• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2012-2018 Rob Clark <robclark@freedesktop.org>
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21  * SOFTWARE.
22  *
23  * Authors:
24  *    Rob Clark <robclark@freedesktop.org>
25  */
26 
27 #ifndef FREEDRENO_PRIV_H_
28 #define FREEDRENO_PRIV_H_
29 
30 #include <errno.h>
31 #include <fcntl.h>
32 #include <stdio.h>
33 #include <stdlib.h>
34 #include <string.h>
35 #include <unistd.h>
36 #include <sys/ioctl.h>
37 #include <sys/mman.h>
38 
39 #include <xf86drm.h>
40 
41 #include "util/hash_table.h"
42 #include "util/list.h"
43 #include "util/log.h"
44 #include "util/simple_mtx.h"
45 #include "util/u_atomic.h"
46 #include "util/u_debug.h"
47 #include "util/u_math.h"
48 
49 #include "freedreno_dev_info.h"
50 #include "freedreno_drmif.h"
51 #include "freedreno_ringbuffer.h"
52 
53 extern simple_mtx_t table_lock;
54 
55 /*
56  * Stupid/simple growable array implementation:
57  */
58 
59 #define MAX_ARRAY_SIZE ((unsigned short)~0)
60 
61 static inline void
grow(void ** ptr,uint16_t nr,uint16_t * max,uint16_t sz)62 grow(void **ptr, uint16_t nr, uint16_t *max, uint16_t sz)
63 {
64    assert((nr + 1) < MAX_ARRAY_SIZE);
65    if ((nr + 1) > *max) {
66       if (*max > MAX_ARRAY_SIZE/2)
67          *max = MAX_ARRAY_SIZE;
68       else if ((*max * 2) < (nr + 1))
69          *max = nr + 5;
70       else
71          *max = *max * 2;
72       *ptr = realloc(*ptr, *max * sz);
73    }
74 }
75 
76 #define DECLARE_ARRAY(type, name)                                              \
77    unsigned short nr_##name, max_##name;                                       \
78    type *name;
79 
80 #define APPEND(x, name, ...)                                                   \
81    ({                                                                          \
82       grow((void **)&(x)->name, (x)->nr_##name, &(x)->max_##name,              \
83            sizeof((x)->name[0]));                                              \
84       (x)->name[(x)->nr_##name] = __VA_ARGS__;                                 \
85       (x)->nr_##name++;                                                        \
86    })
87 
88 #define READ_ONCE(x) (*(volatile __typeof__(x) *)&(x))
89 
90 
91 struct fd_device_funcs {
92    int (*bo_new_handle)(struct fd_device *dev, uint32_t size, uint32_t flags,
93                         uint32_t *handle);
94    struct fd_bo *(*bo_from_handle)(struct fd_device *dev, uint32_t size,
95                                    uint32_t handle);
96    struct fd_pipe *(*pipe_new)(struct fd_device *dev, enum fd_pipe_id id,
97                                unsigned prio);
98    void (*destroy)(struct fd_device *dev);
99 };
100 
101 struct fd_bo_bucket {
102    uint32_t size;
103    struct list_head list;
104 };
105 
106 struct fd_bo_cache {
107    struct fd_bo_bucket cache_bucket[14 * 4];
108    int num_buckets;
109    time_t time;
110 };
111 
112 struct fd_device {
113    int fd;
114    enum fd_version version;
115    int32_t refcnt;
116 
117    /* tables to keep track of bo's, to avoid "evil-twin" fd_bo objects:
118     *
119     *   handle_table: maps handle to fd_bo
120     *   name_table: maps flink name to fd_bo
121     *
122     * We end up needing two tables, because DRM_IOCTL_GEM_OPEN always
123     * returns a new handle.  So we need to figure out if the bo is already
124     * open in the process first, before calling gem-open.
