Re: [PATCH v2 06/25] drm/radeon: Add radeon <--> amdkfd interface

On 20/07/14 20:35, Jerome Glisse wrote:

On Thu, Jul 17, 2014 at 04:29:13PM +0300, Oded Gabbay wrote:

...

This patch adds the interface between the radeon driver and the amdkfd driver. The interface implementation is contained in radeon_kfd.c and radeon_kfd.h.

The interface itself is represented by a pointer to struct kfd_dev. The pointer is located inside radeon_device structure.

All the register accesses that amdkfd need are done using this interface. This allows us to avoid direct register accesses in amdkfd proper, while also avoiding locking between amdkfd and radeon.

The single exception is the doorbells that are used in both of the drivers. However, because they are located in separate pci bar pages, the danger of sharing registers between the drivers is minimal.

Having said that, we are planning to move the doorbells as well to radeon.

The loading of the amdkfd module is done via symbol lookup. According to the code review discussions, this may change in v3 of the patch set.

Signed-off-by: Oded Gabbay oded.gabbay@amd.com

drivers/gpu/drm/radeon/Makefile | 1 + drivers/gpu/drm/radeon/cik.c | 9 + drivers/gpu/drm/radeon/cik_reg.h | 65 +++++ drivers/gpu/drm/radeon/cikd.h | 51 +++- drivers/gpu/drm/radeon/radeon.h | 3 + drivers/gpu/drm/radeon/radeon_drv.c | 5 + drivers/gpu/drm/radeon/radeon_kfd.c | 566 ++++++++++++++++++++++++++++++++++++ drivers/gpu/drm/radeon/radeon_kfd.h | 119 ++++++++ drivers/gpu/drm/radeon/radeon_kms.c | 7 + 9 files changed, 825 insertions(+), 1 deletion(-) create mode 100644 drivers/gpu/drm/radeon/radeon_kfd.c create mode 100644 drivers/gpu/drm/radeon/radeon_kfd.h

diff --git a/drivers/gpu/drm/radeon/Makefile b/drivers/gpu/drm/radeon/Makefile index 1b04002..a1c913d 100644 --- a/drivers/gpu/drm/radeon/Makefile +++ b/drivers/gpu/drm/radeon/Makefile @@ -104,6 +104,7 @@ radeon-y += \ radeon_vce.o \ vce_v1_0.o \ vce_v2_0.o \

• radeon_kfd.o

radeon-$(CONFIG_COMPAT) += radeon_ioc32.o radeon-$(CONFIG_VGA_SWITCHEROO) += radeon_atpx_handler.o diff --git a/drivers/gpu/drm/radeon/cik.c b/drivers/gpu/drm/radeon/cik.c index b4bbc22..6f71095 100644 --- a/drivers/gpu/drm/radeon/cik.c +++ b/drivers/gpu/drm/radeon/cik.c @@ -32,6 +32,7 @@ #include "cik_blit_shaders.h" #include "radeon_ucode.h" #include "clearstate_ci.h" +#include "radeon_kfd.h"

MODULE_FIRMWARE("radeon/BONAIRE_pfp.bin"); MODULE_FIRMWARE("radeon/BONAIRE_me.bin"); @@ -7727,6 +7728,9 @@ restart_ih: while (rptr != wptr) { /* wptr/rptr are in bytes! */ ring_index = rptr / 4;

• radeon_kfd_interrupt(rdev, (const void *) &rdev->ih.ring[ring_index]);
  

• src_id = le32_to_cpu(rdev->ih.ring[ring_index]) & 0xff; src_data = le32_to_cpu(rdev->ih.ring[ring_index + 1]) & 0xfffffff; ring_id = le32_to_cpu(rdev->ih.ring[ring_index + 2]) & 0xff;

@@ -8386,6 +8390,10 @@ static int cik_startup(struct radeon_device *rdev) if (r) return r;

• r = radeon_kfd_resume(rdev);
• if (r)

• return r;
  

• return 0;

}

@@ -8434,6 +8442,7 @@ int cik_resume(struct radeon_device *rdev) */ int cik_suspend(struct radeon_device *rdev) {

• radeon_kfd_suspend(rdev); radeon_pm_suspend(rdev); dce6_audio_fini(rdev); radeon_vm_manager_fini(rdev);

diff --git a/drivers/gpu/drm/radeon/cik_reg.h b/drivers/gpu/drm/radeon/cik_reg.h index ca1bb61..1ab3dbc 100644 --- a/drivers/gpu/drm/radeon/cik_reg.h +++ b/drivers/gpu/drm/radeon/cik_reg.h @@ -147,4 +147,69 @@

