Now that its easier to craft test cases (thanks to 'data <<')
we should start to verify fast-import works as expected.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
During testing its nice to not have to feed the length of a data
chunk to the 'data' command of fast-import. Instead we would
prefer to be able to establish a data chunk much like shell's <<
operator and use a line delimiter to denote the end of the input.
So now if a data command is started as 'data <<EOF' we will look
for a terminator line containing only the string EOF on that line.
Once found, we stop the data command. Everything between the two
lines is used as the data value.
The 'data <<' syntax is slower than 'data n', as we don't know how
many bytes to expect and instead must grow our buffer on the fly.
It also has the problem that the frontend must use a string which
will not appear on a line by itself in the input, and the data
region will always end in an LF. For these reasons real import
frontends are encouraged to continue to use _only_ 'data n'.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
The --objects command line option is rather unnecessary. Internally
we allocate objects in 5000 unit blocks, ensuring that any sort
of malloc overhead is ammortized over the individual objects to
almost nothing. Since most frontends don't know how many objects
they will need for a given import run (and its hard for them to
predict without just doing the run) we probably won't see anyone
using --objects. Further since there's really no major benefit
to using the option, most frontends won't even bother supplying
it even if they could estimate the number of objects. So I'm
removing it.
The --max-objects-per-pack option was probably a mistake to even
have added in the first place. The packfile format is limited
to 4 GiB today; given that objects need at least 3 bytes of data
(and probably need even more) there's no way we are going to exceed
the limit of 1<<32-1 objects before we reach the file size limit.
So I'm removing it (to slightly reduce the complexity of the code)
before anyone gets any wise ideas and tries to use it.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Currently the pack .idx file format uses 32-bit unsigned integers
for the fan-out table and the object offsets. We had previously
defined these as 'unsigned int', but not every system will define
that type to be a 32 bit value. To ensure maximum portability we
should always use 'uint32_t'.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Previously we were using 'unsigned int' to update the hdr_entries
field of the pack header after the file had been completed and
was being hashed. This may not be 32 bits on all platforms.
Instead we want to always uint32_t.
I'm actually cheating here by just using the pack_header like the
rest of Git and letting the struct definition declare the correct
type. Right now that field is still 'unsigned int' (wrong) but a
pending change submitted by Simon 'corecode' Schubert changes it
to uint32_t. After that change is merged in fast-import will do
the right thing all of the time.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Branches are only contained by a packfile if the branch actually
had its most recent commit in that packfile. So new branches are
set to MAX_PACK_ID to ensure they don't cause their commit to list
as part of the first packfile when it closes out if the commit was
actually in existance before fast-import started.
Also corrected the type of last_commit to be umaxint_t to prevent
overflow and wraparound on very large imports. Though that is
highly unlikely to occur as we're talking 4 billion commits, which
no real project has right now.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Apparently the git convention is to declare any function which
takes no arguments as taking void. I did not do this during the
early fast-import development, but should have.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
The length of an atom string cannot be negative. So make it
explicit and declare it as an unsigned value.
The shift width in a mark table node also cannot be negative.
I'm also moving it to after the pointer arrays to prevent any
possible alignment problems on a 64 bit system.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Now that fast-import uses uintmax_t (the largest available unsigned
integer type) for marks we don't want to say its an unsigned 32
bit integer in ASCII base 10 notation. It could be much larger,
especially on 64 bit systems, and especially if a frontend uses
a very large number of marks (1 per file revision on a very, very
large import).
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
To help callers repack very large repositories into a series of
packfiles fast-import now outputs the last commits/tags it wrote to
a packfile when it prints out the packfile name. This information
can be feed to pack-objects --revs to repack. For the first pack
of an initial import this is pretty easy (just feed those SHA1s on
stdin) but for subsequent packs you want to feed the subsequent
pack's final SHA1s but also all prior pack's SHA1s prefixed with
the negation operator. This way the prior pack's data does not
get included into the subsequent pack.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Since object_count is limited to 'unsigned long' (really an
unsigned 32 bit integer value) by the pack file format we may as
well use exactly that type here in fast-import for that counter.
An earlier change by me incorrectly made it uintmax_t.
But since object_count is a counter for the current packfile only,
we don't want to output its value at the end. Instead we should
sum up the individual type counters and report that total, as that
will cover all of the packfiles.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Apparently amd64 has defined 'unsigned long' to be a 64 bit value,
which means -1 was way over the 4 GiB packfile limit. Whoops.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Much like the pack_sha1 the pack_fd is an unnecessary global
variable, we already have the fd stored in our struct packed_git
*pack_data so that the core library functions in sha1_file.c are
able to lookup and decompress object data that we have previously
written. Keeping an extra copy of this value in our own variable
is just a hold-over from earlier versions of fast-import and is
now completely unnecessary.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Because we are renaming the packfile into its file destination we
need to be sure its not open when the rename is called, otherwise
some operating systems (e.g. Windows) may prevent the rename from
occurring.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Because fast-import automatically updates all references (heads
and tags) at the end of its run the repository is corrupt unless
the objects are available in the .git/objects/pack directory prior
to the refs being modified. The easiest way to ensure that is true
is to move the packfile and its associated index directly into the
.git/objects/pack directory as soon as we have finished output to it.
