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 * implementated 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
,
141 CMD( MI_BATCH_BUFFER_START
, SMI
, !F
, 0xFF, S
),
144 static const struct drm_i915_cmd_descriptor render_cmds
[] = {
145 CMD( MI_FLUSH
, SMI
, F
, 1, S
),
146 CMD( MI_ARB_ON_OFF
, SMI
, F
, 1, R
),
147 CMD( MI_PREDICATE
, SMI
, F
, 1, S
),
148 CMD( MI_TOPOLOGY_FILTER
, SMI
, F
, 1, S
),
149 CMD( MI_DISPLAY_FLIP
, SMI
, !F
, 0xFF, R
),
150 CMD( MI_SET_CONTEXT
, SMI
, !F
, 0xFF, R
),
151 CMD( MI_URB_CLEAR
, SMI
, !F
, 0xFF, S
),
152 CMD( MI_STORE_DWORD_IMM
, SMI
, !F
, 0x3F, B
,
155 .mask
= MI_GLOBAL_GTT
,
158 CMD( MI_UPDATE_GTT
, SMI
, !F
, 0xFF, R
),
159 CMD( MI_CLFLUSH
, SMI
, !F
, 0x3FF, B
,
162 .mask
= MI_GLOBAL_GTT
,
165 CMD( MI_REPORT_PERF_COUNT
, SMI
, !F
, 0x3F, B
,
168 .mask
= MI_REPORT_PERF_COUNT_GGTT
,
171 CMD( MI_CONDITIONAL_BATCH_BUFFER_END
, SMI
, !F
, 0xFF, B
,
174 .mask
= MI_GLOBAL_GTT
,
177 CMD( GFX_OP_3DSTATE_VF_STATISTICS
, S3D
, F
, 1, S
),
178 CMD( PIPELINE_SELECT
, S3D
, F
, 1, S
),
179 CMD( MEDIA_VFE_STATE
, S3D
, !F
, 0xFFFF, B
,
182 .mask
= MEDIA_VFE_STATE_MMIO_ACCESS_MASK
,
185 CMD( GPGPU_OBJECT
, S3D
, !F
, 0xFF, S
),
186 CMD( GPGPU_WALKER
, S3D
, !F
, 0xFF, S
),
187 CMD( GFX_OP_3DSTATE_SO_DECL_LIST
, S3D
, !F
, 0x1FF, S
),
188 CMD( GFX_OP_PIPE_CONTROL(5), S3D
, !F
, 0xFF, B
,
191 .mask
= (PIPE_CONTROL_MMIO_WRITE
| PIPE_CONTROL_NOTIFY
),
196 .mask
= (PIPE_CONTROL_GLOBAL_GTT_IVB
|
197 PIPE_CONTROL_STORE_DATA_INDEX
),
199 .condition_offset
= 1,
200 .condition_mask
= PIPE_CONTROL_POST_SYNC_OP_MASK
,
204 static const struct drm_i915_cmd_descriptor hsw_render_cmds
[] = {
205 CMD( MI_SET_PREDICATE
, SMI
, F
, 1, S
),
206 CMD( MI_RS_CONTROL
, SMI
, F
, 1, S
),
207 CMD( MI_URB_ATOMIC_ALLOC
, SMI
, F
, 1, S
),
208 CMD( MI_RS_CONTEXT
, SMI
, F
, 1, S
),
209 CMD( MI_LOAD_SCAN_LINES_INCL
, SMI
, !F
, 0x3F, M
),
210 CMD( MI_LOAD_SCAN_LINES_EXCL
, SMI
, !F
, 0x3F, R
),
211 CMD( MI_LOAD_REGISTER_REG
, SMI
, !F
, 0xFF, R
),
212 CMD( MI_RS_STORE_DATA_IMM
, SMI
, !F
, 0xFF, S
),
213 CMD( MI_LOAD_URB_MEM
, SMI
, !F
, 0xFF, S
),
214 CMD( MI_STORE_URB_MEM
, SMI
, !F
, 0xFF, S
),
215 CMD( GFX_OP_3DSTATE_DX9_CONSTANTF_VS
, S3D
, !F
, 0x7FF, S
),
216 CMD( GFX_OP_3DSTATE_DX9_CONSTANTF_PS
, S3D
, !F
, 0x7FF, S
),
218 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_VS
, S3D
, !F
, 0x1FF, S
),
219 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_GS
, S3D
, !F
, 0x1FF, S
),
220 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_HS
, S3D
, !F
, 0x1FF, S
),
