The copyfrom_source instruction appends data from the preimage buffer
to the end of output. Its arguments are a length and an offset
relative to the beginning of the source view.
With this change, the delta applier is able to reproduce all 5,636,613
blobs in the early history of the ASF repository. Tested with
mkfifo backflow
svn-fe <svn-asf-public-r0:940166 3<backflow |
git fast-import --cat-blob-fd=3 3>backflow
with svn-asf-public-r0:940166 produced by whatever version of
Subversion the dumps in /dump/ on svn.apache.org use (presumably
1.6.something).
Improved-by: Ramkumar Ramachandra <artagnon@gmail.com>
Improved-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Acked-by: Ramkumar Ramachandra <artagnon@gmail.com>
The copyfrom_target instruction copies appends data that is already
present in the current output view to the end of output. (The offset
argument is relative to the beginning of output produced in the
current window.)
The region copied is allowed to run past the end of the existing
output. To support that case, copy one character at a time rather
than calling memcpy or memmove. This allows copyfrom_target to be
used once to repeat a string many times. For example:
COPYFROM_DATA 2
COPYFROM_OUTPUT 10, 0
DATA "ab"
would produce the output "ababababababababababab".
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Acked-by: Ramkumar Ramachandra <artagnon@gmail.com>
By constraining the format of deltas, we can more easily detect
corruption and other breakage.
Requiring deltas not to provide unconsumed data also opens the
possibility of ignoring the declared amount of novel data and simply
streaming the data as needed to fulfill copyfrom_data requests.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Acked-by: Ramkumar Ramachandra <artagnon@gmail.com>
The copyfrom_data instruction copies a few bytes verbatim from the
novel text section of a window to the postimage.
[jn: with memory leak fix from David]
Improved-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Acked-by: Ramkumar Ramachandra <artagnon@gmail.com>
Buffer the instruction section upon encountering it for later
interpretation.
An alternative design would involve parsing the instructions
at this point and buffering them in some processed form. Using
the unprocessed form is simpler.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Acked-by: Ramkumar Ramachandra <artagnon@gmail.com>
Each window of an svndiff0-format delta includes a section for novel
text to be copied to the postimage (in the order it appears in the
window, possibly interspersed with other data).
Slurp in this data when encountering it. It is not actually necessary
to do so --- it would be just as easy to copy from delta to output
as part of interpreting the relevant instructions --- but this way,
the code that interprets svndiff0 instructions can proceed very
quickly because it does not require I/O.
Subversion's svndiff0 parser rejects deltas that do not consume all
the novel text that was provided. Omit that check for now so we can
test the new functionality right away, rather than waiting to learn
instructions that consume data.
Do check for truncated data sections. Subversion's parser rejects
deltas that end in the middle of a declared novel-text section, so it
should be safe for us to reject them, too.
Improved-by: Ramkumar Ramachandra <artagnon@gmail.com>
Improved-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Acked-by: Ramkumar Ramachandra <artagnon@gmail.com>
The source view offset heading each svndiff0 window represents a
number of bytes past the beginning of the preimage. Together with the
source view length, it dictates to the delta applier what portion of
the preimage instructions will refer to. Read that portion right away
using the sliding window code.
Maybe some day we will use mmap to read data more lazily.
Subversion's implementation tolerates source view offsets pointing
past the end of the preimage file but we do not, for simplicity.
This does not teach the delta applier to read instructions or copy
data from the source view. Deltas that could produce nonempty output
will still be rejected.
Improved-by: Ramkumar Ramachandra <artagnon@gmail.com>
Improved-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Acked-by: Ramkumar Ramachandra <artagnon@gmail.com>
Each window in a subversion delta (svndiff0-format file) starts with a
window header, consisting of five integers with variable-length
representation:
source view offset
source view length
output length
instructions length
auxiliary data length
Parse it. The result is not usable for deltas with nonempty postimage
yet; in fact, this only adds support for deltas without any
instructions or auxiliary data. This is a good place to stop, though,
since that little support lets us add some simple passing tests
concerning error handling to the test suite.
