<!DOCTYPE html><html><head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> </head> <body> Am 20.08.24 um 18:00 schrieb Thomas Hellström: <blockquote type="cite" cite="mid:bb0a31ea3d82ee370873ca5f1c66ec4eeafabffe.camel@linux.intel.com"> <blockquote type="cite"> <pre class="moz-quote-pre" wrap="">Or why exactly should shrinking fail? </pre> </blockquote> <pre class="moz-quote-pre" wrap="">A common example would be not having runtime pm and the particular bo needs it to unbind, we want to try the next bo. Example: i915 GGTT bound bos and Lunar Lake PL_TT bos.</pre> </blockquote> WHAT? So you basically block shrinking BOs because you can't unbind them because the device is powered down? I would say that this is a serious NO-GO. It basically means that powered down devices can lock down system memory for undefined amount of time. In other words an application can allocate memory, map it into GGTT and then suspend or even get killed and we are not able to recover the memory because there is no activity on the GPU any more? That really sounds like a bug in the driver design to me. <blockquote type="cite" cite="mid:bb0a31ea3d82ee370873ca5f1c66ec4eeafabffe.camel@linux.intel.com"> <pre class="moz-quote-pre" wrap="">And again, all other drm bo shrinkers do this. We just want to do the same.</pre> </blockquote> Do you have pointers? <blockquote type="cite" cite="mid:bb0a31ea3d82ee370873ca5f1c66ec4eeafabffe.camel@linux.intel.com"> <blockquote type="cite"> <blockquote type="cite"> <pre class="moz-quote-pre" wrap="">If we bump LRU we could end up with infinite loops. So IMO we need to be able to loop. I don't really care wether we do this as an explicit loop or whether we use the LRU walker, but I think from a maintainability point-of-view it is better to keep LRU walking in a single place. If we return an unlocked object, we'd need to refcount and drop the lru lock, but maybe that's not a bad thing. But what's the main drawback of exporting the existing helper. </pre> </blockquote> <pre class="moz-quote-pre" wrap="">Well that we re-creates exactly the mid-layer mess I worked so hard to remove from TTM. </pre> </blockquote> <pre class="moz-quote-pre" wrap="">It doesn't IMO. I agree the first attempt did. This affects only the LRU iteration itself and I'm even fine to get rid of the callback using a for_ macro.</pre> </blockquote> Well, I mean using a for_each approach is objectively better than having a callback and a state bag. But the fundamental question is if drivers are allowed to reject shrinking. And I think the answer is no, they need to be designed in a way where shrinking is always possible. What can be that we can't get the necessary locks to evict and object (because it's about to be used etc...), but that are the per-requisites TTM should be checking. <blockquote type="cite" cite="mid:bb0a31ea3d82ee370873ca5f1c66ec4eeafabffe.camel@linux.intel.com"> <blockquote type="cite"> <blockquote type="cite"> <pre class="moz-quote-pre" wrap="">In any case, I don't think TTM should enforce a different way of shrinking by the means of a severely restricted helper? </pre> </blockquote> <pre class="moz-quote-pre" wrap="">Well, as far as I can see that is exactly what TTM should do. I mean the main advantage to make a common component is to enforce correct behavior. </pre> </blockquote> <pre class="moz-quote-pre" wrap="">But if all other drivers don't agree this as correct behavior and instead want to keep behavior that is proven to work, that's a dead end.</pre> </blockquote> Well no, even if all drivers agree to (for example) drop security precautions it's still not something acceptable. And same thing here, if we block shrinking because drivers think they want their runtime PM implemented in a certain way then upstream needs to block this and push back. As far as I can see it's mandatory to have shrinkers not depend on runtime PM, cause otherwise you run into resources handling which depends on the well behavior of userspace and that in turn in something we can't allow. Regards, Christian. <blockquote type="cite" cite="mid:bb0a31ea3d82ee370873ca5f1c66ec4eeafabffe.camel@linux.intel.com"> <pre class="moz-quote-pre" wrap=""> /Thomas </pre> <blockquote type="cite"> <pre class="moz-quote-pre" wrap="">Regards, Christian. </pre> <blockquote type="cite"> <pre class="moz-quote-pre" wrap="">/Thomas </pre> </blockquote> </blockquote> </blockquote> </body> </html>