[PATCH] mm/madvise: introduce process_madvise() syscall: an external memory hinting API

[Minchan Kim]

There is usecase that System Management Software(SMS) want to give a
memory hint like MADV_[COLD|PAGEEOUT] to other processes and in the case
of Android, it is the ActivityManagerService.

The information required to make the reclaim decision is not known to
the app. Instead, it is known to the centralized userspace
daemon(ActivityManagerService), and that daemon must be able to initiate
reclaim on its own without any app involvement.

To solve the issue, this patch introduces a new syscall process_madvise(2). 
It uses pidfd of an external process to give the hint.

 int process_madvise(int pidfd, void *addr, size_t length, int advice,
			unsigned long flags);

Since it could affect other process's address range, only privileged
process(CAP_SYS_PTRACE) or something else(e.g., being the same UID) gives
it the right to ptrace the process could use it successfully.  The flag
argument is reserved for future use if we need to extend the API.

I think supporting all hints madvise has/will supported/support to
process_madvise is rather risky.  Because we are not sure all hints make
sense from external process and implementation for the hint may rely on
the caller being in the current context so it could be error-prone.  Thus,
I just limited hints as MADV_[COLD|PAGEOUT] in this patch.

If someone want to add other hints, we could hear hear the usecase and
review it for each hint.  It's safer for maintenance rather than
introducing a buggy syscall but hard to fix it later.

Q.1 - Why does any external entity have better knowledge?

Quote from Sandeep

"For Android, every application (including the special SystemServer) are
forked from Zygote.  The reason of course is to share as many libraries
and classes between the two as possible to benefit from the preloading
during boot.

After applications start, (almost) all of the APIs end up calling into
this SystemServer process over IPC (binder) and back to the application.

In a fully running system, the SystemServer monitors every single process
periodically to calculate their PSS / RSS and also decides which process
is "important" to the user for interactivity.

So, because of how these processes start _and_ the fact that the
SystemServer is looping to monitor each process, it does tend to *know*
which address range of the application is not used / useful.

Besides, we can never rely on applications to clean things up themselves.
We've had the "hey app1, the system is low on memory, please trim your
memory usage down" notifications for a long time[1].  They rely on
applications honoring the broadcasts and very few do.

So, if we want to avoid the inevitable killing of the application and
restarting it, some way to be able to tell the OS about unimportant memory
in these applications will be useful.

- ssp

Q.2 - How to guarantee the race(i.e., object validation) between when
giving a hint from an external process and get the hint from the target
process?

process_madvise operates on the target process's address space as it
exists at the instant that process_madvise is called.  If the space target
process can run between the time the process_madvise process inspects the
target process address space and the time that process_madvise is actually
called, process_madvise may operate on memory regions that the calling
process does not expect.  It's the responsibility of the process calling
process_madvise to close this race condition.  For example, the calling
process can suspend the target process with ptrace, SIGSTOP, or the
freezer cgroup so that it doesn't have an opportunity to change its own
address space before process_madvise is called.  Another option is to
operate on memory regions that the caller knows a priori will be unchanged
in the target process.  Yet another option is to accept the race for
certain process_madvise calls after reasoning that mistargeting will do no
harm.  The suggested API itself does not provide synchronization.  It also
apply other APIs like move_pages, process_vm_write.

The race isn't really a problem though.  Why is it so wrong to require
that callers do their own synchronization in some manner?  Nobody objects
to write(2) merely because it's possible for two processes to open the
same file and clobber each other's writes --- instead, we tell people to
use flock or something.  Think about mmap.  It never guarantees newly
allocated address space is still valid when the user tries to access it
because other threads could unmap the memory right before.  That's where
we need synchronization by using other API or design from userside.  It
shouldn't be part of API itself.  If someone needs more fine-grained
synchronization rather than process level, there were two ideas suggested
- cookie[2] and anon-fd[3].  Both are applicable via using last reserved
argument of the API but I don't think it's necessary right now since we
have already ways to prevent the race so don't want to add additional
complexity with more fine-grained optimization model.

To make the API extend, it reserved an unsigned long as last argument so
we could support it in future if someone really needs it.

Q.3 - Why doesn't ptrace work?

Injecting an madvise in the target process using ptrace would not work for
us because such injected madvise would have to be executed by the target
process, which means that process would have to be runnable and that
creates the risk of the abovementioned race and hinting a wrong VMA.
Furthermore, we want to act the hint in caller's context, not the
callee's, because the callee is usually limited in cpuset/cgroups or even
freezed state so they can't act by themselves quick enough, which causes
more thrashing/kill.  It doesn't work if the target process are
ptraced(e.g., strace, debugger, minidump) because a process can have at
most one ptracer.

[1] https://developer.android.com/topic/performance/memory";

[2] process_getinfo for getting the cookie which is updated whenever
    vma of process address layout are changed - Daniel Colascione -
    https://lore.kernel.org/lkml/20190520035254.57579-1-minchan@xxxxxxxxxx/T/#m7694416fd179b2066a2c62b5b139b14e3894e224

[3] anonymous fd which is used for the object(i.e., address range)
    validation - Michal Hocko -
    https://lore.kernel.org/lkml/20200120112722.GY18451@xxxxxxxxxxxxxx/

Link: http://lkml.kernel.org/r/20200302193630.68771-3-minchan@xxxxxxxxxx
Signed-off-by: Minchan Kim <minchan@xxxxxxxxxx>
Reviewed-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
Reviewed-by: Vlastimil Babka <vbabka@xxxxxxx>
Cc: Alexander Duyck <alexander.h.duyck@xxxxxxxxxxxxxxx>
Cc: Brian Geffon <bgeffon@xxxxxxxxxx>
Cc: Christian Brauner <christian@xxxxxxxxxx>
Cc: Daniel Colascione <dancol@xxxxxxxxxx>
Cc: Jann Horn <jannh@xxxxxxxxxx>
Cc: Jens Axboe <axboe@xxxxxxxxx>
Cc: Joel Fernandes <joel@xxxxxxxxxxxxxxxxx>
Cc: Johannes Weiner <hannes@xxxxxxxxxxx>
Cc: John Dias <joaodias@xxxxxxxxxx>
Cc: Kirill Tkhai <ktkhai@xxxxxxxxxxxxx>
Cc: Michal Hocko <mhocko@xxxxxxxx>
Cc: Oleksandr Natalenko <oleksandr@xxxxxxxxxx>
Cc: Sandeep Patil <sspatil@xxxxxxxxxx>
Cc: SeongJae Park <sj38.park@xxxxxxxxx>
Cc: SeongJae Park <sjpark@xxxxxxxxx>
Cc: Shakeel Butt <shakeelb@xxxxxxxxxx>
Cc: Sonny Rao <sonnyrao@xxxxxxxxxx>
Cc: Tim Murray <timmurray@xxxxxxxxxx>
Cc: <linux-man@xxxxxxxxxxxxxxx>
Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>