#define CIK_LB_DESKTOP_HEIGHT 0x6b0c

+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];

+};

#endif diff --git a/drivers/gpu/drm/radeon/cikd.h b/drivers/gpu/drm/radeon/cikd.h index 0c6e1b5..0a2a403 100644 --- a/drivers/gpu/drm/radeon/cikd.h +++ b/drivers/gpu/drm/radeon/cikd.h @@ -1137,6 +1137,9 @@ #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_CACHED 0 +#define MTYPE_NONCACHED 3

#define SX_DEBUG_1 0x9060

@@ -1447,6 +1450,16 @@ #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)

We need documentation for this queue/pipe priority to know their granularity and how they are use exactly.

Done in v3

...

#define CP_HQD_PQ_BASE 0xC934 #define CP_HQD_PQ_BASE_HI 0xC938 #define CP_HQD_PQ_RPTR 0xC93C @@ -1474,12 +1487,32 @@ #define PRIV_STATE (1 << 30) #define KMD_QUEUE (1 << 31)

-#define CP_HQD_DEQUEUE_REQUEST 0xC974 +#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 DEQUEUE_REQUEST_RESET 2

#define CP_MQD_CONTROL 0xC99C #define MQD_VMID(x) ((x) << 0) #define MQD_VMID_MASK (0xf << 0)

+#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 SH_STATIC_MEM_CONFIG 0x9604u

Same here documentation is needed on all those register.

This is a bit more problematic. I need to find out what I can reveal. I prefer to add this later (v4 or a single patch)

...

#define DB_RENDER_CONTROL 0x28000

#define PA_SC_RASTER_CONFIG 0x28350 @@ -2069,4 +2102,20 @@ #define VCE_CMD_IB_AUTO 0x00000005 #define VCE_CMD_SEMAPHORE 0x00000006

+#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/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h index 5136855..94b38a7 100644 --- a/drivers/gpu/drm/radeon/radeon.h +++ b/drivers/gpu/drm/radeon/radeon.h @@ -2342,6 +2342,9 @@ struct radeon_device {

struct dev_pm_domain vga_pm_domain; bool have_disp_power_ref;

• /* HSA KFD interface */
• struct kfd_dev *kfd;

};

bool radeon_is_px(struct drm_device *dev); diff --git a/drivers/gpu/drm/radeon/radeon_drv.c b/drivers/gpu/drm/radeon/radeon_drv.c index cb14213..efaa086 100644 --- a/drivers/gpu/drm/radeon/radeon_drv.c +++ b/drivers/gpu/drm/radeon/radeon_drv.c @@ -39,6 +39,8 @@ #include <linux/pm_runtime.h> #include <linux/vga_switcheroo.h> #include "drm_crtc_helper.h" +#include "radeon_kfd.h"

/*

• KMS wrapper.
• • 2.0.0 - initial interface

@@ -630,12 +632,15 @@ static int __init radeon_init(void) #endif }

• radeon_kfd_init();
• /* let modprobe override vga console setting */ return drm_pci_init(driver, pdriver);

}

static void __exit radeon_exit(void) {

• radeon_kfd_fini(); drm_pci_exit(driver, pdriver); radeon_unregister_atpx_handler();

} diff --git a/drivers/gpu/drm/radeon/radeon_kfd.c b/drivers/gpu/drm/radeon/radeon_kfd.c new file mode 100644 index 0000000..0385239 --- /dev/null +++ b/drivers/gpu/drm/radeon/radeon_kfd.c @@ -0,0 +1,566 @@ +/*

• • 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/module.h> +#include <linux/fdtable.h> +#include <linux/uaccess.h> +#include <drm/drmP.h> +#include "radeon.h" +#include "cikd.h" +#include "cik_reg.h" +#include "radeon_kfd.h"

+#define CIK_PIPE_PER_MEC (4)

+struct kgd_mem {

• struct radeon_bo *bo;
• u32 domain;
• struct radeon_bo_va *bo_va;

+};

+static int allocate_mem(struct kgd_dev *kgd, size_t size, size_t alignment,

• enum kgd_memory_pool pool, struct kgd_mem **memory_handle);
  

+static void free_mem(struct kgd_dev *kgd, struct kgd_mem *memory_handle);

+static int gpumap_mem(struct kgd_dev *kgd, struct kgd_mem *mem, uint64_t *vmid0_address); +static void ungpumap_mem(struct kgd_dev *kgd, struct kgd_mem *mem);