221 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_DS
, S3D
, !F
, 0x1FF, S
),
222 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_PS
, S3D
, !F
, 0x1FF, S
),
225 static const struct drm_i915_cmd_descriptor video_cmds
[] = {
226 CMD( MI_ARB_ON_OFF
, SMI
, F
, 1, R
),
227 CMD( MI_STORE_DWORD_IMM
, SMI
, !F
, 0xFF, B
,
230 .mask
= MI_GLOBAL_GTT
,
233 CMD( MI_UPDATE_GTT
, SMI
, !F
, 0x3F, R
),
234 CMD( MI_FLUSH_DW
, SMI
, !F
, 0x3F, B
,
237 .mask
= MI_FLUSH_DW_NOTIFY
,
242 .mask
= MI_FLUSH_DW_USE_GTT
,
244 .condition_offset
= 0,
245 .condition_mask
= MI_FLUSH_DW_OP_MASK
,
249 .mask
= MI_FLUSH_DW_STORE_INDEX
,
251 .condition_offset
= 0,
252 .condition_mask
= MI_FLUSH_DW_OP_MASK
,
254 CMD( MI_CONDITIONAL_BATCH_BUFFER_END
, SMI
, !F
, 0xFF, B
,
257 .mask
= MI_GLOBAL_GTT
,
261 * MFX_WAIT doesn't fit the way we handle length for most commands.
262 * It has a length field but it uses a non-standard length bias.
263 * It is always 1 dword though, so just treat it as fixed length.
265 CMD( MFX_WAIT
, SMFX
, F
, 1, S
),
268 static const struct drm_i915_cmd_descriptor vecs_cmds
[] = {
269 CMD( MI_ARB_ON_OFF
, SMI
, F
, 1, R
),
270 CMD( MI_STORE_DWORD_IMM
, SMI
, !F
, 0xFF, B
,
273 .mask
= MI_GLOBAL_GTT
,
276 CMD( MI_UPDATE_GTT
, SMI
, !F
, 0x3F, R
),
277 CMD( MI_FLUSH_DW
, SMI
, !F
, 0x3F, B
,
280 .mask
= MI_FLUSH_DW_NOTIFY
,
285 .mask
= MI_FLUSH_DW_USE_GTT
,
287 .condition_offset
= 0,
288 .condition_mask
= MI_FLUSH_DW_OP_MASK
,
292 .mask
= MI_FLUSH_DW_STORE_INDEX
,
294 .condition_offset
= 0,
295 .condition_mask
= MI_FLUSH_DW_OP_MASK
,
297 CMD( MI_CONDITIONAL_BATCH_BUFFER_END
, SMI
, !F
, 0xFF, B
,
300 .mask
= MI_GLOBAL_GTT
,
305 static const struct drm_i915_cmd_descriptor blt_cmds
[] = {
306 CMD( MI_DISPLAY_FLIP
, SMI
, !F
, 0xFF, R
),
307 CMD( MI_STORE_DWORD_IMM
, SMI
, !F
, 0x3FF, B
,
310 .mask
= MI_GLOBAL_GTT
,
313 CMD( MI_UPDATE_GTT
, SMI
, !F
, 0x3F, R
),
314 CMD( MI_FLUSH_DW
, SMI
, !F
, 0x3F, B
,
317 .mask
= MI_FLUSH_DW_NOTIFY
,
322 .mask
= MI_FLUSH_DW_USE_GTT
,
324 .condition_offset
= 0,
325 .condition_mask
= MI_FLUSH_DW_OP_MASK
,
329 .mask
= MI_FLUSH_DW_STORE_INDEX
,
331 .condition_offset
= 0,
332 .condition_mask
= MI_FLUSH_DW_OP_MASK
,
334 CMD( COLOR_BLT
, S2D
, !F
, 0x3F, S
),
335 CMD( SRC_COPY_BLT
, S2D
, !F
, 0x3F, S
),
338 static const struct drm_i915_cmd_descriptor hsw_blt_cmds
[] = {
339 CMD( MI_LOAD_SCAN_LINES_INCL
, SMI
, !F
, 0x3F, M
),
340 CMD( MI_LOAD_SCAN_LINES_EXCL
, SMI
, !F
, 0x3F, R
),
355 static const struct drm_i915_cmd_table gen7_render_cmds
[] = {
356 { common_cmds
, ARRAY_SIZE(common_cmds
) },
357 { render_cmds
