Message ID | 20230915083415.263187-1-knayak@gitlab.com (mailing list archive) |
---|---|
State | Superseded |
Headers | show |
Series | [v3] revision: add `--ignore-missing-links` user option | expand |
Karthik Nayak <karthik.188@gmail.com> writes: > From: Karthik Nayak <karthik.188@gmail.com> > > The revision backend is used by multiple porcelain commands such as > git-rev-list(1) and git-log(1). The backend currently supports ignoring > missing links by setting the `ignore_missing_links` bit. This allows the > revision walk to skip any objects links which are missing. Expose this > bit via an `--ignore-missing-links` user option. Given the above "we merely surface a feature that already exists and supported to be used by the end users from the command line" claim ... > diff --git a/builtin/rev-list.c b/builtin/rev-list.c > index ff715d6918..5239d83c76 100644 > --- a/builtin/rev-list.c > +++ b/builtin/rev-list.c > @@ -266,7 +266,8 @@ static int finish_object(struct object *obj, const char *name UNUSED, > { > struct rev_list_info *info = cb_data; > if (oid_object_info_extended(the_repository, &obj->oid, NULL, 0) < 0) { > - finish_object__ma(obj); > + if (!info->revs->ignore_missing_links) > + finish_object__ma(obj); > return 1; > } ... this hunk is a bit unexpected. As a low-level plumbing command, shouldn't it be left to the user who gives --ignore-missing-links from their command line to specify how the missing "obj" here should be dealt with by giving the "--missing=<foo>" option? While giving "allow-promisor" may not make much sense, "--missing=allow-any" may of course make sense (it is the same as hardcoding the decision not to call finish_object__ma() at all), and so may "--missing=print". Stepping back a bit, with "--missing=print", is this change still needed? The missing objects discovered will be shown at the end, with the setting, no? Thanks.
On Fri, Sep 15, 2023 at 8:54 PM Junio C Hamano <gitster@pobox.com> wrote: > > Karthik Nayak <karthik.188@gmail.com> writes: > > > From: Karthik Nayak <karthik.188@gmail.com> > > > > The revision backend is used by multiple porcelain commands such as > > git-rev-list(1) and git-log(1). The backend currently supports ignoring > > missing links by setting the `ignore_missing_links` bit. This allows the > > revision walk to skip any objects links which are missing. Expose this > > bit via an `--ignore-missing-links` user option. > > Given the above "we merely surface a feature that already exists and > supported to be used by the end users from the command line" claim ... > > > diff --git a/builtin/rev-list.c b/builtin/rev-list.c > > index ff715d6918..5239d83c76 100644 > > --- a/builtin/rev-list.c > > +++ b/builtin/rev-list.c > > @@ -266,7 +266,8 @@ static int finish_object(struct object *obj, const char *name UNUSED, > > { > > struct rev_list_info *info = cb_data; > > if (oid_object_info_extended(the_repository, &obj->oid, NULL, 0) < 0) { > > - finish_object__ma(obj); > > + if (!info->revs->ignore_missing_links) > > + finish_object__ma(obj); > > return 1; > > } > > ... this hunk is a bit unexpected. As a low-level plumbing command, > shouldn't it be left to the user who gives --ignore-missing-links > from their command line to specify how the missing "obj" here should > be dealt with by giving the "--missing=<foo>" option? While giving > "allow-promisor" may not make much sense, "--missing=allow-any" may > of course make sense (it is the same as hardcoding the decision not > to call finish_object__ma() at all), and so may "--missing=print". > This is to be expected, in my opinion. In terms of revision.c and setting the `revs->ignore_missing_links` bit, the traversal will go throw all objects (commits and otherwise) and call `show_commit` or `show_object` on them. Here there is a difference for commits and non-commit objects. 1. Commit objects: commits are parsed in revision.c and after that the `show_commit` function is called only when the object is available. 2. Non-commit objects: while trees are parsed in revision.c, blobs are never parsed and hence, ` show_object` can be called on missing blobs. This is left to the user to handle. In our case, we error out in `rev-list.c`, which is not what we want when using the `--ignore-missing-links` option. Hence, this addition. There is an argument to be made around compatibility between the `--missing` option and `--ignore-missing-links` option, but since the former only works with non-commit objects I think the latter should be independent, and also the latter is about ignoring all missing links. I also don't think the user should again specify what to do with missing links by adding `--missing=allow-any` as `--ignore-missing-links` is a superset of it. > Stepping back a bit, with "--missing=print", is this change still > needed? The missing objects discovered will be shown at the end, > with the setting, no? > The main difference is that the `--missing` options works entirely with non-commit objects (I'm assuming this was built with promisor notes in mind). So if a commit is missing, git-rev-list(1) will still barf an error, but this error handling is not in `builtin/rev-list.c` rather is in a layer above in `revision.c`. So it wouldn't be trivial for the `--missing` option to support missing commit links. So that's why we expose `--ignore-missing-links` which ensures any kind of object (commits included) if missing, is ignored. Thanks for the review!
