int kbase_platform_dvfs_event(struct kbase_device *kbdev, u32 utilisation)
{
unsigned long flags;
struct exynos_context *platform;
KBASE_DEBUG_ASSERT(kbdev != NULL);
platform = (struct exynos_context *) kbdev->platform_context;
spin_lock_irqsave(&mali_dvfs_spinlock, flags);
if (platform->time_tick < MALI_DVFS_TIME_INTERVAL) {
platform->time_tick++;
platform->time_busy += kbdev->pm.metrics.time_busy;
platform->time_idle += kbdev->pm.metrics.time_idle;
} else {
platform->time_busy = kbdev->pm.metrics.time_busy;
platform->time_idle = kbdev->pm.metrics.time_idle;
platform->time_tick = 0;
}
if ((platform->time_tick == MALI_DVFS_TIME_INTERVAL) &&
(platform->time_idle + platform->time_busy > 0))
platform->utilisation = (100*platform->time_busy) / (platform->time_idle + platform->time_busy);
mali_dvfs_status_current.utilisation = utilisation;
#ifdef MALI_DEBUG
printk(KERN_INFO "\n[mali_devfreq]utilization: %d\n", utilisation);
#endif
spin_unlock_irqrestore(&mali_dvfs_spinlock, flags);
queue_work_on(0, mali_dvfs_wq, &mali_dvfs_work);
/*add error handle here*/
return MALI_TRUE;
}
void kbase_pm_register_vsync_callback(struct kbase_device *kbdev)
{
KBASE_DEBUG_ASSERT(kbdev != NULL);
/* no VSync metrics will be available */
kbdev->pm.metrics.platform_data = NULL;
}
void kbase_timeline_job_slot_done(kbase_device *kbdev, kbase_context *kctx,
kbase_jd_atom *katom, int js,
kbasep_js_atom_done_code done_code)
{
lockdep_assert_held(&kbdev->js_data.runpool_irq.lock);
if (done_code & KBASE_JS_ATOM_DONE_EVICTED_FROM_NEXT) {
KBASE_TIMELINE_JOB_START_NEXT(kctx, js, 0);
} else {
/* Job finished in JSn_HEAD */
base_atom_id atom_number = kbase_jd_atom_id(kctx, katom);
KBASE_TIMELINE_JOB_START_HEAD(kctx, js, 0);
KBASE_TIMELINE_JOB_STOP(kctx, js, atom_number);
/* see if we need to trace the job in JSn_NEXT moving to JSn_HEAD */
if (kbdev->timeline.slot_atoms_submitted[js] > 1) {
/* Tag events with next_katom's kctx */
kbase_jm_slot *slot = &kbdev->jm_slots[js];
kbase_jd_atom *next_katom;
kbase_context *next_kctx;
KBASE_DEBUG_ASSERT(kbasep_jm_nr_jobs_submitted(slot) > 0);
/* Peek the next atom - note that the atom in JSn_HEAD will already
* have been dequeued */
next_katom = kbasep_jm_peek_idx_submit_slot(slot, 0);
next_kctx = next_katom->kctx;
KBASE_TIMELINE_JOB_START_NEXT(next_kctx, js, 0);
KBASE_TIMELINE_JOB_START_HEAD(next_kctx, js, 1);
KBASE_TIMELINE_JOB_START(next_kctx, js, kbase_jd_atom_id(next_kctx, next_katom));
}
}
--kbdev->timeline.slot_atoms_submitted[js];
KBASE_TIMELINE_ATOMS_SUBMITTED(kctx, js, kbdev->timeline.slot_atoms_submitted[js]);
}
const struct kbase_pm_ca_policy
*kbase_pm_ca_get_policy(struct kbase_device *kbdev)
{
KBASE_DEBUG_ASSERT(kbdev != NULL);
return kbdev->pm.backend.ca_current_policy;
}
int kbasep_jd_debugfs_ctx_add(struct kbase_context *kctx)
{
/* Refer below for format string, %u is 10 chars max */
