diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c | 452 |
1 files changed, 144 insertions, 308 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c index 0723f800e815..a7b17c8deb00 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c @@ -19,19 +19,9 @@ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR * OTHER DEALINGS IN THE SOFTWARE. */ -#undef pr_fmt -#define pr_fmt(fmt) "kfd2kgd: " fmt - -#include <linux/module.h> -#include <linux/fdtable.h> -#include <linux/uaccess.h> -#include <linux/firmware.h> #include <linux/mmu_context.h> -#include <drm/drmP.h> #include "amdgpu.h" #include "amdgpu_amdkfd.h" -#include "amdgpu_ucode.h" -#include "soc15_hw_ip.h" #include "gc/gc_10_1_0_offset.h" #include "gc/gc_10_1_0_sh_mask.h" #include "navi10_enum.h" @@ -43,6 +33,7 @@ #include "v10_structs.h" #include "nv.h" #include "nvd.h" +#include "gfxhub_v2_0.h" enum hqd_dequeue_request_type { NO_ACTION = 0, @@ -51,63 +42,6 @@ enum hqd_dequeue_request_type { SAVE_WAVES }; -/* - * Register access functions - */ - -static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid, - uint32_t sh_mem_config, - uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit, - uint32_t sh_mem_bases); -static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid, - unsigned int vmid); -static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id); -static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id, - uint32_t queue_id, uint32_t __user *wptr, - uint32_t wptr_shift, uint32_t wptr_mask, - struct mm_struct *mm); -static int kgd_hqd_dump(struct kgd_dev *kgd, - uint32_t pipe_id, uint32_t queue_id, - uint32_t (**dump)[2], uint32_t *n_regs); -static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd, - uint32_t __user *wptr, struct mm_struct *mm); -static int kgd_hqd_sdma_dump(struct kgd_dev *kgd, - uint32_t engine_id, uint32_t queue_id, - uint32_t (**dump)[2], uint32_t *n_regs); -static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address, - uint32_t pipe_id, uint32_t queue_id); -static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd); -static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd, - enum kfd_preempt_type reset_type, - unsigned int utimeout, uint32_t pipe_id, - uint32_t queue_id); -static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd, - unsigned int utimeout); -#if 0 -static uint32_t get_watch_base_addr(struct amdgpu_device *adev); -#endif -static int kgd_address_watch_disable(struct kgd_dev *kgd); -static int kgd_address_watch_execute(struct kgd_dev *kgd, - unsigned int watch_point_id, - uint32_t cntl_val, - uint32_t addr_hi, - uint32_t addr_lo); -static int kgd_wave_control_execute(struct kgd_dev *kgd, - uint32_t gfx_index_val, - uint32_t sq_cmd); -static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd, - unsigned int watch_point_id, - unsigned int reg_offset); - -static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd, - uint8_t vmid); -static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd, - uint8_t vmid); -static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid, - uint64_t page_table_base); -static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid); -static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid); - /* Because of REG_GET_FIELD() being used, we put this function in the * asic specific file. */ @@ -140,37 +74,6 @@ static int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd, return 0; } -static const struct kfd2kgd_calls kfd2kgd = { - .program_sh_mem_settings = kgd_program_sh_mem_settings, - .set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping, - .init_interrupts = kgd_init_interrupts, - .hqd_load = kgd_hqd_load, - .hqd_sdma_load = kgd_hqd_sdma_load, - .hqd_dump = kgd_hqd_dump, - .hqd_sdma_dump = kgd_hqd_sdma_dump, - .hqd_is_occupied = kgd_hqd_is_occupied, - .hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied, - .hqd_destroy = kgd_hqd_destroy, - .hqd_sdma_destroy = kgd_hqd_sdma_destroy, - .address_watch_disable = kgd_address_watch_disable, - .address_watch_execute = kgd_address_watch_execute, - .wave_control_execute = kgd_wave_control_execute, - .address_watch_get_offset = kgd_address_watch_get_offset, - .get_atc_vmid_pasid_mapping_pasid = - get_atc_vmid_pasid_mapping_pasid, - .get_atc_vmid_pasid_mapping_valid = - get_atc_vmid_pasid_mapping_valid, - .invalidate_tlbs = invalidate_tlbs, - .invalidate_tlbs_vmid = invalidate_tlbs_vmid, - .set_vm_context_page_table_base = set_vm_context_page_table_base, - .get_tile_config = amdgpu_amdkfd_get_tile_config, -}; - -struct kfd2kgd_calls *amdgpu_amdkfd_gfx_10_0_get_functions() -{ - return (struct kfd2kgd_calls *)&kfd2kgd; -} - static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd) { return (struct amdgpu_device *)kgd; @@ -204,13 +107,13 @@ static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id, lock_srbm(kgd, mec, pipe, queue_id, 0); } -static uint32_t get_queue_mask(struct amdgpu_device *adev, +static uint64_t get_queue_mask(struct amdgpu_device *adev, uint32_t pipe_id, uint32_t queue_id) { - unsigned int bit = (pipe_id * adev->gfx.mec.num_queue_per_pipe + - queue_id) & 31; + unsigned int bit = pipe_id * adev->gfx.mec.num_queue_per_pipe + + queue_id; - return ((uint32_t)1) << bit; + return 1ull << bit; } static void release_queue(struct kgd_dev *kgd) @@ -251,11 +154,6 @@ static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid, ATC_VMID0_PASID_MAPPING__VALID_MASK; pr_debug("pasid 0x%x vmid %d, reg value %x\n", pasid, vmid, pasid_mapping); - /* - * need to do this twice, once for gfx and once for mmhub - * for ATC add 16 to VMID for mmhub, for IH different registers. - * ATC_VMID0..15 registers are separate from ATC_VMID16..31. - */ pr_debug("ATHUB, reg %x\n", SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid); WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid, @@ -307,11 +205,11 @@ static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id) return 0; } -static uint32_t get_sdma_base_addr(struct amdgpu_device *adev, +static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev, unsigned int engine_id, unsigned int queue_id) { - uint32_t base[2] = { + uint32_t sdma_engine_reg_base[2] = { SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL, /* On gfx10, mmSDMA1_xxx registers are defined NOT based @@ -323,12 +221,12 @@ static uint32_t get_sdma_base_addr(struct amdgpu_device *adev, SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL }; - uint32_t retval; - retval = base[engine_id] + queue_id * (mmSDMA0_RLC1_RB_CNTL - - mmSDMA0_RLC0_RB_CNTL); + uint32_t retval = sdma_engine_reg_base[engine_id] + + queue_id * (mmSDMA0_RLC1_RB_CNTL - mmSDMA0_RLC0_RB_CNTL); - pr_debug("sdma base address: 0x%x\n", retval); + pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id, + queue_id, retval); return retval; } @@ -370,21 +268,6 @@ static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id, pr_debug("Load hqd of pipe %d queue %d\n", pipe_id, queue_id); acquire_queue(kgd, pipe_id, queue_id); - /* HIQ is set during driver init period with vmid set to 0*/ - if (m->cp_hqd_vmid == 0) { - uint32_t value, mec, pipe; - - mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1; - pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec); - - pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n", - mec, pipe, queue_id); - value = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS)); - value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1, - ((mec << 5) | (pipe << 3) | queue_id | 0x80)); - WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS), value); - } - /* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */ mqd_hqd = &m->cp_mqd_base_addr_lo; hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR); @@ -434,9 +317,10 @@ static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id, lower_32_bits((uint64_t)wptr)); WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI), upper_32_bits((uint64_t)wptr)); - pr_debug("%s setting CP_PQ_WPTR_POLL_CNTL1 to %x\n", __func__, get_queue_mask(adev, pipe_id, queue_id)); + pr_debug("%s setting CP_PQ_WPTR_POLL_CNTL1 to %x\n", __func__, + (uint32_t)get_queue_mask(adev, pipe_id, queue_id)); WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1), - get_queue_mask(adev, pipe_id, queue_id)); + (uint32_t)get_queue_mask(adev, pipe_id, queue_id)); } /* Start the EOP fetcher */ @@ -452,6 +336,59 @@ static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id, return 0; } +static int kgd_hiq_mqd_load(struct kgd_dev *kgd, void *mqd, + uint32_t pipe_id, uint32_t queue_id, + uint32_t doorbell_off) +{ + struct amdgpu_device *adev = get_amdgpu_device(kgd); + struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring; + struct v10_compute_mqd *m; + uint32_t mec, pipe; + int r; + + m = get_mqd(mqd); + + acquire_queue(kgd, pipe_id, queue_id); + + mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1; + pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec); + + pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n", + mec, pipe, queue_id); + + spin_lock(&adev->gfx.kiq.ring_lock); + r = amdgpu_ring_alloc(kiq_ring, 7); + if (r) { + pr_err("Failed to alloc KIQ (%d).