Thread (83 messages) 83 messages, 10 authors, 2021-03-27

Re: [PATCH] Makefile: fix bugs in coccicheck and speed it up

From: Ævar Arnfjörð Bjarmason <hidden>
Date: 2021-03-05 17:21:49

On Fri, Mar 05 2021, Jeff King wrote:

I sent out a v2 that should address your feedback in :
https://lore.kernel.org/git/20210305170724.23859-1-avarab@gmail.com/ (local)

Just brief notes:
On Tue, Mar 02, 2021 at 09:51:03PM +0100, Ævar Arnfjörð Bjarmason wrote:
quoted
What we were doing before was processing all our *.c files, and for
each of those *.c files we'd recursively look around for includes and
see if we needed to search/replace in those too.

That we did that dates back to [1] when we were only processing *.c
files, and it was always very redundant. We'd e.g. visit the likes of
strbuf.h lots of times since it's widely used as an include.
OK. I don't offhand know if processing the includes along with the C
files buys us anything else. Coccinelle's behavior is often quite a
mystery, but if we think this produces the same results with less time,
I'm for it.

BTW, this "dates back to when we were only processing *.c files"
statement confused me a bit. Doesn't that include the current state
before this patch?  I.e., this hunk:
quoted
-FOUND_C_SOURCES = $(filter %.c,$(shell $(FIND_SOURCE_FILES)))
+FOUND_C_SOURCES = $(filter %.c %.h,$(shell $(FIND_SOURCE_FILES)))
seems to be an important part of the change, without which moving away
from --all-includes would be wrong.
Updated in the new 2/4 commit message.
quoted
Then in the most recent attempt to optimize coccicheck in [2] this
anti-pattern finally turned into a bug.

Namely: before this change, if your coccicheck rule applied to
e.g. making a change in strbuf.h itself we'd get *lots* of duplicate
hunks applying the exact same change, as concurrent spatch processes
invoked by xargs raced one another. In one instance I ended up with 27
copies of the same hunk in a strbuf.patch.

Setting SPATCH_BATCH_SIZE=0 and processing all the files in one giant
batch mitigated this. I suspect the author of [2] either mostly ran in
that mode, or didn't test on changes that impacted widely used header
files.
This "turned into a bug" I didn't understand, though. That commit [2]
just switched from always feeding files one at a time to letting you
feed more than one.  So yes, feeding all at once (with
SPATCH_BATCH_SIZE=0) mitigated the bug. But wouldn't any bug that shows
up with SPATCH_BATCH_SIZE=1 have already existed previously?

IOW, I don't think the batch-size stuff made anything worse. It made one
specific case better, but that was not even its purpose.

That is maybe splitting hairs, but I want to make sure I understand all
of what you're claiming here. But also...
Explained in the new 2/4. Your commit introduced a regression, because
spatch does racy magic background locking, running N of them at a time
applying the same rule tripped it up.

I suspect it was racy before, but we didn't have overlapping rules in
multiple *.cocci files.
quoted hunk ↗ jump to hunk
quoted
I'm entirely removing SPATCH_BATCH_SIZE. If you want to tweak it you
can tweak SPATCH_XARGS_FLAGS to e.g. "-n 256", or "-P 4 -n 128". But
in my testing it isn't worth it to tweak SPATCH_XARGS_FLAGS for a full
"make coccicheck".
You hard-code this to 32 now. But I'm not sure if that's going to be the
universal best value.

Applying just part of your patch, but leaving SPATCH_BATCH_SIZE in
place, like so:
diff --git a/Makefile b/Makefile
index dfb0f1000f..88cb157547 100644
--- a/Makefile
+++ b/Makefile
@@ -1201,7 +1201,7 @@ SP_EXTRA_FLAGS = -Wno-universal-initializer
 # For the 'coccicheck' target; setting SPATCH_BATCH_SIZE higher will
 # usually result in less CPU usage at the cost of higher peak memory.
 # Setting it to 0 will feed all files in a single spatch invocation.
-SPATCH_FLAGS = --all-includes --patch .
+SPATCH_FLAGS = --no-includes --patch .
 SPATCH_BATCH_SIZE = 1
 
 include config.mak.uname
@@ -2859,7 +2859,7 @@ check: config-list.h command-list.h
 		exit 1; \
 	fi
 
-FOUND_C_SOURCES = $(filter %.c,$(shell $(FIND_SOURCE_FILES)))
+FOUND_C_SOURCES = $(filter %.c %.h,$(shell $(FIND_SOURCE_FILES)))
 COCCI_SOURCES = $(filter-out $(THIRD_PARTY_SOURCES),$(FOUND_C_SOURCES))
 
