[PATCH 11/83] hsa/radeon: Add scheduler code
Jerome Glisse
j.glisse at gmail.com
Fri Jul 11 11:25:22 PDT 2014
On Fri, Jul 11, 2014 at 12:50:11AM +0300, Oded Gabbay wrote:
> This patch adds the code base of the scheduler, which handles queue
> creation, deletion and scheduling on the CP of the GPU.
>
> Signed-off-by: Oded Gabbay <oded.gabbay at amd.com>
I would rather see all this squashed, this gave feeling that driver
can access register which is latter remove. I know jungling with
patch squashing can be daunting but really it makes reviewing hard
here because i have to jump back and forth to see if thing i am looking
at really matter in the final version.
Cheers,
Jérôme
> ---
> drivers/gpu/hsa/radeon/Makefile | 3 +-
> drivers/gpu/hsa/radeon/cik_regs.h | 213 +++++++
> drivers/gpu/hsa/radeon/kfd_device.c | 1 +
> drivers/gpu/hsa/radeon/kfd_registers.c | 50 ++
> drivers/gpu/hsa/radeon/kfd_sched_cik_static.c | 800 ++++++++++++++++++++++++++
> drivers/gpu/hsa/radeon/kfd_vidmem.c | 61 ++
> 6 files changed, 1127 insertions(+), 1 deletion(-)
> create mode 100644 drivers/gpu/hsa/radeon/cik_regs.h
> create mode 100644 drivers/gpu/hsa/radeon/kfd_registers.c
> create mode 100644 drivers/gpu/hsa/radeon/kfd_sched_cik_static.c
> create mode 100644 drivers/gpu/hsa/radeon/kfd_vidmem.c
>
> diff --git a/drivers/gpu/hsa/radeon/Makefile b/drivers/gpu/hsa/radeon/Makefile
> index 989518a..28da10c 100644
> --- a/drivers/gpu/hsa/radeon/Makefile
> +++ b/drivers/gpu/hsa/radeon/Makefile
> @@ -4,6 +4,7 @@
>
> radeon_kfd-y := kfd_module.o kfd_device.o kfd_chardev.o \
> kfd_pasid.o kfd_topology.o kfd_process.o \
> - kfd_doorbell.o
> + kfd_doorbell.o kfd_sched_cik_static.o kfd_registers.o \
> + kfd_vidmem.o
>
> obj-$(CONFIG_HSA_RADEON) += radeon_kfd.o
> diff --git a/drivers/gpu/hsa/radeon/cik_regs.h b/drivers/gpu/hsa/radeon/cik_regs.h
> new file mode 100644
> index 0000000..d0cdc57
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/cik_regs.h
> @@ -0,0 +1,213 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#ifndef CIK_REGS_H
> +#define CIK_REGS_H
> +
> +#define BIF_DOORBELL_CNTL 0x530Cu
> +
> +#define SRBM_GFX_CNTL 0xE44
> +#define PIPEID(x) ((x) << 0)
> +#define MEID(x) ((x) << 2)
> +#define VMID(x) ((x) << 4)
> +#define QUEUEID(x) ((x) << 8)
> +
> +#define SQ_CONFIG 0x8C00
> +
> +#define SH_MEM_BASES 0x8C28
> +/* if PTR32, these are the bases for scratch and lds */
> +#define PRIVATE_BASE(x) ((x) << 0) /* scratch */
> +#define SHARED_BASE(x) ((x) << 16) /* LDS */
> +#define SH_MEM_APE1_BASE 0x8C2C
> +/* if PTR32, this is the base location of GPUVM */
> +#define SH_MEM_APE1_LIMIT 0x8C30
> +/* if PTR32, this is the upper limit of GPUVM */
> +#define SH_MEM_CONFIG 0x8C34
> +#define PTR32 (1 << 0)
> +#define ALIGNMENT_MODE(x) ((x) << 2)
> +#define SH_MEM_ALIGNMENT_MODE_DWORD 0
> +#define SH_MEM_ALIGNMENT_MODE_DWORD_STRICT 1
> +#define SH_MEM_ALIGNMENT_MODE_STRICT 2
> +#define SH_MEM_ALIGNMENT_MODE_UNALIGNED 3
> +#define DEFAULT_MTYPE(x) ((x) << 4)
> +#define APE1_MTYPE(x) ((x) << 7)
> +
> +/* valid for both DEFAULT_MTYPE and APE1_MTYPE */
> +#define MTYPE_NONCACHED 3
> +
> +
> +#define SH_STATIC_MEM_CONFIG 0x9604u
> +
> +#define TC_CFG_L1_LOAD_POLICY0 0xAC68
> +#define TC_CFG_L1_LOAD_POLICY1 0xAC6C
> +#define TC_CFG_L1_STORE_POLICY 0xAC70
> +#define TC_CFG_L2_LOAD_POLICY0 0xAC74
> +#define TC_CFG_L2_LOAD_POLICY1 0xAC78
> +#define TC_CFG_L2_STORE_POLICY0 0xAC7C
> +#define TC_CFG_L2_STORE_POLICY1 0xAC80
> +#define TC_CFG_L2_ATOMIC_POLICY 0xAC84
> +#define TC_CFG_L1_VOLATILE 0xAC88
> +#define TC_CFG_L2_VOLATILE 0xAC8C
> +
> +#define CP_PQ_WPTR_POLL_CNTL 0xC20C
> +#define WPTR_POLL_EN (1 << 31)
> +
> +#define CP_ME1_PIPE0_INT_CNTL 0xC214
> +#define CP_ME1_PIPE1_INT_CNTL 0xC218
> +#define CP_ME1_PIPE2_INT_CNTL 0xC21C
> +#define CP_ME1_PIPE3_INT_CNTL 0xC220
> +#define CP_ME2_PIPE0_INT_CNTL 0xC224
> +#define CP_ME2_PIPE1_INT_CNTL 0xC228
> +#define CP_ME2_PIPE2_INT_CNTL 0xC22C
> +#define CP_ME2_PIPE3_INT_CNTL 0xC230
> +#define DEQUEUE_REQUEST_INT_ENABLE (1 << 13)
> +#define WRM_POLL_TIMEOUT_INT_ENABLE (1 << 17)
> +#define PRIV_REG_INT_ENABLE (1 << 23)
> +#define TIME_STAMP_INT_ENABLE (1 << 26)
> +#define GENERIC2_INT_ENABLE (1 << 29)
> +#define GENERIC1_INT_ENABLE (1 << 30)
> +#define GENERIC0_INT_ENABLE (1 << 31)
> +#define CP_ME1_PIPE0_INT_STATUS 0xC214
> +#define CP_ME1_PIPE1_INT_STATUS 0xC218
> +#define CP_ME1_PIPE2_INT_STATUS 0xC21C
> +#define CP_ME1_PIPE3_INT_STATUS 0xC220
> +#define CP_ME2_PIPE0_INT_STATUS 0xC224
> +#define CP_ME2_PIPE1_INT_STATUS 0xC228
> +#define CP_ME2_PIPE2_INT_STATUS 0xC22C
> +#define CP_ME2_PIPE3_INT_STATUS 0xC230
