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 } ),
127 CMD( MI_STORE_REGISTER_MEM(1), SMI
, !F
, 0xFF, W
| B
,
128 .reg
= { .offset
= 1, .mask
= 0x007FFFFC },
131 .mask
= MI_GLOBAL_GTT
,
134 CMD( MI_LOAD_REGISTER_MEM
, SMI
, !F
, 0xFF, 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_DISPLAY_FLIP
, SMI
, !F
, 0xFF, R
),
155 CMD( MI_SET_APPID
, SMI
, F
, 1, S
),
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.
399 * Some registers that userspace accesses are 64 bits. The register
400 * access commands only allow 32-bit accesses. Hence, we have to include
401 * entries for both halves of the 64-bit registers.
404 /* Convenience macro for adding 64-bit registers */
405 #define REG64(addr) (addr), (addr + sizeof(u32))
407 static const u32 gen7_render_regs
[] = {
408 REG64(HS_INVOCATION_COUNT
),
409 REG64(DS_INVOCATION_COUNT
),
410 REG64(IA_VERTICES_COUNT
),
411 REG64(IA_PRIMITIVES_COUNT
),
412 REG64(VS_INVOCATION_COUNT
),
413 REG64(GS_INVOCATION_COUNT
),
414 REG64(GS_PRIMITIVES_COUNT
),
415 REG64(CL_INVOCATION_COUNT
),
416 REG64(CL_PRIMITIVES_COUNT
),
417 REG64(PS_INVOCATION_COUNT
),
418 REG64(PS_DEPTH_COUNT
),
419 OACONTROL
, /* Only allowed for LRI and SRM. See below. */
420 REG64(MI_PREDICATE_SRC0
),
421 REG64(MI_PREDICATE_SRC1
),
422 GEN7_3DPRIM_END_OFFSET
,
423 GEN7_3DPRIM_START_VERTEX
,
424 GEN7_3DPRIM_VERTEX_COUNT
,
425 GEN7_3DPRIM_INSTANCE_COUNT
,
426 GEN7_3DPRIM_START_INSTANCE
,
427 GEN7_3DPRIM_BASE_VERTEX
,
428 REG64(GEN7_SO_NUM_PRIMS_WRITTEN(0)),
429 REG64(GEN7_SO_NUM_PRIMS_WRITTEN(1)),
430 REG64(GEN7_SO_NUM_PRIMS_WRITTEN(2)),
431 REG64(GEN7_SO_NUM_PRIMS_WRITTEN(3)),
432 REG64(GEN7_SO_PRIM_STORAGE_NEEDED(0)),
433 REG64(GEN7_SO_PRIM_STORAGE_NEEDED(1)),
434 REG64(GEN7_SO_PRIM_STORAGE_NEEDED(2)),
435 REG64(GEN7_SO_PRIM_STORAGE_NEEDED(3)),
436 GEN7_SO_WRITE_OFFSET(0),
437 GEN7_SO_WRITE_OFFSET(1),
438 GEN7_SO_WRITE_OFFSET(2),
439 GEN7_SO_WRITE_OFFSET(3),
445 static const u32 gen7_blt_regs
[] = {
449 static const u32 ivb_master_regs
[] = {
452 GEN7_PIPE_DE_LOAD_SL(PIPE_A
),
453 GEN7_PIPE_DE_LOAD_SL(PIPE_B
),
454 GEN7_PIPE_DE_LOAD_SL(PIPE_C
),
457 static const u32 hsw_master_regs
[] = {
464 static u32
gen7_render_get_cmd_length_mask(u32 cmd_header
)
466 u32 client
= (cmd_header
& INSTR_CLIENT_MASK
) >> INSTR_CLIENT_SHIFT
;
468 (cmd_header
& INSTR_SUBCLIENT_MASK
) >> INSTR_SUBCLIENT_SHIFT
;
470 if (client
== INSTR_MI_CLIENT
)
472 else if (client
== INSTR_RC_CLIENT
) {
473 if (subclient
== INSTR_MEDIA_SUBCLIENT
)
479 DRM_DEBUG_DRIVER("CMD: Abnormal rcs cmd length! 0x%08X\n", cmd_header
);
483 static u32
gen7_bsd_get_cmd_length_mask(u32 cmd_header
)
485 u32 client
= (cmd_header
& INSTR_CLIENT_MASK
) >> INSTR_CLIENT_SHIFT
;
