KVM: ppc: Implement in-kernel exit timing statistics
[pandora-kernel.git] / arch / powerpc / kvm / timing.c
diff --git a/arch/powerpc/kvm/timing.c b/arch/powerpc/kvm/timing.c
new file mode 100644 (file)
index 0000000..f42d272
--- /dev/null
@@ -0,0 +1,262 @@
+/*
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
+ *
+ * Copyright IBM Corp. 2007
+ *
+ * Authors: Hollis Blanchard <hollisb@us.ibm.com>
+ *          Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>
+ */
+
+#include <linux/kvm_host.h>
+#include <linux/fs.h>
+#include <linux/seq_file.h>
+#include <linux/debugfs.h>
+#include <linux/uaccess.h>
+
+#include "timing.h"
+#include <asm/time.h>
+#include <asm-generic/div64.h>
+
+void kvmppc_init_timing_stats(struct kvm_vcpu *vcpu)
+{
+       int i;
+
+       /* pause guest execution to avoid concurrent updates */
+       local_irq_disable();
+       mutex_lock(&vcpu->mutex);
+
+       vcpu->arch.last_exit_type = 0xDEAD;
+       for (i = 0; i < __NUMBER_OF_KVM_EXIT_TYPES; i++) {
+               vcpu->arch.timing_count_type[i] = 0;
+               vcpu->arch.timing_max_duration[i] = 0;
+               vcpu->arch.timing_min_duration[i] = 0xFFFFFFFF;
+               vcpu->arch.timing_sum_duration[i] = 0;
+               vcpu->arch.timing_sum_quad_duration[i] = 0;
+       }
+       vcpu->arch.timing_last_exit = 0;
+       vcpu->arch.timing_exit.tv64 = 0;
+       vcpu->arch.timing_last_enter.tv64 = 0;
+
+       mutex_unlock(&vcpu->mutex);
+       local_irq_enable();
+}
+
+static void add_exit_timing(struct kvm_vcpu *vcpu,
+                                       u64 duration, int type)
+{
+       u64 old;
+
+       do_div(duration, tb_ticks_per_usec);
+       if (unlikely(duration > 0xFFFFFFFF)) {
+               printk(KERN_ERR"%s - duration too big -> overflow"
+                       " duration %lld type %d exit #%d\n",
+                       __func__, duration, type,
+                       vcpu->arch.timing_count_type[type]);
+               return;
+       }
+
+       vcpu->arch.timing_count_type[type]++;
+
+       /* sum */
+       old = vcpu->arch.timing_sum_duration[type];
+       vcpu->arch.timing_sum_duration[type] += duration;
+       if (unlikely(old > vcpu->arch.timing_sum_duration[type])) {
+               printk(KERN_ERR"%s - wrap adding sum of durations"
+                       " old %lld new %lld type %d exit # of type %d\n",
+                       __func__, old, vcpu->arch.timing_sum_duration[type],
+                       type, vcpu->arch.timing_count_type[type]);
+       }
+
+       /* square sum */
+       old = vcpu->arch.timing_sum_quad_duration[type];
+       vcpu->arch.timing_sum_quad_duration[type] += (duration*duration);
+       if (unlikely(old > vcpu->arch.timing_sum_quad_duration[type])) {
+               printk(KERN_ERR"%s - wrap adding sum of squared durations"
+                       " old %lld new %lld type %d exit # of type %d\n",
+                       __func__, old,
+                       vcpu->arch.timing_sum_quad_duration[type],
+                       type, vcpu->arch.timing_count_type[type]);
+       }
+
+       /* set min/max */
+       if (unlikely(duration < vcpu->arch.timing_min_duration[type]))
+               vcpu->arch.timing_min_duration[type] = duration;
+       if (unlikely(duration > vcpu->arch.timing_max_duration[type]))
+               vcpu->arch.timing_max_duration[type] = duration;
+}
+
+void kvmppc_update_timing_stats(struct kvm_vcpu *vcpu)
+{
+       u64 exit = vcpu->arch.timing_last_exit;
+       u64 enter = vcpu->arch.timing_last_enter.tv64;
+
+       /* save exit time, used next exit when the reenter time is known */
+       vcpu->arch.timing_last_exit = vcpu->arch.timing_exit.tv64;
+
+       if (unlikely(vcpu->arch.last_exit_type == 0xDEAD || exit == 0))
+               return; /* skip incomplete cycle (e.g. after reset) */
+
+       /* update statistics for average and standard deviation */
+       add_exit_timing(vcpu, (enter - exit), vcpu->arch.last_exit_type);
+       /* enter -> timing_last_exit is time spent in guest - log this too */
+       add_exit_timing(vcpu, (vcpu->arch.timing_last_exit - enter),
+                       TIMEINGUEST);
+}
+
+static const char *kvm_exit_names[__NUMBER_OF_KVM_EXIT_TYPES] = {
+       [MMIO_EXITS] =                  "MMIO",
+       [DCR_EXITS] =                   "DCR",
+       [SIGNAL_EXITS] =                "SIGNAL",
+       [ITLB_REAL_MISS_EXITS] =        "ITLBREAL",
+       [ITLB_VIRT_MISS_EXITS] =        "ITLBVIRT",
+       [DTLB_REAL_MISS_EXITS] =        "DTLBREAL",
+       [DTLB_VIRT_MISS_EXITS] =        "DTLBVIRT",
+       [SYSCALL_EXITS] =               "SYSCALL",
+       [ISI_EXITS] =                   "ISI",
+       [DSI_EXITS] =                   "DSI",
