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>
28 #include "gt/intel_engine.h"
31 #include "i915_memcpy.h"
34 * DOC: batch buffer command parser
37 * Certain OpenGL features (e.g. transform feedback, performance monitoring)
38 * require userspace code to submit batches containing commands such as
39 * MI_LOAD_REGISTER_IMM to access various registers. Unfortunately, some
40 * generations of the hardware will noop these commands in "unsecure" batches
41 * (which includes all userspace batches submitted via i915) even though the
42 * commands may be safe and represent the intended programming model of the
45 * The software command parser is similar in operation to the command parsing
46 * done in hardware for unsecure batches. However, the software parser allows
47 * some operations that would be noop'd by hardware, if the parser determines
48 * the operation is safe, and submits the batch as "secure" to prevent hardware
52 * At a high level, the hardware (and software) checks attempt to prevent
53 * granting userspace undue privileges. There are three categories of privilege.
55 * First, commands which are explicitly defined as privileged or which should
56 * only be used by the kernel driver. The parser generally rejects such
57 * commands, though it may allow some from the drm master process.
59 * Second, commands which access registers. To support correct/enhanced
60 * userspace functionality, particularly certain OpenGL extensions, the parser
61 * provides a whitelist of registers which userspace may safely access (for both
62 * normal and drm master processes).
64 * Third, commands which access privileged memory (i.e. GGTT, HWS page, etc).
65 * The parser always rejects such commands.
67 * The majority of the problematic commands fall in the MI_* range, with only a
68 * few specific commands on each engine (e.g. PIPE_CONTROL and MI_FLUSH_DW).
71 * Each engine maintains tables of commands and registers which the parser
72 * uses in scanning batch buffers submitted to that engine.
74 * Since the set of commands that the parser must check for is significantly
75 * smaller than the number of commands supported, the parser tables contain only
76 * those commands required by the parser. This generally works because command
77 * opcode ranges have standard command length encodings. So for commands that
78 * the parser does not need to check, it can easily skip them. This is
79 * implemented via a per-engine length decoding vfunc.
81 * Unfortunately, there are a number of commands that do not follow the standard
82 * length encoding for their opcode range, primarily amongst the MI_* commands.
83 * To handle this, the parser provides a way to define explicit "skip" entries
84 * in the per-engine command tables.
86 * Other command table entries map fairly directly to high level categories
87 * mentioned above: rejected, master-only, register whitelist. The parser
88 * implements a number of checks, including the privileged memory checks, via a
89 * general bitmasking mechanism.
93 * A command that requires special handling by the command parser.
95 struct drm_i915_cmd_descriptor {
97 * Flags describing how the command parser processes the command.
99 * CMD_DESC_FIXED: The command has a fixed length if this is set,
100 * a length mask if not set
101 * CMD_DESC_SKIP: The command is allowed but does not follow the
102 * standard length encoding for the opcode range in
104 * CMD_DESC_REJECT: The command is never allowed
105 * CMD_DESC_REGISTER: The command should be checked against the
106 * register whitelist for the appropriate ring
107 * CMD_DESC_MASTER: The command is allowed if the submitting process
111 #define CMD_DESC_FIXED (1<<0)
112 #define CMD_DESC_SKIP (1<<1)
113 #define CMD_DESC_REJECT (1<<2)
114 #define CMD_DESC_REGISTER (1<<3)
115 #define CMD_DESC_BITMASK (1<<4)
116 #define CMD_DESC_MASTER (1<<5)
119 * The command's unique identification bits and the bitmask to get them.
120 * This isn't strictly the opcode field as defined in the spec and may
121 * also include type, subtype, and/or subop fields.
129 * The command's length. The command is either fixed length (i.e. does
130 * not include a length field) or has a length field mask. The flag
131 * CMD_DESC_FIXED indicates a fixed length. Otherwise, the command has
132 * a length mask. All command entries in a command table must include
133 * length information.
141 * Describes where to find a register address in the command to check
142 * against the ring's register whitelist. Only valid if flags has the
143 * CMD_DESC_REGISTER bit set.
145 * A non-zero step value implies that the command may access multiple
146 * registers in sequence (e.g. LRI), in that case step gives the
147 * distance in dwords between individual offset fields.
