Re: [PATCH v2 2/3] mm: introduce process_mrelease system call

On Wed, Jul 21, 2021 at 8:43 AM Suren Baghdasaryan [off-list ref] wrote:

On Wed, Jul 21, 2021 at 1:02 AM David Hildenbrand [off-list ref] wrote:

quoted

On 18.07.21 23:41, Suren Baghdasaryan wrote:

quoted

In modern systems it's not unusual to have a system component monitoring
memory conditions of the system and tasked with keeping system memory
pressure under control. One way to accomplish that is to kill
non-essential processes to free up memory for more important ones.
Examples of this are Facebook's OOM killer daemon called oomd and
Android's low memory killer daemon called lmkd.
For such system component it's important to be able to free memory
quickly and efficiently. Unfortunately the time process takes to free
up its memory after receiving a SIGKILL might vary based on the state
of the process (uninterruptible sleep), size and OPP level of the core
the process is running. A mechanism to free resources of the target
process in a more predictable way would improve system's ability to
control its memory pressure.
Introduce process_mrelease system call that releases memory of a dying
process from the context of the caller. This way the memory is freed in
a more controllable way with CPU affinity and priority of the caller.
The workload of freeing the memory will also be charged to the caller.
The operation is allowed only on a dying process.

Previously I proposed a number of alternatives to accomplish this:
- https://lore.kernel.org/patchwork/patch/1060407 extending
pidfd_send_signal to allow memory reaping using oom_reaper thread;
- https://lore.kernel.org/patchwork/patch/1338196 extending
pidfd_send_signal to reap memory of the target process synchronously from
the context of the caller;
- https://lore.kernel.org/patchwork/patch/1344419/ to add MADV_DONTNEED
support for process_madvise implementing synchronous memory reaping.

To me, this looks a lot cleaner. Although I do wonder why we need two
separate mechanisms to achieve the end goal

1. send sigkill
2. process_mrelease

As 2. doesn't make sense without 1. it somehow feels like it would be
optimal to achieve both steps in a single syscall. But I remember there
were discussions around that.

Yep, we recently discussed the approach in this thread:
https://lore.kernel.org/patchwork/patch/1450952/#1652452

quoted

quoted

The end of the last discussion culminated with suggestion to introduce a
dedicated system call (https://lore.kernel.org/patchwork/patch/1344418/#1553875)
The reasoning was that the new variant of process_madvise
   a) does not work on an address range
   b) is destructive
   c) doesn't share much code at all with the rest of process_madvise
 From the userspace point of view it was awkward and inconvenient to provide
memory range for this operation that operates on the entire address space.
Using special flags or address values to specify the entire address space
was too hacky.

The API is as follows,

           int process_mrelease(int pidfd, unsigned int flags);

         DESCRIPTION
           The process_mrelease() system call is used to free the memory of
           a process which was sent a SIGKILL signal.

           The pidfd selects the process referred to by the PID file
           descriptor.
           (See pidofd_open(2) for further information)

           The flags argument is reserved for future use; currently, this
           argument must be specified as 0.

         RETURN VALUE
           On success, process_mrelease() returns 0. On error, -1 is
           returned and errno is set to indicate the error.

         ERRORS
           EBADF  pidfd is not a valid PID file descriptor.

           EAGAIN Failed to release part of the address space.

           EINVAL flags is not 0.

           EINVAL The task does not have a pending SIGKILL or its memory is
                  shared with another process with no pending SIGKILL.

           ENOSYS This system call is not supported by kernels built with no
                  MMU support (CONFIG_MMU=n).

           ESRCH  The target process does not exist (i.e., it has terminated
                  and been waited on).

Signed-off-by: Suren Baghdasaryan <surenb@google.com>
---
  mm/oom_kill.c | 55 +++++++++++++++++++++++++++++++++++++++++++++++++++
  1 file changed, 55 insertions(+)

diff --git a/mm/oom_kill.c b/mm/oom_kill.c
index d04a13dc9fde..7fbfa70d4e97 100644
--- a/mm/oom_kill.c
+++ b/mm/oom_kill.c

@@ -28,6 +28,7 @@
  #include <linux/sched/task.h>
  #include <linux/sched/debug.h>
  #include <linux/swap.h>
+#include <linux/syscalls.h>
  #include <linux/timex.h>
  #include <linux/jiffies.h>
  #include <linux/cpuset.h>

@@ -755,10 +756,64 @@ static int __init oom_init(void)
      return 0;
  }
  subsys_initcall(oom_init)
+
+SYSCALL_DEFINE2(process_mrelease, int, pidfd, unsigned int, flags)
+{
+     struct pid *pid;
+     struct task_struct *task;
+     struct mm_struct *mm = NULL;
+     unsigned int f_flags;
+     long ret = 0;

Nit: reverse Christmas tree.

Ack. Will reorder like this:

struct mm_struct *mm = NULL;
struct task_struct *task;
unsigned int f_flags;
struct pid *pid;
long ret = 0;

quoted

quoted

+
+     if (flags != 0)
+             return -EINVAL;
+
+     pid = pidfd_get_pid(pidfd, &f_flags);
+     if (IS_ERR(pid))
+             return PTR_ERR(pid);
+
+     task = get_pid_task(pid, PIDTYPE_PID);
+     if (!task) {
+             ret = -ESRCH;
+             goto put_pid;
+     }
+
+     /*
+      * If the task is dying and in the process of releasing its memory
+      * then get its mm.
+      */
+     task_lock(task);
+     if (task_will_free_mem(task) && (task->flags & PF_KTHREAD) == 0) {
+             mm = task->mm;
+             mmget(mm);
+     }

AFAIU, while holding the task_lock, task->mm won't change and we cannot
see a concurrent exit_mm()->mmput(). So the mm structure and the VMAs
won't go away while holding the task_lock(). I do wonder if we need the
mmget() at all here.

We do mmget() here to ensure mm is stable when it is passed later to
__oom_reap_task_mm(mm)/mmap_read_lock(mm)/mmap_read_unlock(mm). Note
that during those calls we do not hold task_lock anymore.

quoted

Also, I wonder if it would be worth dropping the task_lock() while
reaping - to unblock anybody else wanting to lock the task.

As I mentioned above, we do not hold task_lock during reaping. We
release it right after we call task_will_free_mem(), which checks that
the task is exiting. task_lock is held during the call to
task_will_free_mem() to satisfy the requirement listed in that
function's comment: "Caller has to make sure that task->mm is stable
(hold task_lock or it operates on the current)".

quoted

Getting a hold of the mm and locking the mmap_lock would be sufficient I guess.

That's exactly what I do here. The simplified sequence is:

       task_lock
       if (task_will_free_mem())
               mm=mmget()
       task_unlock
       if (!mm) return;

       mmap_read_lock(mm)
       __oom_reap_task_mm(mm)
       mmap_read_unlock(mm)
       mmput(mm)

Or did I misunderstand your comments?

Let me take a closer look at the locking sequence here and will follow
up afterwards.
Thanks for the review!

quoted

In general, looks quite good to me.

--
Thanks,

David / dhildenb