\n", r); + goto out_unlock; + } + + amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5)); + amdgpu_ring_write(kiq_ring, + PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */ + PACKET3_MAP_QUEUES_VMID(m->cp_hqd_vmid) | /* VMID */ + PACKET3_MAP_QUEUES_QUEUE(queue_id) | + PACKET3_MAP_QUEUES_PIPE(pipe) | + PACKET3_MAP_QUEUES_ME((mec - 1)) | + PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */ + PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */ + PACKET3_MAP_QUEUES_ENGINE_SEL(1) | /* engine_sel: hiq */ + PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */ + amdgpu_ring_write(kiq_ring, + PACKET3_MAP_QUEUES_DOORBELL_OFFSET(doorbell_off)); + amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_lo); + amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_hi); + amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_lo); + amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_hi); + amdgpu_ring_commit(kiq_ring); + +out_unlock: + spin_unlock(&adev->gfx.kiq.ring_lock); + release_queue(kgd); + + return r; +} + static int kgd_hqd_dump(struct kgd_dev *kgd, uint32_t pipe_id, uint32_t queue_id, uint32_t (**dump)[2], uint32_t *n_regs) @@ -489,72 +426,67 @@ static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd, { struct amdgpu_device *adev = get_amdgpu_device(kgd); struct v10_sdma_mqd *m; - uint32_t sdma_base_addr, sdmax_gfx_context_cntl; + uint32_t sdma_rlc_reg_offset; unsigned long end_jiffies; uint32_t data; uint64_t data64; uint64_t __user *wptr64 = (uint64_t __user *)wptr; m = get_sdma_mqd(mqd); - sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id, + sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id, m->sdma_queue_id); - pr_debug("sdma load base addr %x for engine %d, queue %d\n", sdma_base_addr, m->sdma_engine_id, m->sdma_queue_id); - sdmax_gfx_context_cntl = m->sdma_engine_id ? - SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_GFX_CONTEXT_CNTL) : - SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_GFX_CONTEXT_CNTL); - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)); end_jiffies = msecs_to_jiffies(2000) + jiffies; while (true) { - data = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS); + data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS); if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK) break; - if (time_after(jiffies, end_jiffies)) + if (time_after(jiffies, end_jiffies)) { + pr_err("SDMA RLC not idle in %s\n", __func__); return -ETIME; + } usleep_range(500, 1000); } - data = RREG32(sdmax_gfx_context_cntl); - data = REG_SET_FIELD(data, SDMA0_GFX_CONTEXT_CNTL, - RESUME_CTX, 0); - WREG32(sdmax_gfx_context_cntl, data); - WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL_OFFSET, + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL_OFFSET, m->sdmax_rlcx_doorbell_offset); data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL, ENABLE, 1); - WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, data); - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR, m->sdmax_rlcx_rb_rptr); - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI, + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data); + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR, + m->sdmax_rlcx_rb_rptr); + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI, m->sdmax_rlcx_rb_rptr_hi); - WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1); + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1); if (read_user_wptr(mm, wptr64, data64)) { - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR, + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR, lower_32_bits(data64)); - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI, + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI, upper_32_bits(data64)); } else { - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR, + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR, m->sdmax_rlcx_rb_rptr); - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI, + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI, m->sdmax_rlcx_rb_rptr_hi); } - WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0); + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0); - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base); - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE_HI, + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base); + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI, m->sdmax_rlcx_rb_base_hi); - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_LO, + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO, m->sdmax_rlcx_rb_rptr_addr_lo); - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_HI, + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI, m->sdmax_rlcx_rb_rptr_addr_hi); data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL, RB_ENABLE, 1); - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, data); + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data); return 0; } @@ -564,28 +496,26 @@ static int kgd_hqd_sdma_dump(struct kgd_dev *kgd, uint32_t (**dump)[2], uint32_t *n_regs) { struct amdgpu_device *adev = get_amdgpu_device(kgd); - uint32_t sdma_base_addr = get_sdma_base_addr(adev, engine_id, queue_id); + uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, + engine_id, queue_id); uint32_t