Hi, Andy, On Fri, Jul 31, 2020 at 06:28:37PM -0700, Andy Lutomirski wrote: > On Mon, Jul 13, 2020 at 4:48 PM Fenghua Yu <fenghua.yu@xxxxxxxxx> wrote: > > > > A #GP fault is generated when ENQCMD instruction is executed without > > a valid PASID value programmed in the current thread's PASID MSR. The > > #GP fault handler will initialize the MSR if a PASID has been allocated > > for this process. > > Let's take a step back here. Why are we trying to avoid IPIs? If you > call munmap(), you IPI other CPUs running tasks in the current mm. If > you do perf_event_open() and thus acquire RDPMC permission, you IPI > other CPUs running tasks in the current mm. If you call modify_ldt(), > you IPI other CPUs running tasks in the current mm. These events can > all happen more than once per process. > > Now we have ENQCMD. An mm can be assigned a PASID *once* in the model > that these patches support. Why not just send an IPI using > essentially identical code to the LDT sync or the CR4.PCE sync? ldt (or the other two cases) is different from ENQCMD: the PASID MSR is per-task and is supported by xsaves. The per-task PASID MSR needs to updated to ALL tasks. That means IPI, which only updates running tasks' MSRs, is not enough. All tasks' MSRs need to be updated when a PASID is allocated. This difference increases the complexity of sending IPI to running tasks and updating sleeping tasks's MSRs with locking etc. Of course, it's doable not to update the MSRs in all task when a new PASID is allocated to the mm. But that means we need to discard xsaves support for the MSR and create our own switch function to load the MSR. That increases complexity. We tried similar IPI way to update the PASID in about 200 lines of code. As Dave Hansen pointed, it's too complex. The current lazy updating the MSR only takes essential 3 lines of code in #GP. Does it make sense to still use the current fix up method to update the MSR? Thanks. -Fenghua _______________________________________________ amd-gfx mailing list amd-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/amd-gfx