, ARRAY_SIZE(render_cmds
) },
360 static const struct drm_i915_cmd_table hsw_render_ring_cmds
[] = {
361 { common_cmds
, ARRAY_SIZE(common_cmds
) },
362 { render_cmds
, ARRAY_SIZE(render_cmds
) },
363 { hsw_render_cmds
, ARRAY_SIZE(hsw_render_cmds
) },
366 static const struct drm_i915_cmd_table gen7_video_cmds
[] = {
367 { common_cmds
, ARRAY_SIZE(common_cmds
) },
368 { video_cmds
, ARRAY_SIZE(video_cmds
) },
371 static const struct drm_i915_cmd_table hsw_vebox_cmds
[] = {
372 { common_cmds
, ARRAY_SIZE(common_cmds
) },
373 { vecs_cmds
, ARRAY_SIZE(vecs_cmds
) },
376 static const struct drm_i915_cmd_table gen7_blt_cmds
[] = {
377 { common_cmds
, ARRAY_SIZE(common_cmds
) },
378 { blt_cmds
, ARRAY_SIZE(blt_cmds
) },
381 static const struct drm_i915_cmd_table hsw_blt_ring_cmds
[] = {
382 { common_cmds
, ARRAY_SIZE(common_cmds
) },
383 { blt_cmds
, ARRAY_SIZE(blt_cmds
) },
384 { hsw_blt_cmds
, ARRAY_SIZE(hsw_blt_cmds
) },
388 * Register whitelists, sorted by increasing register offset.
390 * Some registers that userspace accesses are 64 bits. The register
391 * access commands only allow 32-bit accesses. Hence, we have to include
392 * entries for both halves of the 64-bit registers.
395 /* Convenience macro for adding 64-bit registers */
396 #define REG64(addr) (addr), (addr + sizeof(u32))
398 static const u32 gen7_render_regs
[] = {
399 REG64(HS_INVOCATION_COUNT
),
400 REG64(DS_INVOCATION_COUNT
),
401 REG64(IA_VERTICES_COUNT
),
402 REG64(IA_PRIMITIVES_COUNT
),
403 REG64(VS_INVOCATION_COUNT
),
404 REG64(GS_INVOCATION_COUNT
),
405 REG64(GS_PRIMITIVES_COUNT
),
406 REG64(CL_INVOCATION_COUNT
),
407 REG64(CL_PRIMITIVES_COUNT
),
408 REG64(PS_INVOCATION_COUNT
),
409 REG64(PS_DEPTH_COUNT
),
410 OACONTROL
, /* Only allowed for LRI and SRM. See below. */
411 GEN7_3DPRIM_END_OFFSET
,
412 GEN7_3DPRIM_START_VERTEX
,
413 GEN7_3DPRIM_VERTEX_COUNT
,
414 GEN7_3DPRIM_INSTANCE_COUNT
,
415 GEN7_3DPRIM_START_INSTANCE
,
416 GEN7_3DPRIM_BASE_VERTEX
,
417 REG64(GEN7_SO_NUM_PRIMS_WRITTEN(0)),
418 REG64(GEN7_SO_NUM_PRIMS_WRITTEN(1)),
419 REG64(GEN7_SO_NUM_PRIMS_WRITTEN(2)),
420 REG64(GEN7_SO_NUM_PRIMS_WRITTEN(3)),
421 REG64(GEN7_SO_PRIM_STORAGE_NEEDED(0)),
422 REG64(GEN7_SO_PRIM_STORAGE_NEEDED(1)),
423 REG64(GEN7_SO_PRIM_STORAGE_NEEDED(2)),
424 REG64(GEN7_SO_PRIM_STORAGE_NEEDED(3)),
425 GEN7_SO_WRITE_OFFSET(0),
426 GEN7_SO_WRITE_OFFSET(1),
427 GEN7_SO_WRITE_OFFSET(2),
428 GEN7_SO_WRITE_OFFSET(3),
431 static const u32 gen7_blt_regs
[] = {
435 static const u32 ivb_master_regs
[] = {
438 GEN7_PIPE_DE_LOAD_SL(PIPE_A
),
439 GEN7_PIPE_DE_LOAD_SL(PIPE_B
),
