cpacfstats: Handle CPU hotplug

cpacfstatsd now correctly handles offline cpus and dynamically attaches to
cpus once they get online.  If events are enabled when a hotplug event
occurs, cpacfstatsd uses a pseudo-counter to notify user applications about
the occurence of this event and a potential data inaccuracy.

cpacfstats shows if a hotplug event has been detected since at least one
counter was activated.  As soon as all counters are deactivated, the
hotplug detection state is reset.

Signed-off-by: Juergen Christ <jchrist@linux.ibm.com>
Reviewed-By: Harald Freudenberger <freude@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>

README.md

@@ -306,7 +306,6 @@ build options:
 |                |                    | hsavmcore                             |
 | zlib           | `HAVE_ZLIB`        | zgetdump, dump2tar                    |
 | ncurses        | `HAVE_NCURSES`     | hyptop                                |
-| pfm            | `HAVE_PFM`         | cpacfstats                            |
 | net-snmp       | `HAVE_SNMP`        | osasnmpd                              |
 | glibc-static   | `HAVE_LIBC_STATIC` | zfcpdump                              |
 | openssl        | `HAVE_OPENSSL`     | genprotimg, zkey, libekmfweb,         |
@@ -320,6 +319,7 @@ build options:
 | libxml2        | `HAVE_LIBXML2`     | libkmipclient                         |
 | systemd        | `HAVE_SYSTEMD`     | hsavmcore                             |
 | liblockfile    | `HAVE_LOCKFILE`    | ap-check                              |
+| libudev        | `HAVE_LIBUDEV`     | cpacfstatsd                           |
 
 This table lists additional build or install options:
 
@@ -403,14 +403,13 @@ the different tools are provided:
   As of s390-tools-1.13.0, the minimum required kernel level is 2.6.38.
 
 * cpacfstats:
-  For building the cpacfstats tools you need libpfm version 4 or
-  newer installed (libpfm-devel.rpm). Tip: you may skip the cpacfstats
-  build by adding `HAVE_PFN=0` to the make invocation. To run the
-  cpacfstats daemon the kernel needs to have performance
-  events enabled (check for `CONFIG_PERF_EVENTS=y`) and you need libpfm
-  version 4 or newer installed. A new group 'cpacfstats' needs to be
-  created and all users intending to use the tool should be added to
-  this group.
+  For building the cpacfstats tools you need libudev installed
+  (systemd-devel.rpm). Tip: you may skip the cpacfstats build by
+  adding `HAVE_LIBUDEV=0` to the make invocation. To run the
+  cpacfstats daemon the kernel needs to have performance events
+  enabled (check for `CONFIG_PERF_EVENTS=y`). A new group 'cpacfstats'
+  needs to be created and all users intending to use the tool should
+  be added to this group.
 
 * zdev:
   Depending on the boot loader and initial RAM-disk mechanism used by a

cpacfstats/Makefile

@@ -1,11 +1,28 @@
 include ../common.mak
 
+ifeq (${HAVE_LIBUDEV},0)
+
+all:
+	$(SKIP) HAVE_LIBUDEV=0
+
+install:
+	$(SKIP) HAVE_LIBUDEV=0
+
+else
+
+check_dep:
+	$(call check_dep, \
+		"cpacfstatsd", \
+		"libudev.h", \
+		"systemd-devel", \
+		"HAVE_LIBUDEV=0")
+
 ALL_CPPFLAGS += -DVERSION=$(VERSION)
 
-all:		cpacfstats cpacfstatsd
+all:		check_dep cpacfstats cpacfstatsd
 
 cpacfstatsd:	cpacfstatsd.o stats_sock.o perf_crypto.o
-		$(LINK) $(ALL_LDFLAGS) $^ $(LDLIBS) -o $@
+		$(LINK) $(ALL_LDFLAGS) $^ $(LDLIBS) -ludev -lpthread -o $@
 
 cpacfstats:	cpacfstats.o stats_sock.o
 		$(LINK) $(ALL_LDFLAGS) $^ $(LDLIBS) -o $@
@@ -16,6 +33,8 @@ install: 	all
 		$(INSTALL) -m 644 cpacfstatsd.8 $(DESTDIR)$(MANDIR)/man8
 		$(INSTALL) -m 644 cpacfstats.1  $(DESTDIR)$(MANDIR)/man1
 
+endif
+
 clean:
 		rm -f *.o *~ cpacfstatsd cpacfstats
 

cpacfstats/cpacfstats.c

@@ -91,9 +91,20 @@ static int recv_answer(int s, int *ctr, int *state, uint64_t *value)
 	return rc;
 }
 
