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(GPGPU_THREADS_DISPATCHED),
409 REG64(HS_INVOCATION_COUNT),
410 REG64(DS_INVOCATION_COUNT),
411 REG64(IA_VERTICES_COUNT),
412 REG64(IA_PRIMITIVES_COUNT),
413 REG64(VS_INVOCATION_COUNT),
414 REG64(GS_INVOCATION_COUNT),
415 REG64(GS_PRIMITIVES_COUNT),
416 REG64(CL_INVOCATION_COUNT),
417 REG64(CL_PRIMITIVES_COUNT),
418 REG64(PS_INVOCATION_COUNT),
419 REG64(PS_DEPTH_COUNT),
420 OACONTROL, /* Only allowed for LRI and SRM. See below. */
421 REG64(MI_PREDICATE_SRC0),
422 REG64(MI_PREDICATE_SRC1),
423 GEN7_3DPRIM_END_OFFSET,
424 GEN7_3DPRIM_START_VERTEX,
425 GEN7_3DPRIM_VERTEX_COUNT,
426 GEN7_3DPRIM_INSTANCE_COUNT,
427 GEN7_3DPRIM_START_INSTANCE,
428 GEN7_3DPRIM_BASE_VERTEX,
429 REG64(GEN7_SO_NUM_PRIMS_WRITTEN(0)),
430 REG64(GEN7_SO_NUM_PRIMS_WRITTEN(1)),
431 REG64(GEN7_SO_NUM_PRIMS_WRITTEN(2)),
432 REG64(GEN7_SO_NUM_PRIMS_WRITTEN(3)),
433 REG64(GEN7_SO_PRIM_STORAGE_NEEDED(0)),
434 REG64(GEN7_SO_PRIM_STORAGE_NEEDED(1)),
435 REG64(GEN7_SO_PRIM_STORAGE_NEEDED(2)),
436 REG64(GEN7_SO_PRIM_STORAGE_NEEDED(3)),
437 GEN7_SO_WRITE_OFFSET(0),
438 GEN7_SO_WRITE_OFFSET(1),
439 GEN7_SO_WRITE_OFFSET(2),
440 GEN7_SO_WRITE_OFFSET(3),
446 static const u32 gen7_blt_regs[] = {
450 static const u32 ivb_master_regs[] = {
453 GEN7_PIPE_DE_LOAD_SL(PIPE_A),
454 GEN7_PIPE_DE_LOAD_SL(PIPE_B),
455 GEN7_PIPE_DE_LOAD_SL(PIPE_C),
458 static const u32 hsw_master_regs[] = {
465 static u32 gen7_render_get_cmd_length_mask(u32 cmd_header)
467 u32 client = (cmd_header & INSTR_CLIENT_MASK) >> INSTR_CLIENT_SHIFT;
469 (cmd_header & INSTR_SUBCLIENT_MASK) >> INSTR_SUBCLIENT_SHIFT;
471 if (client == INSTR_MI_CLIENT)
473 else if (client == INSTR_RC_CLIENT) {
474 if (subclient == INSTR_MEDIA_SUBCLIENT)
480 DRM_DEBUG_DRIVER("CMD: Abnormal rcs cmd length! 0x%08X\n", cmd_header);
484 static u32 gen7_bsd_get_cmd_length_mask(u32 cmd_header)
486 u32 client = (cmd_header & INSTR_CLIENT_MASK) >> INSTR_CLIENT_SHIFT;
488 (cmd_header & INSTR_SUBCLIENT_MASK) >> INSTR_SUBCLIENT_SHIFT;
489 u32 op = (cmd_header & INSTR_26_TO_24_MASK) >> INSTR_26_TO_24_SHIFT;
491 if (client == INSTR_MI_CLIENT)
493 else if (client == INSTR_RC_CLIENT) {
494 if (subclient == INSTR_MEDIA_SUBCLIENT) {
503 DRM_DEBUG_DRIVER("CMD: Abnormal bsd cmd length! 0x%08X\n", cmd_header);
507 static u32 gen7_blt_get_cmd_length_mask(u32 cmd_header)
509 u32 client = (cmd_header & INSTR_CLIENT_MASK) >> INSTR_CLIENT_SHIFT;
511 if (client == INSTR_MI_CLIENT)
513 else if (client == INSTR_BC_CLIENT)
516 DRM_DEBUG_DRIVER("CMD: Abnormal blt cmd length! 0x%08X\n", cmd_header);
520 static bool validate_cmds_sorted(struct intel_engine_cs *ring,
521 const struct drm_i915_cmd_table *cmd_tables,
527 if (!cmd_tables || cmd_table_count == 0)
530 for (i = 0; i < cmd_table_count; i++) {
531 const struct drm_i915_cmd_table *table = &cmd_tables[i];
535 for (j = 0; j < table->count; j++) {
536 const struct drm_i915_cmd_descriptor *desc =
538 u32 curr = desc->cmd.value & desc->cmd.mask;
540 if (curr < previous) {