155 #define MAX_CMD_DESC_BITMASKS 3
157 * Describes command checks where a particular dword is masked and
158 * compared against an expected value. If the command does not match
159 * the expected value, the parser rejects it. Only valid if flags has
160 * the CMD_DESC_BITMASK bit set. Only entries where mask is non-zero
163 * If the check specifies a non-zero condition_mask then the parser
164 * only performs the check when the bits specified by condition_mask
171 u32 condition_offset;
173 } bits[MAX_CMD_DESC_BITMASKS];
177 * A table of commands requiring special handling by the command parser.
179 * Each engine has an array of tables. Each table consists of an array of
180 * command descriptors, which must be sorted with command opcodes in
183 struct drm_i915_cmd_table {
184 const struct drm_i915_cmd_descriptor *table;
188 #define STD_MI_OPCODE_SHIFT (32 - 9)
189 #define STD_3D_OPCODE_SHIFT (32 - 16)
190 #define STD_2D_OPCODE_SHIFT (32 - 10)
191 #define STD_MFX_OPCODE_SHIFT (32 - 16)
192 #define MIN_OPCODE_SHIFT 16
194 #define CMD(op, opm, f, lm, fl, ...) \
196 .flags = (fl) | ((f) ? CMD_DESC_FIXED : 0), \
197 .cmd = { (op), ~0u << (opm) }, \
198 .length = { (lm) }, \
202 /* Convenience macros to compress the tables */
203 #define SMI STD_MI_OPCODE_SHIFT
204 #define S3D STD_3D_OPCODE_SHIFT
205 #define S2D STD_2D_OPCODE_SHIFT
206 #define SMFX STD_MFX_OPCODE_SHIFT
208 #define S CMD_DESC_SKIP
209 #define R CMD_DESC_REJECT
210 #define W CMD_DESC_REGISTER
211 #define B CMD_DESC_BITMASK
212 #define M CMD_DESC_MASTER
214 /* Command Mask Fixed Len Action
215 ---------------------------------------------------------- */
216 static const struct drm_i915_cmd_descriptor common_cmds[] = {
217 CMD( MI_NOOP, SMI, F, 1, S ),
218 CMD( MI_USER_INTERRUPT, SMI, F, 1, R ),
219 CMD( MI_WAIT_FOR_EVENT, SMI, F, 1, M ),
220 CMD( MI_ARB_CHECK, SMI, F, 1, S ),
221 CMD( MI_REPORT_HEAD, SMI, F, 1, S ),
222 CMD( MI_SUSPEND_FLUSH, SMI, F, 1, S ),
223 CMD( MI_SEMAPHORE_MBOX, SMI, !F, 0xFF, R ),
224 CMD( MI_STORE_DWORD_INDEX, SMI, !F, 0xFF, R ),
225 CMD( MI_LOAD_REGISTER_IMM(1), SMI, !F, 0xFF, W,
226 .reg = { .offset = 1, .mask = 0x007FFFFC, .step = 2 } ),
227 CMD( MI_STORE_REGISTER_MEM, SMI, F, 3, W | B,
228 .reg = { .offset = 1, .mask = 0x007FFFFC },
231 .mask = MI_GLOBAL_GTT,
234 CMD( MI_LOAD_REGISTER_MEM, SMI, F, 3, W | B,
235 .reg = { .offset = 1, .mask = 0x007FFFFC },
238 .mask = MI_GLOBAL_GTT,
242 * MI_BATCH_BUFFER_START requires some special handling. It's not
243 * really a 'skip' action but it doesn't seem like it's worth adding
244 * a new action. See i915_parse_cmds().
