Loading...
1/*
2 * Copyright © 2013 Intel Corporation
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Brad Volkin <bradley.d.volkin@intel.com>
25 *
26 */
27
28#include "i915_drv.h"
29
30/**
31 * DOC: batch buffer command parser
32 *
33 * Motivation:
34 * Certain OpenGL features (e.g. transform feedback, performance monitoring)
35 * require userspace code to submit batches containing commands such as
36 * MI_LOAD_REGISTER_IMM to access various registers. Unfortunately, some
37 * generations of the hardware will noop these commands in "unsecure" batches
38 * (which includes all userspace batches submitted via i915) even though the
39 * commands may be safe and represent the intended programming model of the
40 * device.
41 *
42 * The software command parser is similar in operation to the command parsing
43 * done in hardware for unsecure batches. However, the software parser allows
44 * some operations that would be noop'd by hardware, if the parser determines
45 * the operation is safe, and submits the batch as "secure" to prevent hardware
46 * parsing.
47 *
48 * Threats:
49 * At a high level, the hardware (and software) checks attempt to prevent
50 * granting userspace undue privileges. There are three categories of privilege.
51 *
52 * First, commands which are explicitly defined as privileged or which should
53 * only be used by the kernel driver. The parser generally rejects such
54 * commands, though it may allow some from the drm master process.
55 *
56 * Second, commands which access registers. To support correct/enhanced
57 * userspace functionality, particularly certain OpenGL extensions, the parser
58 * provides a whitelist of registers which userspace may safely access (for both
59 * normal and drm master processes).
60 *
61 * Third, commands which access privileged memory (i.e. GGTT, HWS page, etc).
62 * The parser always rejects such commands.
63 *
64 * The majority of the problematic commands fall in the MI_* range, with only a
65 * few specific commands on each ring (e.g. PIPE_CONTROL and MI_FLUSH_DW).
66 *
67 * Implementation:
68 * Each ring maintains tables of commands and registers which the parser uses in
69 * scanning batch buffers submitted to that ring.
70 *
71 * Since the set of commands that the parser must check for is significantly
72 * smaller than the number of commands supported, the parser tables contain only
73 * those commands required by the parser. This generally works because command
74 * opcode ranges have standard command length encodings. So for commands that
75 * the parser does not need to check, it can easily skip them. This is
76 * implemented via a per-ring length decoding vfunc.
77 *
78 * Unfortunately, there are a number of commands that do not follow the standard
79 * length encoding for their opcode range, primarily amongst the MI_* commands.
80 * To handle this, the parser provides a way to define explicit "skip" entries
81 * in the per-ring command tables.
82 *
83 * Other command table entries map fairly directly to high level categories
84 * mentioned above: rejected, master-only, register whitelist. The parser
85 * implements a number of checks, including the privileged memory checks, via a
86 * general bitmasking mechanism.
87 */
88
89#define STD_MI_OPCODE_MASK 0xFF800000
90#define STD_3D_OPCODE_MASK 0xFFFF0000
91#define STD_2D_OPCODE_MASK 0xFFC00000
92#define STD_MFX_OPCODE_MASK 0xFFFF0000
93
94#define CMD(op, opm, f, lm, fl, ...) \
95 { \
96 .flags = (fl) | ((f) ? CMD_DESC_FIXED : 0), \
97 .cmd = { (op), (opm) }, \
98 .length = { (lm) }, \
99 __VA_ARGS__ \
100 }
101
102/* Convenience macros to compress the tables */
103#define SMI STD_MI_OPCODE_MASK
104#define S3D STD_3D_OPCODE_MASK
105#define S2D STD_2D_OPCODE_MASK
106#define SMFX STD_MFX_OPCODE_MASK
107#define F true
108#define S CMD_DESC_SKIP
109#define R CMD_DESC_REJECT
110#define W CMD_DESC_REGISTER
111#define B CMD_DESC_BITMASK
112#define M CMD_DESC_MASTER
113
114/* Command Mask Fixed Len Action
115 ---------------------------------------------------------- */
116static const struct drm_i915_cmd_descriptor common_cmds[] = {
117 CMD( MI_NOOP, SMI, F, 1, S ),
118 CMD( MI_USER_INTERRUPT, SMI, F, 1, R ),
119 CMD( MI_WAIT_FOR_EVENT, SMI, F, 1, M ),
120 CMD( MI_ARB_CHECK, SMI, F, 1, S ),
121 CMD( MI_REPORT_HEAD, SMI, F, 1, S ),
122 CMD( MI_SUSPEND_FLUSH, SMI, F, 1, S ),
123 CMD( MI_SEMAPHORE_MBOX, SMI, !F, 0xFF, R ),
124 CMD( MI_STORE_DWORD_INDEX, SMI, !F, 0xFF, R ),
125 CMD( MI_LOAD_REGISTER_IMM(1), SMI, !F, 0xFF, W,
126 .reg = { .offset = 1, .mask = 0x007FFFFC, .step = 2 } ),
127 CMD( MI_STORE_REGISTER_MEM, SMI, F, 3, W | B,
128 .reg = { .offset = 1, .mask = 0x007FFFFC },
129 .bits = {{
130 .offset = 0,
131 .mask = MI_GLOBAL_GTT,
132 .expected = 0,
133 }}, ),
134 CMD( MI_LOAD_REGISTER_MEM, SMI, F, 3, W | B,
135 .reg = { .offset = 1, .mask = 0x007FFFFC },
136 .bits = {{
137 .offset = 0,
138 .mask = MI_GLOBAL_GTT,
139 .expected = 0,
140 }}, ),
141 /*
142 * MI_BATCH_BUFFER_START requires some special handling. It's not
143 * really a 'skip' action but it doesn't seem like it's worth adding
144 * a new action. See i915_parse_cmds().
145 */
146 CMD( MI_BATCH_BUFFER_START, SMI, !F, 0xFF, S ),
147};
148
149static const struct drm_i915_cmd_descriptor render_cmds[] = {
150 CMD( MI_FLUSH, SMI, F, 1, S ),
151 CMD( MI_ARB_ON_OFF, SMI, F, 1, R ),
152 CMD( MI_PREDICATE, SMI, F, 1, S ),
153 CMD( MI_TOPOLOGY_FILTER, SMI, F, 1, S ),
154 CMD( MI_SET_APPID, SMI, F, 1, S ),
155 CMD( MI_DISPLAY_FLIP, SMI, !F, 0xFF, R ),
156 CMD( MI_SET_CONTEXT, SMI, !F, 0xFF, R ),
157 CMD( MI_URB_CLEAR, SMI, !F, 0xFF, S ),
158 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0x3F, B,
159 .bits = {{
160 .offset = 0,
161 .mask = MI_GLOBAL_GTT,
162 .expected = 0,
163 }}, ),
164 CMD( MI_UPDATE_GTT, SMI, !F, 0xFF, R ),
165 CMD( MI_CLFLUSH, SMI, !F, 0x3FF, B,
166 .bits = {{
167 .offset = 0,
168 .mask = MI_GLOBAL_GTT,
169 .expected = 0,
170 }}, ),
171 CMD( MI_REPORT_PERF_COUNT, SMI, !F, 0x3F, B,
172 .bits = {{
173 .offset = 1,
174 .mask = MI_REPORT_PERF_COUNT_GGTT,
175 .expected = 0,
176 }}, ),
177 CMD( MI_CONDITIONAL_BATCH_BUFFER_END, SMI, !F, 0xFF, B,
178 .bits = {{
179 .offset = 0,
180 .mask = MI_GLOBAL_GTT,
181 .expected = 0,
182 }}, ),
183 CMD( GFX_OP_3DSTATE_VF_STATISTICS, S3D, F, 1, S ),
184 CMD( PIPELINE_SELECT, S3D, F, 1, S ),
185 CMD( MEDIA_VFE_STATE, S3D, !F, 0xFFFF, B,
186 .bits = {{
187 .offset = 2,
188 .mask = MEDIA_VFE_STATE_MMIO_ACCESS_MASK,
189 .expected = 0,
190 }}, ),
191 CMD( GPGPU_OBJECT, S3D, !F, 0xFF, S ),
192 CMD( GPGPU_WALKER, S3D, !F, 0xFF, S ),
193 CMD( GFX_OP_3DSTATE_SO_DECL_LIST, S3D, !F, 0x1FF, S ),
194 CMD( GFX_OP_PIPE_CONTROL(5), S3D, !F, 0xFF, B,
195 .bits = {{
196 .offset = 1,
197 .mask = (PIPE_CONTROL_MMIO_WRITE | PIPE_CONTROL_NOTIFY),
198 .expected = 0,
199 },
200 {
201 .offset = 1,
202 .mask = (PIPE_CONTROL_GLOBAL_GTT_IVB |
203 PIPE_CONTROL_STORE_DATA_INDEX),
204 .expected = 0,
205 .condition_offset = 1,
206 .condition_mask = PIPE_CONTROL_POST_SYNC_OP_MASK,
207 }}, ),
208};
209
210static const struct drm_i915_cmd_descriptor hsw_render_cmds[] = {
211 CMD( MI_SET_PREDICATE, SMI, F, 1, S ),
212 CMD( MI_RS_CONTROL, SMI, F, 1, S ),
213 CMD( MI_URB_ATOMIC_ALLOC, SMI, F, 1, S ),
214 CMD( MI_SET_APPID, SMI, F, 1, S ),
215 CMD( MI_RS_CONTEXT, SMI, F, 1, S ),
216 CMD( MI_LOAD_SCAN_LINES_INCL, SMI, !F, 0x3F, M ),
217 CMD( MI_LOAD_SCAN_LINES_EXCL, SMI, !F, 0x3F, R ),
218 CMD( MI_LOAD_REGISTER_REG, SMI, !F, 0xFF, R ),
219 CMD( MI_RS_STORE_DATA_IMM, SMI, !F, 0xFF, S ),
220 CMD( MI_LOAD_URB_MEM, SMI, !F, 0xFF, S ),
221 CMD( MI_STORE_URB_MEM, SMI, !F, 0xFF, S ),
222 CMD( GFX_OP_3DSTATE_DX9_CONSTANTF_VS, S3D, !F, 0x7FF, S ),
223 CMD( GFX_OP_3DSTATE_DX9_CONSTANTF_PS, S3D, !F, 0x7FF, S ),
224
225 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_VS, S3D, !F, 0x1FF, S ),
226 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_GS, S3D, !F, 0x1FF, S ),
227 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_HS, S3D, !F, 0x1FF, S ),
228 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_DS, S3D, !F, 0x1FF, S ),
229 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_PS, S3D, !F, 0x1FF, S ),
230};
231
232static const struct drm_i915_cmd_descriptor video_cmds[] = {
233 CMD( MI_ARB_ON_OFF, SMI, F, 1, R ),
234 CMD( MI_SET_APPID, SMI, F, 1, S ),
235 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0xFF, B,
236 .bits = {{
237 .offset = 0,
238 .mask = MI_GLOBAL_GTT,
239 .expected = 0,
240 }}, ),
241 CMD( MI_UPDATE_GTT, SMI, !F, 0x3F, R ),
242 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, B,
243 .bits = {{
244 .offset = 0,
245 .mask = MI_FLUSH_DW_NOTIFY,
246 .expected = 0,
247 },
248 {
249 .offset = 1,
250 .mask = MI_FLUSH_DW_USE_GTT,
251 .expected = 0,
252 .condition_offset = 0,
253 .condition_mask = MI_FLUSH_DW_OP_MASK,
254 },
255 {
256 .offset = 0,
257 .mask = MI_FLUSH_DW_STORE_INDEX,
258 .expected = 0,
259 .condition_offset = 0,
260 .condition_mask = MI_FLUSH_DW_OP_MASK,
261 }}, ),
262 CMD( MI_CONDITIONAL_BATCH_BUFFER_END, SMI, !F, 0xFF, B,
263 .bits = {{
264 .offset = 0,
265 .mask = MI_GLOBAL_GTT,
266 .expected = 0,
267 }}, ),
268 /*
269 * MFX_WAIT doesn't fit the way we handle length for most commands.
270 * It has a length field but it uses a non-standard length bias.
271 * It is always 1 dword though, so just treat it as fixed length.