+static void print_virtual_counter_answer(int ctr, int state, uint64_t value)
+{
+	if (ctr == HOTPLUG_DETECTED) {
+		if (state >= 0 && value > 0)
+			printf(" hotplug detected\n");
+	}
+}
 
 static void print_answer(int ctr, int state, uint64_t value)
 {
+	if (ctr > ALL_COUNTER) {
+		print_virtual_counter_answer(ctr, state, value);
+		return;
+	}
 	if (state < 0)
 		printf(" %s counter: error state %d\n",
 		       counter_str[ctr], state);
@@ -232,6 +243,18 @@ int main(int argc, char *argv[])
 		}
 		print_answer(j, state, value);
 	}
+	/* receive hotplug state */
+	if (recv_answer(s, &j, &state, &value) != 0) {
+		eprint("Error on receiving hotplug state from daemon\n");
+		close(s);
+		exit(1);
+	}
+	if (state < 0) {
+		eprint("Received bad status code %d from daemon\n", state);
+		close(s);
+		exit(1);
+	}
+	print_answer(j, state, value);
 
 	/* close connection */
 	close(s);

cpacfstats/cpacfstats.h

@@ -23,7 +23,8 @@ int eprint(const char *format, ...);
 
 /*
  * Counter names
- * ALL_COUNTER must always be the last member of the enum
+ * ALL_COUNTER specifies the number of physical counters.  Virtual
+ * counters might be added afterwards.
  */
 enum ctr_e {
 	DES_FUNCTIONS = 0,
@@ -31,7 +32,8 @@ enum ctr_e {
 	SHA_FUNCTIONS,
 	PRNG_FUNCTIONS,
 	ECC_FUNCTIONS,
-	ALL_COUNTER
+	ALL_COUNTER,
+	HOTPLUG_DETECTED
 };
 
 enum type_e {
@@ -53,7 +55,7 @@ enum state_e {
 };
 
 /*
- * query send from clent to daemon
+ * query send from client to daemon
  * Consist of:
  * enum counter
  * enum command
@@ -102,17 +104,19 @@ struct msg {
 } __packed;
 
 int open_socket(int mode);
-int send_msg(int sfd, struct msg *m);
-int recv_msg(int sfd, struct msg *m);
+int send_msg(int sfd, struct msg *m, int timeout);
+int recv_msg(int sfd, struct msg *m, int timeout);
 
 /* perf_crypto.c */
 
 int  perf_init(void);
+void perf_stop(void);
 void perf_close(void);
 int  perf_enable_ctr(enum ctr_e ctr);
 int  perf_disable_ctr(enum ctr_e ctr);
 int  perf_reset_ctr(enum ctr_e ctr);
 int  perf_read_ctr(enum ctr_e ctr, uint64_t *value);
 int  perf_ecc_supported(void);
+int  perf_ctr_state(enum ctr_e ctr);
 
 #endif

cpacfstats/cpacfstatsd.c

@@ -28,6 +28,7 @@
 #include "lib/zt_common.h"
 #include "cpacfstats.h"
 
+static volatile int stopsig;
 
 static const char *const name = "cpacfstatsd";
 
@@ -43,9 +44,6 @@ static const char *const usage =
 
 static int daemonized;
 
-static int ctr_state[ALL_COUNTER];
-
-
 static int recv_query(int s, enum ctr_e *ctr, enum cmd_e *cmd)
 {
 	struct msg m;
@@ -89,32 +87,35 @@ static int send_answer(int s, int ctr, int state, uint64_t value)
 
 static int do_enable(int s, enum ctr_e ctr)
 {
-	uint64_t value;
+	uint64_t value = 0;
 	int i, rc = 0;
+	int state;
 
 	for (i = 0; i < ALL_COUNTER; i++) {
 		if (i == (int) ctr || ctr == ALL_COUNTER) {
-			if (ctr_state[i] == DISABLED) {
+			state = perf_ctr_state(i);
+			if (state == DISABLED) {
 				rc = perf_enable_ctr(i);
 				if (rc != 0) {
 					send_answer(s, i, rc, 0);
 					break;
 				}
-				ctr_state[i] = ENABLED;
+				state = ENABLED;
 			}
-			if (ctr_state[i] == UNSUPPORTED) {
-				send_answer(s, i, UNSUPPORTED, 0);
-			} else {
+			if (state != UNSUPPORTED) {
 				rc = perf_read_ctr(i, &value);
 				if (rc != 0) {
 					send_answer(s, i, rc, 0);
 					break;
 				}
-				send_answer(s, i, ENABLED, value);
 			}
+			send_answer(s, i, state, value);
 		}
 	}
-
+	if (rc == 0) {
+		rc = perf_read_ctr(HOTPLUG_DETECTED, &value);
+		send_answer(s, HOTPLUG_DETECTED, rc, value);
+	}
 	return rc;
 }
 
