[PATCH] drm/amdgpu: support reserve bad page for virt

Zhang, Hawking Hawking.Zhang at amd.com
Wed Jun 3 15:26:27 UTC 2020


[AMD Public Use]

Please check my comments inline.

Regards,
Hawking

-----Original Message-----
From: Stanley.Yang <Stanley.Yang at amd.com> 
Sent: Wednesday, June 3, 2020 22:10
To: amd-gfx at lists.freedesktop.org
Cc: Zhang, Hawking <Hawking.Zhang at amd.com>; Chen, Guchun <Guchun.Chen at amd.com>; Liu, Monk <Monk.Liu at amd.com>; Clements, John <John.Clements at amd.com>; Zhou1, Tao <Tao.Zhou1 at amd.com>; Li, Dennis <Dennis.Li at amd.com>; Yang, Stanley <Stanley.Yang at amd.com>
Subject: [PATCH] drm/amdgpu: support reserve bad page for virt

Signed-off-by: Stanley.Yang <Stanley.Yang at amd.com>
Change-Id: Ia0ad9453ac3ac929f95c73cbee5b7a8fc42a9816
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |   7 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c   | 164 +++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h   |  30 +++-
 3 files changed, 196 insertions(+), 5 deletions(-)  mode change 100644 => 100755 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c
 mode change 100644 => 100755 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index b633171281f8..e8986e007206 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -2001,8 +2001,10 @@ static int amdgpu_device_ip_init(struct amdgpu_device *adev)
 		}
 	}
 
-	if (amdgpu_sriov_vf(adev))
+	if (amdgpu_sriov_vf(adev)) {
+		amdgpu_virt_init_err_handler_data(adev);

[Hawking]: 
1. It might be better to rename the function to amdgpu_virt_init_ras_err_handler_data
2. we shall apply either asic type check or more general data structure check before call into amdgpu_virt_init_ras_err_handler_data. Please check my comments in amdgpu_virt_init_ras_err_handler_data function


 		amdgpu_virt_init_data_exchange(adev);
+	}
 
 	r = amdgpu_ib_pool_init(adev);
 	if (r) {
@@ -2306,6 +2308,9 @@ static int amdgpu_device_ip_fini(struct amdgpu_device *adev)  {
 	int i, r;
 
+	if (amdgpu_sriov_vf(adev))
+		amdgpu_release_virt_err_handler_data(adev);

[Hawking]: 
1. It might be better to rename the function to amdgpu_virt_release_ras_err_handler_data
2. we shall apply either asic type check or more general data structure check In case it introduce regression on MI100 and V340L

+
 	amdgpu_ras_pre_fini(adev);
 
 	if (adev->gmc.xgmi.num_physical_nodes > 1) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c
old mode 100644
new mode 100755
index f3b38c9e04ca..c1554562a2ce
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c
@@ -26,6 +26,7 @@
 #include <drm/drm_drv.h>
 
 #include "amdgpu.h"
+#include "amdgpu_ras.h"
 
 bool amdgpu_virt_mmio_blocked(struct amdgpu_device *adev)  { @@ -255,12 +256,164 @@ int amdgpu_virt_fw_reserve_get_checksum(void *obj,
 	return ret;
 }
 
+int amdgpu_virt_init_err_handler_data(struct amdgpu_device *adev) {

[Hawking]: 
1. same as above, rename it to amdgpu_virt_init_ras_err_handler_data

+	struct amdgpu_virt *virt = &adev->virt;
+	struct virt_ras_err_handler_data **data = &virt->virt_eh_data;
+	/* GPU will be marked bad on host if bp count more then 10,
+	 * so alloc 512 is enough.
+	 */
+	unsigned int align_space = 512;
+	void *bps = NULL;
+	struct amdgpu_bo **bps_bo = NULL;
+

[Hawking]:
1. This will result to regression on MI100 SRIOV if we don't apply check. Or leverage general data structure for the purpose, thoughts?

+	*data = kmalloc(sizeof(struct virt_ras_err_handler_data), GFP_KERNEL);

[Hawking]:
1. rename the data structure to amdgpu_virt_ras_err_handler_data

+	if (!*data)
+		return -ENOMEM;
+
+	bps = kmalloc(align_space * sizeof((*data)->bps), GFP_KERNEL);
+	bps_bo = kmalloc(align_space * sizeof((*data)->bps_bo), GFP_KERNEL);
+
+	if (!bps || !bps_bo) {
+		kfree(bps);
+		kfree(bps_bo);
+		return -ENOMEM;
+	}
+
+	(*data)->bps = bps;
+	(*data)->bps_bo = bps_bo;
+	(*data)->count = 0;
+	(*data)->last_reserved = 0;
+	return 0;
+}
+
+static void amdgpu_virt_release_bp(struct amdgpu_device *adev) {
+	struct amdgpu_virt *virt = &adev->virt;
+	struct virt_ras_err_handler_data *data = virt->virt_eh_data;

