On 10/16/2025 6:36 PM, Alex Deucher wrote:
On Wed, Oct 15, 2025 at 5:56 PM Ellen Pan <[email protected]> wrote:1. Added VF logic in amdgpu_virt to init IP discovery using the offsets from dynamic(v2) critical regions; 2. Added VF logic in amdgpu_virt to init bios image using the offsets from dynamic(v2) critical regions; Signed-off-by: Ellen Pan <[email protected]> --- drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c | 34 ++++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c | 23 ++++++++++-- drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c | 37 +++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h | 2 + 4 files changed, 83 insertions(+), 13 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c index 00e96419fcda..070fd61f8463 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_bios.c @@ -96,11 +96,12 @@ void amdgpu_bios_release(struct amdgpu_device *adev) * part of the system bios. On boot, the system bios puts a * copy of the igp rom at the start of vram if a discrete card is * present. - * For SR-IOV, the vbios image is also put in VRAM in the VF. + * For SR-IOV, if dynamic critical region is not enabled, + * the vbios image is also put at the start of VRAM in the VF. */ static bool amdgpu_read_bios_from_vram(struct amdgpu_device *adev) { - uint8_t __iomem *bios; + uint8_t __iomem *bios = NULL; resource_size_t vram_base; resource_size_t size = 256 * 1024; /* ??? */ @@ -114,18 +115,33 @@ static bool amdgpu_read_bios_from_vram(struct amdgpu_device *adev) adev->bios = NULL; vram_base = pci_resource_start(adev->pdev, 0); - bios = ioremap_wc(vram_base, size); - if (!bios) - return false; adev->bios = kmalloc(size, GFP_KERNEL); - if (!adev->bios) { - iounmap(bios); + if (!adev->bios) return false; + + /* For SRIOV with dynamic critical region is enabled, + * the vbios image is put at a dynamic offset of VRAM in the VF. + * If dynamic critical region is disabled, follow the existing logic as on baremetal. + */ + if (amdgpu_sriov_vf(adev) && adev->virt.is_dynamic_crit_regn_enabled) { + if (amdgpu_virt_get_dynamic_data_info(adev, + AMD_SRIOV_MSG_VBIOS_IMG_TABLE_ID, adev->bios, &size)) { + amdgpu_bios_release(adev); + return false; + } + } else { + bios = ioremap_wc(vram_base, size); + if (!bios) { + amdgpu_bios_release(adev); + return false; + } + + memcpy_fromio(adev->bios, bios, size); + iounmap(bios); } + adev->bios_size = size; - memcpy_fromio(adev->bios, bios, size); - iounmap(bios); if (!check_atom_bios(adev, size)) { amdgpu_bios_release(adev); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c index 4e75334f3b3a..1809deb86797 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c @@ -303,14 +303,29 @@ static int amdgpu_discovery_read_binary_from_mem(struct amdgpu_device *adev, * then it is not required to be reserved. */ if (sz_valid) { - uint64_t pos = vram_size - DISCOVERY_TMR_OFFSET; - amdgpu_device_vram_access(adev, pos, (uint32_t *)binary, - adev->discovery.size, false); - adev->discovery.reserve_tmr = true; + if (amdgpu_sriov_vf(adev) && adev->virt.is_dynamic_crit_regn_enabled) { + /* For SRIOV VFs with dynamic critical region enabled, + * we will get the IPD binary via below call. + * If dynamic critical is disabled, fall through to normal seq. + */ + if (amdgpu_virt_get_dynamic_data_info(adev, + AMD_SRIOV_MSG_IPD_TABLE_ID, binary, + (uint64_t *)&adev->mman.discovery_tmr_size)) {I think this is adev->discovery.size now after Lijo's latest changes. @Lazar, Lijo I think we can remove adev->mman.discovery_tmr_size. It's no longer used by anything. With that fixed, Reviewed-by: Alex Deucher <[email protected]>
Will do. Other than that, and one comment below, this patch looks good to me as well.
>
+ ret = -EINVAL; + goto exit;
The message printed at exit: label is not valid for this case. It's better to have another error message for this case.
Thanks, Lijo
+ } + } else { + uint64_t pos = vram_size - DISCOVERY_TMR_OFFSET; + + amdgpu_device_vram_access(adev, pos, (uint32_t *)binary, + adev->discovery.size, false); + adev->discovery.reserve_tmr = true; + } } else { ret = amdgpu_discovery_read_binary_from_sysmem(adev, binary); } +exit: if (ret) dev_err(adev->dev, "failed to read discovery info from memory, vram size read: %llx", diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c index 12659990abe0..15157ed5df29 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c @@ -999,6 +999,14 @@ int amdgpu_virt_init_critical_region(struct amdgpu_device *adev) goto out; } + /* Validation for critical region info */ + if (adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_IPD_TABLE_ID].size_kb > DISCOVERY_TMR_SIZE) { + dev_err(adev->dev, "Invalid IP discovery size: 0x%x\n", + adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_IPD_TABLE_ID].size_kb); + r = -EINVAL; + goto out; + } + /* reserved memory starts from crit region base offset with the size of 5MB */ adev->mman.fw_vram_usage_start_offset = adev->virt.crit_regn.offset; adev->mman.fw_vram_usage_size = adev->virt.crit_regn.size_kb << 10; @@ -1017,6 +1025,35 @@ int amdgpu_virt_init_critical_region(struct amdgpu_device *adev) return r; } +int amdgpu_virt_get_dynamic_data_info(struct amdgpu_device *adev, + int data_id, uint8_t *binary, uint64_t *size) +{ + uint32_t data_offset = 0; + uint32_t data_size = 0; + enum amd_sriov_msg_table_id_enum data_table_id = data_id; + + if (data_table_id >= AMD_SRIOV_MSG_MAX_TABLE_ID) + return -EINVAL; + + data_offset = adev->virt.crit_regn_tbl[data_table_id].offset; + data_size = adev->virt.crit_regn_tbl[data_table_id].size_kb << 10; + + /* Validate on input params */ + if (!binary || !size || *size < (uint64_t)data_size) + return -EINVAL; + + /* Proceed to copy the dynamic content */ + amdgpu_device_vram_access(adev, + (uint64_t)data_offset, (uint32_t *)binary, data_size, false); + *size = (uint64_t)data_size; // update the size as out param. + + dev_dbg(adev->dev, + "Got %s info from dynamic crit_region_table at offset 0x%x with size of 0x%x bytes.\n", + amdgpu_virt_dynamic_crit_table_name[data_id], data_offset, data_size); + + return 0; +} + void amdgpu_virt_init(struct amdgpu_device *adev) { bool is_sriov = false; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h index 8d03a8620de9..2a13cc892a13 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h @@ -442,6 +442,8 @@ void amdgpu_virt_fini_data_exchange(struct amdgpu_device *adev); void amdgpu_virt_init(struct amdgpu_device *adev); int amdgpu_virt_init_critical_region(struct amdgpu_device *adev); +int amdgpu_virt_get_dynamic_data_info(struct amdgpu_device *adev, + int data_id, uint8_t *binary, uint64_t *size); bool amdgpu_virt_can_access_debugfs(struct amdgpu_device *adev); int amdgpu_virt_enable_access_debugfs(struct amdgpu_device *adev); -- 2.34.1