246 CMD( MI_BATCH_BUFFER_START, SMI, !F, 0xFF, S ),
249 static const struct drm_i915_cmd_descriptor render_cmds[] = {
250 CMD( MI_FLUSH, SMI, F, 1, S ),
251 CMD( MI_ARB_ON_OFF, SMI, F, 1, R ),
252 CMD( MI_PREDICATE, SMI, F, 1, S ),
253 CMD( MI_TOPOLOGY_FILTER, SMI, F, 1, S ),
254 CMD( MI_SET_APPID, SMI, F, 1, S ),
255 CMD( MI_DISPLAY_FLIP, SMI, !F, 0xFF, R ),
256 CMD( MI_SET_CONTEXT, SMI, !F, 0xFF, R ),
257 CMD( MI_URB_CLEAR, SMI, !F, 0xFF, S ),
258 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0x3F, B,
261 .mask = MI_GLOBAL_GTT,
264 CMD( MI_UPDATE_GTT, SMI, !F, 0xFF, R ),
265 CMD( MI_CLFLUSH, SMI, !F, 0x3FF, B,
268 .mask = MI_GLOBAL_GTT,
271 CMD( MI_REPORT_PERF_COUNT, SMI, !F, 0x3F, B,
274 .mask = MI_REPORT_PERF_COUNT_GGTT,
277 CMD( MI_CONDITIONAL_BATCH_BUFFER_END, SMI, !F, 0xFF, B,
280 .mask = MI_GLOBAL_GTT,
283 CMD( GFX_OP_3DSTATE_VF_STATISTICS, S3D, F, 1, S ),
284 CMD( PIPELINE_SELECT, S3D, F, 1, S ),
285 CMD( MEDIA_VFE_STATE, S3D, !F, 0xFFFF, B,
288 .mask = MEDIA_VFE_STATE_MMIO_ACCESS_MASK,
291 CMD( GPGPU_OBJECT, S3D, !F, 0xFF, S ),
292 CMD( GPGPU_WALKER, S3D, !F, 0xFF, S ),
293 CMD( GFX_OP_3DSTATE_SO_DECL_LIST, S3D, !F, 0x1FF, S ),
294 CMD( GFX_OP_PIPE_CONTROL(5), S3D, !F, 0xFF, B,
297 .mask = (PIPE_CONTROL_MMIO_WRITE | PIPE_CONTROL_NOTIFY),
302 .mask = (PIPE_CONTROL_GLOBAL_GTT_IVB |
303 PIPE_CONTROL_STORE_DATA_INDEX),
305 .condition_offset = 1,
306 .condition_mask = PIPE_CONTROL_POST_SYNC_OP_MASK,
310 static const struct drm_i915_cmd_descriptor hsw_render_cmds[] = {
311 CMD( MI_SET_PREDICATE, SMI, F, 1, S ),
312 CMD( MI_RS_CONTROL, SMI, F, 1, S ),
313 CMD( MI_URB_ATOMIC_ALLOC, SMI, F, 1, S ),
314 CMD( MI_SET_APPID, SMI, F, 1, S ),
315 CMD( MI_RS_CONTEXT, SMI, F, 1, S ),
316 CMD( MI_LOAD_SCAN_LINES_INCL, SMI, !F, 0x3F, M ),
317 CMD( MI_LOAD_SCAN_LINES_EXCL, SMI, !F, 0x3F, R ),
318 CMD( MI_LOAD_REGISTER_REG, SMI, !F, 0xFF, W,
319 .reg = { .offset = 1, .mask = 0x007FFFFC, .step = 1 } ),
320 CMD( MI_RS_STORE_DATA_IMM, SMI, !F, 0xFF, S ),
321 CMD( MI_LOAD_URB_MEM, SMI, !F, 0xFF, S ),
322 CMD( MI_STORE_URB_MEM, SMI, !F, 0xFF, S ),
323 CMD( GFX_OP_3DSTATE_DX9_CONSTANTF_VS, S3D, !F, 0x7FF, S ),
324 CMD( GFX_OP_3DSTATE_DX9_CONSTANTF_PS, S3D, !F, 0x7FF, S ),
326 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_VS, S3D, !F, 0x1FF, S ),
327 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_GS, S3D, !F, 0x1FF, S ),
328 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_HS, S3D, !F, 0x1FF, S ),
329 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_DS, S3D, !F, 0x1FF, S ),
330 CMD( GFX_OP_3DSTATE_BINDING_TABLE_EDIT_PS, S3D, !F, 0x1FF, S ),
333 static const struct drm_i915_cmd_descriptor video_cmds[] = {
334 CMD( MI_ARB_ON_OFF, SMI, F, 1, R ),
335 CMD( MI_SET_APPID, SMI, F, 1, S ),
336 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0xFF, B,
339 .mask = MI_GLOBAL_GTT,
342 CMD( MI_UPDATE_GTT, SMI, !F, 0x3F, R ),
343 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, B,
346 .mask = MI_FLUSH_DW_NOTIFY,
351 .mask = MI_FLUSH_DW_USE_GTT,
353 .condition_offset = 0,
354 .condition_mask = MI_FLUSH_DW_OP_MASK,
358 .mask = MI_FLUSH_DW_STORE_INDEX,
360 .condition_offset = 0,
361 .condition_mask = MI_FLUSH_DW_OP_MASK,
363 CMD( MI_CONDITIONAL_BATCH_BUFFER_END, SMI, !F, 0xFF, B,
366 .mask = MI_GLOBAL_GTT,
370 * MFX_WAIT doesn't fit the way we handle length for most commands.
