Re: [PATCH v4 07/10] tests/intel/xe_err_injection: Add GT UC Unicast GAM Walker Command Parity Error Injection

["Purkait, Soham" <soham.purkait@intel.com>]

Hi Ravi,

On 03-09-2026 13:03, Ravi Kishore Koppuravuri wrote:
> MMIO based GT Uncorrectable Unicast GAM Walker command parity error
> injection to verify the Xe driver error handling and recovery flows with
> the help of DRM Netlink API suite
>
> Cc: Purkait Soham <soham.purkait@intel.com>
> Cc: Riana Tauro <riana.tauro@intel.com>
> Cc: Gupta Anshuman <anshuman.gupta@intel.com>
> Cc: Jadav Raag <raag.jadav@intel.com>
> Cc: Nilawar Badal <badal.nilawar@intel.com>
> Cc: Koujalagi Mallesh <mallesh.koujalagi@intel.com>
> Cc: Harish Chegondi <harish.chegondi@intel.com>
> Signed-off-by: Ravi Kishore Koppuravuri <ravi.kishore.koppuravuri@intel.com>
> ---
> v4:
>    - Added Steer Semaphore for disabling and enabling multicast
>      operation(soham)
>    - Increased device recovery check timeout from 5secs to 10secs
>    - Used updated library names with igt_drm_netlink_* prefix (Nikula
>      Jani & Kamil)
> ---
>   tests/intel/xe_err_injection.c | 342 +++++++++++++++++++++++++++++++++
>   tests/intel/xe_err_injection.h |  31 +++
>   tests/meson.build              |   1 +
>   3 files changed, 374 insertions(+)
>   create mode 100644 tests/intel/xe_err_injection.c
>   create mode 100644 tests/intel/xe_err_injection.h
>
> diff --git a/tests/intel/xe_err_injection.c b/tests/intel/xe_err_injection.c
> new file mode 100644
> index 000000000..683994d1c
> --- /dev/null
> +++ b/tests/intel/xe_err_injection.c
> @@ -0,0 +1,342 @@
> +// SPDX-License-Identifier: MIT
> +/*
> + * Copyright © 2026 Intel Corporation
> + */
> +
> +/**
> + * TEST: MMIO based Error Injection
> + * Category: RAS
> + * Mega feature: Telemetry
> + * Sub-category: Driver
> + * Test category: Error Injection
> + */
> +
> +#include <stdio.h>
> +#include <stdlib.h>
> +#include <string.h>
> +#include <time.h>
> +#include <fcntl.h>
> +#include <unistd.h>
> +
> +#include "igt.h"
> +#include "lib/igt_drm_netlink.h"
> +#include "lib/intel_reg.h"
> +#include "lib/intel_compute.h"
> +
> +#include "xe_drm.h"
> +#include "xe/xe_ioctl.h"
> +#include "xe/xe_mmio.h"
> +#include "xe/xe_query.h"
> +#include "xe_err_injection.h"
> +
> +enum {
> +	RECOVERY_SUCCESS = 1,
> +	RECOVERY_FAILED,
> +	RECOVERY_TIMEOUT,
> +};
> +
> +static void run_xe_compute_on_all_engines(int fd, bool state)
> +{
> +	struct drm_xe_engine_class_instance *hwe;
> +
> +	if (state == POST_ARMING_INJECTION_WL)
> +		igt_info("Running compute-square on all engines post injection\n");
> +	else if (state == POST_RECOVERY_WL)
