1 /*
2  * Copyright (C) 2013 Red Hat
3  * Author: Rob Clark <robdclark@gmail.com>
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 
25 #ifndef __MSM_DRM_H__
26 #define __MSM_DRM_H__
27 
28 #include "drm.h"
29 
30 #if defined(__cplusplus)
31 extern "C" {
32 #endif
33 
34 /* Please note that modifications to all structs defined here are
35  * subject to backwards-compatibility constraints:
36  *  1) Do not use pointers, use uint64_t instead for 32 bit / 64 bit
37  *     user/kernel compatibility
38  *  2) Keep fields aligned to their size
39  *  3) Because of how drm_ioctl() works, we can add new fields at
40  *     the end of an ioctl if some care is taken: drm_ioctl() will
41  *     zero out the new fields at the tail of the ioctl, so a zero
42  *     value should have a backwards compatible meaning.  And for
43  *     output params, userspace won't see the newly added output
44  *     fields.. so that has to be somehow ok.
45  */
46 
47 #define MSM_PIPE_NONE        0x00
48 #define MSM_PIPE_2D0         0x01
49 #define MSM_PIPE_2D1         0x02
50 #define MSM_PIPE_3D0         0x10
51 
52 /* The pipe-id just uses the lower bits, so can be OR'd with flags in
53  * the upper 16 bits (which could be extended further, if needed, maybe
54  * we extend/overload the pipe-id some day to deal with multiple rings,
55  * but even then I don't think we need the full lower 16 bits).
56  */
57 #define MSM_PIPE_ID_MASK     0xffff
58 #define MSM_PIPE_ID(x)       ((x) & MSM_PIPE_ID_MASK)
59 #define MSM_PIPE_FLAGS(x)    ((x) & ~MSM_PIPE_ID_MASK)
60 
61 /* timeouts are specified in clock-monotonic absolute times (to simplify
62  * restarting interrupted ioctls).  The following struct is logically the
63  * same as 'struct timespec' but 32/64b ABI safe.
64  */
65 struct drm_msm_timespec {
66 	int64_t tv_sec;          /* seconds */
67 	int64_t tv_nsec;         /* nanoseconds */
68 };
69 
70 #define MSM_PARAM_GPU_ID     0x01
71 #define MSM_PARAM_GMEM_SIZE  0x02
72 #define MSM_PARAM_CHIP_ID    0x03
73 #define MSM_PARAM_MAX_FREQ   0x04
74 #define MSM_PARAM_TIMESTAMP  0x05
75 #define MSM_PARAM_GMEM_BASE  0x06
76 #define MSM_PARAM_NR_RINGS   0x07
77 #define MSM_PARAM_PP_PGTABLE 0x08  /* => 1 for per-process pagetables, else 0 */
78 #define MSM_PARAM_FAULTS     0x09
79 #define MSM_PARAM_SUSPENDS   0x0a
80 
81 struct drm_msm_param {
82 	uint32_t pipe;           /* in, MSM_PIPE_x */
83 	uint32_t param;          /* in, MSM_PARAM_x */
84 	uint64_t value;          /* out (get_param) or in (set_param) */
85 };
86 
87 /*
88  * GEM buffers:
89  */
90 
91 #define MSM_BO_SCANOUT       0x00000001     /* scanout capable */
92 #define MSM_BO_GPU_READONLY  0x00000002
93 #define MSM_BO_CACHE_MASK    0x000f0000
94 /* cache modes */
95 #define MSM_BO_CACHED        0x00010000
96 #define MSM_BO_WC            0x00020000
97 #define MSM_BO_UNCACHED      0x00040000
98 
99 #define MSM_BO_FLAGS         (MSM_BO_SCANOUT | \
100                               MSM_BO_GPU_READONLY | \
101                               MSM_BO_CACHED | \
102                               MSM_BO_WC | \
103                               MSM_BO_UNCACHED)
104 
105 struct drm_msm_gem_new {
106 	uint64_t size;           /* in */
107 	uint32_t flags;          /* in, mask of MSM_BO_x */
108 	uint32_t handle;         /* out */
109 };
110 
111 /* Get or set GEM buffer info.  The requested value can be passed
112  * directly in 'value', or for data larger than 64b 'value' is a
113  * pointer to userspace buffer, with 'len' specifying the number of
114  * bytes copied into that buffer.  For info returned by pointer,