371 * It has a length field but it uses a non-standard length bias.
372 * It is always 1 dword though, so just treat it as fixed length.
374 CMD( MFX_WAIT, SMFX, F, 1, S ),
377 static const struct drm_i915_cmd_descriptor vecs_cmds[] = {
378 CMD( MI_ARB_ON_OFF, SMI, F, 1, R ),
379 CMD( MI_SET_APPID, SMI, F, 1, S ),
380 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0xFF, B,
383 .mask = MI_GLOBAL_GTT,
386 CMD( MI_UPDATE_GTT, SMI, !F, 0x3F, R ),
387 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, B,
390 .mask = MI_FLUSH_DW_NOTIFY,
395 .mask = MI_FLUSH_DW_USE_GTT,
397 .condition_offset = 0,
398 .condition_mask = MI_FLUSH_DW_OP_MASK,
402 .mask = MI_FLUSH_DW_STORE_INDEX,
404 .condition_offset = 0,
405 .condition_mask = MI_FLUSH_DW_OP_MASK,
407 CMD( MI_CONDITIONAL_BATCH_BUFFER_END, SMI, !F, 0xFF, B,
410 .mask = MI_GLOBAL_GTT,
415 static const struct drm_i915_cmd_descriptor blt_cmds[] = {
416 CMD( MI_DISPLAY_FLIP, SMI, !F, 0xFF, R ),
417 CMD( MI_STORE_DWORD_IMM, SMI, !F, 0x3FF, B,
420 .mask = MI_GLOBAL_GTT,
423 CMD( MI_UPDATE_GTT, SMI, !F, 0x3F, R ),
424 CMD( MI_FLUSH_DW, SMI, !F, 0x3F, B,
427 .mask = MI_FLUSH_DW_NOTIFY,
432 .mask = MI_FLUSH_DW_USE_GTT,
434 .condition_offset = 0,
435 .condition_mask = MI_FLUSH_DW_OP_MASK,
439 .mask = MI_FLUSH_DW_STORE_INDEX,
441 .condition_offset = 0,
442 .condition_mask = MI_FLUSH_DW_OP_MASK,
444 CMD( COLOR_BLT, S2D, !F, 0x3F, S ),
445 CMD( SRC_COPY_BLT, S2D, !F, 0x3F, S ),
448 static const struct drm_i915_cmd_descriptor hsw_blt_cmds[] = {
449 CMD( MI_LOAD_SCAN_LINES_INCL, SMI, !F, 0x3F, M ),
450 CMD( MI_LOAD_SCAN_LINES_EXCL, SMI, !F, 0x3F, R ),
453 static const struct drm_i915_cmd_descriptor noop_desc =
454 CMD(MI_NOOP, SMI, F, 1, S);
468 static const struct drm_i915_cmd_table gen7_render_cmds[] = {
469 { common_cmds, ARRAY_SIZE(common_cmds) },
470 { render_cmds, ARRAY_SIZE(render_cmds) },
473 static const struct drm_i915_cmd_table hsw_render_ring_cmds[] = {
474 { common_cmds, ARRAY_SIZE(common_cmds) },
475 { render_cmds, ARRAY_SIZE(render_cmds) },
476 { hsw_render_cmds, ARRAY_SIZE(hsw_render_cmds) },
479 static const struct drm_i915_cmd_table gen7_video_cmds[] = {
480 { common_cmds, ARRAY_SIZE(common_cmds) },
481 { video_cmds, ARRAY_SIZE(video_cmds) },
484 static const struct drm_i915_cmd_table hsw_vebox_cmds[] = {
485 { common_cmds, ARRAY_SIZE(common_cmds) },
486 { vecs_cmds, ARRAY_SIZE(vecs_cmds) },
489 static const struct drm_i915_cmd_table gen7_blt_cmds[] = {
490 { common_cmds, ARRAY_SIZE(common_cmds) },
491 { blt_cmds, ARRAY_SIZE(blt_cmds) },
494 static const struct drm_i915_cmd_table hsw_blt_ring_cmds[] = {
495 { common_cmds, ARRAY_SIZE(common_cmds) },
496 { blt_cmds, ARRAY_SIZE(blt_cmds) },
497 { hsw_blt_cmds, ARRAY_SIZE(hsw_blt_cmds) },
501 * Register whitelists, sorted by increasing register offset.
