| 1 | // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB |
| 2 | /* |
| 3 | * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved. |
| 4 | * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved. |
| 5 | */ |
| 6 | |
| 7 | #include "rxe.h" |
| 8 | |
| 9 | static struct workqueue_struct *rxe_wq; |
| 10 | |
| 11 | int rxe_alloc_wq(void) |
| 12 | { |
| 13 | rxe_wq = alloc_workqueue("rxe_wq" , WQ_UNBOUND, WQ_MAX_ACTIVE); |
| 14 | if (!rxe_wq) |
| 15 | return -ENOMEM; |
| 16 | |
| 17 | return 0; |
| 18 | } |
| 19 | |
| 20 | void rxe_destroy_wq(void) |
| 21 | { |
| 22 | destroy_workqueue(wq: rxe_wq); |
| 23 | } |
| 24 | |
| 25 | /* Check if task is idle i.e. not running, not scheduled in |
| 26 | * work queue and not draining. If so move to busy to |
| 27 | * reserve a slot in do_task() by setting to busy and taking |
| 28 | * a qp reference to cover the gap from now until the task finishes. |
| 29 | * state will move out of busy if task returns a non zero value |
| 30 | * in do_task(). If state is already busy it is raised to armed |
| 31 | * to indicate to do_task that additional pass should be made |
| 32 | * over the task. |
| 33 | * Context: caller should hold task->lock. |
| 34 | * Returns: true if state transitioned from idle to busy else false. |
| 35 | */ |
| 36 | static bool __reserve_if_idle(struct rxe_task *task) |
| 37 | { |
| 38 | WARN_ON(rxe_read(task->qp) <= 0); |
| 39 | |
| 40 | if (task->state == TASK_STATE_IDLE) { |
| 41 | rxe_get(task->qp); |
| 42 | task->state = TASK_STATE_BUSY; |
| 43 | task->num_sched++; |
| 44 | return true; |
| 45 | } |
| 46 | |
| 47 | if (task->state == TASK_STATE_BUSY) |
| 48 | task->state = TASK_STATE_ARMED; |
| 49 | |
| 50 | return false; |
| 51 | } |
| 52 | |
| 53 | /* check if task is idle or drained and not currently |
| 54 | * scheduled in the work queue. This routine is |
| 55 | * called by rxe_cleanup_task or rxe_disable_task to |
| 56 | * see if the queue is empty. |
| 57 | * Context: caller should hold task->lock. |
| 58 | * Returns true if done else false. |
| 59 | */ |
| 60 | static bool __is_done(struct rxe_task *task) |
| 61 | { |
| 62 | if (work_pending(&task->work)) |
| 63 | return false; |
| 64 | |
| 65 | if (task->state == TASK_STATE_IDLE || |
| 66 | task->state == TASK_STATE_DRAINED) { |
| 67 | return true; |
| 68 | } |
| 69 | |
| 70 | return false; |
| 71 | } |
| 72 | |
| 73 | /* a locked version of __is_done */ |
| 74 | static bool is_done(struct rxe_task *task) |
| 75 | { |
| 76 | unsigned long flags; |
| 77 | int done; |
| 78 | |
| 79 | spin_lock_irqsave(&task->lock, flags); |
| 80 | done = __is_done(task); |
| 81 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 82 | |
| 83 | return done; |
| 84 | } |
| 85 | |
| 86 | /* do_task is a wrapper for the three tasks (requester, |
| 87 | * completer, responder) and calls them in a loop until |
| 88 | * they return a non-zero value. It is called indirectly |
| 89 | * when rxe_sched_task schedules the task. They must |
| 90 | * call __reserve_if_idle to move the task to busy before |
| 91 | * calling or scheduling. The task can also be moved to |
| 92 | * drained or invalid by calls to rxe_cleanup_task or |
| 93 | * rxe_disable_task. In that case tasks which get here |
| 94 | * are not executed but just flushed. The tasks are |
| 95 | * designed to look to see if there is work to do and |