440 GEN7_PIPE_DE_LOAD_SL(PIPE_C
),
443 static const u32 hsw_master_regs
[] = {
450 static u32
gen7_render_get_cmd_length_mask(u32 cmd_header
)
452 u32 client
= (cmd_header
& INSTR_CLIENT_MASK
) >> INSTR_CLIENT_SHIFT
;
454 (cmd_header
& INSTR_SUBCLIENT_MASK
) >> INSTR_SUBCLIENT_SHIFT
;
456 if (client
== INSTR_MI_CLIENT
)
458 else if (client
== INSTR_RC_CLIENT
) {
459 if (subclient
== INSTR_MEDIA_SUBCLIENT
)
465 DRM_DEBUG_DRIVER("CMD: Abnormal rcs cmd length! 0x%08X\n", cmd_header
);
469 static u32
gen7_bsd_get_cmd_length_mask(u32 cmd_header
)
471 u32 client
= (cmd_header
& INSTR_CLIENT_MASK
) >> INSTR_CLIENT_SHIFT
;
473 (cmd_header
& INSTR_SUBCLIENT_MASK
) >> INSTR_SUBCLIENT_SHIFT
;
475 if (client
== INSTR_MI_CLIENT
)
477 else if (client
== INSTR_RC_CLIENT
) {
478 if (subclient
== INSTR_MEDIA_SUBCLIENT
)
484 DRM_DEBUG_DRIVER("CMD: Abnormal bsd cmd length! 0x%08X\n", cmd_header
);
488 static u32
gen7_blt_get_cmd_length_mask(u32 cmd_header
)
490 u32 client
= (cmd_header
& INSTR_CLIENT_MASK
) >> INSTR_CLIENT_SHIFT
;
492 if (client
== INSTR_MI_CLIENT
)
494 else if (client
== INSTR_BC_CLIENT
)
497 DRM_DEBUG_DRIVER("CMD: Abnormal blt cmd length! 0x%08X\n", cmd_header
);
501 static bool validate_cmds_sorted(struct intel_engine_cs
*ring
,
502 const struct drm_i915_cmd_table
*cmd_tables
,
508 if (!cmd_tables
|| cmd_table_count
== 0)
511 for (i
= 0; i
< cmd_table_count
; i
++) {
512 const struct drm_i915_cmd_table
*table
= &cmd_tables
[i
];
516 for (j
= 0; j
< table
->count
; j
++) {
517 const struct drm_i915_cmd_descriptor
*desc
=
519 u32 curr
= desc
->cmd
.value
& desc
->cmd
.mask
;
521 if (curr
< previous
) {
522 DRM_ERROR("CMD: table not sorted ring=%d table=%d entry=%d cmd=0x%08X prev=0x%08X\n",
523 ring
->id
, i
, j
, curr
, previous
);
534 static bool check_sorted(int ring_id
, const u32
*reg_table
, int reg_count
)
540 for (i
= 0; i
< reg_count
; i
++) {
541 u32 curr
= reg_table
[i
];
543 if (curr
< previous
) {
544 DRM_ERROR("CMD: table not sorted ring=%d entry=%d reg=0x%08X prev=0x%08X\n",
545 ring_id
, i
, curr
, previous
);
555 static bool validate_regs_sorted(struct intel_engine_cs
*ring
)
557 return check_sorted(ring
->id
, ring
->reg_table
, ring
->reg_count
) &&
558 check_sorted(ring
->id
, ring
->master_reg_table
,
559 ring
->master_reg_count
);
563 const struct drm_i915_cmd_descriptor
*desc
;
564 struct hlist_node node
;
568 * Different command ranges have different numbers of bits for the opcode. For
569 * example, MI commands use bits 31:23 while 3D commands use bits 31:16. The