505 * An individual whitelist entry granting access to register addr. If
506 * mask is non-zero the argument of immediate register writes will be
507 * AND-ed with mask, and the command will be rejected if the result
508 * doesn't match value.
510 * Registers with non-zero mask are only allowed to be written using
513 struct drm_i915_reg_descriptor {
519 /* Convenience macro for adding 32-bit registers. */
520 #define REG32(_reg, ...) \
521 { .addr = (_reg), __VA_ARGS__ }
524 * Convenience macro for adding 64-bit registers.
526 * Some registers that userspace accesses are 64 bits. The register
527 * access commands only allow 32-bit accesses. Hence, we have to include
528 * entries for both halves of the 64-bit registers.
530 #define REG64(_reg) \
532 { .addr = _reg ## _UDW }
534 #define REG64_IDX(_reg, idx) \
535 { .addr = _reg(idx) }, \
536 { .addr = _reg ## _UDW(idx) }
538 static const struct drm_i915_reg_descriptor gen7_render_regs[] = {
539 REG64(GPGPU_THREADS_DISPATCHED),
540 REG64(HS_INVOCATION_COUNT),
541 REG64(DS_INVOCATION_COUNT),
542 REG64(IA_VERTICES_COUNT),
543 REG64(IA_PRIMITIVES_COUNT),
544 REG64(VS_INVOCATION_COUNT),
545 REG64(GS_INVOCATION_COUNT),
546 REG64(GS_PRIMITIVES_COUNT),
547 REG64(CL_INVOCATION_COUNT),
548 REG64(CL_PRIMITIVES_COUNT),
549 REG64(PS_INVOCATION_COUNT),
550 REG64(PS_DEPTH_COUNT),
551 REG64_IDX(RING_TIMESTAMP, RENDER_RING_BASE),
552 REG64(MI_PREDICATE_SRC0),
553 REG64(MI_PREDICATE_SRC1),
554 REG32(GEN7_3DPRIM_END_OFFSET),
555 REG32(GEN7_3DPRIM_START_VERTEX),
556 REG32(GEN7_3DPRIM_VERTEX_COUNT),
557 REG32(GEN7_3DPRIM_INSTANCE_COUNT),
558 REG32(GEN7_3DPRIM_START_INSTANCE),
559 REG32(GEN7_3DPRIM_BASE_VERTEX),
560 REG32(GEN7_GPGPU_DISPATCHDIMX),
561 REG32(GEN7_GPGPU_DISPATCHDIMY),
562 REG32(GEN7_GPGPU_DISPATCHDIMZ),
563 REG64_IDX(RING_TIMESTAMP, BSD_RING_BASE),
564 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 0),
565 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 1),
566 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 2),
567 REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 3),
568 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 0),
569 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 1),
570 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 2),
571 REG64_IDX(GEN7_SO_PRIM_STORAGE_NEEDED, 3),
572 REG32(GEN7_SO_WRITE_OFFSET(0)),
573 REG32(GEN7_SO_WRITE_OFFSET(1)),
574 REG32(GEN7_SO_WRITE_OFFSET(2)),
575 REG32(GEN7_SO_WRITE_OFFSET(3)),
576 REG32(GEN7_L3SQCREG1),
577 REG32(GEN7_L3CNTLREG2),
578 REG32(GEN7_L3CNTLREG3),
579 REG64_IDX(RING_TIMESTAMP, BLT_RING_BASE),
582 static const struct drm_i915_reg_descriptor hsw_render_regs[] = {
583 REG64_IDX(HSW_CS_GPR, 0),
584 REG64_IDX(HSW_CS_GPR, 1),
585 REG64_IDX(HSW_CS_GPR, 2),
586 REG64_IDX(HSW_CS_GPR, 3),
587 REG64_IDX(HSW_CS_GPR, 4),
588 REG64_IDX(HSW_CS_GPR, 5),
589 REG64_IDX(HSW_CS_GPR, 6),
590 REG64_IDX(HSW_CS_GPR, 7),
591 REG64_IDX(HSW_CS_GPR, 8),
592 REG64_IDX(HSW_CS_GPR, 9),
593 REG64_IDX(HSW_CS_GPR, 10),
594 REG64_IDX(HSW_CS_GPR, 11),
595 REG64_IDX(HSW_CS_GPR, 12),
596 REG64_IDX(HSW_CS_GPR, 13),
597 REG64_IDX(HSW_CS_GPR, 14),