> +		igt_info("Running compute-square on all engines post recovery\n");
> +
> +	xe_for_each_engine(fd, hwe) {
> +		if (hwe->engine_class != DRM_XE_ENGINE_CLASS_COMPUTE)
> +			continue;
> +
> +		igt_require_f(xe_run_intel_compute_kernel_on_engine(fd, hwe, NULL,
> +								    EXECENV_PREF_SYSTEM),
> +								    "GPU does not support "
> +								    "compute on engine\n");
> +	}
> +}
> +
> +static void write_reg(struct xe_mmio *mmio, uint32_t reg, uint32_t value)
> +{
> +	xe_mmio_write32(mmio, reg, value);
> +}
> +
> +static void modify_reg_bit(struct xe_mmio *mmio, uint32_t reg_addr,
> +			   uint32_t bit_mask, bool set)
> +{
> +	uint32_t regval;
> +
> +	regval = xe_mmio_read32(mmio, reg_addr);
> +	if (set)
> +		regval |= bit_mask;
> +	else
> +		regval &= ~bit_mask;
> +
> +	xe_mmio_write32(mmio, reg_addr, regval);
> +}
> +
> +static void acquire_steering_semaphore(struct xe_mmio *mmio)
> +{
> +	uint32_t regval;
> +
> +	while (true) {
> +		regval = xe_mmio_read32(mmio, STEER_SEMAPHORE_MGSR_3D);
> +		if (regval == STEER_SEMAPHORE_MGSR_3D_VALUE) {
> +			xe_mmio_write32(mmio, STEER_SEMAPHORE_MGSR_3D, 0x0);
> +			return;
> +		}
> +	}
> +}
> +
> +static void release_steering_semaphore(struct xe_mmio *mmio)
> +{
> +	xe_mmio_write32(mmio, STEER_SEMAPHORE_MGSR_3D,
> +			STEER_SEMAPHORE_MGSR_3D_VALUE);
> +}
> +
> +static int acquire_forcewake(int fd)
> +{
> +	int fw_handle;
> +
> +	fw_handle = igt_debugfs_open(fd, "forcewake_all", O_RDONLY);
> +	igt_assert_lte(0, fw_handle);
> +	return fw_handle;
> +}
> +
> +static void release_forcewake(int fw_handle)
> +{
> +	if (fw_handle >= 0)
> +		close(fw_handle);
> +}
> +
> +static uint32_t get_counter(uint32_t node_id, uint32_t error_id)
> +{
> +	struct app_context ctx;
> +	int ret;
> +	uint32_t error_value = UINT32_MAX;
> +
> +	ret = igt_drm_netlink_init(&ctx);
> +	if (ret < 0) {
> +		igt_warn("Failed to initialize netlink socket for error command (ret=%d)\n", ret);
> +		return error_value;
> +	}
> +
> +	ctx.node_id = node_id;
> +	ctx.error_id = error_id;
> +
> +	ret = igt_drm_netlink_get_error_counter(&ctx);
> +	if (ret < 0) {
> +		igt_warn("get_error_counter failed from error command path (ret=%d)\n", ret);
> +	} else {
> +		error_value = ctx.error_value;
> +		igt_info("get_error_counter: node_id=%u error_id=%u value=%u\n",
> +			 ctx.node_id, ctx.error_id, ctx.error_value);
> +	}
> +
> +	igt_drm_netlink_cleanup(&ctx);
> +
> +	return error_value;
> +}
> +
> +static int check_dmesg(const char *marker, int elapsed_secs)
> +{
> +	static const char * const recovery_messages[] = {
> +		"AER: device recovery successful",
> +		"AER: device recovery failed",
> +	};

