| Message ID | pull.1068.v2.git.1635454237.gitgitgadget@gmail.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Allow clean/smudge filters to handle huge files in the LLP64 data model | expand |
On 2021-10-28 at 20:50:30, Johannes Schindelin via GitGitGadget wrote: > This patch series came in via the Git for Windows fork > [https://github.com/git-for-windows/git/pull/3487], and I intend to merge it > before v2.34.0-rc0, therefore I appreciate every careful review you gentle > people can spare. > > The x86_64 variant of Windows uses the LLP64 data model, where the long data > type is 32-bit. This is very different from the LP64 data model used e.g. by > x86_64 Linux, where unsigned long is 64-bit. > > Most notably, this means that sizeof(unsigned long) != sizeof(size_t) in > general. > > However, since Git was born in the Linux ecosystem, where that inequality > does not hold true, it is understandable that unsigned long is used in many > code locations where size_t should have been used. As a consequence, quite a > few things are broken e.g. on Windows, when it comes to 4GB file contents or > larger. > > Using Git LFS [https://git-lfs.github.io/] trying to work around such issues > is one such a broken scenario. You cannot git checkout, say, 5GB files. Huge > files will be truncated to whatever the file size is modulo 4GB (in the case > of a 5GB file, it would be truncated to 1GB). > > This patch series primarily fixes the Git LFS scenario, by allowing clean > filters to accept 5GB files, and by allowing smudge filters to produce 5GB > files. > > The much larger project to teach Git to use size_t instead of unsigned long > in all the appropriate places is hardly scratched by this patch series. > > Side note: The fix for the clean filter included in this series does not > actually affect Git LFS! The reason is that Git LFS marks its filter as > required, and therefore Git streams the file contents to Git LFS via a file > descriptor (which is unaffected by LLP64). A "clean" filter that is not > marked as required, however, lets Git take the code path that is fixed by > this patch series. This series seems sane to me. I'm delighted we can fix this issue with so little code, since it's a been a very inconvenient problem for a lot of Windows users. I might suggest when we make the giant transition project in the future that we use size_t for things that are going to be in memory and off_t or, if necessary, intmax_t for general object sizes so it's clear which one we want. However, that has no effect on this series since this intentionally has a limited scope.
"brian m. carlson" <sandals@crustytoothpaste.net> writes: > This series seems sane to me. I'm delighted we can fix this issue with > so little code, since it's a been a very inconvenient problem for a lot > of Windows users. Yup. I do find the goal this series aims is quite worthy, and the size of the series is "reviewable in a single sitting" bite-size, which was pleasant. Thanks.
This patch series came in via the Git for Windows fork [https://github.com/git-for-windows/git/pull/3487], and I intend to merge it before v2.34.0-rc0, therefore I appreciate every careful review you gentle people can spare. The x86_64 variant of Windows uses the LLP64 data model, where the long data type is 32-bit. This is very different from the LP64 data model used e.g. by x86_64 Linux, where unsigned long is 64-bit. Most notably, this means that sizeof(unsigned long) != sizeof(size_t) in general. However, since Git was born in the Linux ecosystem, where that inequality does not hold true, it is understandable that unsigned long is used in many code locations where size_t should have been used. As a consequence, quite a few things are broken e.g. on Windows, when it comes to 4GB file contents or larger. Using Git LFS [https://git-lfs.github.io/] trying to work around such issues is one such a broken scenario. You cannot git checkout, say, 5GB files. Huge files will be truncated to whatever the file size is modulo 4GB (in the case of a 5GB file, it would be truncated to 1GB). This patch series primarily fixes the Git LFS scenario, by allowing clean filters to accept 5GB files, and by allowing smudge filters to produce 5GB files. The much larger project to teach Git to use size_t instead of unsigned long in all the appropriate places is hardly scratched by this patch series. Side note: The fix for the clean filter included in this series does not actually affect Git LFS! The reason is that