i = 0, reg; #undef HQD_N_REGS #define HQD_N_REGS (19+6+7+10) - pr_debug("sdma dump engine id %d queue_id %d\n", engine_id, queue_id); - pr_debug("sdma base addr %x\n", sdma_base_addr); - *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL); if (*dump == NULL) return -ENOMEM; for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++) - DUMP_REG(sdma_base_addr + reg); + DUMP_REG(sdma_rlc_reg_offset + reg); for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++) - DUMP_REG(sdma_base_addr + reg); + DUMP_REG(sdma_rlc_reg_offset + reg); for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN; reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++) - DUMP_REG(sdma_base_addr + reg); + DUMP_REG(sdma_rlc_reg_offset + reg); for (reg = mmSDMA0_RLC0_MIDCMD_DATA0; reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++) - DUMP_REG(sdma_base_addr + reg); + DUMP_REG(sdma_rlc_reg_offset + reg); WARN_ON_ONCE(i != HQD_N_REGS); *n_regs = i; @@ -619,14 +549,14 @@ static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd) { struct amdgpu_device *adev = get_amdgpu_device(kgd); struct v10_sdma_mqd *m; - uint32_t sdma_base_addr; + uint32_t sdma_rlc_reg_offset; uint32_t sdma_rlc_rb_cntl; m = get_sdma_mqd(mqd); - sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id, + sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id, m->sdma_queue_id); - sdma_rlc_rb_cntl = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL); + sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL); if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK) return true; @@ -747,157 +677,52 @@ static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd, { struct amdgpu_device *adev = get_amdgpu_device(kgd); struct v10_sdma_mqd *m; - uint32_t sdma_base_addr; + uint32_t sdma_rlc_reg_offset; uint32_t temp; unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies; m = get_sdma_mqd(mqd); - sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id, + sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id, m->sdma_queue_id); - temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL); + temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL); temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK; - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, temp); + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp); while (true) { - temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS); + temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS); if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK) break; - if (time_after(jiffies, end_jiffies)) + if (time_after(jiffies, end_jiffies)) { + pr_err("SDMA RLC not idle in %s\n", __func__); return -ETIME; + } usleep_range(500, 1000); } - WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, 0); - WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, - RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL) | + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0); + WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, + RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) | SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK); - m->sdmax_rlcx_rb_rptr = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR); + m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR); m->sdmax_rlcx_rb_rptr_hi = - RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI); + RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI); return 0; } -static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd, - uint8_t vmid) +static bool get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd, + uint8_t vmid, uint16_t *p_pasid) { - uint32_t reg; + uint32_t value; struct amdgpu_device *adev = (struct amdgpu_device *) kgd; - reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid); - return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK; -} - -static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd, - uint8_t vmid) -{ - uint32_t reg; - struct amdgpu_device *adev = (struct amdgpu_device *) kgd; - - reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) - + vmid); - return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK; -} - -static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid) -{ - struct amdgpu_device *adev = (struct amdgpu_device *) kgd; - uint32_t req = (1 << vmid) | - (0 << GCVM_INVALIDATE_ENG0_REQ__FLUSH_TYPE__SHIFT) |/* legacy */ - GCVM_INVALIDATE_ENG0_REQ__INVALIDATE_L2_PTES_MASK | - GCVM_INVALIDATE_ENG0_REQ__INVALIDATE_L2_PDE0_MASK | - GCVM_INVALIDATE_ENG0_REQ__INVALIDATE_L2_PDE1_MASK | - GCVM_INVALIDATE_ENG0_REQ__INVALIDATE_L2_PDE2_MASK | - GCVM_INVALIDATE_ENG0_REQ__INVALIDATE_L1_PTES_MASK; - - mutex_lock(&adev->srbm_mutex); - - /* Use light weight invalidation. - * - * TODO 1: agree on the right set of invalidation registers for - * KFD use. Use the last one for now. Invalidate only GCHUB as - * SDMA is now moved to GCHUB - * - * TODO 2: support range-based invalidation, requires kfg2kgd - * interface change - */ - WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_INVALIDATE_ENG0_ADDR_RANGE_LO32), - 0xffffffff); - WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_INVALIDATE_ENG0_ADDR_RANGE_HI32), - 0x0000001f); - - WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_INVALIDATE_ENG0_REQ), req); - - while (!(RREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_INVALIDATE_ENG0_ACK)) & - (1 << vmid))) - cpu_relax(); - - mutex_unlock(&adev->srbm_mutex); -} - -static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid) -{ - signed long r; - uint32_t seq; - struct amdgpu_ring *ring = &adev->gfx.kiq.ring; - - spin_lock(&adev->gfx.kiq.ring_lock); - amdgpu_ring_alloc(ring, 12); /* fence + invalidate_tlbs package*/ - amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0)); - amdgpu_ring_write(ring, - PACKET3_INVALIDATE_TLBS_DST_SEL(1) | - PACKET3_INVALIDATE_TLBS_PASID(pasid)); - amdgpu_fence_emit_polling(ring, &seq); - amdgpu_ring_commit(ring); - spin_unlock(&adev->gfx.kiq.ring_lock); - - r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout); - if (r < 1) { - DRM_ERROR("wait for kiq fence error: %ld.\n", r); - return -ETIME; - } - - return 0; -} + *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK; -static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid) -{ - struct amdgpu_device *adev = (struct amdgpu_device *) kgd; - int vmid; - struct amdgpu_ring *ring = &adev->gfx.kiq.ring; - - if (amdgpu_emu_mode == 0 && ring->sched.ready) - return invalidate_tlbs_with_kiq(adev, pasid); - - for (vmid = 0; vmid < 16; vmid++) { - if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) - continue; - if (get_atc_vmid_pasid_mapping_valid(kgd, vmid)) { - if (get_atc_vmid_pasid_mapping_pasid(kgd, vmid) - == pasid) { - write_vmid_invalidate_request(kgd, vmid); - break; - } - } - } - - return 0; -} - -static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid) -{ - struct amdgpu_device *adev = (struct amdgpu_device *) kgd; - - if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) { - pr_err("non kfd vmid %d\n", vmid); - return 0; - } - - write_vmid_invalidate_request(kgd, vmid); - return 0; + return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK); } static int kgd_address_watch_disable(struct kgd_dev *kgd) @@ -950,7 +775,6 @@ static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid, uint64_t page_table_base) { struct amdgpu_device *adev = get_amdgpu_device(kgd); - uint64_t base = page_table_base | AMDGPU_PTE_VALID; if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) { pr_err("trying to set page table base for wrong VMID %u\n", @@ -958,18 +782,30 @@ static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid, return; } - /* TODO: take advantage of per-process address space size. For - * now, all processes share the same address space size, like - * on GFX8 and older. - */ - WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_START_ADDR_LO32) + (vmid*2), 0); - WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_START_ADDR_HI32) + (vmid*2), 0); - - WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_END_ADDR_LO32) + (vmid*2), - lower_32_bits(adev->vm_manager.max_pfn - 1)); - WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_END_ADDR_HI32) + (vmid*2), - upper_32_bits(adev->vm_manager.max_pfn - 1)); - - WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_LO32) + (vmid*2), lower_32_bits(base)); - WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_HI32) + (vmid*2), upper_32_bits(base)); + /* SDMA is on gfxhub as well for Navi1* series */ + gfxhub_v2_0_setup_vm_pt_regs(adev, vmid, page_table_base); } + +const struct kfd2kgd_calls gfx_v10_kfd2kgd = { + .program_sh_mem_settings = kgd_program_sh_mem_settings, + .set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping, + .init_interrupts = kgd_init_interrupts, + .hqd_load = kgd_hqd_load, + .hiq_mqd_load = kgd_hiq_mqd_load, + .hqd_sdma_load = kgd_hqd_sdma_load, + .hqd_dump = kgd_hqd_dump, + .hqd_sdma_dump = kgd_hqd_sdma_dump, + .hqd_is_occupied = kgd_hqd_is_occupied, + .hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied, + .hqd_destroy = kgd_hqd_destroy, + .hqd_sdma_destroy = kgd_hqd_sdma_destroy, + .address_watch_disable = kgd_address_watch_disable, + .address_watch_execute = kgd_address_watch_execute, + .wave_control_execute = kgd_wave_control_execute, + .address_watch_get_offset = kgd_address_watch_get_offset, + .get_atc_vmid_pasid_mapping_info = + get_atc_vmid_pasid_mapping_info, + .get_tile_config = amdgpu_amdkfd_get_tile_config, + .set_vm_context_page_table_base = set_vm_context_page_table_base, + .get_hive_id = amdgpu_amdkfd_get_hive_id, +}; |