487 (cmd_header
& INSTR_SUBCLIENT_MASK
) >> INSTR_SUBCLIENT_SHIFT
;
488 u32 op
= (cmd_header
& INSTR_26_TO_24_MASK
) >> INSTR_26_TO_24_SHIFT
;
490 if (client
== INSTR_MI_CLIENT
)
492 else if (client
== INSTR_RC_CLIENT
) {
493 if (subclient
== INSTR_MEDIA_SUBCLIENT
) {
502 DRM_DEBUG_DRIVER("CMD: Abnormal bsd cmd length! 0x%08X\n", cmd_header
);
506 static u32
gen7_blt_get_cmd_length_mask(u32 cmd_header
)
508 u32 client
= (cmd_header
& INSTR_CLIENT_MASK
) >> INSTR_CLIENT_SHIFT
;
510 if (client
== INSTR_MI_CLIENT
)
512 else if (client
== INSTR_BC_CLIENT
)
515 DRM_DEBUG_DRIVER("CMD: Abnormal blt cmd length! 0x%08X\n", cmd_header
);
519 static bool validate_cmds_sorted(struct intel_engine_cs
*ring
,
520 const struct drm_i915_cmd_table
*cmd_tables
,
526 if (!cmd_tables
|| cmd_table_count
== 0)
529 for (i
= 0; i
< cmd_table_count
; i
++) {
530 const struct drm_i915_cmd_table
*table
= &cmd_tables
[i
];
534 for (j
= 0; j
< table
->count
; j
++) {
535 const struct drm_i915_cmd_descriptor
*desc
=
537 u32 curr
= desc
->cmd
.value
& desc
->cmd
.mask
;
539 if (curr
< previous
) {
540 DRM_ERROR("CMD: table not sorted ring=%d table=%d entry=%d cmd=0x%08X prev=0x%08X\n",
541 ring
->id
, i
, j
, curr
, previous
);
552 static bool check_sorted(int ring_id
, const u32
*reg_table
, int reg_count
)
558 for (i
= 0; i
< reg_count
; i
++) {
559 u32 curr
= reg_table
[i
];
561 if (curr
< previous
) {
562 DRM_ERROR("CMD: table not sorted ring=%d entry=%d reg=0x%08X prev=0x%08X\n",
563 ring_id
, i
, curr
, previous
);
573 static bool validate_regs_sorted(struct intel_engine_cs
*ring
)
575 return check_sorted(ring
->id
, ring
->reg_table
, ring
->reg_count
) &&
576 check_sorted(ring
->id
, ring
->master_reg_table
,
577 ring
->master_reg_count
);
581 const struct drm_i915_cmd_descriptor
*desc
;
582 struct hlist_node node
;
586 * Different command ranges have different numbers of bits for the opcode. For
587 * example, MI commands use bits 31:23 while 3D commands use bits 31:16. The
588 * problem is that, for example, MI commands use bits 22:16 for other fields
589 * such as GGTT vs PPGTT bits. If we include those bits in the mask then when
590 * we mask a command from a batch it could hash to the wrong bucket due to
591 * non-opcode bits being set. But if we don't include those bits, some 3D
592 * commands may hash to the same bucket due to not including opcode bits that
593 * make the command unique. For now, we will risk hashing to the same bucket.
595 * If we attempt to generate a perfect hash, we should be able to look at bits
596 * 31:29 of a command from a batch buffer and use the full mask for that
597 * client. The existing INSTR_CLIENT_MASK/SHIFT defines can be used for this.
