#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.
*/
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
uint64_t cur_freq_khz;
uint64_t max_freq_khz;
uint64_t min_freq_khz;
/* 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);
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;
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;
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
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,
// 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);
Linux Interface
*****************************************************************/
+static unsigned cpus_using_v3_governor;
+static DEFINE_MUTEX(v3_governor_mutex);
+
+/* KCH: this will tell us when there is an actual frequency transition */
+static int v3_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
+ void *data)
+{
+ struct cpufreq_freqs *freq = data;
+
+ if (per_cpu(core_state, freq->cpu).mode != V3_PSTATE_EXTERNAL_CONTROL) {
+ return 0;
+ }
+
+ if (val == CPUFREQ_POSTCHANGE) {
+ DEBUG("P-State: frequency change took effect on cpu %u (now %u kHz)\n",
+ freq->cpu, freq->new);
+ per_cpu(core_state, freq->cpu).cur_freq_khz = freq->new;
+ }
+
+ return 0;
+
+}
+
+
+static struct notifier_block v3_cpufreq_notifier_block = {
+ .notifier_call = v3_cpufreq_notifier
+};
/*
*/
static int governor_run(struct cpufreq_policy *policy, unsigned int event)
{
+ unsigned cpu = policy->cpu;
switch (event) {
/* we can't use cpufreq_driver_target here as it can result
- * in a circular dependency, so we'll just do nothing.
+ * in a circular dependency, so we'll keep the current frequency as is
*/
case CPUFREQ_GOV_START:
+ BUG_ON(!policy->cur);
+
+ mutex_lock(&v3_governor_mutex);
+
+ if (cpus_using_v3_governor == 0) {
+ cpufreq_register_notifier(&v3_cpufreq_notifier_block,
+ CPUFREQ_TRANSITION_NOTIFIER);
+ }
+
+ cpus_using_v3_governor++;
+
+ per_cpu(core_state, cpu).set_freq_khz = policy->cur;
+ per_cpu(core_state, cpu).cur_freq_khz = policy->cur;
+ per_cpu(core_state, cpu).max_freq_khz = policy->max;
+ per_cpu(core_state, cpu).min_freq_khz = policy->min;
+
+ mutex_unlock(&v3_governor_mutex);
+ break;
case CPUFREQ_GOV_STOP:
+ mutex_lock(&v3_governor_mutex);
+
+ cpus_using_v3_governor--;
+
+ if (cpus_using_v3_governor == 0) {
+ cpufreq_unregister_notifier(
+ &v3_cpufreq_notifier_block,
+ CPUFREQ_TRANSITION_NOTIFIER);
+ }
+
+ per_cpu(core_state, cpu).set_freq_khz = 0;
+ per_cpu(core_state, cpu).cur_freq_khz = 0;
+ per_cpu(core_state, cpu).max_freq_khz = 0;
+ per_cpu(core_state, cpu).min_freq_khz = 0;
+
+ mutex_unlock(&v3_governor_mutex);
+ break;
case CPUFREQ_GOV_LIMITS:
/* do nothing */
break;
default:
- ERROR("Undefined governor command\n");
+ ERROR("Undefined governor command (%u)\n", event);
return -1;
}
static struct workqueue_struct *pstate_wq;
-
typedef struct {
struct work_struct work;
uint64_t freq;
} pstate_work_t;
+
+
static inline void pstate_register_linux_governor(void)
{
cpufreq_register_governor(&stub_governor);
}
strncpy(govname, policy->governor->name, MAX_GOV_NAME_LEN);
+ govname[MAX_GOV_NAME_LEN-1] = 0;
get_cpu_var(core_state).linux_governor = govname;
put_cpu_var(core_state);
argv[3] = NULL;
/* KCH: we can't wait here to actually see if we succeeded, we're in interrupt context */
- return call_usermodehelper_fns("/bin/sh", argv, envp, UMH_NO_WAIT, NULL, gov_switch_cleanup, NULL);
+#if LINUX_VERSION_CODE <= KERNEL_VERSION(3,9,0)
+ return call_usermodehelper_fns("/bin/sh", argv, envp, UMH_NO_WAIT, NULL, gov_switch_cleanup, NULL);
+#else
+ {
+ struct subprocess_info *sp;
+
+ sp = call_usermodehelper_setup("/bin/sh", argv, envp, GFP_ATOMIC, NULL, gov_switch_cleanup, NULL);
+ if (!sp) {
+ goto out_freeargv;
+ }
+
+ return call_usermodehelper_exec(sp,0);
+ }
+#endif
+
out_freeargv:
palacios_free(argv);
return -1;
{
char * gov;
unsigned int cpu = get_cpu();
+ put_cpu();
/* KCH: we assume the v3vee governor is already
* registered with kernel by this point
DEBUG("setting the new governor (%s)\n", PALACIOS_GOVNAME);
