Date:   Wed, 22 Sep 2021 09:36:09 +0200
From:   Mauro Carvalho Chehab <mchehab+huawei@...nel.org>
To:     Greg Kroah-Hartman <gregkh@...uxfoundation.org>
Cc:     Linux Doc Mailing List <linux-doc@...r.kernel.org>,
        linux-kernel@...r.kernel.org, Jonathan Corbet <corbet@....net>,
        Anton Vorontsov <anton@...msg.org>,
        Colin Cross <ccross@...roid.com>,
        John Fastabend <john.fastabend@...il.com>,
        KP Singh <kpsingh@...nel.org>,
        Kees Cook <keescook@...omium.org>,
        Martin KaFai Lau <kafai@...com>,
        Song Liu <songliubraving@...com>,
        Tony Luck <tony.luck@...el.com>, Yonghong Song <yhs@...com>,
        bpf@...r.kernel.org, netdev@...r.kernel.org
Subject: Re: [PATCH v3 0/7] get_abi.pl: Check for missing symbols at the ABI
 specs

Em Wed, 22 Sep 2021 08:22:33 +0200
Greg Kroah-Hartman <gregkh@...uxfoundation.org> escreveu:

> On Wed, Sep 22, 2021 at 07:43:42AM +0200, Greg Kroah-Hartman wrote:
> > On Tue, Sep 21, 2021 at 08:16:33PM +0200, Mauro Carvalho Chehab wrote:
> > > Em Tue, 21 Sep 2021 18:52:42 +0200
> > > Greg Kroah-Hartman <gregkh@...uxfoundation.org> escreveu:
> > > 
> > > > On Sat, Sep 18, 2021 at 11:52:10AM +0200, Mauro Carvalho Chehab wrote:
> > > > > Hi Greg,
> > > > > 
> > > > > Add a new feature at get_abi.pl to optionally check for existing symbols
> > > > > under /sys that won't match a "What:" inside Documentation/ABI.
> > > > > 
> > > > > Such feature is very useful to detect missing documentation for ABI.
> > > > > 
> > > > > This series brings a major speedup, plus it fixes a few border cases when
> > > > > matching regexes that end with a ".*" or \d+.
> > > > > 
> > > > > patch 1 changes get_abi.pl logic to handle multiple What: lines, in
> > > > > order to make the script more robust;
> > > > > 
> > > > > patch 2 adds the basic logic. It runs really quicky (up to 2
> > > > > seconds), but it doesn't use sysfs softlinks.
> > > > > 
> > > > > Patch 3 adds support for parsing softlinks. It makes the script a
> > > > > lot slower, making it take a couple of minutes to process the entire
> > > > > sysfs files. It could be optimized in the future by using a graph,
> > > > > but, for now, let's keep it simple.
> > > > > 
> > > > > Patch 4 adds an optional parameter to allow filtering the results
> > > > > using a regex given by the user. When this parameter is used
> > > > > (which should be the normal usecase), it will only try to find softlinks
> > > > > if the sysfs node matches a regex.
> > > > > 
> > > > > Patch 5 improves the report by avoiding it to ignore What: that
> > > > > ends with a wildcard.
> > > > > 
> > > > > Patch 6 is a minor speedup.  On a Dell Precision 5820, after patch 6, 
> > > > > results are:
> > > > > 
> > > > > 	$ time ./scripts/get_abi.pl undefined |sort >undefined && cat undefined| perl -ne 'print "$1\n" if (m#.*/(\S+) not found#)'|sort|uniq -c|sort -nr >undefined_symbols; wc -l undefined; wc -l undefined_symbols
> > > > > 
> > > > > 	real	2m35.563s
> > > > > 	user	2m34.346s
> > > > > 	sys	0m1.220s
> > > > > 	7595 undefined
> > > > > 	896 undefined_symbols
> > > > > 
> > > > > Patch 7 makes a *huge* speedup: it basically switches a linear O(n^3)
> > > > > search for links by a logic which handle symlinks using BFS. It
> > > > > also addresses a border case that was making 'msi-irqs/\d+' regex to
> > > > > be misparsed. 
> > > > > 
> > > > > After patch 7, it is 11 times faster:
> > > > > 
> > > > > 	$ time ./scripts/get_abi.pl undefined |sort >undefined && cat undefined| perl -ne 'print "$1\n" if (m#.*/(\S+) not found#)'|sort|uniq -c|sort -nr >undefined_symbols; wc -l undefined; wc -l undefined_symbols
> > > > > 
> > > > > 	real	0m14.137s
> > > > > 	user	0m12.795s
> > > > > 	sys	0m1.348s
> > > > > 	7030 undefined
> > > > > 	794 undefined_symbols
> > > > > 
> > > > > (the difference on the number of undefined symbols are due to the fix for
> > > > > it to properly handle 'msi-irqs/\d+' regex)
> > > > > 
> > > > > -
> > > > > 
> > > > > While this series is independent from Documentation/ABI changes, it
> > > > > works best when applied from this tree, which also contain ABI fixes
> > > > > and a couple of additions of frequent missed symbols on my machine:
> > > > > 
> > > > >     https://git.kernel.org/pub/scm/linux/kernel/git/mchehab/devel.git/log/?h=get_undefined_abi_v3  
> > > > 
> > > > I've taken all of these, but get_abi.pl seems to be stuck in an endless
> > > > loop or something.  I gave up and stopped it after 14 minutes.  It had
> > > > stopped printing out anything after finding all of the pci attributes
> > > > that are not documented :)
> > > 
> > > It is probably not an endless loop, just there are too many vars to
> > > check on your system, which could make it really slow.
> > 
> > Ah, yes, I ran it overnight and got the following:
> > 
> > $ time ./scripts/get_abi.pl undefined |sort >undefined && cat undefined| perl -ne 'print "$1\n" if (m#.*/(\S+) not found#)'|sort|uniq -c|sort -nr >undefined_symbols; wc -l undefined; wc -l undefined_symbols
> > 
> > real	29m39.503s
> > user	29m37.556s
> > sys	0m0.851s
> > 26669 undefined
> > 765 undefined_symbols
> > 
> > > The way the search algorithm works is that reduces the number of regex 
> > > expressions that will be checked for a given file entry at sysfs. It 
> > > does that by looking at the devnode name. For instance, when it checks for
> > > this file:
> > > 
> > > 	/sys/bus/pci/drivers/iosf_mbi_pci/bind
> > > 
> > > The logic will seek only the "What:" expressions that end with "bind".
> > > Currently, there are just two What expressions for it[1]:
> > > 
> > > 	What: /sys/bus/fsl\-mc/drivers/.*/bind
> > > 	What: /sys/bus/pci/drivers/.*/bind
> > > 
> > > It will then run an O(n²) algorithm to seek:
> > > 
> > > 		foreach my $a (@names) {
> > >                        foreach my $w (split /\xac/, $what) {
> > >                                if ($a =~ m#^$w$#) {
> > > 					exact = 1;
> > >                                         last;
> > >                                 }
> > > 			}
> > > 		}
> > > 
> > > Which runs quickly, when there are few regexs to seek. There are, 
> > > however, some What: expressions that end with a wildcard. Those are
> > > harder to process. Right now, they're all grouped together, which
> > > makes them slower. Most of the processing time are spent on those.
> > > 
> > > I'm working right now on some strategy to also speed up the search 
> > > for them. Once I get something better, I'll send a patch series.
> > > 
> > > --
> > > 
> > > [1] On a side note, there are currently some problems with the What:
> > >     definitions for bind/unbind, as:
> > > 
> > > 	- it doesn't match all PCI devices;
> > > 	- it doesn't match ACPI and other buses that also export
> > > 	  bind/unbind.
> > > 
> > > > 
> > > > Anything I can do to help debug this?
> > > >
> > > 
> > > There are two parameters that can help to identify the issue:
> > > 
> > > a) You can add a "--show-hints" parameter. This turns on some 
> > >    prints that may help to identify what the script is doing.
> > >    It is not really a debug option, but it helps to identify
> > >    when some regexes are failing.
> > > 
> > > b) You can limit the What expressions that will be parsed with:
> > > 	   --search-string <something>
> > > 
> > > You can combine both. For instance, if you want to make it
> > > a lot more verbose, you could run it as:
> > > 
> > > 	./scripts/get_abi.pl undefined --search-string /sys --show-hints
> > 
> > Let me run this and time stamp it to see where it is getting hung up on.
> > Give it another 30 minutes :)
> 
> Hm, that didn't make too much sense as to what it was stalled on.  I've
> attached the compressed file if you are curious.

