linux-kernel - Re: [PATCH v19 22/30] drm/shmem-helper: Add common memory shrinker

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date: Mon, 29 Jan 2024 09:55:11 +0100
From: Boris Brezillon <boris.brezillon@...labora.com>
To: Dmitry Osipenko <dmitry.osipenko@...labora.com>
Cc: David Airlie <airlied@...il.com>, Gerd Hoffmann <kraxel@...hat.com>,
 Gurchetan Singh <gurchetansingh@...omium.org>, Chia-I Wu
 <olvaffe@...il.com>, Daniel Vetter <daniel@...ll.ch>, Maarten Lankhorst
 <maarten.lankhorst@...ux.intel.com>, Maxime Ripard <mripard@...nel.org>,
 Thomas Zimmermann <tzimmermann@...e.de>, Christian König
 <christian.koenig@....com>, Qiang Yu <yuq825@...il.com>, Steven Price
 <steven.price@....com>, Emma Anholt <emma@...olt.net>, Melissa Wen
 <mwen@...lia.com>, dri-devel@...ts.freedesktop.org,
 linux-kernel@...r.kernel.org, kernel@...labora.com,
 virtualization@...ts.linux-foundation.org
Subject: Re: [PATCH v19 22/30] drm/shmem-helper: Add common memory shrinker

On Mon, 29 Jan 2024 09:16:04 +0300
Dmitry Osipenko <dmitry.osipenko@...labora.com> wrote:

> On 1/26/24 21:12, Boris Brezillon wrote:
> > On Fri, 26 Jan 2024 19:27:49 +0300
> > Dmitry Osipenko <dmitry.osipenko@...labora.com> wrote:
> >   
> >> On 1/26/24 12:55, Boris Brezillon wrote:  
> >>> On Fri, 26 Jan 2024 00:56:47 +0300
> >>> Dmitry Osipenko <dmitry.osipenko@...labora.com> wrote:
> >>>     
> >>>> On 1/25/24 13:19, Boris Brezillon wrote:    
> >>>>> On Fri,  5 Jan 2024 21:46:16 +0300
> >>>>> Dmitry Osipenko <dmitry.osipenko@...labora.com> wrote:
> >>>>>       
> >>>>>> +static bool drm_gem_shmem_is_evictable(struct drm_gem_shmem_object *shmem)
> >>>>>> +{
> >>>>>> +	return (shmem->madv >= 0) && shmem->base.funcs->evict &&
> >>>>>> +		refcount_read(&shmem->pages_use_count) &&
> >>>>>> +		!refcount_read(&shmem->pages_pin_count) &&
> >>>>>> +		!shmem->base.dma_buf && !shmem->base.import_attach &&
> >>>>>> +		!shmem->evicted;      
> >>>>>
> >>>>> Are we missing
> >>>>>
> >>>>>                 && dma_resv_test_signaled(shmem->base.resv,
> >>>>> 					  DMA_RESV_USAGE_BOOKKEEP)
> >>>>>
> >>>>> to make sure the GPU is done using the BO?
> >>>>> The same applies to drm_gem_shmem_is_purgeable() BTW.
> >>>>>
> >>>>> If you don't want to do this test here, we need a way to let drivers
> >>>>> provide a custom is_{evictable,purgeable}() test.
> >>>>>
> >>>>> I guess we should also expose drm_gem_shmem_shrinker_update_lru_locked()
> >>>>> to let drivers move the GEMs that were used most recently (those
> >>>>> referenced by a GPU job) at the end of the evictable LRU.      
> >>>>
> >>>> We have the signaled-check in the common drm_gem_evict() helper:
> >>>>
> >>>> https://elixir.bootlin.com/linux/v6.8-rc1/source/drivers/gpu/drm/drm_gem.c#L1496    
> >>>
> >>> Ah, indeed. I'll need DMA_RESV_USAGE_BOOKKEEP instead of
> >>> DMA_RESV_USAGE_READ in panthor, but I can add it in the driver specific    
> >>> ->evict() hook (though that means calling dma_resv_test_signaled()    
> >>> twice, which is not great, oh well).    
> >>
> >> Maybe we should change drm_gem_evict() to use BOOKKEEP. The
> >> test_signaled(BOOKKEEP) should be a "stronger" check than
> >> test_signaled(READ)?  
> > 
> > It is, just wondering if some users have a good reason to want
> > READ here.
> >   
> >>  
> >>> The problem about the evictable LRU remains though: we need a way to let
> >>> drivers put their BOs at the end of the list when the BO has been used
> >>> by the GPU, don't we?    
> >>
> >> If BO is use, then it won't be evicted, while idling BOs will be
> >> evicted. Hence, the used BOs will be naturally moved down the LRU list
> >> each time shrinker is invoked.
> >>  
> > 
> > That only do the trick if the BOs being used most often are busy when
> > the shrinker kicks in though. Let's take this scenario:
> > 
> > 
> > BO 1					BO 2					shinker
> > 
> > 					busy
> > 					idle (first-pos-in-evictable-LRU)
> > 
> > busy
> > idle (second-pos-in-evictable-LRU)
> > 
> > 					busy
> > 					idle
> > 
> > 					busy
> > 					idle
> > 
> > 					busy
> > 					idle
> > 
> > 										find a BO to evict
> > 										pick BO 2
> > 
> > 					busy (swapin)
> > 					idle
> > 
> > If the LRU had been updated at each busy event, BO 1 should have
> > been picked for eviction. But we evicted the BO that was first
> > recorded idle instead of the one that was least recently
> > recorded busy.  
> 
> You have to swapin(BO) every time BO goes to busy state, and swapin does drm_gem_lru_move_tail(BO). Hence, each time BO goes idle->busy, it's moved down the LRU list.

Ah, that's the bit I was missing. It makes sense now. I guess that's
good enough for now, we can sort out the BOOKKEEP vs READ in a
follow-up series.

Reviewed-by: Boris Brezillon <boris.brezillon@...labora.com>