598 REG64_IDX(HSW_CS_GPR, 15),
600 .mask = ~HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE,
602 REG32(HSW_ROW_CHICKEN3,
603 .mask = ~(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE << 16 |
604 HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE),
608 static const struct drm_i915_reg_descriptor gen7_blt_regs[] = {
609 REG64_IDX(RING_TIMESTAMP, RENDER_RING_BASE),
610 REG64_IDX(RING_TIMESTAMP, BSD_RING_BASE),
612 REG64_IDX(RING_TIMESTAMP, BLT_RING_BASE),
615 static const struct drm_i915_reg_descriptor ivb_master_regs[] = {
618 REG32(GEN7_PIPE_DE_LOAD_SL(PIPE_A)),
619 REG32(GEN7_PIPE_DE_LOAD_SL(PIPE_B)),
620 REG32(GEN7_PIPE_DE_LOAD_SL(PIPE_C)),
623 static const struct drm_i915_reg_descriptor hsw_master_regs[] = {
631 struct drm_i915_reg_table {
632 const struct drm_i915_reg_descriptor *regs;
637 static const struct drm_i915_reg_table ivb_render_reg_tables[] = {
638 { gen7_render_regs, ARRAY_SIZE(gen7_render_regs), false },
639 { ivb_master_regs, ARRAY_SIZE(ivb_master_regs), true },
642 static const struct drm_i915_reg_table ivb_blt_reg_tables[] = {
643 { gen7_blt_regs, ARRAY_SIZE(gen7_blt_regs), false },
644 { ivb_master_regs, ARRAY_SIZE(ivb_master_regs), true },
647 static const struct drm_i915_reg_table hsw_render_reg_tables[] = {
648 { gen7_render_regs, ARRAY_SIZE(gen7_render_regs), false },
649 { hsw_render_regs, ARRAY_SIZE(hsw_render_regs), false },
650 { hsw_master_regs, ARRAY_SIZE(hsw_master_regs), true },
653 static const struct drm_i915_reg_table hsw_blt_reg_tables[] = {
654 { gen7_blt_regs, ARRAY_SIZE(gen7_blt_regs), false },
655 { hsw_master_regs, ARRAY_SIZE(hsw_master_regs), true },
658 static u32 gen7_render_get_cmd_length_mask(u32 cmd_header)
660 u32 client = cmd_header >> INSTR_CLIENT_SHIFT;
662 (cmd_header & INSTR_SUBCLIENT_MASK) >> INSTR_SUBCLIENT_SHIFT;
664 if (client == INSTR_MI_CLIENT)
666 else if (client == INSTR_RC_CLIENT) {
667 if (subclient == INSTR_MEDIA_SUBCLIENT)
673 DRM_DEBUG_DRIVER("CMD: Abnormal rcs cmd length! 0x%08X\n", cmd_header);
677 static u32 gen7_bsd_get_cmd_length_mask(u32 cmd_header)
679 u32 client = cmd_header >> INSTR_CLIENT_SHIFT;
681 (cmd_header & INSTR_SUBCLIENT_MASK) >> INSTR_SUBCLIENT_SHIFT;
682 u32 op = (cmd_header & INSTR_26_TO_24_MASK) >> INSTR_26_TO_24_SHIFT;
684 if (client == INSTR_MI_CLIENT)
686 else if (client == INSTR_RC_CLIENT) {
687 if (subclient == INSTR_MEDIA_SUBCLIENT) {
696 DRM_DEBUG_DRIVER("CMD: Abnormal bsd cmd length! 0x%08X\n", cmd_header);
700 static u32 gen7_blt_get_cmd_length_mask(u32 cmd_header)
702 u32 client = cmd_header >> INSTR_CLIENT_SHIFT;
704 if (client == INSTR_MI_CLIENT)
706 else if (client == INSTR_BC_CLIENT)
709 DRM_DEBUG_DRIVER("CMD: Abnormal blt cmd length! 0x%08X\n", cmd_header);
713 static bool validate_cmds_sorted(const struct intel_engine_cs *engine,
714 const struct drm_i915_cmd_table *cmd_tables,
720 if (!cmd_tables || cmd_table_count == 0)
723 for (i = 0; i < cmd_table_count; i++) {
724 const struct drm_i915_cmd_table *table = &cmd_tables[i];
728 for (j = 0; j < table->count; j++) {
729 const struct drm_i915_cmd_descriptor *desc =
731 u32 curr = desc->cmd.value & desc->cmd.mask;
