drm/amdgpu: Add reset control handling to reset workflow
authorLijo Lazar <lijo.lazar@amd.com>
Tue, 16 Mar 2021 12:31:51 +0000 (20:31 +0800)
committerAlex Deucher <alexander.deucher@amd.com>
Fri, 9 Apr 2021 20:46:14 +0000 (16:46 -0400)
This prefers reset control based handling if it's implemented
for a particular ASIC. If not, it takes the legacy path. It uses
the legacy method of preparing environment (job, scheduler tasks)
and restoring environment.

v2: remove unused variable (Alex)

Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Feifei Xu <Feifei.Xu@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu.h
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c

index 4b72fb3..ec21a71 100644 (file)
@@ -270,6 +270,7 @@ struct amdgpu_bo_va_mapping;
 struct amdgpu_atif;
 struct kfd_vm_fault_info;
 struct amdgpu_hive_info;
+struct amdgpu_reset_context;
 struct amdgpu_reset_control;
 
 enum amdgpu_cp_irq {
@@ -1075,6 +1076,7 @@ struct amdgpu_device {
 
        bool                            in_pci_err_recovery;
        struct pci_saved_state          *pci_state;
+
        struct amdgpu_reset_control     *reset_cntl;
 };
 
@@ -1127,13 +1129,10 @@ bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type);
 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev);
 
 int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
-                                 struct amdgpu_job *job,
-                                 bool *need_full_reset_arg);
+                                struct amdgpu_reset_context *reset_context);
 
-int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
-                         struct list_head *device_list_handle,
-                         bool *need_full_reset_arg,
-                         bool skip_hw_reset);
+int amdgpu_do_asic_reset(struct list_head *device_list_handle,
+                        struct amdgpu_reset_context *reset_context);
 
 int emu_soc_asic_init(struct amdgpu_device *adev);
 
index 3cbe813..a51f470 100644 (file)
@@ -65,6 +65,7 @@
 #include "amdgpu_ras.h"
 #include "amdgpu_pmu.h"
 #include "amdgpu_fru_eeprom.h"
+#include "amdgpu_reset.h"
 
 #include <linux/suspend.h>
 #include <drm/task_barrier.h>
@@ -3421,6 +3422,8 @@ int amdgpu_device_init(struct amdgpu_device *adev,
                goto fence_driver_init;
        }
 
+       amdgpu_reset_init(adev);
+
        /* detect if we are with an SRIOV vbios */
        amdgpu_device_detect_sriov_bios(adev);
 
@@ -3671,6 +3674,9 @@ void amdgpu_device_fini(struct amdgpu_device *adev)
        release_firmware(adev->firmware.gpu_info_fw);
        adev->firmware.gpu_info_fw = NULL;
        adev->accel_working = false;
+
+       amdgpu_reset_fini(adev);
+
        /* free i2c buses */
        if (!amdgpu_device_has_dc_support(adev))
                amdgpu_i2c_fini(adev);
@@ -4239,11 +4245,15 @@ int amdgpu_device_mode1_reset(struct amdgpu_device *adev)
 }
 
 int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
-                                 struct amdgpu_job *job,
-                                 bool *need_full_reset_arg)
+                                struct amdgpu_reset_context *reset_context)
 {
        int i, r = 0;
-       bool need_full_reset  = *need_full_reset_arg;
+       struct amdgpu_job *job = NULL;
+       bool need_full_reset =
+               test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
+
+       if (reset_context->reset_req_dev == adev)
+               job = reset_context->job;
 
        /* no need to dump if device is not in good state during probe period */
        if (!adev->gmc.xgmi.pending_reset)
@@ -4268,6 +4278,10 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
        if(job)
                drm_sched_increase_karma(&job->base);
 
+       r = amdgpu_reset_prepare_hwcontext(adev, reset_context);
+       if (r != -ENOSYS)
+               return r;
+
        /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
        if (!amdgpu_sriov_vf(adev)) {
 
@@ -4286,22 +4300,36 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
 
                if (need_full_reset)
                        r = amdgpu_device_ip_suspend(adev);
-
-               *need_full_reset_arg = need_full_reset;
+               if (need_full_reset)
+                       set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
+               else
+                       clear_bit(AMDGPU_NEED_FULL_RESET,
+                                 &reset_context->flags);
        }
 
        return r;
 }
 
-int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
-                         struct list_head *device_list_handle,
-                         bool *need_full_reset_arg,
-                         bool skip_hw_reset)
+int amdgpu_do_asic_reset(struct list_head *device_list_handle,
+                        struct amdgpu_reset_context *reset_context)
 {
        struct amdgpu_device *tmp_adev = NULL;
-       bool need_full_reset = *need_full_reset_arg, vram_lost = false;
+       bool need_full_reset, skip_hw_reset, vram_lost = false;
        int r = 0;
 
