1/* 2 * Fence mechanism for dma-buf to allow for asynchronous dma access 3 * 4 * Copyright (C) 2012 Canonical Ltd 5 * Copyright (C) 2012 Texas Instruments 6 * 7 * Authors: 8 * Rob Clark <robdclark@gmail.com> 9 * Maarten Lankhorst <maarten.lankhorst@canonical.com> 10 * 11 * This program is free software; you can redistribute it and/or modify it 12 * under the terms of the GNU General Public License version 2 as published by 13 * the Free Software Foundation. 14 * 15 * This program is distributed in the hope that it will be useful, but WITHOUT 16 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 17 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 18 * more details. 19 */ 20 21#ifndef __LINUX_DMA_FENCE_H 22#define __LINUX_DMA_FENCE_H 23 24#include <linux/err.h> 25#include <linux/wait.h> 26#include <linux/list.h> 27#include <linux/bitops.h> 28#include <linux/kref.h> 29#include <linux/sched.h> 30#include <linux/printk.h> 31#include <linux/rcupdate.h> 32 33struct dma_fence; 34struct dma_fence_ops; 35struct dma_fence_cb; 36 37/** 38 * struct dma_fence - software synchronization primitive 39 * @refcount: refcount for this fence 40 * @ops: dma_fence_ops associated with this fence 41 * @rcu: used for releasing fence with kfree_rcu 42 * @cb_list: list of all callbacks to call 43 * @lock: spin_lock_irqsave used for locking 44 * @context: execution context this fence belongs to, returned by 45 * dma_fence_context_alloc() 46 * @seqno: the sequence number of this fence inside the execution context, 47 * can be compared to decide which fence would be signaled later. 48 * @flags: A mask of DMA_FENCE_FLAG_* defined below 49 * @timestamp: Timestamp when the fence was signaled. 50 * @error: Optional, only valid if < 0, must be set before calling 51 * dma_fence_signal, indicates that the fence has completed with an error. 52 * 53 * the flags member must be manipulated and read using the appropriate 54 * atomic ops (bit_*), so taking the spinlock will not be needed most 55 * of the time. 56 * 57 * DMA_FENCE_FLAG_SIGNALED_BIT - fence is already signaled 58 * DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT - enable_signaling might have been called 59 * DMA_FENCE_FLAG_USER_BITS - start of the unused bits, can be used by the 60 * implementer of the fence for its own purposes. Can be used in different 61 * ways by different fence implementers, so do not rely on this. 62 * 63 * Since atomic bitops are used, this is not guaranteed to be the case. 64 * Particularly, if the bit was set, but dma_fence_signal was called right 65 * before this bit was set, it would have been able to set the 66 * DMA_FENCE_FLAG_SIGNALED_BIT, before enable_signaling was called. 67 * Adding a check for DMA_FENCE_FLAG_SIGNALED_BIT after setting 68 * DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT closes this race, and makes sure that 69 * after dma_fence_signal was called, any enable_signaling call will have either 70 * been completed, or never called at all. 71 */ 72struct dma_fence { 73 struct kref refcount; 74 const struct dma_fence_ops *ops; 75 struct rcu_head rcu; 76 struct list_head cb_list; 77 spinlock_t *lock; 78 u64 context; 79 unsigned seqno; 80 unsigned long flags; 81 ktime_t timestamp; 82 int error; 83}; 84 85enum dma_fence_flag_bits { 86 DMA_FENCE_FLAG_SIGNALED_BIT, 87 DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, 88 DMA_FENCE_FLAG_USER_BITS, /* must always be last member */ 89}; 90 91typedef void (*dma_fence_func_t)(struct dma_fence *fence, 92 struct dma_fence_cb *cb); 93 94/** 95 * struct dma_fence_cb - callback for dma_fence_add_callback 96 * @node: used by dma_fence_add_callback to append this struct to fence::cb_list 97 * @func: dma_fence_func_t to call 98 * 99 * This struct will be initialized by dma_fence_add_callback, additional 100 * data can be passed along by embedding dma_fence_cb in another struct. 101 */ 102struct dma_fence_cb { 103 struct list_head node; 104 dma_fence_func_t func; 105}; 106 107/** 108 * struct dma_fence_ops - operations implemented for fence 109 * @get_driver_name: returns the driver name. 110 * @get_timeline_name: return the name of the context this fence belongs to. 111 * @enable_signaling: enable software signaling of fence. 112 * @signaled: [optional] peek whether the fence is signaled, can be null. 113 * @wait: custom wait implementation, or dma_fence_default_wait. 114 * @release: [optional] called on destruction of fence, can be null 115 * @fill_driver_data: [optional] callback to fill in free-form debug info 116 * Returns amount of bytes filled, or -errno. 