Karthik Nayak <karthik.188@gmail.com> writes: >> Given the above "we merely surface a feature that already exists and >> supported to be used by the end users from the command line" claim ... >> >> > diff --git a/builtin/rev-list.c b/builtin/rev-list.c >> > index ff715d6918..5239d83c76 100644 >> > --- a/builtin/rev-list.c >> > +++ b/builtin/rev-list.c >> > @@ -266,7 +266,8 @@ static int finish_object(struct object *obj, const char *name UNUSED, >> > { >> > struct rev_list_info *info = cb_data; >> > if (oid_object_info_extended(the_repository, &obj->oid, NULL, 0) < 0) { >> > - finish_object__ma(obj); >> > + if (!info->revs->ignore_missing_links) >> > + finish_object__ma(obj); >> > return 1; >> > } >> >> ... this hunk is a bit unexpected. As a low-level plumbing command, >> shouldn't it be left to the user who gives --ignore-missing-links >> from their command line to specify how the missing "obj" here should >> be dealt with by giving the "--missing=<foo>" option? While giving >> "allow-promisor" may not make much sense, "--missing=allow-any" may >> of course make sense (it is the same as hardcoding the decision not >> to call finish_object__ma() at all), and so may "--missing=print". >> > > This is to be expected, in my opinion. In terms of revision.c and > setting the `revs->ignore_missing_links` bit, the traversal will > go throw all objects (commits and otherwise) and call > `show_commit` or `show_object` on them. Yes. And the user can choose how to handle such an object here by telling finish_object__ma() with the --missing=<how> option, so letting them do so, instead of robbing the choice from them, would be a more flexible design here, right? > if a commit is > missing, git-rev-list(1) will still barf an error, but this error OK, yeah, I do see the need for setting the ignore-missing-links bit for what you are doing, and --missing and --ignore-missing-links are orthogonal options. Getting rid of the hardcoded skipping of finish_object__ma() would make sense from this angle, too. Thanks.
On Mon, Sep 18, 2023 at 5:56 PM Junio C Hamano <gitster@pobox.com> wrote: > Karthik Nayak <karthik.188@gmail.com> writes: > > This is to be expected, in my opinion. In terms of revision.c and > > setting the `revs->ignore_missing_links` bit, the traversal will > > go throw all objects (commits and otherwise) and call > > `show_commit` or `show_object` on them. > > Yes. And the user can choose how to handle such an object here by > telling finish_object__ma() with the --missing=<how> option, so > letting them do so, instead of robbing the choice from them, would > be a more flexible design here, right? > > > if a commit is > > missing, git-rev-list(1) will still barf an error, but this error > > OK, yeah, I do see the need for setting the ignore-missing-links bit > for what you are doing, and --missing and --ignore-missing-links are > orthogonal options. Getting rid of the hardcoded skipping of > finish_object__ma() would make sense from this angle, too. Well. The only problem is that setting `ignore_missing_links` bit never calls `show_commit` for missing commits (since commits are pre-parsed in revision.c). So to keep that behavior consistent for non-commit objects, I hardcoded the skipping of `finish_object__ma()` in `show_object`. If I remove the hardcoding, it would mean that `--ignore-missing-links` would skip missing commits but for non-commits objects, the user would have to pass `--missing=allow-any` else rev-list would still error out with a missing object error. Don't you think this would be confusing for the user? I'm happy to send a revised version removing this hardcoding if you still think otherwise :)
Karthik Nayak <karthik.188@gmail.com> writes: > If I remove the hardcoding, it would mean that > `--ignore-missing-links` would skip missing commits but for > non-commits objects, the user would have to pass > `--missing=allow-any` else rev-list would still error out with a > missing object error. > > Don't you think this would be confusing for the user? I'm happy > to send a revised version removing this hardcoding if you still > think otherwise :) Yes. This is an example of flexibility and ergonomics at odds, and for a low-level plumbing like rev-list, I would prefer not to limit the flexibility unnecessarily. I do not care about the ability to pass allow-any here. But when you traverse a range A..B with the --ignore-missing-links option, the reporting mechanism based on the --boundary cannot tell which ones are at the usual "traversal boundaries" and which ones are ones beyond the broken links, can it? If you allowed the users to pass 'print', then those reported with '?' prefix would be the missing ones. The ones that are reported with '-' prefix may still be mixture of the two kinds, but you can now subtract one set from the other set to see which ones are true boundaries and which ones are missing. The hardcoded "we do not let __ma() logic to kick in" makes it impossible, which is what I find disturbing. Thanks.
