2 * Copyright © 2013 Intel Corporation
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:
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
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
24 * Brad Volkin <bradley.d.volkin@intel.com>
31 * DOC: batch buffer command parser
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
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
49 * At a high level, the hardware (and software) checks attempt to prevent
50 * granting userspace undue privileges. There are three categories of privilege.
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.
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).
61 * Third, commands which access privileged memory (i.e. GGTT, HWS page, etc).
62 * The parser always rejects such commands.
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).
68 * Each ring maintains tables of commands and registers which the parser uses in
69 * scanning batch buffers submitted to that ring.
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.
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.
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.
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
94 #define CMD(op, opm, f, lm, fl, ...) \
96 .flags = (fl) | ((f) ? CMD_DESC_FIXED : 0), \
97 .cmd = { (op), (opm) }, \
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
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
114 /* Command Mask Fixed Len Action
115 ---------------------------------------------------------- */
116 static 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 },
131 .mask
= MI_GLOBAL_GTT
,
134 CMD( MI_LOAD_REGISTER_MEM
, SMI
, F
, 3, W
| B
,
135 .reg
= { .offset
= 1, .mask
= 0x007FFFFC },
138 .mask
= MI_GLOBAL_GTT
,
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().
146 CMD( MI_BATCH_BUFFER_START
, SMI
, !F
, 0xFF, S
),
149 static 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
,
161 .mask
= MI_GLOBAL_GTT
,
164 CMD( MI_UPDATE_GTT
, SMI
, !F
, 0xFF, R
),
165 CMD( MI_CLFLUSH
, SMI
, !F
, 0x3FF, B
,
168 .mask
= MI_GLOBAL_GTT
,
171 CMD( MI_REPORT_PERF_COUNT
, SMI
, !F
, 0x3F, B
,
174 .mask
= MI_REPORT_PERF_COUNT_GGTT
,
177 CMD( MI_CONDITIONAL_BATCH_BUFFER_END
, SMI
, !F
, 0xFF, B
,
180 .mask
= MI_GLOBAL_GTT
,
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
,
188 .mask
= MEDIA_VFE_STATE_MMIO_ACCESS_MASK
,
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
,
197 .mask
= (PIPE_CONTROL_MMIO_WRITE
| PIPE_CONTROL_NOTIFY
),
202 .mask
= (PIPE_CONTROL_GLOBAL_GTT_IVB
|
203 PIPE_CONTROL_STORE_DATA_INDEX
),
205 .condition_offset
= 1,
206 .condition_mask
= PIPE_CONTROL_POST_SYNC_OP_MASK
,
210 static 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
),
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
),
232 static 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
,
238 .mask
= MI_GLOBAL_GTT
,
241 CMD( MI_UPDATE_GTT
, SMI
, !F
, 0x3F, R
),
242 CMD( MI_FLUSH_DW
, SMI
, !F
, 0x3F, B
,
245 .mask
= MI_FLUSH_DW_NOTIFY
,
250 .mask
= MI_FLUSH_DW_USE_GTT
,
252 .condition_offset
= 0,
253 .condition_mask
= MI_FLUSH_DW_OP_MASK
,
257 .mask
= MI_FLUSH_DW_STORE_INDEX
,
259 .condition_offset
= 0,
260 .condition_mask
= MI_FLUSH_DW_OP_MASK
,
262 CMD( MI_CONDITIONAL_BATCH_BUFFER_END
, SMI
, !F
, 0xFF, B
,
265 .mask
= MI_GLOBAL_GTT
,
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.
