Palacios Public Git Repository

To checkout Palacios execute

  git clone http://v3vee.org/palacios/palacios.web/palacios.git
This will give you the master branch. You probably want the devel branch or one of the release branches. To switch to the devel branch, simply execute
  cd palacios
  git checkout --track -b devel origin/devel
The other branches are similar.


P-State control - minor tweaks and change to proc interface
Peter Dinda [Sun, 31 Aug 2014 20:10:22 +0000 (15:10 -0500)]
linux_module/iface-pstate-ctrl.c

index 97a38dd..1559679 100644 (file)
 #include "linux-exts.h"
 
 /*
-   This P-STATE control implementation includes:
+   This P-STATE control implementation includes the following modes.
+   You can switch between modes at any time.
 
-   - Direct control of Intel and AMD processor pstates
-   - External control of processor states via Linux (unimplemented)
    - Internal control of processor states in Palacios (handoff from Linux)
+     When Palacios acuires this control, this module disables Linux cpufreq control
+     and allows code within Palacios unfettered access to the DVFS hardware. 
+   - Direct control of Intel and AMD processor pstates using code in this module
+     When you acquire this control, this module disables Linux cpufreq control
+     and directly programs the processor itself in response to your requests
+   - External control of processor states via Linux 
+     When you acuire this control, this module uses the Linux cpufreq control
+     to program the processor on your behelf
+   - Host control of processor stastes
+     This is the normal mode of DVFS control (e.g., Linux cpufreq)
 
    Additionally, it provides a user-space interface for manipulating
    p-state regardless of the host's functionality.  This includes
    an ioctl for commanding the implementation and a /proc file for 
-   showing current status and capabilities.
+   showing current status and capabilities.  From user space, you can
+   use the Direct, External, and Host modes.  
 
-   What we mean by "pstate" here is the processor's internal
+   What we mean by "p-state" here is the processor's internal
    configuration.   For AMD, this is defined as being the same as
    the ACPI-defined p-state.  For Intel, it is not.  There, it is the 
-   contents of the perf ctl MSR, which, often, is the frequency id 
-   and voltage id (the multipliers).
+   contents of the perf ctl MSR, which is opaque.   We try hard to 
+   provide "p-states" that go from 0...max, by analogy or equivalence
+   to the ACPI p-states. 
 
 */
 
@@ -78,11 +89,11 @@ struct pstate_core_info {
     uint32_t mode;
 
     // Apply if we are under the DIRECT state
-    uint8_t cur_pstate;
-    uint8_t max_pstate;
-    uint8_t min_pstate;
+    uint64_t cur_pstate;
+    uint64_t max_pstate;
+    uint64_t min_pstate;
 
-    uint8_t cur_hw_pstate;
+    uint64_t cur_hw_pstate;
 
     // Apply if we are under the EXTERNAL state
     uint64_t set_freq_khz; // this is the frequency we're hoping to get
@@ -195,6 +206,10 @@ struct p_state_ctl_reg_amd {
 /* CPUID Fn8000_0007_EDX[HwPstate(7)] = 1 */
 static uint8_t supports_pstates_amd (void)
 {
+    int i;
+    int mapwrong=0;
+    int amd_num_pstates;
+
     uint32_t eax, ebx, ecx, edx;
 
     cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
@@ -211,6 +226,20 @@ static uint8_t supports_pstates_amd (void)
             machine_state.have_feedback,
             machine_state.have_pstate_hw_coord);
 
+    amd_num_pstates = get_cpu_var(processors)->performance->state_count;
+    if (amd_num_pstates) { 
+       for (i=0;i<amd_num_pstates;i++) { 
+           INFO("P-State: %u: freq=%llu ctrl=%llx%s\n",
+                i, 
+                get_cpu_var(processors)->performance->states[i].core_frequency*1000,
+                get_cpu_var(processors)->performance->states[i].control,
+                get_cpu_var(processors)->performance->states[i].control != i ? (mapwrong=1, " ALERT - CTRL MAPPING NOT 1:1") : "");
+       }
+    }
+    if (mapwrong) { 
+       ERROR("P-State: AMD: mapping of pstate and control is not 1:1 on this processor - we will probably not work corrrectly\n");
+    }
+
     return machine_state.have_pstate;
 
