Re: [RFC 2/2] io_uring: report pinned memory usage

[Jens Axboe <axboe@kernel.dk>]

On 6/12/20 10:43 PM, Bijan Mottahedeh wrote:
> On 6/12/2020 8:19 AM, Jens Axboe wrote:
>> On 6/12/20 9:16 AM, Jens Axboe wrote:
>>> On 6/11/20 8:23 PM, Bijan Mottahedeh wrote:
>>>> Long term, it makes sense to separate reporting and enforcing of pinned
>>>> memory usage.
>>>>
>>>> Signed-off-by: Bijan Mottahedeh <bijan.mottahedeh@oracle.com>
>>>>
>>>> It is useful to view
>>>> ---
>>>>   fs/io_uring.c | 4 ++++
>>>>   1 file changed, 4 insertions(+)
>>>>
>>>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>>>> index 4248726..cf3acaa 100644
>>>> --- a/fs/io_uring.c
>>>> +++ b/fs/io_uring.c
>>>> @@ -7080,6 +7080,8 @@ static int io_sq_offload_start(struct io_ring_ctx *ctx,
>>>>   static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
>>>>   {
>>>>   	atomic_long_sub(nr_pages, &user->locked_vm);
>>>> +	if (current->mm)
>>>> +		atomic_long_sub(nr_pages, &current->mm->pinned_vm);
>>>>   }
>>>>   
>>>>   static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
>>>> @@ -7096,6 +7098,8 @@ static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
>>>>   			return -ENOMEM;
>>>>   	} while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
>>>>   					new_pages) != cur_pages);
>>>> +	if (current->mm)
>>>> +		atomic_long_add(nr_pages, &current->mm->pinned_vm);
>>>>   
>>>>   	return 0;
>>>>   }
>>> current->mm should always be valid for these, so I think you can skip the
>>> checking of that and just make it unconditional.
>> Two other issues with this:
>>
>> - It's an atomic64, so seems more appropriate to use the atomic64 helpers
>>    for this one.
>> - The unaccount could potentially be a different mm, if the ring is shared
>>    and one task sets it up while another tears it down. So we'd need something
>>    to ensure consistency here.
>>
> Are you referring to a case where one process creates a ring and sends 
> the ring fd to another process?

Or a simpler case, where someone has submissions and completions running
on separate threads, and it just so happens that the completion side is
the one to exit the ring.

-- 
Jens Axboe