272 */
273 CMD( MFX_WAIT, SMFX, F, 1, S ),
274};
275
276static const struct drm_i915_cmd_descriptor vecs_cmds[] = {
277 CMD( MI_ARB_ON_OFF, SMI, F, 1, R ),
278 CMD( MI_SET_APPID, SMI, F, 1, S ),
279 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0xFF, B,
280 .bits = {{
281 .offset = 0,
282 .mask = MI_GLOBAL_GTT,
283 .expected = 0,
284 }}, ),
285 CMD( MI_UPDATE_GTT, SMI, !F, 0x3F, R ),
286 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, B,
287 .bits = {{
288 .offset = 0,
289 .mask = MI_FLUSH_DW_NOTIFY,
290 .expected = 0,
291 },
292 {
293 .offset = 1,
294 .mask = MI_FLUSH_DW_USE_GTT,
295 .expected = 0,
296 .condition_offset = 0,
297 .condition_mask = MI_FLUSH_DW_OP_MASK,
298 },
299 {
300 .offset = 0,
301 .mask = MI_FLUSH_DW_STORE_INDEX,
302 .expected = 0,
303 .condition_offset = 0,
304 .condition_mask = MI_FLUSH_DW_OP_MASK,
305 }}, ),
306 CMD( MI_CONDITIONAL_BATCH_BUFFER_END, SMI, !F, 0xFF, B,
307 .bits = {{
308 .offset = 0,
309 .mask = MI_GLOBAL_GTT,
310 .expected = 0,
311 }}, ),
312};
313
314static const struct drm_i915_cmd_descriptor blt_cmds[] = {
315 CMD( MI_DISPLAY_FLIP, SMI, !F, 0xFF, R ),
316 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0x3FF, B,
317 .bits = {{
318 .offset = 0,
319 .mask = MI_GLOBAL_GTT,
320 .expected = 0,
321 }}, ),
322 CMD( MI_UPDATE_GTT, SMI, !F, 0x3F, R ),
323 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, B,
324 .bits = {{
325 .offset = 0,
326 .mask = MI_FLUSH_DW_NOTIFY,
327 .expected = 0,
328 },
329 {
330 .offset = 1,
331 .mask = MI_FLUSH_DW_USE_GTT,
332 .expected = 0,
333 .condition_offset = 0,
334 .condition_mask = MI_FLUSH_DW_OP_MASK,
335 },
336 {
337 .offset = 0,
338 .mask = MI_FLUSH_DW_STORE_INDEX,
339 .expected = 0,
340 .condition_offset = 0,
341 .condition_mask = MI_FLUSH_DW_OP_MASK,
342 }}, ),
343 CMD( COLOR_BLT, S2D, !F, 0x3F, S ),
344 CMD( SRC_COPY_BLT, S2D, !F, 0x3F, S ),
345};
346
347static const struct drm_i915_cmd_descriptor hsw_blt_cmds[] = {
348 CMD( MI_LOAD_SCAN_LINES_INCL, SMI, !F, 0x3F, M ),
349 CMD( MI_LOAD_SCAN_LINES_EXCL, SMI, !F, 0x3F, R ),
350};
351
352#undef CMD
353#undef SMI
354#undef S3D
355#undef S2D
356#undef SMFX
357#undef F
358#undef S
359#undef R
360#undef W
361#undef B
362#undef M
363
364static const struct drm_i915_cmd_table gen7_render_cmds[] = {
365 { common_cmds, ARRAY_SIZE(common_cmds) },
366 { render_cmds, ARRAY_SIZE(render_cmds) },
367};
368
369static const struct drm_i915_cmd_table hsw_render_ring_cmds[] = {
370 { common_cmds, ARRAY_SIZE(common_cmds) },
371 { render_cmds, ARRAY_SIZE(render_cmds) },
372 { hsw_render_cmds, ARRAY_SIZE(hsw_render_cmds) },
373};
374
375static const struct drm_i915_cmd_table gen7_video_cmds[] = {
376 { common_cmds, ARRAY_SIZE(common_cmds) },
377 { video_cmds, ARRAY_SIZE(video_cmds) },
378};
379
380static const struct drm_i915_cmd_table hsw_vebox_cmds[] = {
381 { common_cmds, ARRAY_SIZE(common_cmds) },
382 { vecs_cmds, ARRAY_SIZE(vecs_cmds) },
383};
384
385static const struct drm_i915_cmd_table gen7_blt_cmds[] = {
386 { common_cmds, ARRAY_SIZE(common_cmds) },
387 { blt_cmds, ARRAY_SIZE(blt_cmds) },
388};
389
390static const struct drm_i915_cmd_table hsw_blt_ring_cmds[] = {
391 { common_cmds, ARRAY_SIZE(common_cmds) },
392 { blt_cmds, ARRAY_SIZE(blt_cmds) },
393 { hsw_blt_cmds, ARRAY_SIZE(hsw_blt_cmds) },
394};
395
396/*
397 * Register whitelists, sorted by increasing register offset.
398 */
399
400/*
401 * An individual whitelist entry granting access to register addr. If
402 * mask is non-zero the argument of immediate register writes will be
403 * AND-ed with mask, and the command will be rejected if the result
404 * doesn't match value.
405 *
406 * Registers with non-zero mask are only allowed to be written using
407 * LRI.
408 */
409struct drm_i915_reg_descriptor {
410 i915_reg_t addr;
411 u32 mask;
412 u32 value;
413};
414
415/* Convenience macro for adding 32-bit registers. */
416#define REG32(_reg, ...) \
417 { .addr = (_reg), __VA_ARGS__ }
418
419/*
420 * Convenience macro for adding 64-bit registers.
421 *
422 * Some registers that userspace accesses are 64 bits. The register
423 * access commands only allow 32-bit accesses. Hence, we have to include
424 * entries for both halves of the 64-bit registers.
425 */
426#define REG64(_reg) \
427 { .addr = _reg }, \
428 { .addr = _reg ## _UDW }
429
430#define REG64_IDX(_reg, idx) \
431 { .addr = _reg(idx) }, \
432 { .addr = _reg ## _UDW(idx) }
433
434static const struct drm_i915_reg_descriptor gen7_render_regs[] = {
435 REG64(GPGPU_THREADS_DISPATCHED),
436 REG64(HS_INVOCATION_COUNT),
437 REG64(DS_INVOCATION_COUNT),
438 REG64(IA_VERTICES_COUNT),
439 REG64(IA_PRIMITIVES_COUNT),
440 REG64(VS_INVOCATION_COUNT),
441 REG64(GS_INVOCATION_COUNT),
442 REG64(GS_PRIMITIVES_COUNT),
443 REG64(CL_INVOCATION_COUNT),
444 REG64(CL_PRIMITIVES_COUNT),
445 REG64(PS_INVOCATION_COUNT),
446 REG64(PS_DEPTH_COUNT),
447 REG32(OACONTROL), /* Only allowed for LRI and SRM. See below. */
448 REG64(MI_PREDICATE_SRC0),
449 REG64(MI_PREDICATE_SRC1),
450 REG32(GEN7_3DPRIM_END_OFFSET),
451 REG32(GEN7_3DPRIM_START_VERTEX),
452 REG32(GEN7_3DPRIM_VERTEX_COUNT),
453 REG32(GEN7_3DPRIM_INSTANCE_COUNT),
454 REG32(GEN7_3DPRIM_START_INSTANCE),
455 REG32(GEN7_3DPRIM_BASE_VERTEX),
456 REG32(GEN7_GPGPU_DISPATCHDIMX),
457 REG32(GEN7_GPGPU_DISPATCHDIMY),
458 REG32(GEN7_GPGPU_DISPATCHDIMZ),
459 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 0),
460 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 1),
461 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 2),
462 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 3),
463 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 0),
464 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 1),
465 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 2),
466 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 3),
467 REG32(GEN7_SO_WRITE_OFFSET(0)),
468 REG32(GEN7_SO_WRITE_OFFSET(1)),
469 REG32(GEN7_SO_WRITE_OFFSET(2)),
470 REG32(GEN7_SO_WRITE_OFFSET(3)),
471 REG32(GEN7_L3SQCREG1),
472 REG32(GEN7_L3CNTLREG2),
473 REG32(GEN7_L3CNTLREG3),
474 REG32(HSW_SCRATCH1,
475 .mask = ~HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE,
476 .value = 0),
477 REG32(HSW_ROW_CHICKEN3,
478 .mask = ~(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE << 16 |
479 HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE),
480 .value = 0),
481};
482
483static const struct drm_i915_reg_descriptor gen7_blt_regs[] = {
484 REG32(BCS_SWCTRL),
485};
486
487static const struct drm_i915_reg_descriptor ivb_master_regs[] = {
488 REG32(FORCEWAKE_MT),
489 REG32(DERRMR),
490 REG32(GEN7_PIPE_DE_LOAD_SL(PIPE_A)),
491 REG32(GEN7_PIPE_DE_LOAD_SL(PIPE_B)),
492 REG32(GEN7_PIPE_DE_LOAD_SL(PIPE_C)),
493};
494
495static const struct drm_i915_reg_descriptor hsw_master_regs[] = {
496 REG32(FORCEWAKE_MT),
497 REG32(DERRMR),
498};
499
500#undef REG64
501#undef REG32
502
503static u32 gen7_render_get_cmd_length_mask(u32 cmd_header)
504{
505 u32 client = (cmd_header & INSTR_CLIENT_MASK) >> INSTR_CLIENT_SHIFT;
506 u32 subclient =
507 (cmd_header & INSTR_SUBCLIENT_MASK) >> INSTR_SUBCLIENT_SHIFT;
508
509 if (client == INSTR_MI_CLIENT)
510 return 0x3F;
511 else if (client == INSTR_RC_CLIENT) {
512 if (subclient == INSTR_MEDIA_SUBCLIENT)
513 return 0xFFFF;
514 else
515 return 0xFF;
516 }
517
518 DRM_DEBUG_DRIVER("CMD: Abnormal rcs cmd length! 0x%08X\n", cmd_header);
519 return 0;
520}
521
522static u32 gen7_bsd_get_cmd_length_mask(u32 cmd_header)
523{
524 u32 client = (cmd_header & INSTR_CLIENT_MASK) >> INSTR_CLIENT_SHIFT;
525 u32 subclient =
526 (cmd_header & INSTR_SUBCLIENT_MASK) >> INSTR_SUBCLIENT_SHIFT;
527 u32 op = (cmd_header & INSTR_26_TO_24_MASK) >> INSTR_26_TO_24_SHIFT;
528
529 if (client == INSTR_MI_CLIENT)
530 return 0x3F;
531 else if (client == INSTR_RC_CLIENT) {
532 if (subclient == INSTR_MEDIA_SUBCLIENT) {
533 if (op == 6)
534 return 0xFFFF;
535 else
536 return 0xFFF;
537 } else
538 return 0xFF;
539 }
540
541 DRM_DEBUG_DRIVER("CMD: Abnormal bsd cmd length! 0x%08X\n", cmd_header);
542 return 0;
543}
544
545static u32 gen7_blt_get_cmd_length_mask(u32 cmd_header)
546{
547 u32 client = (cmd_header & INSTR_CLIENT_MASK) >> INSTR_CLIENT_SHIFT;
548
549 if (client == INSTR_MI_CLIENT)
550 return 0x3F;
551 else if (client == INSTR_BC_CLIENT)
552 return 0xFF;
553
554 DRM_DEBUG_DRIVER("CMD: Abnormal blt cmd length! 0x%08X\n", cmd_header);
555 return 0;
556}
557
558static bool validate_cmds_sorted(struct intel_engine_cs *ring,
559 const struct drm_i915_cmd_table *cmd_tables,
560 int cmd_table_count)
561{
562 int i;
563 bool ret = true;
564
565 if (!cmd_tables || cmd_table_count == 0)
566 return true;
567
568 for (i = 0; i < cmd_table_count; i++) {
569 const struct drm_i915_cmd_table *table = &cmd_tables[i];
570 u32 previous = 0;
571 int j;
572
573 for (j = 0; j < table->count; j++) {
574 const struct drm_i915_cmd_descriptor *desc =
575 &table->table[j];
576 u32 curr = desc->cmd.value & desc->cmd.mask;
577
578 if (curr < previous) {
579 DRM_ERROR("CMD: table not sorted ring=%d table=%d entry=%d cmd=0x%08X prev=0x%08X\n",
580 ring->id, i, j, curr, previous);
581 ret = false;
582 }
583
584 previous = curr;
585 }
586 }
587
588 return ret;
589}
590
591static bool check_sorted(int ring_id,
592 const struct drm_i915_reg_descriptor *reg_table,
593 int reg_count)
594{
595 int i;
596 u32 previous = 0;
597 bool ret = true;
598
599 for (i = 0; i < reg_count; i++) {
600 u32 curr = i915_mmio_reg_offset(reg_table[i].addr);
601
602 if (curr < previous) {
603 DRM_ERROR("CMD: table not sorted ring=%d entry=%d reg=0x%08X prev=0x%08X\n",
604 ring_id, i, curr, previous);
605 ret = false;
606 }
607
608 previous = curr;
609 }
610
611 return ret;
612}
613
614static bool validate_regs_sorted(struct intel_engine_cs *ring)
615{
616 return check_sorted(ring->id, ring->reg_table, ring->reg_count) &&
617 check_sorted(ring->id, ring->master_reg_table,
618 ring->master_reg_count);
619}
620
621struct cmd_node {
622 const struct drm_i915_cmd_descriptor *desc;
623 struct hlist_node node;
624};
625
626/*
627 * Different command ranges have different numbers of bits for the opcode. For
628 * example, MI commands use bits 31:23 while 3D commands use bits 31:16. The
629 * problem is that, for example, MI commands use bits 22:16 for other fields
630 * such as GGTT vs PPGTT bits. If we include those bits in the mask then when
631 * we mask a command from a batch it could hash to the wrong bucket due to
632 * non-opcode bits being set. But if we don't include those bits, some 3D
633 * commands may hash to the same bucket due to not including opcode bits that
634 * make the command unique. For now, we will risk hashing to the same bucket.