> +#define DEQUEUE_REQUEST_INT_STATUS (1 << 13)
> +#define WRM_POLL_TIMEOUT_INT_STATUS (1 << 17)
> +#define PRIV_REG_INT_STATUS (1 << 23)
> +#define TIME_STAMP_INT_STATUS (1 << 26)
> +#define GENERIC2_INT_STATUS (1 << 29)
> +#define GENERIC1_INT_STATUS (1 << 30)
> +#define GENERIC0_INT_STATUS (1 << 31)
> +
> +#define CP_HPD_EOP_BASE_ADDR 0xC904
> +#define CP_HPD_EOP_BASE_ADDR_HI 0xC908
> +#define CP_HPD_EOP_VMID 0xC90C
> +#define CP_HPD_EOP_CONTROL 0xC910
> +#define EOP_SIZE(x) ((x) << 0)
> +#define EOP_SIZE_MASK (0x3f << 0)
> +#define CP_MQD_BASE_ADDR 0xC914
> +#define CP_MQD_BASE_ADDR_HI 0xC918
> +#define CP_HQD_ACTIVE 0xC91C
> +#define CP_HQD_VMID 0xC920
> +
> +#define CP_HQD_PERSISTENT_STATE 0xC924u
> +#define DEFAULT_CP_HQD_PERSISTENT_STATE (0x33U << 8)
> +
> +#define CP_HQD_PIPE_PRIORITY 0xC928u
> +#define CP_HQD_QUEUE_PRIORITY 0xC92Cu
> +#define CP_HQD_QUANTUM 0xC930u
> +#define QUANTUM_EN 1U
> +#define QUANTUM_SCALE_1MS (1U << 4)
> +#define QUANTUM_DURATION(x) ((x) << 8)
> +
> +#define CP_HQD_PQ_BASE 0xC934
> +#define CP_HQD_PQ_BASE_HI 0xC938
> +#define CP_HQD_PQ_RPTR 0xC93C
> +#define CP_HQD_PQ_RPTR_REPORT_ADDR 0xC940
> +#define CP_HQD_PQ_RPTR_REPORT_ADDR_HI 0xC944
> +#define CP_HQD_PQ_WPTR_POLL_ADDR 0xC948
> +#define CP_HQD_PQ_WPTR_POLL_ADDR_HI 0xC94C
> +#define CP_HQD_PQ_DOORBELL_CONTROL 0xC950
> +#define DOORBELL_OFFSET(x) ((x) << 2)
> +#define DOORBELL_OFFSET_MASK (0x1fffff << 2)
> +#define DOORBELL_SOURCE (1 << 28)
> +#define DOORBELL_SCHD_HIT (1 << 29)
> +#define DOORBELL_EN (1 << 30)
> +#define DOORBELL_HIT (1 << 31)
> +#define CP_HQD_PQ_WPTR 0xC954
> +#define CP_HQD_PQ_CONTROL 0xC958
> +#define QUEUE_SIZE(x) ((x) << 0)
> +#define QUEUE_SIZE_MASK (0x3f << 0)
> +#define RPTR_BLOCK_SIZE(x) ((x) << 8)
> +#define RPTR_BLOCK_SIZE_MASK (0x3f << 8)
> +#define MIN_AVAIL_SIZE(x) ((x) << 20)
> +#define PQ_ATC_EN (1 << 23)
> +#define PQ_VOLATILE (1 << 26)
> +#define NO_UPDATE_RPTR (1 << 27)
> +#define UNORD_DISPATCH (1 << 28)
> +#define ROQ_PQ_IB_FLIP (1 << 29)
> +#define PRIV_STATE (1 << 30)
> +#define KMD_QUEUE (1 << 31)
> +
> +#define DEFAULT_RPTR_BLOCK_SIZE RPTR_BLOCK_SIZE(5)
> +#define DEFAULT_MIN_AVAIL_SIZE MIN_AVAIL_SIZE(3)
> +
> +#define CP_HQD_IB_BASE_ADDR 0xC95Cu
> +#define CP_HQD_IB_BASE_ADDR_HI 0xC960u
> +#define CP_HQD_IB_RPTR 0xC964u
> +#define CP_HQD_IB_CONTROL 0xC968u
> +#define IB_ATC_EN (1U << 23)
> +#define DEFAULT_MIN_IB_AVAIL_SIZE (3U << 20)
> +
> +#define CP_HQD_DEQUEUE_REQUEST 0xC974
> +#define DEQUEUE_REQUEST_DRAIN 1
> +
> +#define CP_HQD_SEMA_CMD 0xC97Cu
> +#define CP_HQD_MSG_TYPE 0xC980u
> +#define CP_HQD_ATOMIC0_PREOP_LO 0xC984u
> +#define CP_HQD_ATOMIC0_PREOP_HI 0xC988u
> +#define CP_HQD_ATOMIC1_PREOP_LO 0xC98Cu
> +#define CP_HQD_ATOMIC1_PREOP_HI 0xC990u
> +#define CP_HQD_HQ_SCHEDULER0 0xC994u
> +#define CP_HQD_HQ_SCHEDULER1 0xC998u
> +
> +
> +#define CP_MQD_CONTROL 0xC99C
> +#define MQD_VMID(x) ((x) << 0)
> +#define MQD_VMID_MASK (0xf << 0)
> +#define MQD_CONTROL_PRIV_STATE_EN (1U << 8)
> +
> +#define GRBM_GFX_INDEX 0x30800
> +#define INSTANCE_INDEX(x) ((x) << 0)
> +#define SH_INDEX(x) ((x) << 8)
> +#define SE_INDEX(x) ((x) << 16)
> +#define SH_BROADCAST_WRITES (1 << 29)
> +#define INSTANCE_BROADCAST_WRITES (1 << 30)
> +#define SE_BROADCAST_WRITES (1 << 31)
> +
> +#define SQC_CACHES 0x30d20
> +#define SQC_POLICY 0x8C38u
> +#define SQC_VOLATILE 0x8C3Cu
> +
> +#define CP_PERFMON_CNTL 0x36020
> +
> +#define ATC_VMID0_PASID_MAPPING 0x339Cu
> +#define ATC_VMID_PASID_MAPPING_UPDATE_STATUS 0x3398u
> +#define ATC_VMID_PASID_MAPPING_VALID (1U << 31)
> +
> +#define ATC_VM_APERTURE0_CNTL 0x3310u
> +#define ATS_ACCESS_MODE_NEVER 0
> +#define ATS_ACCESS_MODE_ALWAYS 1
> +
> +#define ATC_VM_APERTURE0_CNTL2 0x3318u
> +#define ATC_VM_APERTURE0_HIGH_ADDR 0x3308u
> +#define ATC_VM_APERTURE0_LOW_ADDR 0x3300u
> +#define ATC_VM_APERTURE1_CNTL 0x3314u
> +#define ATC_VM_APERTURE1_CNTL2 0x331Cu
> +#define ATC_VM_APERTURE1_HIGH_ADDR 0x330Cu
> +#define ATC_VM_APERTURE1_LOW_ADDR 0x3304u
> +
> +#endif
> diff --git a/drivers/gpu/hsa/radeon/kfd_device.c b/drivers/gpu/hsa/radeon/kfd_device.c
> index 4e9fe6c..465c822 100644
> --- a/drivers/gpu/hsa/radeon/kfd_device.c
> +++ b/drivers/gpu/hsa/radeon/kfd_device.c
> @@ -28,6 +28,7 @@
> #include "kfd_scheduler.h"
>
> static const struct kfd_device_info bonaire_device_info = {
> + .scheduler_class = &radeon_kfd_cik_static_scheduler_class,
> .max_pasid_bits = 16,
> };
>
> diff --git a/drivers/gpu/hsa/radeon/kfd_registers.c b/drivers/gpu/hsa/radeon/kfd_registers.c
> new file mode 100644
> index 0000000..223debd
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_registers.c