| 96 | * then do part of it before returning here with a return |
| 97 | * value of zero until all the work has been consumed then |
| 98 | * it returns a non-zero value. |
| 99 | * The number of times the task can be run is limited by |
| 100 | * max iterations so one task cannot hold the cpu forever. |
| 101 | * If the limit is hit and work remains the task is rescheduled. |
| 102 | */ |
| 103 | static void do_task(struct rxe_task *task) |
| 104 | { |
| 105 | unsigned int iterations; |
| 106 | unsigned long flags; |
| 107 | int resched = 0; |
| 108 | int cont; |
| 109 | int ret; |
| 110 | |
| 111 | WARN_ON(rxe_read(task->qp) <= 0); |
| 112 | |
| 113 | spin_lock_irqsave(&task->lock, flags); |
| 114 | if (task->state >= TASK_STATE_DRAINED) { |
| 115 | rxe_put(task->qp); |
| 116 | task->num_done++; |
| 117 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 118 | return; |
| 119 | } |
| 120 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 121 | |
| 122 | do { |
| 123 | iterations = RXE_MAX_ITERATIONS; |
| 124 | cont = 0; |
| 125 | |
| 126 | do { |
| 127 | ret = task->func(task->qp); |
| 128 | } while (ret == 0 && iterations-- > 0); |
| 129 | |
| 130 | spin_lock_irqsave(&task->lock, flags); |
| 131 | /* we're not done yet but we ran out of iterations. |
| 132 | * yield the cpu and reschedule the task |
| 133 | */ |
| 134 | if (!ret) { |
| 135 | if (task->state != TASK_STATE_DRAINING) { |
| 136 | task->state = TASK_STATE_IDLE; |
| 137 | resched = 1; |
| 138 | } else { |
| 139 | cont = 1; |
| 140 | } |
| 141 | goto exit; |
| 142 | } |
| 143 | |
| 144 | switch (task->state) { |
| 145 | case TASK_STATE_BUSY: |
| 146 | task->state = TASK_STATE_IDLE; |
| 147 | break; |
| 148 | |
| 149 | /* someone tried to schedule the task while we |
| 150 | * were running, keep going |
| 151 | */ |
| 152 | case TASK_STATE_ARMED: |
| 153 | task->state = TASK_STATE_BUSY; |
| 154 | cont = 1; |
| 155 | break; |
| 156 | |
| 157 | case TASK_STATE_DRAINING: |
| 158 | task->state = TASK_STATE_DRAINED; |
| 159 | break; |
| 160 | |
| 161 | default: |
| 162 | WARN_ON(1); |
| 163 | rxe_dbg_qp(task->qp, "unexpected task state = %d\n" , |
| 164 | task->state); |
| 165 | task->state = TASK_STATE_IDLE; |
| 166 | } |
| 167 | |
| 168 | exit: |
| 169 | if (!cont) { |
| 170 | task->num_done++; |
| 171 | if (WARN_ON(task->num_done != task->num_sched)) |
| 172 | rxe_dbg_qp( |
| 173 | task->qp, |
| 174 | "%ld tasks scheduled, %ld tasks done\n" , |
| 175 | task->num_sched, task->num_done); |
| 176 | } |
| 177 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 178 | } while (cont); |
| 179 | |
| 180 | task->ret = ret; |
| 181 | |
| 182 | if (resched) |
| 183 | rxe_sched_task(task); |
| 184 | |
| 185 | rxe_put(task->qp); |
| 186 | } |
| 187 | |
| 188 | /* wrapper around do_task to fix argument for work queue */ |
| 189 | static void do_work(struct work_struct *work) |
| 190 | { |
| 191 | do_task(container_of(work, struct rxe_task, work)); |
| 192 | } |
| 193 | |
| 194 | int rxe_init_task(struct rxe_task *task, struct rxe_qp *qp, |
| 195 | int (*func)(struct rxe_qp *)) |
| 196 | { |
| 197 | WARN_ON(rxe_read(qp) <= 0); |
| 198 | |
| 199 | task->qp = qp; |
| 200 | task->func = func; |
| 201 | task->state = TASK_STATE_IDLE; |
| 202 | spin_lock_init(&task->lock); |
| 203 | INIT_WORK(&task->work, do_work); |