char dir_name[10 * 2 + 2];
KBASE_DEBUG_ASSERT(kctx != NULL);
/* Create per-context directory */
scnprintf(dir_name, sizeof(dir_name), "%u_%u", kctx->pid, kctx->id);
kctx->jd_ctx_dir = debugfs_create_dir(dir_name, kctx->kbdev->jd_directory);
if (IS_ERR(kctx->jd_ctx_dir))
goto err;
/* Expose all atoms */
if (IS_ERR(debugfs_create_file("atoms", S_IRUGO,
kctx->jd_ctx_dir, kctx, &kbasep_jd_debugfs_atoms_fops)))
goto err_jd_ctx_dir;
return 0;
err_jd_ctx_dir:
debugfs_remove_recursive(kctx->jd_ctx_dir);
err:
return -1;
}
/**
* kbase_context_set_create_flags - Set creation flags on a context
* @kctx: Kbase context
* @flags: Flags to set
*
* Return: 0 on success
*/
int kbase_context_set_create_flags(struct kbase_context *kctx, u32 flags)
{
int err = 0;
struct kbasep_js_kctx_info *js_kctx_info;
unsigned long irq_flags;
KBASE_DEBUG_ASSERT(NULL != kctx);
js_kctx_info = &kctx->jctx.sched_info;
/* Validate flags */
if (flags != (flags & BASE_CONTEXT_CREATE_KERNEL_FLAGS)) {
err = -EINVAL;
goto out;
}
mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
spin_lock_irqsave(&kctx->kbdev->js_data.runpool_irq.lock, irq_flags);
/* Translate the flags */
if ((flags & BASE_CONTEXT_SYSTEM_MONITOR_SUBMIT_DISABLED) == 0)
js_kctx_info->ctx.flags &= ~((u32) KBASE_CTX_FLAG_SUBMIT_DISABLED);
if ((flags & BASE_CONTEXT_HINT_ONLY_COMPUTE) != 0)
js_kctx_info->ctx.flags |= (u32) KBASE_CTX_FLAG_HINT_ONLY_COMPUTE;
/* Latch the initial attributes into the Job Scheduler */
kbasep_js_ctx_attr_set_initial_attrs(kctx->kbdev, kctx);
spin_unlock_irqrestore(&kctx->kbdev->js_data.runpool_irq.lock,
irq_flags);
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
out:
return err;
}
/* Find region enclosing given address. */
struct kbase_va_region *kbase_region_tracker_find_region_enclosing_address(struct kbase_context *kctx, u64 gpu_addr)
{
struct rb_node *rbnode;
struct kbase_va_region *reg;
u64 gpu_pfn = gpu_addr >> PAGE_SHIFT;
KBASE_DEBUG_ASSERT(NULL != kctx);
lockdep_assert_held(&kctx->reg_lock);
rbnode = kctx->reg_rbtree.rb_node;
while (rbnode) {
u64 tmp_start_pfn, tmp_end_pfn;
reg = rb_entry(rbnode, struct kbase_va_region, rblink);
tmp_start_pfn = reg->start_pfn;
tmp_end_pfn = reg->start_pfn + reg->nr_pages;
/* If start is lower than this, go left. */
if (gpu_pfn < tmp_start_pfn)
rbnode = rbnode->rb_left;
/* If end is higher than this, then go right. */
else if (gpu_pfn >= tmp_end_pfn)
rbnode = rbnode->rb_right;
else /* Enclosing */
return reg;
}
return NULL;
}
void kbasep_jd_debugfs_ctx_remove(struct kbase_context *kctx)
{
KBASE_DEBUG_ASSERT(kctx != NULL);
if (!IS_ERR(kctx->jd_ctx_dir))
debugfs_remove_recursive(kctx->jd_ctx_dir);
}
/** Show callback for the @c mem_profile debugfs file.
*
* This function is called to get the contents of the @c mem_profile debugfs
* file. This is a report of current memory usage and distribution in userspace.