> @@ -0,0 +1,50 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#include <linux/io.h>
> +#include "kfd_priv.h"
> +
> +/* In KFD, "reg" is the byte offset of the register. */
> +static void __iomem *reg_address(struct kfd_dev *dev, uint32_t reg)
> +{
> + return dev->regs + reg;
> +}
> +
> +void radeon_kfd_write_reg(struct kfd_dev *dev, uint32_t reg, uint32_t value)
> +{
> + writel(value, reg_address(dev, reg));
> +}
> +
> +uint32_t radeon_kfd_read_reg(struct kfd_dev *dev, uint32_t reg)
> +{
> + return readl(reg_address(dev, reg));
> +}
> +
> +void radeon_kfd_lock_srbm_index(struct kfd_dev *dev)
> +{
> + kfd2kgd->lock_srbm_gfx_cntl(dev->kgd);
> +}
> +
> +void radeon_kfd_unlock_srbm_index(struct kfd_dev *dev)
> +{
> + kfd2kgd->unlock_srbm_gfx_cntl(dev->kgd);
> +}
> diff --git a/drivers/gpu/hsa/radeon/kfd_sched_cik_static.c b/drivers/gpu/hsa/radeon/kfd_sched_cik_static.c
> new file mode 100644
> index 0000000..b986ff9
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_sched_cik_static.c
> @@ -0,0 +1,800 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#include <linux/log2.h>
> +#include <linux/mutex.h>
> +#include <linux/slab.h>
> +#include <linux/types.h>
> +#include <linux/uaccess.h>
> +#include "kfd_priv.h"
> +#include "kfd_scheduler.h"
> +#include "cik_regs.h"
> +
> +/* CIK CP hardware is arranged with 8 queues per pipe and 8 pipes per MEC (microengine for compute).
> + * The first MEC is ME 1 with the GFX ME as ME 0.
> + * We split the CP with the KGD, they take the first N pipes and we take the rest.
> + */
> +#define CIK_QUEUES_PER_PIPE 8
> +#define CIK_PIPES_PER_MEC 4
> +
> +#define CIK_MAX_PIPES (2 * CIK_PIPES_PER_MEC)
> +
> +#define CIK_NUM_VMID 16
> +
> +#define CIK_HPD_SIZE_LOG2 11
> +#define CIK_HPD_SIZE (1U << CIK_HPD_SIZE_LOG2)
> +#define CIK_HPD_ALIGNMENT 256
> +#define CIK_MQD_ALIGNMENT 4
> +
> +#pragma pack(push, 4)
> +
> +struct cik_hqd_registers {
> + u32 cp_mqd_base_addr;
> + u32 cp_mqd_base_addr_hi;
> + u32 cp_hqd_active;
> + u32 cp_hqd_vmid;
> + u32 cp_hqd_persistent_state;
> + u32 cp_hqd_pipe_priority;
> + u32 cp_hqd_queue_priority;
> + u32 cp_hqd_quantum;
> + u32 cp_hqd_pq_base;
> + u32 cp_hqd_pq_base_hi;
> + u32 cp_hqd_pq_rptr;
> + u32 cp_hqd_pq_rptr_report_addr;
> + u32 cp_hqd_pq_rptr_report_addr_hi;
> + u32 cp_hqd_pq_wptr_poll_addr;
> + u32 cp_hqd_pq_wptr_poll_addr_hi;
> + u32 cp_hqd_pq_doorbell_control;
> + u32 cp_hqd_pq_wptr;
> + u32 cp_hqd_pq_control;
> + u32 cp_hqd_ib_base_addr;
> + u32 cp_hqd_ib_base_addr_hi;
> + u32 cp_hqd_ib_rptr;
> + u32 cp_hqd_ib_control;
> + u32 cp_hqd_iq_timer;
> + u32 cp_hqd_iq_rptr;
> + u32 cp_hqd_dequeue_request;
> + u32 cp_hqd_dma_offload;
> + u32 cp_hqd_sema_cmd;
> + u32 cp_hqd_msg_type;
> + u32 cp_hqd_atomic0_preop_lo;
> + u32 cp_hqd_atomic0_preop_hi;
> + u32 cp_hqd_atomic1_preop_lo;
> + u32 cp_hqd_atomic1_preop_hi;
> + u32 cp_hqd_hq_scheduler0;
> + u32 cp_hqd_hq_scheduler1;
> + u32 cp_mqd_control;
> +};
> +
> +struct cik_mqd {
> + u32 header;
> + u32 dispatch_initiator;
> + u32 dimensions[3];
> + u32 start_idx[3];
> + u32 num_threads[3];
> + u32 pipeline_stat_enable;
> + u32 perf_counter_enable;
> + u32 pgm[2];
> + u32 tba[2];
> + u32 tma[2];
> + u32 pgm_rsrc[2];
> + u32 vmid;
> + u32 resource_limits;
> + u32 static_thread_mgmt01[2];
> + u32 tmp_ring_size;
> + u32 static_thread_mgmt23[2];
> + u32 restart[3];
> + u32 thread_trace_enable;
> + u32 reserved1;
> + u32 user_data[16];
> + u32 vgtcs_invoke_count[2];
> + struct cik_hqd_registers queue_state;
> + u32 dequeue_cntr;
> + u32 interrupt_queue[64];
> +};
> +
> +struct cik_mqd_padded {
> + struct cik_mqd mqd;
> + u8 padding[1024 - sizeof(struct cik_mqd)]; /* Pad MQD out to 1KB. (HW requires 4-byte alignment.) */
> +};
> +
> +#pragma pack(pop)
> +
> +struct cik_static_private {
> + struct kfd_dev *dev;
> +
> + struct mutex mutex;
> +
> + unsigned int first_pipe;
> + unsigned int num_pipes;
> +
> + unsigned long free_vmid_mask; /* unsigned long to make set/clear_bit happy */
> +
> + /* Everything below here is offset by first_pipe. E.g. bit 0 in
> + * free_queues is queue 0 in pipe first_pipe
> + */
> +
> + /* Queue q on pipe p is at bit QUEUES_PER_PIPE * p + q. */
> + unsigned long free_queues[DIV_ROUND_UP(CIK_MAX_PIPES * CIK_QUEUES_PER_PIPE, BITS_PER_LONG)];
> +
> + kfd_mem_obj hpd_mem; /* Single allocation for HPDs for all KFD pipes. */
> + kfd_mem_obj mqd_mem; /* Single allocation for all MQDs for all KFD
> + * pipes. This is actually struct cik_mqd_padded. */
> + uint64_t hpd_addr; /* GPU address for hpd_mem. */
> + uint64_t mqd_addr; /* GPU address for mqd_mem. */
> + /*
> + * Pointer for mqd_mem.
