On Fri, Jan 24, 2020 at 06:39:26PM +0000, Chris Wilson wrote:
Quoting Thomas Hellström (VMware) (2020-01-24 13:37:47)
On 1/24/20 2:01 PM, Chris Wilson wrote:
Since drm_global_mutex is a true global mutex across devices, we don't want to acquire it unless absolutely necessary. For maintaining the device local open_count, we can use atomic operations on the counter itself, except when making the transition to/from 0. Here, we tackle the easy portion of delaying acquiring the drm_global_mutex for the final release by using atomic_dec_and_mutex_lock(), leaving the global serialisation across the device opens.
Signed-off-by: Chris Wilson chris@chris-wilson.co.uk Cc: Thomas Hellström (VMware) thomas_os@shipmail.org
For the series:
Reviewed-by: Thomas Hellström thellstrom@vmware.com
Now being opt-in, it is fairly limited in scope and will not randomly break others (touch wood) and the close() racing in BAT didn't throw anything up, so pushed to drm-misc-next. Thanks for the review and suggestions,
Yeah this version looks reasonable compared to the previous few (I'm catching up on dri-devel). I've looked at getting rid of the global_mutex, and all I have is a simple patch with a pile of notes. It's real nasty.
This one here is a neat trick that I missed, and I'm semi-convinced it's safe :-)
Next task is to suggest others might like to use it as well.
My idea for the opt-in was to look at whether ->load/->unload exists. And ofc not bother with any of this for DRIVER_LEGACY. So maybe next step would be to define a
drm_can_noglobal() { return !DRIVER_LEGACY && !->load && !->unload; }
and inline the close helper again and see what breaks? At least from what I've looked trying to duplicate paths and opt-in is going to be real tough on the open side of things. Best I've done thus far is minor pushing of the critical section. -Daniel