273 CMD( MFX_WAIT
, SMFX
, F
, 1, S
),
276 static 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
,
282 .mask
= MI_GLOBAL_GTT
,
285 CMD( MI_UPDATE_GTT
, SMI
, !F
, 0x3F, R
),
286 CMD( MI_FLUSH_DW
, SMI
, !F
, 0x3F, B
,
289 .mask
= MI_FLUSH_DW_NOTIFY
,
294 .mask
= MI_FLUSH_DW_USE_GTT
,
296 .condition_offset
= 0,
297 .condition_mask
= MI_FLUSH_DW_OP_MASK
,
301 .mask
= MI_FLUSH_DW_STORE_INDEX
,
303 .condition_offset
= 0,
304 .condition_mask
= MI_FLUSH_DW_OP_MASK
,
306 CMD( MI_CONDITIONAL_BATCH_BUFFER_END
, SMI
, !F
, 0xFF, B
,
309 .mask
= MI_GLOBAL_GTT
,
314 static 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
,
319 .mask
= MI_GLOBAL_GTT
,
322 CMD( MI_UPDATE_GTT
, SMI
, !F
, 0x3F, R
),
323 CMD( MI_FLUSH_DW
, SMI
, !F
, 0x3F, B
,
326 .mask
= MI_FLUSH_DW_NOTIFY
,
331 .mask
= MI_FLUSH_DW_USE_GTT
,
333 .condition_offset
= 0,
334 .condition_mask
= MI_FLUSH_DW_OP_MASK
,
338 .mask
= MI_FLUSH_DW_STORE_INDEX
,
340 .condition_offset
= 0,
341 .condition_mask
= MI_FLUSH_DW_OP_MASK
,
343 CMD( COLOR_BLT
, S2D
, !F
, 0x3F, S
),
344 CMD( SRC_COPY_BLT
, S2D
, !F
, 0x3F, S
),
347 static 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
),
364 static const struct drm_i915_cmd_table gen7_render_cmds
[] = {
365 { common_cmds
, ARRAY_SIZE(common_cmds
) },
366 { render_cmds
, ARRAY_SIZE(render_cmds
) },
369 static 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
) },
375 static const struct drm_i915_cmd_table gen7_video_cmds
[] = {
376 { common_cmds
, ARRAY_SIZE(common_cmds
) },
377 { video_cmds
, ARRAY_SIZE(video_cmds
) },
380 static const struct drm_i915_cmd_table hsw_vebox_cmds
[] = {
381 { common_cmds
, ARRAY_SIZE(common_cmds
) },
382 { vecs_cmds
, ARRAY_SIZE(vecs_cmds
) },
385 static const struct drm_i915_cmd_table gen7_blt_cmds
[] = {
386 { common_cmds
, ARRAY_SIZE(common_cmds
) },
387 { blt_cmds
, ARRAY_SIZE(blt_cmds
) },
390 static 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
) },
397 * Register whitelists, sorted by increasing register offset.
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.
406 * Registers with non-zero mask are only allowed to be written using
409 struct drm_i915_reg_descriptor
{
415 /* Convenience macro for adding 32-bit registers. */
416 #define REG32(_reg, ...) \
417 { .addr = (_reg), __VA_ARGS__ }
420 * Convenience macro for adding 64-bit registers.
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.
426 #define REG64(_reg) \
428 { .addr = _reg ## _UDW }
430 #define REG64_IDX(_reg, idx) \
431 { .addr = _reg(idx) }, \
432 { .addr = _reg ## _UDW(idx) }
434 static 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 REG64_IDX(RING_TIMESTAMP
, RENDER_RING_BASE
),
448 REG32(OACONTROL
), /* Only allowed for LRI and SRM. See below. */
449 REG64(MI_PREDICATE_SRC0
),
450 REG64(MI_PREDICATE_SRC1
),
451 REG32(GEN7_3DPRIM_END_OFFSET
),
452 REG32(GEN7_3DPRIM_START_VERTEX
),
453 REG32(GEN7_3DPRIM_VERTEX_COUNT
),
454 REG32(GEN7_3DPRIM_INSTANCE_COUNT
),
455 REG32(GEN7_3DPRIM_START_INSTANCE
),
456 REG32(GEN7_3DPRIM_BASE_VERTEX
),
457 REG32(GEN7_GPGPU_DISPATCHDIMX
),
458 REG32(GEN7_GPGPU_DISPATCHDIMY
),