> + * We keep this mapped because multiple processes may need to access it
> + * in parallel and this is simpler than controlling concurrent kmaps
> + */
> + struct cik_mqd_padded *mqds;
> +};
> +
> +struct cik_static_process {
> + unsigned int vmid;
> + pasid_t pasid;
> +};
> +
> +struct cik_static_queue {
> + unsigned int queue; /* + first_pipe * QUEUES_PER_PIPE */
> +
> + uint64_t mqd_addr;
> + struct cik_mqd *mqd;
> +
> + void __user *pq_addr;
> + void __user *rptr_address;
> + doorbell_t __user *wptr_address;
> + uint32_t doorbell_index;
> +
> + uint32_t queue_size_encoded; /* CP_HQD_PQ_CONTROL.QUEUE_SIZE takes the queue size as log2(size) - 3. */
> +};
> +
> +static uint32_t lower_32(uint64_t x)
> +{
> + return (uint32_t)x;
> +}
> +
> +static uint32_t upper_32(uint64_t x)
> +{
> + return (uint32_t)(x >> 32);
> +}
> +
> +/* SRBM_GFX_CNTL provides the MEC/pipe/queue and vmid for many registers that are
> + * In particular, CP_HQD_* and CP_MQD_* are instanced for each queue. CP_HPD_* are instanced for each pipe.
> + * SH_MEM_* are instanced per-VMID.
> + *
> + * We provide queue_select, pipe_select and vmid_select helpers that should be used before accessing
> + * registers from those groups. Note that these overwrite each other, e.g. after vmid_select the current
> + * selected MEC/pipe/queue is undefined.
> + *
> + * SRBM_GFX_CNTL and the registers it indexes are shared with KGD. You must be holding the srbm_gfx_cntl
> + * lock via lock_srbm_index before setting SRBM_GFX_CNTL or accessing any of the instanced registers.
> + */
> +static uint32_t make_srbm_gfx_cntl_mpqv(unsigned int me, unsigned int pipe, unsigned int queue, unsigned int vmid)
> +{
> + return QUEUEID(queue) | VMID(vmid) | MEID(me) | PIPEID(pipe);
> +}
> +
> +static void pipe_select(struct cik_static_private *priv, unsigned int pipe)
> +{
> + unsigned int pipe_in_mec = (pipe + priv->first_pipe) % CIK_PIPES_PER_MEC;
> + unsigned int mec = (pipe + priv->first_pipe) / CIK_PIPES_PER_MEC;
> +
> + WRITE_REG(priv->dev, SRBM_GFX_CNTL, make_srbm_gfx_cntl_mpqv(mec+1, pipe_in_mec, 0, 0));
> +}
> +
> +static void queue_select(struct cik_static_private *priv, unsigned int queue)
> +{
> + unsigned int queue_in_pipe = queue % CIK_QUEUES_PER_PIPE;
> + unsigned int pipe = queue / CIK_QUEUES_PER_PIPE + priv->first_pipe;
> + unsigned int pipe_in_mec = pipe % CIK_PIPES_PER_MEC;
> + unsigned int mec = pipe / CIK_PIPES_PER_MEC;
> +
> +#if 0
> + dev_err(radeon_kfd_chardev(), "queue select %d = %u/%u/%u = 0x%08x\n", queue, mec+1, pipe_in_mec, queue_in_pipe,
> + make_srbm_gfx_cntl_mpqv(mec+1, pipe_in_mec, queue_in_pipe, 0));
> +#endif
> +
> + WRITE_REG(priv->dev, SRBM_GFX_CNTL, make_srbm_gfx_cntl_mpqv(mec+1, pipe_in_mec, queue_in_pipe, 0));
> +}
> +
> +static void vmid_select(struct cik_static_private *priv, unsigned int vmid)
> +{
> + WRITE_REG(priv->dev, SRBM_GFX_CNTL, make_srbm_gfx_cntl_mpqv(0, 0, 0, vmid));
> +}
> +
> +static void lock_srbm_index(struct cik_static_private *priv)
> +{
> + radeon_kfd_lock_srbm_index(priv->dev);
> +}
> +
> +static void unlock_srbm_index(struct cik_static_private *priv)
> +{
> + WRITE_REG(priv->dev, SRBM_GFX_CNTL, 0); /* Be nice to KGD, reset indexed CP registers to the GFX pipe. */
> + radeon_kfd_unlock_srbm_index(priv->dev);
> +}
> +
> +/* One-time setup for all compute pipes. They need to be programmed with the address & size of the HPD EOP buffer. */
> +static void init_pipes(struct cik_static_private *priv)
> +{
> + unsigned int i;
> +
> + lock_srbm_index(priv);
> +
> + for (i = 0; i < priv->num_pipes; i++) {
> + uint64_t pipe_hpd_addr = priv->hpd_addr + i * CIK_HPD_SIZE;
> +
> + pipe_select(priv, i);
> +
> + WRITE_REG(priv->dev, CP_HPD_EOP_BASE_ADDR, lower_32(pipe_hpd_addr >> 8));
> + WRITE_REG(priv->dev, CP_HPD_EOP_BASE_ADDR_HI, upper_32(pipe_hpd_addr >> 8));
> + WRITE_REG(priv->dev, CP_HPD_EOP_VMID, 0);
> + WRITE_REG(priv->dev, CP_HPD_EOP_CONTROL, CIK_HPD_SIZE_LOG2 - 1);
> + }
> +
> + unlock_srbm_index(priv);
> +}
> +
> +/* Program the VMID -> PASID mapping for one VMID.
> + * PASID 0 is special: it means to associate no PASID with that VMID.
> + * This function waits for the VMID/PASID mapping to complete.
> + */
> +static void set_vmid_pasid_mapping(struct cik_static_private *priv, unsigned int vmid, pasid_t pasid)
> +{
> + /* We have to assume that there is no outstanding mapping.
> + * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because a mapping
> + * is in progress or because a mapping finished and the SW cleared it.
> + * So the protocol is to always wait & clear.
> + */
> +
> + uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid | ATC_VMID_PASID_MAPPING_VALID;
> +
> + WRITE_REG(priv->dev, ATC_VMID0_PASID_MAPPING + vmid*sizeof(uint32_t), pasid_mapping);
> +
> + while (!(READ_REG(priv->dev, ATC_VMID_PASID_MAPPING_UPDATE_STATUS) & (1U << vmid)))
> + cpu_relax();
> + WRITE_REG(priv->dev, ATC_VMID_PASID_MAPPING_UPDATE_STATUS, 1U << vmid);
> +}
> +
> +static uint32_t compute_sh_mem_bases_64bit(unsigned int top_address_nybble)
> +{
> + /* In 64-bit mode, we can only control the top 3 bits of the LDS, scratch and GPUVM apertures.