541 DRM_ERROR("CMD: table not sorted ring=%d table=%d entry=%d cmd=0x%08X prev=0x%08X\n",
542 ring->id, i, j, curr, previous);
553 static bool check_sorted(int ring_id, const u32 *reg_table, int reg_count)
559 for (i = 0; i < reg_count; i++) {
560 u32 curr = reg_table[i];
562 if (curr < previous) {
563 DRM_ERROR("CMD: table not sorted ring=%d entry=%d reg=0x%08X prev=0x%08X\n",
564 ring_id, i, curr, previous);
574 static bool validate_regs_sorted(struct intel_engine_cs *ring)
576 return check_sorted(ring->id, ring->reg_table, ring->reg_count) &&
577 check_sorted(ring->id, ring->master_reg_table,
578 ring->master_reg_count);
582 const struct drm_i915_cmd_descriptor *desc;
583 struct hlist_node node;
587 * Different command ranges have different numbers of bits for the opcode. For
588 * example, MI commands use bits 31:23 while 3D commands use bits 31:16. The
589 * problem is that, for example, MI commands use bits 22:16 for other fields
590 * such as GGTT vs PPGTT bits. If we include those bits in the mask then when
591 * we mask a command from a batch it could hash to the wrong bucket due to
592 * non-opcode bits being set. But if we don't include those bits, some 3D
593 * commands may hash to the same bucket due to not including opcode bits that
594 * make the command unique. For now, we will risk hashing to the same bucket.
596 * If we attempt to generate a perfect hash, we should be able to look at bits
597 * 31:29 of a command from a batch buffer and use the full mask for that
598 * client. The existing INSTR_CLIENT_MASK/SHIFT defines can be used for this.
600 #define CMD_HASH_MASK STD_MI_OPCODE_MASK
602 static int init_hash_table(struct intel_engine_cs *ring,
603 const struct drm_i915_cmd_table *cmd_tables,
609 hash_init(ring->cmd_hash);
611 for (i = 0; i < cmd_table_count; i++) {
612 const struct drm_i915_cmd_table *table = &cmd_tables[i];
614 for (j = 0; j < table->count; j++) {
615 const struct drm_i915_cmd_descriptor *desc =
617 struct cmd_node *desc_node =
618 kmalloc(sizeof(*desc_node), M_DRM, M_WAITOK);
623 desc_node->desc = desc;
624 hash_add(ring->cmd_hash, &desc_node->node,
625 desc->cmd.value & CMD_HASH_MASK);
633 static void fini_hash_table(struct intel_engine_cs *ring)
636 struct hlist_node *tmp;
637 struct cmd_node *desc_node;
640 hash_for_each_safe(ring->cmd_hash, i, tmp, desc_node, node) {
641 hash_del(&desc_node->node);
648 * i915_cmd_parser_init_ring() - set cmd parser related fields for a ringbuffer
649 * @ring: the ringbuffer to initialize
651 * Optionally initializes fields related to batch buffer command parsing in the
652 * struct intel_engine_cs based on whether the platform requires software
655 * Return: non-zero if initialization fails
657 int i915_cmd_parser_init_ring(struct intel_engine_cs *ring)
659 const struct drm_i915_cmd_table *cmd_tables;
663 if (!IS_GEN7(ring->dev))
668 if (IS_HASWELL(ring->dev)) {
669 cmd_tables = hsw_render_ring_cmds;
671 ARRAY_SIZE(hsw_render_ring_cmds);