459 REG32(GEN7_GPGPU_DISPATCHDIMZ
),
460 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN
, 0),
461 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN
, 1),
462 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN
, 2),
463 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN
, 3),
464 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED
, 0),
465 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED
, 1),
466 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED
, 2),
467 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED
, 3),
468 REG32(GEN7_SO_WRITE_OFFSET(0)),
469 REG32(GEN7_SO_WRITE_OFFSET(1)),
470 REG32(GEN7_SO_WRITE_OFFSET(2)),
471 REG32(GEN7_SO_WRITE_OFFSET(3)),
472 REG32(GEN7_L3SQCREG1
),
473 REG32(GEN7_L3CNTLREG2
),
474 REG32(GEN7_L3CNTLREG3
),
477 static const struct drm_i915_reg_descriptor hsw_render_regs
[] = {
478 REG64_IDX(HSW_CS_GPR
, 0),
479 REG64_IDX(HSW_CS_GPR
, 1),
480 REG64_IDX(HSW_CS_GPR
, 2),
481 REG64_IDX(HSW_CS_GPR
, 3),
482 REG64_IDX(HSW_CS_GPR
, 4),
483 REG64_IDX(HSW_CS_GPR
, 5),
484 REG64_IDX(HSW_CS_GPR
, 6),
485 REG64_IDX(HSW_CS_GPR
, 7),
486 REG64_IDX(HSW_CS_GPR
, 8),
487 REG64_IDX(HSW_CS_GPR
, 9),
488 REG64_IDX(HSW_CS_GPR
, 10),
489 REG64_IDX(HSW_CS_GPR
, 11),
490 REG64_IDX(HSW_CS_GPR
, 12),
491 REG64_IDX(HSW_CS_GPR
, 13),
492 REG64_IDX(HSW_CS_GPR
, 14),
493 REG64_IDX(HSW_CS_GPR
, 15),
495 .mask
= ~HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE
,
497 REG32(HSW_ROW_CHICKEN3
,
498 .mask
= ~(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE
<< 16 |
499 HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE
),
503 static const struct drm_i915_reg_descriptor gen7_blt_regs
[] = {
507 static const struct drm_i915_reg_descriptor ivb_master_regs
[] = {
510 REG32(GEN7_PIPE_DE_LOAD_SL(PIPE_A
)),
511 REG32(GEN7_PIPE_DE_LOAD_SL(PIPE_B
)),
512 REG32(GEN7_PIPE_DE_LOAD_SL(PIPE_C
)),
515 static const struct drm_i915_reg_descriptor hsw_master_regs
[] = {
523 struct drm_i915_reg_table
{
524 const struct drm_i915_reg_descriptor
*regs
;
529 static const struct drm_i915_reg_table ivb_render_reg_tables
[] = {
530 { gen7_render_regs
, ARRAY_SIZE(gen7_render_regs
), false },
531 { ivb_master_regs
, ARRAY_SIZE(ivb_master_regs
), true },
534 static const struct drm_i915_reg_table ivb_blt_reg_tables
[] = {
535 { gen7_blt_regs
, ARRAY_SIZE(gen7_blt_regs
), false },
536 { ivb_master_regs
, ARRAY_SIZE(ivb_master_regs
), true },
539 static const struct drm_i915_reg_table hsw_render_reg_tables
[] = {
540 { gen7_render_regs
, ARRAY_SIZE(gen7_render_regs
), false },
541 { hsw_render_regs
, ARRAY_SIZE(hsw_render_regs
), false },
542 { hsw_master_regs
, ARRAY_SIZE(hsw_master_regs
), true },
545 static const struct drm_i915_reg_table hsw_blt_reg_tables
[] = {
546 { gen7_blt_regs
, ARRAY_SIZE(gen7_blt_regs
), false },
547 { hsw_master_regs
, ARRAY_SIZE(hsw_master_regs
), true },
550 static u32
gen7_render_get_cmd_length_mask(u32 cmd_header
)
552 u32 client
= (cmd_header
& INSTR_CLIENT_MASK
) >> INSTR_CLIENT_SHIFT
;
554 (cmd_header
& INSTR_SUBCLIENT_MASK
) >> INSTR_SUBCLIENT_SHIFT
;
556 if (client
== INSTR_MI_CLIENT
)
