Message-ID: <d1023695-200f-0673-0491-899e37f1c002@redhat.com>
Date:   Tue, 27 Jun 2023 23:06:04 +0200
From:   Danilo Krummrich <dakr@...hat.com>
To:     "Liam R. Howlett" <Liam.Howlett@...cle.com>
Cc:     Matthew Wilcox <willy@...radead.org>,
        Peng Zhang <perlyzhang@...il.com>,
        maple-tree@...ts.infradead.org, linux-mm@...ck.org,
        Andrew Morton <akpm@...ux-foundation.org>,
        linux-kernel@...r.kernel.org,
        Peng Zhang <zhangpeng.00@...edance.com>,
        David Airlie <airlied@...hat.com>,
        Boris Brezillon <boris.brezillon@...labora.com>
Subject: Re: [PATCH v2 14/16] maple_tree: Refine mas_preallocate() node
 calculations

On 6/27/23 18:11, Liam R. Howlett wrote:
> * Danilo Krummrich <dakr@...hat.com> [230627 10:58]:
>> Hi Liam,
>>
>> On 6/27/23 03:58, Liam R. Howlett wrote:
>>> * Danilo Krummrich <dakr@...hat.com> [230626 14:37]:
>>>> On 6/26/23 16:52, Matthew Wilcox wrote:
>>>>> On Mon, Jun 26, 2023 at 04:27:54PM +0200, Danilo Krummrich wrote:
>>>>>> On 6/26/23 15:19, Matthew Wilcox wrote:
>>>>>>> On Mon, Jun 26, 2023 at 02:38:06AM +0200, Danilo Krummrich wrote:
>>>>>>>> On the other hand, unless I miss something (and if so, please let me know),
>>>>>>>> something is bogus with the API then.
>>>>>>>>
>>>>>>>> While the documentation of the Advanced API of the maple tree explicitly
>>>>>>>> claims that the user of the API is responsible for locking, this should be
>>>>>>>> limited to the bounds set by the maple tree implementation. Which means, the
>>>>>>>> user must decide for either the internal (spin-) lock or an external lock
>>>>>>>> (which possibly goes away in the future) and acquire and release it
>>>>>>>> according to the rules maple tree enforces through lockdep checks.
>>>>>>>>
>>>>>>>> Let's say one picks the internal lock. How is one supposed to ensure the
>>>>>>>> tree isn't modified using the internal lock with mas_preallocate()?
>>>>>>>>
>>>>>>>> Besides that, I think the documentation should definitely mention this
>>>>>>>> limitation and give some guidance for the locking.
>>>>>>>>
>>>>>>>> Currently, from an API perspective, I can't see how anyone not familiar with
>>>>>>>> the implementation details would be able to recognize this limitation.
>>>>>>>>
>>>>>>>> In terms of the GPUVA manager, unfortunately, it seems like I need to drop
>>>>>>>> the maple tree and go back to using a rb-tree, since it seems there is no
>>>>>>>> sane way doing a worst-case pre-allocation that does not suffer from this
>>>>>>>> limitation.
>>>>>>>
>>>>>>> I haven't been paying much attention here (too many other things going
>>>>>>> on), but something's wrong.
>>>>>>>
>>>>>>> First, you shouldn't need to preallocate.  Preallocation is only there
>>>>>>
>>>>>> Unfortunately, I think we really have a case where we have to. Typically GPU
>>>>>> mappings are created in a dma-fence signalling critical path and that is
>>>>>> where such mappings need to be added to the maple tree. Hence, we can't do
>>>>>> any sleeping allocations there.
>>>>>
>>>>> OK, so there are various ways to hadle this, depending on what's
>>>>> appropriate for your case.
>>>>>
>>>>> The simplest is to use GFP_ATOMIC.  Essentially, you're saying to the MM
>>>>> layer "This is too hard, let me tap into the emergency reserves".  It's
>>>>> mildly frowned upon, so let's see if we can do better.
>>>>>
>>>>> If you know where the allocation needs to be stored, but want it to act as
>>>>> NULL until the time is right, you can store a ZERO entry.  That will read
>>>>> as NULL until you store to it.  A pure overwriting store will not cause
>>>>> any memory allocation since all the implementation has to do is change
>>>>> a pointer.  The XArray wraps this up nicely behind an xa_reserve() API.
>>>>> As you're discovering, the Maple Tree API isn't fully baked yet.
>>>>>
>>>>
>>>> Unfortunately, GFP_ATOMIC seems the be the only option. I think storing
>>>> entries in advance would not work. Typically userspace submits a job to the
>>>> kernel issuing one or multiple requests to map and unmap memory in an ioctl.
>>>> Such a job is then put into a queue and processed asynchronously in a
>>>> dma-fence signalling critical section. Hence, at the we'd store entries in
>>>> advance we could have an arbitrary amount of pending jobs potentially still