+       [EMULATED_INST_EXITS] =         "EMULINST",
+       [EMULATED_MTMSRWE_EXITS] =      "EMUL_WAIT",
+       [EMULATED_WRTEE_EXITS] =        "EMUL_WRTEE",
+       [EMULATED_MTSPR_EXITS] =        "EMUL_MTSPR",
+       [EMULATED_MFSPR_EXITS] =        "EMUL_MFSPR",
+       [EMULATED_MTMSR_EXITS] =        "EMUL_MTMSR",
+       [EMULATED_MFMSR_EXITS] =        "EMUL_MFMSR",
+       [EMULATED_TLBSX_EXITS] =        "EMUL_TLBSX",
+       [EMULATED_TLBWE_EXITS] =        "EMUL_TLBWE",
+       [EMULATED_RFI_EXITS] =          "EMUL_RFI",
+       [DEC_EXITS] =                   "DEC",
+       [EXT_INTR_EXITS] =              "EXTINT",
+       [HALT_WAKEUP] =                 "HALT",
+       [USR_PR_INST] =                 "USR_PR_INST",
+       [FP_UNAVAIL] =                  "FP_UNAVAIL",
+       [DEBUG_EXITS] =                 "DEBUG",
+       [TIMEINGUEST] =                 "TIMEINGUEST"
+};
+
+static int kvmppc_exit_timing_show(struct seq_file *m, void *private)
+{
+       struct kvm_vcpu *vcpu = m->private;
+       int i;
+       u64 min, max;
+
+       for (i = 0; i < __NUMBER_OF_KVM_EXIT_TYPES; i++) {
+               if (vcpu->arch.timing_min_duration[i] == 0xFFFFFFFF)
+                       min = 0;
+               else
+                       min = vcpu->arch.timing_min_duration[i];
+               if (vcpu->arch.timing_max_duration[i] == 0)
+                       max = 0;
+               else
+                       max = vcpu->arch.timing_max_duration[i];
+
+               seq_printf(m, "%12s: count %10d min %10lld "
+                       "max %10lld sum %20lld sum_quad %20lld\n",
+                       kvm_exit_names[i], vcpu->arch.timing_count_type[i],
+                       vcpu->arch.timing_min_duration[i],
+                       vcpu->arch.timing_max_duration[i],
+                       vcpu->arch.timing_sum_duration[i],
+                       vcpu->arch.timing_sum_quad_duration[i]);
+       }
+       return 0;
+}
+
+static ssize_t kvmppc_exit_timing_write(struct file *file,
+                                      const char __user *user_buf,
+                                      size_t count, loff_t *ppos)
+{
+       size_t len;
+       int err;
+       const char __user *p;
+       char c;
+
+       len = 0;
+       p = user_buf;
+       while (len < count) {
+               if (get_user(c, p++))
+                       err = -EFAULT;
+               if (c == 0 || c == '\n')
+                       break;
+               len++;
+       }
+
+       if (len > 1) {
+               err = -EINVAL;
+               goto done;
+       }
+
+       if (copy_from_user(&c, user_buf, sizeof(c))) {
+               err = -EFAULT;
+               goto done;
+       }
+
+       if (c == 'c') {
+               struct seq_file *seqf = (struct seq_file *)file->private_data;
+               struct kvm_vcpu *vcpu = seqf->private;
+               /* write does not affect out buffers previsously generated with
+                * show. Seq file is locked here to prevent races of init with
+                * a show call */
+               mutex_lock(&seqf->lock);
+               kvmppc_init_timing_stats(vcpu);
+               mutex_unlock(&seqf->lock);
+               err = count;
+       } else {
+               err = -EINVAL;
+               goto done;
+       }
+
+done:
+       return err;
+}
+
+static int kvmppc_exit_timing_open(struct inode *inode, struct file *file)
+{
+       return single_open(file, kvmppc_exit_timing_show, inode->i_private);
+}
+
+static struct file_operations kvmppc_exit_timing_fops = {
+       .owner   = THIS_MODULE,
+       .open    = kvmppc_exit_timing_open,
+       .read    = seq_read,
+       .write   = kvmppc_exit_timing_write,
+       .llseek  = seq_lseek,
+       .release = single_release,
+};
+
+void kvmppc_create_vcpu_debugfs(struct kvm_vcpu *vcpu, unsigned int id)
+{
+       static char dbg_fname[50];
+       struct dentry *debugfs_file;
+
+       snprintf(dbg_fname, sizeof(dbg_fname), "vm%u_vcpu%03u_timing",
+                current->pid, id);
+       debugfs_file = debugfs_create_file(dbg_fname, 0666,
+                                       kvm_debugfs_dir, vcpu,
+                                       &kvmppc_exit_timing_fops);
+
+       if (!debugfs_file) {
+               printk(KERN_ERR"%s: error creating debugfs file %s\n",
+                       __func__, dbg_fname);
+               return;
+       }
+
+       vcpu->arch.debugfs_exit_timing = debugfs_file;
+}
+
+void kvmppc_remove_vcpu_debugfs(struct kvm_vcpu *vcpu)
+{
+       if (vcpu->arch.debugfs_exit_timing) {
+               debugfs_remove(vcpu->arch.debugfs_exit_timing);
+               vcpu->arch.debugfs_exit_timing = NULL;
+       }
+}