ipmi: add parameter to limit CPU usage in kipmid

In some cases kipmid can use a lot of CPU. This adds a way to tune the CPU used by kipmid to help in those cases. By setting kipmid_max_busy_us to a value between 100 and 500, it is possible to bring down kipmid CPU load to practically 0 without loosing too much ipmi throughput performance. Not setting the value, or setting the value to zero, operation is unaffected. Signed-off-by: Martin Wilck <martin.wilck@ts.fujitsu.com> Signed-off-by: Corey Minyard <cminyard@mvista.com> Cc: Jean Delvare <jdelvare@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

ipmi: add parameter to limit CPU usage in kipmid
In some cases kipmid can use a lot of CPU. This adds a way to tune the CPU used by kipmid to help in those cases. By setting kipmid_max_busy_us to a value between 100 and 500, it is possible to bring down kipmid CPU load to practically 0 without loosing too much ipmi throughput performance. Not setting the value, or setting the value to zero, operation is unaffected. Signed-off-by: Martin Wilck <martin.wilck@ts.fujitsu.com> Signed-off-by: Corey Minyard <cminyard@mvista.com> Cc: Jean Delvare <jdelvare@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Martin Wilck · Linus Torvalds
1 parent f1eb1332b8
Showing 2 changed files with 76 additions and 2 deletions Side-by-side Diff
Documentation/IPMI.txt
drivers/char/ipmi/ipmi_si_intf.c
@@ -365,6 +365,7 @@
        regshifts=<shift1>,<shift2>,...
        slave_addrs=<addr1>,<addr2>,...
        force_kipmid=<enable1>,<enable2>,...
+       kipmid_max_busy_us=<ustime1>,<ustime2>,...
        unload_when_empty=[0|1]
  
 Each of these except si_trydefaults is a list, the first item for the
@@ -433,6 +434,7 @@
        ipmi_si.regshifts=<shift1>,<shift2>,...
        ipmi_si.slave_addrs=<addr1>,<addr2>,...
        ipmi_si.force_kipmid=<enable1>,<enable2>,...
+       ipmi_si.kipmid_max_busy_us=<ustime1>,<ustime2>,...
  
 It works the same as the module parameters of the same names.
  
@@ -449,6 +451,16 @@
 force this thread on or off.  If you force it off and don't have
 interrupts, the driver will run VERY slowly.  Don't blame me,
 these interfaces suck.
+
+Unfortunately, this thread can use a lot of CPU depending on the
+interface's performance.  This can waste a lot of CPU and cause
+various issues with detecting idle CPU and using extra power.  To
+avoid this, the kipmid_max_busy_us sets the maximum amount of time, in
+microseconds, that kipmid will spin before sleeping for a tick.  This
+value sets a balance between performance and CPU waste and needs to be
+tuned to your needs.  Maybe, someday, auto-tuning will be added, but
+that's not a simple thing and even the auto-tuning would need to be
+tuned to the user's desired performance.
  
 The driver supports a hot add and remove of interfaces.  This way,
 interfaces can be added or removed after the kernel is up and running.
@@ -295,6 +295,9 @@
 static int force_kipmid[SI_MAX_PARMS];
 static int num_force_kipmid;
  
+static unsigned int kipmid_max_busy_us[SI_MAX_PARMS];
+static int num_max_busy_us;
+
 static int unload_when_empty = 1;
  
 static int try_smi_init(struct smi_info *smi);
  
  
  
  
  
  
@@ -925,23 +928,77 @@
 	}
 }
  
+/*
+ * Use -1 in the nsec value of the busy waiting timespec to tell that
+ * we are spinning in kipmid looking for something and not delaying
+ * between checks
+ */
+static inline void ipmi_si_set_not_busy(struct timespec *ts)
+{
+	ts->tv_nsec = -1;
+}
+static inline int ipmi_si_is_busy(struct timespec *ts)
+{
+	return ts->tv_nsec != -1;
+}
+
+static int ipmi_thread_busy_wait(enum si_sm_result smi_result,
+				 const struct smi_info *smi_info,
+				 struct timespec *busy_until)
+{
+	unsigned int max_busy_us = 0;
+
+	if (smi_info->intf_num < num_max_busy_us)
+		max_busy_us = kipmid_max_busy_us[smi_info->intf_num];
+	if (max_busy_us == 0 || smi_result != SI_SM_CALL_WITH_DELAY)
+		ipmi_si_set_not_busy(busy_until);
+	else if (!ipmi_si_is_busy(busy_until)) {
+		getnstimeofday(busy_until);
+		timespec_add_ns(busy_until, max_busy_us*NSEC_PER_USEC);
+	} else {
+		struct timespec now;
+		getnstimeofday(&now);
+		if (unlikely(timespec_compare(&now, busy_until) > 0)) {
+			ipmi_si_set_not_busy(busy_until);
+			return 0;
+		}
+	}
+	return 1;
+}
+
+
+/*
+ * A busy-waiting loop for speeding up IPMI operation.
+ *
+ * Lousy hardware makes this hard.  This is only enabled for systems
+ * that are not BT and do not have interrupts.  It starts spinning
+ * when an operation is complete or until max_busy tells it to stop
+ * (if that is enabled).  See the paragraph on kimid_max_busy_us in
+ * Documentation/IPMI.txt for details.
+ */
 static int ipmi_thread(void *data)
 {
 	struct smi_info *smi_info = data;
 	unsigned long flags;
 	enum si_sm_result smi_result;
+	struct timespec busy_until;
  