>>>> messing with the same address space region.
>>>
>>> What I think you are saying is that you have a number of requests
>>> flooding in, which may overwrite the same areas, but are queued up to be
>>> written after they are queued.  These operations look to be kept in
>>> order according to the code in nouveau_job_submit[1].  Is this correct?
>>
>> That's all correct.
>>
>> (Although Nouveau isn't a good example in this case. Some aspects of it do
>> and some aspects of it do not apply to the problem we're discussing here.)
>>
>>>
>>> So then, your issue isn't that you don't know where they will land, but
>>> don't know if the area that you reserved is already split into other
>>> areas?  For instance, before the range 5-10 is backed by whatever
>>> happens in the fence, it may have already become 5-6 & 8-10 by something
>>> that came after (from userspace) but hasn't been processed by the
>>> kernel that will live at 7?  So you can't write 5-10 right away because
>>> you can't be sure 5-10 is going to exist once you reach the kernel fence
>>> code that stores the entry?
>>>
>>> Is my understanding of your issue correct?
>>
>> Yes, it is.
>>
>> However, the problem already starts while trying to reserve an area. In
>> order to satisfy a user request, such a request is broken down into
>> operations such as unmap mappings which are in the way entirely, remap
>> mappings which intersect with the requested mapping and finally map the
>> requested mapping. The execution of such a sequence must appear atomic and
>> hence be locked accordingly. When trying to reserve an area we'd need to
>> take that lock. But since this lock would be used in the dma-fence
>> signalling critical path as well we'd not be allowed to do sleeping
>> allocations while holding this lock.
>>
>> Potentially, this could be solved with a retry loop though. Drop the lock
>> while allocating, take it again and check whether we still got enough nodes
>> allocated. Analogous to what the maple tree does in mas_store_gfp(), I
>> guess.
>>
>>>
>>> Oh, and I guess the queued requests would have to remain ordered between
>>> threads or whatever is on the other side?  I mean, you can't have two
>>> threads firing different things into the kernel at the same region
>>> because I would think the results would be unpredictable?
>>
>> Once a job is queued up in the kernel they remain ordered.
>>
>> However, user threads could concurrently push jobs to the kernel altering
>> the same region of the address space - it just would not make any sense for
>> userspace to do that.
>>
>> In general userspace is responsible for the semantics of the address space.
>> The kernel basically just takes any (valid) request and make it happen. It
>> also assures waiting and signalling of fences which might be bound to
>> certain jobs and obviously keeps track of the VA space to be able to clean
>> things up once a client disappears.
>>
>>>
>>> Can these overlapping entries partially overlap one region and another?
>>> That is, can you have three in-flight writes that does something like:
>>> store 1-10, store 10-20, store 5-15?
>>
>> Absolutely, yes.
>>
>>>
>>> How stable of an output is needed?  Does each kernel write need to be
>>> 100% correct or is there a point where the userspace updates stop and
>>> only then it is needed to be stable?
>>
>> It needs to be 100% correct all the time. The reason is that, as mentioned
>> above, every job can carry in- and out-fences, such that userspace can order
>> these jobs against the execution of shaders.
> 
> But each job is split into parts, so the fences surround these groups of
> operations?

Yes, each job can have multiple requests to map or unmap something and 
each of them gets broken down into operations to make them happen. The 
fences are per job.

> 
> Since ordering is kept, you must reach a point before entering the
> fences which could call the mas_preallocate() to ensure enough nodes
> exist to install the new mapping, and then no other operations will be
> happening. I guess what you are saying is each fence has more than one
> tree operation?
> 

I guess you assume that in the asynchronous path, where jobs are fetched 
from the queue for execution, there must be a point of time before we 
enter the fence signalling critical path. This is not the case.