@@ -122,67 +123,75 @@ static int do_enable(int s, enum ctr_e ctr)
 static int do_disable(int s, enum ctr_e ctr)
 {
 	int i, rc = 0;
+	uint64_t value;
 
 	for (i = 0; i < ALL_COUNTER; i++) {
 		if (i == (int) ctr || ctr == ALL_COUNTER) {
-			if (ctr_state[i] == ENABLED) {
+			if (perf_ctr_state(i) == ENABLED) {
 				rc = perf_disable_ctr(i);
 				if (rc != 0) {
 					send_answer(s, i, rc, 0);
 					break;
 				}
-				ctr_state[i] = 0;
 			}
-			send_answer(s, i, ctr_state[i], 0);
+			send_answer(s, i, perf_ctr_state(i), 0);
 		}
 	}
-
+	if (rc == 0) {
+		rc = perf_read_ctr(HOTPLUG_DETECTED, &value);
+		send_answer(s, HOTPLUG_DETECTED, rc, value);
+	}
 	return rc;
 }
 
 
 static int do_reset(int s, enum ctr_e ctr)
 {
-	int i, rc = 0;
+	int i, rc = 0, state;
+	uint64_t value;
 
 	for (i = 0; i < ALL_COUNTER; i++) {
 		if (i == (int) ctr || ctr == ALL_COUNTER) {
-			if (ctr_state[i] == ENABLED) {
+			state = perf_ctr_state(i);
+			if (state == ENABLED) {
 				rc = perf_reset_ctr(i);
 				if (rc != 0) {
 					send_answer(s, i, rc, 0);
 					break;
 				}
-				send_answer(s, i, ENABLED, 0);
-			} else {
-				send_answer(s, i, ctr_state[i], 0);
 			}
+			send_answer(s, i, state, 0);
 		}
 	}
-
+	if (rc == 0) {
+		rc = perf_read_ctr(HOTPLUG_DETECTED, &value);
+		send_answer(s, HOTPLUG_DETECTED, rc, value);
+	}
 	return rc;
 }
 
 static int do_print(int s, enum ctr_e ctr)
 {
-	int i, rc = 0;
-	uint64_t value;
+	int i, rc = 0, state;
+	uint64_t value = 0;
 
 	for (i = 0; i < ALL_COUNTER; i++) {
 		if (i == (int) ctr || ctr == ALL_COUNTER) {
-			if (ctr_state[i] == ENABLED) {
+			state = perf_ctr_state(i);
+			if (state == ENABLED) {
 				rc = perf_read_ctr(i, &value);
 				if (rc != 0) {
 					send_answer(s, i, rc, 0);
 					break;
 				}
-				send_answer(s, i, ENABLED, value);
-			} else {
-				send_answer(s, i, ctr_state[i], 0);
 			}
+			send_answer(s, i, state, value);
 		}
 	}
-
+	if (rc == 0) {
+		rc = perf_read_ctr(HOTPLUG_DETECTED, &value);
+		send_answer(s, HOTPLUG_DETECTED, rc, value);
+	}
 	return rc;
 }
 
@@ -324,18 +333,8 @@ static void remove_pidfile(void)
 
 void signalhandler(int sig)
 {
-	if (sig == SIGTERM)
-		eprint("Caught signal SIGTERM, terminating...\n");
-	else if (sig == SIGINT)
-		eprint("Caught signal SIGINT, terminating...\n");
-	else
-		eprint("Caught signal %d, terminating...\n", sig);
-
-	remove_sock();
-	perf_close();
-	remove_pidfile();
-
-	exit(0);
+	perf_stop();
+	stopsig = sig;
 }
 
 
@@ -416,9 +415,6 @@ int main(int argc, char *argv[])
 	}
 	atexit(perf_close);
 
-	if (!perf_ecc_supported())
-		ctr_state[ECC_FUNCTIONS] = UNSUPPORTED;
-
 	sfd = open_socket(SERVER);
 	if (sfd < 0) {
 		eprint("Couldn't initialize server socket\n");
@@ -442,7 +438,7 @@ int main(int argc, char *argv[])
 
 	eprint("Running\n");
 
