It turns out that udev under certain circumstances will concurrently try
to load the same modules over-and-over excessively. This isn't a kernel
bug, but it ends up affecting the kernel, to the point that under
certain circumstances we can fail to boot, because the kernel uses a lot
of memory to read all the module data all at once.
Note that it isn't a memory leak, it's just basically a thundering herd
problem happening at bootup with a lot of CPUs, with the worst cases
then being pretty bad.
Admittedly the worst situations are somewhat contrived: lots and lots of
CPUs, not a lot of memory, and KASAN enabled to make it all slower and
as such (unintentionally) exacerbate the problem.
Luis explains: [1]
"My best assessment of the situation is that each CPU in udev ends up
triggering a load of duplicate set of modules, not just one, but *a
lot*. Not sure what heuristics udev uses to load a set of modules per
CPU."
Petr Pavlu chimes in: [2]
"My understanding is that udev workers are forked. An initial kmod
context is created by the main udevd process but no sharing happens
after the fork. It means that the mentioned memory pool logic doesn't
really kick in.
Multiple parallel load requests come from multiple udev workers, for
instance, each handling an udev event for one CPU device and making
the exactly same requests as all others are doing at the same time.
The optimization idea would be to recognize these duplicate requests
at the udevd/kmod level and converge them"
Note that module loading has tried to mitigate this issue before, see
for example commit 064f4536d1 ("module: avoid allocation if module is
already present and ready"), which has a few ASCII graphs on memory use
due to this same issue.
However, while that noticed that the module was already loaded, and
exited with an error early before spending any more time on setting up
the module, it didn't handle the case of multiple concurrent module
loads all being active - but not complete - at the same time.
Yes, one of them will eventually win the race and finalize its copy, and
the others will then notice that the module already exists and error
out, but while this all happens, we have tons of unnecessary concurrent
work being done.
Again, the real fix is for udev to not do that (maybe it should use
threads instead of fork, and have actual shared data structures and not
cause duplicate work). That real fix is apparently not trivial.
But it turns out that the kernel already has a pretty good model for
dealing with concurrent access to the same file: the i_writecount of the
inode.
In fact, the module loading already indirectly uses 'i_writecount' ,
because 'kernel_file_read()' will in fact do
ret = deny_write_access(file);
if (ret)
return ret;
...
allow_write_access(file);
around the read of the file data. We do not allow concurrent writes to
the file, and return -ETXTBUSY if the file was open for writing at the
same time as the module data is loaded from it.
And the solution to the reader concurrency problem is to simply extend
this "no concurrent writers" logic to simply be "exclusive access".
Note that "exclusive" in this context isn't really some absolute thing:
it's only exclusion from writers and from other "special readers" that
do this writer denial. So we simply introduce a variation of that
"deny_write_access()" logic that not only denies write access, but also
requires that this is the _only_ such access that denies write access.
Which means that you can't start loading a module that is already being
loaded as a module by somebody else, or you will get the same -ETXTBSY
error that you would get if there were writers around.
[ It also means that you can't try to load a currently executing
executable as a module, for the same reason: executables do that same
"deny_write_access()" thing, and that's obviously where the whole
ETXTBSY logic traditionally came from.
This is not a problem for kernel modules, since the set of normal
executable files and kernel module files is entirely disjoint. ]
This new function is called "exclusive_deny_write_access()", and the
implementation is trivial, in that it's just an atomic decrement of
i_writecount if it was 0 before.
To use that new exclusivity check, all we then do is wrap the module
loading with that exclusive_deny_write_access()() / allow_write_access()
pair. The actual patch is a bit bigger than that, because we want to
surround not just the "load file data" part, but the whole module setup,
to get maximum exclusion.
So this ends up splitting up "finit_module()" into a few helper
functions to make it all very clear and legible.
In Luis' test-case (bringing up 255 vcpu's in a virtual machine [3]),
the "wasted vmalloc" space (ie module data read into a vmalloc'ed area
in order to be loaded as a module, but then discarded because somebody
else loaded the same module instead) dropped from 1.8GiB to 474kB. Yes,
that's gigabytes to kilobytes.
It doesn't drop completely to zero, because even with this change, you
can still end up having completely serial pointless module loads, where
one udev process has loaded a module fully (and thus the kernel has
released that exclusive lock on the module file), and then another udev
process tries to load the same module again.
So while we cannot fully get rid of the fundamental bug in user space,
we _can_ get rid of the excessive concurrent thundering herd effect.
A couple of final side notes on this all:
- This tweak only affects the "finit_module()" system call, which gives
the kernel a file descriptor with the module data.
You can also just feed the module data as raw data from user space
with "init_module()" (note the lack of 'f' at the beginning), and
obviously for that case we do _not_ have any "exclusive read" logic.
So if you absolutely want to do things wrong in user space, and try
to load the same module multiple times, and error out only later when
the kernel ends up saying "you can't load the same module name
twice", you can still do that.
And in fact, some distros will do exactly that, because they will
uncompress the kernel module data in user space before feeding it to
the kernel (mainly because they haven't started using the new kernel
side decompression yet).
So this is not some absolute "you can't do concurrent loads of the
same module". It's literally just a very simple heuristic that will
catch it early in case you try to load the exact same module file at
the same time, and in that case avoid a potentially nasty situation.
- There is another user of "deny_write_access()": the verity code that
enables fs-verity on a file (the FS_IOC_ENABLE_VERITY ioctl).
If you use fs-verity and you care about verifying the kernel modules
(which does make sense), you should do it *before* loading said
kernel module. That may sound obvious, but now the implementation
basically requires it. Because if you try to do it concurrently, the
kernel may refuse to load the module file that is being set up by the
fs-verity code.
- This all will obviously mean that if you insist on loading the same
module in parallel, only one module load will succeed, and the others
will return with an error.
That was true before too, but what is different is that the -ETXTBSY
error can be returned *before* the success case of another process
fully loading and instantiating the module.
Again, that might sound obvious, and it is indeed the whole point of
the whole change: we are much quicker to notice the whole "you're
already in the process of loading this module".
So it's very much intentional, but it does mean that if you just
spray the kernel with "finit_module()", and expect that the module is
immediately loaded afterwards without checking the return value, you
are doing something horribly horribly wrong.
I'd like to say that that would never happen, but the whole _reason_
for this commit is that udev is currently doing something horribly
horribly wrong, so ...
Link: https://lore.kernel.org/all/ZEGopJ8VAYnE7LQ2@bombadil.infradead.org/ [1]
Link: https://lore.kernel.org/all/23bd0ce6-ef78-1cd8-1f21-0e706a00424a@suse.com/ [2]
Link: https://lore.kernel.org/lkml/ZG%2Fa+nrt4%2FAAUi5z@bombadil.infradead.org/ [3]
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Lucas De Marchi <lucas.demarchi@intel.com>
Cc: Petr Pavlu <petr.pavlu@suse.com>
Tested-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