*
* @param sfile The debugfs entry
* @param data Data associated with the entry
*
* @return 0 if successfully prints data in debugfs entry file
* -1 if it encountered an error
*/
static int kbasep_mem_profile_seq_show(struct seq_file *sfile, void *data)
{
struct kbase_context *kctx = sfile->private;
KBASE_DEBUG_ASSERT(kctx != NULL);
/* MALI_SEC_INTEGRATION */
{
struct kbase_device *kbdev = kctx->kbdev;
atomic_inc(&kctx->mem_profile_showing_state);
if(kbdev->vendor_callbacks->mem_profile_check_kctx)
if (!kbdev->vendor_callbacks->mem_profile_check_kctx(kctx)) {
atomic_dec(&kctx->mem_profile_showing_state);
return 0;
}
}
/* MALI_SEC_INTEGRATION */
if (kctx->destroying_context) {
atomic_dec(&kctx->mem_profile_showing_state);
return 0;
}
spin_lock(&kctx->mem_profile_lock);
/* MALI_SEC_INTEGRATION */
if (kctx->mem_profile_data) {
seq_write(sfile, kctx->mem_profile_data, kctx->mem_profile_size);
seq_putc(sfile, '\n');
}
spin_unlock(&kctx->mem_profile_lock);
atomic_dec(&kctx->mem_profile_showing_state);
return 0;
}
/**
* Workqueue for handling cache cleaning
*/
void kbasep_cache_clean_worker(struct work_struct *data)
{
struct kbase_device *kbdev;
unsigned long flags;
kbdev = container_of(data, struct kbase_device, hwcnt.cache_clean_work);
mutex_lock(&kbdev->cacheclean_lock);
kbasep_instr_hwcnt_cacheclean(kbdev);
spin_lock_irqsave(&kbdev->hwcnt.lock, flags);
/* Wait for our condition, and any reset to complete */
while (kbdev->hwcnt.state == KBASE_INSTR_STATE_RESETTING ||
kbdev->hwcnt.state == KBASE_INSTR_STATE_CLEANING) {
spin_unlock_irqrestore(&kbdev->hwcnt.lock, flags);
wait_event(kbdev->hwcnt.cache_clean_wait,
(kbdev->hwcnt.state != KBASE_INSTR_STATE_RESETTING
&& kbdev->hwcnt.state != KBASE_INSTR_STATE_CLEANING));
spin_lock_irqsave(&kbdev->hwcnt.lock, flags);
}
KBASE_DEBUG_ASSERT(kbdev->hwcnt.state == KBASE_INSTR_STATE_CLEANED);
/* All finished and idle */
kbdev->hwcnt.state = KBASE_INSTR_STATE_IDLE;
kbdev->hwcnt.triggered = 1;
wake_up(&kbdev->hwcnt.wait);
spin_unlock_irqrestore(&kbdev->hwcnt.lock, flags);
mutex_unlock(&kbdev->cacheclean_lock);
}
void kbase_gpuprops_set(kbase_device *kbdev)
{
kbase_gpu_props *gpu_props;
struct midg_raw_gpu_props *raw;
KBASE_DEBUG_ASSERT(NULL != kbdev);
gpu_props = &kbdev->gpu_props;
raw = &gpu_props->props.raw_props;
/* Initialize the base_gpu_props structure from the hardware */
kbase_gpuprops_get_props(&gpu_props->props, kbdev);
/* Populate the derived properties */
kbase_gpuprops_calculate_props(&gpu_props->props, kbdev);
/* Populate kbase-only fields */
gpu_props->l2_props.associativity = KBASE_UBFX32(raw->l2_features, 8U, 8);
gpu_props->l2_props.external_bus_width = KBASE_UBFX32(raw->l2_features, 24U, 8);
gpu_props->l3_props.associativity = KBASE_UBFX32(raw->l3_features, 8U, 8);
gpu_props->l3_props.external_bus_width = KBASE_UBFX32(raw->l3_features, 24U, 8);
gpu_props->mem.core_group = KBASE_UBFX32(raw->mem_features, 0U, 1);
gpu_props->mem.supergroup = KBASE_UBFX32(raw->mem_features, 1U, 1);
gpu_props->mmu.va_bits = KBASE_UBFX32(raw->mmu_features, 0U, 8);
gpu_props->mmu.pa_bits = KBASE_UBFX32(raw->mmu_features, 8U, 8);
gpu_props->num_cores = hweight64(raw->shader_present);
gpu_props->num_core_groups = hweight64(raw->l2_present);
gpu_props->num_supergroups = hweight64(raw->l3_present);
gpu_props->num_address_spaces = hweight32(raw->as_present);
gpu_props->num_job_slots = hweight32(raw->js_present);
}
STATIC INLINE mali_bool kbase_js_affinity_is_violating(struct kbase_device *kbdev, u64 *affinities)
{
/* This implementation checks whether the two slots involved in Generic thread creation
* have intersecting affinity. This is due to micro-architectural issues where a job in
* slot A targetting cores used by slot B could prevent the job in slot B from making
* progress until the job in slot A has completed.