733 if (curr < previous) {
734 DRM_ERROR("CMD: %s [%d] command table not sorted: "
735 "table=%d entry=%d cmd=0x%08X prev=0x%08X\n",
736 engine->name, engine->id,
737 i, j, curr, previous);
748 static bool check_sorted(const struct intel_engine_cs *engine,
749 const struct drm_i915_reg_descriptor *reg_table,
756 for (i = 0; i < reg_count; i++) {
757 u32 curr = i915_mmio_reg_offset(reg_table[i].addr);
759 if (curr < previous) {
760 DRM_ERROR("CMD: %s [%d] register table not sorted: "
761 "entry=%d reg=0x%08X prev=0x%08X\n",
762 engine->name, engine->id,
773 static bool validate_regs_sorted(struct intel_engine_cs *engine)
776 const struct drm_i915_reg_table *table;
778 for (i = 0; i < engine->reg_table_count; i++) {
779 table = &engine->reg_tables[i];
780 if (!check_sorted(engine, table->regs, table->num_regs))
788 const struct drm_i915_cmd_descriptor *desc;
789 struct hlist_node node;
793 * Different command ranges have different numbers of bits for the opcode. For
794 * example, MI commands use bits 31:23 while 3D commands use bits 31:16. The
795 * problem is that, for example, MI commands use bits 22:16 for other fields
796 * such as GGTT vs PPGTT bits. If we include those bits in the mask then when
797 * we mask a command from a batch it could hash to the wrong bucket due to
798 * non-opcode bits being set. But if we don't include those bits, some 3D
799 * commands may hash to the same bucket due to not including opcode bits that
800 * make the command unique. For now, we will risk hashing to the same bucket.
802 static inline u32 cmd_header_key(u32 x)
804 switch (x >> INSTR_CLIENT_SHIFT) {
806 case INSTR_MI_CLIENT:
807 return x >> STD_MI_OPCODE_SHIFT;
808 case INSTR_RC_CLIENT:
809 return x >> STD_3D_OPCODE_SHIFT;
810 case INSTR_BC_CLIENT:
811 return x >> STD_2D_OPCODE_SHIFT;
815 static int init_hash_table(struct intel_engine_cs *engine,
816 const struct drm_i915_cmd_table *cmd_tables,
821 hash_init(engine->cmd_hash);
823 for (i = 0; i < cmd_table_count; i++) {
824 const struct drm_i915_cmd_table *table = &cmd_tables[i];
826 for (j = 0; j < table->count; j++) {
827 const struct drm_i915_cmd_descriptor *desc =
829 struct cmd_node *desc_node =
830 kmalloc(sizeof(*desc_node), GFP_KERNEL);
835 desc_node->desc = desc;
836 hash_add(engine->cmd_hash, &desc_node->node,
837 cmd_header_key(desc->cmd.value));
844 static void fini_hash_table(struct intel_engine_cs *engine)
846 struct hlist_node *tmp;
847 struct cmd_node *desc_node;
850 hash_for_each_safe(engine->cmd_hash, i, tmp, desc_node, node) {
851 hash_del(&desc_node->node);
857 * intel_engine_init_cmd_parser() - set cmd parser related fields for an engine
858 * @engine: the engine to initialize
860 * Optionally initializes fields related to batch buffer command parsing in the
861 * struct intel_engine_cs based on whether the platform requires software
864 void intel_engine_init_cmd_parser(struct intel_engine_cs *engine)
866 const struct drm_i915_cmd_table *cmd_tables;
870 if (!IS_GEN(engine->i915, 7))
873 switch (engine->class) {
875 if (IS_HASWELL(engine->i915)) {
876 cmd_tables = hsw_render_ring_cmds;
