void kbasep_8401_workaround_term(kbase_device *kbdev)
{
kbasep_js_device_data *js_devdata;
int i;
u16 restored_as;
OSK_ASSERT(kbdev);
OSK_ASSERT(kbdev->workaround_kctx);
js_devdata = &kbdev->js_data;
for(i = 0; i < KBASE_8401_WORKAROUND_COMPUTEJOB_COUNT; i++)
{
osk_kunmap(kbdev->workaround_compute_job_pa[i], kbdev->workaround_compute_job_va[i]);
}
kbase_phy_pages_free(kbdev, &kbdev->workaround_kctx->pgd_allocator, KBASE_8401_WORKAROUND_COMPUTEJOB_COUNT, kbdev->workaround_compute_job_pa);
kbase_destroy_context(kbdev->workaround_kctx);
kbdev->workaround_kctx = NULL;
/* Free up the workaround address space */
kbdev->nr_hw_address_spaces++;
if ( kbdev->nr_user_address_spaces == (kbdev->nr_hw_address_spaces - 1) )
{
/* Only update nr_user_address_spaces if it was unchanged - to ensure
* HW workarounds that have modified this will still work */
++(kbdev->nr_user_address_spaces);
}
OSK_ASSERT( kbdev->nr_user_address_spaces <= kbdev->nr_hw_address_spaces );
/* Recalculate the free address spaces bit-pattern */
restored_as = (1U << kbdev->nr_hw_address_spaces);
js_devdata->as_free |= restored_as;
}
static base_jd_event_code kbase_dump_cpu_gpu_time(kbase_jd_atom *katom)
{
kbase_va_region *reg;
osk_phy_addr addr;
u64 pfn;
u32 offset;
char *page;
struct timespec ts;
base_dump_cpu_gpu_counters data;
u64 system_time;
u64 cycle_counter;
mali_addr64 jc = katom->jc;
kbase_context *kctx = katom->kctx;
u32 hi1, hi2;
memset(&data, 0, sizeof(data));
kbase_pm_context_active(kctx->kbdev);
/* Read hi, lo, hi to ensure that overflow from lo to hi is handled correctly */
do {
hi1 = kbase_reg_read(kctx->kbdev, GPU_CONTROL_REG(CYCLE_COUNT_HI), NULL);
cycle_counter = kbase_reg_read(kctx->kbdev, GPU_CONTROL_REG(CYCLE_COUNT_LO), NULL);
hi2 = kbase_reg_read(kctx->kbdev, GPU_CONTROL_REG(CYCLE_COUNT_HI), NULL);
cycle_counter |= (((u64)hi1) << 32);
} while (hi1 != hi2);
/* Read hi, lo, hi to ensure that overflow from lo to hi is handled correctly */
do {
hi1 = kbase_reg_read(kctx->kbdev, GPU_CONTROL_REG(TIMESTAMP_HI), NULL);
system_time = kbase_reg_read(kctx->kbdev, GPU_CONTROL_REG(TIMESTAMP_LO), NULL);
hi2 = kbase_reg_read(kctx->kbdev, GPU_CONTROL_REG(TIMESTAMP_HI), NULL);
system_time |= (((u64)hi1) << 32);
} while (hi1 != hi2);
/* Record the CPU's idea of current time */
getnstimeofday(&ts);
kbase_pm_context_idle(kctx->kbdev);
data.sec = ts.tv_sec;
data.usec = ts.tv_nsec / 1000;
data.system_time = system_time;
data.cycle_counter = cycle_counter;
pfn = jc >> 12;
offset = jc & 0xFFF;
if (offset > 0x1000-sizeof(data))
{
/* Wouldn't fit in the page */
return BASE_JD_EVENT_JOB_CANCELLED;
}
reg = kbase_region_tracker_find_region_enclosing_address(kctx, jc);
if (!reg)
{
return BASE_JD_EVENT_JOB_CANCELLED;
}
if (! (reg->flags & KBASE_REG_GPU_WR) )
{
/* Region is not writable by GPU so we won't write to it either */
return BASE_JD_EVENT_JOB_CANCELLED;
}
if (!reg->phy_pages)
{
return BASE_JD_EVENT_JOB_CANCELLED;
}
addr = reg->phy_pages[pfn - reg->start_pfn];
if (!addr)
{
return BASE_JD_EVENT_JOB_CANCELLED;
}
page = osk_kmap(addr);
if (!page)
{
return BASE_JD_EVENT_JOB_CANCELLED;
}
memcpy(page+offset, &data, sizeof(data));