/* set the new one to ours */
+
if (governor_switch(PALACIOS_GOVNAME, cpu) < 0) {
ERROR("Could not set governor to (%s)\n", PALACIOS_GOVNAME);
return -1;
-static int linux_get_pstate(void)
+static uint64_t linux_get_pstate(void)
{
struct cpufreq_policy * policy = NULL;
struct cpufreq_frequency_table *table;
- int cpu = get_cpu();
unsigned int i = 0;
unsigned int count = 0;
+ unsigned int cpu = get_cpu();
+ put_cpu();
policy = palacios_alloc(sizeof(struct cpufreq_policy));
}
-static int linux_get_freq(void)
+static uint64_t linux_get_freq(void)
{
+ uint64_t freq;
struct cpufreq_policy * policy = NULL;
- int cpu = get_cpu();
+ unsigned int cpu = get_cpu();
+ put_cpu();
policy = palacios_alloc(sizeof(struct cpufreq_policy));
if (!policy) {
return -1;
}
- return policy->cur;
+ freq=policy->cur;
+
+ palacios_free(policy);
+
+ return freq;
}
static void
{
pstate_work_t * pwork = (pstate_work_t*)work;
struct cpufreq_policy * policy = NULL;
- int cpu = get_cpu();
+ uint64_t freq;
+ unsigned int cpu = get_cpu();
put_cpu();
+ mutex_lock(&v3_governor_mutex);
+
policy = palacios_alloc(sizeof(struct cpufreq_policy));
if (!policy) {
ERROR("Could not allocate space for cpufreq policy\n");
goto out1;
}
- INFO("P-state: setting frequency on core %u to %llu\n", cpu, pwork->freq);
- cpufreq_driver_target(policy, pwork->freq, CPUFREQ_RELATION_H);
+ freq = pwork->freq;
+ get_cpu_var(core_state).set_freq_khz = freq;
- get_cpu_var(core_state).cur_freq_khz = pwork->freq;
+ if (freq < get_cpu_var(core_state).min_freq_khz) {
+ freq = get_cpu_var(core_state).min_freq_khz;
+ }
+ if (freq > get_cpu_var(core_state).max_freq_khz) {
+ freq = get_cpu_var(core_state).max_freq_khz;
+ }
put_cpu_var(core_state);
+ INFO("P-state: requesting frequency change on core %u to %llu\n", cpu, freq);
+ __cpufreq_driver_target(policy, freq, CPUFREQ_RELATION_L);
+
out1:
palacios_free(policy);
out:
palacios_free(work);
+ mutex_unlock(&v3_governor_mutex);
}
-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;
pstate_work_t * work = NULL;
- int cpu = get_cpu();
unsigned int i = 0;
unsigned int count = 0;
int state_set = 0;
int last_valid = 0;
+ unsigned int cpu = get_cpu();
put_cpu();
policy = palacios_alloc(sizeof(struct cpufreq_policy));
struct cpufreq_policy * policy = NULL;
pstate_work_t * work = NULL;
uint64_t freq;
- int cpu = get_cpu();
+ unsigned int cpu = get_cpu();
put_cpu();
policy = palacios_alloc(sizeof(struct cpufreq_policy));
static int linux_restore_defaults(void)
{
- unsigned int cpu = get_cpu();
char * gov = NULL;
+ unsigned int cpu = get_cpu();
+ put_cpu();
gov = get_cpu_var(core_state).linux_governor;
put_cpu_var(core_state);
{
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;
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,
get_cpu_var(processors)->performance->states[i].control);
}
put_cpu_var(processors);
+ */
}
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);
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");
}
.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\tmW\tuS(X)\tuS(B)\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\t%llu\t%llu\t%llu\n",
+ i,
+ get_cpu_var(processors)->performance->states[i].control,
+ get_cpu_var(processors)->performance->states[i].core_frequency*1000,
+ get_cpu_var(processors)->performance->states[i].power,
+ get_cpu_var(processors)->performance->states[i].transition_latency,
+ get_cpu_var(processors)->performance->states[i].bus_master_latency);
+ }
+ }
+ 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());
+ PAL_PROC_CREATE(proc,"v3-dvfs",0444,palacios_get_procdir(),&pstate_fops);
if (!proc) {
ERROR("Failed to create proc entry for p-state control\n");
return -1;
}
- proc->proc_fops = &pstate_fops;
+ INFO("/proc/v3vee/v3-dvfs successfully created\n");
+
+ PAL_PROC_CREATE(prochw,"v3-dvfs-hw",0444,palacios_get_procdir(),&pstate_hw_fops);
+
+ if (!prochw) {
+ ERROR("Failed to create proc entry for p-state hw info\n");
+ return -1;
+ }
+
+ 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());
}
