the end of the window.
The reason Junio and I had hard time tracking this down in current
Git repositories is we were both probably packing with offset deltas,
which minimized the odds of the delta base reference spanning over
the end of the mmap window. Stepping back and repacking with
version 1.3.3 (which only supported reference deltas) increased
the likelyhood of seeing the bug.
The correct technique (as used in sha1_file.c) is to invoke
use_pack() after unpack_object_header_gently to ensure we have
enough data available for the delta base decoding.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
When multiple mmaps start getting used for all pack file access it
is not possible to get all data associated with a specific object
in one contiguous memory region. This limitation prevents simply
passing a single address and length to SHA1_Update or to inflate.
Instead we need to loop until we have processed all data of interest.
As we loop over the data we are always interested in reusing the same
window 'cursor', as the prior window will no longer be of any use
to us. This allows the use_pack() call to automatically decrement
the use count of the prior window before setting up access for us
to the next window.
Within each loop we need to make use of the available length output
parameter of use_pack() to tell us how many bytes are available in
the current memory region, as we cannot tell otherwise.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Part of the implementation concept of the sliding mmap window for
pack access is to permit multiple windows per pack to be mapped
independently. Since the inuse_cnt is associated with the mmap and
not with the file, this value is in struct pack_window and needs to
be incremented/decremented for each pack_window accessed by any code.
To faciliate that implementation we need to replace all uses of
use_packed_git() and unuse_packed_git() with a different API that
follows struct pack_window objects rather than struct packed_git.
The way this works is when we need to start accessing a pack for
the first time we should setup a new window 'cursor' by declaring
a local and setting it to NULL:
struct pack_windows *w_curs = NULL;
To obtain the memory region which contains a specific section of
the pack file we invoke use_pack(), supplying the address of our
current window cursor:
unsigned int len;
unsigned char *addr = use_pack(p, &w_curs, offset, &len);
the returned address `addr` will be the first byte at `offset`
within the pack file. The optional variable len will also be
updated with the number of bytes remaining following the address.
Multiple calls to use_pack() with the same window cursor will
update the window cursor, moving it from one window to another
when necessary. In this way each window cursor variable maintains
only one struct pack_window inuse at a time.
Finally before exiting the scope which originally declared the window
cursor we must invoke unuse_pack() to unuse the current window (which
may be different from the one that was first obtained from use_pack):
unuse_pack(&w_curs);
This implementation is still not complete with regards to multiple
windows, as only one window per pack file is supported right now.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
The idea behind the sliding mmap window pack reader implementation
is to have multiple mmap regions active against the same pack file,
thereby allowing the process to mmap in only the active/hot sections
of the pack and reduce overall virtual address space usage.
To implement this we need to refactor the mmap related data
(pack_base, pack_use_cnt) out of struct packed_git and move them
into a new struct pack_window.
We are refactoring the code to support a single struct pack_window
per packfile, thereby emulating the prior behavior of mmap'ing the
entire pack file.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This adds a new option --reflog to pack-objects and revision
machinery; do not bother documenting it for now, since this is
only useful for local repacking.
When the option is passed, objects reachable from reflog entries
are marked as interesting while computing the set of objects to
pack.
Signed-off-by: Junio C Hamano <junkio@cox.net>
This is a mechanical clean-up of the way *.c files include
system header files.
(1) sources under compat/, platform sha-1 implementations, and
xdelta code are exempt from the following rules;
(2) the first #include must be "git-compat-util.h" or one of
our own header file that includes it first (e.g. config.h,
builtin.h, pkt-line.h);
(3) system headers that are included in "git-compat-util.h"
need not be included in individual C source files.
(4) "git-compat-util.h" does not have to include subsystem
specific header files (e.g. expat.h).
Signed-off-by: Junio C Hamano <junkio@cox.net>
The final 'nr_result' and 'written' values must always be the same
otherwise we're in deep trouble. So let's remove a redundent report.
And for paranoia sake let's make sure those two variables are actually
equal after all objects are written (one never knows).
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
The heuristics to give up deltification when both the source and the
target are both in the same pack affects negatively when we are
repacking the subset of objects in the existing pack. This caused
any incremental updates to use suboptimal packs. Tweak the heuristics
to avoid this problem.
Signed-off-by: Junio C Hamano <junkio@cox.net>
This adds documentation for --progress and --all-progress, remove a
duplicate --progress handling and make usage string more readable.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Currently git-push displays progress status for the local packing of
objects to send, but nothing once it starts to push it over the
connection. Having progress status in that later case is especially
nice when pushing lots of objects over a slow network link.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This is the missing part to git-pack-objects allowing it to reuse delta
data to/from any of the two delta types. It can reuse delta from any
type, and it outputs base offsets when --allow-delta-base-offset is
provided and the base is also included in the pack. Otherwise it
outputs base sha1 references just like it always did.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This is enabled with --delta-base-offset only, and doesn't work with
pack data reuse yet.