125     */
126    struct hash_table *handle_table, *name_table;
127 
128    const struct fd_device_funcs *funcs;
129 
130    struct fd_bo_cache bo_cache;
131    struct fd_bo_cache ring_cache;
132 
133    bool has_cached_coherent;
134 
135    bool closefd; /* call close(fd) upon destruction */
136 
137    /* just for valgrind: */
138    int bo_size;
139 
140    /**
141     * List of deferred submits, protected by submit_lock.  The deferred
142     * submits are tracked globally per-device, even if they execute in
143     * different order on the kernel side (ie. due to different priority
144     * submitqueues, etc) to preserve the order that they are passed off
145     * to the kernel.  Once the kernel has them, it is the fences' job
146     * to preserve correct order of execution.
147     */
148    struct list_head deferred_submits;
149    unsigned deferred_cmds;
150    simple_mtx_t submit_lock;
151 };
152 
153 #define foreach_submit(name, list) \
154    list_for_each_entry(struct fd_submit, name, list, node)
155 #define foreach_submit_safe(name, list) \
156    list_for_each_entry_safe(struct fd_submit, name, list, node)
157 #define last_submit(list) \
158    list_last_entry(list, struct fd_submit, node)
159 
160 void fd_bo_cache_init(struct fd_bo_cache *cache, int coarse);
161 void fd_bo_cache_cleanup(struct fd_bo_cache *cache, time_t time);
162 struct fd_bo *fd_bo_cache_alloc(struct fd_bo_cache *cache, uint32_t *size,
163                                 uint32_t flags);
164 int fd_bo_cache_free(struct fd_bo_cache *cache, struct fd_bo *bo);
165 
166 /* for where @table_lock is already held: */
167 void fd_bo_del_locked(struct fd_bo *bo);
168 void fd_device_del_locked(struct fd_device *dev);
169 void fd_pipe_del_locked(struct fd_pipe *pipe);
170 
171 struct fd_pipe_funcs {
172    struct fd_ringbuffer *(*ringbuffer_new_object)(struct fd_pipe *pipe,
173                                                   uint32_t size);
174    struct fd_submit *(*submit_new)(struct fd_pipe *pipe);
175 
176    /**
177     * Flush any deferred submits (if deferred submits are supported by
178     * the pipe implementation)
179     */
180    void (*flush)(struct fd_pipe *pipe, uint32_t fence);
181 
182    int (*get_param)(struct fd_pipe *pipe, enum fd_param_id param,
183                     uint64_t *value);
184    int (*wait)(struct fd_pipe *pipe, const struct fd_fence *fence,
185                uint64_t timeout);
186    void (*destroy)(struct fd_pipe *pipe);
187 };
188 
189 struct fd_pipe_control {
190    uint32_t fence;
191 };
192 #define control_ptr(pipe, member) \
193    (pipe)->control_mem, offsetof(struct fd_pipe_control, member), 0, 0
194 
195 struct fd_pipe {
196    struct fd_device *dev;
197    enum fd_pipe_id id;
198    struct fd_dev_id dev_id;
199 
200    /**
201     * Note refcnt is *not* atomic, but protected by table_lock, since the
202     * table_lock is held in fd_bo_add_fence(), which is the hotpath.
203     */
204    int32_t refcnt;
205 
206    /**
207     * Previous fence seqno allocated for this pipe.  The fd_pipe represents
208     * a single timeline, fences allocated by this pipe can be compared to
209     * each other, but fences from different pipes are not comparable (as
210     * there could be preemption of multiple priority level submitqueues at
211     * play)
212     */
213    uint32_t last_fence;
214 
215    struct fd_bo *control_mem;
216    volatile struct fd_pipe_control *control;
217 
