From: David Hildenbrand <david@redhat.com>
To: Luis Chamberlain <mcgrof@kernel.org>,
Linus Torvalds <torvalds@linux-foundation.org>
Cc: Johan Hovold <johan@kernel.org>,
Lucas De Marchi <lucas.demarchi@intel.com>,
Petr Pavlu <petr.pavlu@suse.com>,
gregkh@linuxfoundation.org, rafael@kernel.org, song@kernel.org,
lucas.de.marchi@gmail.com, christophe.leroy@csgroup.eu,
peterz@infradead.org, rppt@kernel.org, dave@stgolabs.net,
willy@infradead.org, vbabka@suse.cz, mhocko@suse.com,
dave.hansen@linux.intel.com, colin.i.king@gmail.com,
jim.cromie@gmail.com, catalin.marinas@arm.com, jbaron@akamai.com,
rick.p.edgecombe@intel.com, yujie.liu@intel.com,
tglx@linutronix.de, hch@lst.de, patches@lists.linux.dev,
linux-modules@vger.kernel.org, linux-mm@kvack.org,
linux-kernel@vger.kernel.org, pmladek@suse.com,
prarit@redhat.com, lennart@poettering.net
Subject: Re: [PATCH 2/2] module: add support to avoid duplicates early on load
Date: Wed, 31 May 2023 09:51:41 +0200 [thread overview]
Message-ID: <499e30cc-d015-8353-1364-50d17da58f47@redhat.com> (raw)
In-Reply-To: <ZHYitt7P7W+8ZlSB@bombadil.infradead.org>
On 30.05.23 18:22, Luis Chamberlain wrote:
> On Mon, May 29, 2023 at 09:55:15PM -0400, Linus Torvalds wrote:
>> On Mon, May 29, 2023 at 11:18 AM Johan Hovold <johan@kernel.org> wrote:
>>>
>>> I took a closer look at some of the modules that failed to load and
>>> noticed a pattern in that they have dependencies that are needed by more
>>> than one device.
>>
>> Ok, this is a "maybe something like this" RFC series of two patches -
>> one trivial one to re-organize things a bit so that we can then do the
>> real one which uses a filter based on the inode pointer to return an
>> "idempotent return value" for module loads that share the same inode.
>>
>> It's entirely untested, and since I'm on the road I'm going to not
>> really be able to test it. It compiles for me, and the code looks
>> fairly straightforward, but it's probably buggy.
>>
>> It's very loosely based on Luis' attempt, but it
>> (a) is internal to module loading
>> (b) uses a reliable cookie
>> (c) doesn't leave the cookie around randomly for later
>> (d) has seen absolutely no testing
>>
>> Put another way: if somebody wants to play with this, please treat it
>> as a starting point, not the final thing. You might need to debug
>> things, and fix silly mistakes.
>>
>> The idea is to just have a simple hash list of currently executing
>> module loads, protected by a trivial spinlock. Every module loader
>> adds itself to the right hash list, and if they were the *first* one
>> (ie no other pending module loads for that inode), will actually do
>> the module load.
>>
>> Everybody who *isn't* the first one will just wait for completion and
>> return the same error code that the first one returned.
>
> That's also a hell much more snazzier MODULE_DEBUG_AUTOLOAD_DUPS if we
> ever wanted to do something similar there if we wanted to also
> join request_module() calls, instead of it hiding under debug.
>
>> This is technically bogus. The first one might fail due to arguments.
>
> For boot it's fine, as I can't think of boot wanting to support trying
> to load a module with different arguments but who knows. But I can't
> see it sensible to issue concurrent multiple requests for modules
> with different arguments without waiting in userspace for the first
> to fail.
>
> Even post-boot, doing that sounds rather insane, but it would certainly
> be a compromise and should probably be clearly documented. I think just
> a comment acknolwedging that corner case seems sensible.
>
> Because we won't be able to get the arguments until we process the
> module, so it would be too late for this optimization on kread. So it is
> why I had also stuck to the original feature being in kread, as then it
> provides a uniq kread call and the caller is aware of it. But indeed I
> had not considered the effects of arguments.
>
> Lucas, any thoughts from modules kmod userspace perspective into
> supporting anyone likely issuing concurrent modules requests with
> differing arguments?
>
>> So the cookie shouldn't be just the inode, it should be the inode and
>> a hash of the arguments or something like that.
>
> Personally I think it's a fine optimization without the arguments.
>
>> But it is what it is,
>> and apart from possible show-stopper bugs this is no worse than the
>> failed "exclusive write deny" attempt. IOW - maybe worth trying?
>
> The only thing I can think of is allowing threads other than the
> first one to complete before the one that actually loaded the
> module. I thought about this race for module auto-loading, see
> the comment in kmod_dup_request_announce(), so that just
> further delays the completion to other thread with a stupid
> queue_work(). That seems more important for module auto-loading
> duplicates than for boot finit_module() duplicates. But not sure
> if odering matters in the end due to a preemtible kernel and maybe
> that concern is hysteria.
>
>> And if *that* didn't sell people on this patch series, I don't know
>> what will. I should be in marketing! Two drink minimums, here I come!