673 cmd_tables = gen7_render_cmds;
674 cmd_table_count = ARRAY_SIZE(gen7_render_cmds);
677 ring->reg_table = gen7_render_regs;
678 ring->reg_count = ARRAY_SIZE(gen7_render_regs);
680 if (IS_HASWELL(ring->dev)) {
681 ring->master_reg_table = hsw_master_regs;
682 ring->master_reg_count = ARRAY_SIZE(hsw_master_regs);
684 ring->master_reg_table = ivb_master_regs;
685 ring->master_reg_count = ARRAY_SIZE(ivb_master_regs);
688 ring->get_cmd_length_mask = gen7_render_get_cmd_length_mask;
691 cmd_tables = gen7_video_cmds;
692 cmd_table_count = ARRAY_SIZE(gen7_video_cmds);
693 ring->get_cmd_length_mask = gen7_bsd_get_cmd_length_mask;
696 if (IS_HASWELL(ring->dev)) {
697 cmd_tables = hsw_blt_ring_cmds;
698 cmd_table_count = ARRAY_SIZE(hsw_blt_ring_cmds);
700 cmd_tables = gen7_blt_cmds;
701 cmd_table_count = ARRAY_SIZE(gen7_blt_cmds);
704 ring->reg_table = gen7_blt_regs;
705 ring->reg_count = ARRAY_SIZE(gen7_blt_regs);
707 if (IS_HASWELL(ring->dev)) {
708 ring->master_reg_table = hsw_master_regs;
709 ring->master_reg_count = ARRAY_SIZE(hsw_master_regs);
711 ring->master_reg_table = ivb_master_regs;
712 ring->master_reg_count = ARRAY_SIZE(ivb_master_regs);
715 ring->get_cmd_length_mask = gen7_blt_get_cmd_length_mask;
718 cmd_tables = hsw_vebox_cmds;
719 cmd_table_count = ARRAY_SIZE(hsw_vebox_cmds);
720 /* VECS can use the same length_mask function as VCS */
721 ring->get_cmd_length_mask = gen7_bsd_get_cmd_length_mask;
724 DRM_ERROR("CMD: cmd_parser_init with unknown ring: %d\n",
729 BUG_ON(!validate_cmds_sorted(ring, cmd_tables, cmd_table_count));
730 BUG_ON(!validate_regs_sorted(ring));
733 WARN_ON(!hash_empty(ring->cmd_hash));
736 ret = init_hash_table(ring, cmd_tables, cmd_table_count);
738 DRM_ERROR("CMD: cmd_parser_init failed!\n");
739 fini_hash_table(ring);
743 ring->needs_cmd_parser = true;
749 * i915_cmd_parser_fini_ring() - clean up cmd parser related fields
750 * @ring: the ringbuffer to clean up
752 * Releases any resources related to command parsing that may have been
753 * initialized for the specified ring.
755 void i915_cmd_parser_fini_ring(struct intel_engine_cs *ring)
757 if (!ring->needs_cmd_parser)
760 fini_hash_table(ring);
763 static const struct drm_i915_cmd_descriptor*
764 find_cmd_in_table(struct intel_engine_cs *ring,
768 struct cmd_node *desc_node;
770 hash_for_each_possible(ring->cmd_hash, desc_node, node,
771 cmd_header & CMD_HASH_MASK) {
772 const struct drm_i915_cmd_descriptor *desc = desc_node->desc;
773 u32 masked_cmd = desc->cmd.mask & cmd_header;
774 u32 masked_value = desc->cmd.value & desc->cmd.mask;
776 if (masked_cmd == masked_value)
785 * Returns a pointer to a descriptor for the command specified by cmd_header.
787 * The caller must supply space for a default descriptor via the default_desc
788 * parameter. If no descriptor for the specified command exists in the ring's
789 * command parser tables, this function fills in default_desc based on the
790 * ring's default length encoding and returns default_desc.