570 * problem is that, for example, MI commands use bits 22:16 for other fields
571 * such as GGTT vs PPGTT bits. If we include those bits in the mask then when
572 * we mask a command from a batch it could hash to the wrong bucket due to
573 * non-opcode bits being set. But if we don't include those bits, some 3D
574 * commands may hash to the same bucket due to not including opcode bits that
575 * make the command unique. For now, we will risk hashing to the same bucket.
577 * If we attempt to generate a perfect hash, we should be able to look at bits
578 * 31:29 of a command from a batch buffer and use the full mask for that
579 * client. The existing INSTR_CLIENT_MASK/SHIFT defines can be used for this.
581 #define CMD_HASH_MASK STD_MI_OPCODE_MASK
583 static int init_hash_table(struct intel_engine_cs
*ring
,
584 const struct drm_i915_cmd_table
*cmd_tables
,
589 hash_init(ring
->cmd_hash
);
591 for (i
= 0; i
< cmd_table_count
; i
++) {
592 const struct drm_i915_cmd_table
*table
= &cmd_tables
[i
];
594 for (j
= 0; j
< table
->count
; j
++) {
595 const struct drm_i915_cmd_descriptor
*desc
=
597 struct cmd_node
*desc_node
=
598 kmalloc(sizeof(*desc_node
), GFP_KERNEL
);
603 desc_node
->desc
= desc
;
604 hash_add(ring
->cmd_hash
, &desc_node
->node
,
605 desc
->cmd
.value
& CMD_HASH_MASK
);
612 static void fini_hash_table(struct intel_engine_cs
*ring
)
614 struct hlist_node
*tmp
;
615 struct cmd_node
*desc_node
;
618 hash_for_each_safe(ring
->cmd_hash
, i
, tmp
, desc_node
, node
) {
619 hash_del(&desc_node
->node
);
625 * i915_cmd_parser_init_ring() - set cmd parser related fields for a ringbuffer
626 * @ring: the ringbuffer to initialize
628 * Optionally initializes fields related to batch buffer command parsing in the
629 * struct intel_engine_cs based on whether the platform requires software
632 * Return: non-zero if initialization fails
634 int i915_cmd_parser_init_ring(struct intel_engine_cs
*ring
)
636 const struct drm_i915_cmd_table
*cmd_tables
;
640 if (!IS_GEN7(ring
->dev
))
645 if (IS_HASWELL(ring
->dev
)) {
646 cmd_tables
= hsw_render_ring_cmds
;
648 ARRAY_SIZE(hsw_render_ring_cmds
);
650 cmd_tables
= gen7_render_cmds
;
651 cmd_table_count
= ARRAY_SIZE(gen7_render_cmds
);
654 ring
->reg_table
= gen7_render_regs
;
655 ring
->reg_count
= ARRAY_SIZE(gen7_render_regs
);
657 if (IS_HASWELL(ring
->dev
)) {
658 ring
->master_reg_table
= hsw_master_regs
;
659 ring
->master_reg_count
= ARRAY_SIZE(hsw_master_regs
);
661 ring
->master_reg_table
= ivb_master_regs
;
662 ring
->master_reg_count
= ARRAY_SIZE(ivb_master_regs
);
665 ring
->get_cmd_length_mask