The idea is to allow for the fetch protocol to use an extension flag
to notify the remote end that --delta-base-offset can be used with
git-pack-objects. Eventually git-repack will always provide this flag.
With this, all delta base objects are now pushed before deltas that depend
on them. This is a requirements for OBJ_OFS_DELTA. This is not a
requirement for OBJ_REF_DELTA but always doing so makes the code simpler.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This adds a new object, namely OBJ_OFS_DELTA, renames OBJ_DELTA to
OBJ_REF_DELTA to better make the distinction between those two delta
objects, and adds support for the handling of those new delta objects
in sha1_file.c only.
The OBJ_OFS_DELTA contains a relative offset from the delta object's
position in a pack instead of the 20-byte SHA1 reference to identify
the base object. Since the base is likely to be not so far away, the
relative offset is more likely to have a smaller encoding on average
than an absolute offset. And for those delta objects the base must
always be stored first because there is no way to know the distance of
later objects when streaming a pack. Hence this relative offset is
always meant to be negative.
The offset encoding is slightly denser than the one used for object
size -- credits to <linux@horizon.com> (whoever this is) for bringing
it to my attention.
This allows for pack size reduction between 3.2% (Linux-2.6) to over 5%
(linux-historic). Runtime pack access should be faster too since delta
replay does skip a search in the pack index for each delta in a chain.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Those cleanups are mainly to set the table for the support of deltas
with base objects referenced by offsets instead of sha1. This means
that many pack lookup functions are converted to take a pack/offset
tuple instead of a sha1.
This eliminates many struct pack_entry usages since this structure
carried redundent information in many cases, and it increased stack
footprint needlessly for a couple recursively called functions that used
to declare a local copy of it for every recursion loop.
In the process, packed_object_info_detail() has been reorganized as well
so to look much saner and more amenable to deltas with offset support.
Finally the appropriate adjustments have been made to functions that
depend on the above changes. But there is no functionality changes yet
simply some code refactoring at this point.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
This teaches the internal rev-list logic to understand options
that are needed for pack handling: --all, --unpacked, and --thin.
It also moves two functions from builtin-rev-list to list-objects
so that the two programs can share more code.
Signed-off-by: Junio C Hamano <junkio@cox.net>
Instead of piping the rev-list output from its standard input,
you can say:
pack-objects --all --unpacked --revs pack
and feed the rev parameters you would otherwise give the
rev-list on its command line from the standard input.
In other words:
echo 'master..next' | pack-objects --revs pack
and
rev-list --objects master..next | pack-objects pack
are equivalent.
Signed-off-by: Junio C Hamano <junkio@cox.net>
When copying from an existing pack and when copying from a loose
object with new style header, the code makes sure that the piece
we are going to copy out inflates well and inflate() consumes
the data in full while doing so.
The check to see if the xdelta really apply is quite expensive
as you described, because you would need to have the image of
the base object which can be represented as a delta against
something else.
Signed-off-by: Junio C Hamano <junkio@cox.net>
When revalidating an entry from an existing pack entry->size and
entry->type are not necessarily the size of the final object
when the entry is deltified, but for base objects they must
match.
Signed-off-by: Junio C Hamano <junkio@cox.net>
When reusing data from an existing pack and from a new style
loose objects, we used to just copy it staight into the
resulting pack. Instead make sure they are not corrupt, but
do so only when we are not streaming to stdout, in which case
the receiving end will do the validation either by unpacking
the stream or by constructing the .idx file.
Signed-off-by: Junio C Hamano <junkio@cox.net>
This abstracts away the size of the hash values when copying them
from memory location to memory location, much as the introduction
of hashcmp abstracted away hash value comparsion.
A few call sites were using char* rather than unsigned char* so
I added the cast rather than open hashcpy to be void*. This is a
reasonable tradeoff as most call sites already use unsigned char*
and the existing hashcmp is also declared to be unsigned char*.
[jc: Splitted the patch to "master" part, to be followed by a
patch for merge-recursive.c which is not in "master" yet.
Fixed the cast in the latter hunk to combine-diff.c which was
wrong in the original.
Also converted ones left-over in combine-diff.c, diff-lib.c and
upload-pack.c ]
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Introduces global inline:
hashcmp(const unsigned char *sha1, const unsigned char *sha2)
Uses memcmp for comparison and returns the result based on the length of
the hash name (a future runtime decision).
Acked-by: Alex Riesen <raa.lkml@gmail.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Nicolas Pitre