+static int kmap_mem(struct kgd_dev *kgd, struct kgd_mem *mem, void **ptr); +static void unkmap_mem(struct kgd_dev *kgd, struct kgd_mem *mem);

+static uint64_t get_vmem_size(struct kgd_dev *kgd); +static uint64_t get_gpu_clock_counter(struct kgd_dev *kgd);

+static uint32_t get_max_engine_clock_in_mhz(struct kgd_dev *kgd);

+/*

• • 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_memory(struct kgd_dev *kgd); +static int kgd_init_pipeline(struct kgd_dev *kgd, uint32_t pipe_id, uint32_t hpd_size, uint64_t hpd_gpu_addr); +static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id, uint32_t queue_id, uint32_t __user *wptr); +static bool kgd_hqd_is_occupies(struct kgd_dev *kgd, uint64_t queue_address, uint32_t pipe_id, uint32_t queue_id); +static int kgd_hqd_destroy(struct kgd_dev *kgd, bool is_reset, unsigned int timeout,

• 		uint32_t pipe_id, uint32_t queue_id);
  

+static const struct kfd2kgd_calls kfd2kgd = {

• .allocate_mem = allocate_mem,
• .free_mem = free_mem,
• .gpumap_mem = gpumap_mem,
• .ungpumap_mem = ungpumap_mem,
• .kmap_mem = kmap_mem,
• .unkmap_mem = unkmap_mem,
• .get_vmem_size = get_vmem_size,
• .get_gpu_clock_counter = get_gpu_clock_counter,
• .get_max_engine_clock_in_mhz = get_max_engine_clock_in_mhz,
• .program_sh_mem_settings = kgd_program_sh_mem_settings,
• .set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
• .init_memory = kgd_init_memory,
• .init_pipeline = kgd_init_pipeline,
• .hqd_load = kgd_hqd_load,
• .hqd_is_occupies = kgd_hqd_is_occupies,
• .hqd_destroy = kgd_hqd_destroy,

+};

+static const struct kgd2kfd_calls *kgd2kfd;

+bool radeon_kfd_init(void) +{

• bool (*kgd2kfd_init_p)(unsigned, const struct kfd2kgd_calls*,

• 		const struct kgd2kfd_calls**);
  

• kgd2kfd_init_p = symbol_request(kgd2kfd_init);
• if (kgd2kfd_init_p == NULL)

• return false;
  

• if (!kgd2kfd_init_p(KFD_INTERFACE_VERSION, &kfd2kgd, &kgd2kfd)) {

• symbol_put(kgd2kfd_init);
  

• kgd2kfd = NULL;
  

• return false;
  

• }
• return true;

+}

+void radeon_kfd_fini(void) +{

• if (kgd2kfd) {

• kgd2kfd->exit();
  

• symbol_put(kgd2kfd_init);
  

• }

+}

+void radeon_kfd_device_probe(struct radeon_device *rdev) +{

• if (kgd2kfd)

• rdev->kfd = kgd2kfd->probe((struct kgd_dev *)rdev, rdev->pdev);
  

+}

+void radeon_kfd_device_init(struct radeon_device *rdev) +{

• if (rdev->kfd) {

• struct kgd2kfd_shared_resources gpu_resources = {
  

• 	.compute_vmid_bitmap = 0xFF00,
  

• 	.first_compute_pipe = 1,
  

• 	.compute_pipe_count = 8 - 1,
  

• };
  

• radeon_doorbell_get_kfd_info(rdev,
  

• 		&gpu_resources.doorbell_physical_address,
  

• 		&gpu_resources.doorbell_aperture_size,
  

• 		&gpu_resources.doorbell_start_offset);
  

• kgd2kfd->device_init(rdev->kfd, &gpu_resources);
  

• }

+}

+void radeon_kfd_device_fini(struct radeon_device *rdev) +{

• if (rdev->kfd) {

• kgd2kfd->device_exit(rdev->kfd);
  

• rdev->kfd = NULL;
  

• }

+}

+void radeon_kfd_interrupt(struct radeon_device *rdev, const void *ih_ring_entry) +{

• if (rdev->kfd)

• kgd2kfd->interrupt(rdev->kfd, ih_ring_entry);
  

+}

+void radeon_kfd_suspend(struct radeon_device *rdev) +{

• if (rdev->kfd)

• kgd2kfd->suspend(rdev->kfd);
  

+}

+int radeon_kfd_resume(struct radeon_device *rdev) +{

• int r = 0;
• if (rdev->kfd)

• r = kgd2kfd->resume(rdev->kfd);
  

• return r;

+}

All of the above wrapper function should be move to header file and mark as inline this would allow for compiler optimization. I still would like to see the possibility to build radeon without hsa.

