From: Tong Liu01 Tong.Liu01@amd.com

Move TMR region from top of FB to 2MB for FFBM, so we need to reserve TMR region firstly to make sure TMR can be allocated at 2MB

Signed-off-by: Tong Liu01 Tong.Liu01@amd.com Acked-by: Christian König christian.koenig@amd.com Signed-off-by: Alex Deucher alexander.deucher@amd.com (cherry picked from commit 4864f2ee9ee2acf4a1009b58fbc62f17fa086d4e) Signed-off-by: Mario Limonciello mario.limonciello@amd.com --- .../gpu/drm/amd/amdgpu/amdgpu_atomfirmware.c | 104 ++++++++++++++---- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 51 +++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h | 5 + drivers/gpu/drm/amd/include/atomfirmware.h | 63 +++++++++-- 4 files changed, 191 insertions(+), 32 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_atomfirmware.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_atomfirmware.c index b81b77a9efa6..9b97fa39d47a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_atomfirmware.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_atomfirmware.c @@ -101,39 +101,97 @@ void amdgpu_atomfirmware_scratch_regs_init(struct amdgpu_device *adev) } }

+static int amdgpu_atomfirmware_allocate_fb_v2_1(struct amdgpu_device *adev, + struct vram_usagebyfirmware_v2_1 *fw_usage, int *usage_bytes) +{ + uint32_t start_addr, fw_size, drv_size; + + start_addr = le32_to_cpu(fw_usage->start_address_in_kb); + fw_size = le16_to_cpu(fw_usage->used_by_firmware_in_kb); + drv_size = le16_to_cpu(fw_usage->used_by_driver_in_kb); + + DRM_DEBUG("atom firmware v2_1 requested %08x %dkb fw %dkb drv\n", + start_addr, + fw_size, + drv_size); + + if ((start_addr & ATOM_VRAM_OPERATION_FLAGS_MASK) == + (uint32_t)(ATOM_VRAM_BLOCK_SRIOV_MSG_SHARE_RESERVATION << + ATOM_VRAM_OPERATION_FLAGS_SHIFT)) { + /* Firmware request VRAM reservation for SR-IOV */ + adev->mman.fw_vram_usage_start_offset = (start_addr & + (~ATOM_VRAM_OPERATION_FLAGS_MASK)) << 10; + adev->mman.fw_vram_usage_size = fw_size << 10; + /* Use the default scratch size */ + *usage_bytes = 0; + } else { + *usage_bytes = drv_size << 10; + } + return 0; +} + +static int amdgpu_atomfirmware_allocate_fb_v2_2(struct amdgpu_device *adev, + struct vram_usagebyfirmware_v2_2 *fw_usage, int *usage_bytes) +{ + uint32_t fw_start_addr, fw_size, drv_start_addr, drv_size; + + fw_start_addr = le32_to_cpu(fw_usage->fw_region_start_address_in_kb); + fw_size = le16_to_cpu(fw_usage->used_by_firmware_in_kb); + + drv_start_addr = le32_to_cpu(fw_usage->driver_region0_start_address_in_kb); + drv_size = le32_to_cpu(fw_usage->used_by_driver_region0_in_kb); + + DRM_DEBUG("atom requested fw start at %08x %dkb and drv start at %08x %dkb\n", + fw_start_addr, + fw_size, + drv_start_addr, + drv_size); + + if ((fw_start_addr & (ATOM_VRAM_BLOCK_NEEDS_NO_RESERVATION << 30)) == 0) { + /* Firmware request VRAM reservation for SR-IOV */ + adev->mman.fw_vram_usage_start_offset = (fw_start_addr & + (~ATOM_VRAM_OPERATION_FLAGS_MASK)) << 10; + adev->mman.fw_vram_usage_size = fw_size << 10; + } + + if ((drv_start_addr & (ATOM_VRAM_BLOCK_NEEDS_NO_RESERVATION << 30)) == 0) { + /* driver request VRAM reservation for SR-IOV */ + adev->mman.drv_vram_usage_start_offset = (drv_start_addr & + (~ATOM_VRAM_OPERATION_FLAGS_MASK)) << 10; + adev->mman.drv_vram_usage_size = drv_size << 10; + } + + *usage_bytes = 0; + return 0; +} + int amdgpu_atomfirmware_allocate_fb_scratch(struct amdgpu_device *adev) { struct atom_context *ctx = adev->mode_info.atom_context; int index = get_index_into_master_table(atom_master_list_of_data_tables_v2_1, vram_usagebyfirmware); - struct vram_usagebyfirmware_v2_1 *firmware_usage; - uint32_t start_addr, size; + struct vram_usagebyfirmware_v2_1 *fw_usage_v2_1; + struct vram_usagebyfirmware_v2_2 *fw_usage_v2_2; uint16_t data_offset; + uint8_t frev, crev; int usage_bytes = 0;