*
* @note It just so happens that this restriction also allows
* BASE_HW_ISSUE_8987 to be worked around by placing on job slot 2 the
* atoms from ctxs with KBASE_CTX_FLAG_HINT_ONLY_COMPUTE flag set
*/
u64 affinity_set_left;
u64 affinity_set_right;
u64 intersection;
KBASE_DEBUG_ASSERT(affinities != NULL);
affinity_set_left = affinities[1];
if (kbase_hw_has_issue(kbdev, BASE_HW_ISSUE_8987)) {
/* The left set also includes those on the Fragment slot when
* we are using the HW workaround for BASE_HW_ISSUE_8987 */
affinity_set_left |= affinities[0];
}
affinity_set_right = affinities[2];
/* A violation occurs when any bit in the left_set is also in the right_set */
intersection = affinity_set_left & affinity_set_right;
return (mali_bool) (intersection != (u64) 0u);
}
/* This function inserts a region into the tree. */
static void kbase_region_tracker_insert(struct kbase_context *kctx, struct kbase_va_region *new_reg)
{
u64 start_pfn = new_reg->start_pfn;
struct rb_node **link = &(kctx->reg_rbtree.rb_node);
struct rb_node *parent = NULL;
/* Find the right place in the tree using tree search */
while (*link) {
struct kbase_va_region *old_reg;
parent = *link;
old_reg = rb_entry(parent, struct kbase_va_region, rblink);
/* RBTree requires no duplicate entries. */
KBASE_DEBUG_ASSERT(old_reg->start_pfn != start_pfn);
if (old_reg->start_pfn > start_pfn)
link = &(*link)->rb_left;
else
link = &(*link)->rb_right;
}
/* Put the new node there, and rebalance tree */
rb_link_node(&(new_reg->rblink), parent, link);
rb_insert_color(&(new_reg->rblink), &(kctx->reg_rbtree));
}
void kbasep_jd_debugfs_term(struct kbase_device *kbdev)
{
KBASE_DEBUG_ASSERT(kbdev != NULL);
if (!IS_ERR(kbdev->jd_directory))
debugfs_remove_recursive(kbdev->jd_directory);
}
static inline bool kbase_js_affinity_is_violating(
struct kbase_device *kbdev,
u64 *affinities)
{
/* This implementation checks whether the two slots involved in Generic
* thread creation have intersecting affinity. This is due to micro-
* architectural issues where a job in slot A targetting cores used by
* slot B could prevent the job in slot B from making progress until the
* job in slot A has completed.
*/
u64 affinity_set_left;
u64 affinity_set_right;
u64 intersection;
KBASE_DEBUG_ASSERT(affinities != NULL);
affinity_set_left = affinities[1];
affinity_set_right = affinities[2];
/* A violation occurs when any bit in the left_set is also in the
* right_set */
intersection = affinity_set_left & affinity_set_right;
return (bool) (intersection != (u64) 0u);
}
void kbase_pm_ca_set_policy(struct kbase_device *kbdev, const struct kbase_pm_ca_policy *new_policy)
{
const struct kbase_pm_ca_policy *old_policy;
unsigned long flags;
KBASE_DEBUG_ASSERT(kbdev != NULL);
KBASE_DEBUG_ASSERT(new_policy != NULL);
KBASE_TRACE_ADD(kbdev, PM_CA_SET_POLICY, NULL, NULL, 0u, new_policy->id);
/* During a policy change we pretend the GPU is active */
/* A suspend won't happen here, because we're in a syscall from a userspace thread */
kbase_pm_context_active(kbdev);
wake_lock(&kbdev->pm.kbase_wake_lock);
mutex_lock(&kbdev->pm.lock);
/* Remove the policy to prevent IRQ handlers from working on it */
spin_lock_irqsave(&kbdev->pm.power_change_lock, flags);
old_policy = kbdev->pm.ca_current_policy;
kbdev->pm.ca_current_policy = NULL;
spin_unlock_irqrestore(&kbdev->pm.power_change_lock, flags);
if (old_policy->term)
old_policy->term(kbdev);
if (new_policy->init)
new_policy->init(kbdev);
spin_lock_irqsave(&kbdev->pm.power_change_lock, flags);
kbdev->pm.ca_current_policy = new_policy;
/* If any core power state changes were previously attempted, but couldn't
* be made because the policy was changing (current_policy was NULL), then
* re-try them here. */
kbase_pm_update_cores_state_nolock(kbdev);
kbdev->pm.ca_current_policy->update_core_status(kbdev, kbdev->shader_ready_bitmap, kbdev->shader_transitioning_bitmap);
spin_unlock_irqrestore(&kbdev->pm.power_change_lock, flags);
mutex_unlock(&kbdev->pm.lock);
wake_unlock(&kbdev->pm.kbase_wake_lock);
/* Now the policy change is finished, we release our fake context active reference */
kbase_pm_context_idle(kbdev);
}
void kbase_pm_halt(kbase_device *kbdev)
{
KBASE_DEBUG_ASSERT(kbdev != NULL);
mutex_lock(&kbdev->pm.lock);
kbase_pm_do_poweroff(kbdev);
mutex_unlock(&kbdev->pm.lock);
}
bool kbase_js_affinity_would_violate(struct kbase_device *kbdev, int js,
u64 affinity)
{
struct kbasep_js_device_data *js_devdata;
u64 new_affinities[BASE_JM_MAX_NR_SLOTS];
KBASE_DEBUG_ASSERT(kbdev != NULL);
KBASE_DEBUG_ASSERT(js < BASE_JM_MAX_NR_SLOTS);
js_devdata = &kbdev->js_data;
memcpy(new_affinities, js_devdata->runpool_irq.slot_affinities,
sizeof(js_devdata->runpool_irq.slot_affinities));
new_affinities[js] |= affinity;
return kbase_js_affinity_is_violating(kbdev, new_affinities);
}
/**
* @brief Show callback for the @c JD atoms debugfs file.
