These changes are the raw update to linux-4.4.6-rt14. Kernel sources
[kvmfornfv.git] / kernel / drivers / gpu / drm / radeon / radeon_vce.c
index 0de5711..7eb1ae7 100644 (file)
 #define VCE_IDLE_TIMEOUT_MS    1000
 
 /* Firmware Names */
+#define FIRMWARE_TAHITI        "radeon/TAHITI_vce.bin"
 #define FIRMWARE_BONAIRE       "radeon/BONAIRE_vce.bin"
 
+MODULE_FIRMWARE(FIRMWARE_TAHITI);
 MODULE_FIRMWARE(FIRMWARE_BONAIRE);
 
 static void radeon_vce_idle_work_handler(struct work_struct *work);
@@ -63,6 +65,14 @@ int radeon_vce_init(struct radeon_device *rdev)
        INIT_DELAYED_WORK(&rdev->vce.idle_work, radeon_vce_idle_work_handler);
 
        switch (rdev->family) {
+       case CHIP_TAHITI:
+       case CHIP_PITCAIRN:
+       case CHIP_VERDE:
+       case CHIP_OLAND:
+       case CHIP_ARUBA:
+               fw_name = FIRMWARE_TAHITI;
+               break;
+
        case CHIP_BONAIRE:
        case CHIP_KAVERI:
        case CHIP_KABINI:
@@ -118,13 +128,17 @@ int radeon_vce_init(struct radeon_device *rdev)
        rdev->vce.fw_version = (start << 24) | (mid << 16) | (end << 8);
 
        /* we can only work with this fw version for now */
-       if (rdev->vce.fw_version != ((40 << 24) | (2 << 16) | (2 << 8)))
+       if ((rdev->vce.fw_version != ((40 << 24) | (2 << 16) | (2 << 8))) &&
+           (rdev->vce.fw_version != ((50 << 24) | (0 << 16) | (1 << 8))) &&
+           (rdev->vce.fw_version != ((50 << 24) | (1 << 16) | (2 << 8))))
                return -EINVAL;
 
        /* allocate firmware, stack and heap BO */
 
-       size = RADEON_GPU_PAGE_ALIGN(rdev->vce_fw->size) +
-              RADEON_VCE_STACK_SIZE + RADEON_VCE_HEAP_SIZE;
+       if (rdev->family < CHIP_BONAIRE)
+               size = vce_v1_0_bo_size(rdev);
+       else
+               size = vce_v2_0_bo_size(rdev);
        r = radeon_bo_create(rdev, size, PAGE_SIZE, true,
                             RADEON_GEM_DOMAIN_VRAM, 0, NULL, NULL,
                             &rdev->vce.vcpu_bo);
@@ -225,13 +239,17 @@ int radeon_vce_resume(struct radeon_device *rdev)
                return r;
        }
 
-       memcpy(cpu_addr, rdev->vce_fw->data, rdev->vce_fw->size);
+       memset(cpu_addr, 0, radeon_bo_size(rdev->vce.vcpu_bo));
+       if (rdev->family < CHIP_BONAIRE)
+               r = vce_v1_0_load_fw(rdev, cpu_addr);
+       else
+               memcpy(cpu_addr, rdev->vce_fw->data, rdev->vce_fw->size);
 
        radeon_bo_kunmap(rdev->vce.vcpu_bo);
 
        radeon_bo_unreserve(rdev->vce.vcpu_bo);
 
-       return 0;
+       return r;
 }
 
 /**
@@ -343,31 +361,31 @@ int radeon_vce_get_create_msg(struct radeon_device *rdev, int ring,
 
        /* stitch together an VCE create msg */
        ib.length_dw = 0;
-       ib.ptr[ib.length_dw++] = 0x0000000c; /* len */
-       ib.ptr[ib.length_dw++] = 0x00000001; /* session cmd */
-       ib.ptr[ib.length_dw++] = handle;
-
-       ib.ptr[ib.length_dw++] = 0x00000030; /* len */
-       ib.ptr[ib.length_dw++] = 0x01000001; /* create cmd */
-       ib.ptr[ib.length_dw++] = 0x00000000;
-       ib.ptr[ib.length_dw++] = 0x00000042;
-       ib.ptr[ib.length_dw++] = 0x0000000a;
-       ib.ptr[ib.length_dw++] = 0x00000001;
-       ib.ptr[ib.length_dw++] = 0x00000080;
-       ib.ptr[ib.length_dw++] = 0x00000060;
-       ib.ptr[ib.length_dw++] = 0x00000100;
-       ib.ptr[ib.length_dw++] = 0x00000100;
-       ib.ptr[ib.length_dw++] = 0x0000000c;
-       ib.ptr[ib.length_dw++] = 0x00000000;
-
-       ib.ptr[ib.length_dw++] = 0x00000014; /* len */
-       ib.ptr[ib.length_dw++] = 0x05000005; /* feedback buffer */
-       ib.ptr[ib.length_dw++] = upper_32_bits(dummy);
-       ib.ptr[ib.length_dw++] = dummy;
-       ib.ptr[ib.length_dw++] = 0x00000001;
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x0000000c); /* len */
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000001); /* session cmd */
+       ib.ptr[ib.length_dw++] = cpu_to_le32(handle);
+
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000030); /* len */
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x01000001); /* create cmd */
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000000);
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000042);
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x0000000a);
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000001);
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000080);
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000060);
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000100);
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000100);
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x0000000c);
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000000);
+
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000014); /* len */
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x05000005); /* feedback buffer */
+       ib.ptr[ib.length_dw++] = cpu_to_le32(upper_32_bits(dummy));
+       ib.ptr[ib.length_dw++] = cpu_to_le32(dummy);
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000001);
 
        for (i = ib.length_dw; i < ib_size_dw; ++i)
-               ib.ptr[i] = 0x0;
+               ib.ptr[i] = cpu_to_le32(0x0);
 
        r = radeon_ib_schedule(rdev, &ib, NULL, false);
        if (r) {
@@ -410,21 +428,21 @@ int radeon_vce_get_destroy_msg(struct radeon_device *rdev, int ring,
 
