Message ID | cover-v2-0.7-00000000000-20210912T001420Z-avarab@gmail.com (mailing list archive) |
---|---|
Headers | show |
Series | strvec: use size_t to store nr and alloc | expand |
On Sun, Sep 12, 2021 at 02:15:48AM +0200, Ævar Arnfjörð Bjarmason wrote: > This is what I'd been sitting on locally since that recent thread, I > polished it up a bit since Jeff King posted his version. > > The potential overflow bug I mentioned is in rebase.c. See > 5/7. "Potential" because it's not a bug now, but that code > intentionally considers a strvec, and then iterates it from nr-1 to 0, > and if it reaches 0 intentionally counts down one more to -1 to > indicate that it's visited all elements. > > We then check that with i >= 0, except of course if it becomes > unsigned that doesn't become -1, but rather it wraps around. You can also just use ssize_t, or you can compare against SIZE_MAX to catch the wraparound (there's some prior art in sort_revindex()). That said, I don't mind rewriting loops to count up rather than down. It usually makes them easier to follow (and in your patch 5, I do not see any reason we would need to count down rather than up; we do not even care where we find "-q", only that we found it. > The rest of this is all changes to have that s/int/size_t/ radiate > outwards, i.e. when we assign that value to a variable somewhere its > now a "size_t" instead of an "int" etc. I'm a little "meh" on some of these, for a few reasons: - anything calling into setup_revisions() eventually is just kicking the can anyway. And these are generally not buggy in the first place, since they're bounded argv creations. - passing a strvec instead of the broken-down pair is a less flexible interface. It's one thing if the callee benefits from seeing the strvec (say, because they may push more items onto it). But I think with strbufs, we have a general guideline that if a function _can_ take the bare pointer, then it should. (Sorry, I don't have a succinct reference to CodingGuidelines or anything like that; I feel like this is wisdom we came up with on the list in the early days of strbufs). - if we are going to pass a strvec, it should almost certainly be const, to make it clear how we intend to use it. So if we we wanted to try to reduce the int/size_t conversions here (and I don't mind doing it, but am not altogether sure it is a good use of time, because the rabbit hole runs deep), I think we ought to be switching to size_t everywhere-ish along whole call chains. Or possibly providing a checked size_to_int() which will safely catch and abort. These cases are largely stupid things that real people would never come across. The real goal is making sure we don't get hit with a memory safety bug (under-allocation, converting a big size_t to a negative int, etc). -Peff
Jeff King <peff@peff.net> writes: > I'm a little "meh" on some of these, for a few reasons: > > - anything calling into setup_revisions() eventually is just kicking > the can anyway. And these are generally not buggy in the first place, > since they're bounded argv creations. > > - passing a strvec instead of the broken-down pair is a less flexible > interface. It's one thing if the callee benefits from seeing the > strvec (say, because they may push more items onto it). But I think > with strbufs, we have a general guideline that if a function _can_ > take the bare pointer, then it should. (Sorry, I don't have a > succinct reference to CodingGuidelines or anything like that; I feel > like this is wisdom we came up with on the list in the early days of > strbufs). > > - if we are going to pass a strvec, it should almost certainly be > const, to make it clear how we intend to use it. All true. > These cases are largely stupid things that real people would never come > across. The real goal is making sure we don't get hit with a memory > safety bug (under-allocation, converting a big size_t to a negative int, > etc). Yes. Ævar, I do not mean any disrespect to you, but I have to say that topics like this one are starting to wear my concentration and patience down really thin and making me really slow down. Perhaps I am biased by not yet having seen what you eventually want to build on top, and because of that I do not understood why and how these "clean ups" are so valuable to have right now (as opposed to just letting the sleeping dog lie), which you would of course have a much better chance to know than I do. But with that "bias", many of the recent patches from you look more like pointless churn, mixed with fixes to theoretical problems, than clean-ups with real benefits. What makes it worse is that there are occasional real gems among these "meh" patches, which means I have to read all of them anyway to sift wheat from chaff X-<.
