| Message ID | pull.846.git.1611637582625.gitgitgadget@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | strbuf.c: optimize program logic | expand |
On Mon, Jan 25, 2021 at 10:17:41PM -0800, Junio C Hamano wrote: > "阿德烈 via GitGitGadget" <gitgitgadget@gmail.com> writes: > > > From: ZheNing Hu <adlternative@gmail.com> > > > > the usage in strbuf.h tell us"Alloc is somehow a > > "private" member that should not be messed with. > > use `strbuf_avail()`instead." > > When we use the word "private", it generally means it is private to > the implementation of the API. IOW, it is usually fine for the > implementation of the API (i.e. for strbuf API, what you see in > strbuf.c) to use private members. > > In any case, these changes are _not_ optimizations. Yeah, I had both of those thoughts, too. :) Though... > Replacing (alloc - len - 1) with strbuf_avail() is at best an > equivalent rewrite (which is a good thing from readability's point > of view, but not an optimization). We know sb->alloc during the > loop is never 0, but the compiler may miss the fact, so the inlined > implementation of _avail, i.e. > > static inline size_t strbuf_avail(const struct strbuf *sb) > { > return sb->alloc ? sb->alloc - sb->len - 1 : 0; > } > > may not incur call overhead, but may be pessimizing the executed > code. > > If you compare the code in the loop in the second hunk below with > what _setlen() does, I think you'll see the overhead of _setlen() > relative to the original code is even higher, so it may also be > pessimizing, not optimizing. > > So, overall, I am not all that enthused to see this patch. I would generally value readability/consistency here over trying to micro-optimize an if-zero check. However, if strbuf_avail() ever did return 0, I'm not sure the loop would make forward progress: strbuf_grow(sb, hint ? hint : 8192); for (;;) { ssize_t want = strbuf_avail(sb); ssize_t got = read_in_full(fd, sb->buf + sb->len, want); if (got < 0) { if (oldalloc == 0) strbuf_release(sb); else strbuf_setlen(sb, oldlen); return -1; } strbuf_setlen(sb, sb->len + got); if (got < want) break; strbuf_grow(sb, 8192); } we'd just ask to read 0 bytes over and over. That almost makes me want to add: if (!want) BUG("strbuf did not actually grow!?"); or possibly to teach the "if (got < want)" condition to check for a zero return (though I guess that would probably just end up confusing us into thinking we hit EOF). > One thing I noticed is that, whether open coded like sb->len += got > or made into parameter to strbuf_setlen(sb, sb->len + got), we are > not careful about sb->len growing too large and overflowing with the > addition. That may potentially be an interesting thing to look > into, but at the same time, unlike the usual "compute the number of > bytes we need to allocate and then call xmalloc()" pattern, where we > try to be careful in the "compute" step by using st_add() macros, > this code actually keep growing the buffer, so by the time the size_t > overflows and wraps around, we'd certainly have exhausted the memory > already, so it won't be an issue. I think "len" is OK here. An invariant of strbuf is that "len" is smaller than "alloc" for obvious reasons. So as long as the actual strbuf_grow() is safe, then extending "len". I'm not sure that strbuf_grow() is safe, though. It relies on ALLOC_GROW, which does not use st_add(), etc. -Peff PS The original patch does not seem to have made it to the list for some reason (I didn't get a copy, and neither did lore.kernel.org).
胡哲宁 <adlternative@gmail.com> writes: > Junio C Hamano <gitster@pobox.com> 于2021年1月26日周二 下午2:17写道： >> >> "阿德烈 via GitGitGadget" <gitgitgadget@gmail.com> writes: >> >> > From: ZheNing Hu <adlternative@gmail.com> >> > >> > the usage in strbuf.h tell us"Alloc is somehow a >> > "private" member that should not be messed with. >> > use `strbuf_avail()`instead." >> >> When we use the word "private", it generally means it is private to >> the implementation of the API. IOW, it is usually fine for the >> implementation of the API (i.e. for strbuf API, what you see in >> strbuf.c) to use private members. >> > Well, I just think most other functions in strbuf.c follow the use > of `strbuf_avail()` instead of "sb->alloc-sb->len-1", and the > "sb->alloc-sb->len-1" that appears in `strbuf_read()` is not so uniform. I actually wouldn't have minded if this were sold as "code clean-up to use _avail() when we open-code in the implementation of strbuf API in codepaths that are not performance critical." I am not sure about the _setlen() side of the thing. It is quite obvious what is going on in the original, and it falls into "when it is written one way that is good enough, replacing it with another that is not significantly better often ends up being mere code churn.", I would think. Thanks.
Jeff King <peff@peff.net> writes: > PS The original patch does not seem to have made it to the list for some > reason (I didn't get a copy, and neither did lore.kernel.org). Yes, it seems today's one of these days where vger is hiccuping. I see out of order deliveries of responses to messages I've yet to see.