599 #define CMD_HASH_MASK STD_MI_OPCODE_MASK
601 static int init_hash_table(struct intel_engine_cs
*ring
,
602 const struct drm_i915_cmd_table
*cmd_tables
,
607 hash_init(ring
->cmd_hash
);
609 for (i
= 0; i
< cmd_table_count
; i
++) {
610 const struct drm_i915_cmd_table
*table
= &cmd_tables
[i
];
612 for (j
= 0; j
< table
->count
; j
++) {
613 const struct drm_i915_cmd_descriptor
*desc
=
615 struct cmd_node
*desc_node
=
616 kmalloc(sizeof(*desc_node
), GFP_KERNEL
);
621 desc_node
->desc
= desc
;
622 hash_add(ring
->cmd_hash
, &desc_node
->node
,
623 desc
->cmd
.value
& CMD_HASH_MASK
);
630 static void fini_hash_table(struct intel_engine_cs
*ring
)
632 struct hlist_node
*tmp
;
633 struct cmd_node
*desc_node
;
636 hash_for_each_safe(ring
->cmd_hash
, i
, tmp
, desc_node
, node
) {
637 hash_del(&desc_node
->node
);
643 * i915_cmd_parser_init_ring() - set cmd parser related fields for a ringbuffer
644 * @ring: the ringbuffer to initialize
646 * Optionally initializes fields related to batch buffer command parsing in the
647 * struct intel_engine_cs based on whether the platform requires software
650 * Return: non-zero if initialization fails
652 int i915_cmd_parser_init_ring(struct intel_engine_cs
*ring
)
654 const struct drm_i915_cmd_table
*cmd_tables
;
658 if (!IS_GEN7(ring
->dev
))
663 if (IS_HASWELL(ring
->dev
)) {
664 cmd_tables
= hsw_render_ring_cmds
;
666 ARRAY_SIZE(hsw_render_ring_cmds
);
668 cmd_tables
= gen7_render_cmds
;
669 cmd_table_count
= ARRAY_SIZE(gen7_render_cmds
);
672 ring
->reg_table
= gen7_render_regs
;
673 ring
->reg_count
= ARRAY_SIZE(gen7_render_regs
);
675 if (IS_HASWELL(ring
->dev
)) {
676 ring
->master_reg_table
= hsw_master_regs
;
677 ring
->master_reg_count
= ARRAY_SIZE(hsw_master_regs
);
679 ring
->master_reg_table
= ivb_master_regs
;
680 ring
->master_reg_count
= ARRAY_SIZE(ivb_master_regs
);
683 ring
->get_cmd_length_mask
= gen7_render_get_cmd_length_mask
;
686 cmd_tables
= gen7_video_cmds
;
687 cmd_table_count
= ARRAY_SIZE(gen7_video_cmds
);
688 ring
->get_cmd_length_mask
= gen7_bsd_get_cmd_length_mask
;
691 if (IS_HASWELL(ring
->dev
)) {
692 cmd_tables
= hsw_blt_ring_cmds
;
693 cmd_table_count
= ARRAY_SIZE(hsw_blt_ring_cmds
);
695 cmd_tables
= gen7_blt_cmds
;
696 cmd_table_count
= ARRAY_SIZE(gen7_blt_cmds
);
699 ring
->reg_table
= gen7_blt_regs
;
700 ring
->reg_count
= ARRAY_SIZE(gen7_blt_regs
);
702 if (IS_HASWELL(ring
->dev
)) {
703 ring
->master_reg_table
= hsw_master_regs
;
704 ring
->master_reg_count
= ARRAY_SIZE(hsw_master_regs
);
706 ring
->master_reg_table
= ivb_master_regs
;
707 ring
->master_reg_count
= ARRAY_SIZE(ivb_master_regs
);
710 ring
->get_cmd_length_mask
= gen7_blt_get_cmd_length_mask
;
713 cmd_tables
= hsw_vebox_cmds
;
714 cmd_table_count
= ARRAY_SIZE(hsw_vebox_cmds
);
715 /* VECS can use the same length_mask function as VCS */
716 ring
->get_cmd_length_mask
= gen7_bsd_get_cmd_length_mask
;
719 DRM_ERROR("CMD: cmd_parser_init with unknown ring: %d\n",
724 BUG_ON(!validate_cmds_sorted(ring
, cmd_tables
, cmd_table_count
));
725 BUG_ON(!validate_regs_sorted(ring
));
727 WARN_ON(!hash_empty(ring
->cmd_hash
));
729 ret
= init_hash_table(ring
, cmd_tables
, cmd_table_count
);
731 DRM_ERROR("CMD: cmd_parser_init failed!\n");
732 fini_hash_table(ring
);
736 ring
->needs_cmd_parser
= true;
742 * i915_cmd_parser_fini_ring() - clean up cmd parser related fields
743 * @ring: the ringbuffer to clean up
745 * Releases any resources related to command parsing that may have been
746 * initialized for the specified ring.
748 void i915_cmd_parser_fini_ring(struct intel_engine_cs
*ring
)
750 if (!ring
->needs_cmd_parser
)
753 fini_hash_table(ring
);
756 static const struct drm_i915_cmd_descriptor
*
757 find_cmd_in_table(struct intel_engine_cs
*ring
,
760 struct cmd_node
*desc_node
;
762 hash_for_each_possible(ring
->cmd_hash
, desc_node
, node
,
763 cmd_header
& CMD_HASH_MASK
) {
764 const struct drm_i915_cmd_descriptor
*desc
= desc_node
->desc
;
765 u32 masked_cmd
= desc
->cmd
.mask
& cmd_header
;
766 u32 masked_value
= desc
->cmd
.value
& desc
->cmd
.mask
;
768 if (masked_cmd
== masked_value
)
776 * Returns a pointer to a descriptor for the command specified by cmd_header.