> + * The hardware fills in the remaining 59 bits according to the following pattern:
> + * LDS: X0000000'00000000 - X0000001'00000000 (4GB)
> + * Scratch: X0000001'00000000 - X0000002'00000000 (4GB)
> + * GPUVM: Y0010000'00000000 - Y0020000'00000000 (1TB)
> + *
> + * (where X/Y is the configurable nybble with the low-bit 0)
> + *
> + * LDS and scratch will have the same top nybble programmed in the top 3 bits of SH_MEM_BASES.PRIVATE_BASE.
> + * GPUVM can have a different top nybble programmed in the top 3 bits of SH_MEM_BASES.SHARED_BASE.
> + * We don't bother to support different top nybbles for LDS/Scratch and GPUVM.
> + */
> +
> + BUG_ON((top_address_nybble & 1) || top_address_nybble > 0xE);
> +
> + return PRIVATE_BASE(top_address_nybble << 12) | SHARED_BASE(top_address_nybble << 12);
> +}
> +
> +/* Initial programming for all ATS registers.
> + * - enable ATS for all compute VMIDs
> + * - clear the VMID/PASID mapping for all compute VMIDS
> + * - program the shader core flat address settings:
> + * -- 64-bit mode
> + * -- unaligned access allowed
> + * -- noncached (this is the only CPU-coherent mode in CIK)
> + * -- APE 1 disabled
> + */
> +static void init_ats(struct cik_static_private *priv)
> +{
> + unsigned int i;
> +
> + /* Enable self-ringing doorbell recognition and direct the BIF to send
> + * untranslated writes to the IOMMU before comparing to the aperture.*/
> + WRITE_REG(priv->dev, BIF_DOORBELL_CNTL, 0);
> +
> + WRITE_REG(priv->dev, ATC_VM_APERTURE0_CNTL, ATS_ACCESS_MODE_ALWAYS);
> + WRITE_REG(priv->dev, ATC_VM_APERTURE0_CNTL2, priv->free_vmid_mask);
> + WRITE_REG(priv->dev, ATC_VM_APERTURE0_LOW_ADDR, 0);
> + WRITE_REG(priv->dev, ATC_VM_APERTURE0_HIGH_ADDR, 0);
> +
> + WRITE_REG(priv->dev, ATC_VM_APERTURE1_CNTL, 0);
> + WRITE_REG(priv->dev, ATC_VM_APERTURE1_CNTL2, 0);
> + WRITE_REG(priv->dev, ATC_VM_APERTURE1_LOW_ADDR, 0);
> + WRITE_REG(priv->dev, ATC_VM_APERTURE1_HIGH_ADDR, 0);
> +
> + lock_srbm_index(priv);
> +
> + for (i = 0; i < CIK_NUM_VMID; i++) {
> + if (priv->free_vmid_mask & (1U << i)) {
> + uint32_t sh_mem_config;
> +
> + set_vmid_pasid_mapping(priv, i, 0);
> +
> + vmid_select(priv, i);
> +
> + sh_mem_config = ALIGNMENT_MODE(SH_MEM_ALIGNMENT_MODE_UNALIGNED);
> + sh_mem_config |= DEFAULT_MTYPE(MTYPE_NONCACHED);
> +
> + WRITE_REG(priv->dev, SH_MEM_CONFIG, sh_mem_config);
> +
> + /* Configure apertures:
> + * LDS: 0x60000000'00000000 - 0x60000001'00000000 (4GB)
> + * Scratch: 0x60000001'00000000 - 0x60000002'00000000 (4GB)
> + * GPUVM: 0x60010000'00000000 - 0x60020000'00000000 (1TB)
> + */
> + WRITE_REG(priv->dev, SH_MEM_BASES, compute_sh_mem_bases_64bit(6));
> +
> + /* Scratch aperture is not supported for now. */
> + WRITE_REG(priv->dev, SH_STATIC_MEM_CONFIG, 0);
> +
> + /* APE1 disabled for now. */
> + WRITE_REG(priv->dev, SH_MEM_APE1_BASE, 1);
> + WRITE_REG(priv->dev, SH_MEM_APE1_LIMIT, 0);
> + }
> + }
> +
> + unlock_srbm_index(priv);
> +}
> +
> +static void exit_ats(struct cik_static_private *priv)
> +{
> + unsigned int i;
> +
> + for (i = 0; i < CIK_NUM_VMID; i++)
> + if (priv->free_vmid_mask & (1U << i))
> + set_vmid_pasid_mapping(priv, i, 0);
> +
> + WRITE_REG(priv->dev, ATC_VM_APERTURE0_CNTL, ATS_ACCESS_MODE_NEVER);
> + WRITE_REG(priv->dev, ATC_VM_APERTURE0_CNTL2, 0);
> +}
> +
> +static struct cik_static_private *kfd_scheduler_to_private(struct kfd_scheduler *scheduler)
> +{
> + return (struct cik_static_private *)scheduler;
> +}
> +
> +static struct cik_static_process *kfd_process_to_private(struct kfd_scheduler_process *process)
> +{
> + return (struct cik_static_process *)process;
> +}
> +
> +static struct cik_static_queue *kfd_queue_to_private(struct kfd_scheduler_queue *queue)
> +{
> + return (struct cik_static_queue *)queue;
> +}
> +
> +static int cik_static_create(struct kfd_dev *dev, struct kfd_scheduler **scheduler)
> +{
> + struct cik_static_private *priv;
> + unsigned int i;
> + int err;
> + void *hpdptr;
> +
> + priv = kmalloc(sizeof(*priv), GFP_KERNEL);
> + if (priv == NULL)
> + return -ENOMEM;
> +
> + mutex_init(&priv->mutex);
> +
> + priv->dev = dev;
> +
> + priv->first_pipe = dev->shared_resources.first_compute_pipe;
> + priv->num_pipes = dev->shared_resources.compute_pipe_count;
> +
> + for (i = 0; i < priv->num_pipes * CIK_QUEUES_PER_PIPE; i++)
> + __set_bit(i, priv->free_queues);
> +
> + priv->free_vmid_mask = dev->shared_resources.compute_vmid_bitmap;
> +
> + /*
> + * Allocate memory for the HPDs. This is hardware-owned per-pipe data.
> + * The driver never accesses this memory after zeroing it. It doesn't even have
> + * to be saved/restored on suspend/resume because it contains no data when there
> + * are no active queues.
> + */
> + err = radeon_kfd_vidmem_alloc(dev,
> + CIK_HPD_SIZE * priv->num_pipes * 2,
> + PAGE_SIZE,
> + KFD_MEMPOOL_SYSTEM_WRITECOMBINE,
> + &priv->hpd_mem);
> + if (err)
> + goto err_hpd_alloc;
> +
> + err = radeon_kfd_vidmem_kmap(dev, priv->hpd_mem, &hpdptr);
> + if (err)
> + goto err_hpd_kmap;
> + memset(hpdptr, 0, CIK_HPD_SIZE * priv->num_pipes);
> + radeon_kfd_vidmem_unkmap(dev, priv->hpd_mem);
> +
> + /*
> + * Allocate memory for all the MQDs.
