1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2023-2024 Intel Corporation 4 */ 5 6 #include <drm/drm_managed.h> 7 8 #include "xe_assert.h" 9 #include "xe_device.h" 10 #include "xe_gt_sriov_printk.h" 11 #include "xe_sriov.h" 12 #include "xe_sriov_printk.h" 13 #include "xe_sriov_vf.h" 14 15 /** 16 * DOC: VF restore procedure in PF KMD and VF KMD 17 * 18 * Restoring previously saved state of a VF is one of core features of 19 * SR-IOV. All major VM Management applications allow saving and restoring 20 * the VM state, and doing that to a VM which uses SRIOV VF as one of 21 * the accessible devices requires support from KMD on both PF and VF side. 22 * VMM initiates all required operations through VFIO module, which then 23 * translates them into PF KMD calls. This description will focus on these 24 * calls, leaving out the module which initiates these steps (VFIO). 25 * 26 * In order to start the restore procedure, GuC needs to keep the VF in 27 * proper state. The PF driver can ensure GuC set it to VF_READY state 28 * by provisioning the VF, which in turn can be done after Function Level 29 * Reset of said VF (or after it was freshly created - in that case FLR 30 * is not needed). The FLR procedure ends with GuC sending message 31 * `GUC_PF_NOTIFY_VF_FLR_DONE`, and then provisioning data is sent to GuC. 32 * After the provisioning is completed, the VF needs to be paused, and 33 * at that point the actual restore can begin. 34 * 35 * During VF Restore, state of several resources is restored. These may 36 * include local memory content (system memory is restored by VMM itself), 37 * values of MMIO registers, stateless compression metadata and others. 38 * The final resource which also needs restoring is state of the VF 39 * submission maintained within GuC. For that, `GUC_PF_OPCODE_VF_RESTORE` 40 * message is used, with reference to the state blob to be consumed by 41 * GuC. 42 * 43 * Next, when VFIO is asked to set the VM into running state, the PF driver 44 * sends `GUC_PF_TRIGGER_VF_RESUME` to GuC. When sent after restore, this 45 * changes VF state within GuC to `VF_RESFIX_BLOCKED` rather than the 46 * usual `VF_RUNNING`. At this point GuC triggers an interrupt to inform 47 * the VF KMD within the VM that it was migrated. 48 * 49 * As soon as Virtual GPU of the VM starts, the VF driver within receives 50 * the MIGRATED interrupt and schedules post-migration recovery worker. 51 * That worker queries GuC for new provisioning (using MMIO communication), 52 * and applies fixups to any non-virtualized resources used by the VF. 53 * 54 * When the VF driver is ready to continue operation on the newly connected 55 * hardware, it sends `VF2GUC_NOTIFY_RESFIX_DONE` which causes it to 56 * enter the long awaited `VF_RUNNING` state, and therefore start handling 57 * CTB messages and scheduling workloads from the VF:: 58 * 59 * PF GuC VF 60 * [ ] | | 61 * [ ] PF2GUC_VF_CONTROL(pause) | | 62 * [ ]---------------------------> [ ] | 63 * [ ] [ ] GuC sets new VF state to | 64 * [ ] [ ]------- VF_READY_PAUSED | 65 * [ ] [ ] | | 66 * [ ] [ ] <----- | 67 * [ ] success [ ] | 68 * [ ] <---------------------------[ ] | 69 * [ ] | | 70 * [ ] PF loads resources from the | | 71 * [ ]------- saved image supplied | | 72 * [ ] | | | 73 * [ ] <----- | | 74 * [ ] | | 75 * [ ] GUC_PF_OPCODE_VF_RESTORE | | 76 * [ ]---------------------------> [ ] | 77 * [ ] [ ] GuC loads contexts and CTB | 78 * [ ] [ ]------- state from image | 79 * [ ] [ ] | | 80 * [ ] [ ] <----- | 81 * [ ] [ ] | 82 * [ ] [ ] GuC sets new VF state to | 83 * [ ] [ ]------- VF_RESFIX_PAUSED | 84 * [ ] [ ] | | 85 * [ ] success [ ] <----- | 86 * [ ] <---------------------------[ ] | 87 * [ ] | | 88 * [ ] GUC_PF_TRIGGER_VF_RESUME | | 89 * [ ]---------------------------> [ ] | 90 * [ ] [ ] GuC sets new VF state to | 91 * [ ] [ ]------- VF_RESFIX_BLOCKED | 92 * [ ] [ ] | | 93 * [ ] [ ] <----- | 94 * [ ] [ ] | 95 * [ ] [ ] GUC_INTR_SW_INT_0 | 96 * [ ] success [ ]---------------------------> [ ] 97 * [ ] <---------------------------[ ] [ ] 98 * | | VF2GUC_QUERY_SINGLE_KLV [ ] 99 * | [ ] <---------------------------[ ] 100 * | [ ] [ ] 101 * | [ ] new VF provisioning [ ] 102 * | [ ]---------------------------> [ ] 103 * | | [ ] 104 * | | VF driver applies post [ ] 105 * | | migration fixups -------[ ] 106 * | | | [ ] 107 * | | -----> [ ] 108 * | | [ ] 109 * | | VF2GUC_NOTIFY_RESFIX_DONE [ ] 110 * | [ ] <---------------------------[ ] 111 * | [ ] [ ] 112 * | [ ] GuC sets new VF state to [ ] 113 * | [ ]------- VF_RUNNING [ ] 114 * | [ ] | [ ] 115 * | [ ] <----- [ ] 116 * | [ ] success [ ] 117 * | [ ]---------------------------> [ ] 118 * | | | 119 * | | | 120 */ 121 122 static void migration_worker_func(struct work_struct *w); 123 124 /** 125 * xe_sriov_vf_init_early - Initialize SR-IOV VF specific data. 126 * @xe: the &xe_device to initialize 127 */ 128 void xe_sriov_vf_init_early(struct xe_device *xe) 129 { 130 INIT_WORK(&xe->sriov.vf.migration.worker, migration_worker_func); 131 } 132 133 static void vf_post_migration_recovery(struct xe_device *xe) 134 { 135 drm_dbg(&xe->drm, "migration recovery in progress\n"); 136 /* FIXME: add the recovery steps */ 137 drm_notice(&xe->drm, "migration recovery ended\n"); 138 } 139 140 static void migration_worker_func(struct work_struct *w) 141 { 142 struct xe_device *xe = container_of(w, struct xe_device, 143 sriov.vf.migration.worker); 144 145 vf_post_migration_recovery(xe); 146 } 147 148 static bool vf_ready_to_recovery_on_all_gts(struct xe_device *xe) 149 { 150 struct xe_gt *gt; 151 unsigned int id; 152 153 for_each_gt(gt, xe, id) { 154 if (!test_bit(id, &xe->sriov.vf.migration.gt_flags)) { 155 xe_gt_sriov_dbg_verbose(gt, "still not ready to recover\n"); 156 return false; 157 } 158 } 159 return true; 160 } 161 162 /** 163 * xe_sriov_vf_start_migration_recovery - Start VF migration recovery. 164 * @xe: the &xe_device to start recovery on 165 * 166 * This function shall be called only by VF. 167 */ 168 void xe_sriov_vf_start_migration_recovery(struct xe_device *xe) 169 { 170 bool started; 171 172 xe_assert(xe, IS_SRIOV_VF(xe)); 173 174 if (!vf_ready_to_recovery_on_all_gts(xe)) 175 return; 176 177 WRITE_ONCE(xe->sriov.vf.migration.gt_flags, 0); 178 /* Ensure other threads see that no flags are set now. */ 179 smp_mb(); 180 181 started = queue_work(xe->sriov.wq, &xe->sriov.vf.migration.worker); 182 drm_info(&xe->drm, "VF migration recovery %s\n", started ? 183 "scheduled" : "already in progress"); 184 } 185