= gen7_render_get_cmd_length_mask
;
668 cmd_tables
= gen7_video_cmds
;
669 cmd_table_count
= ARRAY_SIZE(gen7_video_cmds
);
670 ring
->get_cmd_length_mask
= gen7_bsd_get_cmd_length_mask
;
673 if (IS_HASWELL(ring
->dev
)) {
674 cmd_tables
= hsw_blt_ring_cmds
;
675 cmd_table_count
= ARRAY_SIZE(hsw_blt_ring_cmds
);
677 cmd_tables
= gen7_blt_cmds
;
678 cmd_table_count
= ARRAY_SIZE(gen7_blt_cmds
);
681 ring
->reg_table
= gen7_blt_regs
;
682 ring
->reg_count
= ARRAY_SIZE(gen7_blt_regs
);
684 if (IS_HASWELL(ring
->dev
)) {
685 ring
->master_reg_table
= hsw_master_regs
;
686 ring
->master_reg_count
= ARRAY_SIZE(hsw_master_regs
);
688 ring
->master_reg_table
= ivb_master_regs
;
689 ring
->master_reg_count
= ARRAY_SIZE(ivb_master_regs
);
692 ring
->get_cmd_length_mask
= gen7_blt_get_cmd_length_mask
;
695 cmd_tables
= hsw_vebox_cmds
;
696 cmd_table_count
= ARRAY_SIZE(hsw_vebox_cmds
);
697 /* VECS can use the same length_mask function as VCS */
698 ring
->get_cmd_length_mask
= gen7_bsd_get_cmd_length_mask
;
701 DRM_ERROR("CMD: cmd_parser_init with unknown ring: %d\n",
706 BUG_ON(!validate_cmds_sorted(ring
, cmd_tables
, cmd_table_count
));
707 BUG_ON(!validate_regs_sorted(ring
));
709 ret
= init_hash_table(ring
, cmd_tables
, cmd_table_count
);
711 DRM_ERROR("CMD: cmd_parser_init failed!\n");
712 fini_hash_table(ring
);
716 ring
->needs_cmd_parser
= true;
722 * i915_cmd_parser_fini_ring() - clean up cmd parser related fields
723 * @ring: the ringbuffer to clean up
725 * Releases any resources related to command parsing that may have been
726 * initialized for the specified ring.
728 void i915_cmd_parser_fini_ring(struct intel_engine_cs
*ring
)
730 if (!ring
->needs_cmd_parser
)
733 fini_hash_table(ring
);
736 static const struct drm_i915_cmd_descriptor
*
737 find_cmd_in_table(struct intel_engine_cs
*ring
,
740 struct cmd_node
*desc_node
;
742 hash_for_each_possible(ring
->cmd_hash
, desc_node
, node
,
743 cmd_header
& CMD_HASH_MASK
) {
744 const struct drm_i915_cmd_descriptor
*desc
= desc_node
->desc
;
745 u32 masked_cmd
= desc
->cmd
.mask
& cmd_header
;
746 u32 masked_value
= desc
->cmd
.value
& desc
->cmd
.mask
;
748 if (masked_cmd
== masked_value
)
756 * Returns a pointer to a descriptor for the command specified by cmd_header.
758 * The caller must supply space for a default descriptor via the default_desc
759 * parameter. If no descriptor for the specified command exists in the ring's
760 * command parser tables, this function fills in default_desc based on the
761 * ring's default length encoding and returns default_desc.