> + * These are per-queue data that is hardware owned but with driver init.
> + * The driver has to copy this data into HQD registers when a
> + * pipe is (re)activated.
> + */
> + err = radeon_kfd_vidmem_alloc(dev,
> + sizeof(struct cik_mqd_padded) * priv->num_pipes * CIK_QUEUES_PER_PIPE,
> + PAGE_SIZE,
> + KFD_MEMPOOL_SYSTEM_CACHEABLE,
> + &priv->mqd_mem);
> + if (err)
> + goto err_mqd_alloc;
> + radeon_kfd_vidmem_kmap(dev, priv->mqd_mem, (void **)&priv->mqds);
> + if (err)
> + goto err_mqd_kmap;
> +
> + *scheduler = (struct kfd_scheduler *)priv;
> +
> + return 0;
> +
> +err_mqd_kmap:
> + radeon_kfd_vidmem_free(dev, priv->mqd_mem);
> +err_mqd_alloc:
> +err_hpd_kmap:
> + radeon_kfd_vidmem_free(dev, priv->hpd_mem);
> +err_hpd_alloc:
> + mutex_destroy(&priv->mutex);
> + kfree(priv);
> + return err;
> +}
> +
> +static void cik_static_destroy(struct kfd_scheduler *scheduler)
> +{
> + struct cik_static_private *priv = kfd_scheduler_to_private(scheduler);
> +
> + radeon_kfd_vidmem_unkmap(priv->dev, priv->mqd_mem);
> + radeon_kfd_vidmem_free(priv->dev, priv->mqd_mem);
> + radeon_kfd_vidmem_free(priv->dev, priv->hpd_mem);
> +
> + mutex_destroy(&priv->mutex);
> +
> + kfree(priv);
> +}
> +
> +static void cik_static_start(struct kfd_scheduler *scheduler)
> +{
> + struct cik_static_private *priv = kfd_scheduler_to_private(scheduler);
> +
> + radeon_kfd_vidmem_gpumap(priv->dev, priv->hpd_mem, &priv->hpd_addr);
> + radeon_kfd_vidmem_gpumap(priv->dev, priv->mqd_mem, &priv->mqd_addr);
> +
> + init_pipes(priv);
> + init_ats(priv);
> +}
> +
> +static void cik_static_stop(struct kfd_scheduler *scheduler)
> +{
> + struct cik_static_private *priv = kfd_scheduler_to_private(scheduler);
> +
> + exit_ats(priv);
> +
> + radeon_kfd_vidmem_ungpumap(priv->dev, priv->hpd_mem);
> + radeon_kfd_vidmem_ungpumap(priv->dev, priv->mqd_mem);
> +}
> +
> +static bool allocate_vmid(struct cik_static_private *priv, unsigned int *vmid)
> +{
> + bool ok = false;
> +
> + mutex_lock(&priv->mutex);
> +
> + if (priv->free_vmid_mask != 0) {
> + unsigned int v = __ffs64(priv->free_vmid_mask);
> +
> + clear_bit(v, &priv->free_vmid_mask);
> + *vmid = v;
> +
> + ok = true;
> + }
> +
> + mutex_unlock(&priv->mutex);
> +
> + return ok;
> +}
> +
> +static void release_vmid(struct cik_static_private *priv, unsigned int vmid)
> +{
> + /* It's okay to race against allocate_vmid because this only adds bits to free_vmid_mask.
> + * And set_bit/clear_bit are atomic wrt each other. */
> + set_bit(vmid, &priv->free_vmid_mask);
> +}
> +
> +static void setup_vmid_for_process(struct cik_static_private *priv, struct cik_static_process *p)
> +{
> + set_vmid_pasid_mapping(priv, p->vmid, p->pasid);
> +
> + /*
> + * SH_MEM_CONFIG and others need to be programmed differently
> + * for 32/64-bit processes. And maybe other reasons.
> + */
> +}
> +
> +static int
> +cik_static_register_process(struct kfd_scheduler *scheduler, struct kfd_process *process,
> + struct kfd_scheduler_process **scheduler_process)
> +{
> + struct cik_static_private *priv = kfd_scheduler_to_private(scheduler);
> +
> + struct cik_static_process *hwp;
> +
> + hwp = kmalloc(sizeof(*hwp), GFP_KERNEL);
> + if (hwp == NULL)
> + return -ENOMEM;
> +
> + if (!allocate_vmid(priv, &hwp->vmid)) {
> + kfree(hwp);
> + return -ENOMEM;
> + }
> +
> + hwp->pasid = process->pasid;
> +
> + setup_vmid_for_process(priv, hwp);
> +
> + *scheduler_process = (struct kfd_scheduler_process *)hwp;
> +
> + return 0;
> +}
> +
> +static void cik_static_deregister_process(struct kfd_scheduler *scheduler,
> + struct kfd_scheduler_process *scheduler_process)
> +{
> + struct cik_static_private *priv = kfd_scheduler_to_private(scheduler);
> + struct cik_static_process *pp = kfd_process_to_private(scheduler_process);
> +
> + release_vmid(priv, pp->vmid);
> + kfree(pp);
> +}
> +
> +static bool allocate_hqd(struct cik_static_private *priv, unsigned int *queue)
> +{
> + bool ok = false;
> + unsigned int q;
> +
> + mutex_lock(&priv->mutex);
> +
> + q = find_first_bit(priv->free_queues, priv->num_pipes * CIK_QUEUES_PER_PIPE);
> +
> + if (q != priv->num_pipes * CIK_QUEUES_PER_PIPE) {
> + clear_bit(q, priv->free_queues);
> + *queue = q;
> +
> + ok = true;
> + }
> +
> + mutex_unlock(&priv->mutex);
> +
> + return ok;
> +}
> +
> +static void release_hqd(struct cik_static_private *priv, unsigned int queue)
> +{
> + /* It's okay to race against allocate_hqd because this only adds bits to free_queues.