792 static const struct drm_i915_cmd_descriptor*
793 find_cmd(struct intel_engine_cs *ring,
795 struct drm_i915_cmd_descriptor *default_desc)
797 const struct drm_i915_cmd_descriptor *desc;
800 desc = find_cmd_in_table(ring, cmd_header);
804 mask = ring->get_cmd_length_mask(cmd_header);
808 BUG_ON(!default_desc);
809 default_desc->flags = CMD_DESC_SKIP;
810 default_desc->length.mask = mask;
815 static bool valid_reg(const u32 *table, int count, u32 addr)
817 if (table && count != 0) {
820 for (i = 0; i < count; i++) {
821 if (table[i] == addr)
829 static u32 *vmap_batch(struct drm_i915_gem_object *obj,
830 unsigned start, unsigned len)
834 int first_page = start >> PAGE_SHIFT;
835 int last_page = (len + start + 4095) >> PAGE_SHIFT;
836 int npages = last_page - first_page;
837 struct vm_page **pages;
839 pages = drm_malloc_ab(npages, sizeof(*pages));
841 DRM_DEBUG_DRIVER("Failed to get space for pages\n");
847 pages[i] = obj->pages[first_page + i];
852 addr = vmap(pages, i, 0, PAGE_KERNEL);
854 DRM_DEBUG_DRIVER("Failed to vmap pages\n");
861 drm_free_large(pages);
865 /* Returns a vmap'd pointer to dest_obj, which the caller must unmap */
866 static u32 *copy_batch(struct drm_i915_gem_object *dest_obj,
867 struct drm_i915_gem_object *src_obj,
868 u32 batch_start_offset,
871 int needs_clflush = 0;
872 char *src_base, *src;
876 if (batch_len > dest_obj->base.size ||
877 batch_len + batch_start_offset > src_obj->base.size)
878 return ERR_PTR(-E2BIG);
880 ret = i915_gem_obj_prepare_shmem_read(src_obj, &needs_clflush);
882 DRM_DEBUG_DRIVER("CMD: failed to prepare shadow batch\n");
886 src_base = (char *)vmap_batch(src_obj, batch_start_offset, batch_len);
888 DRM_DEBUG_DRIVER("CMD: Failed to vmap batch\n");
893 ret = i915_gem_object_get_pages(dest_obj);
895 DRM_DEBUG_DRIVER("CMD: Failed to get pages for shadow batch\n");
898 i915_gem_object_pin_pages(dest_obj);
900 ret = i915_gem_object_set_to_cpu_domain(dest_obj, true);
902 DRM_DEBUG_DRIVER("CMD: Failed to set shadow batch to CPU\n");
906 dst = vmap_batch(dest_obj, 0, batch_len);
908 DRM_DEBUG_DRIVER("CMD: Failed to vmap shadow batch\n");
909 i915_gem_object_unpin_pages(dest_obj);
914 src = src_base + offset_in_page(batch_start_offset);
916 drm_clflush_virt_range(src, batch_len);
918 memcpy(dst, src, batch_len);
923 i915_gem_object_unpin_pages(src_obj);
925 return ret ? ERR_PTR(ret) : dst;
929 * i915_needs_cmd_parser() - should a given ring use software command parsing?
930 * @ring: the ring in question
932 * Only certain platforms require software batch buffer command parsing, and
933 * only when enabled via module parameter.
935 * Return: true if the ring requires software command parsing
937 bool i915_needs_cmd_parser(struct intel_engine_cs *ring)
939 if (!ring->needs_cmd_parser)
942 if (!USES_PPGTT(ring->dev))
945 return (i915.enable_cmd_parser == 1);
948 static bool check_cmd(const struct intel_engine_cs *ring,
949 const struct drm_i915_cmd_descriptor *desc,
951 const bool is_master,
954 if (desc->flags & CMD_DESC_REJECT) {
955 DRM_DEBUG_DRIVER("CMD: Rejected command: 0x%08X\n", *cmd);
959 if ((desc->flags & CMD_DESC_MASTER) && !is_master) {
960 DRM_DEBUG_DRIVER("CMD: Rejected master-only command: 0x%08X\n",
965 if (desc->flags & CMD_DESC_REGISTER) {
966 u32 reg_addr = cmd[desc->reg.offset] & desc->reg.mask;
969 * OACONTROL requires some special handling for writes. We
970 * want to make sure that any batch which enables OA also
971 * disables it before the end of the batch. The goal is to
972 * prevent one process from snooping on the perf data from
973 * another process. To do that, we need to check the value
974 * that will be written to the register. Hence, limit
975 * OACONTROL writes to only MI_LOAD_REGISTER_IMM commands.