558 else if (client
== INSTR_RC_CLIENT
) {
559 if (subclient
== INSTR_MEDIA_SUBCLIENT
)
565 DRM_DEBUG_DRIVER("CMD: Abnormal rcs cmd length! 0x%08X\n", cmd_header
);
569 static u32
gen7_bsd_get_cmd_length_mask(u32 cmd_header
)
571 u32 client
= (cmd_header
& INSTR_CLIENT_MASK
) >> INSTR_CLIENT_SHIFT
;
573 (cmd_header
& INSTR_SUBCLIENT_MASK
) >> INSTR_SUBCLIENT_SHIFT
;
574 u32 op
= (cmd_header
& INSTR_26_TO_24_MASK
) >> INSTR_26_TO_24_SHIFT
;
576 if (client
== INSTR_MI_CLIENT
)
578 else if (client
== INSTR_RC_CLIENT
) {
579 if (subclient
== INSTR_MEDIA_SUBCLIENT
) {
588 DRM_DEBUG_DRIVER("CMD: Abnormal bsd cmd length! 0x%08X\n", cmd_header
);
592 static u32
gen7_blt_get_cmd_length_mask(u32 cmd_header
)
594 u32 client
= (cmd_header
& INSTR_CLIENT_MASK
) >> INSTR_CLIENT_SHIFT
;
596 if (client
== INSTR_MI_CLIENT
)
598 else if (client
== INSTR_BC_CLIENT
)
601 DRM_DEBUG_DRIVER("CMD: Abnormal blt cmd length! 0x%08X\n", cmd_header
);
605 static bool validate_cmds_sorted(struct intel_engine_cs
*engine
,
606 const struct drm_i915_cmd_table
*cmd_tables
,
612 if (!cmd_tables
|| cmd_table_count
== 0)
615 for (i
= 0; i
< cmd_table_count
; i
++) {
616 const struct drm_i915_cmd_table
*table
= &cmd_tables
[i
];
620 for (j
= 0; j
< table
->count
; j
++) {
621 const struct drm_i915_cmd_descriptor
*desc
=
623 u32 curr
= desc
->cmd
.value
& desc
->cmd
.mask
;
625 if (curr
< previous
) {
626 DRM_ERROR("CMD: table not sorted ring=%d table=%d entry=%d cmd=0x%08X prev=0x%08X\n",
627 engine
->id
, i
, j
, curr
, previous
);
638 static bool check_sorted(int ring_id
,
639 const struct drm_i915_reg_descriptor
*reg_table
,
646 for (i
= 0; i
< reg_count
; i
++) {
647 u32 curr
= i915_mmio_reg_offset(reg_table
[i
].addr
);
649 if (curr
< previous
) {
650 DRM_ERROR("CMD: table not sorted ring=%d entry=%d reg=0x%08X prev=0x%08X\n",
651 ring_id
, i
, curr
, previous
);
661 static bool validate_regs_sorted(struct intel_engine_cs
*engine
)
664 const struct drm_i915_reg_table
*table
;
666 for (i
= 0; i
< engine
->reg_table_count
; i
++) {
667 table
= &engine
->reg_tables
[i
];
668 if (!check_sorted(engine
->id
, table
->regs
, table
->num_regs
))
676 const struct drm_i915_cmd_descriptor
*desc
;
677 struct hlist_node node
;
681 * Different command ranges have different numbers of bits for the opcode. For
682 * example, MI commands use bits 31:23 while 3D commands use bits 31:16. The
683 * problem is that, for example, MI commands use bits 22:16 for other fields
684 * such as GGTT vs PPGTT bits. If we include those bits in the mask then when
685 * we mask a command from a batch it could hash to the wrong bucket due to
686 * non-opcode bits being set. But if we don't include those bits, some 3D
687 * commands may hash to the same bucket due to not including opcode bits that
688 * make the command unique. For now, we will risk hashing to the same bucket.
690 * If we attempt to generate a perfect hash, we should be able to look at bits
691 * 31:29 of a command from a batch buffer and use the full mask for that
692 * client. The existing INSTR_CLIENT_MASK/SHIFT defines can be used for this.