That is problematic as they don't compile in the header file. Anyway, these functions are rarely called so a compiler optimization is quite useless here. Radeon can definitely build without amdkfd. This file will always be built as it will be part of radeon.

...

+static u32 pool_to_domain(enum kgd_memory_pool p) +{

• switch (p) {
• case KGD_POOL_FRAMEBUFFER: return RADEON_GEM_DOMAIN_VRAM;
• default: return RADEON_GEM_DOMAIN_GTT;
• }

+}

+static int allocate_mem(struct kgd_dev *kgd, size_t size, size_t alignment,

• enum kgd_memory_pool pool, struct kgd_mem **memory_handle)
  

+{

• struct radeon_device *rdev = (struct radeon_device *)kgd;
• struct kgd_mem *mem;
• int r;
• mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
• if (!mem)

• return -ENOMEM;
  

• mem->domain = pool_to_domain(pool);
• r = radeon_bo_create(rdev, size, alignment, true, mem->domain, NULL, &mem->bo);
• if (r) {

• kfree(mem);
  

• return r;
  

• }
• *memory_handle = mem;
• return 0;

+}

+static void free_mem(struct kgd_dev *kgd, struct kgd_mem *mem) +{

• /* Assume that KFD will never free gpumapped or kmapped memory. This is not quite settled. */
• radeon_bo_unref(&mem->bo);
• kfree(mem);

+}

+static int gpumap_mem(struct kgd_dev *kgd, struct kgd_mem *mem, uint64_t *vmid0_address) +{

• int r;
• r = radeon_bo_reserve(mem->bo, true);
• /*

• * ttm_bo_reserve can only fail if the buffer reservation lock
  

• * is held in circumstances that would deadlock
  

• */
  

• BUG_ON(r != 0);
• r = radeon_bo_pin(mem->bo, mem->domain, vmid0_address);
• radeon_bo_unreserve(mem->bo);
• return r;

+}

NACK NACK NACK, no radeon_bo_pin this is not acceptable. Buffer pining should be done very seldomly and i would say only radeon module can do it and only for buffer object under its control. We certainly can not accept to do that for buffer object that are under userspace management.

Changed to new method in v3, as discussed in main thread.

...

+static void ungpumap_mem(struct kgd_dev *kgd, struct kgd_mem *mem) +{

• int r;
• r = radeon_bo_reserve(mem->bo, true);
• /*

• * ttm_bo_reserve can only fail if the buffer reservation lock
  

• * is held in circumstances that would deadlock
  

• */
  

• BUG_ON(r != 0);
• r = radeon_bo_unpin(mem->bo);
• /*

• * This unpin only removed NO_EVICT placement flags
  

• * and should never fail
  

• */
  

• BUG_ON(r != 0);
• radeon_bo_unreserve(mem->bo);

+}

+static int kmap_mem(struct kgd_dev *kgd, struct kgd_mem *mem, void **ptr) +{

• int r;
• r = radeon_bo_reserve(mem->bo, true);
• /*

• * ttm_bo_reserve can only fail if the buffer reservation lock
  

• * is held in circumstances that would deadlock
  

• */
  

• BUG_ON(r != 0);
• r = radeon_bo_kmap(mem->bo, ptr);
• radeon_bo_unreserve(mem->bo);
• return r;

+}

+static void unkmap_mem(struct kgd_dev *kgd, struct kgd_mem *mem) +{

• int r;
• r = radeon_bo_reserve(mem->bo, true);
• /*

• * ttm_bo_reserve can only fail if the buffer reservation lock
  

• * is held in circumstances that would deadlock
  

• */
  

• BUG_ON(r != 0);
• radeon_bo_kunmap(mem->bo);
• radeon_bo_unreserve(mem->bo);

+}

+static uint64_t get_vmem_size(struct kgd_dev *kgd) +{

• struct radeon_device *rdev = (struct radeon_device *)kgd;
• BUG_ON(kgd == NULL);
• return rdev->mc.real_vram_size;

+}

+static uint64_t get_gpu_clock_counter(struct kgd_dev *kgd) +{

• struct radeon_device *rdev = (struct radeon_device *)kgd;
• return rdev->asic->get_gpu_clock_counter(rdev);

+}

+static uint32_t get_max_engine_clock_in_mhz(struct kgd_dev *kgd) +{

• struct radeon_device *rdev = (struct radeon_device *)kgd;
• /* The sclk is in quantas of 10kHz */
• return rdev->pm.dpm.dyn_state.max_clock_voltage_on_ac.sclk / 100;

+}

+/*

• • kfd/radeon registers access interface
• */

+inline uint32_t lower_32(uint64_t x) +{

• return (uint32_t)x;

+}

+inline uint32_t upper_32(uint64_t x) +{

• return (uint32_t)(x >> 32);

+}

Use appropriate macro (upper_32_bits, lower_32_bits) instead of those inline function.

