On Thu, 7 Dec 2023 10:23:55 +0000 Steven Price <steven.price@xxxxxxx> wrote: > On 07/12/2023 08:56, Boris Brezillon wrote: > > On Thu, 7 Dec 2023 09:12:43 +0100 > > Boris Brezillon <boris.brezillon@xxxxxxxxxxxxx> wrote: > > > >> On Wed, 6 Dec 2023 16:55:42 +0000 > >> Steven Price <steven.price@xxxxxxx> wrote: > >> > >>> On 04/12/2023 17:32, Boris Brezillon wrote: > >>>> The panthor driver is designed in a modular way, where each logical > >>>> block is dealing with a specific HW-block or software feature. In order > >>>> for those blocks to communicate with each other, we need a central > >>>> panthor_device collecting all the blocks, and exposing some common > >>>> features, like interrupt handling, power management, reset, ... > >>>> > >>>> This what this panthor_device logical block is about. > >>>> > >>>> v3: > >>>> - Add acks for the MIT+GPL2 relicensing > >>>> - Fix 32-bit support > >>>> - Shorten the sections protected by panthor_device::pm::mmio_lock to fix > >>>> lock ordering issues. > >>>> - Rename panthor_device::pm::lock into panthor_device::pm::mmio_lock to > >>>> better reflect what this lock is protecting > >>>> - Use dev_err_probe() > >>>> - Make sure we call drm_dev_exit() when something fails half-way in > >>>> panthor_device_reset_work() > >>>> - Replace CSF_GPU_LATEST_FLUSH_ID_DEFAULT with a constant '1' and a > >>>> comment to explain. Also remove setting the dummy flush ID on suspend. > >>>> - Remove drm_WARN_ON() in panthor_exception_name() > >>>> - Check pirq->suspended in panthor_xxx_irq_raw_handler() > >>>> > >>>> Signed-off-by: Boris Brezillon <boris.brezillon@xxxxxxxxxxxxx> > >>>> Signed-off-by: Steven Price <steven.price@xxxxxxx> > >>>> Acked-by: Steven Price <steven.price@xxxxxxx> # MIT+GPL2 relicensing,Arm > >>>> Acked-by: Grant Likely <grant.likely@xxxxxxxxxx> # MIT+GPL2 relicensing,Linaro > >>>> Acked-by: Boris Brezillon <boris.brezillon@xxxxxxxxxxxxx> # MIT+GPL2 relicensing,Collabora > >>> > >>> We still have the "FIXME: this is racy" > >> > >> Yeah, I still didn't find a proper solution for that. > > > > This [1] should fix the race condition in the unplug path. > > > > [1]https://gitlab.freedesktop.org/panfrost/linux/-/commit/b79b28069e524ae7fea22a9a158b870eab2d5f9a > > Looks like it should do the job. I'm surprised that we're the only ones > to face this though. Most drivers just have one path where they call drm_dev_unplug(): the device removal callback, which is only called once per device. The only exception where I see more than one occurrence are the amdgpu and powervr drivers. I guess amdgpu has some tricks to serialize _unplug operations, and powervr is probably buggy as you pointed out. > > Looking at the new imagination driver it appears there's a similar problem: > > pvr_device_lost() uses a boolean 'lost' to track multiple calls but that > boolean isn't protected by any specific lock (AFAICT). Indeed > pvr_device_lost() calls drm_dev_unplug() while in a drm_dev_enter() > critical section (see pvr_mmu_flush_exec()). If I'm not mistaken that's > the same problem we discussed and isn't allowed? It is, indeed. That means there's a deadlock when pvr_device_lost() is called from the MMU code. And I guess the race condition with concurrent pvr_device_lost() callers exists too (unless I missed something, and calls to pvr_device_lost() are serialized with another lock). > drm_dev_unplug() will > synchronise on the SRCU that drm_dev_enter() is holding. > > +CC: Frank, Donald, Matt from Imagination. > > Steve >