-	while (1) {
+	while (!stopsig) {
 		enum ctr_e ctr;
 		enum cmd_e cmd;
 		int s;
@@ -480,5 +476,13 @@ cleanup:
 		close(s);
 	}
 
+	if (stopsig == SIGTERM)
+		eprint("Caught signal SIGTERM, terminating...\n");
+	else if (stopsig == SIGINT)
+		eprint("Caught signal SIGINT, terminating...\n");
+	else
+		eprint("Caught signal %d, terminating...\n", stopsig);
+	remove_pidfile();
+
 	return 0;
 }

cpacfstats/perf_crypto.c

@@ -16,6 +16,9 @@
 #include <inttypes.h>
 #include <limits.h>
 #include <linux/perf_event.h>
+#include <pthread.h>
+#include <poll.h>
+#include <signal.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -23,14 +26,19 @@
 #include <sys/ioctl.h>
 #include <unistd.h>
 
+#include <libudev.h>
+
 #include "cpacfstats.h"
+#include "../include/lib/zt_common.h"
+
+#define NUM_COUNTER ALL_COUNTER
 
 /* correlation between counter and perf counter string */
 static const struct {
 	char pmu[20];
 	char pfm_name[60];
 	enum ctr_e ctr;
-} pmf_counter_name[ALL_COUNTER] = {
+} pmf_counter_name[NUM_COUNTER] = {
 	{"cpum_cf", "DEA_FUNCTIONS", DES_FUNCTIONS},
 	{"cpum_cf", "AES_FUNCTIONS", AES_FUNCTIONS},
 	{"cpum_cf", "SHA_FUNCTIONS", SHA_FUNCTIONS},
@@ -38,25 +46,71 @@ static const struct {
 	{"cpum_cf", "ECC_FUNCTION_COUNT", ECC_FUNCTIONS}
 };
 
-/*
- * We need one filedescriptor per CPU per counter.
- * So perf_init builds this:
- *
- * ctr_fds - is an array of pointers to file descriptor arrays.
- * Each file descriptor array has space for number of logical CPUs + 1
- * filedescriptors (int values). The last element of each file descriptor
- * array is always 0, assuming there will never appear a filedescriptor
- * with value 0:
- *
- * ctr_fds:
- *   ctr_fds[0]             -> [file descriptor 0] [fd1] ... [fd cpus-1][0]
- *   ctr_fds[1]             -> [file descriptor 0] [fd1] ... [fd cpus-1][0]
- *   ...
- *   ctr_fds[ALL_COUNTER-1] -> [file descriptor 0] [fd1] ... [fd cpus-1][0]
- */
-static int *ctr_fds[ALL_COUNTER];
+static struct pmf_data {
+	int pmutype;
+	int eventid;
+} pmf_counter_data[NUM_COUNTER];
 
-static int ecc_supported;
+struct percpucounter {
+	int                   ctr_fds[NUM_COUNTER];
+	unsigned int          cpunum;
+	struct percpucounter *next;
+};
+
+static struct percpucounter *root;
+pthread_mutex_t rootmux = PTHREAD_MUTEX_INITIALIZER;
+static volatile int hotplugdetected;
+static unsigned int enabledcounter;
+pthread_t hotplugthread;
+
+#define foreachcpu(PCPU) if (pthread_mutex_lock(&rootmux)) return -1; \
+	for ((PCPU) = root; (PCPU) != NULL; (PCPU) = (PCPU)->next)
+#define endforeachcpu() pthread_mutex_unlock(&rootmux)
+
+static int ctr_state[NUM_COUNTER];
+
+static volatile int stoprequested;
+
+static struct percpucounter *allocpercpucounter(unsigned int cpunum)
+{
+	struct percpucounter *ppc;
+
+	ppc = malloc(sizeof(struct percpucounter));
+	if (ppc) {
+		int i;
+
+		for (i = 0; i < NUM_COUNTER; ++i)
+			ppc->ctr_fds[i] = -1;
+		ppc->cpunum = cpunum;
+		ppc->next = NULL;
+	}
+	return ppc;
+}
+
+static void freepercpucounter(struct percpucounter *pcpu)
+{
+	int i;
+
+	for (i = 0; i < NUM_COUNTER; ++i)
+		(void)close(pcpu->ctr_fds[i]);
+	free(pcpu);
+}
+
+static struct percpucounter *findcpu(unsigned int cpunum, int unlinkflag)
+{
+	struct percpucounter **prev = &root, *walk = root;
+
+	while (walk) {
+		if (walk->cpunum == cpunum) {
+			if (unlinkflag)
+				*prev = walk->next;
+			return walk;
+		}
+		prev = &(walk->next);
+		walk = walk->next;
+	}
+	return NULL;
+}
 
 static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
 			    int cpu, int group_fd, unsigned long flags)
@@ -85,12 +139,10 @@ static int perf_counter_supported(const char *pmu, const char *counter)
 	return !access(buf, R_OK);
 }
 