On 12/09/2021 01:15, Ævar Arnfjörð Bjarmason wrote: > This is a proposed v2 of Jeff King's one-patch change to change > strvec's nr/alloc from "int" to "size_t". As noted below I think it's > worthwhile to not only change that in the struct, but also in code > that directly references the "nr" member. > > On Sat, Sep 11 2021, Philip Oakley wrote: > >> On 11/09/2021 17:13, Ævar Arnfjörð Bjarmason wrote: >>> On Sat, Sep 11 2021, Jeff King wrote: >>> >>>> We converted argv_array (which later became strvec) to use size_t in >>>> 819f0e76b1 (argv-array: use size_t for count and alloc, 2020-07-28) in >>>> order to avoid the possibility of integer overflow. But later, commit >>>> d70a9eb611 (strvec: rename struct fields, 2020-07-28) accidentally >>>> converted these back to ints! >>>> >>>> Those two commits were part of the same patch series. I'm pretty sure >>>> what happened is that they were originally written in the opposite order >>>> and then cleaned up and re-ordered during an interactive rebase. And >>>> when resolving the inevitable conflict, I mistakenly took the "rename" >>>> patch completely, accidentally dropping the type change. >>>> >>>> We can correct it now; better late than never. >>>> >>>> Signed-off-by: Jeff King <peff@peff.net> >>>> --- >>>> This was posted previously in the midst of another thread, but I don't >>>> think was picked up. There was some positive reaction, but one "do we >>>> really need this?" to which I responded in detail: >>>> >>>> https://lore.kernel.org/git/YTIBnT8Ue1HZXs82@coredump.intra.peff.net/ >>>> >>>> I don't really think any of that needs to go into the commit message, >>>> but if that's a hold-up, I can try to summarize it (though I think >>>> referring to the commit which _already_ did this and was accidentally >>>> reverted would be sufficient). >>> Thanks, I have a WIP version of this outstanding starting with this >>> patch that I was planning to submit sometime, but I'm happy to have you >>> pursue it, especially with the ~100 outstanding patches I have in >>> master..seen. >>> >>> It does feel somewhere between iffy and a landmine waiting to be stepped >>> on to only convert the member itself, and not any of the corresponding >>> "int" variables that track it to "size_t". >>> >>> If you do the change I suggested in >>> https://lore.kernel.org/git/87v93i8svd.fsf@evledraar.gmail.com/ you'll >>> find that there's at least one first-order reference to this that now >>> uses "int" that if converted to "size_t" will result in a wrap-around >>> error, we're lucky that one has a test failure. >>> >>> I can tell you what that bug is, but maybe it's better if you find it >>> yourself :) I.e. I found *that* one, but I'm not sure I found them >>> all. I just s/int nr/size_t *nr/ and eyeballed the wall off compiler >>> errors & the code context (note: pointer, obviously broken, but makes >>> the compiler yell). >>> >>> That particular bug will be caught by the compiler as it involves a >= 0 >>> comparison against unsigned, but we may not not have that everywhere... >> I'm particularly interested in the int -> size_t change problem as part >> of the wider 4GB limitations for the LLP64 systems [0] such as the >> RaspPi, git-lfs (on windows [1]), and Git-for-Windows[2]. It is a big >> problem. > Okey, fine, no fun excercise for the reader then ;) There's a lot of weeds in there ;-) In some ways it feels like the SHA1->SHA256 transition, but without the consensus, as the 'shifting foundations' problem only affect a subgroup of a subgroup (large Windows files and repositories). > > This is what I'd been sitting on locally since that recent thread, I > polished it up a bit since Jeff King posted his version. > > The potential overflow bug I mentioned is in rebase.c. See > 5/7. "Potential" because it's not a bug now, but that code > intentionally considers a strvec, and then iterates it from nr-1 to 0, > and if it reaches 0 intentionally counts down one more to -1 to > indicate that it's visited all elements. It's these tidbits about how the problems surface, their detection and resolution that are really useful. Along with general awareness raising. At least here the issue is reasonably tightly focussed, and even then, testing is hard. > > We then check that with i >= 0, except of course if it becomes > unsigned that doesn't become -1, but rather it wraps around. > > The rest of this is all changes to have that s/int/size_t/ radiate > outwards, i.e. when we assign that value to a variable somewhere its > now a "size_t" instead of an "int" etc. In the LLP64 case, I'm somewhat concerned about the possible pushback of a wide spread s/int/size_t/ on the codebase's look & feel. (aside) I don't think there is even a `1S` to match the the `1L` and `1U` shorthands used in various places. None of that is part of the series, but the patches are beneficial to the codes portability. > >> [0] >> http://nickdesaulniers.github.io/blog/2016/05/30/data-models-and-word-size/ >> [1] https://github.com/git-lfs/git-lfs/issues/2434 Git on Windows >> client corrupts files > 4Gb >> [2] https://github.com/git-for-windows/git/pull/2179 [DRAFT] for >> testing : Fix 4Gb limit for large files on Git for Windows > Jeff King (1): > strvec: use size_t to store nr and alloc > > Ævar Arnfjörð Bjarmason (6): > remote-curl: pass "struct strvec *" instead of int/char ** pair > pack-objects: pass "struct strvec *" instead of int/char ** pair > sequencer.