694 #define CMD_HASH_MASK STD_MI_OPCODE_MASK
696 static int init_hash_table(struct intel_engine_cs
*engine
,
697 const struct drm_i915_cmd_table
*cmd_tables
,
703 hash_init(engine
->cmd_hash
);
705 for (i
= 0; i
< cmd_table_count
; i
++) {
706 const struct drm_i915_cmd_table
*table
= &cmd_tables
[i
];
708 for (j
= 0; j
< table
->count
; j
++) {
709 const struct drm_i915_cmd_descriptor
*desc
=
711 struct cmd_node
*desc_node
=
712 kmalloc(sizeof(*desc_node
), GFP_KERNEL
);
717 desc_node
->desc
= desc
;
718 hash_add(engine
->cmd_hash
, &desc_node
->node
,
719 desc
->cmd
.value
& CMD_HASH_MASK
);
727 static void fini_hash_table(struct intel_engine_cs
*engine
)
730 struct hlist_node
*tmp
;
731 struct cmd_node
*desc_node
;
734 hash_for_each_safe(engine
->cmd_hash
, i
, tmp
, desc_node
, node
) {
735 hash_del(&desc_node
->node
);
742 * i915_cmd_parser_init_ring() - set cmd parser related fields for a ringbuffer
743 * @ring: the ringbuffer to initialize
745 * Optionally initializes fields related to batch buffer command parsing in the
746 * struct intel_engine_cs based on whether the platform requires software
749 * Return: non-zero if initialization fails
751 int i915_cmd_parser_init_ring(struct intel_engine_cs
*engine
)
753 const struct drm_i915_cmd_table
*cmd_tables
;
757 if (!IS_GEN7(engine
->dev
))
760 switch (engine
->id
) {
762 if (IS_HASWELL(engine
->dev
)) {
763 cmd_tables
= hsw_render_ring_cmds
;
765 ARRAY_SIZE(hsw_render_ring_cmds
);
767 cmd_tables
= gen7_render_cmds
;
768 cmd_table_count
= ARRAY_SIZE(gen7_render_cmds
);
771 if (IS_HASWELL(engine
->dev
)) {
772 engine
->reg_tables
= hsw_render_reg_tables
;
773 engine
->reg_table_count
= ARRAY_SIZE(hsw_render_reg_tables
);
775 engine
->reg_tables
= ivb_render_reg_tables
;
776 engine
->reg_table_count
= ARRAY_SIZE(ivb_render_reg_tables
);
779 engine
->get_cmd_length_mask
= gen7_render_get_cmd_length_mask
;
782 cmd_tables
= gen7_video_cmds
;
783 cmd_table_count
= ARRAY_SIZE(gen7_video_cmds
);
784 engine
->get_cmd_length_mask
= gen7_bsd_get_cmd_length_mask
;
787 if (IS_HASWELL(engine
->dev
)) {
788 cmd_tables
= hsw_blt_ring_cmds
;
789 cmd_table_count
= ARRAY_SIZE(hsw_blt_ring_cmds
);
791 cmd_tables
= gen7_blt_cmds
;
792 cmd_table_count
= ARRAY_SIZE(gen7_blt_cmds
);
795 if (IS_HASWELL(engine
->dev
)) {
796 engine
->reg_tables
= hsw_blt_reg_tables
;
797 engine
->reg_table_count
= ARRAY_SIZE(hsw_blt_reg_tables
);
799 engine
->reg_tables
= ivb_blt_reg_tables
;
800 engine
->reg_table_count
= ARRAY_SIZE(ivb_blt_reg_tables
);
803 engine
->get_cmd_length_mask
= gen7_blt_get_cmd_length_mask
;
806 cmd_tables
= hsw_vebox_cmds
;
807 cmd_table_count
= ARRAY_SIZE(hsw_vebox_cmds
);
808 /* VECS can use the same length_mask function as VCS */
809 engine
->get_cmd_length_mask
= gen7_bsd_get_cmd_length_mask
;
812 DRM_ERROR("CMD: cmd_parser_init with unknown ring: %d\n",
817 BUG_ON(!validate_cmds_sorted(engine
, cmd_tables
, cmd_table_count
));
818 BUG_ON(!validate_regs_sorted(engine
));
821 WARN_ON(!hash_empty(engine
->cmd_hash
));
824 ret
= init_hash_table(engine
, cmd_tables
, cmd_table_count
);
826 DRM_ERROR("CMD: cmd_parser_init failed!\n");
827 fini_hash_table(engine
);
831 engine
->needs_cmd_parser
= true;
837 * i915_cmd_parser_fini_ring() - clean up cmd parser related fields
838 * @ring: the ringbuffer to clean up
840 * Releases any resources related to command parsing that may have been
841 * initialized for the specified ring.
843 void i915_cmd_parser_fini_ring(struct intel_engine_cs
*engine
)
845 if (!engine
->needs_cmd_parser
)
848 fini_hash_table(engine
);
851 static const struct drm_i915_cmd_descriptor
*
852 find_cmd_in_table(struct intel_engine_cs
*engine
,
856 struct cmd_node
*desc_node
;
858 hash_for_each_possible(engine
->cmd_hash
, desc_node
, node
,
859 cmd_header
& CMD_HASH_MASK
) {
860 const struct drm_i915_cmd_descriptor
*desc
= desc_node
->desc
;
861 u32 masked_cmd
= desc
->cmd
.mask
& cmd_header
;
862 u32 masked_value
= desc
->cmd
.value
& desc
->cmd
.mask
;
864 if (masked_cmd
== masked_value
)
873 * Returns a pointer to a descriptor for the command specified by cmd_header.