117 * @fence_value_str: [optional] fills in the value of the fence as a string 118 * @timeline_value_str: [optional] fills in the current value of the timeline 119 * as a string 120 * 121 * Notes on enable_signaling: 122 * For fence implementations that have the capability for hw->hw 123 * signaling, they can implement this op to enable the necessary 124 * irqs, or insert commands into cmdstream, etc. This is called 125 * in the first wait() or add_callback() path to let the fence 126 * implementation know that there is another driver waiting on 127 * the signal (ie. hw->sw case). 128 * 129 * This function can be called called from atomic context, but not 130 * from irq context, so normal spinlocks can be used. 131 * 132 * A return value of false indicates the fence already passed, 133 * or some failure occurred that made it impossible to enable 134 * signaling. True indicates successful enabling. 135 * 136 * fence->error may be set in enable_signaling, but only when false is 137 * returned. 138 * 139 * Calling dma_fence_signal before enable_signaling is called allows 140 * for a tiny race window in which enable_signaling is called during, 141 * before, or after dma_fence_signal. To fight this, it is recommended 142 * that before enable_signaling returns true an extra reference is 143 * taken on the fence, to be released when the fence is signaled. 144 * This will mean dma_fence_signal will still be called twice, but 145 * the second time will be a noop since it was already signaled. 146 * 147 * Notes on signaled: 148 * May set fence->error if returning true. 149 * 150 * Notes on wait: 151 * Must not be NULL, set to dma_fence_default_wait for default implementation. 152 * the dma_fence_default_wait implementation should work for any fence, as long 153 * as enable_signaling works correctly. 154 * 155 * Must return -ERESTARTSYS if the wait is intr = true and the wait was 156 * interrupted, and remaining jiffies if fence has signaled, or 0 if wait 157 * timed out. Can also return other error values on custom implementations, 158 * which should be treated as if the fence is signaled. For example a hardware 159 * lockup could be reported like that. 160 * 161 * Notes on release: 162 * Can be NULL, this function allows additional commands to run on 163 * destruction of the fence. Can be called from irq context. 164 * If pointer is set to NULL, kfree will get called instead. 165 */ 166 167struct dma_fence_ops { 168 const char * (*get_driver_name)(struct dma_fence *fence); 169 const char * (*get_timeline_name)(struct dma_fence *fence); 170 bool (*enable_signaling)(struct dma_fence *fence); 171 bool (*signaled)(struct dma_fence *fence); 172 signed long (*wait)(struct dma_fence *fence, 173 bool intr, signed long timeout); 174 void (*release)(struct dma_fence *fence); 175 176 int (*fill_driver_data)(struct dma_fence *fence, void *data, int size); 177 void (*fence_value_str)(struct dma_fence *fence, char *str, int size); 178 void (*timeline_value_str)(struct dma_fence *fence, 179 char *str, int size); 180}; 181 182void dma_fence_init(struct dma_fence *fence, const struct dma_fence_ops *ops, 183 spinlock_t *lock, u64 context, unsigned seqno); 184 185void dma_fence_release(struct kref *kref); 186void dma_fence_free(struct dma_fence *fence); 187 188/** 189 * dma_fence_put - decreases refcount of the fence 190 * @fence: [in] fence to reduce refcount of 191 */ 192static inline void dma_fence_put(struct dma_fence *fence) 193{ 194 if (fence) 195 kref_put(&fence->refcount, dma_fence_release); 196} 197 198/** 199 * dma_fence_get - increases refcount of the fence 200 * @fence: [in] fence to increase refcount of 201 * 202 * Returns the same fence, with refcount increased by 1. 203 */ 204static inline struct dma_fence *dma_fence_get(struct dma_fence *fence) 205{ 206 if (fence) 207 kref_get(&fence->refcount); 208 return fence; 209} 210 211/** 212 * dma_fence_get_rcu - get a fence from a reservation_object_list with 213 * rcu read lock 214 * @fence: [in] fence to increase refcount of 215 * 216 * Function returns NULL if no refcount could be obtained, or the fence. 217 */ 218static inline struct dma_fence *dma_fence_get_rcu(struct dma_fence *fence) 219{ 220 if (kref_get_unless_zero(&fence->refcount)) 221 return fence; 222 else 223 return NULL; 224} 225 226/** 227 * dma_fence_get_rcu_safe - acquire a reference to an RCU tracked fence 228 * @fencep: [in] pointer to fence to increase refcount of 229 * 230 * Function returns NULL if no refcount could be obtained, or the fence. 