635 *
636 * If we attempt to generate a perfect hash, we should be able to look at bits
637 * 31:29 of a command from a batch buffer and use the full mask for that
638 * client. The existing INSTR_CLIENT_MASK/SHIFT defines can be used for this.
639 */
640#define CMD_HASH_MASK STD_MI_OPCODE_MASK
641
642static int init_hash_table(struct intel_engine_cs *ring,
643 const struct drm_i915_cmd_table *cmd_tables,
644 int cmd_table_count)
645{
646 int i, j;
647
648 hash_init(ring->cmd_hash);
649
650 for (i = 0; i < cmd_table_count; i++) {
651 const struct drm_i915_cmd_table *table = &cmd_tables[i];
652
653 for (j = 0; j < table->count; j++) {
654 const struct drm_i915_cmd_descriptor *desc =
655 &table->table[j];
656 struct cmd_node *desc_node =
657 kmalloc(sizeof(*desc_node), GFP_KERNEL);
658
659 if (!desc_node)
660 return -ENOMEM;
661
662 desc_node->desc = desc;
663 hash_add(ring->cmd_hash, &desc_node->node,
664 desc->cmd.value & CMD_HASH_MASK);
665 }
666 }
667
668 return 0;
669}
670
671static void fini_hash_table(struct intel_engine_cs *ring)
672{
673 struct hlist_node *tmp;
674 struct cmd_node *desc_node;
675 int i;
676
677 hash_for_each_safe(ring->cmd_hash, i, tmp, desc_node, node) {
678 hash_del(&desc_node->node);
679 kfree(desc_node);
680 }
681}
682
683/**
684 * i915_cmd_parser_init_ring() - set cmd parser related fields for a ringbuffer
685 * @ring: the ringbuffer to initialize
686 *
687 * Optionally initializes fields related to batch buffer command parsing in the
688 * struct intel_engine_cs based on whether the platform requires software
689 * command parsing.
690 *
691 * Return: non-zero if initialization fails
692 */
693int i915_cmd_parser_init_ring(struct intel_engine_cs *ring)
694{
695 const struct drm_i915_cmd_table *cmd_tables;
696 int cmd_table_count;
697 int ret;
698
699 if (!IS_GEN7(ring->dev))
700 return 0;
701
702 switch (ring->id) {
703 case RCS:
704 if (IS_HASWELL(ring->dev)) {
705 cmd_tables = hsw_render_ring_cmds;
706 cmd_table_count =
707 ARRAY_SIZE(hsw_render_ring_cmds);
708 } else {
709 cmd_tables = gen7_render_cmds;
710 cmd_table_count = ARRAY_SIZE(gen7_render_cmds);
711 }
712
713 ring->reg_table = gen7_render_regs;
714 ring->reg_count = ARRAY_SIZE(gen7_render_regs);
715
716 if (IS_HASWELL(ring->dev)) {
717 ring->master_reg_table = hsw_master_regs;
718 ring->master_reg_count = ARRAY_SIZE(hsw_master_regs);
719 } else {
720 ring->master_reg_table = ivb_master_regs;
721 ring->master_reg_count = ARRAY_SIZE(ivb_master_regs);
722 }
723
724 ring->get_cmd_length_mask = gen7_render_get_cmd_length_mask;
725 break;
726 case VCS:
727 cmd_tables = gen7_video_cmds;
728 cmd_table_count = ARRAY_SIZE(gen7_video_cmds);
729 ring->get_cmd_length_mask = gen7_bsd_get_cmd_length_mask;
730 break;
731 case BCS:
732 if (IS_HASWELL(ring->dev)) {
733 cmd_tables = hsw_blt_ring_cmds;
734 cmd_table_count = ARRAY_SIZE(hsw_blt_ring_cmds);
735 } else {
736 cmd_tables = gen7_blt_cmds;
737 cmd_table_count = ARRAY_SIZE(gen7_blt_cmds);
738 }
739
740 ring->reg_table = gen7_blt_regs;
741 ring->reg_count = ARRAY_SIZE(gen7_blt_regs);
742
743 if (IS_HASWELL(ring->dev)) {
744 ring->master_reg_table = hsw_master_regs;
745 ring->master_reg_count = ARRAY_SIZE(hsw_master_regs);
746 } else {
747 ring->master_reg_table = ivb_master_regs;
748 ring->master_reg_count = ARRAY_SIZE(ivb_master_regs);
749 }
750
751 ring->get_cmd_length_mask = gen7_blt_get_cmd_length_mask;
752 break;
753 case VECS:
754 cmd_tables = hsw_vebox_cmds;
755 cmd_table_count = ARRAY_SIZE(hsw_vebox_cmds);
756 /* VECS can use the same length_mask function as VCS */
757 ring->get_cmd_length_mask = gen7_bsd_get_cmd_length_mask;
758 break;
759 default:
760 DRM_ERROR("CMD: cmd_parser_init with unknown ring: %d\n",
761 ring->id);
762 BUG();
763 }
764
765 BUG_ON(!validate_cmds_sorted(ring, cmd_tables, cmd_table_count));
766 BUG_ON(!validate_regs_sorted(ring));
767
768 WARN_ON(!hash_empty(ring->cmd_hash));
769
770 ret = init_hash_table(ring, cmd_tables, cmd_table_count);
771 if (ret) {
772 DRM_ERROR("CMD: cmd_parser_init failed!\n");
773 fini_hash_table(ring);
774 return ret;
775 }
776
777 ring->needs_cmd_parser = true;
778
779 return 0;
780}
781
782/**
783 * i915_cmd_parser_fini_ring() - clean up cmd parser related fields
784 * @ring: the ringbuffer to clean up
785 *
786 * Releases any resources related to command parsing that may have been
787 * initialized for the specified ring.
788 */
789void i915_cmd_parser_fini_ring(struct intel_engine_cs *ring)
790{
791 if (!ring->needs_cmd_parser)
792 return;
793
794 fini_hash_table(ring);
795}
796
797static const struct drm_i915_cmd_descriptor*
798find_cmd_in_table(struct intel_engine_cs *ring,
799 u32 cmd_header)
800{
801 struct cmd_node *desc_node;
802
803 hash_for_each_possible(ring->cmd_hash, desc_node, node,
804 cmd_header & CMD_HASH_MASK) {
805 const struct drm_i915_cmd_descriptor *desc = desc_node->desc;
806 u32 masked_cmd = desc->cmd.mask & cmd_header;
807 u32 masked_value = desc->cmd.value & desc->cmd.mask;
808
809 if (masked_cmd == masked_value)
810 return desc;
811 }
812
813 return NULL;
814}
815
816/*
817 * Returns a pointer to a descriptor for the command specified by cmd_header.
818 *
819 * The caller must supply space for a default descriptor via the default_desc
820 * parameter. If no descriptor for the specified command exists in the ring's
821 * command parser tables, this function fills in default_desc based on the
822 * ring's default length encoding and returns default_desc.
823 */
824static const struct drm_i915_cmd_descriptor*
825find_cmd(struct intel_engine_cs *ring,
826 u32 cmd_header,
827 struct drm_i915_cmd_descriptor *default_desc)
828{
829 const struct drm_i915_cmd_descriptor *desc;
830 u32 mask;
831
832 desc = find_cmd_in_table(ring, cmd_header);
833 if (desc)
834 return desc;
835
836 mask = ring->get_cmd_length_mask(cmd_header);
837 if (!mask)
838 return NULL;
839
840 BUG_ON(!default_desc);
841 default_desc->flags = CMD_DESC_SKIP;
842 default_desc->length.mask = mask;
843
844 return default_desc;
845}
846
847static const struct drm_i915_reg_descriptor *
848find_reg(const struct drm_i915_reg_descriptor *table,
849 int count, u32 addr)
850{
851 if (table) {
852 int i;
853
854 for (i = 0; i < count; i++) {
855 if (i915_mmio_reg_offset(table[i].addr) == addr)
856 return &table[i];
857 }
858 }
859
860 return NULL;
861}
862
863static u32 *vmap_batch(struct drm_i915_gem_object *obj,
864 unsigned start, unsigned len)
865{
866 int i;
867 void *addr = NULL;
868 struct sg_page_iter sg_iter;
869 int first_page = start >> PAGE_SHIFT;
870 int last_page = (len + start + 4095) >> PAGE_SHIFT;
871 int npages = last_page - first_page;
872 struct page **pages;
873
874 pages = drm_malloc_ab(npages, sizeof(*pages));
875 if (pages == NULL) {
876 DRM_DEBUG_DRIVER("Failed to get space for pages\n");
877 goto finish;
878 }
879
880 i = 0;
881 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, first_page) {
882 pages[i++] = sg_page_iter_page(&sg_iter);
883 if (i == npages)
884 break;
885 }
886
887 addr = vmap(pages, i, 0, PAGE_KERNEL);
888 if (addr == NULL) {
889 DRM_DEBUG_DRIVER("Failed to vmap pages\n");
890 goto finish;
891 }
892
893finish:
894 if (pages)
895 drm_free_large(pages);
896 return (u32*)addr;
897}
898
899/* Returns a vmap'd pointer to dest_obj, which the caller must unmap */
900static u32 *copy_batch(struct drm_i915_gem_object *dest_obj,
901 struct drm_i915_gem_object *src_obj,
902 u32 batch_start_offset,
903 u32 batch_len)
904{
905 int needs_clflush = 0;
906 void *src_base, *src;
907 void *dst = NULL;
908 int ret;
909
910 if (batch_len > dest_obj->base.size ||
911 batch_len + batch_start_offset > src_obj->base.size)
912 return ERR_PTR(-E2BIG);
913
914 if (WARN_ON(dest_obj->pages_pin_count == 0))
915 return ERR_PTR(-ENODEV);
916
917 ret = i915_gem_obj_prepare_shmem_read(src_obj, &needs_clflush);
918 if (ret) {
919 DRM_DEBUG_DRIVER("CMD: failed to prepare shadow batch\n");
920 return ERR_PTR(ret);
921 }
922
923 src_base = vmap_batch(src_obj, batch_start_offset, batch_len);
924 if (!src_base) {
925 DRM_DEBUG_DRIVER("CMD: Failed to vmap batch\n");
926 ret = -ENOMEM;
927 goto unpin_src;
928 }
929
930 ret = i915_gem_object_set_to_cpu_domain(dest_obj, true);
931 if (ret) {
932 DRM_DEBUG_DRIVER("CMD: Failed to set shadow batch to CPU\n");
933 goto unmap_src;
934 }
935
936 dst = vmap_batch(dest_obj, 0, batch_len);
937 if (!dst) {
938 DRM_DEBUG_DRIVER("CMD: Failed to vmap shadow batch\n");
939 ret = -ENOMEM;
940 goto unmap_src;
941 }
942
943 src = src_base + offset_in_page(batch_start_offset);
944 if (needs_clflush)
945 drm_clflush_virt_range(src, batch_len);
946
947 memcpy(dst, src, batch_len);
948
949unmap_src:
950 vunmap(src_base);
951unpin_src:
952 i915_gem_object_unpin_pages(src_obj);
953
954 return ret ? ERR_PTR(ret) : dst;
955}
956
957/**
958 * i915_needs_cmd_parser() - should a given ring use software command parsing?
959 * @ring: the ring in question
960 *
961 * Only certain platforms require software batch buffer command parsing, and
962 * only when enabled via module parameter.