- if (amdgpu_atom_parse_data_header(ctx, index, NULL, NULL, NULL, &data_offset)) { - firmware_usage = (struct vram_usagebyfirmware_v2_1 *)(ctx->bios + data_offset); - DRM_DEBUG("atom firmware requested %08x %dkb fw %dkb drv\n", - le32_to_cpu(firmware_usage->start_address_in_kb), - le16_to_cpu(firmware_usage->used_by_firmware_in_kb), - le16_to_cpu(firmware_usage->used_by_driver_in_kb)); - - start_addr = le32_to_cpu(firmware_usage->start_address_in_kb); - size = le16_to_cpu(firmware_usage->used_by_firmware_in_kb); - - if ((uint32_t)(start_addr & ATOM_VRAM_OPERATION_FLAGS_MASK) == - (uint32_t)(ATOM_VRAM_BLOCK_SRIOV_MSG_SHARE_RESERVATION << - ATOM_VRAM_OPERATION_FLAGS_SHIFT)) { - /* Firmware request VRAM reservation for SR-IOV */ - adev->mman.fw_vram_usage_start_offset = (start_addr & - (~ATOM_VRAM_OPERATION_FLAGS_MASK)) << 10; - adev->mman.fw_vram_usage_size = size << 10; - /* Use the default scratch size */ - usage_bytes = 0; - } else { - usage_bytes = le16_to_cpu(firmware_usage->used_by_driver_in_kb) << 10; + if (amdgpu_atom_parse_data_header(ctx, index, NULL, &frev, &crev, &data_offset)) { + if (frev == 2 && crev == 1) { + fw_usage_v2_1 = + (struct vram_usagebyfirmware_v2_1 *)(ctx->bios + data_offset); + amdgpu_atomfirmware_allocate_fb_v2_1(adev, + fw_usage_v2_1, + &usage_bytes); + } else if (frev >= 2 && crev >= 2) { + fw_usage_v2_2 = + (struct vram_usagebyfirmware_v2_2 *)(ctx->bios + data_offset); + amdgpu_atomfirmware_allocate_fb_v2_2(adev, + fw_usage_v2_2, + &usage_bytes); } } + ctx->scratch_size_bytes = 0; if (usage_bytes == 0) usage_bytes = 20 * 1024; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index b64938ed8cb6..c66e7c79d653 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -1537,6 +1537,23 @@ static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev) NULL, &adev->mman.fw_vram_usage_va); }

+/* + * Driver Reservation functions + */ +/** + * amdgpu_ttm_drv_reserve_vram_fini - free drv reserved vram + * + * @adev: amdgpu_device pointer + * + * free drv reserved vram if it has been reserved. + */ +static void amdgpu_ttm_drv_reserve_vram_fini(struct amdgpu_device *adev) +{ + amdgpu_bo_free_kernel(&adev->mman.drv_vram_usage_reserved_bo, + NULL, + NULL); +} + /** * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw * @@ -1563,6 +1580,31 @@ static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev) &adev->mman.fw_vram_usage_va); }

+/** + * amdgpu_ttm_drv_reserve_vram_init - create bo vram reservation from driver + * + * @adev: amdgpu_device pointer + * + * create bo vram reservation from drv. + */ +static int amdgpu_ttm_drv_reserve_vram_init(struct amdgpu_device *adev) +{ + uint64_t vram_size = adev->gmc.visible_vram_size; + + adev->mman.drv_vram_usage_reserved_bo = NULL; + + if (adev->mman.drv_vram_usage_size == 0 || + adev->mman.drv_vram_usage_size > vram_size) + return 0; + + return amdgpu_bo_create_kernel_at(adev, + adev->mman.drv_vram_usage_start_offset, + adev->mman.drv_vram_usage_size, + AMDGPU_GEM_DOMAIN_VRAM, + &adev->mman.drv_vram_usage_reserved_bo, + NULL); +} + /* * Memoy training reservation functions */ @@ -1730,6 +1772,14 @@ int amdgpu_ttm_init(struct amdgpu_device *adev) return r; }

+ /* + *The reserved vram for driver must be pinned to the specified + *place on the VRAM, so reserve it early. + */ + r = amdgpu_ttm_drv_reserve_vram_init(adev); + if (r) + return r; + /* * only NAVI10 and onwards ASIC support for IP discovery. * If IP discovery enabled, a block of memory should be @@ -1855,6 +1905,7 @@ void amdgpu_ttm_fini(struct amdgpu_device *adev) amdgpu_bo_free_kernel(&adev->mman.sdma_access_bo, NULL, &adev->mman.sdma_access_ptr); amdgpu_ttm_fw_reserve_vram_fini(adev); + amdgpu_ttm_drv_reserve_vram_fini(adev);

if (drm_dev_enter(adev_to_drm(adev), &idx)) {

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h index a37207011a69..b391c8d076ff 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h @@ -86,6 +86,11 @@ struct amdgpu_mman { struct amdgpu_bo *fw_vram_usage_reserved_bo; void *fw_vram_usage_va;