>
> Sold:
>
> on 255 vcpus 0 duplicates found with this setup:
>
> root@kmod ~ # cat /sys/kernel/debug/modules/stats
> Mods ever loaded 66
> Mods failed on kread 0
> Mods failed on decompress 0
> Mods failed on becoming 0
> Mods failed on load 0
> Total module size 11268096
> Total mod text size 4149248
> Failed kread bytes 0
> Failed decompress bytes 0
> Failed becoming bytes 0
> Failed kmod bytes 0
> Virtual mem wasted bytes 0
> Average mod size 170729
> Average mod text size 62868
>
> So:
>
> Tested-by: Luis Chamberlain <mcgrof@kernel.org>
>
> In terms of bootup timing:
>
> Before:
> Startup finished in 41.653s (kernel) + 44.305s (userspace) = 1min 25.958s
> graphical.target reached after 44.178s in userspace.
>
> After:
> Startup finished in 23.995s (kernel) + 40.350s (userspace) = 1min 4.345s
> graphical.target reached after 40.226s in userspace.
I'll try grabbing the system where we saw the KASAN-related issues [1]
and give it a churn with and without the two patches. Might take a bit
(~1 day), unfortunately.
[1] https://lkml.kernel.org/r/20221013180518.217405-1-david@redhat.com
--
Thanks,
David / dhildenb
next prev parent reply other threads:[~2023-05-31 7:54 UTC|newest]
Thread overview: 53+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-05-24 21:36 [PATCH 0/2] module: avoid all memory pressure due to duplicates Luis Chamberlain
2023-05-24 21:36 ` [PATCH 1/2] fs/kernel_read_file: add support for duplicate detection Luis Chamberlain
2023-05-24 21:52 ` Linus Torvalds
2023-05-24 21:56 ` Linus Torvalds
2023-05-24 22:07 ` Luis Chamberlain
2023-05-25 4:00 ` Linus Torvalds
2023-05-25 18:08 ` Luis Chamberlain
2023-05-25 18:35 ` Luis Chamberlain
2023-05-25 18:50 ` Linus Torvalds
2023-05-25 19:32 ` Luis Chamberlain
2023-05-25 7:01 ` Christian Brauner
2023-05-24 21:36 ` [PATCH 2/2] module: add support to avoid duplicates early on load Luis Chamberlain
2023-05-25 11:40 ` Petr Pavlu
2023-05-25 16:07 ` Linus Torvalds
2023-05-25 16:42 ` Greg KH
2023-05-25 18:22 ` Luis Chamberlain
2023-05-25 17:52 ` Linus Torvalds
2023-05-25 18:45 ` Lucas De Marchi
2023-05-25 21:12 ` Linus Torvalds
2023-05-25 22:02 ` Luis Chamberlain
2023-05-26 1:39 ` Linus Torvalds
2023-05-29 8:58 ` Johan Hovold
2023-05-29 11:00 ` Linus Torvalds
2023-05-29 12:44 ` Johan Hovold
2023-05-29 15:18 ` Johan Hovold
2023-05-30 1:55 ` Linus Torvalds
2023-05-30 9:40 ` Johan Hovold
2023-06-05 12:25 ` Johan Hovold
2023-05-30 16:22 ` Luis Chamberlain
2023-05-30 17:16 ` Lucas De Marchi
2023-05-30 19:41 ` Luis Chamberlain
2023-05-30 22:17 ` Linus Torvalds
2023-05-31 5:30 ` Lucas De Marchi
2023-05-31 0:31 ` Luis Chamberlain
2023-05-31 7:51 ` David Hildenbrand [this message]
2023-05-31 16:57 ` Luis Chamberlain
2023-06-02 15:19 ` David Hildenbrand
2023-06-02 16:04 ` Luis Chamberlain
2023-06-05 11:26 ` David Hildenbrand
2023-06-05 15:17 ` Luis Chamberlain
2023-06-05 15:28 ` Luis Chamberlain
2023-06-28 18:52 ` Luis Chamberlain
2023-06-28 20:14 ` Linus Torvalds
2023-06-28 22:07 ` Linus Torvalds
2023-06-28 23:17 ` Linus Torvalds
2023-06-29 0:18 ` Luis Chamberlain
2023-06-02 16:06 ` Linus Torvalds
2023-06-02 16:37 ` David Hildenbrand
2023-05-30 22:45 ` Dan Williams
2023-06-04 14:26 ` Rudi Heitbaum
2023-05-29 17:47 ` Linus Torvalds
2023-05-30 10:01 ` Johan Hovold
2023-05-25 16:54 ` Lucas De Marchi
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=499e30cc-d015-8353-1364-50d17da58f47@redhat.com \
--to=david@redhat.com \
--cc=catalin.marinas@arm.com \
--cc=christophe.leroy@csgroup.eu \
--cc=colin.i.king@gmail.com \
--cc=dave.hansen@linux.intel.com \
--cc=dave@stgolabs.net \
--cc=gregkh@linuxfoundation.org \
--cc=hch@lst.de \
--cc=jbaron@akamai.com \
--cc=jim.cromie@gmail.com \
--cc=johan@kernel.org \
--cc=lennart@poettering.net \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-mm@kvack.org \
--cc=linux-modules@vger.kernel.org \
--cc=lucas.de.marchi@gmail.com \
--cc=lucas.demarchi@intel.com \
--cc=mcgrof@kernel.org \
--cc=mhocko@suse.com \
--cc=patches@lists.linux.dev \
--cc=peterz@infradead.org \
--cc=petr.pavlu@suse.com \
--cc=pmladek@suse.com \
--cc=prarit@redhat.com \
--cc=rafael@kernel.org \
--cc=rick.p.edgecombe@intel.com \
--cc=rppt@kernel.org \
--cc=song@kernel.org \
--cc=tglx@linutronix.de \
--cc=torvalds@linux-foundation.org \
--cc=vbabka@suse.cz \
--cc=willy@infradead.org \
--cc=yujie.liu@intel.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).