963 *
964 * Return: true if the ring requires software command parsing
965 */
966bool i915_needs_cmd_parser(struct intel_engine_cs *ring)
967{
968 if (!ring->needs_cmd_parser)
969 return false;
970
971 if (!USES_PPGTT(ring->dev))
972 return false;
973
974 return (i915.enable_cmd_parser == 1);
975}
976
977static bool check_cmd(const struct intel_engine_cs *ring,
978 const struct drm_i915_cmd_descriptor *desc,
979 const u32 *cmd, u32 length,
980 const bool is_master,
981 bool *oacontrol_set)
982{
983 if (desc->flags & CMD_DESC_REJECT) {
984 DRM_DEBUG_DRIVER("CMD: Rejected command: 0x%08X\n", *cmd);
985 return false;
986 }
987
988 if ((desc->flags & CMD_DESC_MASTER) && !is_master) {
989 DRM_DEBUG_DRIVER("CMD: Rejected master-only command: 0x%08X\n",
990 *cmd);
991 return false;
992 }
993
994 if (desc->flags & CMD_DESC_REGISTER) {
995 /*
996 * Get the distance between individual register offset
997 * fields if the command can perform more than one
998 * access at a time.
999 */
1000 const u32 step = desc->reg.step ? desc->reg.step : length;
1001 u32 offset;
1002
1003 for (offset = desc->reg.offset; offset < length;
1004 offset += step) {
1005 const u32 reg_addr = cmd[offset] & desc->reg.mask;
1006 const struct drm_i915_reg_descriptor *reg =
1007 find_reg(ring->reg_table, ring->reg_count,
1008 reg_addr);
1009
1010 if (!reg && is_master)
1011 reg = find_reg(ring->master_reg_table,
1012 ring->master_reg_count,
1013 reg_addr);
1014
1015 if (!reg) {
1016 DRM_DEBUG_DRIVER("CMD: Rejected register 0x%08X in command: 0x%08X (ring=%d)\n",
1017 reg_addr, *cmd, ring->id);
1018 return false;
1019 }
1020
1021 /*
1022 * OACONTROL requires some special handling for
1023 * writes. We want to make sure that any batch which
1024 * enables OA also disables it before the end of the
1025 * batch. The goal is to prevent one process from
1026 * snooping on the perf data from another process. To do
1027 * that, we need to check the value that will be written
1028 * to the register. Hence, limit OACONTROL writes to
1029 * only MI_LOAD_REGISTER_IMM commands.
1030 */
1031 if (reg_addr == i915_mmio_reg_offset(OACONTROL)) {
1032 if (desc->cmd.value == MI_LOAD_REGISTER_MEM) {
1033 DRM_DEBUG_DRIVER("CMD: Rejected LRM to OACONTROL\n");
1034 return false;
1035 }
1036
1037 if (desc->cmd.value == MI_LOAD_REGISTER_IMM(1))
1038 *oacontrol_set = (cmd[offset + 1] != 0);
1039 }
1040
1041 /*
1042 * Check the value written to the register against the
1043 * allowed mask/value pair given in the whitelist entry.
1044 */
1045 if (reg->mask) {
1046 if (desc->cmd.value == MI_LOAD_REGISTER_MEM) {
1047 DRM_DEBUG_DRIVER("CMD: Rejected LRM to masked register 0x%08X\n",
1048 reg_addr);
1049 return false;
1050 }
1051
1052 if (desc->cmd.value == MI_LOAD_REGISTER_IMM(1) &&
1053 (offset + 2 > length ||
1054 (cmd[offset + 1] & reg->mask) != reg->value)) {
1055 DRM_DEBUG_DRIVER("CMD: Rejected LRI to masked register 0x%08X\n",
1056 reg_addr);
1057 return false;
1058 }
1059 }
1060 }
1061 }
1062
1063 if (desc->flags & CMD_DESC_BITMASK) {
1064 int i;
1065
1066 for (i = 0; i < MAX_CMD_DESC_BITMASKS; i++) {
1067 u32 dword;
1068
1069 if (desc->bits[i].mask == 0)
1070 break;
1071
1072 if (desc->bits[i].condition_mask != 0) {
1073 u32 offset =
1074 desc->bits[i].condition_offset;
1075 u32 condition = cmd[offset] &
1076 desc->bits[i].condition_mask;
1077
1078 if (condition == 0)
1079 continue;
1080 }
1081
1082 dword = cmd[desc->bits[i].offset] &
1083 desc->bits[i].mask;
1084
1085 if (dword != desc->bits[i].expected) {
1086 DRM_DEBUG_DRIVER("CMD: Rejected command 0x%08X for bitmask 0x%08X (exp=0x%08X act=0x%08X) (ring=%d)\n",
1087 *cmd,
1088 desc->bits[i].mask,
1089 desc->bits[i].expected,
1090 dword, ring->id);
1091 return false;
1092 }
1093 }
1094 }
1095
1096 return true;
1097}
1098
1099#define LENGTH_BIAS 2
1100
1101/**
1102 * i915_parse_cmds() - parse a submitted batch buffer for privilege violations
1103 * @ring: the ring on which the batch is to execute
1104 * @batch_obj: the batch buffer in question
1105 * @shadow_batch_obj: copy of the batch buffer in question
1106 * @batch_start_offset: byte offset in the batch at which execution starts
1107 * @batch_len: length of the commands in batch_obj
1108 * @is_master: is the submitting process the drm master?
1109 *
1110 * Parses the specified batch buffer looking for privilege violations as
1111 * described in the overview.
1112 *
1113 * Return: non-zero if the parser finds violations or otherwise fails; -EACCES
1114 * if the batch appears legal but should use hardware parsing
1115 */
1116int i915_parse_cmds(struct intel_engine_cs *ring,
1117 struct drm_i915_gem_object *batch_obj,
1118 struct drm_i915_gem_object *shadow_batch_obj,
1119 u32 batch_start_offset,
1120 u32 batch_len,
1121 bool is_master)
1122{
1123 u32 *cmd, *batch_base, *batch_end;
1124 struct drm_i915_cmd_descriptor default_desc = { 0 };
1125 bool oacontrol_set = false; /* OACONTROL tracking. See check_cmd() */
1126 int ret = 0;
1127
1128 batch_base = copy_batch(shadow_batch_obj, batch_obj,
1129 batch_start_offset, batch_len);
1130 if (IS_ERR(batch_base)) {
1131 DRM_DEBUG_DRIVER("CMD: Failed to copy batch\n");
1132 return PTR_ERR(batch_base);
1133 }
1134
1135 /*
1136 * We use the batch length as size because the shadow object is as
1137 * large or larger and copy_batch() will write MI_NOPs to the extra
1138 * space. Parsing should be faster in some cases this way.
1139 */
1140 batch_end = batch_base + (batch_len / sizeof(*batch_end));
1141
1142 cmd = batch_base;
1143 while (cmd < batch_end) {
1144 const struct drm_i915_cmd_descriptor *desc;
1145 u32 length;
1146
1147 if (*cmd == MI_BATCH_BUFFER_END)
1148 break;
1149
1150 desc = find_cmd(ring, *cmd, &default_desc);
1151 if (!desc) {
1152 DRM_DEBUG_DRIVER("CMD: Unrecognized command: 0x%08X\n",
1153 *cmd);
1154 ret = -EINVAL;
1155 break;
1156 }
1157
1158 /*
1159 * If the batch buffer contains a chained batch, return an
1160 * error that tells the caller to abort and dispatch the
1161 * workload as a non-secure batch.
1162 */
1163 if (desc->cmd.value == MI_BATCH_BUFFER_START) {
1164 ret = -EACCES;
1165 break;
1166 }
1167
1168 if (desc->flags & CMD_DESC_FIXED)
1169 length = desc->length.fixed;
1170 else
1171 length = ((*cmd & desc->length.mask) + LENGTH_BIAS);
1172
1173 if ((batch_end - cmd) < length) {
1174 DRM_DEBUG_DRIVER("CMD: Command length exceeds batch length: 0x%08X length=%u batchlen=%td\n",
1175 *cmd,
1176 length,
1177 batch_end - cmd);
1178 ret = -EINVAL;
1179 break;
1180 }
1181
1182 if (!check_cmd(ring, desc, cmd, length, is_master,
1183 &oacontrol_set)) {
1184 ret = -EINVAL;
1185 break;
1186 }
1187
1188 cmd += length;
1189 }
1190
1191 if (oacontrol_set) {
1192 DRM_DEBUG_DRIVER("CMD: batch set OACONTROL but did not clear it\n");
1193 ret = -EINVAL;
1194 }
1195
1196 if (cmd >= batch_end) {
1197 DRM_DEBUG_DRIVER("CMD: Got to the end of the buffer w/o a BBE cmd!\n");
1198 ret = -EINVAL;
1199 }
1200
1201 vunmap(batch_base);
1202
1203 return ret;
1204}
1205
1206/**
1207 * i915_cmd_parser_get_version() - get the cmd parser version number
1208 *
1209 * The cmd parser maintains a simple increasing integer version number suitable
1210 * for passing to userspace clients to determine what operations are permitted.
1211 *
1212 * Return: the current version number of the cmd parser
1213 */
1214int i915_cmd_parser_get_version(void)
1215{
1216 /*
1217 * Command parser version history
1218 *
1219 * 1. Initial version. Checks batches and reports violations, but leaves
1220 * hardware parsing enabled (so does not allow new use cases).
1221 * 2. Allow access to the MI_PREDICATE_SRC0 and
1222 * MI_PREDICATE_SRC1 registers.
1223 * 3. Allow access to the GPGPU_THREADS_DISPATCHED register.
1224 * 4. L3 atomic chicken bits of HSW_SCRATCH1 and HSW_ROW_CHICKEN3.
1225 * 5. GPGPU dispatch compute indirect registers.
1226 */
1227 return 5;
1228}
1/*
2 * Copyright © 2013 Intel Corporation
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Brad Volkin <bradley.d.volkin@intel.com>
25 *
26 */
27
28#include "i915_drv.h"
29
30/**
31 * DOC: batch buffer command parser
32 *
33 * Motivation:
34 * Certain OpenGL features (e.g. transform feedback, performance monitoring)
35 * require userspace code to submit batches containing commands such as
36 * MI_LOAD_REGISTER_IMM to access various registers. Unfortunately, some
37 * generations of the hardware will noop these commands in "unsecure" batches
38 * (which includes all userspace batches submitted via i915) even though the
39 * commands may be safe and represent the intended programming model of the
40 * device.
41 *
42 * The software command parser is similar in operation to the command parsing
43 * done in hardware for unsecure batches. However, the software parser allows
44 * some operations that would be noop'd by hardware, if the parser determines
45 * the operation is safe, and submits the batch as "secure" to prevent hardware
46 * parsing.
47 *
48 * Threats:
49 * At a high level, the hardware (and software) checks attempt to prevent
50 * granting userspace undue privileges. There are three categories of privilege.
51 *
52 * First, commands which are explicitly defined as privileged or which should
53 * only be used by the kernel driver. The parser generally rejects such
54 * commands, though it may allow some from the drm master process.
55 *
56 * Second, commands which access registers. To support correct/enhanced
57 * userspace functionality, particularly certain OpenGL extensions, the parser
58 * provides a whitelist of registers which userspace may safely access (for both
59 * normal and drm master processes).
60 *
61 * Third, commands which access privileged memory (i.e. GGTT, HWS page, etc).
62 * The parser always rejects such commands.
63 *
64 * The majority of the problematic commands fall in the MI_* range, with only a
65 * few specific commands on each engine (e.g. PIPE_CONTROL and MI_FLUSH_DW).
66 *
67 * Implementation:
68 * Each engine maintains tables of commands and registers which the parser
69 * uses in scanning batch buffers submitted to that engine.
70 *
71 * Since the set of commands that the parser must check for is significantly
72 * smaller than the number of commands supported, the parser tables contain only
73 * those commands required by the parser. This generally works because command
74 * opcode ranges have standard command length encodings. So for commands that
75 * the parser does not need to check, it can easily skip them. This is
76 * implemented via a per-engine length decoding vfunc.
77 *
78 * Unfortunately, there are a number of commands that do not follow the standard
79 * length encoding for their opcode range, primarily amongst the MI_* commands.
80 * To handle this, the parser provides a way to define explicit "skip" entries
81 * in the per-engine command tables.
82 *
83 * Other command table entries map fairly directly to high level categories
84 * mentioned above: rejected, master-only, register whitelist. The parser
85 * implements a number of checks, including the privileged memory checks, via a
86 * general bitmasking mechanism.