osk_sync_to_cpu(addr+offset, page+offset, sizeof(data));
osk_kunmap(addr, page);
return BASE_JD_EVENT_DONE;
}
mali_error kbasep_8401_workaround_init(kbase_device *kbdev)
{
kbase_context *workaround_kctx;
u32 count;
int i;
OSK_ASSERT(kbdev);
OSK_ASSERT(kbdev->workaround_kctx == NULL);
/* For this workaround we reserve one address space to allow us to
* submit a special job independent of other contexts */
kbdev->nr_address_spaces--;
workaround_kctx = kbase_create_context(kbdev);
if(!workaround_kctx)
{
return MALI_ERROR_FUNCTION_FAILED;
}
/* Allocate the pages required to contain the job */
count = kbase_phy_pages_alloc(workaround_kctx->kbdev,
&workaround_kctx->pgd_allocator,
KBASE_8401_WORKAROUND_COMPUTEJOB_COUNT,
kbdev->workaround_compute_job_pa);
if(count < KBASE_8401_WORKAROUND_COMPUTEJOB_COUNT)
{
goto page_release;
}
/* Get virtual address of mapped memory and write a compute job for each page */
for(i = 0; i < KBASE_8401_WORKAROUND_COMPUTEJOB_COUNT; i++)
{
kbdev->workaround_compute_job_va[i] = osk_kmap(kbdev->workaround_compute_job_pa[i]);
if(NULL == kbdev->workaround_compute_job_va[i])
{
goto page_free;
}
/* Generate the compute job data */
kbasep_8401_workaround_update_job_pointers((u32*)kbdev->workaround_compute_job_va[i], i);
}
/* Insert pages to the gpu mmu. */
kbase_mmu_insert_pages(workaround_kctx,
/* vpfn = page number */
(u64)WORKAROUND_PAGE_OFFSET,
/* physical address */
kbdev->workaround_compute_job_pa,
/* number of pages */
KBASE_8401_WORKAROUND_COMPUTEJOB_COUNT,
/* flags */
KBASE_REG_CPU_RW|KBASE_REG_GPU_RW);
kbdev->workaround_kctx = workaround_kctx;
return MALI_ERROR_NONE;
page_free:
while(i--)
{
osk_kunmap(kbdev->workaround_compute_job_pa[i], kbdev->workaround_compute_job_va[i]);
}
page_release:
kbase_phy_pages_free(kbdev, &workaround_kctx->pgd_allocator, count, kbdev->workaround_compute_job_pa);
kbase_destroy_context(workaround_kctx);
return MALI_ERROR_FUNCTION_FAILED;
}
mali_error kbasep_8401_workaround_init(kbase_device *kbdev)
{
kbasep_js_device_data *js_devdata;
kbase_context *workaround_kctx;
u32 count;
int i;
u16 as_present_mask;
OSK_ASSERT(kbdev);
OSK_ASSERT(kbdev->workaround_kctx == NULL);
js_devdata = &kbdev->js_data;
/* For this workaround we reserve one address space to allow us to
* submit a special job independent of other contexts */
--(kbdev->nr_hw_address_spaces);
if ( kbdev->nr_user_address_spaces == (kbdev->nr_hw_address_spaces + 1) )
{
/* Only update nr_user_address_spaces if it was unchanged - to ensure
* HW workarounds that have modified this will still work */
--(kbdev->nr_user_address_spaces);
}
OSK_ASSERT( kbdev->nr_user_address_spaces <= kbdev->nr_hw_address_spaces );
/* Recalculate the free address spaces bit-pattern */
as_present_mask = (1U << kbdev->nr_hw_address_spaces) - 1;
js_devdata->as_free &= as_present_mask;
workaround_kctx = kbase_create_context(kbdev);
if(!workaround_kctx)
{
return MALI_ERROR_FUNCTION_FAILED;
}
/* Allocate the pages required to contain the job */
count = kbase_phy_pages_alloc(workaround_kctx->kbdev,
&workaround_kctx->pgd_allocator,
KBASE_8401_WORKAROUND_COMPUTEJOB_COUNT,
kbdev->workaround_compute_job_pa);
