These changes are the raw update to linux-4.4.6-rt14. Kernel sources
[kvmfornfv.git] / kernel / drivers / gpu / drm / amd / amdkfd / kfd_device_queue_manager_vi.c
index 20553dc..7e9cae9 100644 (file)
  */
 
 #include "kfd_device_queue_manager.h"
+#include "gca/gfx_8_0_enum.h"
+#include "gca/gfx_8_0_sh_mask.h"
+#include "gca/gfx_8_0_enum.h"
+#include "oss/oss_3_0_sh_mask.h"
 
 static bool set_cache_memory_policy_vi(struct device_queue_manager *dqm,
                                   struct qcm_process_device *qpd,
@@ -32,14 +36,43 @@ static bool set_cache_memory_policy_vi(struct device_queue_manager *dqm,
 static int register_process_vi(struct device_queue_manager *dqm,
                                        struct qcm_process_device *qpd);
 static int initialize_cpsch_vi(struct device_queue_manager *dqm);
+static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
+                               struct qcm_process_device *qpd);
 
-void device_queue_manager_init_vi(struct device_queue_manager_ops *ops)
+void device_queue_manager_init_vi(struct device_queue_manager_asic_ops *ops)
 {
-       pr_warn("amdkfd: VI DQM is not currently supported\n");
-
        ops->set_cache_memory_policy = set_cache_memory_policy_vi;
        ops->register_process = register_process_vi;
        ops->initialize = initialize_cpsch_vi;
+       ops->init_sdma_vm = init_sdma_vm;
+}
+
+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 ||
+               top_address_nybble == 0);
+
+       return top_address_nybble << 12 |
+                       (top_address_nybble << 12) <<
+                       SH_MEM_BASES__SHARED_BASE__SHIFT;
 }
 
 static bool set_cache_memory_policy_vi(struct device_queue_manager *dqm,
@@ -49,13 +82,83 @@ static bool set_cache_memory_policy_vi(struct device_queue_manager *dqm,
                                   void __user *alternate_aperture_base,
                                   uint64_t alternate_aperture_size)
 {
-       return false;
+       uint32_t default_mtype;
+       uint32_t ape1_mtype;
+
+       default_mtype = (default_policy == cache_policy_coherent) ?
+                       MTYPE_CC :
+                       MTYPE_NC;
+
+       ape1_mtype = (alternate_policy == cache_policy_coherent) ?
+                       MTYPE_CC :
+                       MTYPE_NC;
+
+       qpd->sh_mem_config = (qpd->sh_mem_config &
+                       SH_MEM_CONFIG__ADDRESS_MODE_MASK) |
+               SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
+                               SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
+               default_mtype << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
+               ape1_mtype << SH_MEM_CONFIG__APE1_MTYPE__SHIFT |
+               SH_MEM_CONFIG__PRIVATE_ATC_MASK;
+
+       return true;
 }
 
 static int register_process_vi(struct device_queue_manager *dqm,
                                        struct qcm_process_device *qpd)
 {
-       return -1;
+       struct kfd_process_device *pdd;
+       unsigned int temp;
+
+       BUG_ON(!dqm || !qpd);
+
+       pdd = qpd_to_pdd(qpd);
+
+       /* check if sh_mem_config register already configured */
+       if (qpd->sh_mem_config == 0) {
+               qpd->sh_mem_config =
+                       SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
+                               SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
+                       MTYPE_CC << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
+                       MTYPE_CC << SH_MEM_CONFIG__APE1_MTYPE__SHIFT |
+                       SH_MEM_CONFIG__PRIVATE_ATC_MASK;
+
+               qpd->sh_mem_ape1_limit = 0;
+               qpd->sh_mem_ape1_base = 0;
+       }
+
+       if (qpd->pqm->process->is_32bit_user_mode) {
+               temp = get_sh_mem_bases_32(pdd);
+               qpd->sh_mem_bases = temp << SH_MEM_BASES__SHARED_BASE__SHIFT;
+               qpd->sh_mem_config |= SH_MEM_ADDRESS_MODE_HSA32 <<
+                                       SH_MEM_CONFIG__ADDRESS_MODE__SHIFT;
+       } else {
+               temp = get_sh_mem_bases_nybble_64(pdd);
+               qpd->sh_mem_bases = compute_sh_mem_bases_64bit(temp);
+               qpd->sh_mem_config |= SH_MEM_ADDRESS_MODE_HSA64 <<
+                       SH_MEM_CONFIG__ADDRESS_MODE__SHIFT;
+       }
+
+       pr_debug("kfd: is32bit process: %d sh_mem_bases nybble: 0x%X and register 0x%X\n",
+               qpd->pqm->process->is_32bit_user_mode, temp, qpd->sh_mem_bases);
+
+       return 0;
+}
+
+static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
+                               struct qcm_process_device *qpd)
+{
+       uint32_t value = (1 << SDMA0_RLC0_VIRTUAL_ADDR__ATC__SHIFT);
+
+       if (q->process->is_32bit_user_mode)
+               value |= (1 << SDMA0_RLC0_VIRTUAL_ADDR__PTR32__SHIFT) |
+                               get_sh_mem_bases_32(qpd_to_pdd(qpd));
+       else
+               value |= ((get_sh_mem_bases_nybble_64(qpd_to_pdd(qpd))) <<
+                               SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE__SHIFT) &
+                               SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE_MASK;
+
+       q->properties.sdma_vm_addr = value;
 }
 
 static int initialize_cpsch_vi(struct device_queue_manager *dqm)