> + * And set_bit/clear_bit are atomic wrt each other. */
> + set_bit(queue, priv->free_queues);
> +}
> +
> +static void init_mqd(const struct cik_static_queue *queue, const struct cik_static_process *process)
> +{
> + struct cik_mqd *mqd = queue->mqd;
> +
> + memset(mqd, 0, sizeof(*mqd));
> +
> + mqd->header = 0xC0310800;
> + mqd->pipeline_stat_enable = 1;
> + mqd->static_thread_mgmt01[0] = 0xffffffff;
> + mqd->static_thread_mgmt01[1] = 0xffffffff;
> + mqd->static_thread_mgmt23[0] = 0xffffffff;
> + mqd->static_thread_mgmt23[1] = 0xffffffff;
> +
> + mqd->queue_state.cp_mqd_base_addr = lower_32(queue->mqd_addr);
> + mqd->queue_state.cp_mqd_base_addr_hi = upper_32(queue->mqd_addr);
> + mqd->queue_state.cp_mqd_control = MQD_CONTROL_PRIV_STATE_EN;
> +
> + mqd->queue_state.cp_hqd_pq_base = lower_32((uintptr_t)queue->pq_addr >> 8);
> + mqd->queue_state.cp_hqd_pq_base_hi = upper_32((uintptr_t)queue->pq_addr >> 8);
> + mqd->queue_state.cp_hqd_pq_control = QUEUE_SIZE(queue->queue_size_encoded) | DEFAULT_RPTR_BLOCK_SIZE
> + | DEFAULT_MIN_AVAIL_SIZE | PQ_ATC_EN;
> + mqd->queue_state.cp_hqd_pq_rptr_report_addr = lower_32((uintptr_t)queue->rptr_address);
> + mqd->queue_state.cp_hqd_pq_rptr_report_addr_hi = upper_32((uintptr_t)queue->rptr_address);
> + mqd->queue_state.cp_hqd_pq_doorbell_control = DOORBELL_OFFSET(queue->doorbell_index) | DOORBELL_EN;
> + mqd->queue_state.cp_hqd_vmid = process->vmid;
> + mqd->queue_state.cp_hqd_active = 1;
> +
> + mqd->queue_state.cp_hqd_persistent_state = DEFAULT_CP_HQD_PERSISTENT_STATE;
> +
> + /* The values for these 3 are from WinKFD. */
> + mqd->queue_state.cp_hqd_quantum = QUANTUM_EN | QUANTUM_SCALE_1MS | QUANTUM_DURATION(10);
> + mqd->queue_state.cp_hqd_pipe_priority = 1;
> + mqd->queue_state.cp_hqd_queue_priority = 15;
> +
> + mqd->queue_state.cp_hqd_ib_control = IB_ATC_EN | DEFAULT_MIN_IB_AVAIL_SIZE;
> +}
> +
> +/* Write the HQD registers and activate the queue.
> + * Requires that SRBM_GFX_CNTL has already been programmed for the queue.
> + */
> +static void load_hqd(struct cik_static_private *priv, struct cik_static_queue *queue)
> +{
> + struct kfd_dev *dev = priv->dev;
> + const struct cik_hqd_registers *qs = &queue->mqd->queue_state;
> +
> + WRITE_REG(dev, CP_MQD_BASE_ADDR, qs->cp_mqd_base_addr);
> + WRITE_REG(dev, CP_MQD_BASE_ADDR_HI, qs->cp_mqd_base_addr_hi);
> + WRITE_REG(dev, CP_MQD_CONTROL, qs->cp_mqd_control);
> +
> + WRITE_REG(dev, CP_HQD_PQ_BASE, qs->cp_hqd_pq_base);
> + WRITE_REG(dev, CP_HQD_PQ_BASE_HI, qs->cp_hqd_pq_base_hi);
> + WRITE_REG(dev, CP_HQD_PQ_CONTROL, qs->cp_hqd_pq_control);
> + /* DOORBELL_CONTROL before WPTR because WPTR writes are dropped if DOORBELL_HIT is set. */
> + WRITE_REG(dev, CP_HQD_PQ_DOORBELL_CONTROL, qs->cp_hqd_pq_doorbell_control);
> + WRITE_REG(dev, CP_HQD_PQ_WPTR, qs->cp_hqd_pq_wptr);
> + WRITE_REG(dev, CP_HQD_PQ_RPTR, qs->cp_hqd_pq_rptr);
> + WRITE_REG(dev, CP_HQD_PQ_RPTR_REPORT_ADDR, qs->cp_hqd_pq_rptr_report_addr);
> + WRITE_REG(dev, CP_HQD_PQ_RPTR_REPORT_ADDR_HI, qs->cp_hqd_pq_rptr_report_addr_hi);
> +
> + WRITE_REG(dev, CP_HQD_VMID, qs->cp_hqd_vmid);
> + WRITE_REG(dev, CP_HQD_PERSISTENT_STATE, qs->cp_hqd_persistent_state);
> + WRITE_REG(dev, CP_HQD_QUANTUM, qs->cp_hqd_quantum);
> + WRITE_REG(dev, CP_HQD_PIPE_PRIORITY, qs->cp_hqd_pipe_priority);
> + WRITE_REG(dev, CP_HQD_QUEUE_PRIORITY, qs->cp_hqd_queue_priority);
> +
> + WRITE_REG(dev, CP_HQD_IB_CONTROL, qs->cp_hqd_ib_control);
> + WRITE_REG(dev, CP_HQD_IB_BASE_ADDR, qs->cp_hqd_ib_base_addr);
> + WRITE_REG(dev, CP_HQD_IB_BASE_ADDR_HI, qs->cp_hqd_ib_base_addr_hi);
> + WRITE_REG(dev, CP_HQD_IB_RPTR, qs->cp_hqd_ib_rptr);
> + WRITE_REG(dev, CP_HQD_SEMA_CMD, qs->cp_hqd_sema_cmd);
> + WRITE_REG(dev, CP_HQD_MSG_TYPE, qs->cp_hqd_msg_type);
> + WRITE_REG(dev, CP_HQD_ATOMIC0_PREOP_LO, qs->cp_hqd_atomic0_preop_lo);
> + WRITE_REG(dev, CP_HQD_ATOMIC0_PREOP_HI, qs->cp_hqd_atomic0_preop_hi);
> + WRITE_REG(dev, CP_HQD_ATOMIC1_PREOP_LO, qs->cp_hqd_atomic1_preop_lo);
> + WRITE_REG(dev, CP_HQD_ATOMIC1_PREOP_HI, qs->cp_hqd_atomic1_preop_hi);
> + WRITE_REG(dev, CP_HQD_HQ_SCHEDULER0, qs->cp_hqd_hq_scheduler0);
> + WRITE_REG(dev, CP_HQD_HQ_SCHEDULER1, qs->cp_hqd_hq_scheduler1);
> +
> + WRITE_REG(dev, CP_HQD_ACTIVE, 1);
> +}
> +
> +static void activate_queue(struct cik_static_private *priv, struct cik_static_queue *queue)
> +{
> + bool wptr_shadow_valid;
> + doorbell_t wptr_shadow;
> +
> + /* Avoid sleeping while holding the SRBM lock. */
> + wptr_shadow_valid = !get_user(wptr_shadow, queue->wptr_address);
> +
> + lock_srbm_index(priv);
> + queue_select(priv, queue->queue);
> +
> + load_hqd(priv, queue);
> +
> + /* Doorbell and wptr are special because there is a race when reactivating a queue.
> + * Since doorbell writes to deactivated queues are ignored by hardware, the application
> + * shadows the doorbell into memory at queue->wptr_address.
> + *
> + * We want the queue to automatically resume processing as if it were always active,
> + * so we want to copy from queue->wptr_address into the wptr/doorbell.
> + *
> + * The race is that the app could write a new wptr into the doorbell before we
> + * write the shadowed wptr, resulting in an old wptr written later.