778 * The caller must supply space for a default descriptor via the default_desc
779 * parameter. If no descriptor for the specified command exists in the ring's
780 * command parser tables, this function fills in default_desc based on the
781 * ring's default length encoding and returns default_desc.
783 static const struct drm_i915_cmd_descriptor
*
784 find_cmd(struct intel_engine_cs
*ring
,
786 struct drm_i915_cmd_descriptor
*default_desc
)
788 const struct drm_i915_cmd_descriptor
*desc
;
791 desc
= find_cmd_in_table(ring
, cmd_header
);
795 mask
= ring
->get_cmd_length_mask(cmd_header
);
799 BUG_ON(!default_desc
);
800 default_desc
->flags
= CMD_DESC_SKIP
;
801 default_desc
->length
.mask
= mask
;
806 static bool valid_reg(const u32
*table
, int count
, u32 addr
)
808 if (table
&& count
!= 0) {
811 for (i
= 0; i
< count
; i
++) {
812 if (table
[i
] == addr
)
820 static u32
*vmap_batch(struct drm_i915_gem_object
*obj
)
824 struct sg_page_iter sg_iter
;
827 pages
= drm_malloc_ab(obj
->base
.size
>> PAGE_SHIFT
, sizeof(*pages
));
829 DRM_DEBUG_DRIVER("Failed to get space for pages\n");
834 for_each_sg_page(obj
->pages
->sgl
, &sg_iter
, obj
->pages
->nents
, 0) {
835 pages
[i
] = sg_page_iter_page(&sg_iter
);
839 addr
= vmap(pages
, i
, 0, PAGE_KERNEL
);
841 DRM_DEBUG_DRIVER("Failed to vmap pages\n");
847 drm_free_large(pages
);
852 * i915_needs_cmd_parser() - should a given ring use software command parsing?
853 * @ring: the ring in question
855 * Only certain platforms require software batch buffer command parsing, and
856 * only when enabled via module parameter.
858 * Return: true if the ring requires software command parsing
860 bool i915_needs_cmd_parser(struct intel_engine_cs
*ring
)
862 if (!ring
->needs_cmd_parser
)
865 if (!USES_PPGTT(ring
->dev
))
868 return (i915
.enable_cmd_parser
== 1);
871 static bool check_cmd(const struct intel_engine_cs
*ring
,
872 const struct drm_i915_cmd_descriptor
*desc
,
874 const bool is_master
,
877 if (desc
->flags
& CMD_DESC_REJECT
) {
878 DRM_DEBUG_DRIVER("CMD: Rejected command: 0x%08X\n", *cmd
);
882 if ((desc
->flags
& CMD_DESC_MASTER
) && !is_master
) {
883 DRM_DEBUG_DRIVER("CMD: Rejected master-only command: 0x%08X\n",
888 if (desc
->flags
& CMD_DESC_REGISTER
) {
889 u32 reg_addr
= cmd
[desc
->reg
.offset
] & desc
->reg
.mask
;
892 * OACONTROL requires some special handling for writes. We
893 * want to make sure that any batch which enables OA also
894 * disables it before the end of the batch. The goal is to
895 * prevent one process from snooping on the perf data from
896 * another process. To do that, we need to check the value
897 * that will be written to the register. Hence, limit
898 * OACONTROL writes to only MI_LOAD_REGISTER_IMM commands.