*
* This function is called to get the contents of the @c JD atoms debugfs file.
* This is a report of all atoms managed by kbase_jd_context::atoms .
*
* @param sfile The debugfs entry
* @param data Data associated with the entry
*
* @return 0 if successfully prints data in debugfs entry file, failure
* otherwise
*/
static int kbasep_jd_debugfs_atoms_show(struct seq_file *sfile, void *data)
{
struct kbase_context *kctx = sfile->private;
struct kbase_jd_atom *atoms;
unsigned long irq_flags;
int i;
KBASE_DEBUG_ASSERT(kctx != NULL);
/* Print table heading */
seq_puts(sfile, "atom id,core reqs,status,coreref status,predeps,start time,time on gpu\n");
atoms = kctx->jctx.atoms;
/* General atom states */
mutex_lock(&kctx->jctx.lock);
/* JS-related states */
spin_lock_irqsave(&kctx->kbdev->js_data.runpool_irq.lock, irq_flags);
for (i = 0; i != BASE_JD_ATOM_COUNT; ++i) {
struct kbase_jd_atom *atom = &atoms[i];
s64 start_timestamp = 0;
if (atom->status == KBASE_JD_ATOM_STATE_UNUSED)
continue;
/* start_timestamp is cleared as soon as the atom leaves UNUSED state
* and set before a job is submitted to the h/w, a non-zero value means
* it is valid */
if (ktime_to_ns(atom->start_timestamp))
start_timestamp = ktime_to_ns(
ktime_sub(ktime_get(), atom->start_timestamp));
/* MALI_SEC_INTEGRATION */
#ifdef CONFIG_ARM64
seq_printf(sfile,
"%i,%u,%u,%u,%ld,%ld,%lli,%llu\n",
i, atom->core_req, atom->status, atom->coreref_state,
atom->dep[0].atom ? atom->dep[0].atom - atoms : 0,
atom->dep[1].atom ? atom->dep[1].atom - atoms : 0,
(signed long long)start_timestamp,
(unsigned long long)(atom->time_spent_us ?
atom->time_spent_us * 1000 : start_timestamp)
);
#else
seq_printf(sfile,
"%i,%u,%u,%u,%d,%d,%lli,%llu\n",
i, atom->core_req, atom->status, atom->coreref_state,
atom->dep[0].atom ? atom->dep[0].atom - atoms : 0,
atom->dep[1].atom ? atom->dep[1].atom - atoms : 0,
(signed long long)start_timestamp,
(unsigned long long)(atom->time_spent_us ?