-static int perf_event_encode(struct perf_event_attr *attr,
+static int perf_event_encode(int *pmutype, int *eventid,
 			     const char *pmu, const char *event)
 {
 	FILE *f;
-	int eventid;
-	int pmutype;
 	char buf[PATH_MAX];
 
 	if (snprintf(buf, PATH_MAX, "/sys/bus/event_source/devices/%s/events/%s",
@@ -104,7 +156,7 @@ static int perf_event_encode(struct perf_event_attr *attr,
 		       errno, strerror(errno));
 		return -1;
 	}
-	if (fscanf(f, "event=0x%x\n", &eventid) != 1) {
+	if (fscanf(f, "event=0x%x\n", eventid) != 1) {
 		fclose(f);
 		eprint("Event file %s has invalid format\n", buf);
 		return -1;
@@ -121,21 +173,185 @@ static int perf_event_encode(struct perf_event_attr *attr,
 		       errno, strerror(errno));
 		return -1;
 	}
-	if (fscanf(f, "%d\n", &pmutype) != 1) {
+	if (fscanf(f, "%d\n", pmutype) != 1) {
 		fclose(f);
 		eprint("Type file %s has invalid format\n", buf);
 		return -1;
 	}
-	attr->type = pmutype;
-	attr->config = eventid;
 	return 0;
 }
 