> + *
> + * The hardware solves this ignoring CP_HQD_WPTR writes after a doorbell write.
> + * So the KFD can activate the doorbell then write the shadow wptr to CP_HQD_WPTR
> + * knowing it will be ignored if the user has written a more-recent doorbell.
> + */
> + if (wptr_shadow_valid)
> + WRITE_REG(priv->dev, CP_HQD_PQ_WPTR, wptr_shadow);
> +
> + unlock_srbm_index(priv);
> +}
> +
> +static void drain_hqd(struct cik_static_private *priv)
> +{
> + WRITE_REG(priv->dev, CP_HQD_DEQUEUE_REQUEST, DEQUEUE_REQUEST_DRAIN);
> +}
> +
> +static void wait_hqd_inactive(struct cik_static_private *priv)
> +{
> + while (READ_REG(priv->dev, CP_HQD_ACTIVE) != 0)
> + cpu_relax();
> +}
> +
> +static void deactivate_queue(struct cik_static_private *priv, struct cik_static_queue *queue)
> +{
> + lock_srbm_index(priv);
> + queue_select(priv, queue->queue);
> +
> + drain_hqd(priv);
> + wait_hqd_inactive(priv);
> +
> + unlock_srbm_index(priv);
> +}
> +
> +#define BIT_MASK_64(high, low) (((1ULL << (high)) - 1) & ~((1ULL << (low)) - 1))
> +#define RING_ADDRESS_BAD_BIT_MASK (~BIT_MASK_64(48, 8))
> +#define RWPTR_ADDRESS_BAD_BIT_MASK (~BIT_MASK_64(48, 2))
> +
> +#define MAX_QUEUE_SIZE (1ULL << 32)
> +#define MIN_QUEUE_SIZE (1ULL << 10)
> +
> +static int
> +cik_static_create_queue(struct kfd_scheduler *scheduler,
> + struct kfd_scheduler_process *process,
> + struct kfd_scheduler_queue *queue,
> + void __user *ring_address,
> + uint64_t ring_size,
> + void __user *rptr_address,
> + void __user *wptr_address,
> + unsigned int doorbell)
> +{
> + struct cik_static_private *priv = kfd_scheduler_to_private(scheduler);
> + struct cik_static_process *hwp = kfd_process_to_private(process);
> + struct cik_static_queue *hwq = kfd_queue_to_private(queue);
> +
> + if ((uint64_t)ring_address & RING_ADDRESS_BAD_BIT_MASK
> + || (uint64_t)rptr_address & RWPTR_ADDRESS_BAD_BIT_MASK
> + || (uint64_t)wptr_address & RWPTR_ADDRESS_BAD_BIT_MASK)
> + return -EINVAL;
> +
> + if (ring_size > MAX_QUEUE_SIZE || ring_size < MIN_QUEUE_SIZE || !is_power_of_2(ring_size))
> + return -EINVAL;
> +
> + if (!allocate_hqd(priv, &hwq->queue))
> + return -ENOMEM;
> +
> + hwq->mqd_addr = priv->mqd_addr + sizeof(struct cik_mqd_padded) * hwq->queue;
> + hwq->mqd = &priv->mqds[hwq->queue].mqd;
> + hwq->pq_addr = ring_address;
> + hwq->rptr_address = rptr_address;
> + hwq->wptr_address = wptr_address;
> + hwq->doorbell_index = doorbell;
> + hwq->queue_size_encoded = ilog2(ring_size) - 3;
> +
> + init_mqd(hwq, hwp);
> + activate_queue(priv, hwq);
> +
> + return 0;
> +}
> +
> +static void
> +cik_static_destroy_queue(struct kfd_scheduler *scheduler, struct kfd_scheduler_queue *queue)
> +{
> + struct cik_static_private *priv = kfd_scheduler_to_private(scheduler);
> + struct cik_static_queue *hwq = kfd_queue_to_private(queue);
> +
> + deactivate_queue(priv, hwq);
> +
> + release_hqd(priv, hwq->queue);
> +}
> +
> +const struct kfd_scheduler_class radeon_kfd_cik_static_scheduler_class = {
> + .name = "CIK static scheduler",
> + .create = cik_static_create,
> + .destroy = cik_static_destroy,
> + .start = cik_static_start,
> + .stop = cik_static_stop,
> + .register_process = cik_static_register_process,
> + .deregister_process = cik_static_deregister_process,
> + .queue_size = sizeof(struct cik_static_queue),
> + .create_queue = cik_static_create_queue,
> + .destroy_queue = cik_static_destroy_queue,
> +};
> diff --git a/drivers/gpu/hsa/radeon/kfd_vidmem.c b/drivers/gpu/hsa/radeon/kfd_vidmem.c
> new file mode 100644
> index 0000000..c8d3770
> --- /dev/null
> +++ b/drivers/gpu/hsa/radeon/kfd_vidmem.c
> @@ -0,0 +1,61 @@
> +/*
> + * Copyright 2014 Advanced Micro Devices, Inc.
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice shall be included in
> + * all copies or substantial portions of the Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
> + * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> + * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> + * OTHER DEALINGS IN THE SOFTWARE.
> + */
> +
> +#include "kfd_priv.h"
> +
> +int radeon_kfd_vidmem_alloc(struct kfd_dev *kfd, size_t size, size_t alignment,
> + enum kfd_mempool pool, kfd_mem_obj *mem_obj)
> +{
> + return kfd2kgd->allocate_mem(kfd->kgd,
> + size,
> + alignment,
> + (enum kgd_memory_pool)pool,
> + (struct kgd_mem **)mem_obj);
> +}
> +
> +void radeon_kfd_vidmem_free(struct kfd_dev *kfd, kfd_mem_obj mem_obj)
> +{
> + kfd2kgd->free_mem(kfd->kgd, (struct kgd_mem *)mem_obj);
> +}
> +
> +int radeon_kfd_vidmem_gpumap(struct kfd_dev *kfd, kfd_mem_obj mem_obj,
> + uint64_t *vmid0_address)
> +{
> + return kfd2kgd->gpumap_mem(kfd->kgd,
> + (struct kgd_mem *)mem_obj,
> + vmid0_address);
> +}
> +
> +void radeon_kfd_vidmem_ungpumap(struct kfd_dev *kfd, kfd_mem_obj mem_obj)
> +{
> + kfd2kgd->ungpumap_mem(kfd->kgd, (struct kgd_mem *)mem_obj);
> +}
> +
> +int radeon_kfd_vidmem_kmap(struct kfd_dev *kfd, kfd_mem_obj mem_obj, void **ptr)
> +{
> + return kfd2kgd->kmap_mem(kfd->kgd, (struct kgd_mem *)mem_obj, ptr);
> +}
> +
> +void radeon_kfd_vidmem_unkmap(struct kfd_dev *kfd, kfd_mem_obj mem_obj)
> +{
> + kfd2kgd->unkmap_mem(kfd->kgd, (struct kgd_mem *)mem_obj);
> +}
> --
> 1.9.1
>
More information about the dri-devel
mailing list