From: Sasha Levin <sashal@kernel.org>
To: stable@vger.kernel.org
Cc: chong li <chongli2@amd.com>,
	Alex Deucher <alexander.deucher@amd.com>,
	Sasha Levin <sashal@kernel.org>
Subject: [PATCH 7.2.y 1/2] drm/amdgpu: reduce early full GPU access during SR-IOV init
Date: Tue, 22 Sep 2026 15:33:59 -0400
Message-ID: <20260922193400.4156340-1-sashal@kernel.org>
In-Reply-To: <2026092210-sprinkled-coherent-e132@gregkh>
References: <2026092210-sprinkled-coherent-e132@gregkh>
Precedence: bulk
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit

From: chong li <chongli2@amd.com>

[ Upstream commit cba4928cdffaa0f9012acff0ec4f896320107077 ]

Allow early FB reads to fall back to BAR0 when the VRAM aperture is not ready.
This lets SR-IOV VFs consume host-provided init data before requesting full GPU access.

For ASICs that support request_init_data,
defer full GPU access until after non-GPU early init to shorten the full-access window.

Legacy ASICs(before NV12) do not send request_init_data;
the host dumps init data only during full GPU access,
so keep the original early full-access request path for them.

Signed-off-by: chong li <chongli2@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Stable-dep-of: 04de4007d323 ("drm/amdgpu: Fix GPU PCIe link capability reporting")
Signed-off-by: Sasha Levin <sashal@kernel.org>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 88 ++++++++++++++++++++--
 drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c      |  7 +-
 2 files changed, 88 insertions(+), 7 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index ddfc51a92d79d..5934fe1b295b3 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -750,6 +750,64 @@ void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos,
 	drm_dev_exit(idx);
 }
 
+#ifdef CONFIG_64BIT
+/*
+ * During early SR-IOV VF init, host-provided init data can live in FB before
+ * the normal VRAM aperture mapping is ready.  Use a temporary BAR0 mapping for
+ * reads only, and verify it matches the VRAM aperture when aperture information
+ * is already available.
+ */
+static int amdgpu_device_read_fb_via_bar0(struct amdgpu_device *adev,
+					  u64 offset, void *buf, size_t size)
+{
+	resource_size_t aper_base, aper_size, bar_start, bar_size, map_base;
+	void __iomem *vram;
+	size_t map_offset, map_size;
+	unsigned long flags;
+	u64 end;
+
+	if (!buf || !size)
+		return -EINVAL;
+
+	flags = pci_resource_flags(adev->pdev, 0);
+	if ((flags & IORESOURCE_UNSET) || !(flags & IORESOURCE_MEM))
+		return -EINVAL;
+
+	bar_start = pci_resource_start(adev->pdev, 0);
+	bar_size = pci_resource_len(adev->pdev, 0);
+	if (!bar_size)
+		return -ENODEV;
+
+	aper_base = adev->gmc.aper_base;
+	aper_size = adev->gmc.visible_vram_size ? adev->gmc.visible_vram_size :
+						  adev->gmc.aper_size;
+
+	if (aper_base || aper_size) {
+		if (aper_base != bar_start || aper_size > bar_size)
+			return -EINVAL;
+	} else {
+		aper_base = bar_start;
+		aper_size = bar_size;
+	}
+
+	if (check_add_overflow(offset, size, &end) || end > aper_size)
+		return -EINVAL;
+
+	map_offset = offset_in_page(offset);
+	map_base = aper_base + (offset & PAGE_MASK);
+	map_size = PAGE_ALIGN(map_offset + size);
+
+	vram = ioremap_wc(map_base, map_size);
+	if (!vram)
+		return -ENOMEM;
+
+	memcpy_fromio(buf, (u8 __iomem *)vram + map_offset, size);
+	iounmap(vram);
+
+	return 0;
+}
+#endif
+
 /**
  * amdgpu_device_aper_access - access vram by vram aperture
  *
@@ -769,8 +827,12 @@ size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos,
 	size_t count = 0;
 	uint64_t last;
 
-	if (!adev->mman.aper_base_kaddr)
+	if (!adev->mman.aper_base_kaddr) {
+		/* Writes still require the regular aperture/MM path. */
+		if (!write && !amdgpu_device_read_fb_via_bar0(adev, pos, buf, size))
+			return size;
 		return 0;
+	}
 
 	last = min(pos + size, adev->gmc.visible_vram_size);
 	if (last > pos) {
@@ -2011,16 +2073,25 @@ static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
 {
 	struct amdgpu_ip_block *ip_block;
 	struct pci_dev *parent;
-	bool total, skip_bios;
+	bool total, skip_bios, early_full_gpu_access = false;
 	uint32_t bios_flags;
 	int i, r;
 
 	amdgpu_device_enable_virtual_display(adev);
 
 	if (amdgpu_sriov_vf(adev)) {
-		r = amdgpu_virt_request_full_gpu(adev, true);
-		if (r)
-			return r;
+		/*
+		 * Legacy hosts do not provide init data before early init, so
+		 * keep the original early full GPU access request for them.  Newer
+		 * hosts publish the init data through VF FB, which lets us defer
+		 * full GPU access until after non-GPU early init work is done.
+		 */
+		early_full_gpu_access = (adev->virt.req_init_data_ver == 0);
+		if (early_full_gpu_access) {
+			r = amdgpu_virt_request_full_gpu(adev, true);
+			if (r)
+				return r;
+		}
 
 		r = amdgpu_virt_init_critical_region(adev);
 		if (r)
@@ -2183,6 +2254,13 @@ static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
 	if (!total)
 		return -ENODEV;
 
+	/* Request full GPU access only for the remaining SR-IOV init work. */
+	if (amdgpu_sriov_vf(adev) && !early_full_gpu_access) {
+		r = amdgpu_virt_request_full_gpu(adev, true);
+		if (r)
+			return r;
+	}
+
 	if (adev->gmc.xgmi.supported)
 		amdgpu_xgmi_early_init(adev);
 
diff --git a/drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c b/drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c
index 9a40107a0869d..fec08f235b920 100644
--- a/drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c
+++ b/drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c
@@ -185,8 +185,11 @@ static int xgpu_ai_send_access_requests(struct amdgpu_device *adev,
 	} else if (req == IDH_REQ_GPU_INIT_DATA){
 		/* Dummy REQ_GPU_INIT_DATA handling */
 		r = xgpu_ai_poll_msg(adev, IDH_REQ_GPU_INIT_DATA_READY);
-		/* version set to 0 since dummy */
-		adev->virt.req_init_data_ver = 0;	
+		/*
+		 * AI uses the GPU_CRIT_REGION_V1 layout in practice, so fix the
+		 * dummy version value to match the actual init-data format.
+		 */
+		adev->virt.req_init_data_ver = GPU_CRIT_REGION_V1;
 	}
 
 	return 0;
-- 
2.53.0