+static int activatecpu(unsigned int cpu)
+{
+	struct perf_event_attr pfm_event;
+	struct percpucounter *ppc;
+	int fd, i, rc = 0;
+
+	ppc = allocpercpucounter(cpu);
+	if (ppc == NULL) {
+		eprint("Failed to allocate per cpu counter data");
+		return -1;
+	}
+	if (pthread_mutex_lock(&rootmux)) {
+		freepercpucounter(ppc);
+		return -1;
+	}
+	for (i = 0; i < NUM_COUNTER; ++i) {
+		if (ctr_state[i] == UNSUPPORTED)
+			continue;
+		memset(&pfm_event, 0, sizeof(pfm_event));
+		pfm_event.size = sizeof(pfm_event);
+		pfm_event.type = pmf_counter_data[i].pmutype;
+		pfm_event.config = pmf_counter_data[i].eventid;
+
+		/* fetch file descriptor for this perf event
+		 * the counter event should start disabled
+		 */
+		pfm_event.disabled = ctr_state[i] == DISABLED;
+		fd = perf_event_open(
+			&pfm_event,
+			-1,  /* pid -1 means all processes */
+			cpu,
+			-1,  /* group filedescriptor */
+			0);  /* flags */
+		if (fd < 0) {
+			eprint("Perf_event_open() failed with errno=%d [%s]\n",
+				errno, strerror(errno));
+			rc = -1;
+			ctr_state[i] = UNSUPPORTED;
+		} else {
+			ppc->ctr_fds[i] = fd;
+		}
+	}
+	ppc->next = root;
+	root = ppc;
+	if (enabledcounter)
+		hotplugdetected = 1;
+	pthread_mutex_unlock(&rootmux);
+	return rc;
+}
+
+static void deactivatecpu(unsigned int cpunum)
+{
+	struct percpucounter *pcpu;
+	int i;
+
+	if (pthread_mutex_lock(&rootmux))
+		return;
+	pcpu = findcpu(cpunum, 1);
+	if (pcpu != NULL) {
+		for (i = 0; i < NUM_COUNTER; ++i)
+			(void)close(pcpu->ctr_fds[i]);
+		free(pcpu);
+		if (enabledcounter)
+			hotplugdetected = 1;
+	}
+	pthread_mutex_unlock(&rootmux);
+}
+
+static int addallcpus(void)
+{
+	unsigned int start, end;
+	int scanned, rc = 0;
+	FILE *fp;
+
+	/* comma separated list of intervals */
+	if ((fp = fopen("/sys/devices/system/cpu/online", "r")) == NULL) {
+		eprint("Failed to get online cpus (%d:%s)\n",
+			errno, strerror(errno));
+		return -1;
+	}
+		
+	while (!feof(fp)) {
+		/* scan all intervals of online cpus */
+		scanned = fscanf(fp, "%u-%u", &start, &end);
+		/* take care of singleton intervals */
+		if (scanned == 1)
+			end = start;
+		for (; start <= end; ++start) {
+			if (activatecpu(start)) {
+				rc = -1;
+				goto out;
+			}
+		}
+		/* Skip comma separator */
+		(void)fgetc(fp);
+	}
+out:
+	fclose(fp);
+	return rc;
+}
+
+static int perf_load_counter_data(void)
+{
+	int i, res = 0;
+
+	for (i = 0; i < NUM_COUNTER; ++i) {
+		if (ctr_state[i] != UNSUPPORTED)
+			res |= perf_event_encode(&pmf_counter_data[i].pmutype,
+						 &pmf_counter_data[i].eventid,
+						 pmf_counter_name[i].pmu,
+						 pmf_counter_name[i].pfm_name);
+	}
+	return res;
+}
+
+static void *hotplughandler(void *UNUSED(unused))
+{
+	struct udev *hotplug;
+	struct udev_monitor *monitor;
+	struct pollfd item;
+	
+	hotplug = udev_new();
+	if (!hotplug) {
+		eprint("Failed to create hotplug device\n");
+		return NULL;
+	}
+	monitor = udev_monitor_new_from_netlink(hotplug, "udev");
+	udev_monitor_filter_add_match_subsystem_devtype(monitor, "cpu", NULL);
+	udev_monitor_enable_receiving(monitor);
+	item.fd = udev_monitor_get_fd(monitor);
+	item.events = POLLIN;
+	item.revents = 0;
+	while (!stoprequested) {
+		struct udev_device *dev;
+		const char *path, *action;
+		unsigned int cpunum;
+		int rc, on, off;
+
+		errno = 0;
+		rc = poll(&item, 1, -1);
+		if (rc == -1) {
+			if (errno == EINTR)
+				continue;
+			break;
+		}
+		dev = udev_monitor_receive_device(monitor);
+		if (dev == NULL)
+			continue;
+		action = udev_device_get_action(dev);
+		if (action == NULL)
+			continue;
+		off = strcmp(action, "offline") == 0;
+		on = strcmp(action, "online") == 0;
+		if (!on && !off)
+			continue;
+		path = udev_device_get_devpath(dev);
+		if (sscanf(path, "/devices/system/cpu/cpu%u", &cpunum) != 1)
+			continue;
+		if (on && activatecpu(cpunum))
+			eprint("Failed to attach to hotplugged CPU %u\n", cpunum);
+		if (off)
+			deactivatecpu(cpunum);
+	}
+	udev_monitor_unref(monitor);
+	udev_unref(hotplug);
+	return NULL;
+}
+
 int perf_init(void)
 {
-	int i, cpus, ctr, cpu, *fds;
-
-	memset(ctr_fds, 0, sizeof(ctr_fds));
+	int ecc_supported;
 
 	/*  initialize performance monitoring library */
 	if (!perf_supported()) {
@@ -145,97 +361,44 @@ int perf_init(void)
 
 	/* Check if ECC is supported on current hardware */
 	ecc_supported = perf_counter_supported("cpum_cf", "ECC_FUNCTION_COUNT");
+	if (!ecc_supported)
+		ctr_state[ECC_FUNCTIONS] = UNSUPPORTED;
 
