<html xmlns:v="urn:schemas-microsoft-com:vml" xmlns:o="urn:schemas-microsoft-com:office:office" xmlns:w="urn:schemas-microsoft-com:office:word" xmlns:m="http://schemas.microsoft.com/office/2004/12/omml" xmlns="http://www.w3.org/TR/REC-html40"> <head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> <meta name="Generator" content="Microsoft Word 15 (filtered medium)"> <style></style> </head> <body lang="EN-US" link="blue" vlink="purple"> <div class="WordSection1"> Hi Maarten: Thanks for the reply. 😊 Have your patches been merged into the drm-intel? (The new i915_gem_object_ggtt_pin() you mentioned in the email). Also, we might go on-demand pin recently in the suspend/resume support. In that case, I guess I need to review my patches again based on the new mechanism. Wrt your concern, my understanding is ww dance is for recording the related locks in a section. When a deadlock happens, we can rewind the locks recorded in a ww dance and rewind them. In the current code, the context is always pinned. So we don’t need to worry about the context as what you said. For preparing_shadow_wa_ctx, the gem object will be newly created whenever scan a execlist context and it will be destroyed when the context is scheduled-out. I think in that case, we need a separate ww dance.<o:p></o:p> Thanks, Zhi.<o:p></o:p> -----Original Message----- From: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Sent: Wednesday, September 30, 2020 2:55 PM To: Wang, Zhi A <zhi.a.wang@intel.com>; Joonas Lahtinen <joonas.lahtinen@linux.intel.com>; Nikula, Jani <jani.nikula@intel.com>; intel-gfx <intel-gfx@lists.freedesktop.org>; Yuan, Hang <hang.yuan@intel.com>; Lv, Zhiyuan <zhiyuan.lv@intel.com>; intel-gvt-dev <intel-gvt-dev@lists.freedesktop.org> Subject: Re: [Intel-gfx] [PULL] topic/gvt-ww-lock Hey, Op 22-09-2020 om 13:51 schreef Wang, Zhi A: > > Hi, > > Here's the patch which introduces GVT-g ww lock support against drm-intel-gt-next branch. > > Thanks > > -- > > The following changes since commit 4316b19dee27cc5cd34a95fdbc0a3a5237507701: > > drm/i915: Fix uninitialised variable in > intel_context_create_request. (2020-09-21 11:09:46 +0200) > > are available in the git repository at: > > <a href="https://github.com/intel/gvt-linux.git" target="_BLANK">https://github.com/intel/gvt-linux.git</a> tags/gvt-ww-lock-09-22-2020 > > for you to fetch changes up to e1b49ae155451aaa039eeb23fce808c71a038283: > > drm/i915/gvt: Introduce per object locking in GVT scheduler. > (2020-09-22 12:09:56 +0300) > > ---------------------------------------------------------------- > gvt-ww-lock-09-22-2020: > > - Introduce ww lock support into GVT-g. > > ---------------------------------------------------------------- > Zhi Wang (1): > drm/i915/gvt: Introduce per object locking in GVT scheduler. > > drivers/gpu/drm/i915/gvt/scheduler.c | 68 > ++++++++++++++++++++++++++++++------ > 1 file changed, 57 insertions(+), 11 deletions(-) I've taken a look at this branch, and I have some feedback. Adding ww_ctx blindly like is done doesn't work. For example I'm working on a patch that adds ww handling to i915_gem_object_ggtt_pin(), so you don't have to perform manually in prepare_shadow_wa_ctx(). intel_context_pin() already has a ww dance, you can keep calling it without open coding it. If you do decide to call it manually, intel_context_put would have to be called after ctx_fini. Cheers, Maarten<o:p></o:p> </div> </body> </html>