Author: avg
Date: Mon May 7 12:22:25 2018
New Revision: 333321
URL: https://svnweb.freebsd.org/changeset/base/333321
Log:
x86 cpususpend_handler: call wbinvd after setting suspend state bits
Without a subsequent wbinvd the changes to suspended_cpus (and
resuming_cpus) can be lost at least on AMD systems that use MOESI cache
coherency protocol. That can happen because one of APs ends up as an
Owner of the corresponding cache line(s) and the changes may never reach
the main memory before the AP is reset.
While here, move clearing of suspended_cpus a little bit earlier as the
fact of returning from savectx (with zero return value) means that the
CPU has fully restored it execution context.
Also, rework the comment that describes the need for resuming_cpus.
This change fixed suspend to RAM a previously broken AMD-based system.
Reviewed by: kib
Discussed with: bde
MFC after: 3 weeks
Differential Revision: https://reviews.freebsd.org/D15295
Modified:
head/sys/x86/x86/mp_x86.c
Modified: head/sys/x86/x86/mp_x86.c
==============================================================================
--- head/sys/x86/x86/mp_x86.c Mon May 7 09:42:35 2018 (r333320)
+++ head/sys/x86/x86/mp_x86.c Mon May 7 12:22:25 2018 (r333321)
@@ -1425,16 +1425,36 @@ cpususpend_handler(void)
#else
npxsuspend(susppcbs[cpu]->sp_fpususpend);
#endif
- wbinvd();
- CPU_SET_ATOMIC(cpu, &suspended_cpus);
/*
- * Hack for xen, which does not use resumectx() so never
- * uses the next clause: set resuming_cpus early so that
- * resume_cpus() can wait on the same bitmap for acpi and
- * xen. resuming_cpus now means eventually_resumable_cpus.
+ * suspended_cpus is cleared shortly after each AP is restarted
+ * by a Startup IPI, so that the BSP can proceed to restarting
+ * the next AP.
+ *
+ * resuming_cpus gets cleared when the AP completes
+ * initialization after having been released by the BSP.
+ * resuming_cpus is probably not the best name for the
+ * variable, because it is actually a set of processors that
+ * haven't resumed yet and haven't necessarily started resuming.
+ *
+ * Note that suspended_cpus is meaningful only for ACPI suspend
+ * as it's not really used for Xen suspend since the APs are
+ * automatically restored to the running state and the correct
+ * context. For the same reason resumectx is never called in
+ * that case.
*/
+ CPU_SET_ATOMIC(cpu, &suspended_cpus);
CPU_SET_ATOMIC(cpu, &resuming_cpus);
+
+ /*
+ * Invalidate the cache after setting the global status bits.
+ * The last AP to set its bit may end up being an Owner of the
+ * corresponding cache line in MOESI protocol. The AP may be
+ * stopped before the cache line is written to the main memory.
+ */
+ wbinvd();
} else {
+ /* Indicate that we have restarted and restored the context. */
+ CPU_CLR_ATOMIC(cpu, &suspended_cpus);
#ifdef __amd64__
fpuresume(susppcbs[cpu]->sp_fpususpend);
#else
@@ -1444,9 +1464,6 @@ cpususpend_handler(void)
initializecpu();
PCPU_SET(switchtime, 0);
PCPU_SET(switchticks, ticks);
-
- /* Indicate that we are resuming */
- CPU_CLR_ATOMIC(cpu, &suspended_cpus);
}
/* Wait for resume directive */