-	/* get number of logical processors */
-	cpus = sysconf(_SC_NPROCESSORS_ONLN);
-
-	/* for each counter */
-	for (ctr = 0; ctr < ALL_COUNTER; ctr++) {
-
-		/* Skip ECC counters completely if unsupported */
-		if (ctr == ECC_FUNCTIONS && !ecc_supported)
-			continue;
-
-		/*
-		 * allocate an array of ints to store for each CPU
-		 * one filedescriptor + a terminating 0
-		 */
-		fds = (int *) calloc(sizeof(int), cpus+1);
-		if (!fds) {
-			eprint("Malloc() of %d byte failed, errno=%d [%s]\n",
-			       (int)(sizeof(int) * (cpus+1)),
-			       errno, strerror(errno));
-			return -1;
-		}
-
-		ctr_fds[ctr] = fds;
-
-		/* search for the counter's corresponding pfm name */
-		for (i = ALL_COUNTER-1; i >= 0; i--)
-			if ((int) pmf_counter_name[i].ctr == ctr)
-				break;
-		if (i < 0) {
-			eprint("Pfm ctr name not found for counter %d, please adjust pmf_counter_name[] in %s\n",
-			       ctr, __FILE__);
-			return -1;
-		}
-
-		for (cpu = 0; cpu < cpus; cpu++) {
-			struct perf_event_attr pfm_event;
-			int fd;
-
-			memset(&pfm_event, 0, sizeof(pfm_event));
-			pfm_event.size = sizeof(pfm_event);
-			if (perf_event_encode(&pfm_event,
-					      pmf_counter_name[i].pmu,
-					      pmf_counter_name[i].pfm_name)) {
-				eprint("Failed to initialize counter %s for pmu %s\n",
-				       pmf_counter_name[i].pfm_name,
-				       pmf_counter_name[i].pmu);
-				return -1;
-			}
-
-			/* fetch file descriptor for this perf event
-			 * the counter event should start disabled
-			 */
-			pfm_event.disabled = 1;
-			fd = perf_event_open(
-				&pfm_event,
-				-1,  /* pid -1 means all processes */
-				cpu,
-				-1,  /* group filedescriptor */
-				0);  /* flags */
-			if (fd < 0) {
-				eprint("Perf_event_open() failed with errno=%d [%s]\n",
-				       errno, strerror(errno));
-				return -1;
-			}
-			fds[cpu] = fd;
-		}
+	if (perf_load_counter_data())
+		return -1;
+	if (pthread_create(&hotplugthread, NULL, hotplughandler, NULL)) {
+		eprint("Failed to start hotplug handler thread\n");
+		return -1;
 	}
+	return addallcpus();
+}
 
-	return 0;
+
+void perf_stop(void)
+{
+	stoprequested = 1;
 }
 
 
 void perf_close(void)
 {
-	int ctr, *fds;
+	struct percpucounter *walk, *next;
 
-	for (ctr = 0; ctr < ALL_COUNTER; ctr++) {
-		for (fds = ctr_fds[ctr]; fds && *fds; fds++) {
-			close(*fds);
-			*fds = 0;
-		}
-		free(ctr_fds[ctr]);
-		ctr_fds[ctr] = NULL;
+	pthread_kill(hotplugthread, SIGINT);
+	pthread_join(hotplugthread, NULL);
+	walk = root;
+	while (walk) {
+		next = walk->next;
+		freepercpucounter(walk);
+		walk = next;
 	}
 }
 
 
 int perf_enable_ctr(enum ctr_e ctr)
 {
-	int *fds, ec, rc = 0;
+	struct percpucounter *pcpu;
+	int ec, rc = 0;
 
 	if (ctr == ALL_COUNTER) {
 		for (ctr = 0; ctr < ALL_COUNTER; ctr++) {
@@ -243,15 +406,20 @@ int perf_enable_ctr(enum ctr_e ctr)
 			if (rc != 0)
 				return rc;
 		}
-	} else {
-		for (fds = ctr_fds[ctr]; fds && *fds; fds++) {
-			ec = ioctl(*fds, PERF_EVENT_IOC_ENABLE, 0);
+	} else if (ctr < ALL_COUNTER) {
+		foreachcpu(pcpu) {
+			ec = ioctl(pcpu->ctr_fds[ctr], PERF_EVENT_IOC_ENABLE, 0);
 			if (ec < 0) {
 				eprint("Ioctl(PERF_EVENT_IOC_ENABLE) failed with errno=%d [%s]\n",
 				       errno, strerror(errno));
 				rc = -1;
 			}
 		}
+		ctr_state[ctr] = ENABLED;
+		++enabledcounter;
+		endforeachcpu();
+	} else {
+		rc = -1;
 	}
 
 	return rc;
@@ -260,7 +428,8 @@ int perf_enable_ctr(enum ctr_e ctr)
 
 int perf_disable_ctr(enum ctr_e ctr)
 {
-	int *fds, ec, rc = 0;
+	struct percpucounter *pcpu;
+	int ec, rc = 0;
 
