Hi Noah,

Hope you don't mind me adding elfutils-devel back to the CC.

On Fri, 2022-05-27 at 15:32 -0400, Noah Sanci wrote:
> On Fri, May 27, 2022 at 12:29 PM Mark Wielaard <mark@klomp.org>
> wrote:
> 
> > On Wed, 2022-05-25 at 14:44 -0400, Noah Sanci via Elfutils-devel
> > wrote:
> > > PR28577: Make run-debuginfod-fd-prefetch-caches.sh test
> > >              something
> > > 
> > > Update to the run-debuginfod-fd-prefetch to make the test more
> > > sound.
> > 
> > Thanks. And according to Evgeny this makes the test pass for him:
> > https://sourceware.org/bugzilla/show_bug.cgi?id=29180#c2
> 
> Good to hear.
> > But I find it a little hard to review, it would have been helpful
> > to
> > get a bit more comments/documentation what exactly is being tested
> > now.
> 
> Fair, I should have made what was being tested/why clearer.

Thanks. I also understand now that I read the test logic the wrong way
around.

> > > diff --git a/tests/run-debuginfod-fd-prefetch-caches.sh
> > > b/tests/run-
> > > debuginfod-fd-prefetch-caches.sh
> > > index 66a7a083..50ffc4ac 100755
> > > --- a/tests/run-debuginfod-fd-prefetch-caches.sh
> > > +++ b/tests/run-debuginfod-fd-prefetch-caches.sh
> > > @@ -21,22 +21,22 @@
> > >  # for test case debugging, uncomment:
> > >  set -x
> > >  unset VALGRIND_CMD
> > > +mkdir R Z
> > 
> > OK, some boilerplate.
> > > -mkdir R
> > > -cp -rvp ${abs_srcdir}/debuginfod-rpms R
> > > -if [ "$zstd" = "false" ]; then  # nuke the zstd fedora 31 ones
> > > -    rm -vrf R/debuginfod-rpms/fedora31
> > > -fi
> > 
> > This gets moved to below.
> 
> The two above changes should look better now
> 
> > >  DB=${PWD}/.debuginfod_tmp.sqlite
> > >  tempfiles $DB
> > > @@ -47,88 +47,68 @@ export
> > > DEBUGINFOD_CACHE_PATH=${PWD}/.client_cache
> > >  ######
> > >  rm -rf $DEBUGINFOD_CACHE_PATH
> > >  rm -rf $DB
> > > +
> > >  # Set mb values to ensure the caches certainly have enough space
> > >  # To store the test files
> > >  env LD_LIBRARY_PATH=$ldpath DEBUGINFOD_URLS=
> > > ${abs_builddir}/../debuginfod/debuginfod $VERBOSE -p $PORT1 -d
> > > $DB \
> > > -    --fdcache-fds=$FDCACHE_FDS --fdcache-prefetch-
> > > fds=$PREFETCH_FDS --fdcache-mintmp 0 -vvvvv -g 0 -t 0 -R R \
> > > -    --fdcache-mbs=50  --fdcache-prefetch-mbs=50 \
> > > -    --fdcache-prefetch=$PREFETCH  > vlog$PORT1 2>&1 &
> > > +    --fdcache-fds=$FDCACHE_FDS --fdcache-prefetch-
> > > fds=$PREFETCH_FDS  -vvvvv -g 0 -t 0 -R R \
> > > +    -Z .tar.bz2=bzcat Z  --fdcache-mbs=50  --fdcache-prefetch-
> > > mbs=100 \
> > > +    --fdcache-prefetch=$PREFETCH -F > vlog$PORT1 2>&1 &
> > 
> > Now run with --fdcache-fds=1 --fdcache-mbs=50 and
> > -fdcache-prefetch-fds=1 fdcache-prefetch=1 -fdcache-prefetch-
> > mbs=100
> > 
> > It is these last 3 we want to test.
> 
> This particular instance of debuginfod was intended to test
> fdcache-fds and prefetch-fds.
> The test is just testing prefetch-fds. fdcache-fds was tested in a
> different diff, but because the
> fdcache is just the basic lru the test was removed. Mb values are set
> high enough so that no
> files will be excluded for being too large. Maybe there should be a
> test prior to this for just
> fdcache-prefetch as it determines how much is prefetched.

OK. I think no extra test is needed.

> > >  PID1=$!
> > >  tempfiles vlog$PORT1
> > >  errfiles vlog$PORT1
> > > +
> > >  # Server must become ready
> > >  wait_ready $PORT1 'ready' 1
> > >  ################################################################
> > > ########
> > > -cp -rvp ${abs_srcdir}/debuginfod-rpms R
> > > -if [ "$zstd" = "false" ]; then  # nuke the zstd fedora 31 ones
> > > -    rm -vrf R/debuginfod-rpms/fedora31
> > > -fi
> > >  kill -USR1 $PID1
> > >  wait_ready $PORT1 'thread_work_total{role="traverse"}' 1
> > > +wait_ready $PORT1 'thread_work_pending{role="scan"}' 0
> > > +wait_ready $PORT1 'thread_busy{role="scan"}' 0
> > 
> > OK, boilerplate to wait for initial scan.
> 
> Yes, Frank told me that just checking traverse is insufficient to
> conclude a
> scan has completed.
> > > -archive_hits=5
> > > -SHA=f4a1a8062be998ae93b8f1cd744a398c6de6dbb1
> > > -for i in $archive_hits
> > > -do
> > > -  archive_test c36708a78618d597dee15d0dc989f093ca5f9120
> > > /usr/src/debug/hello2-1.0-2.x86_64/hello.c $SHA
> > > -done
> > > +export DEBUGINFOD_URLS=http://127.0.0.1:$PORT1/
> > > +# load prefetch cache
> > > +testrun ${abs_top_builddir}/debuginfod/debuginfod-find debuginfo
> > > cee13b2ea505a7f37bd20d271c6bc7e5f8d2dfcb
> > 
> > OK, which particular file is this?
> 
> pk.pkg.tar/usr/src/debug/hello.c, now elaborated in comment