+       /* Try reset handler method first */
+       tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
+                                   reset_list);
+       r = amdgpu_reset_perform_reset(tmp_adev, reset_context);
+
+       if (r != -ENOSYS)
+               return r;
+
+       /* Reset handler not implemented, use the default method */
+       need_full_reset =
+               test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
+       skip_hw_reset = test_bit(AMDGPU_SKIP_HW_RESET, &reset_context->flags);
+
        /*
         * ASIC reset has to be done on all XGMI hive nodes ASAP
         * to allow proper links negotiation in FW (within 1 sec)
@@ -4385,7 +4413,8 @@ int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
                                 */
                                amdgpu_register_gpu_instance(tmp_adev);
 
-                               if (!hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1)
+                               if (!reset_context->hive &&
+                                   tmp_adev->gmc.xgmi.num_physical_nodes > 1)
                                        amdgpu_xgmi_add_device(tmp_adev);
 
                                r = amdgpu_device_ip_late_init(tmp_adev);
@@ -4413,8 +4442,10 @@ int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
                                }
 
                                /* Update PSP FW topology after reset */
-                               if (hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1)
-                                       r = amdgpu_xgmi_update_topology(hive, tmp_adev);
+                               if (reset_context->hive &&
+                                   tmp_adev->gmc.xgmi.num_physical_nodes > 1)
+                                       r = amdgpu_xgmi_update_topology(
+                                               reset_context->hive, tmp_adev);
                        }
                }
 
@@ -4438,7 +4469,10 @@ out:
        }
 
 end:
-       *need_full_reset_arg = need_full_reset;
+       if (need_full_reset)
+               set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
+       else
+               clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
        return r;
 }
 
@@ -4575,10 +4609,9 @@ static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev)
        return 0;
 }
 
-void amdgpu_device_recheck_guilty_jobs(struct amdgpu_device *adev,
-                              struct amdgpu_hive_info *hive,
-                              struct list_head *device_list_handle,
-                              bool *need_full_reset)
+void amdgpu_device_recheck_guilty_jobs(
+       struct amdgpu_device *adev, struct list_head *device_list_handle,
+       struct amdgpu_reset_context *reset_context)
 {
        int i, r = 0;
 
@@ -4614,8 +4647,10 @@ retry:
                                if (r)
                                        adev->asic_reset_res = r;
                        } else {
-                               r  = amdgpu_do_asic_reset(hive, device_list_handle,
-                                               need_full_reset, false);
+                               clear_bit(AMDGPU_SKIP_HW_RESET,
+                                         &reset_context->flags);
+                               r = amdgpu_do_asic_reset(device_list_handle,
+                                                        reset_context);
                                if (r && r == -EAGAIN)
                                        goto retry;
                        }
@@ -4657,7 +4692,6 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
                              struct amdgpu_job *job)
 {
        struct list_head device_list, *device_list_handle =  NULL;
-       bool need_full_reset = false;
        bool job_signaled = false;
        struct amdgpu_hive_info *hive = NULL;
        struct amdgpu_device *tmp_adev = NULL;
@@ -4665,6 +4699,9 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
        bool need_emergency_restart = false;
        bool audio_suspended = false;
        int tmp_vram_lost_counter;
+       struct amdgpu_reset_context reset_context;
+
+       memset(&reset_context, 0, sizeof(reset_context));
 
        /*
         * Special case: RAS triggered and full reset isn't supported
@@ -4705,6 +4742,12 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
                mutex_lock(&hive->hive_lock);
        }
 
+       reset_context.method = AMD_RESET_METHOD_NONE;
+       reset_context.reset_req_dev = adev;
+       reset_context.job = job;
+       reset_context.hive = hive;
+       clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
+
        /*
         * lock the device before we try to operate the linked list
         * if didn't get the device lock, don't touch the linked list since
@@ -4805,9 +4848,7 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
 