The fence signalling critical path is entered once the corresponding 
out-fence is published to userspace and hence becomes visible to 
userspace. This happens when the job submit ioctl() (which is where the 
job is queued up for execution) returns. Hence, all jobs in the queue 
potentially entered their fence signalling critical path already before 
they could have been fetched from the queue.

The job submit ioctl() is where we would need to call mas_preallocate(), 
but this is also where we could have concurrent modifications to the 
tree from previously submitted jobs that have been fetched from the 
queue for asynchronous execution.

> As long as you are not mapping more than a range, then this should be possible
> in a single write and thus a single preallocation.  You can do this by
> not actually writing unmaps/remaps to the tree within the fence.  Once
> the out-fence is reached, then the operation looks atomic.

As mentioned above there can be an arbitrary amount of map and unmap 
requests per job.

Also, we can have cases where we, for instance, have a mapping A[0,10] 
either already in the tree (or pending in the job queue) and a user 
requests B[4,7].

The expected result would be: A'[0,4], B[4,7] and A''[7,10]. Hence, one 
job can cause multiple writes per single map request even.

At the time the job asking to map B is submitted, A might not even be in 
the tree yet.

> 
> Reading your patch, it is not clear this is accurate for VM_BIND of
> asynchronous syncobjs.  Is the fence spanning multiple syncobjs with
> various ranges to map?  Or are these things the split-up tasks of
> unmap/remap, etc that will eventually boil down to what appears to be a
> single write?
> 

The VM_BIND ioctl() is the ioctl() mentioned above to submit (bind) 
jobs. Userspace can pass syncobjs together with a job, which can be both 
syncobj which contain fences to wait for before executing the job and 
syncobjs to install the job's fence to for userspace to wait for them or 
pass them into another kernel interface to synchronize them against 
something else.

Otherwise, as mentioned above, each job can have multiple requests to 
map or unmap something and each of them gets broken down into operations 
to make such a request happen, which themselves can be re-maps, unmaps 
and maps. For instance, an unmap request for a given range, gets broken 
down into re-maps of mappings it intersects and unmaps of mappings it 
entirely spans over.

Thanks,
Danilo

>>
>> This is also why there could be jobs queued up, where all of them apply
>> changes to the same region within the VA space, since there might be shader
>> executions (or just memory copies) ordered right between them.
>>
>> - Danilo
>>
>>>
>>>>
>>>> So, the only way to go seems to be to use mas_store_gfp() with GFP_ATOMIC
>>>> directly in the fence signalling critical path. I guess mas_store_gfp() does
>>>> not BUG_ON() if it can't get atomic pages?
>>>>
>>>> Also, I just saw that the tree is limited in it's height (MAPLE_HEIGHT_MAX).
>>>> Do you think it could be a sane alternative to pre-allocate with
>>>> MAPLE_HEIGHT_MAX rather than to rely on atomic pages? Or maybe a compromise
>>>> of pre-allocating just a couple of nodes and then rely on atomic pages for
>>>> the rest?
>>>>
>>>> FYI, we're talking about a magnitude of hundreds of thousands of entries to
>>>> be stored in the tree.
>>>>
>>>
>>> Since you are not tracking gaps, you will get 16 entries per node.  The
>>> maximum height is 31, so that would be 16^31, assuming a gap between
>>> each entry (the worst case), you can cut that in 1/2.  To assure you can
>>> successfully allocate storage for a new entries, you'd need to allocate
>>> 30 * 3 + 1, or 91 nodes, which is 6 pages.  That'll be highly wasteful
>>> as almost all of these would be freed, and sometimes all of them.
>>>
>>> You estimate less than 1M entries, that would never go over 6 levels (8.3M
>>> entries with the worst-case).  5 levels would get you 500K in the worst
>>> case, but realistically you'll be in the 5 levels almost always.  So,
>>> 5*3+1 = 17 nodes, or 2 pages (1 node over 1 page).. assuming 4k pages.
>>>
>>> [1] https://lore.kernel.org/linux-mm/20230620004217.4700-8-dakr@redhat.com/T/#Z2e.:..:20230620004217.4700-4-dakr::40redhat.com:1drivers:gpu:drm:drm_gem.c
>>>
>>
> 

Powered by blists - more mailing lists