 	if (ctr == ALL_COUNTER) {
 		for (ctr = 0; ctr < ALL_COUNTER; ctr++) {
@@ -268,15 +437,22 @@ int perf_disable_ctr(enum ctr_e ctr)
 			if (rc != 0)
 				return rc;
 		}
-	} else {
-		for (fds = ctr_fds[ctr]; fds && *fds; fds++) {
-			ec = ioctl(*fds, PERF_EVENT_IOC_DISABLE, 0);
+	} else if (ctr < ALL_COUNTER) {
+		foreachcpu(pcpu) {
+			ec = ioctl(pcpu->ctr_fds[ctr], PERF_EVENT_IOC_DISABLE, 0);
 			if (ec < 0) {
 				eprint("Ioctl(PERF_EVENT_IOC_DISABLE) failed with errno=%d [%s]\n",
 				       errno, strerror(errno));
 				rc = -1;
 			}
 		}
+		ctr_state[ctr] = DISABLED;
+		--enabledcounter;
+		if (enabledcounter == 0)
+			hotplugdetected = 0;
+		endforeachcpu();
+	} else {
+		rc = -1;
 	}
 
 	return rc;
@@ -285,7 +461,8 @@ int perf_disable_ctr(enum ctr_e ctr)
 
 int perf_reset_ctr(enum ctr_e ctr)
 {
-	int *fds, ec, rc = 0;
+	struct percpucounter *pcpu;
+	int ec, rc = 0;
 
 	if (ctr == ALL_COUNTER) {
 		for (ctr = 0; ctr < ALL_COUNTER; ctr++) {
@@ -293,15 +470,18 @@ int perf_reset_ctr(enum ctr_e ctr)
 			if (rc != 0)
 				return rc;
 		}
-	} else {
-		for (fds = ctr_fds[ctr]; fds && *fds; fds++) {
-			ec = ioctl(*fds, PERF_EVENT_IOC_RESET, 0);
+	} else if (ctr < ALL_COUNTER) {
+		foreachcpu(pcpu) {
+			ec = ioctl(pcpu->ctr_fds[ctr], PERF_EVENT_IOC_RESET, 0);
 			if (ec < 0) {
 				eprint("Ioctl(PERF_EVENT_IOC_RESET) failed with errno=%d [%s]\n",
 				       errno, strerror(errno));
 				rc = -1;
 			}
 		}
+		endforeachcpu();
+	} else {
+		rc = -1;
 	}
 
 	return rc;
@@ -310,15 +490,22 @@ int perf_reset_ctr(enum ctr_e ctr)
 
 int perf_read_ctr(enum ctr_e ctr, uint64_t *value)
 {
-	int *fds, ec, rc = -1;
+	struct percpucounter *pcpu;
+	int ec, rc = -1;
 	uint64_t val;
 
 	if (!value)
 		return -1;
+	if (ctr == HOTPLUG_DETECTED) {
+		*value = hotplugdetected;
+		return 0;
+	}
+	if (ctr >= ALL_COUNTER)
+		return -1;
 	*value = 0;
 
-	for (fds = ctr_fds[ctr]; fds && *fds; fds++) {
-		ec = read(*fds, &val, sizeof(val));
+	foreachcpu(pcpu) {
+		ec = read(pcpu->ctr_fds[ctr], &val, sizeof(val));
 		if (ec != sizeof(val)) {
 			eprint("Read() on perf file descriptor failed with errno=%d [%s]\n",
 			       errno, strerror(errno));
@@ -327,11 +514,18 @@ int perf_read_ctr(enum ctr_e ctr, uint64_t *value)
 			rc = 0;
 		}
 	}
+	endforeachcpu();
 
 	return rc;
 }
 
 int  perf_ecc_supported(void)
 {
-	return ecc_supported;
+	return ctr_state[ECC_FUNCTIONS] != UNSUPPORTED;
+}
+
+int perf_ctr_state(enum ctr_e ctr) {
+	if (ctr < ALL_COUNTER)
+		return ctr_state[ctr];
+	return UNSUPPORTED;
 }

cpacfstats/stats_sock.c

@@ -13,6 +13,7 @@
 #include <endian.h>
 #include <errno.h>
 #include <grp.h>
+#include <poll.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <sys/socket.h>
@@ -244,7 +245,7 @@ int recv_msg(int sfd, struct msg *m, int timeout)
 		return -1;
 	}
 
-	n = timeout ? __timedread(sfd, ((char *)m) + sizeof(m->head), len) :
+	n = timeout ? __timedread(sfd, ((char *)m) + sizeof(m->head), len, timeout) :
 		__read(sfd, ((char *)m) + sizeof(m->head), len);
 	if (n != len) {
 		eprint("Recv() error: recv()=%d expected %d, errno=%d [%s]\n",