 %.cocci.patch: %.cocci $(COCCI_SOURCES)
Then I get this with various batch sizes (using "make -j16" on an 8-core
machine, so all of the patches are being run at the same time):

  size |  real  |  user  | sys
  -------------------------------
    1  |  1m28s | 10m00s | 0m56s
    2  |  1m03s |  7m49s | 0m33s
    4  |  0m51s |  6m28s | 0m20s
    8  |  0m45s |  6m02s | 0m14s
   16  |  0m45s |  6m17s | 0m10s
   32  |  0m44s |  6m33s | 0m07s
   64  |  0m45s |  6m48s | 0m06s
    0  |  1m08s | 10m08s | 0m03s

So there's a sweet spot at 8. Doing "32" isn't that much worse (runtime
is about the same, but we use more CPU), but it gets progressively worse
as the batch size increases.

That's the same sweet spot before your patch, too, btw. I know because I
happened to be timing it the other day, as coccinelle 1.1.0 hit debian
unstable. When I originally wrote the SPATCH_BATCH_SIZE feature, I think
I was using 1.0.4 or 1.0.7. Back then SPATCH_BATCH_SIZE=0 was a clear
win, assuming you had the memory. But in 1.0.8 it ran for many minutes
without finishing. I found back then that "2" was the sweet spot. But
now it's "8".

All of which is to say: the timing difference between my "8" and "32"
cases isn't that exciting. But the performance of spatch has been
sufficiently baffling to me that I think it's worth keeping the knob.

Your XARGS_FLAGS does accomplish something similar (and is in fact more
flexible, though at the cost of being less abstract).  I'm OK to replace
one with the other, but shouldn't it happen in a separate patch? It's
completely orthogonal to the --no-includes behavior.
*nod*, set to a default of 8 in the new 4/4.
quoted hunk ↗ jump to hunk
quoted
I'm also the whole "cat $@.log" introduced in [3]. Since we don't call
this in a loop anymore (and xargs will early-exit) we can just rely on
standard V=1 for debugging issues.
I think this is missing the point of the cat. We've redirected spatch's
stderr to the logfile. So if there's an error, you have no idea what it
was without manually figuring out which file to cat. And V=1 doesn't
help that.

We could stop doing that redirection, but the problem is that spatch is
very verbose. So the point was to store the output but only dump it
when we see an error.

So this part of the patch seems like a pure regression to me. E.g.,
introduce a syntax error like:
diff --git a/contrib/coccinelle/free.cocci b/contrib/coccinelle/free.cocci
index 4490069df9..c6c6562a0a 100644
--- a/contrib/coccinelle/free.cocci
+++ b/contrib/coccinelle/free.cocci
@@ -11,7 +11,7 @@ expression E;
   free(E);
 
 @@
-expression E;
+expression E
 @@
 - free(E);
 + FREE_AND_NULL(E);
and run "make coccicheck" before and after your patch:

  [before]
  $ make coccicheck
  GIT_VERSION = 2.31.0.rc1.8.gbe7935ed8b.dirty
      SPATCH contrib/coccinelle/free.cocci
  init_defs_builtins: /usr/bin/../lib/coccinelle/standard.h
  meta: parse error: 
    File "contrib/coccinelle/free.cocci", line 15, column 0, charpos = 99
    around = '@@',
    whole content = @@

  xargs: spatch: exited with status 255; aborting
  make: *** [Makefile:2866: contrib/coccinelle/free.cocci.patch] Error 1

  [after]
  $ make coccicheck
    SPATCH contrib/coccinelle/free.cocci
  make: *** [Makefile:2875: contrib/coccinelle/free.cocci.patch] Error 124
The "cat $@.log" is back, but there's still some issues with it on
master now (I didn't change this) it'll spew a lot at you with xargs
since we emit the whole error output for every file, seemingly, before
cat-ing the file:

    make -j1 coccicheck SPATCH_FLAGS=--blahblah V=1 2>&1|grep -i example.*use|wc -l
    464

Well, it's a bit better now on my series with a default batch size of 8:

    $ make -j1 coccicheck SPATCH_FLAGS=--blahblah V=1 2>&1|grep -i example.*use|wc -l
    88

I got tired of dealing with the combination of shellscritp and make for
the day. But maybe something to do as a follow-up if anyone's
interested.
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help