231 * This function handles acquiring a reference to a fence that may be 232 * reallocated within the RCU grace period (such as with SLAB_TYPESAFE_BY_RCU), 233 * so long as the caller is using RCU on the pointer to the fence. 234 * 235 * An alternative mechanism is to employ a seqlock to protect a bunch of 236 * fences, such as used by struct reservation_object. When using a seqlock, 237 * the seqlock must be taken before and checked after a reference to the 238 * fence is acquired (as shown here). 239 * 240 * The caller is required to hold the RCU read lock. 241 */ 242static inline struct dma_fence * 243dma_fence_get_rcu_safe(struct dma_fence * __rcu *fencep) 244{ 245 do { 246 struct dma_fence *fence; 247 248 fence = rcu_dereference(*fencep); 249 if (!fence || !dma_fence_get_rcu(fence)) 250 return NULL; 251 252 /* The atomic_inc_not_zero() inside dma_fence_get_rcu() 253 * provides a full memory barrier upon success (such as now). 254 * This is paired with the write barrier from assigning 255 * to the __rcu protected fence pointer so that if that 256 * pointer still matches the current fence, we know we 257 * have successfully acquire a reference to it. If it no 258 * longer matches, we are holding a reference to some other 259 * reallocated pointer. This is possible if the allocator 260 * is using a freelist like SLAB_TYPESAFE_BY_RCU where the 261 * fence remains valid for the RCU grace period, but it 262 * may be reallocated. When using such allocators, we are 263 * responsible for ensuring the reference we get is to 264 * the right fence, as below. 265 */ 266 if (fence == rcu_access_pointer(*fencep)) 267 return rcu_pointer_handoff(fence); 268 269 dma_fence_put(fence); 270 } while (1); 271} 272 273int dma_fence_signal(struct dma_fence *fence); 274int dma_fence_signal_locked(struct dma_fence *fence); 275signed long dma_fence_default_wait(struct dma_fence *fence, 276 bool intr, signed long timeout); 277int dma_fence_add_callback(struct dma_fence *fence, 278 struct dma_fence_cb *cb, 279 dma_fence_func_t func); 280bool dma_fence_remove_callback(struct dma_fence *fence, 281 struct dma_fence_cb *cb); 282void dma_fence_enable_sw_signaling(struct dma_fence *fence); 283 284/** 285 * dma_fence_is_signaled_locked - Return an indication if the fence 286 * is signaled yet. 287 * @fence: [in] the fence to check 288 * 289 * Returns true if the fence was already signaled, false if not. Since this 290 * function doesn't enable signaling, it is not guaranteed to ever return 291 * true if dma_fence_add_callback, dma_fence_wait or 292 * dma_fence_enable_sw_signaling haven't been called before. 293 * 294 * This function requires fence->lock to be held. 295 */ 296static inline bool 297dma_fence_is_signaled_locked(struct dma_fence *fence) 298{ 299 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) 300 return true; 301 302 if (fence->ops->signaled && fence->ops->signaled(fence)) { 303 dma_fence_signal_locked(fence); 304 return true; 305 } 306 307 return false; 308} 309 310/** 311 * dma_fence_is_signaled - Return an indication if the fence is signaled yet. 312 * @fence: [in] the fence to check 313 * 314 * Returns true if the fence was already signaled, false if not. Since this 315 * function doesn't enable signaling, it is not guaranteed to ever return 316 * true if dma_fence_add_callback, dma_fence_wait or 317 * dma_fence_enable_sw_signaling haven't been called before. 318 * 319 * It's recommended for seqno fences to call dma_fence_signal when the 320 * operation is complete, it makes it possible to prevent issues from 321 * wraparound between time of issue and time of use by checking the return 322 * value of this function before calling hardware-specific wait instructions. 323 */ 324static inline bool 325dma_fence_is_signaled(struct dma_fence *fence) 326{ 327 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) 328 return true; 329 330 if (fence->ops->signaled && fence->ops->signaled(fence)) { 331 dma_fence_signal(fence); 332 return true; 333 } 334 335 return false; 336} 337 338/** 339 * dma_fence_is_later - return if f1 is chronologically later than f2 340 * @f1: [in] the first fence from the same context 341 * @f2: [in] the second fence from the same context 342 * 343 * Returns true if f1 is chronologically later than f2. Both fences must be 344 * from the same context, since a seqno is not re-used across contexts. 