if(count < KBASE_8401_WORKAROUND_COMPUTEJOB_COUNT)
{
goto page_release;
}
/* Get virtual address of mapped memory and write a compute job for each page */
for(i = 0; i < KBASE_8401_WORKAROUND_COMPUTEJOB_COUNT; i++)
{
kbdev->workaround_compute_job_va[i] = osk_kmap(kbdev->workaround_compute_job_pa[i]);
if(NULL == kbdev->workaround_compute_job_va[i])
{
goto page_free;
}
/* Generate the compute job data */
kbasep_8401_workaround_update_job_pointers((u32*)kbdev->workaround_compute_job_va[i], i);
}
/* Insert pages to the gpu mmu. */
kbase_mmu_insert_pages(workaround_kctx,
/* vpfn = page number */
(u64)WORKAROUND_PAGE_OFFSET,
/* physical address */
kbdev->workaround_compute_job_pa,
/* number of pages */
KBASE_8401_WORKAROUND_COMPUTEJOB_COUNT,
/* flags */
KBASE_REG_GPU_RD|KBASE_REG_CPU_RD|KBASE_REG_CPU_WR|KBASE_REG_GPU_WR);
kbdev->workaround_kctx = workaround_kctx;
return MALI_ERROR_NONE;
page_free:
while(i--)
{
osk_kunmap(kbdev->workaround_compute_job_pa[i], kbdev->workaround_compute_job_va[i]);
}
page_release:
kbase_phy_pages_free(kbdev, &workaround_kctx->pgd_allocator, count, kbdev->workaround_compute_job_pa);
kbase_destroy_context(workaround_kctx);
return MALI_ERROR_FUNCTION_FAILED;
}
static base_jd_event_code kbase_dump_cpu_gpu_time(kbase_context *kctx, mali_addr64 jc)
{
kbase_va_region *reg;
osk_phy_addr addr;
u64 pfn;
u32 offset;
char *page;
osk_timeval tv;
base_dump_cpu_gpu_counters data;
u64 system_time;
u64 cycle_counter;
u32 hi1, hi2;
OSK_MEMSET(&data, 0, sizeof(data));
/* Read hi, lo, hi to ensure that overflow from lo to hi is handled correctly */
do {
hi1 = kbase_reg_read(kctx->kbdev, GPU_CONTROL_REG(CYCLE_COUNT_HI), NULL);
cycle_counter = kbase_reg_read(kctx->kbdev, GPU_CONTROL_REG(CYCLE_COUNT_LO), NULL);
hi2 = kbase_reg_read(kctx->kbdev, GPU_CONTROL_REG(CYCLE_COUNT_HI), NULL);
cycle_counter |= (((u64)hi1) << 32);
} while (hi1 != hi2);
/* Read hi, lo, hi to ensure that overflow from lo to hi is handled correctly */
do {
hi1 = kbase_reg_read(kctx->kbdev, GPU_CONTROL_REG(TIMESTAMP_HI), NULL);
system_time = kbase_reg_read(kctx->kbdev, GPU_CONTROL_REG(TIMESTAMP_LO), NULL);
hi2 = kbase_reg_read(kctx->kbdev, GPU_CONTROL_REG(TIMESTAMP_HI), NULL);
system_time |= (((u64)hi1) << 32);
} while (hi1 != hi2);
/* Record the CPU's idea of current time */
osk_gettimeofday(&tv);
data.sec = tv.tv_sec;
data.usec = tv.tv_usec;
data.system_time = system_time;
data.cycle_counter = cycle_counter;
pfn = jc >> 12;
offset = jc & 0xFFF;
if (offset > 0x1000-sizeof(data))
{
/* Wouldn't fit in the page */
return BASE_JD_EVENT_JOB_CANCELLED;
}
reg = kbase_region_lookup(kctx, jc);
if (!reg)
{
return BASE_JD_EVENT_JOB_CANCELLED;
}
if (! (reg->flags & KBASE_REG_GPU_RW) )
{
/* Region is not writable by GPU so we won't write to it either */
return BASE_JD_EVENT_JOB_CANCELLED;
}
if (!reg->phy_pages)
{
return BASE_JD_EVENT_JOB_CANCELLED;
}
addr = reg->phy_pages[pfn - reg->start_pfn];
if (!addr)
{
return BASE_JD_EVENT_JOB_CANCELLED;
}
page = osk_kmap(addr);
if (!page)
{
return BASE_JD_EVENT_JOB_CANCELLED;
}
memcpy(page+offset, &data, sizeof(data));
osk_kunmap(addr, page);
return BASE_JD_EVENT_DONE;
}
