42,7 → 42,6 |
#include <arch.h> |
#include <config.h> |
#include <macros.h> |
#include <func.h> |
#include <arch/types.h> |
#include <synch/synch.h> |
#include <synch/waitq.h> |
52,40 → 51,74 |
#include <arch/sun4v/ipi.h> |
#include <time/delay.h> |
|
/** hypervisor code of the "running" state of the CPU */ |
#define CPU_STATE_RUNNING 2 |
|
/** maximum possible number of processor cores */ |
#define MAX_NUM_CORES 8 |
|
/** needed in the CPU_START hypercall */ |
extern void kernel_image_start(void); |
|
/** needed in the CPU_START hypercall */ |
extern void *trap_table; |
|
/** number of execution units detected */ |
uint8_t exec_unit_count = 0; |
/** Determine number of processors. */ |
void smp_init(void) |
{ |
md_node_t node = md_get_root(); |
count_t cpu_count = 0; |
|
/** execution units (processor cores) */ |
exec_unit_t exec_units[MAX_NUM_CORES]; |
/* walk through MD, find the current CPU node & its clock-frequency */ |
while(md_next_node(&node, "cpu")) { |
cpu_count++; |
} |
|
/** CPU structures */ |
extern cpu_t *cpus; |
config.cpu_count = cpu_count; |
} |
|
/** maximum number of strands per a physical core detected */ |
unsigned int max_core_strands = 0; |
|
#ifdef CONFIG_SIMICS_SMP_HACK |
/** |
* Copies a piece of HelenOS code to the place where OBP had its IPI handler. |
* By sending an IPI by the BSP to the AP the code will be executed. |
* The code will jump to the first instruction of the kernel. This is |
* a workaround how to make APs execute HelenOS code on Simics. |
*/ |
static void simics_smp_hack_init(void) { |
/** Wake application processors up. */ |
void kmp(void *arg) |
{ |
#if 1 |
(void) arg; |
|
uint64_t myid; |
__hypercall_fast_ret1(0, 0, 0, 0, 0, CPU_MYID, &myid); |
|
/* stop the CPUs before making them execute our code */ |
uint64_t i; |
for (i = 0; i < config.cpu_count; i++) { |
if (i == myid) |
continue; |
|
if (__hypercall_fast1(CPU_STOP, i) != 0) |
continue; |
|
uint64_t state; |
__hypercall_fast_ret1(i, 0, 0, 0, 0, CPU_STATE, &state); |
while (state == CPU_STATE_RUNNING) { |
__hypercall_fast_ret1(i, 0, 0, 0, 0, CPU_STATE, &state); |
} |
} |
|
/* wake the processors up, one by one */ |
uint64_t state; |
for (i = 1; i < config.cpu_count; i++) { |
__hypercall_fast_ret1(i, 0, 0, 0, 0, CPU_STATE, &state); |
printf("Starting CPU %d, error code = %d.\n", i, __hypercall_fast4( |
CPU_START, |
i, |
(uint64_t) KA2PA(kernel_image_start), |
KA2PA(trap_table), |
bootinfo.physmem_start |
)); |
|
if (waitq_sleep_timeout(&ap_completion_wq, 10000000, SYNCH_FLAGS_NONE) == |
ESYNCH_TIMEOUT) |
printf("%s: waiting for processor (cpuid = %" PRIu32 |
") timed out\n", __func__, i); |
|
} |
#else |
|
asm volatile ( |
"setx temp_cpu_mondo_handler, %g4, %g6 \n" |
//"setx 0x80246ad8, %g4, %g7 \n" |
"setx 0x80200f80, %g4, %g7 \n" |
|
"ldx [%g6], %g4 \n" |
135,272 → 168,14 |
"flush %i7" |
|
); |
} |
#endif |
|
/** |
* Finds out which execution units belong to particular CPUs. By execution unit |
* we mean the physical core the logical processor is backed by. Since each |
* Niagara physical core has just one integer execution unit and we will |
* ignore other execution units than the integer ones, we will use the terms |
* "integer execution unit", "execution unit" and "physical core" |
* interchangeably. |
* |
* The physical cores are detected by browsing the children of the CPU node |
* in the machine description and looking for a node representing an integer |
* execution unit. Once the integer execution unit of a particular CPU is |
* known, the ID of the CPU is added to the list of cpuids of the corresponding |
* execution unit structure (exec_unit_t). If an execution unit is encountered |
* for the first time, a new execution unit structure (exec_unit_t) must be |
* created first and added to the execution units array (exec_units). |