Improved-by: Ramkumar Ramachandra <artagnon@gmail.com>
Improved-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
A delta in the subversion delta (svndiff0) format consists of the
magic bytes SVN\0 followed by a sequence of windows of a certain well
specified format (starting with five integers).
Add an svndiff0_apply function and test-svn-fe -d commandline tool to
parse such a delta in the special case of not including any windows.
Later patches will add features to turn this into a fully functional
delta applier for svn-fe to use to parse the streams produced by
"svnrdump dump" and "svnadmin dump --deltas".
The content of symlinks starts with the word "link " in Subversion's
worldview, so we need to be able to prepend that text to input for the
sake of delta application. So initialization of the input state of
the delta preimage is left to the calling program, giving callers a
chance to seed the buffer with text of their choice.
Improved-by: Ramkumar Ramachandra <artagnon@gmail.com>
Improved-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
buffer_read_binary is a thin wrapper around fread, but its signature
is wrong:
- fread can fill an arbitrary in-memory buffer. buffer_read_binary
is limited to buffers whose size is representable by a 32-bit
integer.
- The result from fread is the number of bytes actually read.
buffer_read_binary only reports the number of bytes read by
incrementing sb->len by that amount and returns void.
Fix both: let buffer_read_binary accept a size_t instead of uint32_t
for the number of bytes to read and as a convenience return the number
of bytes actually read.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Each section of a Subversion-format delta only requires examining (and
keeping in random-access memory) a small portion of the preimage. At
any moment, this portion starts at a certain file offset and has a
well-defined length, and as the delta is applied, the portion advances
from the beginning to the end of the preimage. Add a move_window
function to keep track of this view into the preimage.
You can use it like this:
buffer_init(f, NULL);
struct sliding_view window = SLIDING_VIEW_INIT(f);
move_window(&window, 3, 7); /* (1) */
move_window(&window, 5, 5); /* (2) */
move_window(&window, 12, 2); /* (3) */
strbuf_release(&window.buf);
buffer_deinit(f);
The data structure is called sliding_view instead of _window to
prevent confusion with svndiff0 Windows.
In this example, (1) reads 10 bytes and discards the first 3;
(2) discards the first 2, which are not needed any more; and (3) skips
2 bytes and reads 2 new bytes to work with.
When move_window returns, the file position indicator is at position
window->off + window->width and the data from positions window->off to
the current file position are stored in window->buf.
This function performs only sequential access from the input file and
never seeks, so it can be safely used on pipes and sockets.
On end-of-file, move_window silently reads less than the caller
requested. On other errors, it prints a message and returns -1.
Helped-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
As the svn import infrastructure evolves, it's getting to be a pain to
tell by eye what files were added or removed from a dependency line
like
VCSSVN_OBJS = vcs-svn/string_pool.o vcs-svn/line_buffer.o \
vcs-svn/repo_tree.o vcs-svn/fast_export.o vcs-svn/svndump.o
So use a style with one entry per line instead, like the existing
BUILTIN_OBJS:
# protect against environment
VCSSVN_OBJS =
...
VCSSVN_OBJS += vcs-svn/string_pool.o
VCSSVN_OBJS += vcs-svn/line_buffer.o
...
which is readable on its own and produces nice, clear diffs.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
gcc -m32 correctly warns:
vcs-svn/fast_export.c: In function 'fast_export_commit':
vcs-svn/fast_export.c:54:2: warning: format '%llu' expects
argument of type 'long long unsigned int', but argument 2
has type 'unsigned int' [-Wformat]
Fix it.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Pass the log message by strbuf instead of as a C-style string and use
fwrite instead of printf to write it to fast-import so embedded '\0'
bytes can be preserved.
Currently "git log" doesn't show the embedded NULs but "git cat-file
commit" can.