115  * calling the GEM_INFO ioctl with null 'value' will return the
116  * required buffer size in 'len'
117  */
118 #define MSM_INFO_GET_OFFSET	0x00   /* get mmap() offset, returned by value */
119 #define MSM_INFO_GET_IOVA	0x01   /* get iova, returned by value */
120 #define MSM_INFO_SET_NAME	0x02   /* set the debug name (by pointer) */
121 #define MSM_INFO_GET_NAME	0x03   /* get debug name, returned by pointer */
122 
123 struct drm_msm_gem_info {
124 	uint32_t handle;         /* in */
125 	uint32_t info;           /* in - one of MSM_INFO_* */
126 	uint64_t value;          /* in or out */
127 	uint32_t len;            /* in or out */
128 	uint32_t pad;
129 };
130 
131 #define MSM_PREP_READ        0x01
132 #define MSM_PREP_WRITE       0x02
133 #define MSM_PREP_NOSYNC      0x04
134 
135 #define MSM_PREP_FLAGS       (MSM_PREP_READ | MSM_PREP_WRITE | MSM_PREP_NOSYNC)
136 
137 struct drm_msm_gem_cpu_prep {
138 	uint32_t handle;         /* in */
139 	uint32_t op;             /* in, mask of MSM_PREP_x */
140 	struct drm_msm_timespec timeout;   /* in */
141 };
142 
143 struct drm_msm_gem_cpu_fini {
144 	uint32_t handle;         /* in */
145 };
146 
147 /*
148  * Cmdstream Submission:
149  */
150 
151 /* The value written into the cmdstream is logically:
152  *
153  *   ((relocbuf->gpuaddr + reloc_offset) << shift) | or
154  *
155  * When we have GPU's w/ >32bit ptrs, it should be possible to deal
156  * with this by emit'ing two reloc entries with appropriate shift
157  * values.  Or a new MSM_SUBMIT_CMD_x type would also be an option.
158  *
159  * NOTE that reloc's must be sorted by order of increasing submit_offset,
160  * otherwise EINVAL.
161  */
162 struct drm_msm_gem_submit_reloc {
163 	uint32_t submit_offset;  /* in, offset from submit_bo */
164 	uint32_t or;             /* in, value OR'd with result */
165 	int32_t shift;          /* in, amount of left shift (can be negative) */
166 	uint32_t reloc_idx;      /* in, index of reloc_bo buffer */
167 	uint64_t reloc_offset;   /* in, offset from start of reloc_bo */
168 };
169 
170 /* submit-types:
171  *   BUF - this cmd buffer is executed normally.
172  *   IB_TARGET_BUF - this cmd buffer is an IB target.  Reloc's are
173  *      processed normally, but the kernel does not setup an IB to
174  *      this buffer in the first-level ringbuffer
175  *   CTX_RESTORE_BUF - only executed if there has been a GPU context
176  *      switch since the last SUBMIT ioctl
177  */
178 #define MSM_SUBMIT_CMD_BUF             0x0001
179 #define MSM_SUBMIT_CMD_IB_TARGET_BUF   0x0002
180 #define MSM_SUBMIT_CMD_CTX_RESTORE_BUF 0x0003
181 struct drm_msm_gem_submit_cmd {
182 	uint32_t type;           /* in, one of MSM_SUBMIT_CMD_x */
183 	uint32_t submit_idx;     /* in, index of submit_bo cmdstream buffer */
184 	uint32_t submit_offset;  /* in, offset into submit_bo */
185 	uint32_t size;           /* in, cmdstream size */
186 	uint32_t pad;
187 	uint32_t nr_relocs;      /* in, number of submit_reloc's */
188 	uint64_t relocs;         /* in, ptr to array of submit_reloc's */
189 };
190 
191 /* Each buffer referenced elsewhere in the cmdstream submit (ie. the
192  * cmdstream buffer(s) themselves or reloc entries) has one (and only
193  * one) entry in the submit->bos[] table.
194  *
195  * As a optimization, the current buffer (gpu virtual address) can be
196  * passed back through the 'presumed' field.  If on a subsequent reloc,
197  * userspace passes back a 'presumed' address that is still valid,
198  * then patching the cmdstream for this entry is skipped.  This can
199  * avoid kernel needing to map/access the cmdstream bo in the common
200  * case.