[Hawking]:
1. same as above, please rename the virt_ras_err_handler_data to amdgpu_virt_ras_err_handler_data

+	struct amdgpu_bo *bo;
+	int i;
+
+	if (!data)
+		return;
+
+	for (i = data->last_reserved - 1; i >= 0; i--) {
+		bo = data->bps_bo[i];
+		amdgpu_bo_free_kernel(&bo, NULL, NULL);
+		data->bps_bo[i] = bo;
+		data->last_reserved = i;
+	}
+}
+
+void amdgpu_release_virt_err_handler_data(struct amdgpu_device *adev) {

[Hawking]:
1. same as above, rename the function to amdgpu_virt_release_ras_err_handler_data

+	struct amdgpu_virt *virt = &adev->virt;
+	struct virt_ras_err_handler_data *data = virt->virt_eh_data;
+
+	if (!data)
+		return;
+
+	amdgpu_virt_release_bp(adev);
+
+	kfree(data->bps);
+	kfree(data->bps_bo);
+	kfree(data);
+	virt->virt_eh_data = NULL;
+}
+
+static void amdgpu_virt_ras_add_bps(struct amdgpu_device *adev,
+		struct eeprom_table_record *bps, int pages) {
+	struct amdgpu_virt *virt = &adev->virt;
+	struct virt_ras_err_handler_data *data = virt->virt_eh_data;
+
+	if (!data)
+		return;
+
+	memcpy(&data->bps[data->count], bps, pages * sizeof(*data->bps));
+	data->count += pages;
+}
+
+static void amdgpu_virt_ras_reserve_bps(struct amdgpu_device *adev) {
+	struct amdgpu_virt *virt = &adev->virt;
+	struct virt_ras_err_handler_data *data = virt->virt_eh_data;
+	struct amdgpu_bo *bo = NULL;
+	uint64_t bp;
+	int i;
+
+	if (!data)
+		return;
+
+	for (i = data->last_reserved; i < data->count; i++) {
+		bp = data->bps[i].retired_page;
+
+		/* There are two cases of reserve error should be ignored:
+		 * 1) a ras bad page has been allocated (used by someone);
+		 * 2) a ras bad page has been reserved (duplicate error injection
+		 *    for one page);
+		 */
+		if (amdgpu_bo_create_kernel_at(adev, bp << AMDGPU_GPU_PAGE_SHIFT,
+					       AMDGPU_GPU_PAGE_SIZE,
+					       AMDGPU_GEM_DOMAIN_VRAM,
+					       &bo, NULL))
+			DRM_WARN("RAS WARN: reserve vram for retired page %llx fail\n", bp);

[Hawking]:
1. can we make this as debug message?

+
+		data->bps_bo[i] = bo;
+		data->last_reserved = i + 1;
+		bo = NULL;
+	}
+}
+
+static bool amdgpu_virt_check_bad_page(struct amdgpu_device *adev,
+		uint64_t retired_page)
+{
+	struct amdgpu_virt *virt = &adev->virt;
+	struct virt_ras_err_handler_data *data = virt->virt_eh_data;
+	int i;
+
+	if (!data)
+		return true;
+
+	for (i = 0; i < data->count; i++)
+		if (retired_page == data->bps[i].retired_page)
+			return true;
+
+	return false;
+}
+
+static void amdgpu_virt_add_bad_page(struct amdgpu_device *adev,
+		uint64_t bp_block_offset, uint32_t bp_block_size) {
+	struct eeprom_table_record bp;
+	uint64_t retired_page;
+	uint32_t bp_idx, bp_cnt;
+
+	if (bp_block_size) {
+		bp_cnt = bp_block_size / sizeof(uint64_t);
+		for (bp_idx = 0; bp_idx < bp_cnt; bp_idx++) {
+			retired_page = *(uint64_t *)(adev->fw_vram_usage.va +
+					bp_block_offset + bp_idx * sizeof(uint64_t));
+			bp.retired_page = retired_page;
+
+			if (amdgpu_virt_check_bad_page(adev, retired_page))
+				continue;
+
+			amdgpu_virt_ras_add_bps(adev, &bp, 1);
+
+			amdgpu_virt_ras_reserve_bps(adev);
+		}
+	}
+}
+
 void amdgpu_virt_init_data_exchange(struct amdgpu_device *adev)  {
 	uint32_t pf2vf_size = 0;
 	uint32_t checksum = 0;
 	uint32_t checkval;
 	char *str;
+	uint64_t bp_block_offset = 0;
+	uint32_t bp_block_size = 0;
 
 	adev->virt.fw_reserve.p_pf2vf = NULL;
 	adev->virt.fw_reserve.p_vf2pf = NULL;
@@ -275,6 +428,17 @@ void amdgpu_virt_init_data_exchange(struct amdgpu_device *adev)
 