 
@@ -245,6 +274,12 @@ static uint64_t get_pstate_amd(void)
 static void set_pstate_amd(uint64_t p)
 {
     struct p_state_ctl_reg_amd pctl;
+
+    if (p>get_cpu_var(core_state).max_pstate) { 
+       p=get_cpu_var(core_state).max_pstate;
+    }
+    put_cpu_var(core_state);
+
     pctl.val = 0;
     pctl.reg.cmd = p;
 
@@ -299,7 +334,7 @@ static struct pstate_core_funcs amd_funcs =
    This implementation uses SpeedStep, but does check
    to see if the other features (MPERF/APERF, Turbo/IDA, HWP)
    are available.
-   */
+*/
 
 /* Intel System Programmer's Manual Vol. 3B, 14-2 */
 #define MSR_MPERF_IA32         0x000000e7
@@ -425,6 +460,9 @@ static uint8_t supports_pstates_intel(void)
     machine_state.have_mwait_int =  !!(ecx & 1<<1);
 
 
+    // Note we test all the available hardware features documented as of August 2014
+    // We are only currently using speed_step, however.
+
     INFO("P-State: Intel: Speedstep=%d, PstateHWCoord=%d, Opportunistic=%d PolicyHint=%d HWP=%d HDC=%d, MwaitExt=%d MwaitInt=%d \n",
             machine_state.have_speedstep, 
             machine_state.have_pstate_hw_coord, 
@@ -553,12 +591,12 @@ static void set_pstate_intel(uint64_t p)
     // fid bits
 
     rdmsrl(MSR_PERF_CTL_IA32, val);
-    INFO("P-State: Pre-Set: 0x%llx\n", val);
+    //INFO("P-State: Pre-Set: 0x%llx\n", val);
 
     val &= ~0xffffULL;
     val |= ctrl & 0xffffULL;
 
-    INFO("P-State: Set: 0x%llx\n", val);
+    //INFO("P-State: Set: 0x%llx\n", val);
 
     wrmsrl(MSR_PERF_CTL_IA32, val);
 
@@ -949,7 +987,7 @@ static int linux_setup_palacios_governor(void)
 
 
 
-static int linux_get_pstate(void)
+static uint64_t linux_get_pstate(void)
 {
     struct cpufreq_policy * policy = NULL;
     struct cpufreq_frequency_table *table;
@@ -988,8 +1026,9 @@ static int linux_get_pstate(void)
 }
 
 
-static int linux_get_freq(void)
+static uint64_t linux_get_freq(void)
 {
+    uint64_t freq;
     struct cpufreq_policy * policy = NULL;
     unsigned int cpu = get_cpu();
     put_cpu();
@@ -1005,7 +1044,11 @@ static int linux_get_freq(void)
         return -1;
     }
 
-    return policy->cur;
+    freq=policy->cur;
+
+    palacios_free(policy);
+
+    return freq;
 }
 
 static void  
@@ -1052,7 +1095,7 @@ out:
 } 
 
 
-static int linux_set_pstate(uint8_t p)
+static int linux_set_pstate(uint64_t p)
 {
     struct cpufreq_policy * policy = NULL;
     struct cpufreq_frequency_table *table;
@@ -1203,10 +1246,9 @@ static void init_core(void)
 {
     unsigned cpu;
     struct cpufreq_policy *p;
-    unsigned int i;
 