| 204 | |
| 205 | return 0; |
| 206 | } |
| 207 | |
| 208 | /* rxe_cleanup_task is only called from rxe_do_qp_cleanup in |
| 209 | * process context. The qp is already completed with no |
| 210 | * remaining references. Once the queue is drained the |
| 211 | * task is moved to invalid and returns. The qp cleanup |
| 212 | * code then calls the task functions directly without |
| 213 | * using the task struct to drain any late arriving packets |
| 214 | * or work requests. |
| 215 | */ |
| 216 | void rxe_cleanup_task(struct rxe_task *task) |
| 217 | { |
| 218 | unsigned long flags; |
| 219 | |
| 220 | spin_lock_irqsave(&task->lock, flags); |
| 221 | if (!__is_done(task) && task->state < TASK_STATE_DRAINED) { |
| 222 | task->state = TASK_STATE_DRAINING; |
| 223 | } else { |
| 224 | task->state = TASK_STATE_INVALID; |
| 225 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 226 | return; |
| 227 | } |
| 228 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 229 | |
| 230 | /* now the task cannot be scheduled or run just wait |
| 231 | * for the previously scheduled tasks to finish. |
| 232 | */ |
| 233 | while (!is_done(task)) |
| 234 | cond_resched(); |
| 235 | |
| 236 | spin_lock_irqsave(&task->lock, flags); |
| 237 | task->state = TASK_STATE_INVALID; |
| 238 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 239 | } |
| 240 | |
| 241 | /* schedule the task to run later as a work queue entry. |
| 242 | * the queue_work call can be called holding |
| 243 | * the lock. |
| 244 | */ |
| 245 | void rxe_sched_task(struct rxe_task *task) |
| 246 | { |
| 247 | unsigned long flags; |
| 248 | |
| 249 | WARN_ON(rxe_read(task->qp) <= 0); |
| 250 | |
| 251 | spin_lock_irqsave(&task->lock, flags); |
| 252 | if (__reserve_if_idle(task)) |
| 253 | queue_work(wq: rxe_wq, work: &task->work); |
| 254 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 255 | } |
| 256 | |
| 257 | /* rxe_disable/enable_task are only called from |
| 258 | * rxe_modify_qp in process context. Task is moved |
| 259 | * to the drained state by do_task. |
| 260 | */ |
| 261 | void rxe_disable_task(struct rxe_task *task) |
| 262 | { |
| 263 | unsigned long flags; |
| 264 | |
| 265 | WARN_ON(rxe_read(task->qp) <= 0); |
| 266 | |
| 267 | spin_lock_irqsave(&task->lock, flags); |
| 268 | if (!__is_done(task) && task->state < TASK_STATE_DRAINED) { |
| 269 | task->state = TASK_STATE_DRAINING; |
| 270 | } else { |
| 271 | task->state = TASK_STATE_DRAINED; |
| 272 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 273 | return; |
| 274 | } |
| 275 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 276 | |
| 277 | while (!is_done(task)) |
| 278 | cond_resched(); |
| 279 | |
| 280 | spin_lock_irqsave(&task->lock, flags); |
| 281 | task->state = TASK_STATE_DRAINED; |
| 282 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 283 | } |
| 284 | |
| 285 | void rxe_enable_task(struct rxe_task *task) |
| 286 | { |
| 287 | unsigned long flags; |
| 288 | |
| 289 | WARN_ON(rxe_read(task->qp) <= 0); |
| 290 | |
| 291 | spin_lock_irqsave(&task->lock, flags); |
| 292 | if (task->state == TASK_STATE_INVALID) { |
| 293 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 294 | return; |
| 295 | } |
| 296 | |
| 297 | task->state = TASK_STATE_IDLE; |
| 298 | spin_unlock_irqrestore(lock: &task->lock, flags); |
| 299 | } |
| 300 | |