763 static const struct drm_i915_cmd_descriptor
*
764 find_cmd(struct intel_engine_cs
*ring
,
766 struct drm_i915_cmd_descriptor
*default_desc
)
768 const struct drm_i915_cmd_descriptor
*desc
;
771 desc
= find_cmd_in_table(ring
, cmd_header
);
775 mask
= ring
->get_cmd_length_mask(cmd_header
);
779 BUG_ON(!default_desc
);
780 default_desc
->flags
= CMD_DESC_SKIP
;
781 default_desc
->length
.mask
= mask
;
786 static bool valid_reg(const u32
*table
, int count
, u32 addr
)
788 if (table
&& count
!= 0) {
791 for (i
= 0; i
< count
; i
++) {
792 if (table
[i
] == addr
)
800 static u32
*vmap_batch(struct drm_i915_gem_object
*obj
)
804 struct sg_page_iter sg_iter
;
807 pages
= drm_malloc_ab(obj
->base
.size
>> PAGE_SHIFT
, sizeof(*pages
));
809 DRM_DEBUG_DRIVER("Failed to get space for pages\n");
814 for_each_sg_page(obj
->pages
->sgl
, &sg_iter
, obj
->pages
->nents
, 0) {
815 pages
[i
] = sg_page_iter_page(&sg_iter
);
819 addr
= vmap(pages
, i
, 0, PAGE_KERNEL
);
821 DRM_DEBUG_DRIVER("Failed to vmap pages\n");
827 drm_free_large(pages
);
832 * i915_needs_cmd_parser() - should a given ring use software command parsing?
833 * @ring: the ring in question
835 * Only certain platforms require software batch buffer command parsing, and
836 * only when enabled via module paramter.
838 * Return: true if the ring requires software command parsing
840 bool i915_needs_cmd_parser(struct intel_engine_cs
*ring
)
842 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
844 if (!ring
->needs_cmd_parser
)
848 * XXX: VLV is Gen7 and therefore has cmd_tables, but has PPGTT
849 * disabled. That will cause all of the parser's PPGTT checks to
850 * fail. For now, disable parsing when PPGTT is off.
852 if (!dev_priv
->mm
.aliasing_ppgtt
)
855 return (i915
.enable_cmd_parser
== 1);
858 static bool check_cmd(const struct intel_engine_cs
*ring
,
859 const struct drm_i915_cmd_descriptor
*desc
,
861 const bool is_master
,
864 if (desc
->flags
& CMD_DESC_REJECT
) {
865 DRM_DEBUG_DRIVER("CMD: Rejected command: 0x%08X\n", *cmd
);
869 if ((desc
->flags
& CMD_DESC_MASTER
) && !is_master
) {
870 DRM_DEBUG_DRIVER("CMD: Rejected master-only command: 0x%08X\n",
875 if (desc
->flags
& CMD_DESC_REGISTER
) {
876 u32 reg_addr
= cmd
[desc
->reg
.offset
] & desc
->reg
.mask
;
879 * OACONTROL requires some special handling for writes. We
880 * want to make sure that any batch which enables OA also
881 * disables it before the end of the batch. The goal is to
882 * prevent one process from snooping on the perf data from
883 * another process. To do that, we need to check the value
884 * that will be written to the register. Hence, limit
885 * OACONTROL writes to only MI_LOAD_REGISTER_IMM commands.