+	ipmi_si_set_not_busy(&busy_until);
 	set_user_nice(current, 19);
 	while (!kthread_should_stop()) {
+		int busy_wait;
+
 		spin_lock_irqsave(&(smi_info->si_lock), flags);
 		smi_result = smi_event_handler(smi_info, 0);
 		spin_unlock_irqrestore(&(smi_info->si_lock), flags);
+		busy_wait = ipmi_thread_busy_wait(smi_result, smi_info,
+						  &busy_until);
 		if (smi_result == SI_SM_CALL_WITHOUT_DELAY)
 			; /* do nothing */
-		else if (smi_result == SI_SM_CALL_WITH_DELAY)
+		else if (smi_result == SI_SM_CALL_WITH_DELAY && busy_wait)
 			schedule();
 		else
-			schedule_timeout_interruptible(1);
+			schedule_timeout_interruptible(0);
 	}
 	return 0;
 }
@@ -1212,6 +1269,11 @@
 MODULE_PARM_DESC(unload_when_empty, "Unload the module if no interfaces are"
 		 " specified or found, default is 1.  Setting to 0"
 		 " is useful for hot add of devices using hotmod.");
+module_param_array(kipmid_max_busy_us, uint, &num_max_busy_us, 0644);
+MODULE_PARM_DESC(kipmid_max_busy_us,
+		 "Max time (in microseconds) to busy-wait for IPMI data before"
+		 " sleeping. 0 (default) means to wait forever. Set to 100-500"
+		 " if kipmid is using up a lot of CPU time.");
  