Done in v3.

...

+static inline struct radeon_device *get_radeon_device(struct kgd_dev *kgd) +{

• return (struct radeon_device *)kgd;

+}

+static void write_register(struct kgd_dev *kgd, uint32_t offset, uint32_t value) +{

• struct radeon_device *rdev = get_radeon_device(kgd);
• writel(value, (void __iomem *)(rdev->rmmio + offset));

+}

+static uint32_t read_register(struct kgd_dev *kgd, uint32_t offset) +{

• struct radeon_device *rdev = get_radeon_device(kgd);
• return readl((void __iomem *)(rdev->rmmio + offset));

+}

+static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe, uint32_t queue, uint32_t vmid) +{

• struct radeon_device *rdev = get_radeon_device(kgd);
• uint32_t value = PIPEID(pipe) | MEID(mec) | VMID(vmid) | QUEUEID(queue);
• mutex_lock(&rdev->srbm_mutex);
• write_register(kgd, SRBM_GFX_CNTL, value);

+}

+static void unlock_srbm(struct kgd_dev *kgd) +{

• struct radeon_device *rdev = get_radeon_device(kgd);
• write_register(kgd, SRBM_GFX_CNTL, 0);
• mutex_unlock(&rdev->srbm_mutex);

+}

+static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id, uint32_t queue_id) +{

• uint32_t mec = (++pipe_id / CIK_PIPE_PER_MEC) + 1;
• uint32_t pipe = (pipe_id % CIK_PIPE_PER_MEC);
• lock_srbm(kgd, mec, pipe, queue_id, 0);

+}

+static void release_queue(struct kgd_dev *kgd) +{

• unlock_srbm(kgd);

+}

+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)
  

+{

• lock_srbm(kgd, 0, 0, 0, vmid);
• write_register(kgd, SH_MEM_CONFIG, sh_mem_config);
• write_register(kgd, SH_MEM_APE1_BASE, sh_mem_ape1_base);
• write_register(kgd, SH_MEM_APE1_LIMIT, sh_mem_ape1_limit);
• write_register(kgd, SH_MEM_BASES, sh_mem_bases);
• unlock_srbm(kgd);

+}

+static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid, unsigned int vmid) +{

• /* 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_register(kgd, ATC_VMID0_PASID_MAPPING + vmid*sizeof(uint32_t), pasid_mapping);
• while (!(read_register(kgd, ATC_VMID_PASID_MAPPING_UPDATE_STATUS) & (1U << vmid)))

• cpu_relax();
  

• write_register(kgd, ATC_VMID_PASID_MAPPING_UPDATE_STATUS, 1U << vmid);
• return 0;

+}

+static int kgd_init_memory(struct kgd_dev *kgd) +{

• /* Configure apertures:

• * LDS:         0x60000000'00000000 - 0x60000001'00000000 (4GB)
  

• * Scratch:     0x60000001'00000000 - 0x60000002'00000000 (4GB)
  

• * GPUVM:       0x60010000'00000000 - 0x60020000'00000000 (1TB)
  

• */
  

Again this whole aperture business need some explanation somewhere.

Added explanation in v3.

...

• int i;
• uint32_t sh_mem_bases = PRIVATE_BASE(0x6000) | SHARED_BASE(0x6000);
• for (i = 8; i < 16; i++) {

• uint32_t sh_mem_config;
  

• lock_srbm(kgd, 0, 0, 0, i);
  

• sh_mem_config = ALIGNMENT_MODE(SH_MEM_ALIGNMENT_MODE_UNALIGNED);
  

• sh_mem_config |= DEFAULT_MTYPE(MTYPE_NONCACHED);
  

• write_register(kgd, SH_MEM_CONFIG, sh_mem_config);
  

• write_register(kgd, SH_MEM_BASES, sh_mem_bases);
  

• /* Scratch aperture is not supported for now. */
  

• write_register(kgd, SH_STATIC_MEM_CONFIG, 0);
  

• /* APE1 disabled for now. */
  

• write_register(kgd, SH_MEM_APE1_BASE, 1);
  