Jeff King <peff@peff.net> 于2021年1月27日周三 上午2:23写道： > > On Mon, Jan 25, 2021 at 10:17:41PM -0800, Junio C Hamano wrote: > > > "阿德烈 via GitGitGadget" <gitgitgadget@gmail.com> writes: > > > > > From: ZheNing Hu <adlternative@gmail.com> > > > > > > the usage in strbuf.h tell us"Alloc is somehow a > > > "private" member that should not be messed with. > > > use `strbuf_avail()`instead." > > > > When we use the word "private", it generally means it is private to > > the implementation of the API. IOW, it is usually fine for the > > implementation of the API (i.e. for strbuf API, what you see in > > strbuf.c) to use private members. > > > > In any case, these changes are _not_ optimizations. > > Yeah, I had both of those thoughts, too. :) > > Though... > > > Replacing (alloc - len - 1) with strbuf_avail() is at best an > > equivalent rewrite (which is a good thing from readability's point > > of view, but not an optimization). We know sb->alloc during the > > loop is never 0, but the compiler may miss the fact, so the inlined > > implementation of _avail, i.e. > > > > static inline size_t strbuf_avail(const struct strbuf *sb) > > { > > return sb->alloc ? sb->alloc - sb->len - 1 : 0; > > } > > > > may not incur call overhead, but may be pessimizing the executed > > code. > > > > If you compare the code in the loop in the second hunk below with > > what _setlen() does, I think you'll see the overhead of _setlen() > > relative to the original code is even higher, so it may also be > > pessimizing, not optimizing. > > > > So, overall, I am not all that enthused to see this patch. > > I would generally value readability/consistency here over trying to > micro-optimize an if-zero check. > > However, if strbuf_avail() ever did return 0, I'm not sure the loop > would make forward progress: > > strbuf_grow(sb, hint ? hint : 8192); > for (;;) { > ssize_t want = strbuf_avail(sb); > ssize_t got = read_in_full(fd, sb->buf + sb->len, want); > > if (got < 0) { > if (oldalloc == 0) > strbuf_release(sb); > else > strbuf_setlen(sb, oldlen); > return -1; > } > strbuf_setlen(sb, sb->len + got); > if (got < want) > break; > strbuf_grow(sb, 8192); > } > > we'd just ask to read 0 bytes over and over. That almost makes me want > to add: > > if (!want) > BUG("strbuf did not actually grow!?"); > > or possibly to teach the "if (got < want)" condition to check for a zero > return (though I guess that would probably just end up confusing us into > thinking we hit EOF). > I agree.But I always feel that there will be no such strange bug. > > One thing I noticed is that, whether open coded like sb->len += got > > or made into parameter to strbuf_setlen(sb, sb->len + got), we are > > not careful about sb->len growing too large and overflowing with the > > addition. That may potentially be an interesting thing to look > > into, but at the same time, unlike the usual "compute the number of > > bytes we need to allocate and then call xmalloc()" pattern, where we > > try to be careful in the "compute" step by using st_add() macros, > > this code actually keep growing the buffer, so by the time the size_t > > overflows and wraps around, we'd certainly have exhausted the memory > > already, so it won't be an issue. > > I think "len" is OK here. An invariant of strbuf is that "len" is > smaller than "alloc" for obvious reasons. So as long as the actual > strbuf_grow() is safe, then extending "len". > > I'm not sure that strbuf_grow() is safe, though. It relies on > ALLOC_GROW, which does not use st_add(), etc. > I want to say is that `strbuf_grow()` have checked overflow before `ALLOC_GROW`,so that `strbuf_grow()`is maybe also safe too? :) void strbuf_grow(struct strbuf *sb, size_t extra) { int new_buf = !sb->alloc; if (unsigned_add_overflows(extra, 1) || unsigned_add_overflows(sb->len, extra + 1)) die("you want to use way too much memory"); ... } > -Peff > > PS The original patch does not seem to have made it to the list for some > reason (I didn't get a copy, and neither did lore.kernel.org).
On Fri, Jan 29, 2021 at 02:09:12PM +0800, 胡哲宁 wrote: > > I'm not sure that strbuf_grow() is safe, though. It relies on > > ALLOC_GROW, which does not use st_add(), etc. > > > I want to say is that `strbuf_grow()` have checked overflow before > `ALLOC_GROW`,so that `strbuf_grow()`is maybe also safe too? :) > void strbuf_grow(struct strbuf *sb, size_t extra) > { > int new_buf = !sb->alloc; > if (unsigned_add_overflows(extra, 1) || > unsigned_add_overflows(sb->len, extra + 1)) > die("you want to use way too much memory"); > ... Oh, you're right. I misread it as checking only "extra", but of course the second line there is making sure our total requested size does not overflow. I do think ALLOC_GROW() could still overflow internally as it sizes larger than sb->len + extra. But this is quite unlikely on a 64-bit system, as it would imply we're using on the order of 2^63 bytes of RAM before we enter the function. It potentially could be a problem on a 32-bit system, though I'm not sure how much in practice (the numbers are small enough to be reasonable, but I'm not sure it's realistic that a single strbuf could already be using half of the available address space). -Peff
diff --git a/strbuf.c b/strbuf.c index e3397cc4c72..76f560a28d0 100644 --- a/strbuf.c +++ b/strbuf.c @@ -517,7 +517,7 @@ ssize_t strbuf_read(struct strbuf *sb, int fd, size_t hint) strbuf_grow(sb, hint ? hint : 8192); for (;;) { - ssize_t want = sb->alloc - sb->len - 1; + ssize_t want = strbuf_avail(sb); ssize_t got = read_in_full(fd, sb->buf + sb->len, want); if (got < 0) { @@ -527,7 +527,7 @@ ssize_t strbuf_read(struct strbuf *sb, int fd, size_t hint) strbuf_setlen(sb, oldlen); return -1; } - sb->len += got; + strbuf_setlen(sb, sb->len + got); if (got < want) break; strbuf_grow(sb, 8192); @@ -543,7 +543,7 @@ ssize_t strbuf_read_once(struct strbuf *sb, int fd, size_t hint) ssize_t cnt; strbuf_grow(sb, hint ? hint : 8192); - cnt = xread(fd, sb->buf + sb->len, sb->alloc - sb->len - 1); + cnt = xread(fd, sb->buf + sb->len, strbuf_avail(sb)); if (cnt > 0) strbuf_setlen(sb, sb->len + cnt); else if (oldalloc == 0)