Hmm...

	[07:52:44] --> /sys/devices/pci0000:40/0000:40:01.3/0000:4a:00.1/iommu/amd-iommu/cap
	[08:07:52] --> /sys/devices/pci0000:40/0000:40:01.1/0000:41:00.0/0000:42:05.0/iommu/amd-iommu/cap

It sounds it took quite a while handling iommu cap, which sounds weird, as
it should be looking just 3 What expressions:

	[07:43:06] What: /sys/class/iommu/.*/amd\-iommu/cap
	[07:43:06] What: /sys/class/iommu/.*/intel\-iommu/cap
	[07:43:06] What: /sys/devices/pci.*.*.*.*\:.*.*/0000\:.*.*\:.*.*..*/dma/dma.*chan.*/quickdata/cap

Maybe there was a memory starvation while running the script, causing
swaps. Still, it is weird that it would happen there, as the hashes
and arrays used at the script are all allocated before it starts the
search logic. Here, the allocation part takes ~2 seconds.

At least on my Dell Precision 5820 (12 cpu threads), the amount of memory it
uses is not huge:

    $ /usr/bin/time -v ./scripts/get_abi.pl undefined >/dev/null
	Command being timed: "./scripts/get_abi.pl undefined"
	User time (seconds): 12.68
	System time (seconds): 1.29
	Percent of CPU this job got: 99%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 0:13.98
	Average shared text size (kbytes): 0
	Average unshared data size (kbytes): 0
	Average stack size (kbytes): 0
	Average total size (kbytes): 0
	Maximum resident set size (kbytes): 212608
	Average resident set size (kbytes): 0
	Major (requiring I/O) page faults: 0
	Minor (reclaiming a frame) page faults: 52003
	Voluntary context switches: 1
	Involuntary context switches: 56
	Swaps: 0
	File system inputs: 0
	File system outputs: 0
	Socket messages sent: 0
	Socket messages received: 0
	Signals delivered: 0
	Page size (bytes): 4096
	Exit status: 0

Unfortunately, I don't have any amd-based machine here, but I'll
try to run it later on a big arm server and see how it behaves.

> Anyway, this is all in my tree now, and I'll gladly take patches to make
> it go faster :)

Ok!

Thanks,
Mauro

Powered by blists - more mailing lists