875 * The caller must supply space for a default descriptor via the default_desc
876 * parameter. If no descriptor for the specified command exists in the ring's
877 * command parser tables, this function fills in default_desc based on the
878 * ring's default length encoding and returns default_desc.
880 static const struct drm_i915_cmd_descriptor
*
881 find_cmd(struct intel_engine_cs
*engine
,
883 struct drm_i915_cmd_descriptor
*default_desc
)
885 const struct drm_i915_cmd_descriptor
*desc
;
888 desc
= find_cmd_in_table(engine
, cmd_header
);
892 mask
= engine
->get_cmd_length_mask(cmd_header
);
896 BUG_ON(!default_desc
);
897 default_desc
->flags
= CMD_DESC_SKIP
;
898 default_desc
->length
.mask
= mask
;
903 static const struct drm_i915_reg_descriptor
*
904 find_reg(const struct drm_i915_reg_descriptor
*table
,
909 for (i
= 0; i
< count
; i
++) {
910 if (i915_mmio_reg_offset(table
[i
].addr
) == addr
)
917 static const struct drm_i915_reg_descriptor
*
918 find_reg_in_tables(const struct drm_i915_reg_table
*tables
,
919 int count
, bool is_master
, u32 addr
)
922 const struct drm_i915_reg_table
*table
;
923 const struct drm_i915_reg_descriptor
*reg
;
925 for (i
= 0; i
< count
; i
++) {
927 if (!table
->master
|| is_master
) {
928 reg
= find_reg(table
->regs
, table
->num_regs
,
938 static u32
*vmap_batch(struct drm_i915_gem_object
*obj
,
939 unsigned start
, unsigned len
)
943 struct sg_page_iter sg_iter
;
944 int first_page
= start
>> PAGE_SHIFT
;
945 int last_page
= (len
+ start
+ 4095) >> PAGE_SHIFT
;
946 int npages
= last_page
- first_page
;
947 struct vm_page
**pages
;
949 pages
= drm_malloc_ab(npages
, sizeof(*pages
));
951 DRM_DEBUG_DRIVER("Failed to get space for pages\n");
956 for_each_sg_page(obj
->pages
->sgl
, &sg_iter
, obj
->pages
->nents
, first_page
) {
957 pages
[i
++] = sg_page_iter_page(&sg_iter
);
963 addr
= vmap(pages
, i
, 0, PAGE_KERNEL
);
965 DRM_DEBUG_DRIVER("Failed to vmap pages\n");
972 drm_free_large(pages
);
976 /* Returns a vmap'd pointer to dest_obj, which the caller must unmap */
977 static u32
*copy_batch(struct drm_i915_gem_object
*dest_obj
,
978 struct drm_i915_gem_object
*src_obj
,
979 u32 batch_start_offset
,
982 int needs_clflush
= 0;
983 char *src_base
, *src
;
987 if (batch_len
> dest_obj
->base
.size
||
988 batch_len
+ batch_start_offset
> src_obj
->base
.size
)
989 return ERR_PTR(-E2BIG
);
991 if (WARN_ON(dest_obj
->pages_pin_count
== 0))
992 return ERR_PTR(-ENODEV
);
994 ret
= i915_gem_obj_prepare_shmem_read(src_obj
, &needs_clflush
);
996 DRM_DEBUG_DRIVER("CMD: failed to prepare shadow batch\n");
1000 src_base
= (char *)vmap_batch(src_obj
, batch_start_offset
, batch_len
);
1002 DRM_DEBUG_DRIVER("CMD: Failed to vmap batch\n");
1007 ret
= i915_gem_object_set_to_cpu_domain(dest_obj
, true);
1009 DRM_DEBUG_DRIVER("CMD: Failed to set shadow batch to CPU\n");
1013 dst
= vmap_batch(dest_obj
, 0, batch_len
);
1015 DRM_DEBUG_DRIVER("CMD: Failed to vmap shadow batch\n");
1020 src
= src_base
+ offset_in_page(batch_start_offset
);
1022 drm_clflush_virt_range(src
, batch_len
);
1024 memcpy(dst
, src
, batch_len
);
1029 i915_gem_object_unpin_pages(src_obj
);
1031 return ret
? ERR_PTR(ret
) : dst
;
1035 * i915_needs_cmd_parser() - should a given ring use software command parsing?