 
-    DEBUG("P-State Core Init\n");
+    //DEBUG("P-State Core Init\n");
 
     get_cpu_var(core_state).mode = V3_PSTATE_HOST_CONTROL;
     get_cpu_var(core_state).cur_pstate = 0;
@@ -1236,6 +1278,7 @@ static void init_core(void)
         get_cpu_var(core_state).cur_freq_khz=p->cur; } cpufreq_cpu_put(p); 
     put_cpu_var(core_state);
 
+    /*
     for (i=0;i<get_cpu_var(processors)->performance->state_count; i++) { 
         INFO("P-State: %u: freq=%llu ctrl=%llx",
                i, 
@@ -1243,6 +1286,7 @@ static void init_core(void)
                get_cpu_var(processors)->performance->states[i].control);
    }
    put_cpu_var(processors);
+    */
 }
 
 
@@ -1467,8 +1511,6 @@ static int switch_from_internal(void)
 
     if (get_cpu_var(core_state).have_cpufreq) { 
         put_cpu_var(core_state);
-        // ERROR("Unimplemented: switch from internal on machine with cpu freq - will just pretend to do so\n");
-        // The implementation would switch back to default policy and governor
         linux_restore_defaults();
     } else {
         put_cpu_var(core_state);
@@ -1577,31 +1619,19 @@ static int pstate_show(struct seq_file * file, void * v)
         palacios_xcall(cpu,update_hw_pstate,0);
     }
 
-    seq_printf(file, "Arch:\t%s\nPStates:\t%s\n\n",
-            machine_state.arch==INTEL ? "Intel" : 
-            machine_state.arch==AMD ? "AMD" : "Other",
-            machine_state.supports_pstates ? "Yes" : "No");
-
     for (cpu=0;cpu<numcpus;cpu++) { 
         struct pstate_core_info *s = &per_cpu(core_state,cpu);
-        seq_printf(file,"pcore %u: hw pstate 0x%x mode %s of [ host ",cpu,
+        seq_printf(file,"pcore %u: hw pstate 0x%llx mode %s ",cpu,
                 s->cur_hw_pstate,
                 s->mode==V3_PSTATE_HOST_CONTROL ? "host" :
                 s->mode==V3_PSTATE_EXTERNAL_CONTROL ? "external" :
                 s->mode==V3_PSTATE_DIRECT_CONTROL ? "direct" : 
                 s->mode==V3_PSTATE_INTERNAL_CONTROL ? "internal" : "UNKNOWN");
-        if (s->have_cpufreq) { 
-            seq_printf(file,"external ");
-        }
-        if (machine_state.supports_pstates) {
-            seq_printf(file,"direct ");
-        }
-        seq_printf(file,"internal ] ");
         if (s->mode==V3_PSTATE_EXTERNAL_CONTROL) { 
             seq_printf(file,"(min=%llu max=%llu cur=%llu) ", s->min_freq_khz, s->max_freq_khz, s->cur_freq_khz);
         } 
         if (s->mode==V3_PSTATE_DIRECT_CONTROL) { 
-            seq_printf(file,"(min=%u max=%u cur=%u) ", (uint32_t)s->min_pstate, (uint32_t)s->max_pstate, (uint32_t)s->cur_pstate);
+            seq_printf(file,"(min=%llu max=%llu cur=%llu) ",s->min_pstate, s->max_pstate, s->cur_pstate);
         }
         seq_printf(file,"\n");
     }
@@ -1622,9 +1652,89 @@ static struct file_operations pstate_fops = {
     .release = seq_release
 };
 