[ch]: pass "struct strvec *" instead of int/char ** pair > upload-pack.c: pass "struct strvec *" instead of int/char ** pair > rebase: don't have loop over "struct strvec" depend on signed "nr" > strvec API users: change some "int" tracking "nr" to "size_t" > > builtin/pack-objects.c | 6 +++--- > builtin/rebase.c | 26 ++++++++++++-------------- > connect.c | 8 ++++---- > fetch-pack.c | 4 ++-- > ls-refs.c | 2 +- > remote-curl.c | 23 +++++++++++------------ > sequencer.c | 8 ++++---- > sequencer.h | 4 ++-- > serve.c | 2 +- > shallow.c | 5 +++-- > shallow.h | 6 ++++-- > strvec.h | 4 ++-- > submodule.c | 2 +- > upload-pack.c | 7 +++---- > 14 files changed, 53 insertions(+), 54 deletions(-) > > Range-diff against v1: > -: ----------- > 1: 2ef48d734e8 remote-curl: pass "struct strvec *" instead of int/char ** pair > -: ----------- > 2: 7f59a58ed97 pack-objects: pass "struct strvec *" instead of int/char ** pair > -: ----------- > 3: c35cfb9c9c5 sequencer.[ch]: pass "struct strvec *" instead of int/char ** pair > -: ----------- > 4: 2e0b82d4316 upload-pack.c: pass "struct strvec *" instead of int/char ** pair > -: ----------- > 5: be85a0565ef rebase: don't have loop over "struct strvec" depend on signed "nr" > 1: 498f5ed80dc ! 6: ba17290852c strvec: use size_t to store nr and alloc > @@ Commit message > We can correct it now; better late than never. > > Signed-off-by: Jeff King <peff@peff.net> > + Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> > > ## strvec.h ## > @@ strvec.h: extern const char *empty_strvec[]; > -: ----------- > 7: 2edd9708888 strvec API users: change some "int" tracking "nr" to "size_t"
On Sun, Sep 12 2021, Junio C Hamano wrote: > Jeff King <peff@peff.net> writes: > >> I'm a little "meh" on some of these, for a few reasons: >> >> - anything calling into setup_revisions() eventually is just kicking >> the can anyway. And these are generally not buggy in the first place, >> since they're bounded argv creations. >> >> - passing a strvec instead of the broken-down pair is a less flexible >> interface. It's one thing if the callee benefits from seeing the >> strvec (say, because they may push more items onto it). But I think >> with strbufs, we have a general guideline that if a function _can_ >> take the bare pointer, then it should. (Sorry, I don't have a >> succinct reference to CodingGuidelines or anything like that; I feel >> like this is wisdom we came up with on the list in the early days of >> strbufs). >> >> - if we are going to pass a strvec, it should almost certainly be >> const, to make it clear how we intend to use it. > > All true. > >> These cases are largely stupid things that real people would never come >> across. The real goal is making sure we don't get hit with a memory >> safety bug (under-allocation, converting a big size_t to a negative int, >> etc). > > Yes. > > Ævar, I do not mean any disrespect to you, but I have to say that > topics like this one are starting to wear my concentration and > patience down really thin and making me really slow down. I'm sorry. I'll try to lay off on the patch firehose until the delta I've got in master..seen is way down from what it is now. > Perhaps I am biased by not yet having seen what you eventually want > to build on top, and because of that I do not understood why and how > these "clean ups" are so valuable to have right now (as opposed to > just letting the sleeping dog lie), which you would of course have a > much better chance to know than I do. I could go into that, but it's probably best to leave it at: Yeah for these seemingly going nowhere small changes I'm generally taking them somewhere interesting. But figuring out how to split that is still a hard problem. E.g. I've had one series (the fsck error message improvements) that's been stalled for ~3 months due to size/including the "interesting" part, but recently relatively small increments of prep code changes seem to get a lot of review traction. As for this strvec.h s/int/size_t/ topic. I'm not taking that anywhere, Jeff suggested it and came up with the patch, I figured more helpful than "if we change s/int/size_t/g for x, shouldn't we change that for y which whe assign x to?" would be patches I had to do that, which I'd come up with after Jeff suggested this direction in response to another topic.
On Mon, Sep 13, 2021 at 02:29:01PM +0200, Ævar Arnfjörð Bjarmason wrote: > As for this strvec.h s/int/size_t/ topic. I'm not taking that anywhere, > Jeff suggested it and came up with the patch, I figured more helpful > than "if we change s/int/size_t/g for x, shouldn't we change that for y > which whe assign x to?" would be patches I had to do that, which I'd > come up with after Jeff suggested this direction in response to another > topic. I'm not inherently opposed to further int/size_t cleanups. But the trouble is that my single patch stands on its own as an improvement to a real issue, and does not (as far as I know) have any functional downsides (either known or even hypothetical, aside from the obvious mismatch that some callers will still use "int"). But doing wide-spread int/size_t conversion has less obvious immediate benefit, is much easier to get wrong, and may introduce further complications (e.g., differences of opinion in whether we should be passing strvecs around more, or just using size_t in more places). So I don't mind a series in that direction (though I don't necessarily think it is the best use of time), but I'd prefer not to see my original patch tied up in it. -Peff