Merge branch 'csd.2020.09.04a' into HEAD

csd.2020.09.04a: CPU smp_call_function() torture tests.

Documentation/admin-guide/kernel-parameters.txt

···
 			and gids from such clients.  This is intended to ease
 			migration from NFSv2/v3.
 
+	nmi_backtrace.backtrace_idle [KNL]
+			Dump stacks even of idle CPUs in response to an
+			NMI stack-backtrace request.
+
 	nmi_debug=	[KNL,SH] Specify one or more actions to take
 			when a NMI is triggered.
 			Format: [state][,regs][,debounce][,die]

include/linux/smp.h

···
 		struct {
 			struct llist_node llist;
 			unsigned int flags;
+#ifdef CONFIG_64BIT
+			u16 src, dst;
+#endif
 		};
 	};
 	smp_call_func_t func;

include/linux/smp_types.h

···
 		unsigned int	u_flags;
 		atomic_t	a_flags;
 	};
+#ifdef CONFIG_64BIT
+	u16 src, dst;
+#endif
 };
 
 #endif /* __LINUX_SMP_TYPES_H */

kernel/smp.c

···
 #include <linux/sched.h>
 #include <linux/sched/idle.h>
 #include <linux/hypervisor.h>
+#include <linux/sched/clock.h>
+#include <linux/nmi.h>
+#include <linux/sched/debug.h>
 
 #include "smpboot.h"
 #include "sched/smp.h"
···
 	smpcfd_prepare_cpu(smp_processor_id());
 }
 
+#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
+
+static DEFINE_PER_CPU(call_single_data_t *, cur_csd);
+static DEFINE_PER_CPU(smp_call_func_t, cur_csd_func);
+static DEFINE_PER_CPU(void *, cur_csd_info);
+
+#define CSD_LOCK_TIMEOUT (5ULL * NSEC_PER_SEC)
+static atomic_t csd_bug_count = ATOMIC_INIT(0);
+
+/* Record current CSD work for current CPU, NULL to erase. */
+static void csd_lock_record(call_single_data_t *csd)
+{
+	if (!csd) {
+		smp_mb(); /* NULL cur_csd after unlock. */
+		__this_cpu_write(cur_csd, NULL);
+		return;
+	}
+	__this_cpu_write(cur_csd_func, csd->func);
+	__this_cpu_write(cur_csd_info, csd->info);
+	smp_wmb(); /* func and info before csd. */
+	__this_cpu_write(cur_csd, csd);
+	smp_mb(); /* Update cur_csd before function call. */
+		  /* Or before unlock, as the case may be. */
+}
+
+static __always_inline int csd_lock_wait_getcpu(call_single_data_t *csd)
+{
+	unsigned int csd_type;
+
+	csd_type = CSD_TYPE(csd);
+	if (csd_type == CSD_TYPE_ASYNC || csd_type == CSD_TYPE_SYNC)
+		return csd->dst; /* Other CSD_TYPE_ values might not have ->dst. */
+	return -1;
+}
+
+/*
+ * Complain if too much time spent waiting.  Note that only
+ * the CSD_TYPE_SYNC/ASYNC types provide the destination CPU,
+ * so waiting on other types gets much less information.
+ */
+static __always_inline bool csd_lock_wait_toolong(call_single_data_t *csd, u64 ts0, u64 *ts1, int *bug_id)
+{
+	int cpu = -1;
+	int cpux;
+	bool firsttime;
+	u64 ts2, ts_delta;
+	call_single_data_t *cpu_cur_csd;
+	unsigned int flags = READ_ONCE(csd->flags);
+
+	if (!(flags & CSD_FLAG_LOCK)) {
+		if (!unlikely(*bug_id))
+			return true;
+		cpu = csd_lock_wait_getcpu(csd);
+		pr_alert("csd: CSD lock (#%d) got unstuck on CPU#%02d, CPU#%02d released the lock.\n",
+			 *bug_id, raw_smp_processor_id(), cpu);
+		return true;
+	}
+
+	ts2 = sched_clock();
+	ts_delta = ts2 - *ts1;
+	if (likely(ts_delta <= CSD_LOCK_TIMEOUT))
+		return false;
+
+	firsttime = !*bug_id;
+	if (firsttime)
+		*bug_id = atomic_inc_return(&csd_bug_count);
+	cpu = csd_lock_wait_getcpu(csd);
+	if (WARN_ONCE(cpu < 0 || cpu >= nr_cpu_ids, "%s: cpu = %d\n", __func__, cpu))
+		cpux = 0;
+	else
+		cpux = cpu;
+	cpu_cur_csd = smp_load_acquire(&per_cpu(cur_csd, cpux)); /* Before func and info. */
+	pr_alert("csd: %s non-responsive CSD lock (#%d) on CPU#%d, waiting %llu ns for CPU#%02d %pS(%ps).\n",
+		 firsttime ? "Detected" : "Continued", *bug_id, raw_smp_processor_id(), ts2 - ts0,
+		 cpu, csd->func, csd->info);
+	if (cpu_cur_csd && csd != cpu_cur_csd) {
+		pr_alert("\tcsd: CSD lock (#%d) handling prior %pS(%ps) request.\n",
+			 *bug_id, READ_ONCE(per_cpu(cur_csd_func, cpux)),
+			 READ_ONCE(per_cpu(cur_csd_info, cpux)));
+	} else {
+		pr_alert("\tcsd: CSD lock (#%d) %s.\n",
+			 *bug_id, !cpu_cur_csd ? "unresponsive" : "handling this request");
+	}
+	if (cpu >= 0) {
+		if (!trigger_single_cpu_backtrace(cpu))
+			dump_cpu_task(cpu);
+		if (!cpu_cur_csd) {
+			pr_alert("csd: Re-sending CSD lock (#%d) IPI from CPU#%02d to CPU#%02d\n", *bug_id, raw_smp_processor_id(), cpu);
+			arch_send_call_function_single_ipi(cpu);
+		}
+	}
+	dump_stack();
+	*ts1 = ts2;
+
+	return false;
+}
+
 /*
  * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
  *
···
  */
 static __always_inline void csd_lock_wait(call_single_data_t *csd)
 {
+	int bug_id = 0;
+	u64 ts0, ts1;
+
+	ts1 = ts0 = sched_clock();
+	for (;;) {
+		if (csd_lock_wait_toolong(csd, ts0, &ts1, &bug_id))
+			break;
+		cpu_relax();
+	}
+	smp_acquire__after_ctrl_dep();
+}
+
+#else
+static void csd_lock_record(call_single_data_t *csd)
+{
+}
+
+static __always_inline void csd_lock_wait(call_single_data_t *csd)
+{
 	smp_cond_load_acquire(&csd->flags, !(VAL & CSD_FLAG_LOCK));
 }
+#endif
 