900 if (reg_addr
== OACONTROL
) {
901 if (desc
->cmd
.value
== MI_LOAD_REGISTER_MEM
) {
902 DRM_DEBUG_DRIVER("CMD: Rejected LRM to OACONTROL\n");
906 if (desc
->cmd
.value
== MI_LOAD_REGISTER_IMM(1))
907 *oacontrol_set
= (cmd
[2] != 0);
910 if (!valid_reg(ring
->reg_table
,
911 ring
->reg_count
, reg_addr
)) {
913 !valid_reg(ring
->master_reg_table
,
914 ring
->master_reg_count
,
916 DRM_DEBUG_DRIVER("CMD: Rejected register 0x%08X in command: 0x%08X (ring=%d)\n",
925 if (desc
->flags
& CMD_DESC_BITMASK
) {
928 for (i
= 0; i
< MAX_CMD_DESC_BITMASKS
; i
++) {
931 if (desc
->bits
[i
].mask
== 0)
934 if (desc
->bits
[i
].condition_mask
!= 0) {
936 desc
->bits
[i
].condition_offset
;
937 u32 condition
= cmd
[offset
] &
938 desc
->bits
[i
].condition_mask
;
944 dword
= cmd
[desc
->bits
[i
].offset
] &
947 if (dword
!= desc
->bits
[i
].expected
) {
948 DRM_DEBUG_DRIVER("CMD: Rejected command 0x%08X for bitmask 0x%08X (exp=0x%08X act=0x%08X) (ring=%d)\n",
951 desc
->bits
[i
].expected
,
961 #define LENGTH_BIAS 2
964 * i915_parse_cmds() - parse a submitted batch buffer for privilege violations
965 * @ring: the ring on which the batch is to execute
966 * @batch_obj: the batch buffer in question
967 * @batch_start_offset: byte offset in the batch at which execution starts
968 * @is_master: is the submitting process the drm master?
970 * Parses the specified batch buffer looking for privilege violations as
971 * described in the overview.
973 * Return: non-zero if the parser finds violations or otherwise fails; -EACCES
974 * if the batch appears legal but should use hardware parsing
976 int i915_parse_cmds(struct intel_engine_cs
*ring
,
977 struct drm_i915_gem_object
*batch_obj
,
978 u32 batch_start_offset
,
982 u32
*cmd
, *batch_base
, *batch_end
;
983 struct drm_i915_cmd_descriptor default_desc
= { 0 };
984 int needs_clflush
= 0;
985 bool oacontrol_set
= false; /* OACONTROL tracking. See check_cmd() */
987 ret
= i915_gem_obj_prepare_shmem_read(batch_obj
, &needs_clflush
);
989 DRM_DEBUG_DRIVER("CMD: failed to prep read\n");
993 batch_base
= vmap_batch(batch_obj
);
995 DRM_DEBUG_DRIVER("CMD: Failed to vmap batch\n");
996 i915_gem_object_unpin_pages(batch_obj
);
1001 drm_clflush_virt_range((char *)batch_base
, batch_obj
->base
.size
);
1003 cmd
= batch_base
+ (batch_start_offset
/ sizeof(*cmd
));
1004 batch_end
= cmd
+ (batch_obj
->base
.size
/ sizeof(*batch_end
));
1006 while (cmd
< batch_end
) {
1007 const struct drm_i915_cmd_descriptor
*desc
;
1010 if (*cmd
== MI_BATCH_BUFFER_END
)
1013 desc
= find_cmd(ring
, *cmd
, &default_desc
);
1015 DRM_DEBUG_DRIVER("CMD: Unrecognized command: 0x%08X\n",
1022 * If the batch buffer contains a chained batch, return an
1023 * error that tells the caller to abort and dispatch the
1024 * workload as a non-secure batch.
1026 if (desc
->cmd
.value
== MI_BATCH_BUFFER_START
) {
1031 if (desc
->flags
& CMD_DESC_FIXED
)
1032 length
= desc
->length
.fixed
;
1034 length
= ((*cmd
& desc
->length
.mask
) + LENGTH_BIAS
);
1036 if ((batch_end
- cmd
) < length
) {
1037 DRM_DEBUG_DRIVER("CMD: Command length exceeds batch length: 0x%08X length=%u batchlen=%td\n",
1045 if (!check_cmd(ring
, desc
, cmd
, is_master
, &oacontrol_set
)) {
1053 if (oacontrol_set
) {
1054 DRM_DEBUG_DRIVER("CMD: batch set OACONTROL but did not clear it\n");
1058 if (cmd
>= batch_end
) {
1059 DRM_DEBUG_DRIVER("CMD: Got to the end of the buffer w/o a BBE cmd!\n");
1065 i915_gem_object_unpin_pages(batch_obj
);
1071 * i915_cmd_parser_get_version() - get the cmd parser version number
1073 * The cmd parser maintains a simple increasing integer version number suitable
1074 * for passing to userspace clients to determine what operations are permitted.
1076 * Return: the current version number of the cmd parser
1078 int i915_cmd_parser_get_version(void)
1081 * Command parser version history
1083 * 1. Initial version. Checks batches and reports violations, but leaves
1084 * hardware parsing enabled (so does not allow new use cases).
1085 * 2. Allow access to the MI_PREDICATE_SRC0 and
1086 * MI_PREDICATE_SRC1 registers.