* |
* If the function fails to find an execution unit for a CPU (this may happen |
* on machines with older firmware or on Simics), it performs a fallback code |
* which pretends there exists just one execution unit and all CPUs belong to |
* it. |
* |
* Finally, the array of all execution units is reordered such that its element |
* which represents the physical core of the the bootstrap CPU is at index 0. |
* Moreover, the array of CPU IDs within the BSP's physical core structure is |
* reordered such that the element which represents the ID of the BSP is at |
* index 0. This is done because we would like the CPUs to be woken up |
* such that the 0-index CPU of the 0-index execution unit is |
* woken up first. And since the BSP is already woken up, we would like it to be |
* at 0-th position of the 0-th execution unit structure. |
* |
* Apart from that, the code also counts the total number of CPUs and stores |
* it to the global config.cpu_count variable. |
*/ |
static void detect_execution_units(void) |
{ |
/* ID of the bootstrap processor */ |
uint64_t myid; |
|
/* total number of CPUs detected */ |
count_t cpu_count = 0; |
|
/* will be set to 1 if detecting the physical cores fails */ |
bool exec_unit_assign_error = 0; |
|
/* index of the bootstrap physical core in the array of cores */ |
unsigned int bsp_exec_unit_index = 0; |
|
/* index of the BSP ID inside the array of bootstrap core's cpuids */ |
unsigned int bsp_core_strand_index = 0; |
|
__hypercall_fast_ret1(0, 0, 0, 0, 0, CPU_MYID, &myid); |
md_node_t node = md_get_root(); |
|
/* walk through all the CPU nodes in the MD*/ |
while (md_next_node(&node, "cpu")) { |
|
uint64_t cpuid; |
md_get_integer_property(node, "id", &cpuid); |
cpu_count++; |
|
/* |
* if failed in previous CPUs, don't try |
* to detect physical cores any more |
*/ |
if (exec_unit_assign_error) |
continue; |
|
/* detect exec. unit for the CPU represented by current node */ |
uint64_t exec_unit_id = 0; |
md_child_iter_t it = md_get_child_iterator(node); |
|
while (md_next_child(&it)) { |
md_node_t child = md_get_child_node(it); |
const char *exec_unit_type; |
md_get_string_property(child, "type", &exec_unit_type); |
|
/* each physical core has just 1 integer exec. unit */ |
if (strcmp(exec_unit_type, "integer") == 0) { |
exec_unit_id = child; |
break; |
} |
} |
|
/* execution unit detected successfully */ |
if (exec_unit_id != 0) { |
|
/* find the exec. unit in array of existing units */ |
unsigned int i = 0; |
for (i = 0; i < exec_unit_count; i++) { |
if (exec_units[i].exec_unit_id == exec_unit_id) |
break; |
} |
|
/* |
* execution unit just met has not been met before, so |
* create a new entry in array of all execution units |
*/ |
if (i == exec_unit_count) { |
exec_units[i].exec_unit_id = exec_unit_id; |
exec_units[i].strand_count = 0; |
exec_unit_count++; |
} |
|
/* |
* remember the exec. unit and strand of the BSP |
*/ |
if (cpuid == myid) { |
bsp_exec_unit_index = i; |
bsp_core_strand_index = exec_units[i].strand_count; |
} |
|
/* add the CPU just met to the exec. unit's list */ |
exec_units[i].cpuids[exec_units[i].strand_count] = cpuid; |
exec_units[i].strand_count++; |
max_core_strands = |
exec_units[i].strand_count > max_core_strands ? |
exec_units[i].strand_count : max_core_strands; |
|
/* detecting execution unit failed */ |
} else { |
exec_unit_assign_error = 1; |
} |
} |
|
/* save the number of CPUs to a globally accessible variable */ |
config.cpu_count = cpu_count; |
|
/* |
* A fallback code which will be executed if finding out which |
* execution units belong to particular CPUs fails. Pretend there |
* exists just one execution unit and all CPUs belong to it. |
*/ |
if (exec_unit_assign_error) { |
bsp_exec_unit_index = 0; |
exec_unit_count = 1; |
exec_units[0].strand_count = cpu_count; |