Git LFS marks its filter as required, and therefore Git streams the file contents to Git LFS via a file descriptor (which is unaffected by LLP64). A "clean" filter that is not marked as required, however, lets Git take the code path that is fixed by this patch series. Changes since v1: * Removed extraneous "Signed-off-by:" lines from "git-compat-util: introduce more size_t helpers". * Integrated Carlo's patch to allow genzeros to generate large amounts of NULs, even in LLP64 data models. * Using test-tool genzeros instead of dd if=/dev/zero, to help HP NonStop (which appears to use the LP64 data model and therefore should pass the new test cases even without the fixes provided in this patch series). * Accelerating genzeros to have performance characteristics similar to dd if=/dev/zero instead of being ~50x slower. Carlo Marcelo Arenas Belón (1): test-genzeros: allow more than 2G zeros in Windows Johannes Schindelin (2): test-tool genzeros: generate large amounts of data more efficiently git-compat-util: introduce more size_t helpers Matt Cooper (4): t1051: introduce a smudge filter test for extremely large files odb: teach read_blob_entry to use size_t odb: guard against data loss checking out a huge file clean/smudge: allow clean filters to process extremely large files convert.c | 2 +- delta.h | 6 +++--- entry.c | 8 +++++--- entry.h | 2 +- git-compat-util.h | 25 +++++++++++++++++++++++++ object-file.c | 6 +++--- packfile.c | 6 +++--- parallel-checkout.c | 2 +- t/helper/test-genzeros.c | 21 +++++++++++++++++---- t/t1051-large-conversion.sh | 23 +++++++++++++++++++++++ 10 files changed, 82 insertions(+), 19 deletions(-) base-commit: ebf3c04b262aa27fbb97f8a0156c2347fecafafb Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1068%2Fdscho%2Fhuge-file-smudge-clean-v2 Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1068/dscho/huge-file-smudge-clean-v2 Pull-Request: https://github.com/gitgitgadget/git/pull/1068 Range-diff vs v1: -: ----------- > 1: 068f897b973 test-genzeros: allow more than 2G zeros in Windows -: ----------- > 2: 6edcbae372e test-tool genzeros: generate large amounts of data more efficiently 1: 449eb5c205e ! 3: 1bdded86f5d t1051: introduce a smudge filter test for extremely large files @@ t/t1051-large-conversion.sh: test_expect_success 'ident converts on output' ' +# ensures that smudging doesn't mangle large files on 64-bit Windows. +test_expect_failure EXPENSIVE,!LONG_IS_64BIT 'files over 4GB convert on output' ' + test_commit test small "a small file" && -+ test_config filter.makelarge.smudge "dd if=/dev/zero bs=$((1024*1024)) count=$((5*1024)) && cat" && ++ test_config filter.makelarge.smudge \ ++ "test-tool genzeros $((5*1024*1024*1024)) && cat" && + echo "small filter=makelarge" >.gitattributes && + rm small && + git checkout -- small && 2: 5b9d149ba23 ! 4: 3ffd3a001f7 odb: teach read_blob_entry to use size_t @@ t/t1051-large-conversion.sh: test_expect_success 'ident converts on output' ' -test_expect_failure EXPENSIVE,!LONG_IS_64BIT 'files over 4GB convert on output' ' +test_expect_success EXPENSIVE,!LONG_IS_64BIT 'files over 4GB convert on output' ' test_commit test small "a small file" && - test_config filter.makelarge.smudge "dd if=/dev/zero bs=$((1024*1024)) count=$((5*1024)) && cat" && - echo "small filter=makelarge" >.gitattributes && + test_config filter.makelarge.smudge \ + "test-tool genzeros $((5*1024*1024*1024)) && cat" && 3: c81ee778b26 ! 5: 32472ae3f98 git-compat-util: introduce more size_t helpers @@ Commit message We will use them in the next commit. Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> - Signed-off-by: Matt Cooper <vtbassmatt@gmail.com> - Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> ## git-compat-util.h ## @@ 4: bca013a0511 = 6: c6910584108 odb: guard against data loss checking out a huge file 5: 20387ce3557 ! 7: d87d4229bb4 clean/smudge: allow clean filters to process extremely large files @@ t/t1051-large-conversion.sh: test_expect_success EXPENSIVE,!LONG_IS_64BIT 'files +# This clean filter writes down the size of input it receives. By checking against +# the actual size, we ensure that cleaning doesn't mangle large files on 64-bit Windows. +test_expect_success EXPENSIVE,!LONG_IS_64BIT 'files over 4GB convert on input' ' -+ dd if=/dev/zero bs=$((1024*1024)) count=$((5*1024)) >big && ++ test-tool genzeros $((5*1024*1024*1024)) >big && + test_config filter.checklarge.clean "wc -c >big.size" && + echo "big filter=checklarge" >.gitattributes && + git add big &&