1036 * @ring: the ring in question
1038 * Only certain platforms require software batch buffer command parsing, and
1039 * only when enabled via module parameter.
1041 * Return: true if the ring requires software command parsing
1043 bool i915_needs_cmd_parser(struct intel_engine_cs
*engine
)
1045 if (!engine
->needs_cmd_parser
)
1048 if (!USES_PPGTT(engine
->dev
))
1051 return (i915
.enable_cmd_parser
== 1);
1054 static bool check_cmd(const struct intel_engine_cs
*engine
,
1055 const struct drm_i915_cmd_descriptor
*desc
,
1056 const u32
*cmd
, u32 length
,
1057 const bool is_master
,
1058 bool *oacontrol_set
)
1060 if (desc
->flags
& CMD_DESC_REJECT
) {
1061 DRM_DEBUG_DRIVER("CMD: Rejected command: 0x%08X\n", *cmd
);
1065 if ((desc
->flags
& CMD_DESC_MASTER
) && !is_master
) {
1066 DRM_DEBUG_DRIVER("CMD: Rejected master-only command: 0x%08X\n",
1071 if (desc
->flags
& CMD_DESC_REGISTER
) {
1073 * Get the distance between individual register offset
1074 * fields if the command can perform more than one
1077 const u32 step
= desc
->reg
.step
? desc
->reg
.step
: length
;
1080 for (offset
= desc
->reg
.offset
; offset
< length
;
1082 const u32 reg_addr
= cmd
[offset
] & desc
->reg
.mask
;
1083 const struct drm_i915_reg_descriptor
*reg
=
1084 find_reg_in_tables(engine
->reg_tables
,
1085 engine
->reg_table_count
,
1090 DRM_DEBUG_DRIVER("CMD: Rejected register 0x%08X in command: 0x%08X (ring=%d)\n",
1091 reg_addr
, *cmd
, engine
->id
);
1096 * OACONTROL requires some special handling for
1097 * writes. We want to make sure that any batch which
1098 * enables OA also disables it before the end of the
1099 * batch. The goal is to prevent one process from
1100 * snooping on the perf data from another process. To do
1101 * that, we need to check the value that will be written
1102 * to the register. Hence, limit OACONTROL writes to
1103 * only MI_LOAD_REGISTER_IMM commands.
1105 if (reg_addr
== i915_mmio_reg_offset(OACONTROL
)) {
1106 if (desc
->cmd
.value
== MI_LOAD_REGISTER_MEM
) {
1107 DRM_DEBUG_DRIVER("CMD: Rejected LRM to OACONTROL\n");
1111 if (desc
->cmd
.value
== MI_LOAD_REGISTER_IMM(1))
1112 *oacontrol_set
= (cmd
[offset
+ 1] != 0);
1116 * Check the value written to the register against the
1117 * allowed mask/value pair given in the whitelist entry.
1120 if (desc
->cmd
.value
== MI_LOAD_REGISTER_MEM
) {
1121 DRM_DEBUG_DRIVER("CMD: Rejected LRM to masked register 0x%08X\n",
1126 if (desc
->cmd
.value
== MI_LOAD_REGISTER_IMM(1) &&
1127 (offset
+ 2 > length
||
1128 (cmd
[offset
+ 1] & reg
->mask
) != reg
->value
)) {
1129 DRM_DEBUG_DRIVER("CMD: Rejected LRI to masked register 0x%08X\n",
1137 if (desc
->flags
& CMD_DESC_BITMASK
) {
1140 for (i
= 0; i
< MAX_CMD_DESC_BITMASKS
; i
++) {
1143 if (desc
->bits
[i
].mask
== 0)
1146 if (desc
->bits
[i
].condition_mask
!= 0) {
1148 desc
->bits
[i
].condition_offset
;
1149 u32 condition
= cmd
[offset
] &
1150 desc
->bits
[i
].condition_mask
;
1156 dword
= cmd
[desc
->bits
[i
].offset
] &
1159 if (dword
!= desc
->bits
[i
].expected
) {
1160 DRM_DEBUG_DRIVER("CMD: Rejected command 0x%08X for bitmask 0x%08X (exp=0x%08X act=0x%08X) (ring=%d)\n",
1163 desc
->bits
[i
].expected
,
1173 #define LENGTH_BIAS 2
1176 * i915_parse_cmds() - parse a submitted batch buffer for privilege violations
1177 * @ring: the ring on which the batch is to execute
1178 * @batch_obj: the batch buffer in question
1179 * @shadow_batch_obj: copy of the batch buffer in question
1180 * @batch_start_offset: byte offset in the batch at which execution starts
1181 * @batch_len: length of the commands in batch_obj
1182 * @is_master: is the submitting process the drm master?
