On 08/04/2019 02:23, Duy Nguyen wrote:
On Mon, Apr 8, 2019 at 5:52 AM Christian Couder
<christian.couder@xxxxxxxxx> wrote:
Git has a very optimized mechanism to compactly store
objects (blobs, trees, commits, etc.) in packfiles[2]. These files are
created by[3]:
1. listing objects;
2. sorting the list with some good heuristics;
3. traversing the list with a sliding window to find similar objects in
the window, in order to do delta decomposing;
4. compress the objects with zlib and write them to the packfile.
What we are calling pack access code in this document, is the set of
functions responsible for retrieving the objects stored at the
packfiles. This process consists, roughly speaking, in three parts:
1. Locate and read the blob from packfile, using the index file;
2. If the blob is a delta, locate and read the base object to apply the
delta on top of it;
3. Once the full content is read, decompress it (using zlib inflate).
Note: There is a delta cache for the second step so that if another
delta depends on the same base object, it is already in memory. This
cache is global; also, the sliding windows, are global per packfile.
Yeah, but the sliding windows are used only when creating pack files,
not when reading them, right?
These windows are actually for reading. We used to just mmap the whole
pack file in the early days but that was impossible for 4+ GB packs on
32-bit platforms, which was one of the reasons, I think, that sliding
windows were added, to map just the parts we want to read.
Another "32-bit problem" should also be expressly considered during the
GSoC work because of the MS Windows definition of uInt / long to be only
32 bits, leading to much of the Git code failing on the Git for Windows
port and on the Git LFS (for Windows) for packs and files greater than
4Gb. https://github.com/git-for-windows/git/issues/1063
Mainly it is just substitution of size_t for long, but there can be
unexpected coercions when mixed data types get coerced down to a local
32-bit long. This is made worse by it being implementation defined, so
one needs to be explicit about some casts up to pointer/memsized types.
# Points to work on
* Investigate pack access call chains and look for non-thread-safe
operations on then.
* Protect packfile.c read-and-write global variables, such as
pack_open_windows, pack_open_fds and etc., using mutexes.
Do you want to work on making both packfile reading and packfile
writing thread safe? Or just packfile reading?
Packfile writing is probably already or pretty close to thread-safe
(at least the main writing code path in git-pack-objects; the
streaming blobs to a pack, i'm not so sure).
--
Philip