Make this as static global and keep the name generic, something like 
"dmesg_status". Better approach would be like :

static const char * const dmesg_status[] = {
     [RECOVERY_SUCCESS] = "AER: device recovery successful",
     [RECOVERY_FAILED] = "AER: device recovery failed",
};

To be on the safe side, make sure these strings remain unchanged in the 
future, since they are used to detect the corresponding status in dmesg.

thanks,
Soham

> +	char *buff = NULL;
> +	size_t buff_size = 0;
> +	size_t i;
> +	ssize_t line_len;
> +	FILE *fp;
> +	bool marker_seen = false;
> +
> +	fp = popen("dmesg", "r");
> +	if (!fp) {
> +		igt_warn("Unable to open dmesg to check recovery status\n");
> +		return -1;
> +	}
> +
> +	while ((line_len = getline(&buff, &buff_size, fp)) != -1) {
> +		if (!marker_seen) {
> +			if (strstr(buff, marker))
> +				marker_seen = true;
> +			continue;
> +		}
> +
> +		for (i = 0; i < ARRAY_SIZE(recovery_messages); i++) {
> +			if (!strstr(buff, recovery_messages[i]))
> +				continue;
> +
> +			igt_debug("Found \"%s\" in dmesg after %d secs\n",
> +				  recovery_messages[i], elapsed_secs);
> +			free(buff);
> +			pclose(fp);
> +			return RECOVERY_SUCCESS + i;
> +		}
> +	}
> +
> +	free(buff);
> +	pclose(fp);
> +	return RECOVERY_TIMEOUT;
> +}
> +
> +static int check_err_recovery(void)
> +{
> +	time_t start_time = time(NULL);
> +	time_t timeout = 10 * 60; /* 10 secs */
> +	int time_interval = 30;
> +	time_t elapsed_time;
> +	int status;
> +	char marker[128];
> +
> +	snprintf(marker, sizeof(marker),
> +		 "IGT xe_err_injection recovery marker pid=%d start=%lld",
> +		 getpid(), (long long)start_time);
> +	igt_kmsg(KMSG_INFO "%s\n", marker);
> +
> +	while (1) {
> +		elapsed_time = time(NULL) - start_time;
> +		status = check_dmesg(marker, elapsed_time);
> +		if (status < 0) {
> +			igt_warn("Failed to query dmesg for recovery status\n");
> +			return RECOVERY_FAILED;
> +		}
> +		if (status == RECOVERY_SUCCESS || status == RECOVERY_FAILED)
> +			return status;
> +
> +		if (elapsed_time >= timeout) {
> +			igt_warn("Timed out while waiting for error recovery\n");
> +			return RECOVERY_TIMEOUT;
> +		}
> +
> +		sleep(time_interval);
> +	}
> +}
> +
> +static void print_aer_recovery_status(void)
> +{
> +	int recovery_ret;
> +
> +	recovery_ret = check_err_recovery();
> +	igt_assert_f(recovery_ret != RECOVERY_TIMEOUT,
> +		     "AER error recovery timed out\n");
> +	igt_assert_f(recovery_ret != RECOVERY_FAILED,
> +		     "AER device recovery failed\n");
> +	if (recovery_ret == RECOVERY_SUCCESS)
> +		igt_info("AER device recovery successful\n");
> +}
> +
> +static void log_status(bool val, const char *err_name)
> +{
> +	if (val)
> +		igt_info("Injection: %s Success\n", err_name);
> +	else
> +		igt_info("Injection: %s Failed\n", err_name);
> +}
> +
> +static bool gt_uc_wkr_parity_recovered;
> +
> +/**
> + * SUBTEST: test_GT-UC-unicast-wkr-cmd-parity-err
> + * Description: GT Uncorrectable Unicast Walker Command parity error injection
> + * Functionality: error injection
> + */
> +static void wkr_cmd_parity_err_injection(struct xe_mmio *mmio, int fd)
> +{
> +	int fw_handle;
> +	uint32_t error_counter_before_inj;
> +	uint32_t error_counter_after_inj;
> +	/** TODO
> +	 *  Fetching node_id and error_id dynamically is pending to implement.
> +	 *  For now, using
> +	 *  node_id=1 (uncorrectable_errors) and
> +	 *  error_id=1 (core_compute)
> +	 */
> +	uint32_t node_id = NODE_ID_UNCORRECTABLE_ERROR;
> +	uint32_t error_id = ERROR_ID_CORE_COMPUTE;
> +
> +	fw_handle = acquire_forcewake(fd);
> +
> +	error_counter_before_inj = get_counter(node_id, error_id);
> +
> +	acquire_steering_semaphore(mmio);
> +	/* Arm the injection sequence. */
> +	write_reg(mmio, MC_PKT_CTRL_MGSR_3D_ADDRESS, 0x0);
> +	modify_reg_bit(mmio,
> +		       WKR_FABRIC_ERR_INJ_GAMWALK_3D_ADDRESS,
> +		       WKR_FABRIC_ERR_INJ_GAMWALK_3D_VALUE,
> +		       true);
> +	modify_reg_bit(mmio,
> +		       MC_PKT_CTRL_MGSR_3D_ADDRESS,
> +		       MC_PKT_CTRL_MGSR_3D_VALUE,
> +		       true);
> +	release_steering_semaphore(mmio);
> +	igt_info("Injected GT Uncorrectable Unicast Walker Cmd parity error\n");