+ /* driver VRAM reservation */ + u64 drv_vram_usage_start_offset; + u64 drv_vram_usage_size; + struct amdgpu_bo *drv_vram_usage_reserved_bo; + /* PAGE_SIZE'd BO for process memory r/w over SDMA. */ struct amdgpu_bo *sdma_access_bo; void *sdma_access_ptr; diff --git a/drivers/gpu/drm/amd/include/atomfirmware.h b/drivers/gpu/drm/amd/include/atomfirmware.h index ff855cb21d3f..bbe1337a8cee 100644 --- a/drivers/gpu/drm/amd/include/atomfirmware.h +++ b/drivers/gpu/drm/amd/include/atomfirmware.h @@ -705,20 +705,65 @@ struct atom_gpio_pin_lut_v2_1 };

-/* - *************************************************************************** - Data Table vram_usagebyfirmware structure - *************************************************************************** -*/ +/* + * VBIOS/PRE-OS always reserve a FB region at the top of frame buffer. driver should not write + * access that region. driver can allocate their own reservation region as long as it does not + * overlap firwmare's reservation region. + * if (pre-NV1X) atom data table firmwareInfoTable version < 3.3: + * in this case, atom data table vram_usagebyfirmwareTable version always <= 2.1 + * if VBIOS/UEFI GOP is posted: + * VBIOS/UEFIGOP update used_by_firmware_in_kb = total reserved size by VBIOS + * update start_address_in_kb = total_mem_size_in_kb - used_by_firmware_in_kb; + * ( total_mem_size_in_kb = reg(CONFIG_MEMSIZE)<<10) + * driver can allocate driver reservation region under firmware reservation, + * used_by_driver_in_kb = driver reservation size + * driver reservation start address = (start_address_in_kb - used_by_driver_in_kb) + * Comment1[hchan]: There is only one reservation at the beginning of the FB reserved by + * host driver. Host driver would overwrite the table with the following + * used_by_firmware_in_kb = total reserved size for pf-vf info exchange and + * set SRIOV_MSG_SHARE_RESERVATION mask start_address_in_kb = 0 + * else there is no VBIOS reservation region: + * driver must allocate driver reservation region at top of FB. + * driver set used_by_driver_in_kb = driver reservation size + * driver reservation start address = (total_mem_size_in_kb - used_by_driver_in_kb) + * same as Comment1 + * else (NV1X and after): + * if VBIOS/UEFI GOP is posted: + * VBIOS/UEFIGOP update: + * used_by_firmware_in_kb = atom_firmware_Info_v3_3.fw_reserved_size_in_kb; + * start_address_in_kb = total_mem_size_in_kb - used_by_firmware_in_kb; + * (total_mem_size_in_kb = reg(CONFIG_MEMSIZE)<<10) + * if vram_usagebyfirmwareTable version <= 2.1: + * driver can allocate driver reservation region under firmware reservation, + * driver set used_by_driver_in_kb = driver reservation size + * driver reservation start address = start_address_in_kb - used_by_driver_in_kb + * same as Comment1 + * else driver can: + * allocate it reservation any place as long as it does overlap pre-OS FW reservation area + * set used_by_driver_region0_in_kb = driver reservation size + * set driver_region0_start_address_in_kb = driver reservation region start address + * Comment2[hchan]: Host driver can set used_by_firmware_in_kb and start_address_in_kb to + * zero as the reservation for VF as it doesn’t exist. And Host driver should also + * update atom_firmware_Info table to remove the same VBIOS reservation as well. + */

struct vram_usagebyfirmware_v2_1 { - struct atom_common_table_header table_header; - uint32_t start_address_in_kb; - uint16_t used_by_firmware_in_kb; - uint16_t used_by_driver_in_kb; + struct atom_common_table_header table_header; + uint32_t start_address_in_kb; + uint16_t used_by_firmware_in_kb; + uint16_t used_by_driver_in_kb; };

+struct vram_usagebyfirmware_v2_2 { + struct atom_common_table_header table_header; + uint32_t fw_region_start_address_in_kb; + uint16_t used_by_firmware_in_kb; + uint16_t reserved; + uint32_t driver_region0_start_address_in_kb; + uint32_t used_by_driver_region0_in_kb; + uint32_t reserved32[7]; +};

/* ***************************************************************************

From: Lijo Lazar lijo.lazar@amd.com

Use the generic term fw_reserved_memory for FW reserve region. This region may also hold discovery TMR in addition to other reserve regions. This region size could be larger than discovery tmr size, hence don't change the discovery tmr size based on this.