1184 * Parses the specified batch buffer looking for privilege violations as
1185 * described in the overview.
1187 * Return: non-zero if the parser finds violations or otherwise fails; -EACCES
1188 * if the batch appears legal but should use hardware parsing
1190 int i915_parse_cmds(struct intel_engine_cs
*engine
,
1191 struct drm_i915_gem_object
*batch_obj
,
1192 struct drm_i915_gem_object
*shadow_batch_obj
,
1193 u32 batch_start_offset
,
1197 u32
*cmd
, *batch_base
, *batch_end
;
1198 struct drm_i915_cmd_descriptor default_desc
= { 0 };
1199 bool oacontrol_set
= false; /* OACONTROL tracking. See check_cmd() */
1202 batch_base
= copy_batch(shadow_batch_obj
, batch_obj
,
1203 batch_start_offset
, batch_len
);
1204 if (IS_ERR(batch_base
)) {
1205 DRM_DEBUG_DRIVER("CMD: Failed to copy batch\n");
1206 return PTR_ERR(batch_base
);
1210 * We use the batch length as size because the shadow object is as
1211 * large or larger and copy_batch() will write MI_NOPs to the extra
1212 * space. Parsing should be faster in some cases this way.
1214 batch_end
= batch_base
+ (batch_len
/ sizeof(*batch_end
));
1217 while (cmd
< batch_end
) {
1218 const struct drm_i915_cmd_descriptor
*desc
;
1221 if (*cmd
== MI_BATCH_BUFFER_END
)
1224 desc
= find_cmd(engine
, *cmd
, &default_desc
);
1226 DRM_DEBUG_DRIVER("CMD: Unrecognized command: 0x%08X\n",
1233 * If the batch buffer contains a chained batch, return an
1234 * error that tells the caller to abort and dispatch the
1235 * workload as a non-secure batch.
1237 if (desc
->cmd
.value
== MI_BATCH_BUFFER_START
) {
1242 if (desc
->flags
& CMD_DESC_FIXED
)
1243 length
= desc
->length
.fixed
;
1245 length
= ((*cmd
& desc
->length
.mask
) + LENGTH_BIAS
);
1247 if ((batch_end
- cmd
) < length
) {
1248 DRM_DEBUG_DRIVER("CMD: Command length exceeds batch length: 0x%08X length=%u batchlen=%td\n",
1256 if (!check_cmd(engine
, desc
, cmd
, length
, is_master
,
1265 if (oacontrol_set
) {
1266 DRM_DEBUG_DRIVER("CMD: batch set OACONTROL but did not clear it\n");
1270 if (cmd
>= batch_end
) {
1271 DRM_DEBUG_DRIVER("CMD: Got to the end of the buffer w/o a BBE cmd!\n");
1281 * i915_cmd_parser_get_version() - get the cmd parser version number
1283 * The cmd parser maintains a simple increasing integer version number suitable
1284 * for passing to userspace clients to determine what operations are permitted.
1286 * Return: the current version number of the cmd parser
1288 int i915_cmd_parser_get_version(void)
1291 * Command parser version history
1293 * 1. Initial version. Checks batches and reports violations, but leaves
1294 * hardware parsing enabled (so does not allow new use cases).
1295 * 2. Allow access to the MI_PREDICATE_SRC0 and
1296 * MI_PREDICATE_SRC1 registers.
1297 * 3. Allow access to the GPGPU_THREADS_DISPATCHED register.
1298 * 4. L3 atomic chicken bits of HSW_SCRATCH1 and HSW_ROW_CHICKEN3.
1299 * 5. GPGPU dispatch compute indirect registers.
1300 * 6. TIMESTAMP register and Haswell CS GPR registers