• write_register(kgd, SH_MEM_APE1_LIMIT, 0);
  

• unlock_srbm(kgd);
  

• }
• return 0;

+}

+static int kgd_init_pipeline(struct kgd_dev *kgd, uint32_t pipe_id, uint32_t hpd_size, uint64_t hpd_gpu_addr) +{

• uint32_t mec = (++pipe_id / CIK_PIPE_PER_MEC) + 1;
• uint32_t pipe = (pipe_id % CIK_PIPE_PER_MEC);
• lock_srbm(kgd, mec, pipe, 0, 0);
• write_register(kgd, CP_HPD_EOP_BASE_ADDR, lower_32(hpd_gpu_addr >> 8));
• write_register(kgd, CP_HPD_EOP_BASE_ADDR_HI, upper_32(hpd_gpu_addr >> 8));
• write_register(kgd, CP_HPD_EOP_VMID, 0);
• write_register(kgd, CP_HPD_EOP_CONTROL, hpd_size);
• unlock_srbm(kgd);
• return 0;

+}

+static inline struct cik_mqd *get_mqd(void *mqd) +{

• return (struct cik_mqd *)mqd;

+}

+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_shadow, is_wptr_shadow_valid;
• struct cik_mqd *m;
• m = get_mqd(mqd);
• is_wptr_shadow_valid = !get_user(wptr_shadow, wptr);
• acquire_queue(kgd, pipe_id, queue_id);
• write_register(kgd, CP_MQD_BASE_ADDR, m->queue_state.cp_mqd_base_addr);
• write_register(kgd, CP_MQD_BASE_ADDR_HI, m->queue_state.cp_mqd_base_addr_hi);
• write_register(kgd, CP_MQD_CONTROL, m->queue_state.cp_mqd_control);
• write_register(kgd, CP_HQD_PQ_BASE, m->queue_state.cp_hqd_pq_base);
• write_register(kgd, CP_HQD_PQ_BASE_HI, m->queue_state.cp_hqd_pq_base_hi);
• write_register(kgd, CP_HQD_PQ_CONTROL, m->queue_state.cp_hqd_pq_control);
• write_register(kgd, CP_HQD_IB_CONTROL, m->queue_state.cp_hqd_ib_control);
• write_register(kgd, CP_HQD_IB_BASE_ADDR, m->queue_state.cp_hqd_ib_base_addr);
• write_register(kgd, CP_HQD_IB_BASE_ADDR_HI, m->queue_state.cp_hqd_ib_base_addr_hi);
• write_register(kgd, CP_HQD_IB_RPTR, m->queue_state.cp_hqd_ib_rptr);
• write_register(kgd, CP_HQD_PERSISTENT_STATE, m->queue_state.cp_hqd_persistent_state);
• write_register(kgd, CP_HQD_SEMA_CMD, m->queue_state.cp_hqd_sema_cmd);
• write_register(kgd, CP_HQD_MSG_TYPE, m->queue_state.cp_hqd_msg_type);
• write_register(kgd, CP_HQD_ATOMIC0_PREOP_LO, m->queue_state.cp_hqd_atomic0_preop_lo);
• write_register(kgd, CP_HQD_ATOMIC0_PREOP_HI, m->queue_state.cp_hqd_atomic0_preop_hi);
• write_register(kgd, CP_HQD_ATOMIC1_PREOP_LO, m->queue_state.cp_hqd_atomic1_preop_lo);
• write_register(kgd, CP_HQD_ATOMIC1_PREOP_HI, m->queue_state.cp_hqd_atomic1_preop_hi);
• write_register(kgd, CP_HQD_PQ_RPTR_REPORT_ADDR, m->queue_state.cp_hqd_pq_rptr_report_addr);
• write_register(kgd, CP_HQD_PQ_RPTR_REPORT_ADDR_HI, m->queue_state.cp_hqd_pq_rptr_report_addr_hi);
• write_register(kgd, CP_HQD_PQ_RPTR, m->queue_state.cp_hqd_pq_rptr);
• write_register(kgd, CP_HQD_PQ_WPTR_POLL_ADDR, m->queue_state.cp_hqd_pq_wptr_poll_addr);
• write_register(kgd, CP_HQD_PQ_WPTR_POLL_ADDR_HI, m->queue_state.cp_hqd_pq_wptr_poll_addr_hi);
• write_register(kgd, CP_HQD_PQ_DOORBELL_CONTROL, m->queue_state.cp_hqd_pq_doorbell_control);
• write_register(kgd, CP_HQD_VMID, m->queue_state.cp_hqd_vmid);
• write_register(kgd, CP_HQD_QUANTUM, m->queue_state.cp_hqd_quantum);
• write_register(kgd, CP_HQD_PIPE_PRIORITY, m->queue_state.cp_hqd_pipe_priority);
• write_register(kgd, CP_HQD_QUEUE_PRIORITY, m->queue_state.cp_hqd_queue_priority);
• write_register(kgd, CP_HQD_HQ_SCHEDULER0, m->queue_state.cp_hqd_hq_scheduler0);
• write_register(kgd, CP_HQD_HQ_SCHEDULER1, m->queue_state.cp_hqd_hq_scheduler1);
• if (is_wptr_shadow_valid)