977 if (reg_addr == OACONTROL) {
978 if (desc->cmd.value == MI_LOAD_REGISTER_MEM) {
979 DRM_DEBUG_DRIVER("CMD: Rejected LRM to OACONTROL\n");
983 if (desc->cmd.value == MI_LOAD_REGISTER_IMM(1))
984 *oacontrol_set = (cmd[2] != 0);
987 if (!valid_reg(ring->reg_table,
988 ring->reg_count, reg_addr)) {
990 !valid_reg(ring->master_reg_table,
991 ring->master_reg_count,
993 DRM_DEBUG_DRIVER("CMD: Rejected register 0x%08X in command: 0x%08X (ring=%d)\n",
1002 if (desc->flags & CMD_DESC_BITMASK) {
1005 for (i = 0; i < MAX_CMD_DESC_BITMASKS; i++) {
1008 if (desc->bits[i].mask == 0)
1011 if (desc->bits[i].condition_mask != 0) {
1013 desc->bits[i].condition_offset;
1014 u32 condition = cmd[offset] &
1015 desc->bits[i].condition_mask;
1021 dword = cmd[desc->bits[i].offset] &
1024 if (dword != desc->bits[i].expected) {
1025 DRM_DEBUG_DRIVER("CMD: Rejected command 0x%08X for bitmask 0x%08X (exp=0x%08X act=0x%08X) (ring=%d)\n",
1028 desc->bits[i].expected,
1038 #define LENGTH_BIAS 2
1041 * i915_parse_cmds() - parse a submitted batch buffer for privilege violations
1042 * @ring: the ring on which the batch is to execute
1043 * @batch_obj: the batch buffer in question
1044 * @shadow_batch_obj: copy of the batch buffer in question
1045 * @batch_start_offset: byte offset in the batch at which execution starts
1046 * @batch_len: length of the commands in batch_obj
1047 * @is_master: is the submitting process the drm master?
1049 * Parses the specified batch buffer looking for privilege violations as
1050 * described in the overview.
1052 * Return: non-zero if the parser finds violations or otherwise fails; -EACCES
1053 * if the batch appears legal but should use hardware parsing
1055 int i915_parse_cmds(struct intel_engine_cs *ring,
1056 struct drm_i915_gem_object *batch_obj,
1057 struct drm_i915_gem_object *shadow_batch_obj,
1058 u32 batch_start_offset,
1062 u32 *cmd, *batch_base, *batch_end;
1063 struct drm_i915_cmd_descriptor default_desc = { 0 };
1064 bool oacontrol_set = false; /* OACONTROL tracking. See check_cmd() */
1067 batch_base = copy_batch(shadow_batch_obj, batch_obj,
1068 batch_start_offset, batch_len);
1069 if (IS_ERR(batch_base)) {
1070 DRM_DEBUG_DRIVER("CMD: Failed to copy batch\n");
1071 return PTR_ERR(batch_base);
1075 * We use the batch length as size because the shadow object is as
1076 * large or larger and copy_batch() will write MI_NOPs to the extra
1077 * space. Parsing should be faster in some cases this way.
1079 batch_end = batch_base + (batch_len / sizeof(*batch_end));
1082 while (cmd < batch_end) {
1083 const struct drm_i915_cmd_descriptor *desc;
1086 if (*cmd == MI_BATCH_BUFFER_END)
1089 desc = find_cmd(ring, *cmd, &default_desc);
1091 DRM_DEBUG_DRIVER("CMD: Unrecognized command: 0x%08X\n",
1098 * If the batch buffer contains a chained batch, return an
1099 * error that tells the caller to abort and dispatch the
1100 * workload as a non-secure batch.
1102 if (desc->cmd.value == MI_BATCH_BUFFER_START) {
1107 if (desc->flags & CMD_DESC_FIXED)
1108 length = desc->length.fixed;
1110 length = ((*cmd & desc->length.mask) + LENGTH_BIAS);
1112 if ((batch_end - cmd) < length) {
1113 DRM_DEBUG_DRIVER("CMD: Command length exceeds batch length: 0x%08X length=%u batchlen=%td\n",
1121 if (!check_cmd(ring, desc, cmd, is_master, &oacontrol_set)) {
1129 if (oacontrol_set) {
1130 DRM_DEBUG_DRIVER("CMD: batch set OACONTROL but did not clear it\n");
1134 if (cmd >= batch_end) {
1135 DRM_DEBUG_DRIVER("CMD: Got to the end of the buffer w/o a BBE cmd!\n");
1140 i915_gem_object_unpin_pages(shadow_batch_obj);
1146 * i915_cmd_parser_get_version() - get the cmd parser version number
1148 * The cmd parser maintains a simple increasing integer version number suitable
1149 * for passing to userspace clients to determine what operations are permitted.
1151 * Return: the current version number of the cmd parser
1153 int i915_cmd_parser_get_version(void)
1156 * Command parser version history
1158 * 1. Initial version. Checks batches and reports violations, but leaves
1159 * hardware parsing enabled (so does not allow new use cases).
1160 * 2. Allow access to the MI_PREDICATE_SRC0 and
1161 * MI_PREDICATE_SRC1 registers.
1162 * 3. Allow access to the GPGPU_THREADS_DISPATCHED register.