87 */
88
89#define STD_MI_OPCODE_SHIFT (32 - 9)
90#define STD_3D_OPCODE_SHIFT (32 - 16)
91#define STD_2D_OPCODE_SHIFT (32 - 10)
92#define STD_MFX_OPCODE_SHIFT (32 - 16)
93#define MIN_OPCODE_SHIFT 16
94
95#define CMD(op, opm, f, lm, fl, ...) \
96 { \
97 .flags = (fl) | ((f) ? CMD_DESC_FIXED : 0), \
98 .cmd = { (op), ~0u << (opm) }, \
99 .length = { (lm) }, \
100 __VA_ARGS__ \
101 }
102
103/* Convenience macros to compress the tables */
104#define SMI STD_MI_OPCODE_SHIFT
105#define S3D STD_3D_OPCODE_SHIFT
106#define S2D STD_2D_OPCODE_SHIFT
107#define SMFX STD_MFX_OPCODE_SHIFT
108#define F true
109#define S CMD_DESC_SKIP
110#define R CMD_DESC_REJECT
111#define W CMD_DESC_REGISTER
112#define B CMD_DESC_BITMASK
113#define M CMD_DESC_MASTER
114
115/* Command Mask Fixed Len Action
116 ---------------------------------------------------------- */
117static const struct drm_i915_cmd_descriptor common_cmds[] = {
118 CMD( MI_NOOP, SMI, F, 1, S ),
119 CMD( MI_USER_INTERRUPT, SMI, F, 1, R ),
120 CMD( MI_WAIT_FOR_EVENT, SMI, F, 1, M ),
121 CMD( MI_ARB_CHECK, SMI, F, 1, S ),
122 CMD( MI_REPORT_HEAD, SMI, F, 1, S ),
123 CMD( MI_SUSPEND_FLUSH, SMI, F, 1, S ),
124 CMD( MI_SEMAPHORE_MBOX, SMI, !F, 0xFF, R ),
125 CMD( MI_STORE_DWORD_INDEX, SMI, !F, 0xFF, R ),
126 CMD( MI_LOAD_REGISTER_IMM(1), SMI, !F, 0xFF, W,
127 .reg = { .offset = 1, .mask = 0x007FFFFC, .step = 2 } ),
128 CMD( MI_STORE_REGISTER_MEM, SMI, F, 3, W | B,
129 .reg = { .offset = 1, .mask = 0x007FFFFC },
130 .bits = {{
131 .offset = 0,
132 .mask = MI_GLOBAL_GTT,
133 .expected = 0,
134 }}, ),
135 CMD( MI_LOAD_REGISTER_MEM, SMI, F, 3, W | B,
136 .reg = { .offset = 1, .mask = 0x007FFFFC },
137 .bits = {{
138 .offset = 0,
139 .mask = MI_GLOBAL_GTT,
140 .expected = 0,
141 }}, ),
142 /*
143 * MI_BATCH_BUFFER_START requires some special handling. It's not
144 * really a 'skip' action but it doesn't seem like it's worth adding
145 * a new action. See i915_parse_cmds().
146 */
147 CMD( MI_BATCH_BUFFER_START, SMI, !F, 0xFF, S ),
148};
149
150static const struct drm_i915_cmd_descriptor render_cmds[] = {
151 CMD( MI_FLUSH, SMI, F, 1, S ),
152 CMD( MI_ARB_ON_OFF, SMI, F, 1, R ),
153 CMD( MI_PREDICATE, SMI, F, 1, S ),
154 CMD( MI_TOPOLOGY_FILTER, SMI, F, 1, S ),
155 CMD( MI_SET_APPID, SMI, F, 1, S ),
156 CMD( MI_DISPLAY_FLIP, SMI, !F, 0xFF, R ),
157 CMD( MI_SET_CONTEXT, SMI, !F, 0xFF, R ),
158 CMD( MI_URB_CLEAR, SMI, !F, 0xFF, S ),
159 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0x3F, B,
160 .bits = {{
161 .offset = 0,
162 .mask = MI_GLOBAL_GTT,
163 .expected = 0,
164 }}, ),
165 CMD( MI_UPDATE_GTT, SMI, !F, 0xFF, R ),
166 CMD( MI_CLFLUSH, SMI, !F, 0x3FF, B,
167 .bits = {{
168 .offset = 0,
169 .mask = MI_GLOBAL_GTT,
170 .expected = 0,
171 }}, ),
172 CMD( MI_REPORT_PERF_COUNT, SMI, !F, 0x3F, B,
173 .bits = {{
174 .offset = 1,
175 .mask = MI_REPORT_PERF_COUNT_GGTT,
176 .expected = 0,
177 }}, ),
178 CMD( MI_CONDITIONAL_BATCH_BUFFER_END, SMI, !F, 0xFF, B,
179 .bits = {{
180 .offset = 0,
181 .mask = MI_GLOBAL_GTT,
182 .expected = 0,
183 }}, ),
184 CMD( GFX_OP_3DSTATE_VF_STATISTICS, S3D, F, 1, S ),
185 CMD( PIPELINE_SELECT, S3D, F, 1, S ),
186 CMD( MEDIA_VFE_STATE, S3D, !F, 0xFFFF, B,
187 .bits = {{
188 .offset = 2,
189 .mask = MEDIA_VFE_STATE_MMIO_ACCESS_MASK,
190 .expected = 0,
191 }}, ),
192 CMD( GPGPU_OBJECT, S3D, !F, 0xFF, S ),
193 CMD( GPGPU_WALKER, S3D, !F, 0xFF, S ),
194 CMD( GFX_OP_3DSTATE_SO_DECL_LIST, S3D, !F, 0x1FF, S ),
195 CMD( GFX_OP_PIPE_CONTROL(5), S3D, !F, 0xFF, B,
196 .bits = {{
197 .offset = 1,
198 .mask = (PIPE_CONTROL_MMIO_WRITE | PIPE_CONTROL_NOTIFY),
199 .expected = 0,
200 },
201 {
202 .offset = 1,
203 .mask = (PIPE_CONTROL_GLOBAL_GTT_IVB |
204 PIPE_CONTROL_STORE_DATA_INDEX),
205 .expected = 0,
206 .condition_offset = 1,
207 .condition_mask = PIPE_CONTROL_POST_SYNC_OP_MASK,
208 }}, ),
209};
210
211static const struct drm_i915_cmd_descriptor hsw_render_cmds[] = {
212 CMD( MI_SET_PREDICATE, SMI, F, 1, S ),
213 CMD( MI_RS_CONTROL, SMI, F, 1, S ),
214 CMD( MI_URB_ATOMIC_ALLOC, SMI, F, 1, S ),
215 CMD( MI_SET_APPID, SMI, F, 1, S ),
216 CMD( MI_RS_CONTEXT, SMI, F, 1, S ),
217 CMD( MI_LOAD_SCAN_LINES_INCL, SMI, !F, 0x3F, M ),
218 CMD( MI_LOAD_SCAN_LINES_EXCL, SMI, !F, 0x3F, R ),
219 CMD( MI_LOAD_REGISTER_REG, SMI, !F, 0xFF, W,
220 .reg = { .offset = 1, .mask = 0x007FFFFC, .step = 1 } ),
221 CMD( MI_RS_STORE_DATA_IMM, SMI, !F, 0xFF, S ),
222 CMD( MI_LOAD_URB_MEM, SMI, !F, 0xFF, S ),
223 CMD( MI_STORE_URB_MEM, SMI, !F, 0xFF, S ),
224 CMD( GFX_OP_3DSTATE_DX9_CONSTANTF_VS, S3D, !F, 0x7FF, S ),
225 CMD( GFX_OP_3DSTATE_DX9_CONSTANTF_PS, S3D, !F, 0x7FF, S ),
226
227 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_VS, S3D, !F, 0x1FF, S ),
228 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_GS, S3D, !F, 0x1FF, S ),
229 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_HS, S3D, !F, 0x1FF, S ),
230 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_DS, S3D, !F, 0x1FF, S ),
231 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_PS, S3D, !F, 0x1FF, S ),
232};
233
234static const struct drm_i915_cmd_descriptor video_cmds[] = {
235 CMD( MI_ARB_ON_OFF, SMI, F, 1, R ),
236 CMD( MI_SET_APPID, SMI, F, 1, S ),
237 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0xFF, B,
238 .bits = {{
239 .offset = 0,
240 .mask = MI_GLOBAL_GTT,
241 .expected = 0,
242 }}, ),
243 CMD( MI_UPDATE_GTT, SMI, !F, 0x3F, R ),
244 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, B,
245 .bits = {{
246 .offset = 0,
247 .mask = MI_FLUSH_DW_NOTIFY,
248 .expected = 0,
249 },
250 {
251 .offset = 1,
252 .mask = MI_FLUSH_DW_USE_GTT,
253 .expected = 0,
254 .condition_offset = 0,
255 .condition_mask = MI_FLUSH_DW_OP_MASK,
256 },
257 {
258 .offset = 0,
259 .mask = MI_FLUSH_DW_STORE_INDEX,
260 .expected = 0,
261 .condition_offset = 0,
262 .condition_mask = MI_FLUSH_DW_OP_MASK,
263 }}, ),
264 CMD( MI_CONDITIONAL_BATCH_BUFFER_END, SMI, !F, 0xFF, B,
265 .bits = {{
266 .offset = 0,
267 .mask = MI_GLOBAL_GTT,
268 .expected = 0,
269 }}, ),
270 /*
271 * MFX_WAIT doesn't fit the way we handle length for most commands.
272 * It has a length field but it uses a non-standard length bias.
273 * It is always 1 dword though, so just treat it as fixed length.
274 */
275 CMD( MFX_WAIT, SMFX, F, 1, S ),
276};
277
278static const struct drm_i915_cmd_descriptor vecs_cmds[] = {
279 CMD( MI_ARB_ON_OFF, SMI, F, 1, R ),
280 CMD( MI_SET_APPID, SMI, F, 1, S ),
281 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0xFF, B,
282 .bits = {{
283 .offset = 0,
284 .mask = MI_GLOBAL_GTT,
285 .expected = 0,
286 }}, ),
287 CMD( MI_UPDATE_GTT, SMI, !F, 0x3F, R ),
288 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, B,
289 .bits = {{
290 .offset = 0,
291 .mask = MI_FLUSH_DW_NOTIFY,
292 .expected = 0,
293 },
294 {
295 .offset = 1,
296 .mask = MI_FLUSH_DW_USE_GTT,
297 .expected = 0,
298 .condition_offset = 0,
299 .condition_mask = MI_FLUSH_DW_OP_MASK,
300 },
301 {
302 .offset = 0,
303 .mask = MI_FLUSH_DW_STORE_INDEX,
304 .expected = 0,
305 .condition_offset = 0,
306 .condition_mask = MI_FLUSH_DW_OP_MASK,
307 }}, ),
308 CMD( MI_CONDITIONAL_BATCH_BUFFER_END, SMI, !F, 0xFF, B,
309 .bits = {{
310 .offset = 0,
311 .mask = MI_GLOBAL_GTT,
312 .expected = 0,
313 }}, ),
314};
315
316static const struct drm_i915_cmd_descriptor blt_cmds[] = {
317 CMD( MI_DISPLAY_FLIP, SMI, !F, 0xFF, R ),
318 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0x3FF, B,
319 .bits = {{
320 .offset = 0,
321 .mask = MI_GLOBAL_GTT,
322 .expected = 0,
323 }}, ),
324 CMD( MI_UPDATE_GTT, SMI, !F, 0x3F, R ),
325 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, B,
326 .bits = {{
327 .offset = 0,
328 .mask = MI_FLUSH_DW_NOTIFY,
329 .expected = 0,
330 },
331 {
332 .offset = 1,
333 .mask = MI_FLUSH_DW_USE_GTT,
334 .expected = 0,
335 .condition_offset = 0,
336 .condition_mask = MI_FLUSH_DW_OP_MASK,
337 },
338 {
339 .offset = 0,
340 .mask = MI_FLUSH_DW_STORE_INDEX,
341 .expected = 0,
342 .condition_offset = 0,
343 .condition_mask = MI_FLUSH_DW_OP_MASK,
344 }}, ),
345 CMD( COLOR_BLT, S2D, !F, 0x3F, S ),
346 CMD( SRC_COPY_BLT, S2D, !F, 0x3F, S ),
347};
348
349static const struct drm_i915_cmd_descriptor hsw_blt_cmds[] = {
350 CMD( MI_LOAD_SCAN_LINES_INCL, SMI, !F, 0x3F, M ),
351 CMD( MI_LOAD_SCAN_LINES_EXCL, SMI, !F, 0x3F, R ),
352};
353
354static const struct drm_i915_cmd_descriptor noop_desc =
355 CMD(MI_NOOP, SMI, F, 1, S);
356
357#undef CMD
358#undef SMI
359#undef S3D
360#undef S2D
361#undef SMFX
362#undef F
363#undef S
364#undef R
365#undef W
366#undef B
367#undef M
368
369static const struct drm_i915_cmd_table gen7_render_cmds[] = {
370 { common_cmds, ARRAY_SIZE(common_cmds) },
371 { render_cmds, ARRAY_SIZE(render_cmds) },
372};
373
374static const struct drm_i915_cmd_table hsw_render_ring_cmds[] = {
375 { common_cmds, ARRAY_SIZE(common_cmds) },
376 { render_cmds, ARRAY_SIZE(render_cmds) },
377 { hsw_render_cmds, ARRAY_SIZE(hsw_render_cmds) },
378};
379
380static const struct drm_i915_cmd_table gen7_video_cmds[] = {
381 { common_cmds, ARRAY_SIZE(common_cmds) },
382 { video_cmds, ARRAY_SIZE(video_cmds) },
383};
384
385static const struct drm_i915_cmd_table hsw_vebox_cmds[] = {
386 { common_cmds, ARRAY_SIZE(common_cmds) },
387 { vecs_cmds, ARRAY_SIZE(vecs_cmds) },
388};
389
390static const struct drm_i915_cmd_table gen7_blt_cmds[] = {
391 { common_cmds, ARRAY_SIZE(common_cmds) },
392 { blt_cmds, ARRAY_SIZE(blt_cmds) },
393};
394
395static const struct drm_i915_cmd_table hsw_blt_ring_cmds[] = {
396 { common_cmds, ARRAY_SIZE(common_cmds) },
397 { blt_cmds, ARRAY_SIZE(blt_cmds) },
398 { hsw_blt_cmds, ARRAY_SIZE(hsw_blt_cmds) },
399};
400
401/*
402 * Register whitelists, sorted by increasing register offset.