But the only safe way to do this is to create the a temporary .keep
file for that pack, so we use the same tricks that index-pack uses
when its being invoked by receive-pack.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Rather than maintaing our own packfile level sha1 variable we
can make use of the one already available in struct packed_git.
Its meant for the SHA1 of the index but it can also hold the
SHA1 of the packfile itself between final checksumming of the
packfile and creation of the index.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Prior to git having read_in_full() fast-import used its own private
function yread to perform the header reading task. No sense in
keeping that around now that read_in_full is a public, stable
function.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
If a frontend wants to use a mark per file revision and per commit
and is doing a truly huge import (such as a 32 GiB SVN repository)
we may need more than 2**32 unique mark values, especially if the
frontend is unable (or unwilling) to recycle mark values. For mark
idnums we should use the largest unsigned integer type available,
hoping that will be at least 64 bits when we are compiled as a 64
bit executable. This way we may consume huge amounts of memory
storing our mark table, but we'll at least be able to process
the entire import without failing.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
If we previously were using a delta but we needed to checkpoint the
current packfile and switch to a new packfile we need to throw away
the delta and compress the raw object by itself, as delta chains
cannot span non-thin packfiles. Unfortunately the output buffer
in this case needs to grow, as the size of the compressed object
may be quite a bit larger than the size of the compressed delta.
I've also avoided recompressing the object if we are checkpointing
and we didn't use a delta. In this case the output buffer is the
correct size and has already been populated with the right data,
we just need to close out the current packfile and open a new one.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Caller scripts may want to know what packfiles the fast-import
process just wrote out for them. This is now output to stdout,
one packfile name per line, after we checkpoint each packfile.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
When the number of objects or number of bytes gets close to the limit
allowed by the packfile format (or configured on the command line by
our caller) we should automatically checkpoint the current packfile
and start a new one before writing the object out. This does however
require that we abandon the delta (if we had one) as its not valid
in a new packfile.
I also added the simple rule that if we got a delta back but the
delta itself is the same size as or larger than the uncompressed
object to ignore the delta and just store the object data. This
should avoid some really bad behavior caused by our current delta
strategy.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
When we are generating multiple packfiles at once we only need
to scan the blocks of object_entry structs which contain objects
for the current packfile. Because the most recent blocks are at
the front of the linked list, and because all new objects going
into the current file are allocated from the front of that list,
we can stop scanning for objects as soon as we identify one which
doesn't belong to the current packfile.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
If the last packfile is going to be empty (has 0 objects) then it
shouldn't be kept after the import has terminated, as there is no
point to the packfile. So rather than hashing it and making the
index file, just delete the packfile.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
To help importers which are dealing with massive amounts of data
fast-import needs to be able to close the packfile it is currently
writing to and open a new packfile for any additional data that
will be received. A new 'checkpoint' command has been introduced
which can be used by the frontend import process to force this
to occur at any time. This may be useful to ensure a very long
running import doesn't lose any work due to unexpected failures.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
There is little reason to be keeping a global duplicate_count
value when we also keep it per object type. The global counter can
easily be computed at the end, once all processing has completed.
This saves us a couple of machine instructions in an unimportant
part of code. But it looks slightly better to me to not keep
two counters around.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Now that we are starting to see some really large projects (such
as KDE or a fork of FreeBSD) get imported into Git we're running
into the upper limit on packfile object count as well as overall
byte length. The KDE and FreeBSD projects are both likely to
require more than 4 GiB to store their current history, which means
we really need multiple packfiles to handle their content.
This is a fairly simple restructuring of the internal code to help
us support creating multiple packfiles from within fast-import.
We are now adding a 5 digit incrementing suffix to the end of the
basename supplied to us by the caller, permitting up to 99,999
packs to be generated in a single fast-import run.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Now that the sha1_file.c library routines use the sliding mmap
routines to perform efficient access to portions of a packfile
I can remove that code from fast-import.c and just invoke it.
One benefit is we now have reloading support for any packfile which
uses OBJ_OFS_DELTA. Another is we have significantly less code
to maintain.
This code reuse change *requires* that fast-import generate only
an OBJ_OFS_DELTA format packfile, as there is absolutely no index
available to perform OBJ_REF_DELTA lookup in while unpacking
an object. This is probably reasonable to require as the delta
offsets result in smaller packfiles and are faster to unpack,
as no index searching is required. Its also only a temporary
requirement as users could always repack without offsets before
making the import available to older versions of Git.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
I'm bringing master in early so that the OBJ_OFS_DELTA implementation
is available as part of the topic. This way git-fast-import can
learn about this new slightly smaller and faster packfile format,
and can generate them directly rather than needing to have them be
repacked with git-pack-objects.