        /* stitch together an VCE destroy msg */
        ib.length_dw = 0;
-       ib.ptr[ib.length_dw++] = 0x0000000c; /* len */
-       ib.ptr[ib.length_dw++] = 0x00000001; /* session cmd */
-       ib.ptr[ib.length_dw++] = handle;
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x0000000c); /* len */
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000001); /* session cmd */
+       ib.ptr[ib.length_dw++] = cpu_to_le32(handle);
 
-       ib.ptr[ib.length_dw++] = 0x00000014; /* len */
-       ib.ptr[ib.length_dw++] = 0x05000005; /* feedback buffer */
-       ib.ptr[ib.length_dw++] = upper_32_bits(dummy);
-       ib.ptr[ib.length_dw++] = dummy;
-       ib.ptr[ib.length_dw++] = 0x00000001;
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000014); /* len */
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x05000005); /* feedback buffer */
+       ib.ptr[ib.length_dw++] = cpu_to_le32(upper_32_bits(dummy));
+       ib.ptr[ib.length_dw++] = cpu_to_le32(dummy);
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000001);
 
-       ib.ptr[ib.length_dw++] = 0x00000008; /* len */
-       ib.ptr[ib.length_dw++] = 0x02000001; /* destroy cmd */
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x00000008); /* len */
+       ib.ptr[ib.length_dw++] = cpu_to_le32(0x02000001); /* destroy cmd */
 
        for (i = ib.length_dw; i < ib_size_dw; ++i)
-               ib.ptr[i] = 0x0;
+               ib.ptr[i] = cpu_to_le32(0x0);
 
        r = radeon_ib_schedule(rdev, &ib, NULL, false);
        if (r) {
@@ -681,12 +699,12 @@ bool radeon_vce_semaphore_emit(struct radeon_device *rdev,
 {
        uint64_t addr = semaphore->gpu_addr;
 
-       radeon_ring_write(ring, VCE_CMD_SEMAPHORE);
-       radeon_ring_write(ring, (addr >> 3) & 0x000FFFFF);
-       radeon_ring_write(ring, (addr >> 23) & 0x000FFFFF);
-       radeon_ring_write(ring, 0x01003000 | (emit_wait ? 1 : 0));
+       radeon_ring_write(ring, cpu_to_le32(VCE_CMD_SEMAPHORE));
+       radeon_ring_write(ring, cpu_to_le32((addr >> 3) & 0x000FFFFF));
+       radeon_ring_write(ring, cpu_to_le32((addr >> 23) & 0x000FFFFF));
+       radeon_ring_write(ring, cpu_to_le32(0x01003000 | (emit_wait ? 1 : 0)));
        if (!emit_wait)
-               radeon_ring_write(ring, VCE_CMD_END);
+               radeon_ring_write(ring, cpu_to_le32(VCE_CMD_END));
 
        return true;
 }
@@ -701,10 +719,10 @@ bool radeon_vce_semaphore_emit(struct radeon_device *rdev,
 void radeon_vce_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
 {
        struct radeon_ring *ring = &rdev->ring[ib->ring];
-       radeon_ring_write(ring, VCE_CMD_IB);
-       radeon_ring_write(ring, ib->gpu_addr);
-       radeon_ring_write(ring, upper_32_bits(ib->gpu_addr));
-       radeon_ring_write(ring, ib->length_dw);
+       radeon_ring_write(ring, cpu_to_le32(VCE_CMD_IB));
+       radeon_ring_write(ring, cpu_to_le32(ib->gpu_addr));
+       radeon_ring_write(ring, cpu_to_le32(upper_32_bits(ib->gpu_addr)));
+       radeon_ring_write(ring, cpu_to_le32(ib->length_dw));
 }
 
 /**
@@ -720,12 +738,12 @@ void radeon_vce_fence_emit(struct radeon_device *rdev,
        struct radeon_ring *ring = &rdev->ring[fence->ring];
        uint64_t addr = rdev->fence_drv[fence->ring].gpu_addr;
 
-       radeon_ring_write(ring, VCE_CMD_FENCE);
-       radeon_ring_write(ring, addr);
-       radeon_ring_write(ring, upper_32_bits(addr));
-       radeon_ring_write(ring, fence->seq);
-       radeon_ring_write(ring, VCE_CMD_TRAP);
-       radeon_ring_write(ring, VCE_CMD_END);
+       radeon_ring_write(ring, cpu_to_le32(VCE_CMD_FENCE));
+       radeon_ring_write(ring, cpu_to_le32(addr));
+       radeon_ring_write(ring, cpu_to_le32(upper_32_bits(addr)));
+       radeon_ring_write(ring, cpu_to_le32(fence->seq));
+       radeon_ring_write(ring, cpu_to_le32(VCE_CMD_TRAP));
+       radeon_ring_write(ring, cpu_to_le32(VCE_CMD_END));
 }
 
 /**
@@ -747,7 +765,7 @@ int radeon_vce_ring_test(struct radeon_device *rdev, struct radeon_ring *ring)
                          ring->idx, r);
                return r;
        }
-       radeon_ring_write(ring, VCE_CMD_END);
+       radeon_ring_write(ring, cpu_to_le32(VCE_CMD_END));
        radeon_ring_unlock_commit(rdev, ring, false);
 
        for (i = 0; i < rdev->usec_timeout; i++) {