Re: [RFC PATCH 05/19] RMDA/siw: Convert to use vm_account

[Alistair Popple]

Bernard Metzler <BMT@xxxxxxxxxxxxxx> writes:

>> -----Original Message-----
>> From: Jason Gunthorpe <jgg@xxxxxxxxxx>
>> Sent: Tuesday, 24 January 2023 15:37
>> To: Alistair Popple <apopple@xxxxxxxxxx>
>> Cc: linux-mm@xxxxxxxxx; cgroups@xxxxxxxxxxxxxxx; linux-
>> kernel@xxxxxxxxxxxxxxx; jhubbard@xxxxxxxxxx; tjmercier@xxxxxxxxxx;
>> hannes@xxxxxxxxxxx; surenb@xxxxxxxxxx; mkoutny@xxxxxxxx; daniel@xxxxxxxx;
>> Bernard Metzler <BMT@xxxxxxxxxxxxxx>; Leon Romanovsky <leon@xxxxxxxxxx>;
>> linux-rdma@xxxxxxxxxxxxxxx
>> Subject: [EXTERNAL] Re: [RFC PATCH 05/19] RMDA/siw: Convert to use
>> vm_account
>> 
>> On Tue, Jan 24, 2023 at 04:42:34PM +1100, Alistair Popple wrote:
>> 
>> > @@ -385,20 +382,16 @@ struct siw_umem *siw_umem_get(u64 start, u64 len,
>> bool writable)
>> >  	if (!umem)
>> >  		return ERR_PTR(-ENOMEM);
>> >
>> > -	mm_s = current->mm;
>> > -	umem->owning_mm = mm_s;
>> >  	umem->writable = writable;
>> >
>> > -	mmgrab(mm_s);
>> > +	vm_account_init_current(&umem->vm_account);
>> >
>> >  	if (writable)
>> >  		foll_flags |= FOLL_WRITE;
>> >
>> > -	mmap_read_lock(mm_s);
>> > +	mmap_read_lock(current->mm);
>> >
>> > -	mlock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
>> > -
>> > -	if (num_pages + atomic64_read(&mm_s->pinned_vm) > mlock_limit) {
>> > +	if (vm_account_pinned(&umem->vm_account, num_pages)) {
>> >  		rv = -ENOMEM;
>> >  		goto out_sem_up;
>> >  	}
>> > @@ -429,7 +422,6 @@ struct siw_umem *siw_umem_get(u64 start, u64 len,
>> bool writable)
>> >  				goto out_sem_up;
>> >
>> >  			umem->num_pages += rv;
>> > -			atomic64_add(rv, &mm_s->pinned_vm);
>> 
>> Also fixes the race bug
>
> But introduces another one. In that loop, umem->num_pages keeps the
> number of pages currently pinned, not the target number. The current
> patch uses that umem->num_pages to call vm_unaccount_pinned() in
> siw_umem_release(). Bailing out before all pages are pinned would
> mess up that accounting during release. Maybe introduce another
> parameter to siw_umem_release(), or better have another umem member
> 'umem->num_pages_accounted' for correct accounting during release.

Yes, I see the problem thanks for pointing it out. Will fix for the next
version.

> Bernard.
>> 
>> Jason