Signed-off-by: Lijo Lazar lijo.lazar@amd.com Reviewed-by: Le Ma le.ma@amd.com Signed-off-by: Alex Deucher alexander.deucher@amd.com (cherry picked from commit db3b5cb64a9ca301d14ed027e470834316720e42) This change fixes reading IP discovery from debugfs.

It needed to be hand modified because: * GC 9.4.3 support isn't introduced in older kernels until 228ce176434b ("drm/amdgpu: Handle VRAM dependencies on GFXIP9.4.3") * It also changed because of 58ab2c08d708 (drm/amdgpu: use VRAM|GTT for a bunch of kernel allocations) not being present. Link: https://gitlab.freedesktop.org/drm/amd/-/issues/2748 Signed-off-by: Mario Limonciello mario.limonciello@amd.com --- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 33 ++++++++++++++----------- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h | 3 ++- 2 files changed, 21 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index 917147dee8a2..10469f20a10c 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -1625,14 +1625,15 @@ static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev) return 0; }

-static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev) +static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev, + uint32_t reserve_size) { struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;

memset(ctx, 0, sizeof(*ctx));

ctx->c2p_train_data_offset = - ALIGN((adev->gmc.mc_vram_size - adev->mman.discovery_tmr_size - SZ_1M), SZ_1M); + ALIGN((adev->gmc.mc_vram_size - reserve_size - SZ_1M), SZ_1M); ctx->p2c_train_data_offset = (adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET); ctx->train_data_size = @@ -1650,9 +1651,10 @@ static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev) */ static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev) { - int ret; struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx; bool mem_train_support = false; + uint32_t reserve_size = 0; + int ret;

if (!amdgpu_sriov_vf(adev)) { if (amdgpu_atomfirmware_mem_training_supported(adev)) @@ -1668,14 +1670,15 @@ static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev) * Otherwise, fallback to legacy approach to check and reserve tmr block for ip * discovery data and G6 memory training data respectively */ - adev->mman.discovery_tmr_size = - amdgpu_atomfirmware_get_fw_reserved_fb_size(adev); - if (!adev->mman.discovery_tmr_size) - adev->mman.discovery_tmr_size = DISCOVERY_TMR_OFFSET; + if (adev->bios) + reserve_size = + amdgpu_atomfirmware_get_fw_reserved_fb_size(adev); + if (!reserve_size) + reserve_size = DISCOVERY_TMR_OFFSET;

if (mem_train_support) { /* reserve vram for mem train according to TMR location */ - amdgpu_ttm_training_data_block_init(adev); + amdgpu_ttm_training_data_block_init(adev, reserve_size); ret = amdgpu_bo_create_kernel_at(adev, ctx->c2p_train_data_offset, ctx->train_data_size, @@ -1690,13 +1693,14 @@ static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev) }

ret = amdgpu_bo_create_kernel_at(adev, - adev->gmc.real_vram_size - adev->mman.discovery_tmr_size, - adev->mman.discovery_tmr_size, - &adev->mman.discovery_memory, + adev->gmc.real_vram_size - reserve_size, + reserve_size, + &adev->mman.fw_reserved_memory, NULL); if (ret) { DRM_ERROR("alloc tmr failed(%d)!\n", ret); - amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL); + amdgpu_bo_free_kernel(&adev->mman.fw_reserved_memory, + NULL, NULL); return ret; }

@@ -1890,8 +1894,9 @@ void amdgpu_ttm_fini(struct amdgpu_device *adev) /* return the stolen vga memory back to VRAM */ amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL); amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL); - /* return the IP Discovery TMR memory back to VRAM */ - amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL); + /* return the FW reserved memory back to VRAM */ + amdgpu_bo_free_kernel(&adev->mman.fw_reserved_memory, NULL, + NULL); if (adev->mman.stolen_reserved_size) amdgpu_bo_free_kernel(&adev->mman.stolen_reserved_memory, NULL, NULL); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h index b391c8d076ff..0fefa5e3a524 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h @@ -78,7 +78,8 @@ struct amdgpu_mman { /* discovery */ uint8_t *discovery_bin; uint32_t discovery_tmr_size; - struct amdgpu_bo *discovery_memory; + /* fw reserved memory */ + struct amdgpu_bo *fw_reserved_memory;

/* firmware VRAM reservation */ u64 fw_vram_usage_start_offset;