201  */
202 #define MSM_SUBMIT_BO_READ             0x0001
203 #define MSM_SUBMIT_BO_WRITE            0x0002
204 #define MSM_SUBMIT_BO_DUMP             0x0004
205 
206 #define MSM_SUBMIT_BO_FLAGS            (MSM_SUBMIT_BO_READ | \
207 					MSM_SUBMIT_BO_WRITE | \
208 					MSM_SUBMIT_BO_DUMP)
209 
210 struct drm_msm_gem_submit_bo {
211 	uint32_t flags;          /* in, mask of MSM_SUBMIT_BO_x */
212 	uint32_t handle;         /* in, GEM handle */
213 	uint64_t presumed;       /* in/out, presumed buffer address */
214 };
215 
216 /* Valid submit ioctl flags: */
217 #define MSM_SUBMIT_NO_IMPLICIT   0x80000000 /* disable implicit sync */
218 #define MSM_SUBMIT_FENCE_FD_IN   0x40000000 /* enable input fence_fd */
219 #define MSM_SUBMIT_FENCE_FD_OUT  0x20000000 /* enable output fence_fd */
220 #define MSM_SUBMIT_SUDO          0x10000000 /* run submitted cmds from RB */
221 #define MSM_SUBMIT_SYNCOBJ_IN    0x08000000 /* enable input syncobj */
222 #define MSM_SUBMIT_SYNCOBJ_OUT   0x04000000 /* enable output syncobj */
223 #define MSM_SUBMIT_FLAGS                ( \
224 		MSM_SUBMIT_NO_IMPLICIT   | \
225 		MSM_SUBMIT_FENCE_FD_IN   | \
226 		MSM_SUBMIT_FENCE_FD_OUT  | \
227 		MSM_SUBMIT_SUDO          | \
228 		MSM_SUBMIT_SYNCOBJ_IN    | \
229 		MSM_SUBMIT_SYNCOBJ_OUT   | \
230 		0)
231 
232 #define MSM_SUBMIT_SYNCOBJ_RESET 0x00000001 /* Reset syncobj after wait. */
233 #define MSM_SUBMIT_SYNCOBJ_FLAGS        ( \
234 		MSM_SUBMIT_SYNCOBJ_RESET | \
235 		0)
236 
237 struct drm_msm_gem_submit_syncobj {
238 	uint32_t handle;     /* in, syncobj handle. */
239 	uint32_t flags;      /* in, from MSM_SUBMIT_SYNCOBJ_FLAGS */
240 	uint64_t point;      /* in, timepoint for timeline syncobjs. */
241 };
242 
243 /* Each cmdstream submit consists of a table of buffers involved, and
244  * one or more cmdstream buffers.  This allows for conditional execution
245  * (context-restore), and IB buffers needed for per tile/bin draw cmds.
246  */
247 struct drm_msm_gem_submit {
248 	uint32_t flags;          /* MSM_PIPE_x | MSM_SUBMIT_x */
249 	uint32_t fence;          /* out */
250 	uint32_t nr_bos;         /* in, number of submit_bo's */
251 	uint32_t nr_cmds;        /* in, number of submit_cmd's */
252 	uint64_t bos;            /* in, ptr to array of submit_bo's */
253 	uint64_t cmds;           /* in, ptr to array of submit_cmd's */
254 	int32_t fence_fd;       /* in/out fence fd (see MSM_SUBMIT_FENCE_FD_IN/OUT) */
255 	uint32_t queueid;        /* in, submitqueue id */
256 	uint64_t in_syncobjs;    /* in, ptr to array of drm_msm_gem_submit_syncobj */
257 	uint64_t out_syncobjs;   /* in, ptr to array of drm_msm_gem_submit_syncobj */
258 	uint32_t nr_in_syncobjs; /* in, number of entries in in_syncobj */
259 	uint32_t nr_out_syncobjs; /* in, number of entries in out_syncobj. */
260 	uint32_t syncobj_stride; /* in, stride of syncobj arrays. */
261 	uint32_t pad;            /*in, reserved for future use, always 0. */
262 
263 };
264 
265 /* The normal way to synchronize with the GPU is just to CPU_PREP on
266  * a buffer if you need to access it from the CPU (other cmdstream
267  * submission from same or other contexts, PAGE_FLIP ioctl, etc, all
268  * handle the required synchronization under the hood).  This ioctl
269  * mainly just exists as a way to implement the gallium pipe_fence
270  * APIs without requiring a dummy bo to synchronize on.
271  */
272 struct drm_msm_wait_fence {
273 	uint32_t fence;          /* in */
274 	uint32_t pad;
275 	struct drm_msm_timespec timeout;   /* in */
276 	uint32_t queueid;         /* in, submitqueue id */
277 };
278 
279 /* madvise provides a way to tell the kernel in case a buffers contents
280  * can be discarded under memory pressure, which is useful for userspace
281  * bo cache where we want to optimistically hold on to buffer allocate
282  * and potential mmap, but allow the pages to be discarded under memory
283  * pressure.
284  *
285  * Typical usage would involve madvise(DONTNEED) when buffer enters BO
286  * cache, and madvise(WILLNEED) if trying to recycle buffer from BO cache.
287  * In the WILLNEED case, 'retained' indicates to userspace whether the
288  * backing pages still exist.