While at it, stop including system headers from repo_tree.h. git
source files need to include git-compat-util.h (or cache.h or
builtin.h) sooner to ensure the appropriate feature test macros are
defined.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Use strbuf_swap when storing the svn:log and svn:author properties, so
pointers to rather than the contents of buffers get copied. The main
effect should be to make the code a little easier to read.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
All previous users of buffer_read_string have already been converted
to use the more intuitive buffer_read_binary, so remove the old API to
avoid some confusion.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
svn-fe errors out on revision 59151 of the ASF repository:
fatal: invalid dump: unexpected end of file
The proximate cause is a property with an embedded NUL character.
Previously such anomalies were ignored but commit c9d1c8ba
(2010-12-28) introduced a check strlen(val) == len to avoid reading
uninitialized data when a property list ends early and unfortunately
this test does not distinguish between "foo" followed by EOF and the
string "foo\0bar\0baz".
Fix it by using buffer_read_binary to read to a strbuf and checking
the actual length read. Most consumers of properties still use
C-style strings, so in practice an author or log message with embedded
NULs will be truncated, but a least this way svn-fe won't error out
(fixing the regression).
Reported-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
This is a small optimisation (4% reduction in user time) but is the
largest artifact within the parsing portion of svndump.c
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Instead of interning property names and comparing their string_pool
keys, look them up in a table by string length, which should be about
as fast.
Another small step towards removing dependence on string_pool
altogether.
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Instead of interning property names and comparing their string_pool
keys, look them up in a table by string length, which should be about
as fast.
This is a small step towards removing dependence on string_pool.
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Use strbufs and strings instead of interned strings for values of rev,
dump, and node fields that happen to be strings. After this change,
the only remaining string_pool use is for paths in the repo_tree API
and internals.
Functional change: treat an empty author, UUID, or URL as none at all.
So for example, in repos where the first revision has an empty
svn:author property, the first rev will be treated as by "nobody"
rather than by a person with empty name and email address created by
prepending an @ sign to the repository UUID.
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
obj_pool is overkill for this application: all that is needed is a
buffer that can resize from rev to rev to accomodate differently-sized
strings. In the spirit of commit deadcef4 (2010-11-06), use a strbuf
instead.
This is a small step towards removing dependence on obj_pool.h.
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Catch input errors and exit early enough to print a reasonable
diagnosis based on errno.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Currently buffer_copy_bytes does not report to its caller whether
it encountered an early end of file.
Add a return value representing the number of bytes read (but not
the number of bytes copied). This way all three unusual conditions
can be distinguished: input error with buffer_ferror, output error
with ferror(outfile), early end of input by checking the return
value.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Currently there is no way to detect when input ended if it ended
early during buffer_skip_bytes. Tell the calling program how many
bytes were actually skipped for easier debugging.
Existing callers will still ignore early EOF.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Move from uint32_t to off_t as the fundamental unit of length used by
the line_buffer library. Performance would get worse if anything but
I think it's worth it for support of deltas that need to skip large
pieces (> 4 GiB).
Exception: buffer_read_string still takes a uint32_t, since it keeps
its result in an in-core obj_pool.
Callers still have to be updated to take advantage of this.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
The line_buffer library silently flags input errors until
buffer_deinit time; unfortunately, by that point usually errno is
invalid. Expose the error flag so callers can check for and
report errors early for easy debugging.
some_error_prone_operation(...);
if (buffer_ferror(buf))
return error("input error: %s", strerror(errno));
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Restrict the repo_tree API to functions that are actually needed.
- decouple reading the mode and content of dirents from other
operations.
- remove repo_modify_path. It is only used to read the mode from
dirents.
- remove the ability to use repo_read_mode on a missing path. The
existing code only errors out in that case, anyway.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
svn-fe processes each commit in two stages: first decide on the
correct content for all paths and export the relevant blobs, then
export a commit with the result.
But we can keep less state and simplify svn-fe a great deal by
exporting the commit in one step: use 'inline' blobs for each path and
remember nothing. This way, the repo_tree structure could be
eliminated, and we would get support for incremental imports 'for
free'.
Reorganize handle_node along these lines. This is just a code
cleanup; the changes in repo_tree and handle_revision will come later.