• write_register(kgd, CP_HQD_PQ_WPTR, wptr_shadow);
  

• write_register(kgd, CP_HQD_ACTIVE, m->queue_state.cp_hqd_active);
• release_queue(kgd);
• return 0;

+}

+static bool kgd_hqd_is_occupies(struct kgd_dev *kgd, uint64_t queue_address, uint32_t pipe_id, uint32_t queue_id) +{

• uint32_t act;
• bool retval = false;
• uint32_t low, high;
• acquire_queue(kgd, pipe_id, queue_id);
• act = read_register(kgd, CP_HQD_ACTIVE);
• if (act) {

• low = lower_32(queue_address >> 8);
  

• high = upper_32(queue_address >> 8);
  

• if (low == read_register(kgd, CP_HQD_PQ_BASE) &&
  

• 		high == read_register(kgd, CP_HQD_PQ_BASE_HI))
  

• 	retval = true;
  

• }
• release_queue(kgd);
• return retval;

+}

+static int kgd_hqd_destroy(struct kgd_dev *kgd, bool is_reset,

• 		unsigned int timeout, uint32_t pipe_id,
  

• 		uint32_t queue_id)
  

+{

• int status = 0;
• bool sync = (timeout > 0) ? true : false;
• acquire_queue(kgd, pipe_id, queue_id);
• write_register(kgd, CP_HQD_PQ_DOORBELL_CONTROL, 0);
• if (is_reset)

• write_register(kgd, CP_HQD_DEQUEUE_REQUEST, DEQUEUE_REQUEST_RESET);
  

• else

• write_register(kgd, CP_HQD_DEQUEUE_REQUEST, DEQUEUE_REQUEST_DRAIN);
  

• while (read_register(kgd, CP_HQD_ACTIVE) != 0) {

• if (sync && timeout <= 0) {
  

• 	status = -EBUSY;
  

• 	break;
  

• }
  

• msleep(20);
  

• if (sync) {
  

• 	if (timeout >= 20)
  

• 		timeout -= 20;
  

• 	else
  

• 		timeout = 0;
  

• }
  

• }
• release_queue(kgd);
• return status;

+} diff --git a/drivers/gpu/drm/radeon/radeon_kfd.h b/drivers/gpu/drm/radeon/radeon_kfd.h new file mode 100644 index 0000000..5171726 --- /dev/null +++ b/drivers/gpu/drm/radeon/radeon_kfd.h @@ -0,0 +1,119 @@ +/*

• • 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.
• */

+/*

• • radeon_kfd.h defines the private interface between the
• • AMD kernel graphics drivers and the AMD KFD.
• */

+#ifndef RADEON_KFD_H_INCLUDED +#define RADEON_KFD_H_INCLUDED

+#include <linux/types.h>

+struct pci_dev;

+#define KFD_INTERFACE_VERSION 1

+struct kfd_dev; +struct kgd_dev;

+struct kgd_mem;

+struct radeon_device;

+enum kgd_memory_pool {

• KGD_POOL_SYSTEM_CACHEABLE = 1,
• KGD_POOL_SYSTEM_WRITECOMBINE = 2,
• KGD_POOL_FRAMEBUFFER = 3,

+};

+struct kgd2kfd_shared_resources {

• unsigned int compute_vmid_bitmap; /* Bit n == 1 means VMID n is available for KFD. */
• unsigned int first_compute_pipe; /* Compute pipes are counted starting from MEC0/pipe0 as 0. */
• unsigned int compute_pipe_count; /* Number of MEC pipes available for KFD. */
• phys_addr_t doorbell_physical_address; /* Base address of doorbell aperture. */
• size_t doorbell_aperture_size; /* Size in bytes of doorbell aperture. */
• size_t doorbell_start_offset; /* Number of bytes at start of aperture reserved for KGD. */

+};