 static __always_inline void csd_lock(call_single_data_t *csd)
 {
···
 		 * We can unlock early even for the synchronous on-stack case,
 		 * since we're doing this from the same CPU..
 		 */
+		csd_lock_record(csd);
 		csd_unlock(csd);
 		local_irq_save(flags);
 		func(info);
+		csd_lock_record(NULL);
 		local_irq_restore(flags);
 		return 0;
 	}
···
 				entry = &csd_next->llist;
 			}
 
+			csd_lock_record(csd);
 			func(info);
 			csd_unlock(csd);
+			csd_lock_record(NULL);
 		} else {
 			prev = &csd->llist;
 		}
···
 				smp_call_func_t func = csd->func;
 				void *info = csd->info;
 
+				csd_lock_record(csd);
 				csd_unlock(csd);
 				func(info);
+				csd_lock_record(NULL);
 			} else if (type == CSD_TYPE_IRQ_WORK) {
 				irq_work_single(csd);
 			}
···
 
 	csd->func = func;
 	csd->info = info;
+#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
+	csd->src = smp_processor_id();
+	csd->dst = cpu;
+#endif
 
 	err = generic_exec_single(cpu, csd);
 
···
 			csd->flags |= CSD_TYPE_SYNC;
 		csd->func = func;
 		csd->info = info;
+#ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
+		csd->src = smp_processor_id();
+		csd->dst = cpu;
+#endif
 		if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
 			__cpumask_set_cpu(cpu, cfd->cpumask_ipi);
 	}

lib/Kconfig.debug

···
 	  module may be built after the fact on the running kernel to
 	  be tested, if desired.
 
+config CSD_LOCK_WAIT_DEBUG
+	bool "Debugging for csd_lock_wait(), called from smp_call_function*()"
+	depends on DEBUG_KERNEL
+	depends on 64BIT
+	default n
+	help
+	  This option enables debug prints when CPUs are slow to respond
+	  to the smp_call_function*() IPI wrappers.  These debug prints
+	  include the IPI handler function currently executing (if any)
+	  and relevant stack traces.
+
 endmenu # lock debugging
 
 config TRACE_IRQFLAGS

lib/nmi_backtrace.c

···
 	put_cpu();
 }
 
+// Dump stacks even for idle CPUs.
+static bool backtrace_idle;
+module_param(backtrace_idle, bool, 0644);
+
 bool nmi_cpu_backtrace(struct pt_regs *regs)
 {
 	int cpu = smp_processor_id();
 
 	if (cpumask_test_cpu(cpu, to_cpumask(backtrace_mask))) {
-		if (regs && cpu_in_idle(instruction_pointer(regs))) {
+		if (!READ_ONCE(backtrace_idle) && regs && cpu_in_idle(instruction_pointer(regs))) {
 			pr_warn("NMI backtrace for cpu %d skipped: idling at %pS\n",
 				cpu, (void *)instruction_pointer(regs));
 		} else {