> +
> +	run_xe_compute_on_all_engines(fd, POST_ARMING_INJECTION_WL);
> +
> +	release_forcewake(fw_handle);
> +	print_aer_recovery_status();
> +	error_counter_after_inj = get_counter(node_id, error_id);
> +
> +	igt_assert_f(error_counter_before_inj != UINT32_MAX &&
> +		     error_counter_after_inj != UINT32_MAX,
> +		     "Failed to fetch valid error counters: before=%u after=%u\n",
> +		     error_counter_before_inj, error_counter_after_inj);
> +
> +	igt_info("error counter: before injection=%u after injection=%u\n",
> +		 error_counter_before_inj, error_counter_after_inj);
> +
> +	igt_assert_f(error_counter_after_inj > error_counter_before_inj,
> +		     "GT Uncorrectable Unicast Walker Cmd parity error injection "
> +		     "failed: before injection=%u after injection=%u\n",
> +		     error_counter_before_inj, error_counter_after_inj);
> +	igt_info("GT Uncorrectable Unicast Walker Cmd parity error injection successful\n");
> +	gt_uc_wkr_parity_recovered = true;
> +}
> +
> +static void inject_error(const char *injection, struct xe_mmio *mmio, int fd)
> +{
> +	igt_info("Starting Error Injection test: %s\n", injection);
> +	if (strcmp(injection, "test_GT-UC-unicast-wkr-cmd-parity-err") == 0)
> +		wkr_cmd_parity_err_injection(mmio, fd);
> +	else if (strcmp(injection, "test_GT-UC-unicast-wkr-cmd-parity-err-post-recovery-wl") == 0)
> +		run_xe_compute_on_all_engines(fd, POST_RECOVERY_WL);
> +	else
> +		igt_info("Invalid Error Injection specified\n");
> +}
> +
> +/**
> + * SUBTEST: test_GT-UC-unicast-wkr-cmd-parity-err-post-recovery-wl
> + * Description: Run a post-recovery Xe workload after parity error injection.
> + * Functionality: workload validation
> + */
> +
> +int igt_main()
> +{
> +	int fd;
> +	struct xe_mmio mmio;
> +
> +	igt_fixture() {
> +		fd = drm_open_driver(DRIVER_XE);
> +		igt_require(igt_debugfs_exists(fd, "forcewake_all", O_RDONLY));
> +		xe_mmio_access_init(fd, &mmio);
> +		igt_require(xe_mmio_is_initialized(&mmio));
> +	}
> +
> +	igt_describe("Inject GT uncorrectable unicast worker command parity error.");
> +	igt_subtest("test_GT-UC-unicast-wkr-cmd-parity-err")
> +		inject_error("test_GT-UC-unicast-wkr-cmd-parity-err", &mmio, fd);
> +
> +	igt_describe("Run post-recovery workload after GT parity error injection test.");
> +	igt_subtest("test_GT-UC-unicast-wkr-cmd-parity-err-post-recovery-wl") {
> +		igt_require_f(gt_uc_wkr_parity_recovered,
> +			      "Run GT-UC-unicast-wkr-cmd-parity-err first\n");
> +		inject_error("test_GT-UC-unicast-wkr-cmd-parity-err-post-recovery-wl", &mmio, fd);
> +	}
> +
> +	igt_fixture() {
> +		xe_mmio_access_fini(&mmio);
> +		drm_close_driver(fd);
> +	}
> +}
> diff --git a/tests/intel/xe_err_injection.h b/tests/intel/xe_err_injection.h
> new file mode 100644
> index 000000000..a70e6af22
> --- /dev/null
> +++ b/tests/intel/xe_err_injection.h
> @@ -0,0 +1,31 @@
> +/* SPDX-License-Identifier: MIT */
> +/*
> + * Copyright © 2026 Intel Corporation
> + */
> +
> +#ifndef XE_ERR_INJECTION_H
> +#define XE_ERR_INJECTION_H
> +
> +/* Steering Control Register (MC_PKT_CTRL) */
> +#define MC_PKT_CTRL_MGSR_3D_ADDRESS 0x00FD4
> +#define MC_PKT_CTRL_MGSR_3D_VALUE 0x80000000
> +
> +/* GT Uncorrectable worker command parity error injection */
> +#define WKR_FABRIC_ERR_INJ_GAMWALK_3D_ADDRESS 0xF310
> +#define WKR_FABRIC_ERR_INJ_GAMWALK_3D_VALUE 0x1
> +
> +#define POST_RECOVERY_WL 1
> +#define POST_ARMING_INJECTION_WL 0
> +
> +/* STEER_SEMAPHORE_MGSR_3D */
> +#define STEER_SEMAPHORE_MGSR_3D 0x00FD0
> +#define STEER_SEMAPHORE_MGSR_3D_VALUE 0x1
> +
> +/* Node types for error counters*/
> +#define NODE_ID_UNCORRECTABLE_ERROR 1
> +
> +/* Error types for error counter */
> +#define ERROR_ID_CORE_COMPUTE 1
> +
> +#endif /* XE_ERR_INJECTION_H */
> +
> diff --git a/tests/meson.build b/tests/meson.build
> index a62f447df..0f090dcee 100644
> --- a/tests/meson.build
> +++ b/tests/meson.build
> @@ -292,6 +292,7 @@ intel_xe_progs = [
>   	'xe_debugfs',
>   	'xe_dma_buf_sync',
>   	'xe_drm_fdinfo',
> +	'xe_err_injection',
>   	'xe_eu_stall',
>   	'xe_evict',
>   	'xe_evict_ccs',