Thanks.

> > >  metrics=$(curl http://127.0.0.1:$PORT1/metrics)
> > > -regex="fdcache_op_count\{op=\"enqueue\"\} ([0-
> > > 9]+).*fdcache_op_count\{op=\"evict\"\} ([0-9]+).*"
> > > -enqueue=0
> > > -if [[ $metrics =~ $regex ]]
> > > -then
> > > -   enqueue=${BASH_REMATCH[1]}
> > > -   evict=${BASH_REMATCH[2]}
> > > +regex="fdcache_prefetch_count ([0-9])+"
> > > +# Check to see if prefetch cache is loaded
> > > +if [[ $metrics =~ $regex ]]; then
> > > +  if [[  ${BASH_REMATCH[1]} -ne 1 ]]; then
> > > +    err
> > > +  fi
> > 
> > Why the -ne 1 ? What number are we expecting and why?
> 
> Ensuring that prefetch is loaded, but not loaded more than it is
> intended to hold.
> debuginfod will try to load 2 files into prefetch, so check to ensure
> that only 1 is loaded.
> Explained in comment.

Right. I misread the test condition. This makes sense now.

> > >  else
> > > -   err
> > > -fi
> > > -# This is an ad-hoc test than only works when FDCACHE_FDS <
> > > "archive_test"s above (5)
> > > -# otherwise there will be no eviction
> > > -if [[ $(( $enqueue-$FDCACHE_FDS )) -ne $evict ]]
> > > -then
> > > -   err
> > > -fi
> > > -# Test prefetch cache
> > > -archive_test bc1febfd03ca05e030f0d205f7659db29f8a4b30
> > > /usr/src/debug/hello-1.0/hello.c $SHA
> > > +  err
> > > +fi
> > > +# Ensure that searching for the source prefetched finds it in
> > > the prefetch cache
> > > +testrun ${abs_top_builddir}/debuginfod/debuginfod-find source
> > > cee13b2ea505a7f37bd20d271c6bc7e5f8d2dfcb /usr/src/debug/hello.c
> > >  metrics=$(curl http://127.0.0.1:$PORT1/metrics)
> > > -regex="fdcache_prefetch_count ([0-9])+"
> > > -pf_count=0
> > > -for i in $FDCACHE_FDS
> > > -do
> > > -  if [[ $metrics =~ $regex ]]; then
> > > -
> > > -    if [[ $(( $pf_count+$PREFETCH )) -ne ${BASH_REMATCH[1]} ]];
> > > then
> > > -      err
> > > -    else
> > > -      pf_count=${BASH_REMATCH[1]}
> > > -    fi
> > > -
> > > -  else
> > > +regex="fdcache_op_count\{op=\"prefetch_access\"\} ([0-9])+"
> > > +if [[ $metrics =~ $regex ]]; then
> > > +  if [[  ${BASH_REMATCH[1]} -ne 1 ]]; then
> > >      err
> > >    fi
> > > -done
> > > +else
> > > +  err
> > > +fi
> > 
> > Again -ne 1? Shouldn't it be different from the above now?
> 
> Now we're seeing if debuginfod is actually using the prefetch cache
> by checking the access count.
> We're searching for the 1 file we loaded before so we expect 1
> access.
> Explained in comment

Thanks.

> > >  metrics=$(curl http://127.0.0.1:$PORT1/metrics)
> > > -regex="fdcache_bytes ([0-9]+).*fdcache_prefetch_bytes ([0-9]+)"
> > > +regex="fdcache_bytes ([0-9]+)"
> > >  mb=$((1024*1024))
> > >  if [[ $metrics =~ $regex ]]; then
> > >    fdbytes=${BASH_REMATCH[1]}
> > > -  pfbytes=${BASH_REMATCH[2]}
> > > -  if [ $fdbytes -gt $mb ] || [ $pfbytes -gt $mb ]; then
> > > +  if [ $fdbytes -gt $mb ] ; then
> > >      err
> > >    fi
> > 
> > OK, check that the prefetch cache isn't bigger than 10MB. Why that
> > number?
> 
> mb should equal 1mb in bytes as this is the maximum space allocated
> to both the prefetch and fd cache. and curling metrics returns memory
> used in bytes

Thanks. I was off by a factor 10. Sorry.

> Please find the new patch attached.

Thanks for adding all the comments. Looks good.

Cheers,

Mark