atom->time_spent_us * 1000 : start_timestamp)
);
#endif
}
spin_unlock_irqrestore(&kctx->kbdev->js_data.runpool_irq.lock, irq_flags);
mutex_unlock(&kctx->jctx.lock);
return 0;
}
int kbase_pm_context_active_handle_suspend(struct kbase_device *kbdev, enum kbase_pm_suspend_handler suspend_handler)
{
int c;
int old_count;
KBASE_DEBUG_ASSERT(kbdev != NULL);
/* Trace timeline information about how long it took to handle the decision
* to powerup. Sometimes the event might be missed due to reading the count
* outside of mutex, but this is necessary to get the trace timing
* correct. */
old_count = kbdev->pm.active_count;
if (old_count == 0)
kbase_timeline_pm_send_event(kbdev, KBASE_TIMELINE_PM_EVENT_GPU_ACTIVE);
mutex_lock(&kbdev->pm.lock);
if (kbase_pm_is_suspending(kbdev)) {
switch (suspend_handler) {
case KBASE_PM_SUSPEND_HANDLER_DONT_REACTIVATE:
if (kbdev->pm.active_count != 0)
break;
/* FALLTHROUGH */
case KBASE_PM_SUSPEND_HANDLER_DONT_INCREASE:
mutex_unlock(&kbdev->pm.lock);
if (old_count == 0)
kbase_timeline_pm_handle_event(kbdev, KBASE_TIMELINE_PM_EVENT_GPU_ACTIVE);
return 1;
case KBASE_PM_SUSPEND_HANDLER_NOT_POSSIBLE:
/* FALLTHROUGH */
default:
KBASE_DEBUG_ASSERT_MSG(MALI_FALSE, "unreachable");
break;
}
}
c = ++kbdev->pm.active_count;
KBASE_TIMELINE_CONTEXT_ACTIVE(kbdev, c);
KBASE_TRACE_ADD_REFCOUNT(kbdev, PM_CONTEXT_ACTIVE, NULL, NULL, 0u, c);
/* Trace the event being handled */
if (old_count == 0)
kbase_timeline_pm_handle_event(kbdev, KBASE_TIMELINE_PM_EVENT_GPU_ACTIVE);
if (c == 1) {
/* First context active: Power on the GPU and any cores requested by
* the policy */
kbase_pm_update_active(kbdev);
#ifndef MALI_SEC_SEPERATED_UTILIZATION
kbasep_pm_record_gpu_active(kbdev);
#endif
}
mutex_unlock(&kbdev->pm.lock);
return 0;
}
void kbase_pm_suspend(struct kbase_device *kbdev)
{
int nr_keep_gpu_powered_ctxs;
KBASE_DEBUG_ASSERT(kbdev);
mutex_lock(&kbdev->pm.lock);
KBASE_DEBUG_ASSERT(!kbase_pm_is_suspending(kbdev));
kbdev->pm.suspending = MALI_TRUE;
mutex_unlock(&kbdev->pm.lock);
/* From now on, the active count will drop towards zero. Sometimes, it'll
* go up briefly before going down again. However, once it reaches zero it
* will stay there - guaranteeing that we've idled all pm references */
/* Suspend job scheduler and associated components, so that it releases all
* the PM active count references */
kbasep_js_suspend(kbdev);
/* Suspend any counter collection that might be happening */
kbase_instr_hwcnt_suspend(kbdev);
/* Cancel the keep_gpu_powered calls */
for (nr_keep_gpu_powered_ctxs = atomic_read(&kbdev->keep_gpu_powered_count);
nr_keep_gpu_powered_ctxs > 0;
--nr_keep_gpu_powered_ctxs) {
kbase_pm_context_idle(kbdev);
}
/* Wait for the active count to reach zero. This is not the same as
* waiting for a power down, since not all policies power down when this
* reaches zero. */
wait_event(kbdev->pm.zero_active_count_wait, kbdev->pm.active_count == 0);
/* NOTE: We synchronize with anything that was just finishing a
* kbase_pm_context_idle() call by locking the pm.lock below */
/* Force power off the GPU and all cores (regardless of policy), only after
* the PM active count reaches zero (otherwise, we risk turning it off
* prematurely) */
mutex_lock(&kbdev->pm.lock);
