Re: [PATCH RFC 4/5] mm/ksm, proc: introduce remote merge

Hi.

On Thu, May 16, 2019 at 12:00:24PM +0200, Jann Horn wrote:

On Thu, May 16, 2019 at 11:43 AM Oleksandr Natalenko
[off-list ref] wrote:

quoted

Use previously introduced remote madvise knob to mark task's
anonymous memory as mergeable.

To force merging task's VMAs, "merge" hint is used:

   # echo merge > /proc/<pid>/madvise

Force unmerging is done similarly:

   # echo unmerge > /proc/<pid>/madvise

To achieve this, previously introduced ksm_madvise_*() helpers
are used.

Why does this not require PTRACE_MODE_ATTACH_FSCREDS to the target
process? Enabling KSM on another process is hazardous because it
significantly increases the attack surface for side channels.

(Note that if you change this to require PTRACE_MODE_ATTACH_FSCREDS,
you'll want to use mm_access() in the ->open handler and drop the mm
in ->release. mm_access() from a ->write handler is not permitted.)

Sounds reasonable. So, something similar to what mem_open() & friends do
now:

static int madvise_open(...)
...
	struct task_struct *task = get_proc_task(inode);
...
	if (task) {
		mm = mm_access(task, PTRACE_MODE_ATTACH_FSCREDS);
		put_task_struct(task);
		if (!IS_ERR_OR_NULL(mm)) {
			mmgrab(mm);
			mmput(mm);
...

Then:

static ssize_t madvise_write(...)
...
	if (!mmget_not_zero(mm))
		goto out;

	down_write(&mm->mmap_sem);
	if (!mmget_still_valid(mm))
		goto skip_mm;
...
skip_mm:
	up_write(&mm->mmap_sem);

	mmput(mm);
out:
	return ...;

And, finally:

static int madvise_release(...)
...
		mmdrop(mm);
...

Right?

[...]

quoted

@@ -2960,15 +2962,63 @@ static int proc_stack_depth(struct seq_file *m, struct pid_namespace *ns,
 static ssize_t madvise_write(struct file *file, const char __user *buf,
                size_t count, loff_t *ppos)
 {
+       /* For now, only KSM hints are implemented */
+#ifdef CONFIG_KSM
+       char buffer[PROC_NUMBUF];
+       int behaviour;
        struct task_struct *task;
+       struct mm_struct *mm;
+       int err = 0;
+       struct vm_area_struct *vma;
+
+       memset(buffer, 0, sizeof(buffer));
+       if (count > sizeof(buffer) - 1)
+               count = sizeof(buffer) - 1;
+       if (copy_from_user(buffer, buf, count))
+               return -EFAULT;
+
+       if (!memcmp("merge", buffer, min(sizeof("merge")-1, count)))

This means that you also match on something like "mergeblah". Just use strcmp().

I agree. Just to make it more interesting I must say that

   /sys/kernel/mm/transparent_hugepage/enabled

uses memcmp in the very same way, and thus echoing "alwaysssss" or
"madviseeee" works perfectly there, and it was like that from the very
beginning, it seems. Should we fix it, or it became (zomg) a public API?

quoted

+               behaviour = MADV_MERGEABLE;
+       else if (!memcmp("unmerge", buffer, min(sizeof("unmerge")-1, count)))
+               behaviour = MADV_UNMERGEABLE;
+       else
+               return -EINVAL;

        task = get_proc_task(file_inode(file));
        if (!task)
                return -ESRCH;

+       mm = get_task_mm(task);
+       if (!mm) {
+               err = -EINVAL;
+               goto out_put_task_struct;
+       }
+
+       down_write(&mm->mmap_sem);

Should a check for mmget_still_valid(mm) be inserted here? See commit
04f5866e41fb70690e28397487d8bd8eea7d712a.

Yeah, it seems so :/. Thanks for the pointer. I've put it into the
madvise_write snippet above.

quoted

+       switch (behaviour) {
+       case MADV_MERGEABLE:
+       case MADV_UNMERGEABLE:

This switch isn't actually necessary at this point, right?

Yup, but it is there to highlight a possibility of adding other, non-KSM
options. So, let it be, and I'll just re-arrange CONFIG_KSM ifdef
instead.

Thank you.

quoted

+               vma = mm->mmap;
+               while (vma) {
+                       if (behaviour == MADV_MERGEABLE)
+                               ksm_madvise_merge(vma->vm_mm, vma, &vma->vm_flags);
+                       else
+                               ksm_madvise_unmerge(vma, vma->vm_start, vma->vm_end, &vma->vm_flags);
+                       vma = vma->vm_next;
+               }
+               break;
+       }
+       up_write(&mm->mmap_sem);
+
+       mmput(mm);
+
+out_put_task_struct:
        put_task_struct(task);

-       return count;
+       return err ? err : count;
+#else
+       return -EINVAL;
+#endif /* CONFIG_KSM */
 }

-- 
  Best regards,
    Oleksandr Natalenko (post-factum)
    Senior Software Maintenance Engineer