218    const struct fd_pipe_funcs *funcs;
219 };
220 
221 uint32_t fd_pipe_emit_fence(struct fd_pipe *pipe, struct fd_ringbuffer *ring);
222 
223 static inline void
fd_pipe_flush(struct fd_pipe * pipe,uint32_t fence)224 fd_pipe_flush(struct fd_pipe *pipe, uint32_t fence)
225 {
226    if (!pipe->funcs->flush)
227       return;
228    pipe->funcs->flush(pipe, fence);
229 }
230 
231 struct fd_submit_funcs {
232    struct fd_ringbuffer *(*new_ringbuffer)(struct fd_submit *submit,
233                                            uint32_t size,
234                                            enum fd_ringbuffer_flags flags);
235    int (*flush)(struct fd_submit *submit, int in_fence_fd,
236                 struct fd_submit_fence *out_fence);
237    void (*destroy)(struct fd_submit *submit);
238 };
239 
240 struct fd_submit {
241    int32_t refcnt;
242    struct fd_pipe *pipe;
243    const struct fd_submit_funcs *funcs;
244 
245    struct fd_ringbuffer *primary;
246    uint32_t fence;
247    struct list_head node;  /* node in fd_pipe::deferred_submits */
248 };
249 
250 static inline unsigned
fd_dev_count_deferred_cmds(struct fd_device * dev)251 fd_dev_count_deferred_cmds(struct fd_device *dev)
252 {
253    unsigned nr = 0;
254 
255    simple_mtx_assert_locked(&dev->submit_lock);
256 
257    list_for_each_entry (struct fd_submit, submit, &dev->deferred_submits, node) {
258       nr += fd_ringbuffer_cmd_count(submit->primary);
259    }
260 
261    return nr;
262 }
263 
264 struct fd_bo_funcs {
265    int (*offset)(struct fd_bo *bo, uint64_t *offset);
266    int (*cpu_prep)(struct fd_bo *bo, struct fd_pipe *pipe, uint32_t op);
267    void (*cpu_fini)(struct fd_bo *bo);
268    int (*madvise)(struct fd_bo *bo, int willneed);
269    uint64_t (*iova)(struct fd_bo *bo);
270    void (*set_name)(struct fd_bo *bo, const char *fmt, va_list ap);
271    void (*destroy)(struct fd_bo *bo);
272 };
273 
274 struct fd_bo_fence {
275    /* For non-shared buffers, track the last pipe the buffer was active
276     * on, and the per-pipe fence value that indicates when the buffer is
277     * idle:
278     */
279    uint32_t fence;
280    struct fd_pipe *pipe;
281 };
282 
283 struct fd_bo {
284    struct fd_device *dev;
285    uint32_t size;
286    uint32_t handle;
287    uint32_t name;
288    int32_t refcnt;
289    uint32_t reloc_flags; /* flags like FD_RELOC_DUMP to use for relocs to this BO */
290    uint32_t alloc_flags; /* flags that control allocation/mapping, ie. FD_BO_x */
291    uint64_t iova;
292    void *map;
293    const struct fd_bo_funcs *funcs;
294 
295    enum {
296       NO_CACHE = 0,
297       BO_CACHE = 1,
298       RING_CACHE = 2,
299    } bo_reuse : 2;
300 
301    /* Buffers that are shared (imported or exported) may be used in
302     * other processes, so we need to fallback to kernel to determine
303     * busyness.
304     */
305    bool shared : 1;
306 
307    /* We need to be able to disable userspace fence synchronization for
308     * special internal buffers, namely the pipe->control buffer, to avoid
309     * a circular reference loop.
310     */
311    bool nosync : 1;
312 
313    struct list_head list; /* bucket-list entry */
314    time_t free_time;      /* time when added to bucket-list */
315 
316    DECLARE_ARRAY(struct fd_bo_fence, fences);
317 
318    /* In the common case, there is no more than one fence attached.
319     * This provides storage for the fences table until it grows to
320     * be larger than a single element.