 		/* pf2vf message must be in 4K */
 		if (pf2vf_size > 0 && pf2vf_size < 4096) {
+			if (adev->virt.fw_reserve.p_pf2vf->version == 2) {
+				struct amdgim_pf2vf_info_v2 *pf2vf_v2 = NULL;
+
+				pf2vf_v2 = (struct amdgim_pf2vf_info_v2 *)adev->virt.fw_reserve.p_pf2vf;
+				bp_block_offset = ((uint64_t)pf2vf_v2->bp_block_offset_L & 0xFFFFFFFF) |
+						((((uint64_t)pf2vf_v2->bp_block_offset_H) << 32) & 0xFFFFFFFF00000000);
+				bp_block_size = pf2vf_v2->bp_block_size;
+
+				amdgpu_virt_add_bad_page(adev, bp_block_offset, bp_block_size);
+			}
+
 			checkval = amdgpu_virt_fw_reserve_get_checksum(
 				adev->virt.fw_reserve.p_pf2vf, pf2vf_size,
 				adev->virt.fw_reserve.checksum_key, checksum); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h
old mode 100644
new mode 100755
index b90e822cebd7..96d84f036e96
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h
@@ -143,19 +143,27 @@ struct  amdgim_pf2vf_info_v2 {
 	uint32_t vce_enc_max_pixels_count;
 	/* 16x16 pixels/sec, codec independent */
 	uint32_t vce_enc_max_bandwidth;
+	/* Bad pages block position in BYTE */
+	uint32_t bp_block_offset_L;
+	uint32_t bp_block_offset_H;
+	/* Bad pages block size in BYTE */
+	uint32_t bp_block_size;
 	/* MEC FW position in kb from the start of VF visible frame buffer */
-	uint64_t mecfw_kboffset;
+	uint32_t mecfw_kboffset_L;
+	uint32_t mecfw_kboffset_H;

[Hawking]:
1. can you just explain the reason that split the position in lower 32bit and higher 32 bit respectively? 

 	/* MEC FW size in KB */
 	uint32_t mecfw_ksize;
 	/* UVD FW position in kb from the start of VF visible frame buffer */
-	uint64_t uvdfw_kboffset;
+	uint32_t uvdfw_kboffset_L;
+	uint32_t uvdfw_kboffset_H;
 	/* UVD FW size in KB */
 	uint32_t uvdfw_ksize;
 	/* VCE FW position in kb from the start of VF visible frame buffer */
-	uint64_t vcefw_kboffset;
+	uint32_t vcefw_kboffset_L;
+	uint32_t vcefw_kboffset_H;
 	/* VCE FW size in KB */
 	uint32_t vcefw_ksize;
-	uint32_t reserved[AMDGIM_GET_STRUCTURE_RESERVED_SIZE(256, 0, 0, (9 + sizeof(struct amd_sriov_msg_pf2vf_info_header)/sizeof(uint32_t)), 3)];
+	uint32_t reserved[AMDGIM_GET_STRUCTURE_RESERVED_SIZE(256, 0, 0, (18 + 
+sizeof(struct amd_sriov_msg_pf2vf_info_header)/sizeof(uint32_t)), 0)];
 } __aligned(4);
 
 
@@ -254,6 +262,17 @@ typedef struct amdgim_vf2pf_info_v2 amdgim_vf2pf_info ;
 		} \
 	} while (0)
 
+struct virt_ras_err_handler_data {
+	/* point to bad page records array */
+	struct eeprom_table_record *bps;
+	/* point to reserved bo array */
+	struct amdgpu_bo **bps_bo;
+	/* the count of entries */
+	int count;
+	/* last reserved entry's index + 1 */
+	int last_reserved;
+};
+
 /* GPU virtualization */
 struct amdgpu_virt {
 	uint32_t			caps;
@@ -272,6 +291,7 @@ struct amdgpu_virt {
 	uint32_t reg_access_mode;
 	int req_init_data_ver;
 	bool tdr_debug;
+	struct virt_ras_err_handler_data *virt_eh_data;
 };
 
 #define amdgpu_sriov_enabled(adev) \
@@ -323,6 +343,8 @@ void amdgpu_virt_free_mm_table(struct amdgpu_device *adev);  int amdgpu_virt_fw_reserve_get_checksum(void *obj, unsigned long obj_size,
 					unsigned int key,
 					unsigned int chksum);
+void amdgpu_release_virt_err_handler_data(struct amdgpu_device *adev); 
+int amdgpu_virt_init_err_handler_data(struct amdgpu_device *adev);
 void amdgpu_virt_init_data_exchange(struct amdgpu_device *adev);  void amdgpu_detect_virtualization(struct amdgpu_device *adev);
 
--
2.25.1


More information about the amd-gfx mailing list