887 if (reg_addr
== OACONTROL
) {
888 if (desc
->cmd
.value
== MI_LOAD_REGISTER_MEM
)
891 if (desc
->cmd
.value
== MI_LOAD_REGISTER_IMM(1))
892 *oacontrol_set
= (cmd
[2] != 0);
895 if (!valid_reg(ring
->reg_table
,
896 ring
->reg_count
, reg_addr
)) {
898 !valid_reg(ring
->master_reg_table
,
899 ring
->master_reg_count
,
901 DRM_DEBUG_DRIVER("CMD: Rejected register 0x%08X in command: 0x%08X (ring=%d)\n",
910 if (desc
->flags
& CMD_DESC_BITMASK
) {
913 for (i
= 0; i
< MAX_CMD_DESC_BITMASKS
; i
++) {
916 if (desc
->bits
[i
].mask
== 0)
919 if (desc
->bits
[i
].condition_mask
!= 0) {
921 desc
->bits
[i
].condition_offset
;
922 u32 condition
= cmd
[offset
] &
923 desc
->bits
[i
].condition_mask
;
929 dword
= cmd
[desc
->bits
[i
].offset
] &
932 if (dword
!= desc
->bits
[i
].expected
) {
933 DRM_DEBUG_DRIVER("CMD: Rejected command 0x%08X for bitmask 0x%08X (exp=0x%08X act=0x%08X) (ring=%d)\n",
936 desc
->bits
[i
].expected
,
946 #define LENGTH_BIAS 2
949 * i915_parse_cmds() - parse a submitted batch buffer for privilege violations
950 * @ring: the ring on which the batch is to execute
951 * @batch_obj: the batch buffer in question
952 * @batch_start_offset: byte offset in the batch at which execution starts
953 * @is_master: is the submitting process the drm master?
955 * Parses the specified batch buffer looking for privilege violations as
956 * described in the overview.
958 * Return: non-zero if the parser finds violations or otherwise fails
960 int i915_parse_cmds(struct intel_engine_cs
*ring
,
961 struct drm_i915_gem_object
*batch_obj
,
962 u32 batch_start_offset
,
966 u32
*cmd
, *batch_base
, *batch_end
;
967 struct drm_i915_cmd_descriptor default_desc
= { 0 };
968 int needs_clflush
= 0;
969 bool oacontrol_set
= false; /* OACONTROL tracking. See check_cmd() */
971 ret
= i915_gem_obj_prepare_shmem_read(batch_obj
, &needs_clflush
);
973 DRM_DEBUG_DRIVER("CMD: failed to prep read\n");
977 batch_base
= vmap_batch(batch_obj
);
979 DRM_DEBUG_DRIVER("CMD: Failed to vmap batch\n");
980 i915_gem_object_unpin_pages(batch_obj
);
985 drm_clflush_virt_range((char *)batch_base
, batch_obj
->base
.size
);
987 cmd
= batch_base
+ (batch_start_offset
/ sizeof(*cmd
));
988 batch_end
= cmd
+ (batch_obj
->base
.size
/ sizeof(*batch_end
));
990 while (cmd
< batch_end
) {
991 const struct drm_i915_cmd_descriptor
*desc
;
994 if (*cmd
== MI_BATCH_BUFFER_END
)
997 desc
= find_cmd(ring
, *cmd
, &default_desc
);
999 DRM_DEBUG_DRIVER("CMD: Unrecognized command: 0x%08X\n",
1005 if (desc
->flags
& CMD_DESC_FIXED
)
1006 length
= desc
->length
.fixed
;
1008 length
= ((*cmd
& desc
->length
.mask
) + LENGTH_BIAS
);
1010 if ((batch_end
- cmd
) < length
) {
1011 DRM_DEBUG_DRIVER("CMD: Command length exceeds batch length: 0x%08X length=%u batchlen=%td\n",
1019 if (!check_cmd(ring
, desc
, cmd
, is_master
, &oacontrol_set
)) {
1027 if (oacontrol_set
) {
1028 DRM_DEBUG_DRIVER("CMD: batch set OACONTROL but did not clear it\n");
1032 if (cmd
>= batch_end
) {
1033 DRM_DEBUG_DRIVER("CMD: Got to the end of the buffer w/o a BBE cmd!\n");
1039 i915_gem_object_unpin_pages(batch_obj
);
1045 * i915_cmd_parser_get_version() - get the cmd parser version number
1047 * The cmd parser maintains a simple increasing integer version number suitable
1048 * for passing to userspace clients to determine what operations are permitted.
1050 * Return: the current version number of the cmd parser
1052 int i915_cmd_parser_get_version(void)
1055 * Command parser version history
1057 * 1. Initial version. Checks batches and reports violations, but leaves
1058 * hardware parsing enabled (so does not allow new use cases).