On 4/11/22 2:06 PM, Stephen Boyd wrote:
The second problem you have is exhibited by the IPA driver if the "fix" commit (upstream b95b668eaaa2) is back-ported to the Linux 5.10.y LTS branch (along with some other prerequisite commits). We can conclude that applying the above commit makes the bandwidth for an unused interconnect (or perhaps the rate for the IPA core clock) get set to zero. And in that case, an attempt to access IPA hardware leads to the crash you observed. The IPA driver does not implement runtime power management until Linux v5.15. You later said you thought enabling that might ensure the clock and interconnects were active when needed by the IPA driver, and I concur (but there could be a little more to it).Is the runtime PM patch series necessary to enable the IPA clk and interconnects? Things don't look good on 5.10.y and I'm not sure it will be workable. Commit b1d681d8d324 ("interconnect: Add sync state support") was introduced in v5.10 and that seems to be the commit that broke suspend on Lazor.
This isn't a response to your complete message but I'm going to respond to this part. Before runtime PM was in place, the "IPA clock" (which was a logical notion representing the IPA core clock and all the interconnects it uses) was enabled before the IPA hardware was first touched. It was disabled again when system suspend occurred, and re-enabled again on system resume. At one time we did observe the XO clock turning off. I'm not sure that answers your question. But bottom line is that system suspend/resume were supported (and made IPA clock+interconnects get shut off and then on again), but not runtime PM. -Alex