289  */
290 #define MSM_MADV_WILLNEED 0       /* backing pages are needed, status returned in 'retained' */
291 #define MSM_MADV_DONTNEED 1       /* backing pages not needed */
292 #define __MSM_MADV_PURGED 2       /* internal state */
293 
294 struct drm_msm_gem_madvise {
295 	uint32_t handle;         /* in, GEM handle */
296 	uint32_t madv;           /* in, MSM_MADV_x */
297 	uint32_t retained;       /* out, whether backing store still exists */
298 };
299 
300 /*
301  * Draw queues allow the user to set specific submission parameter. Command
302  * submissions specify a specific submitqueue to use.  ID 0 is reserved for
303  * backwards compatibility as a "default" submitqueue
304  */
305 
306 #define MSM_SUBMITQUEUE_FLAGS (0)
307 
308 struct drm_msm_submitqueue {
309 	uint32_t flags;   /* in, MSM_SUBMITQUEUE_x */
310 	uint32_t prio;    /* in, Priority level */
311 	uint32_t id;      /* out, identifier */
312 };
313 
314 #define MSM_SUBMITQUEUE_PARAM_FAULTS   0
315 
316 struct drm_msm_submitqueue_query {
317 	uint64_t data;
318 	uint32_t id;
319 	uint32_t param;
320 	uint32_t len;
321 	uint32_t pad;
322 };
323 
324 #define DRM_MSM_GET_PARAM              0x00
325 /* placeholder:
326 #define DRM_MSM_SET_PARAM              0x01
327  */
328 #define DRM_MSM_GEM_NEW                0x02
329 #define DRM_MSM_GEM_INFO               0x03
330 #define DRM_MSM_GEM_CPU_PREP           0x04
331 #define DRM_MSM_GEM_CPU_FINI           0x05
332 #define DRM_MSM_GEM_SUBMIT             0x06
333 #define DRM_MSM_WAIT_FENCE             0x07
334 #define DRM_MSM_GEM_MADVISE            0x08
335 /* placeholder:
336 #define DRM_MSM_GEM_SVM_NEW            0x09
337  */
338 #define DRM_MSM_SUBMITQUEUE_NEW        0x0A
339 #define DRM_MSM_SUBMITQUEUE_CLOSE      0x0B
340 #define DRM_MSM_SUBMITQUEUE_QUERY      0x0C
341 
342 #define DRM_IOCTL_MSM_GET_PARAM        DRM_IOWR(DRM_COMMAND_BASE + DRM_MSM_GET_PARAM, struct drm_msm_param)
343 #define DRM_IOCTL_MSM_GEM_NEW          DRM_IOWR(DRM_COMMAND_BASE + DRM_MSM_GEM_NEW, struct drm_msm_gem_new)
344 #define DRM_IOCTL_MSM_GEM_INFO         DRM_IOWR(DRM_COMMAND_BASE + DRM_MSM_GEM_INFO, struct drm_msm_gem_info)
345 #define DRM_IOCTL_MSM_GEM_CPU_PREP     DRM_IOW (DRM_COMMAND_BASE + DRM_MSM_GEM_CPU_PREP, struct drm_msm_gem_cpu_prep)
346 #define DRM_IOCTL_MSM_GEM_CPU_FINI     DRM_IOW (DRM_COMMAND_BASE + DRM_MSM_GEM_CPU_FINI, struct drm_msm_gem_cpu_fini)
347 #define DRM_IOCTL_MSM_GEM_SUBMIT       DRM_IOWR(DRM_COMMAND_BASE + DRM_MSM_GEM_SUBMIT, struct drm_msm_gem_submit)
348 #define DRM_IOCTL_MSM_WAIT_FENCE       DRM_IOW (DRM_COMMAND_BASE + DRM_MSM_WAIT_FENCE, struct drm_msm_wait_fence)
349 #define DRM_IOCTL_MSM_GEM_MADVISE      DRM_IOWR(DRM_COMMAND_BASE + DRM_MSM_GEM_MADVISE, struct drm_msm_gem_madvise)
350 #define DRM_IOCTL_MSM_SUBMITQUEUE_NEW    DRM_IOWR(DRM_COMMAND_BASE + DRM_MSM_SUBMITQUEUE_NEW, struct drm_msm_submitqueue)
351 #define DRM_IOCTL_MSM_SUBMITQUEUE_CLOSE  DRM_IOW (DRM_COMMAND_BASE + DRM_MSM_SUBMITQUEUE_CLOSE, uint32_t)
352 #define DRM_IOCTL_MSM_SUBMITQUEUE_QUERY  DRM_IOW (DRM_COMMAND_BASE + DRM_MSM_SUBMITQUEUE_QUERY, struct drm_msm_submitqueue_query)
353 
354 #if defined(__cplusplus)
355 }
356 #endif
357 
358 #endif /* __MSM_DRM_H__ */
359