Re: [PATCH] vfio: remove useless judgement

[Steven Sistare]

On 6/28/2022 9:03 AM, Jason Gunthorpe wrote:
> On Tue, Jun 28, 2022 at 08:48:11AM -0400, Steven Sistare wrote:
>> For cpr, old qemu directly exec's new qemu, so task does not change.
>>
>> To support fork+exec, the ownership test needs to be deleted or modified.
>>
>> Pinned page accounting is another issue, as the parent counts pins in its
>> mm->locked_vm.  If the child unmaps, it cannot simply decrement its own
>> mm->locked_vm counter.
> 
> It is fine already:
> 
> 	mm = async ? get_task_mm(dma->task) : dma->task->mm;
> 	if (!mm)
> 		return -ESRCH; /* process exited */
> 
> 	ret = mmap_write_lock_killable(mm);
> 	if (!ret) {
> 		ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task,
> 					  dma->lock_cap);
> 
> Each 'dma' already stores a pointer to the mm that sourced it and only
> manipulates the counter in that mm. AFAICT 'current' is not used
> during unmap.
Ah yes, no problem then.
Limits become looser, though, as the child can pin an additional RLIMIT_MEMLOCK
of pages.  That is the natural consequence of mm->locked_vm being a per process limit, 
but probably not what the application wants.  Another argument for switching to 
user->locked_vm.

>> As you and I have discussed, the count is also wrong in the direct
>> exec model, because exec clears mm->locked_vm.
> 
> Really? Yikes, I thought exec would generate a new mm?

Yes, exec creates a new mm with locked_vm = 0.  The old locked_vm count is dropped
on the floor.  The existing dma points to the same task, but task->mm has changed,
and dma->task->mm->locked_vm is 0.  An unmap ioctl drives it negative.

I have prototyped a few possible fixes.  One changes vfio to use user->locked_vm.
Another changes to mm->pinned_vm and preserves it during exec.  A third preserves
mm->locked_vm across exec, but that is not practical, because mm->locked_vm mixes
vfio pins and mlocks.  The mlock component must be cleared during exec, and we don't 
have a separate count for it.

>> I am thinking vfio could count pins in struct user locked_vm to handle both 
>> models.  The user struct and its count would persist across direct exec,
>> and be shared by parent and child for fork+exec.  However, that does change
>> the RLIMIT_MEMLOCK value that applications must set, because the limit must
>> accommodate vfio plus other sub-systems that count in user->locked_vm, which
>> includes io_uring, skbuff, xdp, and perf.  Plus, the limit must accommodate all
>> processes of that user, not just a single process.
> 
> We discussed this, for iommufd we are currently planning to go this
> way and will See How it Goes.

Yes, I have followed that thread with interest.

- Steve