<html><head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> </head> <body text="#000000" bgcolor="#FFFFFF"> Great! I will update the patches to also use the barrier in PSP MODE 1 reset case and resend the patches for formal review. Andrey <div class="moz-cite-prefix">On 12/11/19 7:18 AM, Ma, Le wrote: </div> <blockquote type="cite" cite="mid:MN2PR12MB42855499B960506C3BA62198F65A0@MN2PR12MB4285.namprd12.prod.outlook.com"> <meta name="Generator" content="Microsoft Word 15 (filtered medium)"> <style></style> <div class="WordSection1"> [AMD Official Use Only - Internal Distribution Only]<o:p></o:p> <o:p> </o:p> I tried your new patches to run BACO for about 10 loops and the result looks positive, without observing enter/exit baco message failure again.<o:p></o:p> <o:p> </o:p> The time interval between BACO entries or exits in my environment was almost less than 10 us: max 36us, min 2us. I think it’s safe enough according to the sample data we collected in both sides.<o:p></o:p> <o:p> </o:p> And it looks not necessary to continue using system_highpri_wq any more because we require all the nodes enter or exit at the same time, while do not mind how long the time interval is b/t enter and exit. The system_unbound_wq can satisfy our requirement here since it wakes different CPUs up to work at the same time.<o:p></o:p> <o:p> </o:p> Regards,<o:p></o:p> Ma Le<o:p></o:p> <o:p> </o:p> <div> <div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0in 0in 0in"> From: Grodzovsky, Andrey <a class="moz-txt-link-rfc2396E" href="mailto:Andrey.Grodzovsky@amd.com"><Andrey.Grodzovsky@amd.com></a> Sent: Wednesday, December 11, 2019 3:56 AM To: Ma, Le <a class="moz-txt-link-rfc2396E" href="mailto:Le.Ma@amd.com"><Le.Ma@amd.com></a>; <a class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org">amd-gfx@lists.freedesktop.org</a>; Zhou1, Tao <a class="moz-txt-link-rfc2396E" href="mailto:Tao.Zhou1@amd.com"><Tao.Zhou1@amd.com></a>; Deucher, Alexander <a class="moz-txt-link-rfc2396E" href="mailto:Alexander.Deucher@amd.com"><Alexander.Deucher@amd.com></a>; Li, Dennis <a class="moz-txt-link-rfc2396E" href="mailto:Dennis.Li@amd.com"><Dennis.Li@amd.com></a>; Zhang, Hawking <a class="moz-txt-link-rfc2396E" href="mailto:Hawking.Zhang@amd.com"><Hawking.Zhang@amd.com></a> Cc: Chen, Guchun <a class="moz-txt-link-rfc2396E" href="mailto:Guchun.Chen@amd.com"><Guchun.Chen@amd.com></a> Subject: Re: [PATCH 07/10] drm/amdgpu: add concurrent baco reset support for XGMI<o:p></o:p> </div> </div> <o:p> </o:p> I switched the workqueue we were using for xgmi_reset_work from system_highpri_wq to system_unbound_wq - the difference is that workers servicing the queue in system_unbound_wq are not bounded to specific CPU and so the reset jobs for each XGMI node are getting scheduled to different CPU while system_highpri_wq is a bounded work queue. I traced it as bellow for 10 consecutive times and didn't see errors any more. Also the time diff between BACO entries or exits was never more then around 2 uS. <o:p></o:p> Please give this updated patchset a try<o:p></o:p> kworker/u16:2-57 [004] ...1 243.276312: trace_code: func: vega20_baco_set_state, line 91 <----- - Before BEACO enter <...>-60 [007] ...1 243.276312: trace_code: func: vega20_baco_set_state, line 91 <----- - Before BEACO enter kworker/u16:2-57 [004] ...1 243.276384: trace_code: func: vega20_baco_set_state, line 105 <----- - After BEACO enter done <...>-60 [007] ...1 243.276392: trace_code: func: vega20_baco_set_state, line 105 <----- - After BEACO enter done kworker/u16:3-60 [007] ...1 243.276397: trace_code: func: vega20_baco_set_state, line 108 <----- - Before BEACO exit kworker/u16:2-57 [004] ...1 243.276399: trace_code: func: vega20_baco_set_state, line 108 <----- - Before BEACO exit kworker/u16:3-60 [007] ...1 243.288067: trace_code: func: vega20_baco_set_state, line 114 <----- - After BEACO exit done kworker/u16:2-57 [004] ...1 243.295624: trace_code: func: vega20_baco_set_state, line 114 <----- - After BEACO exit done<o:p></o:p> Andrey<o:p></o:p> <div> On 12/9/19 9:45 PM, Ma, Le wrote:<o:p></o:p> </div> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> [AMD Official Use Only - Internal Distribution Only]<o:p></o:p> <o:p> </o:p> I’m fine with your solution if synchronization time interval satisfies BACO requirements and loop test can pass on XGMI system.