Re: [PATCH 2/2] module: add support to avoid duplicates early on load

[David Hildenbrand <david@redhat.com>]

On 31.05.23 18:57, Luis Chamberlain wrote:
> On Wed, May 31, 2023 at 09:51:41AM +0200, David Hildenbrand wrote:
>> 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
> 
> Great, don't forget:
> 
> diff --git a/kernel/module/main.c b/kernel/module/main.c
> index 82b0dcc1fe77..222015093eeb 100644
> --- a/kernel/module/main.c
> +++ b/kernel/module/main.c
> @@ -3066,7 +3066,7 @@ struct idempotent {
>   
>   #define IDEM_HASH_BITS 8
>   static struct hlist_head idem_hash[1 << IDEM_HASH_BITS];
> -static struct spinlock idem_lock;
> +static DEFINE_SPINLOCK(idem_lock);
>   
>   static bool idempotent(struct idempotent *u, const void *cookie)
>   {
> 

Finally was able to run it on that ThinkSystem SR950 with 8 populated
sockets -> 224 cores and 448 logical CPUs.

The KASAN vmap issues on that system were already no longer reproducible with your
(Luis) previous work that's already in master.

I tested a !debug and debug config (both based on corresponding RHEL9 configs), comparing
929ed21dfdb6 ("master") with 929ed21dfdb6 + Linus' patches ("patched").


Unfortunately, boot times vary a lot, and I did not figure out how to reduce
the noise. I captured the "systemd-analyze blame" output as well.


1) !debug config (not enabling KASAN)

a) master

Startup finished in 32.225s (kernel) + 7.399s (initrd) + 20.378s (userspace) = 1min 3ms
multi-user.target reached after 20.352s in userspace.
Startup finished in 43.734s (kernel) + 7.288s (initrd) + 19.827s (userspace) = 1min 10.851s
multi-user.target reached after 19.800s in userspace.
Startup finished in 50.514s (kernel) + 7.171s (initrd) + 24.757s (userspace) = 1min 22.443s
multi-user.target reached after 24.734s in userspace.
Startup finished in 26.722s (kernel) + 7.249s (initrd) + 23.923s (userspace) = 57.895s
multi-user.target reached after 23.892s in userspace.

b) patched

Startup finished in 36.318s (kernel) + 7.177s (initrd) + 21.383s (userspace) = 1min 4.879s
multi-user.target reached after 21.355s in userspace.
Startup finished in 36.318s (kernel) + 7.177s (initrd) + 21.383s (userspace) = 1min 4.879s
multi-user.target reached after 21.355s in userspace.
Startup finished in 1min 34.678s (kernel) + 7.239s (initrd) + 24.066s (userspace) = 2min 5.985s
multi-user.target reached after 24.040s in userspace.
Startup finished in 25.879s (kernel) + 7.144s (initrd) + 29.665s (userspace) = 1min 2.689s
multi-user.target reached after 29.637s in userspace.



2) debug config (enabling KASAN)

a) master

Startup finished in 2min 12.695s (kernel) + 25.058s (initrd) + 1min 13.012s (userspace) = 3min 50.765s
multi-user.target reached after 1min 12.903s in userspace.
Startup finished in 1min 45.400s (kernel) + 24.294s (initrd) + 1min 8.910s (userspace) = 3min 18.606s
multi-user.target reached after 1min 8.786s in userspace.
Startup finished in 2min 4.857s (kernel) + 24.715s (initrd) + 1min 5.088s (userspace) = 3min 34.660s
multi-user.target reached after 1min 4.967s in userspace.
Startup finished in 3min 20.400s (kernel) + 24.703s (initrd) + 1min 5.469s (userspace) = 4min 50.573s
multi-user.target reached after 1min 5.344s in userspace.

b) patched

Startup finished in 2min 5.250s (kernel) + 25.049s (initrd) + 1min 1.961s (userspace) = 3min 32.262s
multi-user.target reached after 1min 1.844s in userspace.
Startup finished in 1min 52.524s (kernel) + 24.897s (initrd) + 1min 5.062s (userspace) = 3min 22.484s
multi-user.target reached after 1min 4.916s in userspace.
Startup finished in 9min 36.817s (kernel) + 24.859s (initrd) + 1min 18.657s (userspace) = 11min 20.335s
multi-user.target reached after 1min 18.455s in userspace.
Startup finished in 30min 20.715s (kernel) + 24.722s (initrd) + 1min 7.039s (userspace) = 31min 52.476s
multi-user.target reached after 1min 6.907s in userspace.


What concerns me a bit, is that on the patched kernel we seem to hit more cases where
boot takes much longer (in both kernel configs).

I'll do some more runs/investigation to see if this is reproducible or just some system oddity.

Staring just at the udev settle time (systemd-analyze blame), it's very similar between both kernel
versions.

-- 
Cheers,

David / dhildenb