403 */
404
405/*
406 * An individual whitelist entry granting access to register addr. If
407 * mask is non-zero the argument of immediate register writes will be
408 * AND-ed with mask, and the command will be rejected if the result
409 * doesn't match value.
410 *
411 * Registers with non-zero mask are only allowed to be written using
412 * LRI.
413 */
414struct drm_i915_reg_descriptor {
415 i915_reg_t addr;
416 u32 mask;
417 u32 value;
418};
419
420/* Convenience macro for adding 32-bit registers. */
421#define REG32(_reg, ...) \
422 { .addr = (_reg), __VA_ARGS__ }
423
424/*
425 * Convenience macro for adding 64-bit registers.
426 *
427 * Some registers that userspace accesses are 64 bits. The register
428 * access commands only allow 32-bit accesses. Hence, we have to include
429 * entries for both halves of the 64-bit registers.
430 */
431#define REG64(_reg) \
432 { .addr = _reg }, \
433 { .addr = _reg ## _UDW }
434
435#define REG64_IDX(_reg, idx) \
436 { .addr = _reg(idx) }, \
437 { .addr = _reg ## _UDW(idx) }
438
439static const struct drm_i915_reg_descriptor gen7_render_regs[] = {
440 REG64(GPGPU_THREADS_DISPATCHED),
441 REG64(HS_INVOCATION_COUNT),
442 REG64(DS_INVOCATION_COUNT),
443 REG64(IA_VERTICES_COUNT),
444 REG64(IA_PRIMITIVES_COUNT),
445 REG64(VS_INVOCATION_COUNT),
446 REG64(GS_INVOCATION_COUNT),
447 REG64(GS_PRIMITIVES_COUNT),
448 REG64(CL_INVOCATION_COUNT),
449 REG64(CL_PRIMITIVES_COUNT),
450 REG64(PS_INVOCATION_COUNT),
451 REG64(PS_DEPTH_COUNT),
452 REG64_IDX(RING_TIMESTAMP, RENDER_RING_BASE),
453 REG32(OACONTROL), /* Only allowed for LRI and SRM. See below. */
454 REG64(MI_PREDICATE_SRC0),
455 REG64(MI_PREDICATE_SRC1),
456 REG32(GEN7_3DPRIM_END_OFFSET),
457 REG32(GEN7_3DPRIM_START_VERTEX),
458 REG32(GEN7_3DPRIM_VERTEX_COUNT),
459 REG32(GEN7_3DPRIM_INSTANCE_COUNT),
460 REG32(GEN7_3DPRIM_START_INSTANCE),
461 REG32(GEN7_3DPRIM_BASE_VERTEX),
462 REG32(GEN7_GPGPU_DISPATCHDIMX),
463 REG32(GEN7_GPGPU_DISPATCHDIMY),
464 REG32(GEN7_GPGPU_DISPATCHDIMZ),
465 REG64_IDX(RING_TIMESTAMP, BSD_RING_BASE),
466 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 0),
467 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 1),
468 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 2),
469 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 3),
470 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 0),
471 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 1),
472 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 2),
473 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 3),
474 REG32(GEN7_SO_WRITE_OFFSET(0)),
475 REG32(GEN7_SO_WRITE_OFFSET(1)),
476 REG32(GEN7_SO_WRITE_OFFSET(2)),
477 REG32(GEN7_SO_WRITE_OFFSET(3)),
478 REG32(GEN7_L3SQCREG1),
479 REG32(GEN7_L3CNTLREG2),
480 REG32(GEN7_L3CNTLREG3),
481 REG64_IDX(RING_TIMESTAMP, BLT_RING_BASE),
482};
483
484static const struct drm_i915_reg_descriptor hsw_render_regs[] = {
485 REG64_IDX(HSW_CS_GPR, 0),
486 REG64_IDX(HSW_CS_GPR, 1),
487 REG64_IDX(HSW_CS_GPR, 2),
488 REG64_IDX(HSW_CS_GPR, 3),
489 REG64_IDX(HSW_CS_GPR, 4),
490 REG64_IDX(HSW_CS_GPR, 5),
491 REG64_IDX(HSW_CS_GPR, 6),
492 REG64_IDX(HSW_CS_GPR, 7),
493 REG64_IDX(HSW_CS_GPR, 8),
494 REG64_IDX(HSW_CS_GPR, 9),
495 REG64_IDX(HSW_CS_GPR, 10),
496 REG64_IDX(HSW_CS_GPR, 11),
497 REG64_IDX(HSW_CS_GPR, 12),
498 REG64_IDX(HSW_CS_GPR, 13),
499 REG64_IDX(HSW_CS_GPR, 14),
500 REG64_IDX(HSW_CS_GPR, 15),
501 REG32(HSW_SCRATCH1,
502 .mask = ~HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE,
503 .value = 0),
504 REG32(HSW_ROW_CHICKEN3,
505 .mask = ~(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE << 16 |
506 HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE),
507 .value = 0),
508};
509
510static const struct drm_i915_reg_descriptor gen7_blt_regs[] = {
511 REG64_IDX(RING_TIMESTAMP, RENDER_RING_BASE),
512 REG64_IDX(RING_TIMESTAMP, BSD_RING_BASE),
513 REG32(BCS_SWCTRL),
514 REG64_IDX(RING_TIMESTAMP, BLT_RING_BASE),
515};
516
517static const struct drm_i915_reg_descriptor ivb_master_regs[] = {
518 REG32(FORCEWAKE_MT),
519 REG32(DERRMR),
520 REG32(GEN7_PIPE_DE_LOAD_SL(PIPE_A)),
521 REG32(GEN7_PIPE_DE_LOAD_SL(PIPE_B)),
522 REG32(GEN7_PIPE_DE_LOAD_SL(PIPE_C)),
523};
524
525static const struct drm_i915_reg_descriptor hsw_master_regs[] = {
526 REG32(FORCEWAKE_MT),
527 REG32(DERRMR),
528};
529
530#undef REG64
531#undef REG32
532
533struct drm_i915_reg_table {
534 const struct drm_i915_reg_descriptor *regs;
535 int num_regs;
536 bool master;
537};
538
539static const struct drm_i915_reg_table ivb_render_reg_tables[] = {
540 { gen7_render_regs, ARRAY_SIZE(gen7_render_regs), false },
541 { ivb_master_regs, ARRAY_SIZE(ivb_master_regs), true },
542};
543
544static const struct drm_i915_reg_table ivb_blt_reg_tables[] = {
545 { gen7_blt_regs, ARRAY_SIZE(gen7_blt_regs), false },
546 { ivb_master_regs, ARRAY_SIZE(ivb_master_regs), true },
547};
548
549static const struct drm_i915_reg_table hsw_render_reg_tables[] = {
550 { gen7_render_regs, ARRAY_SIZE(gen7_render_regs), false },
551 { hsw_render_regs, ARRAY_SIZE(hsw_render_regs), false },
552 { hsw_master_regs, ARRAY_SIZE(hsw_master_regs), true },
553};
554
555static const struct drm_i915_reg_table hsw_blt_reg_tables[] = {
556 { gen7_blt_regs, ARRAY_SIZE(gen7_blt_regs), false },
557 { hsw_master_regs, ARRAY_SIZE(hsw_master_regs), true },
558};
559
560static u32 gen7_render_get_cmd_length_mask(u32 cmd_header)
561{
562 u32 client = (cmd_header & INSTR_CLIENT_MASK) >> INSTR_CLIENT_SHIFT;
563 u32 subclient =
564 (cmd_header & INSTR_SUBCLIENT_MASK) >> INSTR_SUBCLIENT_SHIFT;
565
566 if (client == INSTR_MI_CLIENT)
567 return 0x3F;
568 else if (client == INSTR_RC_CLIENT) {
569 if (subclient == INSTR_MEDIA_SUBCLIENT)
570 return 0xFFFF;
571 else
572 return 0xFF;
573 }
574
575 DRM_DEBUG_DRIVER("CMD: Abnormal rcs cmd length! 0x%08X\n", cmd_header);
576 return 0;
577}
578
579static u32 gen7_bsd_get_cmd_length_mask(u32 cmd_header)
580{
581 u32 client = (cmd_header & INSTR_CLIENT_MASK) >> INSTR_CLIENT_SHIFT;
582 u32 subclient =
583 (cmd_header & INSTR_SUBCLIENT_MASK) >> INSTR_SUBCLIENT_SHIFT;
584 u32 op = (cmd_header & INSTR_26_TO_24_MASK) >> INSTR_26_TO_24_SHIFT;
585
586 if (client == INSTR_MI_CLIENT)
587 return 0x3F;
588 else if (client == INSTR_RC_CLIENT) {
589 if (subclient == INSTR_MEDIA_SUBCLIENT) {
590 if (op == 6)
591 return 0xFFFF;
592 else
593 return 0xFFF;
594 } else
595 return 0xFF;
596 }
597
598 DRM_DEBUG_DRIVER("CMD: Abnormal bsd cmd length! 0x%08X\n", cmd_header);
599 return 0;
600}
601
602static u32 gen7_blt_get_cmd_length_mask(u32 cmd_header)
603{
604 u32 client = (cmd_header & INSTR_CLIENT_MASK) >> INSTR_CLIENT_SHIFT;
605
606 if (client == INSTR_MI_CLIENT)
607 return 0x3F;
608 else if (client == INSTR_BC_CLIENT)
609 return 0xFF;
610
611 DRM_DEBUG_DRIVER("CMD: Abnormal blt cmd length! 0x%08X\n", cmd_header);
612 return 0;
613}
614
615static bool validate_cmds_sorted(const struct intel_engine_cs *engine,
616 const struct drm_i915_cmd_table *cmd_tables,
617 int cmd_table_count)
618{
619 int i;
620 bool ret = true;
621
622 if (!cmd_tables || cmd_table_count == 0)
623 return true;
624
625 for (i = 0; i < cmd_table_count; i++) {
626 const struct drm_i915_cmd_table *table = &cmd_tables[i];
627 u32 previous = 0;
628 int j;
629
630 for (j = 0; j < table->count; j++) {
631 const struct drm_i915_cmd_descriptor *desc =
632 &table->table[j];
633 u32 curr = desc->cmd.value & desc->cmd.mask;
634
635 if (curr < previous) {
636 DRM_ERROR("CMD: %s [%d] command table not sorted: "
637 "table=%d entry=%d cmd=0x%08X prev=0x%08X\n",
638 engine->name, engine->id,
639 i, j, curr, previous);
640 ret = false;
641 }
642
643 previous = curr;
644 }
645 }
646
647 return ret;
648}
649
650static bool check_sorted(const struct intel_engine_cs *engine,
651 const struct drm_i915_reg_descriptor *reg_table,
652 int reg_count)
653{
654 int i;
655 u32 previous = 0;
656 bool ret = true;
657
658 for (i = 0; i < reg_count; i++) {
659 u32 curr = i915_mmio_reg_offset(reg_table[i].addr);
660
661 if (curr < previous) {
662 DRM_ERROR("CMD: %s [%d] register table not sorted: "
663 "entry=%d reg=0x%08X prev=0x%08X\n",
664 engine->name, engine->id,
665 i, curr, previous);
666 ret = false;
667 }
668
669 previous = curr;
670 }
671
672 return ret;
673}
674
675static bool validate_regs_sorted(struct intel_engine_cs *engine)
676{
677 int i;
678 const struct drm_i915_reg_table *table;
679
680 for (i = 0; i < engine->reg_table_count; i++) {
681 table = &engine->reg_tables[i];
682 if (!check_sorted(engine, table->regs, table->num_regs))
683 return false;
684 }
685
686 return true;
687}
688
689struct cmd_node {
690 const struct drm_i915_cmd_descriptor *desc;
691 struct hlist_node node;
692};
693
694/*
695 * Different command ranges have different numbers of bits for the opcode. For
696 * example, MI commands use bits 31:23 while 3D commands use bits 31:16. The
697 * problem is that, for example, MI commands use bits 22:16 for other fields
698 * such as GGTT vs PPGTT bits. If we include those bits in the mask then when
699 * we mask a command from a batch it could hash to the wrong bucket due to
700 * non-opcode bits being set. But if we don't include those bits, some 3D
701 * commands may hash to the same bucket due to not including opcode bits that
702 * make the command unique. For now, we will risk hashing to the same bucket.