<o:p></o:p> <o:p></o:p> Regards,<o:p></o:p> Ma Le<o:p></o:p> <o:p></o:p> <div> <div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0in 0in 0in"> From: Grodzovsky, Andrey <a href="mailto:Andrey.Grodzovsky@amd.com" moz-do-not-send="true"><Andrey.Grodzovsky@amd.com></a> Sent: Monday, December 9, 2019 11:52 PM To: Ma, Le <a href="mailto:Le.Ma@amd.com" moz-do-not-send="true"><Le.Ma@amd.com></a>; <a href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true"> amd-gfx@lists.freedesktop.org</a>; Zhou1, Tao <a href="mailto:Tao.Zhou1@amd.com" moz-do-not-send="true"> <Tao.Zhou1@amd.com></a>; Deucher, Alexander <a href="mailto:Alexander.Deucher@amd.com" moz-do-not-send="true"> <Alexander.Deucher@amd.com></a>; Li, Dennis <a href="mailto:Dennis.Li@amd.com" moz-do-not-send="true"><Dennis.Li@amd.com></a>; Zhang, Hawking <a href="mailto:Hawking.Zhang@amd.com" moz-do-not-send="true"><Hawking.Zhang@amd.com></a> Cc: Chen, Guchun <a href="mailto:Guchun.Chen@amd.com" moz-do-not-send="true"><Guchun.Chen@amd.com></a> Subject: Re: [PATCH 07/10] drm/amdgpu: add concurrent baco reset support for XGMI<o:p></o:p> </div> </div> <o:p></o:p> Thanks a lot Ma for trying - I think I have to have my own system to debug this so I will keep trying enabling XGMI - i still think the is the right and the generic solution for multiple nodes reset synchronization and in fact the barrier should also be used for synchronizing PSP mode 1 XGMI reset too.<o:p></o:p> Andrey<o:p></o:p> <div> On 12/9/19 6:34 AM, Ma, Le wrote:<o:p></o:p> </div> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> [AMD Official Use Only - Internal Distribution Only]<o:p></o:p> <o:p></o:p> Hi Andrey,<o:p></o:p> <o:p></o:p> I tried your patches on my 2P XGMI platform. The baco can work at most time, and randomly got following error:<o:p></o:p> [ 1701.542298] amdgpu: [powerplay] Failed to send message 0x25, response 0x0<o:p></o:p> <o:p></o:p> This error usually means some sync issue exist for xgmi baco case. Feel free to debug your patches on my XGMI platform.<o:p></o:p> <o:p></o:p> Regards,<o:p></o:p> Ma Le<o:p></o:p> <o:p></o:p> <div> <div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0in 0in 0in"> From: Grodzovsky, Andrey <a href="mailto:Andrey.Grodzovsky@amd.com" moz-do-not-send="true"><Andrey.Grodzovsky@amd.com></a> Sent: Saturday, December 7, 2019 5:51 AM To: Ma, Le <a href="mailto:Le.Ma@amd.com" moz-do-not-send="true"><Le.Ma@amd.com></a>; <a href="mailto:amd-gfx@lists.freedesktop.org" moz-do-not-send="true">amd-gfx@lists.freedesktop.org</a>; Zhou1, Tao <a href="mailto:Tao.Zhou1@amd.com" moz-do-not-send="true"><Tao.Zhou1@amd.com></a>; Deucher, Alexander <a href="mailto:Alexander.Deucher@amd.com" moz-do-not-send="true"><Alexander.Deucher@amd.com></a>; Li, Dennis <a href="mailto:Dennis.Li@amd.com" moz-do-not-send="true"><Dennis.Li@amd.com></a>; Zhang, Hawking <a href="mailto:Hawking.Zhang@amd.com" moz-do-not-send="true"><Hawking.Zhang@amd.com></a> Cc: Chen, Guchun <a href="mailto:Guchun.Chen@amd.com" moz-do-not-send="true"><Guchun.Chen@amd.com></a> Subject: Re: [PATCH 07/10] drm/amdgpu: add concurrent baco reset support for XGMI<o:p></o:p> </div> </div> <o:p></o:p> Hey Ma, attached a solution - it's just compiled as I still can't make my XGMI setup work (with bridge connected only one device is visible to the system while the other is not). Please try it on your system if you have a chance. <o:p></o:p> Andrey<o:p></o:p> <div> On 12/4/19 10:14 PM, Ma, Le wrote:<o:p></o:p> </div> <blockquote style="margin-top:5.0pt;margin-bottom:5.0pt"> AFAIK it's enough for even single one node in the hive to to fail the enter the BACO state on time to fail the entire hive reset procedure, no ?<o:p></o:p> [Le]: Yeah, agree that. I’ve been thinking that make all nodes entering baco simultaneously can reduce the possibility of node failure to enter/exit BACO risk. For example, in an XGMI hive with 8 nodes, the total time interval of 8 nodes enter/exit BACO on 8 CPUs is less than the interval that 8 nodes enter BACO serially and exit BACO serially depending on one CPU with yield capability. This interval is usually strict for BACO feature itself. Anyway, we need more looping test later on any method we will choose.<o:p></o:p> Any way - I see our discussion blocks your entire patch set - I think you can go ahead and commit yours way (I think you got an RB from Hawking) and I will look then and see if I can implement my method and if it works will just revert your patch.<o:p></o:p> [Le]: OK, fine.<o:p></o:p> Andrey<o:p></o:p> </blockquote> </blockquote> </blockquote> </div> </blockquote> </body> </html>