diff --git a/Documentation/rev-list-options.txt b/Documentation/rev-list-options.txt index a4a0cb93b2..8ee713db3d 100644 --- a/Documentation/rev-list-options.txt +++ b/Documentation/rev-list-options.txt @@ -227,6 +227,15 @@ explicitly. Upon seeing an invalid object name in the input, pretend as if the bad input was not given. +--ignore-missing-links:: + During traversal, if an object that is referenced does not + exist, instead of dying of a repository corruption, pretend as + if the reference itself does not exist. Running the command + with the `--boundary` option makes these missing commits, + together with the commits on the edge of revision ranges + (i.e. true boundary objects), appear on the output, prefixed + with '-'. + ifndef::git-rev-list[] --bisect:: Pretend as if the bad bisection ref `refs/bisect/bad` diff --git a/builtin/rev-list.c b/builtin/rev-list.c index ff715d6918..5239d83c76 100644 --- a/builtin/rev-list.c +++ b/builtin/rev-list.c @@ -266,7 +266,8 @@ static int finish_object(struct object *obj, const char *name UNUSED, { struct rev_list_info *info = cb_data; if (oid_object_info_extended(the_repository, &obj->oid, NULL, 0) < 0) { - finish_object__ma(obj); + if (!info->revs->ignore_missing_links) + finish_object__ma(obj); return 1; } if (info->revs->verify_objects && !obj->parsed && obj->type != OBJ_COMMIT) diff --git a/revision.c b/revision.c index 2f4c53ea20..cbfcbf6e28 100644 --- a/revision.c +++ b/revision.c @@ -2595,6 +2595,8 @@ static int handle_revision_opt(struct rev_info *revs, int argc, const char **arg revs->limited = 1; } else if (!strcmp(arg, "--ignore-missing")) { revs->ignore_missing = 1; + } else if (!strcmp(arg, "--ignore-missing-links")) { + revs->ignore_missing_links = 1; } else if (opt && opt->allow_exclude_promisor_objects && !strcmp(arg, "--exclude-promisor-objects")) { if (fetch_if_missing) diff --git a/t/t6022-rev-list-alternates.sh b/t/t6022-rev-list-alternates.sh new file mode 100755 index 0000000000..567dd21876 --- /dev/null +++ b/t/t6022-rev-list-alternates.sh @@ -0,0 +1,93 @@ +#!/bin/sh + +test_description='handling of alternates in rev-list' + +TEST_PASSES_SANITIZE_LEAK=true +. ./test-lib.sh + +# We create 5 commits and move them to the alt directory and +# create 5 more commits which will stay in the main odb. +test_expect_success 'create repository and alternate directory' ' + test_commit_bulk 5 && + git clone --reference=. --shared . alt && + test_commit_bulk --start=6 -C alt 5 +' + +# when the alternate odb is provided, all commits are listed along with the boundary +# commit. +test_expect_success 'rev-list passes with alternate object directory' ' + git -C alt rev-list --all --objects --no-object-names >actual.raw && + { + git rev-list --all --objects --no-object-names && + git -C alt rev-list --all --objects --no-object-names --not \ + --alternate-refs + } >expect.raw && + sort actual.raw >actual && + sort expect.raw >expect && + test_cmp expect actual +' + +alt=alt/.git/objects/info/alternates + +hide_alternates () { + test -f "$alt.bak" || mv "$alt" "$alt.bak" +} + +show_alternates () { + test -f "$alt" || mv "$alt.bak" "$alt" +} + +# When the alternate odb is not provided, rev-list fails since the 5th commit's +# parent is not present in the main odb. +test_expect_success 'rev-list fails without alternate object directory' ' + hide_alternates && + test_must_fail git -C alt rev-list HEAD +' + +# With `--ignore-missing-links`, we stop the traversal when we encounter a +# missing link. The boundary commit is not listed as we haven't used the +# `--boundary` options. +test_expect_success 'rev-list only prints main odb commits with --ignore-missing-links' ' + hide_alternates && + + git -C alt rev-list --objects --no-object-names \ + --ignore-missing-links HEAD >actual.raw && + git -C alt cat-file --batch-check="%(objectname)" \ + --batch-all-objects >expect.raw && + + sort actual.raw >actual && + sort expect.raw >expect && + test_must_fail git -C alt rev-list HEAD +' + +# With `--ignore-missing-links` and `--boundary`, we can even print those boundary +# commits. +test_expect_success 'rev-list prints boundary commit with --ignore-missing-links' ' + git -C alt rev-list --ignore-missing-links --boundary HEAD >got && + grep "^-$(git rev-parse HEAD)" got +' + +test_expect_success "setup for rev-list --ignore-missing-links with missing objects" ' + show_alternates && + test_commit -C alt 11 +' + +for obj in "HEAD^{tree}" "HEAD:11.t" +do + # The `--ignore-missing-links` option should ensure that git-rev-list(1) + # doesn't fail when used alongside `--objects` when a tree/blob is + # missing. + test_expect_success "rev-list --ignore-missing-links with missing $type" ' + oid="$(git -C alt rev-parse $obj)" && + path="alt/.git/objects/$(test_oid_to_path $oid)" && + + mv "$path" "$path.hidden" && + test_when_finished "mv $path.hidden $path" && + + git -C alt rev-list --ignore-missing-links --objects HEAD \ + >actual && + ! grep $oid actual + ' +done + +test_done