exec_units[0].exec_unit_id = 1; |
max_core_strands = cpu_count; |
|
/* browse CPUs again, assign them the fictional exec. unit */ |
node = md_get_root(); |
unsigned int i = 0; |
|
while (md_next_node(&node, "cpu")) { |
uint64_t cpuid; |
md_get_integer_property(node, "id", &cpuid); |
if (cpuid == myid) { |
bsp_core_strand_index = i; |
} |
exec_units[0].cpuids[i++] = cpuid; |
} |
} |
|
/* |
* Reorder the execution units array elements and the cpuid array |
* elements so that the BSP will always be the very first CPU of |
* the very first execution unit. |
*/ |
exec_unit_t temp_exec_unit = exec_units[0]; |
exec_units[0] = exec_units[bsp_exec_unit_index]; |
exec_units[bsp_exec_unit_index] = temp_exec_unit; |
|
uint64_t temp_cpuid = exec_units[0].cpuids[0]; |
exec_units[0].cpuids[0] = exec_units[0].cpuids[bsp_exec_unit_index]; |
exec_units[0].cpuids[bsp_core_strand_index] = temp_cpuid; |
|
} |
|
/** |
* Determine number of processors and detect physical cores. On Simics |
* copy the code which will be executed by the AP when the BSP sends an |
* IPI to it in order to make it execute HelenOS code. |
*/ |
void smp_init(void) |
{ |
detect_execution_units(); |
#ifdef CONFIG_SIMICS_SMP_HACK |
simics_smp_hack_init(); |
#endif |
} |
|
/** |
* For each CPU sets the value of cpus[i].arch.id, where i is the |
* index of the CPU in the cpus variable, to the cpuid of the i-th processor |
* to be run. The CPUs are run such that the CPU represented by cpus[0] |
* is run first, cpus[1] is run after it, and cpus[cpu_count - 1] is run as the |
* last one. |
* |
* The CPU IDs are set such that during waking the CPUs up the |
* processor cores will be alternated, i.e. first one CPU from the first core |
* will be run, after that one CPU from the second CPU core will be run,... |
* then one CPU from the last core will be run, after that another CPU |
* from the first core will be run, then another CPU from the second core |
* will be run,... then another CPU from the last core will be run, and so on. |
*/ |
static void init_cpuids(void) |
{ |
unsigned int cur_core_strand; |
unsigned int cur_core; |
unsigned int cur_cpu = 0; |
|
for (cur_core_strand = 0; cur_core_strand < max_core_strands; cur_core_strand++) { |
for (cur_core = 0; cur_core < exec_unit_count; cur_core++) { |
if (cur_core_strand > exec_units[cur_core].strand_count) |
continue; |
|
cpus[cur_cpu++].arch.id = exec_units[cur_core].cpuids[cur_core_strand]; |
} |
} |
} |
|
/** |
* Wakes up a single CPU. |
* |
* @param cpuid ID of the CPU to be woken up |
*/ |
static bool wake_cpu(uint64_t cpuid) |
{ |
|
#ifdef CONFIG_SIMICS_SMP_HACK |
ipi_unicast_to((void (*)(void)) 1234, cpuid); |
#else |
/* stop the CPU before making it execute our code */ |
if (__hypercall_fast1(CPU_STOP, cpuid) != EOK) |
return false; |
|
/* wait for the CPU to stop */ |
uint64_t state; |
__hypercall_fast_ret1(cpuid, 0, 0, 0, 0, |
CPU_STATE, &state); |
while (state == CPU_STATE_RUNNING) { |
__hypercall_fast_ret1(cpuid, 0, 0, 0, 0, |
CPU_STATE, &state); |
} |
|
/* make the CPU run again and execute HelenOS code */ |
if (__hypercall_fast4( |
CPU_START, cpuid, |
(uint64_t) KA2PA(kernel_image_start), |
KA2PA(trap_table), bootinfo.physmem_start |
) != EOK) |
return false; |
#endif |
|
delay(1000); |
printf("Result: %d\n", ipi_unicast_to((void (*)(void)) 1234, 1)); |
if (waitq_sleep_timeout(&ap_completion_wq, 10000000, SYNCH_FLAGS_NONE) == |
ESYNCH_TIMEOUT) |
printf("%s: waiting for processor (cpuid = %" PRIu32 |
") timed out\n", __func__, cpuid); |
|
return true; |
") timed out\n", __func__, 1); |
#endif |
} |
|
/** Wake application processors up. */ |
void kmp(void *arg) |
{ |
init_cpuids(); |
|
unsigned int i; |
|
for (i = 1; i < config.cpu_count; i++) { |
wake_cpu(cpus[i].arch.id); |
} |
} |
|
/** @} |
*/ |