+struct kgd2kfd_calls {

• void (*exit)(void);
• struct kfd_dev* (*probe)(struct kgd_dev *kgd, struct pci_dev *pdev);
• bool (*device_init)(struct kfd_dev *kfd, const struct kgd2kfd_shared_resources *gpu_resources);
• void (*device_exit)(struct kfd_dev *kfd);
• void (*interrupt)(struct kfd_dev *kfd, const void *ih_ring_entry);
• void (*suspend)(struct kfd_dev *kfd);
• int (*resume)(struct kfd_dev *kfd);

+};

+struct kfd2kgd_calls {

• /* Memory management. */
• int (*allocate_mem)(struct kgd_dev *kgd,

• 		size_t size,
  

• 		size_t alignment,
  

• 		enum kgd_memory_pool pool,
  

• 		struct kgd_mem **memory_handle);
  

• void (*free_mem)(struct kgd_dev *kgd, struct kgd_mem *memory_handle);
• int (*gpumap_mem)(struct kgd_dev *kgd, struct kgd_mem *mem, uint64_t *vmid0_address);
• void (*ungpumap_mem)(struct kgd_dev *kgd, struct kgd_mem *mem);
• int (*kmap_mem)(struct kgd_dev *kgd, struct kgd_mem *mem, void **ptr);
• void (*unkmap_mem)(struct kgd_dev *kgd, struct kgd_mem *mem);
• uint64_t (*get_vmem_size)(struct kgd_dev *kgd);
• uint64_t (*get_gpu_clock_counter)(struct kgd_dev *kgd);
• uint32_t (*get_max_engine_clock_in_mhz)(struct kgd_dev *kgd);
• /* Register access functions */
• void (*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);
  

• int (*set_pasid_vmid_mapping)(struct kgd_dev *kgd, unsigned int pasid, unsigned int vmid);
• int (*init_memory)(struct kgd_dev *kgd);
• int (*init_pipeline)(struct kgd_dev *kgd, uint32_t pipe_id, uint32_t hpd_size, uint64_t hpd_gpu_addr);
• int (*hqd_load)(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id, uint32_t queue_id, uint32_t __user *wptr);
• bool (*hqd_is_occupies)(struct kgd_dev *kgd, uint64_t queue_address, uint32_t pipe_id, uint32_t queue_id);
• int (*hqd_destroy)(struct kgd_dev *kgd, bool is_reset, unsigned int timeout,

• 		uint32_t pipe_id, uint32_t queue_id);
  

+};

Such interface should be documented looks at ttm or any other function structure inside the kernel for example on how to document those.

Done in v3.

Oded

...

+bool radeon_kfd_init(void); +void radeon_kfd_fini(void); +bool kgd2kfd_init(unsigned interface_version,

•   const struct kfd2kgd_calls *f2g,
  

•   const struct kgd2kfd_calls **g2f);
  

+void radeon_kfd_suspend(struct radeon_device *rdev); +int radeon_kfd_resume(struct radeon_device *rdev); +void radeon_kfd_interrupt(struct radeon_device *rdev,

• 	const void *ih_ring_entry);
  

+void radeon_kfd_device_probe(struct radeon_device *rdev); +void radeon_kfd_device_init(struct radeon_device *rdev); +void radeon_kfd_device_fini(struct radeon_device *rdev);

+#endif

diff --git a/drivers/gpu/drm/radeon/radeon_kms.c b/drivers/gpu/drm/radeon/radeon_kms.c index 35d9318..929beda 100644 --- a/drivers/gpu/drm/radeon/radeon_kms.c +++ b/drivers/gpu/drm/radeon/radeon_kms.c @@ -34,6 +34,8 @@ #include <linux/slab.h> #include <linux/pm_runtime.h>

+#include "radeon_kfd.h"

#if defined(CONFIG_VGA_SWITCHEROO) bool radeon_has_atpx(void); #else @@ -63,6 +65,8 @@ int radeon_driver_unload_kms(struct drm_device *dev)

pm_runtime_get_sync(dev->dev);

• radeon_kfd_device_fini(rdev);

• radeon_acpi_fini(rdev);

  radeon_modeset_fini(rdev);

@@ -142,6 +146,9 @@ int radeon_driver_load_kms(struct drm_device *dev, unsigned long flags) "Error during ACPI methods call\n"); }

• radeon_kfd_device_probe(rdev);
• radeon_kfd_device_init(rdev);
• if (radeon_is_px(dev)) { pm_runtime_use_autosuspend(dev->dev); pm_runtime_set_autosuspend_delay(dev->dev, 5000);

-- 1.9.1