  
 static void std_irq_cleanup(struct smi_info *info)
...	...	@@ -365,6 +365,7 @@
365	365	regshifts=<shift1>,<shift2>,...
366	366	slave_addrs=<addr1>,<addr2>,...
367	367	force_kipmid=<enable1>,<enable2>,...
	368	+ kipmid_max_busy_us=<ustime1>,<ustime2>,...
368	369	unload_when_empty=[0\|1]
369	370
370	371	Each of these except si_trydefaults is a list, the first item for the
...	...	@@ -433,6 +434,7 @@
433	434	ipmi_si.regshifts=<shift1>,<shift2>,...
434	435	ipmi_si.slave_addrs=<addr1>,<addr2>,...
435	436	ipmi_si.force_kipmid=<enable1>,<enable2>,...
	437	+ ipmi_si.kipmid_max_busy_us=<ustime1>,<ustime2>,...
436	438
437	439	It works the same as the module parameters of the same names.
438	440
...	...	@@ -449,6 +451,16 @@
449	451	force this thread on or off. If you force it off and don't have
450	452	interrupts, the driver will run VERY slowly. Don't blame me,
451	453	these interfaces suck.
	454	+
	455	+Unfortunately, this thread can use a lot of CPU depending on the
	456	+interface's performance. This can waste a lot of CPU and cause
	457	+various issues with detecting idle CPU and using extra power. To
	458	+avoid this, the kipmid_max_busy_us sets the maximum amount of time, in
	459	+microseconds, that kipmid will spin before sleeping for a tick. This
	460	+value sets a balance between performance and CPU waste and needs to be
	461	+tuned to your needs. Maybe, someday, auto-tuning will be added, but
	462	+that's not a simple thing and even the auto-tuning would need to be
	463	+tuned to the user's desired performance.
452	464
453	465	The driver supports a hot add and remove of interfaces. This way,
454	466	interfaces can be added or removed after the kernel is up and running.
...	...	@@ -295,6 +295,9 @@
295	295	static int force_kipmid[SI_MAX_PARMS];
296	296	static int num_force_kipmid;
297	297
	298	+static unsigned int kipmid_max_busy_us[SI_MAX_PARMS];
	299	+static int num_max_busy_us;
	300	+
298	301	static int unload_when_empty = 1;
299	302
300	303	static int try_smi_init(struct smi_info *smi);
301	304
302	305
303	306
304	307
305	308
306	309
...	...	@@ -925,23 +928,77 @@
925	928	}
926	929	}
927	930
	931	+/*
	932	+ * Use -1 in the nsec value of the busy waiting timespec to tell that
	933	+ * we are spinning in kipmid looking for something and not delaying
	934	+ * between checks
	935	+ */
	936	+static inline void ipmi_si_set_not_busy(struct timespec *ts)
	937	+{
	938	+ ts->tv_nsec = -1;
	939	+}
	940	+static inline int ipmi_si_is_busy(struct timespec *ts)
	941	+{
	942	+ return ts->tv_nsec != -1;
	943	+}
	944	+
	945	+static int ipmi_thread_busy_wait(enum si_sm_result smi_result,
	946	+ const struct smi_info *smi_info,
	947	+ struct timespec *busy_until)
	948	+{
	949	+ unsigned int max_busy_us = 0;
	950	+
	951	+ if (smi_info->intf_num < num_max_busy_us)
	952	+ max_busy_us = kipmid_max_busy_us[smi_info->intf_num];
	953	+ if (max_busy_us == 0 \|\| smi_result != SI_SM_CALL_WITH_DELAY)
	954	+ ipmi_si_set_not_busy(busy_until);
	955	+ else if (!ipmi_si_is_busy(busy_until)) {
	956	+ getnstimeofday(busy_until);
	957	+ timespec_add_ns(busy_until, max_busy_us*NSEC_PER_USEC);
	958	+ } else {
	959	+ struct timespec now;
	960	+ getnstimeofday(&now);
	961	+ if (unlikely(timespec_compare(&now, busy_until) > 0)) {
	962	+ ipmi_si_set_not_busy(busy_until);
	963	+ return 0;
	964	+ }
	965	+ }
	966	+ return 1;
	967	+}
	968	+
	969	+
	970	+/*
	971	+ * A busy-waiting loop for speeding up IPMI operation.
	972	+ *
	973	+ * Lousy hardware makes this hard. This is only enabled for systems
	974	+ * that are not BT and do not have interrupts. It starts spinning
	975	+ * when an operation is complete or until max_busy tells it to stop
	976	+ * (if that is enabled). See the paragraph on kimid_max_busy_us in
	977	+ * Documentation/IPMI.txt for details.
	978	+ */
928	979	static int ipmi_thread(void *data)
929	980	{
930	981	struct smi_info *smi_info = data;
931	982	unsigned long flags;
932	983	enum si_sm_result smi_result;
	984	+ struct timespec busy_until;
933	985
	986	+ ipmi_si_set_not_busy(&busy_until);
934	987	set_user_nice(current, 19);
935	988	while (!kthread_should_stop()) {
	989	+ int busy_wait;
	990	+
936	991	spin_lock_irqsave(&(smi_info->si_lock), flags);
937	992	smi_result = smi_event_handler(smi_info, 0);
938	993	spin_unlock_irqrestore(&(smi_info->si_lock), flags);
	994	+ busy_wait = ipmi_thread_busy_wait(smi_result, smi_info,
	995	+ &busy_until);
939	996	if (smi_result == SI_SM_CALL_WITHOUT_DELAY)
940	997	; /* do nothing */
941		- else if (smi_result == SI_SM_CALL_WITH_DELAY)
	998	+ else if (smi_result == SI_SM_CALL_WITH_DELAY && busy_wait)
942	999	schedule();
943	1000	else
944		- schedule_timeout_interruptible(1);
	1001	+ schedule_timeout_interruptible(0);
945	1002	}
946	1003	return 0;
947	1004	}
...	...	@@ -1212,6 +1269,11 @@
1212	1269	MODULE_PARM_DESC(unload_when_empty, "Unload the module if no interfaces are"
1213	1270	" specified or found, default is 1. Setting to 0"
1214	1271	" is useful for hot add of devices using hotmod.");
	1272	+module_param_array(kipmid_max_busy_us, uint, &num_max_busy_us, 0644);
	1273	+MODULE_PARM_DESC(kipmid_max_busy_us,
	1274	+ "Max time (in microseconds) to busy-wait for IPMI data before"
	1275	+ " sleeping. 0 (default) means to wait forever. Set to 100-500"
	1276	+ " if kipmid is using up a lot of CPU time.");
1215	1277
1216	1278
1217	1279	static void std_irq_cleanup(struct smi_info *info)