+static int pstate_hw_show(struct seq_file * file, void * v)
+{
+    int numstates;
+
+    seq_printf(file, "V3VEE DVFS Hardware Info\n(all logical cores assumed identical)\n\n");
+
+    seq_printf(file, "Arch:   \t%s\n"
+                    "PStates:\t%s\n\n",
+            machine_state.arch==INTEL ? "Intel" : 
+            machine_state.arch==AMD ? "AMD" : "Other",
+            machine_state.supports_pstates ? "Yes" : "No");
+
+
+#define YN(x) ((x) ? "Y" : "N")
+
+    if (machine_state.arch==INTEL) {
+       seq_printf(file,"SpeedStep:           \t%s\n",YN(machine_state.have_speedstep));
+       seq_printf(file,"APERF/MPERF:         \t%s\n",YN(machine_state.have_pstate_hw_coord));
+       seq_printf(file,"IDA or TurboCore:    \t%s\n",YN(machine_state.have_opportunistic));
+       seq_printf(file,"Policy Hint:         \t%s\n",YN(machine_state.have_policy_hint));
+       seq_printf(file,"Hardware Policy:     \t%s\n",YN(machine_state.have_hwp));
+       seq_printf(file,"Hardware Duty Cycle: \t%s\n",YN(machine_state.have_hdc));
+       seq_printf(file,"MWAIT extensions:    \t%s\n",YN(machine_state.have_mwait_ext));
+       seq_printf(file,"MWAIT wake on intr:  \t%s\n",YN(machine_state.have_mwait_int));
+    } 
+
+    if (machine_state.arch==AMD) { 
+       seq_printf(file,"PState:              \t%s\n",YN(machine_state.have_pstate));
+       seq_printf(file,"APERF/MPERF:         \t%s\n",YN(machine_state.have_pstate_hw_coord));
+       seq_printf(file,"CoreBoost:           \t%s\n",YN(machine_state.have_coreboost));
+       seq_printf(file,"Feedback:            \t%s\n",YN(machine_state.have_feedback));
+    }
+
+
+    seq_printf(file,"\nPstate\tCtrl\tKHz\n");
+    numstates = get_cpu_var(processors)->performance->state_count;
+    if (!numstates) { 
+       seq_printf(file,"UNKNOWN\n");
+    } else {
+       int i;
+       for (i=0;i<numstates;i++) { 
+           seq_printf(file,
+                      "%u\t%llx\t%llu\n",
+                      i, 
+                      get_cpu_var(processors)->performance->states[i].control,
+                      get_cpu_var(processors)->performance->states[i].core_frequency*1000);
+       }
+    }
+    put_cpu_var(processors);
+
+    seq_printf(file,"\nAvailable Modes:");
+    seq_printf(file," host");
+    if (get_cpu_var(core_state).have_cpufreq) { 
+       seq_printf(file," external");
+    }
+    put_cpu_var(core_state);
+    if (machine_state.supports_pstates) {
+       seq_printf(file," direct");
+    }
+    seq_printf(file," internal\n");
+
+    return 0;
+}
+
+static int pstate_hw_open(struct inode * inode, struct file * file) 
+{
+    return single_open(file, pstate_hw_show, NULL);
+}
+
+
+static struct file_operations pstate_hw_fops = {
+    .owner = THIS_MODULE,
+    .open = pstate_hw_open, 
+    .read = seq_read,
+    .llseek = seq_lseek,
+    .release = seq_release
+};
+
+
 int pstate_proc_setup(void)
 {
     struct proc_dir_entry *proc;
+    struct proc_dir_entry *prochw;
 
     proc = create_proc_entry("v3-dvfs",0444, palacios_get_procdir());
 
@@ -1635,11 +1745,26 @@ int pstate_proc_setup(void)
 
     proc->proc_fops = &pstate_fops;
 
+    INFO("/proc/v3vee/v3-dvfs successfully created\n");
+
+    prochw = create_proc_entry("v3-dvfs-hw",0444,palacios_get_procdir());
+
+
+    if (!prochw) { 
+        ERROR("Failed to create proc entry for p-state hw info\n");
+        return -1;
+    }
+
+    prochw->proc_fops = &pstate_hw_fops;
+
+    INFO("/proc/v3vee/v3-dvfs-hw successfully created\n");
+
     return 0;
 }
 
 void pstate_proc_teardown(void)
 {
+    remove_proc_entry("v3-dvfs-hw",palacios_get_procdir());
     remove_proc_entry("v3-dvfs",palacios_get_procdir());
 }