[AMD Official Use Only - AMD Internal Distribution Only] Hi, Christian
I think we have a discussion about this before . Alex also have a change that allow driver to use different write back address for the fence for each submission for the original issue . From MES point of view , MES will update the fence when the API can be complete successfully, so if the API (ex . remove_queue) fails due to other component issue (ex , CP hang), the MES will not update the fence In this situation , but MES itself still works and can respond to other commands (ex ,,read_reg) . Alex's change allow driver to check the fence for each API without mess around them . If you expect MES to stop responding to further commands after one API fails , that will introduce combability issue since this design already exist on products for customer and MES also need to works for windows . Also MES always need to respond to some commands like RESET etc that might make things worse if we need to change the logic . One possible solution is MES can trigger an Interrupt to indicate which submission has failed with the seq number . In this case driver can get the failure of the submission to MES in time and make its own decision for what to do next , What do you think about this ? Regards Shaoyun.liu -----Original Message----- From: amd-gfx <amd-gfx-boun...@lists.freedesktop.org> On Behalf Of Christian König Sent: Wednesday, May 29, 2024 11:19 AM To: Li, Yunxiang (Teddy) <yunxiang...@amd.com>; Koenig, Christian <christian.koe...@amd.com>; amd-gfx@lists.freedesktop.org Subject: Re: [PATCH v2 03/10] drm/amdgpu: abort fence poll if reset is started Am 29.05.24 um 16:48 schrieb Li, Yunxiang (Teddy): > [AMD Official Use Only - AMD Internal Distribution Only] > >> Yeah, I know. That's one of the reason I've pointed out on the patch >> adding that that this behavior is actually completely broken. >> >> If you run into issues with the MES because of this then please >> suggest a revert of that patch. > I think it just need to be improved to allow this force-signal behavior. The > current behavior is slow/inconvenient, but the old behavior is wrong. Since > MES will continue process submissions even when one submission failed. So > with just one fence location there's no way to tell if a command failed or > not. No the MES behavior is broken. When a submission failed it should stop processing or signal that the operation didn't completed through some other mechanism. Just not writing the fence and continuing results in tons of problems, from the TLB fence all the way to the ring buffer and reset handling. This is a hard requirement and really can't be changed. Regards, Christian.
