On 11/16/21 12:21 PM, Vito Caputo wrote:

On Tue, Nov 16, 2021 at 11:55:41AM -0700, Jens Axboe wrote:

quoted

On 11/16/21 11:36 AM, Matthew Wilcox wrote:

quoted

On Mon, Nov 15, 2021 at 08:35:30PM -0800, Andrew Morton wrote:

quoted

I'd also be interested in seeing feedback from the MM developers.

[...]

quoted

Subject: Increase default MLOCK_LIMIT to 8 MiB

On the one hand, processes can already allocate at least this much
memory that is non-swappable, just by doing things like opening a lot of
files (allocating struct file & fdtable), using a lot of address space
(allocating page tables), so I don't have a problem with it per se.

On the other hand, 64kB is available on anything larger than an IBM XT.
Linux will still boot on machines with 4MB of RAM (eg routers).  For
someone with a machine with only, say, 32MB of memory, this allows a
process to make a quarter of the memory unswappable, and maybe that's
not a good idea.  So perhaps this should scale over a certain range?

Is 8MB a generally useful amount of memory for an iouring user anyway?
If you're just playing with it, sure, but if you have, oh i don't know,
a database, don't you want to pin the entire cache and allow IO to the
whole thing?

8MB is plenty for most casual use cases, which is exactly the ones that
we want to "just work" without requiring weird system level
modifications to increase the memlock limit.

Considering a single fullscreen 32bpp 4K-resolution framebuffer is
~32MiB, I'm not convinced this is really correct in nearly 2022.

You don't need to register any buffers, and I don't expect any basic
uses cases to do so. Which means that the 8MB just need to cover the
ring itself, and you can fit a _lot_ of rings into 8MB. The memlock
limit only applies to buffers if you register them, not for any "normal"
use cases where you just pass buffers for read/write or O_DIRECT
read/write.

If we're going to bump the default at the kernel, I'm with Matthew on
making it autoscale within a sane range, depending on available
memory.

I just don't want to turn this into a bikeshedding conversation. I'm
fine with making it autoscale obviously, but who's going to do the work?

As an upper bound I'd probably look at the highest anticipated
consumer resolutions, and handle a couple fullscreen 32bpp instances
being pinned.

Not sure I see the relevance here.

-- 
Jens Axboe