703 */
704static inline u32 cmd_header_key(u32 x)
705{
706 u32 shift;
707
708 switch (x >> INSTR_CLIENT_SHIFT) {
709 default:
710 case INSTR_MI_CLIENT:
711 shift = STD_MI_OPCODE_SHIFT;
712 break;
713 case INSTR_RC_CLIENT:
714 shift = STD_3D_OPCODE_SHIFT;
715 break;
716 case INSTR_BC_CLIENT:
717 shift = STD_2D_OPCODE_SHIFT;
718 break;
719 }
720
721 return x >> shift;
722}
723
724static int init_hash_table(struct intel_engine_cs *engine,
725 const struct drm_i915_cmd_table *cmd_tables,
726 int cmd_table_count)
727{
728 int i, j;
729
730 hash_init(engine->cmd_hash);
731
732 for (i = 0; i < cmd_table_count; i++) {
733 const struct drm_i915_cmd_table *table = &cmd_tables[i];
734
735 for (j = 0; j < table->count; j++) {
736 const struct drm_i915_cmd_descriptor *desc =
737 &table->table[j];
738 struct cmd_node *desc_node =
739 kmalloc(sizeof(*desc_node), GFP_KERNEL);
740
741 if (!desc_node)
742 return -ENOMEM;
743
744 desc_node->desc = desc;
745 hash_add(engine->cmd_hash, &desc_node->node,
746 cmd_header_key(desc->cmd.value));
747 }
748 }
749
750 return 0;
751}
752
753static void fini_hash_table(struct intel_engine_cs *engine)
754{
755 struct hlist_node *tmp;
756 struct cmd_node *desc_node;
757 int i;
758
759 hash_for_each_safe(engine->cmd_hash, i, tmp, desc_node, node) {
760 hash_del(&desc_node->node);
761 kfree(desc_node);
762 }
763}
764
765/**
766 * intel_engine_init_cmd_parser() - set cmd parser related fields for an engine
767 * @engine: the engine to initialize
768 *
769 * Optionally initializes fields related to batch buffer command parsing in the
770 * struct intel_engine_cs based on whether the platform requires software
771 * command parsing.
772 */
773void intel_engine_init_cmd_parser(struct intel_engine_cs *engine)
774{
775 const struct drm_i915_cmd_table *cmd_tables;
776 int cmd_table_count;
777 int ret;
778
779 if (!IS_GEN7(engine->i915))
780 return;
781
782 switch (engine->id) {
783 case RCS:
784 if (IS_HASWELL(engine->i915)) {
785 cmd_tables = hsw_render_ring_cmds;
786 cmd_table_count =
787 ARRAY_SIZE(hsw_render_ring_cmds);
788 } else {
789 cmd_tables = gen7_render_cmds;
790 cmd_table_count = ARRAY_SIZE(gen7_render_cmds);
791 }
792
793 if (IS_HASWELL(engine->i915)) {
794 engine->reg_tables = hsw_render_reg_tables;
795 engine->reg_table_count = ARRAY_SIZE(hsw_render_reg_tables);
796 } else {
797 engine->reg_tables = ivb_render_reg_tables;
798 engine->reg_table_count = ARRAY_SIZE(ivb_render_reg_tables);
799 }
800
801 engine->get_cmd_length_mask = gen7_render_get_cmd_length_mask;
802 break;
803 case VCS:
804 cmd_tables = gen7_video_cmds;
805 cmd_table_count = ARRAY_SIZE(gen7_video_cmds);
806 engine->get_cmd_length_mask = gen7_bsd_get_cmd_length_mask;
807 break;
808 case BCS:
809 if (IS_HASWELL(engine->i915)) {
810 cmd_tables = hsw_blt_ring_cmds;
811 cmd_table_count = ARRAY_SIZE(hsw_blt_ring_cmds);
812 } else {
813 cmd_tables = gen7_blt_cmds;
814 cmd_table_count = ARRAY_SIZE(gen7_blt_cmds);
815 }
816
817 if (IS_HASWELL(engine->i915)) {
818 engine->reg_tables = hsw_blt_reg_tables;
819 engine->reg_table_count = ARRAY_SIZE(hsw_blt_reg_tables);
820 } else {
821 engine->reg_tables = ivb_blt_reg_tables;
822 engine->reg_table_count = ARRAY_SIZE(ivb_blt_reg_tables);
823 }
824
825 engine->get_cmd_length_mask = gen7_blt_get_cmd_length_mask;
826 break;
827 case VECS:
828 cmd_tables = hsw_vebox_cmds;
829 cmd_table_count = ARRAY_SIZE(hsw_vebox_cmds);
830 /* VECS can use the same length_mask function as VCS */
831 engine->get_cmd_length_mask = gen7_bsd_get_cmd_length_mask;
832 break;
833 default:
834 MISSING_CASE(engine->id);
835 return;
836 }
837
838 if (!validate_cmds_sorted(engine, cmd_tables, cmd_table_count)) {
839 DRM_ERROR("%s: command descriptions are not sorted\n",
840 engine->name);
841 return;
842 }
843 if (!validate_regs_sorted(engine)) {
844 DRM_ERROR("%s: registers are not sorted\n", engine->name);
845 return;
846 }
847
848 ret = init_hash_table(engine, cmd_tables, cmd_table_count);
849 if (ret) {
850 DRM_ERROR("%s: initialised failed!\n", engine->name);
851 fini_hash_table(engine);
852 return;
853 }
854
855 engine->needs_cmd_parser = true;
856}
857
858/**
859 * intel_engine_cleanup_cmd_parser() - clean up cmd parser related fields
860 * @engine: the engine to clean up
861 *
862 * Releases any resources related to command parsing that may have been
863 * initialized for the specified engine.
864 */
865void intel_engine_cleanup_cmd_parser(struct intel_engine_cs *engine)
866{
867 if (!engine->needs_cmd_parser)
868 return;
869
870 fini_hash_table(engine);
871}
872
873static const struct drm_i915_cmd_descriptor*
874find_cmd_in_table(struct intel_engine_cs *engine,
875 u32 cmd_header)
876{
877 struct cmd_node *desc_node;
878
879 hash_for_each_possible(engine->cmd_hash, desc_node, node,
880 cmd_header_key(cmd_header)) {
881 const struct drm_i915_cmd_descriptor *desc = desc_node->desc;
882 if (((cmd_header ^ desc->cmd.value) & desc->cmd.mask) == 0)
883 return desc;
884 }
885
886 return NULL;
887}
888
889/*
890 * Returns a pointer to a descriptor for the command specified by cmd_header.
891 *
892 * The caller must supply space for a default descriptor via the default_desc
893 * parameter. If no descriptor for the specified command exists in the engine's
894 * command parser tables, this function fills in default_desc based on the
895 * engine's default length encoding and returns default_desc.
896 */
897static const struct drm_i915_cmd_descriptor*
898find_cmd(struct intel_engine_cs *engine,
899 u32 cmd_header,
900 const struct drm_i915_cmd_descriptor *desc,
901 struct drm_i915_cmd_descriptor *default_desc)
902{
903 u32 mask;
904
905 if (((cmd_header ^ desc->cmd.value) & desc->cmd.mask) == 0)
906 return desc;
907
908 desc = find_cmd_in_table(engine, cmd_header);
909 if (desc)
910 return desc;
911
912 mask = engine->get_cmd_length_mask(cmd_header);
913 if (!mask)
914 return NULL;
915
916 default_desc->cmd.value = cmd_header;
917 default_desc->cmd.mask = ~0u << MIN_OPCODE_SHIFT;
918 default_desc->length.mask = mask;
919 default_desc->flags = CMD_DESC_SKIP;
920 return default_desc;
921}
922
923static const struct drm_i915_reg_descriptor *
924__find_reg(const struct drm_i915_reg_descriptor *table, int count, u32 addr)
925{
926 int start = 0, end = count;
927 while (start < end) {
928 int mid = start + (end - start) / 2;
929 int ret = addr - i915_mmio_reg_offset(table[mid].addr);
930 if (ret < 0)
931 end = mid;
932 else if (ret > 0)
933 start = mid + 1;
934 else
935 return &table[mid];
936 }
937 return NULL;
938}
939
940static const struct drm_i915_reg_descriptor *
941find_reg(const struct intel_engine_cs *engine, bool is_master, u32 addr)
942{
943 const struct drm_i915_reg_table *table = engine->reg_tables;
944 int count = engine->reg_table_count;
945
946 do {
947 if (!table->master || is_master) {
948 const struct drm_i915_reg_descriptor *reg;
949
950 reg = __find_reg(table->regs, table->num_regs, addr);
951 if (reg != NULL)
952 return reg;
953 }
954 } while (table++, --count);
955
956 return NULL;
957}
958
959/* Returns a vmap'd pointer to dst_obj, which the caller must unmap */
960static u32 *copy_batch(struct drm_i915_gem_object *dst_obj,
961 struct drm_i915_gem_object *src_obj,
962 u32 batch_start_offset,
963 u32 batch_len,
964 bool *needs_clflush_after)
965{
966 unsigned int src_needs_clflush;
967 unsigned int dst_needs_clflush;
968 void *dst, *src;
969 int ret;
970
971 ret = i915_gem_obj_prepare_shmem_read(src_obj, &src_needs_clflush);
972 if (ret)
973 return ERR_PTR(ret);
974
975 ret = i915_gem_obj_prepare_shmem_write(dst_obj, &dst_needs_clflush);
976 if (ret) {
977 dst = ERR_PTR(ret);
978 goto unpin_src;
979 }
980
981 dst = i915_gem_object_pin_map(dst_obj, I915_MAP_WB);
982 if (IS_ERR(dst))
983 goto unpin_dst;
984
985 src = ERR_PTR(-ENODEV);
986 if (src_needs_clflush &&
987 i915_memcpy_from_wc((void *)(uintptr_t)batch_start_offset, NULL, 0)) {
988 src = i915_gem_object_pin_map(src_obj, I915_MAP_WC);
989 if (!IS_ERR(src)) {
990 i915_memcpy_from_wc(dst,
991 src + batch_start_offset,
992 ALIGN(batch_len, 16));
993 i915_gem_object_unpin_map(src_obj);
994 }
995 }
996 if (IS_ERR(src)) {
997 void *ptr;
998 int offset, n;
999
1000 offset = offset_in_page(batch_start_offset);
1001
1002 /* We can avoid clflushing partial cachelines before the write
1003 * if we only every write full cache-lines. Since we know that
1004 * both the source and destination are in multiples of
1005 * PAGE_SIZE, we can simply round up to the next cacheline.
1006 * We don't care about copying too much here as we only
1007 * validate up to the end of the batch.