Due to the API changes in master during the period of development
of git-fast-import, a few minor tweaks to fast-import.c are needed
to produce a working merge. I've done them here as part of the
merge to ensure bisection always works.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Some importers may want to create a branch long before they actually
commit to it, or in some cases they may never commit to the branch
but they still need the ref to be created in the repository after
the import is complete.
This extends the 'reset ' command to automatically create a new
branch if the supplied reference isn't already known as a branch.
While I'm at it I also modified the syntax of the reset command
to terminate with an empty line, like commit and tag operate.
This just makes the command set more consistent.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Some importers are able to determine when branch merges occurred
within their source data. In these cases they will want to supply
the correct commits to fast-import so that a proper merge commit
will exist in Git. This is now supported by supplying a 'merge '
command after the commit message and optional from command.
A merge is not actually performed by fast-import, its assumed that
the frontend performed any sort of merging activity already and
that fast-import should simply be storing its result.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Apparently we did not copy the blob SHA1 into the stack variable
'sha1' when a mark is used to refer to a prior blob. This code
was not previously tested as the Mozilla CVS -> git-fast-import
program always fed us full SHA1s for modified blobs and did not
use the mark feature there.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
The new tree delta implementation caused blob SHA1s to be used
instead of a tree SHA1 when a tree was written out. This really
only appeared to happen when converting an existing file to a tree,
but may have been possible in some other situations.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Since most commits and tag objects are around the same size and we
only generate one at a time we can reuse the same buffer rather than
xmalloc'ing and free'ing the buffer every time we generate a commit.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
We only ever generate at most two tree streams at a time. Since most
trees are around the same size we can simply recycle the buffers from
one tree generation to the next rather than constantly xmalloc'ing
and free'ing them. This should perform slightly better when handling
a large number of trees as malloc has less work to do.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
We now store for every tree entry two modes and two sha1 values;
the base (aka "version 0") and the current/new (aka "version 1").
When we generate a tree object we also regenerate the prior version
object and use that as our base object for a delta. This strategy
saves a significant amount of memory as we can continue to use the
atom pool for file/directory names and only increases each tree
entry by an additional 24 bytes of memory.
Branches should automatically delta against their ancestor tree,
unless the ancestor tree is already at the delta chain limit.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Sometimes an import frontend may need to work with a temporary branch
which will actually contain many different branches over the life
of the import. This is especially useful when the frontend needs
to create a tag from a set of file versions which are otherwise
never a commit.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
When generating a very large pack file (for example close to 1 GB
in size) it may be impossible for the kernel to find a contiguous
free range within a 32 bit address space for the mapping to be
located at. This is especially problematic on large imports where
there is a lot of malloc activity occuring within the same process
and the malloc'd regions may straddle the previously mapped regions,
thereby creating large holes in the address space.
So instead we map only 128 MB of the pack at any given time.
This will likely increase the number of times the file gets mapped
(with additional system time required to update the page tables
more frequently) but will allow the program to handle packs up to
4 GB in size.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
fast-import was encounting a GPF when it ran out of free tree_entry
objects but didn't know this was the cause because the last
tree_entry wasn't terminated with a NULL pointer. The missing NULL
pointer occurred when we allocated additional entries via xmalloc
but didn't set the last tree_entry's "next" pointer to NULL.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
This option can be used to have a record of every commit, the mark
(if supplied) and branch name of the commit recorded into a log file
when the commit is generated. This log can be useful to verify the
results of an import as the commits can be compared to some source
repository matching commits through the mark value.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
The marks table can be used by the frontend to load any commit after
the import and compare it to whatever data the frontend knows about
that commit. If the mark idnums can be easily correlated to some
reference source then its relatively trivial to compare the GIT
tree to the reference to verify the accuracy of the import.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
cvs2svn has three phases: begin_commit, middle_commit, end_commit.
The ancester is computed in the middle_commit phase. So its easier
to generate a stream if the from command appears after the commit
message itself but before the file change commands.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Jon Smirl was finding it difficult to alter cvs2svn to generate
branch commands prior to the first commit of the same branch.
This change moves the 'from' command to be an optional parameter of
the 'commit' command, thereby allowing a new branch to be defined
at the moment it gets used to create the first commit on that branch.
This change makes it impossible to create a branch with no commits
on it as at least one commit is needed to register the branch.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Some architectures (e.g. SPARC) would require that we access pointers
only on pointer-sized alignments. So ensure the pool allocator
rounds out non-pointer sized allocations to the next pointer so we
don't generate bad memory addresses. This could have occurred if
we had previously allocated an atom whose string was not a whole
multiple of the pointer size, for example.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