 retry: /* Rest of adevs pre asic reset from XGMI hive. */
        list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
-               r = amdgpu_device_pre_asic_reset(tmp_adev,
-                                                (tmp_adev == adev) ? job : NULL,
-                                                &need_full_reset);
+               r = amdgpu_device_pre_asic_reset(tmp_adev, &reset_context);
                /*TODO Should we stop ?*/
                if (r) {
                        dev_err(tmp_adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s ",
@@ -4824,7 +4865,7 @@ retry:    /* Rest of adevs pre asic reset from XGMI hive. */
                if (r)
                        adev->asic_reset_res = r;
        } else {
-               r  = amdgpu_do_asic_reset(hive, device_list_handle, &need_full_reset, false);
+               r = amdgpu_do_asic_reset(device_list_handle, &reset_context);
                if (r && r == -EAGAIN)
                        goto retry;
        }
@@ -4843,8 +4884,8 @@ skip_hw_reset:
                 */
                if (amdgpu_gpu_recovery == 2 &&
                        !(tmp_vram_lost_counter < atomic_read(&adev->vram_lost_counter)))
-                       amdgpu_device_recheck_guilty_jobs(tmp_adev, hive,
-                                       device_list_handle, &need_full_reset);
+                       amdgpu_device_recheck_guilty_jobs(
+                               tmp_adev, device_list_handle, &reset_context);
 
                for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
                        struct amdgpu_ring *ring = tmp_adev->rings[i];
@@ -5189,12 +5230,14 @@ pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev)
        struct drm_device *dev = pci_get_drvdata(pdev);
        struct amdgpu_device *adev = drm_to_adev(dev);
        int r, i;
-       bool need_full_reset = true;
+       struct amdgpu_reset_context reset_context;
        u32 memsize;
        struct list_head device_list;
 
        DRM_INFO("PCI error: slot reset callback!!\n");
 
+       memset(&reset_context, 0, sizeof(reset_context));
+
        INIT_LIST_HEAD(&device_list);
        list_add_tail(&adev->reset_list, &device_list);
 
@@ -5217,13 +5260,18 @@ pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev)
                goto out;
        }
 
+       reset_context.method = AMD_RESET_METHOD_NONE;
+       reset_context.reset_req_dev = adev;
+       set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
+       set_bit(AMDGPU_SKIP_HW_RESET, &reset_context.flags);
+
        adev->in_pci_err_recovery = true;
-       r = amdgpu_device_pre_asic_reset(adev, NULL, &need_full_reset);
+       r = amdgpu_device_pre_asic_reset(adev, &reset_context);
        adev->in_pci_err_recovery = false;
        if (r)
                goto out;
 
-       r = amdgpu_do_asic_reset(NULL, &device_list, &need_full_reset, true);
+       r = amdgpu_do_asic_reset(&device_list, &reset_context);
 
 out:
        if (!r) {
index 0e10c39..d8f131e 100644 (file)
@@ -47,6 +47,7 @@
 
 #include "amdgpu_ras.h"
 #include "amdgpu_xgmi.h"
+#include "amdgpu_reset.h"
 
 /*
  * KMS wrapper.
@@ -1349,7 +1350,9 @@ static void amdgpu_drv_delayed_reset_work_handler(struct work_struct *work)
        struct list_head device_list;
        struct amdgpu_device *adev;
        int i, r;
-       bool need_full_reset = true;
+       struct amdgpu_reset_context reset_context;
+
+       memset(&reset_context, 0, sizeof(reset_context));
 
        mutex_lock(&mgpu_info.mutex);
        if (mgpu_info.pending_reset == true) {
@@ -1359,9 +1362,14 @@ static void amdgpu_drv_delayed_reset_work_handler(struct work_struct *work)
        mgpu_info.pending_reset = true;
        mutex_unlock(&mgpu_info.mutex);
 
+       /* Use a common context, just need to make sure full reset is done */
+       reset_context.method = AMD_RESET_METHOD_NONE;
+       set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
+
        for (i = 0; i < mgpu_info.num_dgpu; i++) {
                adev = mgpu_info.gpu_ins[i].adev;
-               r = amdgpu_device_pre_asic_reset(adev, NULL, &need_full_reset);
+               reset_context.reset_req_dev = adev;
+               r = amdgpu_device_pre_asic_reset(adev, &reset_context);
                if (r) {
                        dev_err(adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s ",
                                r, adev_to_drm(adev)->unique);
@@ -1388,7 +1396,10 @@ static void amdgpu_drv_delayed_reset_work_handler(struct work_struct *work)
        list_for_each_entry(adev, &device_list, reset_list)
                amdgpu_unregister_gpu_instance(adev);
 
-       r = amdgpu_do_asic_reset(NULL, &device_list, &need_full_reset, true);
+       /* Use a common context, just need to make sure full reset is done */
+       set_bit(AMDGPU_SKIP_HW_RESET, &reset_context.flags);
+       r = amdgpu_do_asic_reset(&device_list, &reset_context);
+
        if (r) {
                DRM_ERROR("reinit gpus failure");
                return;