1008 */
1009 if (dst_needs_clflush & CLFLUSH_BEFORE)
1010 batch_len = roundup(batch_len,
1011 boot_cpu_data.x86_clflush_size);
1012
1013 ptr = dst;
1014 for (n = batch_start_offset >> PAGE_SHIFT; batch_len; n++) {
1015 int len = min_t(int, batch_len, PAGE_SIZE - offset);
1016
1017 src = kmap_atomic(i915_gem_object_get_page(src_obj, n));
1018 if (src_needs_clflush)
1019 drm_clflush_virt_range(src + offset, len);
1020 memcpy(ptr, src + offset, len);
1021 kunmap_atomic(src);
1022
1023 ptr += len;
1024 batch_len -= len;
1025 offset = 0;
1026 }
1027 }
1028
1029 /* dst_obj is returned with vmap pinned */
1030 *needs_clflush_after = dst_needs_clflush & CLFLUSH_AFTER;
1031
1032unpin_dst:
1033 i915_gem_obj_finish_shmem_access(dst_obj);
1034unpin_src:
1035 i915_gem_obj_finish_shmem_access(src_obj);
1036 return dst;
1037}
1038
1039/**
1040 * intel_engine_needs_cmd_parser() - should a given engine use software
1041 * command parsing?
1042 * @engine: the engine in question
1043 *
1044 * Only certain platforms require software batch buffer command parsing, and
1045 * only when enabled via module parameter.
1046 *
1047 * Return: true if the engine requires software command parsing
1048 */
1049bool intel_engine_needs_cmd_parser(struct intel_engine_cs *engine)
1050{
1051 if (!engine->needs_cmd_parser)
1052 return false;
1053
1054 if (!USES_PPGTT(engine->i915))
1055 return false;
1056
1057 return (i915.enable_cmd_parser == 1);
1058}
1059
1060static bool check_cmd(const struct intel_engine_cs *engine,
1061 const struct drm_i915_cmd_descriptor *desc,
1062 const u32 *cmd, u32 length,
1063 const bool is_master,
1064 bool *oacontrol_set)
1065{
1066 if (desc->flags & CMD_DESC_SKIP)
1067 return true;
1068
1069 if (desc->flags & CMD_DESC_REJECT) {
1070 DRM_DEBUG_DRIVER("CMD: Rejected command: 0x%08X\n", *cmd);
1071 return false;
1072 }
1073
1074 if ((desc->flags & CMD_DESC_MASTER) && !is_master) {
1075 DRM_DEBUG_DRIVER("CMD: Rejected master-only command: 0x%08X\n",
1076 *cmd);
1077 return false;
1078 }
1079
1080 if (desc->flags & CMD_DESC_REGISTER) {
1081 /*
1082 * Get the distance between individual register offset
1083 * fields if the command can perform more than one
1084 * access at a time.
1085 */
1086 const u32 step = desc->reg.step ? desc->reg.step : length;
1087 u32 offset;
1088
1089 for (offset = desc->reg.offset; offset < length;
1090 offset += step) {
1091 const u32 reg_addr = cmd[offset] & desc->reg.mask;
1092 const struct drm_i915_reg_descriptor *reg =
1093 find_reg(engine, is_master, reg_addr);
1094
1095 if (!reg) {
1096 DRM_DEBUG_DRIVER("CMD: Rejected register 0x%08X in command: 0x%08X (exec_id=%d)\n",
1097 reg_addr, *cmd, engine->exec_id);
1098 return false;
1099 }
1100
1101 /*
1102 * OACONTROL requires some special handling for
1103 * writes. We want to make sure that any batch which
1104 * enables OA also disables it before the end of the
1105 * batch. The goal is to prevent one process from
1106 * snooping on the perf data from another process. To do
1107 * that, we need to check the value that will be written
1108 * to the register. Hence, limit OACONTROL writes to
1109 * only MI_LOAD_REGISTER_IMM commands.
1110 */
1111 if (reg_addr == i915_mmio_reg_offset(OACONTROL)) {
1112 if (desc->cmd.value == MI_LOAD_REGISTER_MEM) {
1113 DRM_DEBUG_DRIVER("CMD: Rejected LRM to OACONTROL\n");
1114 return false;
1115 }
1116
1117 if (desc->cmd.value == MI_LOAD_REGISTER_REG) {
1118 DRM_DEBUG_DRIVER("CMD: Rejected LRR to OACONTROL\n");
1119 return false;
1120 }
1121
1122 if (desc->cmd.value == MI_LOAD_REGISTER_IMM(1))
1123 *oacontrol_set = (cmd[offset + 1] != 0);
1124 }
1125
1126 /*
1127 * Check the value written to the register against the
1128 * allowed mask/value pair given in the whitelist entry.
1129 */
1130 if (reg->mask) {
1131 if (desc->cmd.value == MI_LOAD_REGISTER_MEM) {
1132 DRM_DEBUG_DRIVER("CMD: Rejected LRM to masked register 0x%08X\n",
1133 reg_addr);
1134 return false;
1135 }
1136
1137 if (desc->cmd.value == MI_LOAD_REGISTER_REG) {
1138 DRM_DEBUG_DRIVER("CMD: Rejected LRR to masked register 0x%08X\n",
1139 reg_addr);
1140 return false;
1141 }
1142
1143 if (desc->cmd.value == MI_LOAD_REGISTER_IMM(1) &&
1144 (offset + 2 > length ||
1145 (cmd[offset + 1] & reg->mask) != reg->value)) {
1146 DRM_DEBUG_DRIVER("CMD: Rejected LRI to masked register 0x%08X\n",
1147 reg_addr);
1148 return false;
1149 }
1150 }
1151 }
1152 }
1153
1154 if (desc->flags & CMD_DESC_BITMASK) {
1155 int i;
1156
1157 for (i = 0; i < MAX_CMD_DESC_BITMASKS; i++) {
1158 u32 dword;
1159
1160 if (desc->bits[i].mask == 0)
1161 break;
1162
1163 if (desc->bits[i].condition_mask != 0) {
1164 u32 offset =
1165 desc->bits[i].condition_offset;
1166 u32 condition = cmd[offset] &
1167 desc->bits[i].condition_mask;
1168
1169 if (condition == 0)
1170 continue;
1171 }
1172
1173 dword = cmd[desc->bits[i].offset] &
1174 desc->bits[i].mask;
1175
1176 if (dword != desc->bits[i].expected) {
1177 DRM_DEBUG_DRIVER("CMD: Rejected command 0x%08X for bitmask 0x%08X (exp=0x%08X act=0x%08X) (exec_id=%d)\n",
1178 *cmd,
1179 desc->bits[i].mask,
1180 desc->bits[i].expected,
1181 dword, engine->exec_id);
1182 return false;
1183 }
1184 }
1185 }
1186
1187 return true;
1188}
1189
1190#define LENGTH_BIAS 2
1191
1192/**
1193 * i915_parse_cmds() - parse a submitted batch buffer for privilege violations
1194 * @engine: the engine on which the batch is to execute
1195 * @batch_obj: the batch buffer in question
1196 * @shadow_batch_obj: copy of the batch buffer in question
1197 * @batch_start_offset: byte offset in the batch at which execution starts
1198 * @batch_len: length of the commands in batch_obj
1199 * @is_master: is the submitting process the drm master?
1200 *
1201 * Parses the specified batch buffer looking for privilege violations as
1202 * described in the overview.
1203 *
1204 * Return: non-zero if the parser finds violations or otherwise fails; -EACCES
1205 * if the batch appears legal but should use hardware parsing
1206 */
1207int intel_engine_cmd_parser(struct intel_engine_cs *engine,
1208 struct drm_i915_gem_object *batch_obj,
1209 struct drm_i915_gem_object *shadow_batch_obj,
1210 u32 batch_start_offset,
1211 u32 batch_len,
1212 bool is_master)
1213{
1214 u32 *cmd, *batch_end;
1215 struct drm_i915_cmd_descriptor default_desc = noop_desc;
1216 const struct drm_i915_cmd_descriptor *desc = &default_desc;
1217 bool oacontrol_set = false; /* OACONTROL tracking. See check_cmd() */
1218 bool needs_clflush_after = false;
1219 int ret = 0;
1220
1221 cmd = copy_batch(shadow_batch_obj, batch_obj,
1222 batch_start_offset, batch_len,
1223 &needs_clflush_after);
1224 if (IS_ERR(cmd)) {
1225 DRM_DEBUG_DRIVER("CMD: Failed to copy batch\n");
1226 return PTR_ERR(cmd);
1227 }
1228
1229 /*
1230 * We use the batch length as size because the shadow object is as
1231 * large or larger and copy_batch() will write MI_NOPs to the extra
1232 * space. Parsing should be faster in some cases this way.
1233 */
1234 batch_end = cmd + (batch_len / sizeof(*batch_end));
1235 while (cmd < batch_end) {
1236 u32 length;
1237
1238 if (*cmd == MI_BATCH_BUFFER_END)
1239 break;
1240
1241 desc = find_cmd(engine, *cmd, desc, &default_desc);
1242 if (!desc) {
1243 DRM_DEBUG_DRIVER("CMD: Unrecognized command: 0x%08X\n",
1244 *cmd);
1245 ret = -EINVAL;
1246 break;
1247 }
1248
1249 /*
1250 * If the batch buffer contains a chained batch, return an
1251 * error that tells the caller to abort and dispatch the
1252 * workload as a non-secure batch.
1253 */
1254 if (desc->cmd.value == MI_BATCH_BUFFER_START) {
1255 ret = -EACCES;
1256 break;
1257 }
1258
1259 if (desc->flags & CMD_DESC_FIXED)
1260 length = desc->length.fixed;
1261 else
1262 length = ((*cmd & desc->length.mask) + LENGTH_BIAS);
1263
1264 if ((batch_end - cmd) < length) {
1265 DRM_DEBUG_DRIVER("CMD: Command length exceeds batch length: 0x%08X length=%u batchlen=%td\n",
1266 *cmd,
1267 length,
1268 batch_end - cmd);
1269 ret = -EINVAL;
1270 break;
1271 }
1272
1273 if (!check_cmd(engine, desc, cmd, length, is_master,
1274 &oacontrol_set)) {
1275 ret = -EINVAL;
1276 break;
1277 }
1278
1279 cmd += length;
1280 }
1281
1282 if (oacontrol_set) {
1283 DRM_DEBUG_DRIVER("CMD: batch set OACONTROL but did not clear it\n");
1284 ret = -EINVAL;
1285 }
1286
1287 if (cmd >= batch_end) {
1288 DRM_DEBUG_DRIVER("CMD: Got to the end of the buffer w/o a BBE cmd!\n");
1289 ret = -EINVAL;
1290 }
1291
1292 if (ret == 0 && needs_clflush_after)
1293 drm_clflush_virt_range(shadow_batch_obj->mm.mapping, batch_len);
1294 i915_gem_object_unpin_map(shadow_batch_obj);
1295
1296 return ret;
1297}
1298
1299/**
1300 * i915_cmd_parser_get_version() - get the cmd parser version number
1301 * @dev_priv: i915 device private
1302 *
1303 * The cmd parser maintains a simple increasing integer version number suitable
1304 * for passing to userspace clients to determine what operations are permitted.
1305 *
1306 * Return: the current version number of the cmd parser
1307 */
1308int i915_cmd_parser_get_version(struct drm_i915_private *dev_priv)
1309{
1310 struct intel_engine_cs *engine;
1311 enum intel_engine_id id;
1312 bool active = false;
1313
1314 /* If the command parser is not enabled, report 0 - unsupported */
1315 for_each_engine(engine, dev_priv, id) {
1316 if (intel_engine_needs_cmd_parser(engine)) {
1317 active = true;
1318 break;
1319 }
1320 }
1321 if (!active)
1322 return 0;
1323
1324 /*
1325 * Command parser version history
1326 *
1327 * 1. Initial version. Checks batches and reports violations, but leaves
1328 * hardware parsing enabled (so does not allow new use cases).
1329 * 2. Allow access to the MI_PREDICATE_SRC0 and
1330 * MI_PREDICATE_SRC1 registers.
1331 * 3. Allow access to the GPGPU_THREADS_DISPATCHED register.
1332 * 4. L3 atomic chicken bits of HSW_SCRATCH1 and HSW_ROW_CHICKEN3.
1333 * 5. GPGPU dispatch compute indirect registers.
1334 * 6. TIMESTAMP register and Haswell CS GPR registers
1335 * 7. Allow MI_LOAD_REGISTER_REG between whitelisted registers.
1336 */
1337 return 7;
1338}