345 */ 346static inline bool dma_fence_is_later(struct dma_fence *f1, 347 struct dma_fence *f2) 348{ 349 if (WARN_ON(f1->context != f2->context)) 350 return false; 351 352 return (int)(f1->seqno - f2->seqno) > 0; 353} 354 355/** 356 * dma_fence_later - return the chronologically later fence 357 * @f1: [in] the first fence from the same context 358 * @f2: [in] the second fence from the same context 359 * 360 * Returns NULL if both fences are signaled, otherwise the fence that would be 361 * signaled last. Both fences must be from the same context, since a seqno is 362 * not re-used across contexts. 363 */ 364static inline struct dma_fence *dma_fence_later(struct dma_fence *f1, 365 struct dma_fence *f2) 366{ 367 if (WARN_ON(f1->context != f2->context)) 368 return NULL; 369 370 /* 371 * Can't check just DMA_FENCE_FLAG_SIGNALED_BIT here, it may never 372 * have been set if enable_signaling wasn't called, and enabling that 373 * here is overkill. 374 */ 375 if (dma_fence_is_later(f1, f2)) 376 return dma_fence_is_signaled(f1) ? NULL : f1; 377 else 378 return dma_fence_is_signaled(f2) ? NULL : f2; 379} 380 381/** 382 * dma_fence_get_status_locked - returns the status upon completion 383 * @fence: [in] the dma_fence to query 384 * 385 * Drivers can supply an optional error status condition before they signal 386 * the fence (to indicate whether the fence was completed due to an error 387 * rather than success). The value of the status condition is only valid 388 * if the fence has been signaled, dma_fence_get_status_locked() first checks 389 * the signal state before reporting the error status. 390 * 391 * Returns 0 if the fence has not yet been signaled, 1 if the fence has 392 * been signaled without an error condition, or a negative error code 393 * if the fence has been completed in err. 394 */ 395static inline int dma_fence_get_status_locked(struct dma_fence *fence) 396{ 397 if (dma_fence_is_signaled_locked(fence)) 398 return fence->error ?: 1; 399 else 400 return 0; 401} 402 403int dma_fence_get_status(struct dma_fence *fence); 404 405/** 406 * dma_fence_set_error - flag an error condition on the fence 407 * @fence: [in] the dma_fence 408 * @error: [in] the error to store 409 * 410 * Drivers can supply an optional error status condition before they signal 411 * the fence, to indicate that the fence was completed due to an error 412 * rather than success. This must be set before signaling (so that the value 413 * is visible before any waiters on the signal callback are woken). This 414 * helper exists to help catching erroneous setting of #dma_fence.error. 415 */ 416static inline void dma_fence_set_error(struct dma_fence *fence, 417 int error) 418{ 419 BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)); 420 BUG_ON(error >= 0 || error < -MAX_ERRNO); 421 422 fence->error = error; 423} 424 425signed long dma_fence_wait_timeout(struct dma_fence *, 426 bool intr, signed long timeout); 427signed long dma_fence_wait_any_timeout(struct dma_fence **fences, 428 uint32_t count, 429 bool intr, signed long timeout, 430 uint32_t *idx); 431 432/** 433 * dma_fence_wait - sleep until the fence gets signaled 434 * @fence: [in] the fence to wait on 435 * @intr: [in] if true, do an interruptible wait 436 * 437 * This function will return -ERESTARTSYS if interrupted by a signal, 438 * or 0 if the fence was signaled. Other error values may be 439 * returned on custom implementations. 440 * 441 * Performs a synchronous wait on this fence. It is assumed the caller 442 * directly or indirectly holds a reference to the fence, otherwise the 443 * fence might be freed before return, resulting in undefined behavior. 444 */ 445static inline signed long dma_fence_wait(struct dma_fence *fence, bool intr) 446{ 447 signed long ret; 448 449 /* Since dma_fence_wait_timeout cannot timeout with 450 * MAX_SCHEDULE_TIMEOUT, only valid return values are 451 * -ERESTARTSYS and MAX_SCHEDULE_TIMEOUT. 452 */ 453 ret = dma_fence_wait_timeout(fence, intr, MAX_SCHEDULE_TIMEOUT); 454 455 return ret < 0 ? ret : 0; 456} 457 458u64 dma_fence_context_alloc(unsigned num); 459 460#define DMA_FENCE_TRACE(f, fmt, args...) \ 461 do { \ 462 struct dma_fence *__ff = (f); \ 463 if (IS_ENABLED(CONFIG_DMA_FENCE_TRACE)) \ 464 pr_info("f %llu#%u: " fmt, \ 465 __ff->context, __ff->seqno, ##args); \ 466 } while (0) 467 468#define DMA_FENCE_WARN(f, fmt, args...) \ 469 do { \ 470 struct dma_fence *__ff = (f); \ 471 pr_warn("f %llu#%u: " fmt, __ff->context, __ff->seqno, \ 472 ##args); \ 473 } while (0) 474 475#define DMA_FENCE_ERR(f, fmt, args...) \ 476 do { \ 477 struct dma_fence *__ff = (f); \ 478 pr_err("f %llu#%u: " fmt, __ff->context, __ff->seqno, \ 479 ##args); \ 480 } while (0) 481 482#endif /* __LINUX_DMA_FENCE_H */ 483