On Mon, Aug 19, 2024 at 01:38:56PM +0200, Thomas Hellström wrote:Hi, Christian, On Mon, 2024-08-19 at 13:03 +0200, Christian König wrote:Am 06.08.24 um 10:29 schrieb Thomas Hellström:Hi, Christian. On Thu, 2024-07-11 at 14:01 +0200, Christian König wrote:Am 10.07.24 um 20:19 schrieb Matthew Brost:On Wed, Jul 10, 2024 at 02:42:58PM +0200, Christian König wrote:That is something drivers really shouldn't mess with.Thomas uses this in Xe to implement a shrinker [1]. Seems to need to remain available for drivers.No, that is exactly what I try to prevent with that. This is an internally thing of TTM and drivers should never use it directly.That driver-facing LRU walker is a direct response/solution to this comment that you made in the first shrinker series: https://lore.kernel.org/linux-mm/b7491378-defd-4f1c-31e2-29e4c77e2d67@xxxxxxx/T/#ma918844aa8a6efe8768fdcda0c6590d5c93850c9Ah, yeah that was about how we are should be avoiding middle layer design. But a function which returns the next candidate for eviction and a function which calls a callback for eviction is exactly the opposite.That is also mentioned in the cover letter of the recent shrinker series, and this walker has been around in that shrinker series for more than half a year now so if you think it's not the correct driver facing API IMHO that should be addressed by a review comment in that series rather than in posting a conflicting patch?I actually outlined that in the review comments for the patch series. E.g. a walker function with a callback is basically a middle layer. What outlined in the link above is that a function which returns the next eviction candidate is a better approach than a callback.So assuming that we still want the driver to register the shrinker, IMO that helper abstracts away all the nasty locking and pitfalls for a driver-registered shrinker, and is similar in structure to for example the pagewalk helper in mm/pagewalk.c. An alternative that could be tried as a driver-facing API is to provide a for_each_bo_in_lru_lock() macro where the driver open-codes "process_bo()" inside the for loop but I tried this and found it quite fragile since the driver might exit the loop without performing the necessary cleanup.The point is that the shrinker should *never* need to have context. E.g. a walker which allows going over multiple BOs for eviction is exactly the wrong approach for that. The shrinker should evict always only exactly one BO and the next invocation of a shrinker should not depend on the result of the previous one. Or am I missing something vital here?A couple of things, 1) I'd like to think of the middle-layer vs helper like the helper has its own ops, and can be used optionally from the driver. IIRC, the atomic modesetting / pageflip ops are implemented in exactly this way. Sometimes a certain loop operation can't be easily or at least robustly implemented using a for_each_.. approach. Like for example mm/pagewalk.c. In this shrinking case I think it's probably possible using the scoped_guard() in cleanup.h. This way we could get rid of this middle layer discussion by turning the interface inside-out: for_each_bo_on_lru_locked(xxx) driver_shrink(); But I do think the currently suggested approach is less fragile and prone to driver error. FWIW in addition to the above examples, also drm_gem_lru_scan works like this.a iteration state structure (like drm_connector_iter) plus then a macro for the common case that uses cleanup.h should get the job done.
Yeah, completely agree. It basically boils down to which one needs to pack a state bag.
In a mid-layer design it's the driver or the caller of functions, e.g. the much hated init callback in the DRM layer was a perfect example of that.
In a non mid-layer approach it's the framework or the called function, examples are how the fence iterators in the dma_resv or the drm_connector, plane etc.. work.
One big difference between the two approach is who and where things like preparation and cleanup are done, e.g. who takes locks for example.
2) The shrinkers suggest to shrink a number of pages, not a single bo, again drm_gem_lru_scan works like this. i915 works like this. I think we should align with those.Yeah that's how shrinkers work, so if we demidlayer then it realls needs to be a loop in the driver, not a "here's the next bo to nuke" I think.
Hui? Well that's not how I understand shrinkers.
The shrinker gives you the maximum number of objects to scan and not how many pages to free. In return you give the actual number of objects to scanned and pages freed.
As far as I know the number of objects are in the sense of SLUBs or rather different LRU lists.
So for BO shrinking we should probably only free or rather unpin a single BO per callback otherwise we mess up the fairness between shrinkers in the MM layer.
3) Even if we had a function to obtain the driver to shrink, the driver needs to have its say about the suitability for shrinking, so a callback is needed anyway (eviction_valuable). In addition, if shrinking fails for some reason, what would stop that function to return the same bo, again and again just like the problem we previously had in TTM?Yeah I think if consensus moves back to drivers not looking at ttm lru directly, then that entire shrinker looping needs to be as some kind of midlayer in ttm itself. Otherwise I don't think it works, so agreeing with Thomas here.
Well, the idea was to get rid of callbacks like eviction_valuable.
So instead of TTM calling the driver asking if a BO is valuable to evict TTM provides the functionality to iterate over the possible candidates and the driver then says "oh yeah evict that one".
We probably can't do that fully, e.g. in case of contention on a resource we still need a callback to inform the driver and we also need a copy callback. But that's basically it.
So basically all the restartable LRU work was motivated by this use- case in mind, so making that private I must say comes as a pretty major surprise. I could have a look at the for_each_bo_on_lru_locked(xxx) driver_shrink(); approach, but having TTM just blindly return a single bo without context will not work IMO.Another thing to keep in mind is that at least experience from really resource-starved igpu platforms says that cpu consumption for shrinking matters. So really need to not toss list walk state, and also at least from I think msm experience and maybe also i915 (i915-gem's a bit ... too complex to really understand it anymore) is that parallelism matters too. Eventually under memory pressures multiple cpu cores just aboslutely hammer the shrinkers, so being stuck on locks is no good.
Yeah that's basically the reason why I looked into this before as well.
Thomas implementation is actually pretty good at that because it only has the LRU spinlock as contented object and multiple CPUs can walk the LRU at the same time otherwise.
Regards,
Christian.
But maybe let's get this off the ground first. -Sima