[db: backported to apply without text delta support]
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
The repo_tree structure remembers, for each path in each revision, a
mode (regular file, executable, symlink, or directory) and content
(blob mark or directory structure). Maintaining a second copy of all
this information when it's already in the target repository is
wasteful, it does not persist between svn-fe invocations, and most
importantly, there is no convenient way to transfer it from one
machine to another. So it would be nice to get rid of it.
As a first step, let's change the repo_tree API to match fast-import's
read commands more closely. Currently to read the mode for a path,
one uses
repo_modify_path(path, new_mode, new_content);
which changes the mode and content as a side effect. There is no
function to read the content at a path; add one.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
The dereference() function to peel a tree-ish and find the underlying
tree expects arithmetic to (void *) to work on byte addresses. We
should be reading the text of objects through a char * anyway.
Noticed-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
* git://github.com/gitster/git:
vcs-svn: Allow change nodes for root of tree (/)
vcs-svn: Implement Prop-delta handling
vcs-svn: Sharpen parsing of property lines
vcs-svn: Split off function for handling of individual properties
vcs-svn: Make source easier to read on small screens
vcs-svn: More dump format sanity checks
vcs-svn: Reject path nodes without Node-action
vcs-svn: Delay read of per-path properties
vcs-svn: Combine repo_replace and repo_modify functions
vcs-svn: Replace = Delete + Add
vcs-svn: handle_node: Handle deletion case early
vcs-svn: Use mark to indicate nodes with included text
vcs-svn: Unclutter handle_node by introducing have_props var
vcs-svn: Eliminate node_ctx.mark global
vcs-svn: Eliminate node_ctx.srcRev global
vcs-svn: Check for errors from open()
vcs-svn: Allow simple v3 dumps (no deltas yet)
Conflicts:
t/t9010-svn-fe.sh
vcs-svn/svndump.c
It can sometimes be useful to write information temporarily to file,
to read back later. These functions allow a program to use the
line_buffer facilities when doing so.
It works like this:
1. find a unique filename with buffer_tmpfile_init.
2. rewind with buffer_tmpfile_rewind. This returns a stdio
handle for writing.
3. when finished writing, declare so with
buffer_tmpfile_prepare_to_read. The return value indicates
how many bytes were written.
4. read whatever portion of the file is needed.
5. if finished, remove the temporary file with buffer_deinit.
otherwise, go back to step 2,
The svn support would use this to buffer the postimage from delta
application until the length is known and fast-import can receive
the resulting blob.
Based-on-patch-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
buffer_read_char can be used in place of buffer_read_string(1) to
avoid consuming valuable static buffer space. The delta applier will
use this to read variable-length integers one byte at a time.
Underneath, it is fgetc, wrapped so the line_buffer library can
maintain its role as gatekeeper of input.
Later it might be worth checking if fgetc_unlocked is faster ---
most line_buffer functions are not thread-safe anyway.
Helpd-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
buffer_read_string works well for non line-oriented input except for
one problem: it does not tell the caller how many bytes were actually
written. This means that unless one is very careful about checking
for errors (and eof) the calling program cannot tell the difference
between the string "foo" followed by an early end of file and the
string "foo\0bar\0baz".
So introduce a variant that reports the length, too, a thinner wrapper
around strbuf_fread. Its result is written to a strbuf so the caller
does not need to keep track of the number of bytes read.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
POSIX makes the behavior of read(2) from a pipe fairly clear: a read
from an empty pipe will block until there is data available and any
other read will not block, prefering to return a partial result.
Likewise, fread(3) and fgets(3) are clearly specified to act as
though implemented by calling fgetc(3) in a simple loop. But the
buffering behavior of fgetc is less clear.
Luckily, no sane platform is going to implement fgetc by calling the
equivalent of read(2) more than once. fgetc has to be able to
return without filling its buffer to preserve errno when errors are
encountered anyway. So let's assume the simpler behavior (trust) but
add some tests to catch insane platforms that violate that when they
come (verify).
First check that fread can handle a 0-length read from an empty fifo.
Because open(O_RDONLY) blocks until the writing end is open, open the
writing end of the fifo in advance in a subshell.
Next try short inputs from a pipe that is not filled all the way.