kbase_pm_cancel_deferred_poweroff(kbdev);
kbase_pm_do_poweroff(kbdev, MALI_TRUE);
mutex_unlock(&kbdev->pm.lock);
}
/**
* @brief Issue Dump command to hardware
*
* Notes:
* - does not sleep
*/
mali_error kbase_instr_hwcnt_dump_irq(struct kbase_context *kctx)
{
unsigned long flags;
mali_error err = MALI_ERROR_FUNCTION_FAILED;
struct kbase_device *kbdev;
KBASE_DEBUG_ASSERT(NULL != kctx);
kbdev = kctx->kbdev;
KBASE_DEBUG_ASSERT(NULL != kbdev);
spin_lock_irqsave(&kbdev->hwcnt.lock, flags);
if (kbdev->hwcnt.kctx != kctx) {
/* The instrumentation has been setup for another context */
GPU_LOG(DVFS_INFO, DUMMY, 0u, 0u, "hwcnt irq error in %s %d \n", __FUNCTION__, err);
goto unlock;
}
if (kbdev->hwcnt.state != KBASE_INSTR_STATE_IDLE) {
/* HW counters are disabled or another dump is ongoing, or we're resetting */
GPU_LOG(DVFS_INFO, DUMMY, 0u, 0u, "hwcnt disabled or another dump is ongoing in %s %d \n", __FUNCTION__, err);
goto unlock;
}
kbdev->hwcnt.triggered = 0;
/* Mark that we're dumping - the PF handler can signal that we faulted */
kbdev->hwcnt.state = KBASE_INSTR_STATE_DUMPING;
/* Reconfigure the dump address */
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_BASE_LO), kbdev->hwcnt.addr & 0xFFFFFFFF, NULL);
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_BASE_HI), kbdev->hwcnt.addr >> 32, NULL);
/* Start dumping */
KBASE_TRACE_ADD(kbdev, CORE_GPU_PRFCNT_SAMPLE, NULL, NULL, kbdev->hwcnt.addr, 0);
kbase_reg_write(kbdev, GPU_CONTROL_REG(GPU_COMMAND), GPU_COMMAND_PRFCNT_SAMPLE, kctx);
dev_dbg(kbdev->dev, "HW counters dumping done for context %p", kctx);
err = MALI_ERROR_NONE;
unlock:
spin_unlock_irqrestore(&kbdev->hwcnt.lock, flags);
return err;
}
void kbasep_mem_profile_debugfs_add(struct kbase_context *kctx)
{
KBASE_DEBUG_ASSERT(kctx != NULL);
spin_lock_init(&kctx->mem_profile_lock);
debugfs_create_file("mem_profile", S_IRUGO, kctx->kctx_dentry, kctx,
&kbasep_mem_profile_debugfs_fops);
}
void kbase_pm_halt(struct kbase_device *kbdev)
{
KBASE_DEBUG_ASSERT(kbdev != NULL);
mutex_lock(&kbdev->pm.lock);
kbase_pm_cancel_deferred_poweroff(kbdev);
kbase_pm_do_poweroff(kbdev, MALI_FALSE);
mutex_unlock(&kbdev->pm.lock);
}
void kbasep_jd_debugfs_ctx_add(struct kbase_context *kctx)
{
KBASE_DEBUG_ASSERT(kctx != NULL);
/* Expose all atoms */
debugfs_create_file("atoms", S_IRUGO, kctx->kctx_dentry, kctx,
&kbasep_jd_debugfs_atoms_fops);
}
STATIC base_jd_udata kbase_event_process(kbase_context *kctx, kbase_jd_atom *katom)
{
base_jd_udata data;
KBASE_DEBUG_ASSERT(kctx != NULL);
KBASE_DEBUG_ASSERT(katom != NULL);
KBASE_DEBUG_ASSERT(katom->status == KBASE_JD_ATOM_STATE_COMPLETED);
data = katom->udata;
KBASE_TIMELINE_ATOMS_IN_FLIGHT(kctx, atomic_sub_return(1, &kctx->timeline.jd_atoms_in_flight));
katom->status = KBASE_JD_ATOM_STATE_UNUSED;
wake_up(&katom->completed);
return data;
}
void kbasep_mem_profile_debugfs_remove(struct kbase_context *kctx)
{
KBASE_DEBUG_ASSERT(kctx != NULL);
spin_lock(&kctx->mem_profile_lock);
kfree(kctx->mem_profile_data);
kctx->mem_profile_data = NULL;
spin_unlock(&kctx->mem_profile_lock);
}
/* Disable instrumentation and wait for any existing dump to complete
* It's assumed that there's only one privileged context
