<div dir="ltr"><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Sun, Dec 15, 2019 at 10:24 PM Niranjan Vishwanathapura <<a href="mailto:niranjana.vishwanathapura@intel.com">niranjana.vishwanathapura@intel.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">On Sat, Dec 14, 2019 at 10:31:37AM +0000, Chris Wilson wrote:<br>
>Quoting Jason Ekstrand (2019-12-14 00:36:19)<br>
>> On Fri, Dec 13, 2019 at 5:24 PM Niranjan Vishwanathapura <<br>
>> <a href="mailto:niranjana.vishwanathapura@intel.com" target="_blank">niranjana.vishwanathapura@intel.com</a>> wrote:<br>
>><br>
>> On Fri, Dec 13, 2019 at 04:58:42PM -0600, Jason Ekstrand wrote:<br>
>> ><br>
>> > +/**<br>
>> > + * struct drm_i915_gem_vm_bind<br>
>> > + *<br>
>> > + * Bind an object in a vm's page table.<br>
>> ><br>
>> > First off, this is something I've wanted for a while for Vulkan, it's<br>
>> just<br>
>> > never made its way high enough up the priority list. However, it's<br>
>> going<br>
>> > to have to come one way or another soon. I'm glad to see kernel API<br>
>> for<br>
>> > this being proposed.<br>
>> > I do, however, have a few high-level comments/questions about the API:<br>
>> > 1. In order to be useful for sparse memory support, the API has to go<br>
>> the<br>
>> > other way around so that it binds a VA range to a range within the BO. <br>
>> It<br>
>> > also needs to be able to handle overlapping where two different VA<br>
>> ranges<br>
>> > may map to the same underlying bytes in the BO. This likely means that<br>
>> > unbind needs to also take a VA range and only unbind that range.<br>
>> > 2. If this is going to be useful for managing GL's address space where<br>
>> we<br>
>> > have lots of BOs, we probably want it to take a list of ranges so we<br>
>> > aren't making one ioctl for each thing we want to bind.<br>
>><br>
>> Hi Jason,<br>
>><br>
>> Yah, some of these requirements came up.<br>
>><br>
>> <br>
>> Yes, I have raised them every single time an API like this has come across my<br>
>> e-mail inbox for years and they continue to get ignored. Why are we landing an<br>
>> API that we know isn't the API we want especially when it's pretty obvious<br>
>> roughly what the API we want is? It may be less time in the short term, but<br>
>> long-term it means two ioctls and two implementations in i915, IGT tests for<br>
>> both code paths, and code in all UMDs to call one or the other depending on<br>
>> what kernel you're running on, and we have to maintain all that code going<br>
>> forward forever. Sure, that's a price we pay today for a variety of things but<br>
>> that's because they all seemed like the right thing at the time. Landing the<br>
>> wrong API when we know it's the wrong API seems foolish.<br>
><br>
>Exactly. This is not even close to the uAPI we need. Reposting an RFC<br>
>without taking in the concerns last time (or the time before that, or<br>
>the time before that...) suggests that you aren't really requesting for<br>
>comments at all.<br>
<br>
Thanks Jason for detailed exlanation.<br>
Chris, all comments and guidance are much appreciated :)<br>
<br>
I haven't looked in detail, but my concern is that implementing<br>
partial object binding (offset, lenght) from vma down to [un]binding<br>
in ppgtt might be a lot of work to include in this SVM patch series.<br>
I believe we need the partial object binding in non-SVM scenario<br>
as well?<br></blockquote><div><br></div><div>Yes, the Vulkan APIs require both partial binding and aliasing.</div><div><br></div><div>It may be worth pointing out that we're already doing some of this stuff today, although in a rather backwards way. Our non-softpin model for Vulkan uses a memfd which we then map in userspace and turn into a BO via userptr. Due to the way we handle threading in the driver, we end up with multiple BOs pointing to the same overlapping range in the memfd and hence the same pages. That doesn't mean that everything in the path is already set up for what you need but the VA -> page mappings should be. Also, avoiding these kinds of shinanigans is exactly why we want a "real" kernel API for this. :-)<br></div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
Ok, let me change the interface as below.<br>
<br>
struct drm_i915_gem_vm_bind_va<br>
{<br>
/** VA start to bind **/<br>
__u64 start;<br>
<br>
/** Offset in Object to bind for I915_GEM_VM_BIND_SVM_OBJ type **/<br>
__u64 offset;<br>
<br>
/** VA length to [un]bind **/<br>
__u64 length;<br>
<br>
/** Type of memory to [un]bind **/<br>
__u32 type;<br>
#define I915_GEM_VM_BIND_SVM_OBJ 0<br>
#define I915_GEM_VM_BIND_SVM_BUFFER 1<br>
<br>
/** Object handle to [un]bind for I915_GEM_VM_BIND_SVM_OBJ type **/<br>
__u32 handle;<br>
<br>
/** Flags **/<br>
__u32 flags;<br>
#define I915_GEM_VM_BIND_UNBIND (1 << 0)<br>
#define I915_GEM_VM_BIND_READONLY (1 << 1)<br>
}<br>
<br>
struct drm_i915_gem_vm_bind {<br>
/** vm to [un]bind **/<br>
__u32 vm_id;<br>
<br>
/** number of VAs to bind **/<br>
__u32 num_vas;<br>
<br>
/** Array of VAs to bind **/<br>
struct drm_i915_gem_vm_bind_va *bind_vas;<br>
<br>
/** User extensions **/<br>
__u64 extensions;<br>
};<br>
<br>
When synchronization control is added as extension, it applies to all VAs in the array.<br>
Does this looks good?<br></blockquote><div><br></div><div>Yes, I think that header looks good. It's probably fine if synchronization comes later.</div><div><br></div><div>I have two more questions (more for my own education than anything):</div><div><br></div><div> 1. What is the difference between a SVM object and a buffer?</div><div><br></div><div> 2. I see a vm_id but there is no context. What (if any) are the synchronization guarantees between the VM_BIND ioctl and EXECBUF? If I VM_BIND followed by EXECBUF is it guaranteed to happen in that order? What if I EXECBUF and then VM_BIND to unbind something? If I VM_BIND while an execbuf is happening but I have some way of delaying the GPU work from the CPU and I unblock it once the VM_BIND comes back, is that ok?</div><div><br></div><div>If those questions are answered by other patches, feel free to just point me at them instead of answering in detail here.</div><div><br></div><div>--Jason<br></div></div></div>