Lastly (two tests) try very large inputs from a pipe that will not fit
in the relevant buffers. The first of these tests reads a little
more than 8192 bytes, which is BUFSIZ (the size of stdio's buffers)
on this Linux machine. The second reads a little over 64 KiB (the
pipe capacity on Linux) and is not run unless requested by setting
the GIT_REMOTE_SVN_TEST_BIG_FILES environment variable.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Do not expect an implicit newline after each input record.
Use a separate command to exercise buffer_skip_bytes.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Split the line_buffer test into small pieces and move it to its
own file as preparation for adding more tests.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Imitate the input format of test-obj-pool to support arbitrary
sequences of commands rather than alternating read/copy. This should
make it easier to add tests that exercise other line_buffer functions.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Collect the line_buffer state in a newly public line_buffer struct.
Callers can use multiple line_buffers to manage input from multiple
files at a time.
svn-fe's delta applier will use this to stream a delta from svnrdump
and the preimage it applies to from fast-import at the same time.
The tests don't take advantage of the new features, but I think that's
okay. It is easier to find lingering examples of nonreentrant code by
searching for "static" in line_buffer.c.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Prepare for the line_buffer lib to support input from multiple files,
by collecting global state in a struct that can be easily passed
around.
No API change yet.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
obj_pool is inherently global and does not use the standard growing
factor alloc_nr, which makes it feel out of place in the git codebase.
Plus it is overkill for this application: all that is needed is a
buffer that can grow between requests to accomodate larger strings.
Use a strbuf instead.
As a side effect, this improves the error handling: allocation
failures will result in a clean exit instead of segfaults. It would
be nice to add a test case (using ulimit or failmalloc) but that can
wait for another day.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
The data stored in byte_buffer[] is always either discarded or
written to stdout immediately. No need for it to persist between
function calls.
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Lazy fast-import frontend authors that want to rely on the backend to
keep track of the content of the imported trees _almost_ have what
they need in the 'cat-blob' command (v1.7.4-rc0~30^2~3, 2010-11-28).
But it is not quite enough, since
(1) cat-blob can be used to retrieve the content of files, but
not their mode, and
(2) using cat-blob requires the frontend to keep track of a name
(mark number or object id) for each blob to be retrieved
Introduce an 'ls' command to complement cat-blob and take care of the
remaining needs. The 'ls' command finds what is at a given path
within a given tree-ish (tag, commit, or tree):
'ls' SP <dataref> SP <path> LF
or in fast-import's active commit:
'ls' SP <path> LF
The response is a single line sent through the cat-blob channel,
imitating ls-tree output. So for example:
FE> ls :1 Documentation
gfi> 040000 tree 9e6c2b599341d28a2a375f8207507e0a2a627fe9 Documentation
FE> ls 9e6c2b599341d28a2a375f8207507e0a2a627fe9 git-fast-import.txt
gfi> 100644 blob 4f92954396e3f0f97e75b6838a5635b583708870 git-fast-import.txt
FE> ls :1 RelNotes
gfi> 120000 blob b942e49944 RelNotes
FE> cat-blob b942e49944
gfi> b942e49944 blob 32
gfi> Documentation/RelNotes/1.7.4.txt
The most interesting parts of the reply are the first word, which is
a 6-digit octal mode (regular file, executable, symlink, directory,
or submodule), and the part from the second space to the tab, which is
a <dataref> that can be used in later cat-blob, ls, and filemodify (M)
commands to refer to the content (blob, tree, or commit) at that path.
If there is nothing there, the response is "missing some/path".
The intent is for this command to be used to read files from the
active commit, so a frontend can apply patches to them, and to copy
files and directories from previous revisions.
For example, proposed updates to svn-fe use this command in place of
its internal representation of the repository directory structure.
This simplifies the frontend a great deal and means support for
resuming an import in a separate fast-import run (i.e., incremental
import) is basically free.
Signed-off-by: David Barr <david.barr@cordelta.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Improved-by: Junio C Hamano <gitster@pobox.com>
Improved-by: Sverre Rabbelier <srabbelier@gmail.com>