* Safe to do this without lock when doing an OS suspend, because it only
* changes in response to user-space IOCTLs */
void kbase_instr_hwcnt_suspend(kbase_device *kbdev)
{
kbase_context *kctx;
KBASE_DEBUG_ASSERT(kbdev);
KBASE_DEBUG_ASSERT(!kbdev->hwcnt.suspended_kctx);
kctx = kbdev->hwcnt.kctx;
kbdev->hwcnt.suspended_kctx = kctx;
/* Relevant state was saved into hwcnt.suspended_state when enabling the
* counters */
if (kctx)
{
KBASE_DEBUG_ASSERT(kctx->jctx.sched_info.ctx.flags & KBASE_CTX_FLAG_PRIVILEGED);
kbase_instr_hwcnt_disable(kctx);
}
}
void kbase_pm_context_idle(struct kbase_device *kbdev)
{
int c;
int old_count;
KBASE_DEBUG_ASSERT(kbdev != NULL);
/* Trace timeline information about how long it took to handle the decision
* to powerdown. Sometimes the event might be missed due to reading the
* count outside of mutex, but this is necessary to get the trace timing
* correct. */
old_count = kbdev->pm.active_count;
if (old_count == 0)
kbase_timeline_pm_send_event(kbdev, KBASE_TIMELINE_PM_EVENT_GPU_IDLE);
mutex_lock(&kbdev->pm.lock);
c = --kbdev->pm.active_count;
KBASE_TIMELINE_CONTEXT_ACTIVE(kbdev, c);
KBASE_TRACE_ADD_REFCOUNT(kbdev, PM_CONTEXT_IDLE, NULL, NULL, 0u, c);
KBASE_DEBUG_ASSERT(c >= 0);
/* Trace the event being handled */
if (old_count == 0)
kbase_timeline_pm_handle_event(kbdev, KBASE_TIMELINE_PM_EVENT_GPU_IDLE);
if (c == 0) {
/* Last context has gone idle */
kbase_pm_update_active(kbdev);
#ifndef MALI_SEC_SEPERATED_UTILIZATION
kbasep_pm_record_gpu_idle(kbdev);
#endif
/* Wake up anyone waiting for this to become 0 (e.g. suspend). The
* waiters must synchronize with us by locking the pm.lock after
* waiting */
wake_up(&kbdev->pm.zero_active_count_wait);
}
mutex_unlock(&kbdev->pm.lock);
}
static int gpu_get_clock(struct kbase_device *kbdev)
{
struct exynos_context *platform = (struct exynos_context *) kbdev->platform_context;
if (!platform)
return -ENODEV;
KBASE_DEBUG_ASSERT(kbdev != NULL);
fin_pll = clk_get(kbdev->dev, "fin_pll");
if (IS_ERR(fin_pll) || (fin_pll == NULL)) {
GPU_LOG(DVFS_ERROR, DUMMY, 0u, 0u, "%s: failed to clk_get [fin_pll]\n", __func__);
return -1;
}
fout_g3d_pll = clk_get(kbdev->dev, "fout_g3d_pll");
if (IS_ERR(fout_g3d_pll)) {
GPU_LOG(DVFS_ERROR, DUMMY, 0u, 0u, "%s: failed to clk_get [fout_g3d_pll]\n", __func__);
return -1;
}
aclk_g3d = clk_get(kbdev->dev, "aclk_g3d");
if (IS_ERR(aclk_g3d)) {
GPU_LOG(DVFS_ERROR, DUMMY, 0u, 0u, "%s: failed to clk_get [aclk_g3d]\n", __func__);
return -1;
}
mout_g3d = clk_get(kbdev->dev, "mout_g3d");
if (IS_ERR(mout_g3d)) {
GPU_LOG(DVFS_ERROR, DUMMY, 0u, 0u, "%s: failed to clk_get [mout_g3d]\n", __func__);
return -1;
}
sclk_hpm_g3d = clk_get(kbdev->dev, "sclk_hpm_g3d");
if (IS_ERR(sclk_hpm_g3d)) {
GPU_LOG(DVFS_ERROR, DUMMY, 0u, 0u, "%s: failed to clk_get [sclk_hpm_g3d]\n", __func__);
return -1;
}
aclk_lh_g3d0 = clk_get(kbdev->dev, "aclk_lh_g3d0");
if (IS_ERR(aclk_lh_g3d0)) {
GPU_LOG(DVFS_ERROR, DUMMY, 0u, 0u, "%s: failed to clk_get [aclk_lh_g3d0]\n", __func__);
return -1;
}
aclk_lh_g3d1 = clk_get(kbdev->dev, "aclk_lh_g3d1");
if (IS_ERR(aclk_lh_g3d1)) {
GPU_LOG(DVFS_ERROR, DUMMY, 0u, 0u, "%s: failed to clk_get [aclk_lh_g3d1]\n", __func__);
return -1;
}
__raw_writel(0x1, EXYNOS7420_MUX_SEL_G3D);
gpu_enable_clock(platform);
return 0;
}