321     */
322    struct fd_bo_fence _inline_fence;
323 };
324 
325 void fd_bo_add_fence(struct fd_bo *bo, struct fd_pipe *pipe, uint32_t fence);
326 
327 enum fd_bo_state {
328    FD_BO_STATE_IDLE,
329    FD_BO_STATE_BUSY,
330    FD_BO_STATE_UNKNOWN,
331 };
332 enum fd_bo_state fd_bo_state(struct fd_bo *bo);
333 
334 struct fd_bo *fd_bo_new_ring(struct fd_device *dev, uint32_t size);
335 
336 #define enable_debug 0 /* TODO make dynamic */
337 
338 bool fd_dbg(void);
339 
340 #define INFO_MSG(fmt, ...)                                                     \
341    do {                                                                        \
342       if (fd_dbg())                                                            \
343          mesa_logi("%s:%d: " fmt, __FUNCTION__, __LINE__, ##__VA_ARGS__);      \
344    } while (0)
345 #define DEBUG_MSG(fmt, ...)                                                    \
346    do                                                                          \
347       if (enable_debug) {                                                      \
348          mesa_logd("%s:%d: " fmt, __FUNCTION__, __LINE__, ##__VA_ARGS__);      \
349       }                                                                        \
350    while (0)
351 #define WARN_MSG(fmt, ...)                                                     \
352    do {                                                                        \
353       mesa_logw("%s:%d: " fmt, __FUNCTION__, __LINE__, ##__VA_ARGS__);         \
354    } while (0)
355 #define ERROR_MSG(fmt, ...)                                                    \
356    do {                                                                        \
357       mesa_loge("%s:%d: " fmt, __FUNCTION__, __LINE__, ##__VA_ARGS__);         \
358    } while (0)
359 
360 #define U642VOID(x) ((void *)(unsigned long)(x))
361 #define VOID2U64(x) ((uint64_t)(unsigned long)(x))
362 
363 #if HAVE_VALGRIND
364 #include <memcheck.h>
365 
366 /*
367  * For tracking the backing memory (if valgrind enabled, we force a mmap
368  * for the purposes of tracking)
369  */
370 static inline void
VG_BO_ALLOC(struct fd_bo * bo)371 VG_BO_ALLOC(struct fd_bo *bo)
372 {
373    if (bo && RUNNING_ON_VALGRIND) {
374       VALGRIND_MALLOCLIKE_BLOCK(fd_bo_map(bo), bo->size, 0, 1);
375    }
376 }
377 
378 static inline void
VG_BO_FREE(struct fd_bo * bo)379 VG_BO_FREE(struct fd_bo *bo)
380 {
381    VALGRIND_FREELIKE_BLOCK(bo->map, 0);
382 }
383 
384 /*
385  * For tracking bo structs that are in the buffer-cache, so that valgrind
386  * doesn't attribute ownership to the first one to allocate the recycled
387  * bo.
388  *
389  * Note that the list_head in fd_bo is used to track the buffers in cache
390  * so disable error reporting on the range while they are in cache so
391  * valgrind doesn't squawk about list traversal.
392  *
393  */
394 static inline void
VG_BO_RELEASE(struct fd_bo * bo)395 VG_BO_RELEASE(struct fd_bo *bo)
396 {
397    if (RUNNING_ON_VALGRIND) {
398       VALGRIND_DISABLE_ADDR_ERROR_REPORTING_IN_RANGE(bo, bo->dev->bo_size);
399       VALGRIND_MAKE_MEM_NOACCESS(bo, bo->dev->bo_size);
400       VALGRIND_FREELIKE_BLOCK(bo->map, 0);
401    }
402 }
403 static inline void
VG_BO_OBTAIN(struct fd_bo * bo)404 VG_BO_OBTAIN(struct fd_bo *bo)
405 {
406    if (RUNNING_ON_VALGRIND) {
407       VALGRIND_MAKE_MEM_DEFINED(bo, bo->dev->bo_size);
408       VALGRIND_ENABLE_ADDR_ERROR_REPORTING_IN_RANGE(bo, bo->dev->bo_size);
409       VALGRIND_MALLOCLIKE_BLOCK(bo->map, bo->size, 0, 1);
410    }
411 }
412 #else
413 static inline void
VG_BO_ALLOC(struct fd_bo * bo)414 VG_BO_ALLOC(struct fd_bo *bo)
415 {
416 }
417 static inline void
VG_BO_FREE(struct fd_bo * bo)418 VG_BO_FREE(struct fd_bo *bo)
419 {
420 }
421 static inline void
VG_BO_RELEASE(struct fd_bo * bo)422 VG_BO_RELEASE(struct fd_bo *bo)
423 {
424 }
425 static inline void
VG_BO_OBTAIN(struct fd_bo * bo)426 VG_BO_OBTAIN(struct fd_bo *bo)
427 {
428 }
429 #endif
430 
431 #define FD_DEFINE_CAST(parent, child)                                          \
432    static inline struct child *to_##child(struct parent *x)                    \
433    {                                                                           \
434       return (struct child *)x;                                                \
435    }
436 
437 #endif /* FREEDRENO_PRIV_H_ */
438