878 ARRAY_SIZE(hsw_render_ring_cmds);
880 cmd_tables = gen7_render_cmds;
881 cmd_table_count = ARRAY_SIZE(gen7_render_cmds);
884 if (IS_HASWELL(engine->i915)) {
885 engine->reg_tables = hsw_render_reg_tables;
886 engine->reg_table_count = ARRAY_SIZE(hsw_render_reg_tables);
888 engine->reg_tables = ivb_render_reg_tables;
889 engine->reg_table_count = ARRAY_SIZE(ivb_render_reg_tables);
892 engine->get_cmd_length_mask = gen7_render_get_cmd_length_mask;
894 case VIDEO_DECODE_CLASS:
895 cmd_tables = gen7_video_cmds;
896 cmd_table_count = ARRAY_SIZE(gen7_video_cmds);
897 engine->get_cmd_length_mask = gen7_bsd_get_cmd_length_mask;
899 case COPY_ENGINE_CLASS:
900 if (IS_HASWELL(engine->i915)) {
901 cmd_tables = hsw_blt_ring_cmds;
902 cmd_table_count = ARRAY_SIZE(hsw_blt_ring_cmds);
904 cmd_tables = gen7_blt_cmds;
905 cmd_table_count = ARRAY_SIZE(gen7_blt_cmds);
908 if (IS_HASWELL(engine->i915)) {
909 engine->reg_tables = hsw_blt_reg_tables;
910 engine->reg_table_count = ARRAY_SIZE(hsw_blt_reg_tables);
912 engine->reg_tables = ivb_blt_reg_tables;
913 engine->reg_table_count = ARRAY_SIZE(ivb_blt_reg_tables);
916 engine->get_cmd_length_mask = gen7_blt_get_cmd_length_mask;
918 case VIDEO_ENHANCEMENT_CLASS:
919 cmd_tables = hsw_vebox_cmds;
920 cmd_table_count = ARRAY_SIZE(hsw_vebox_cmds);
921 /* VECS can use the same length_mask function as VCS */
922 engine->get_cmd_length_mask = gen7_bsd_get_cmd_length_mask;
925 MISSING_CASE(engine->class);
929 if (!validate_cmds_sorted(engine, cmd_tables, cmd_table_count)) {
930 DRM_ERROR("%s: command descriptions are not sorted\n",
934 if (!validate_regs_sorted(engine)) {
935 DRM_ERROR("%s: registers are not sorted\n", engine->name);
939 ret = init_hash_table(engine, cmd_tables, cmd_table_count);
941 DRM_ERROR("%s: initialised failed!\n", engine->name);
942 fini_hash_table(engine);
946 engine->flags |= I915_ENGINE_NEEDS_CMD_PARSER;
950 * intel_engine_cleanup_cmd_parser() - clean up cmd parser related fields
951 * @engine: the engine to clean up
953 * Releases any resources related to command parsing that may have been
954 * initialized for the specified engine.
956 void intel_engine_cleanup_cmd_parser(struct intel_engine_cs *engine)
958 if (!intel_engine_needs_cmd_parser(engine))
961 fini_hash_table(engine);
964 static const struct drm_i915_cmd_descriptor*
965 find_cmd_in_table(struct intel_engine_cs *engine,
968 struct cmd_node *desc_node;
970 hash_for_each_possible(engine->cmd_hash, desc_node, node,
971 cmd_header_key(cmd_header)) {
972 const struct drm_i915_cmd_descriptor *desc = desc_node->desc;
973 if (((cmd_header ^ desc->cmd.value) & desc->cmd.mask) == 0)
981 * Returns a pointer to a descriptor for the command specified by cmd_header.
983 * The caller must supply space for a default descriptor via the default_desc
984 * parameter. If no descriptor for the specified command exists in the engine's
985 * command parser tables, this function fills in default_desc based on the
986 * engine's default length encoding and returns default_desc.
988 static const struct drm_i915_cmd_descriptor*
989 find_cmd(struct intel_engine_cs *engine,
991 const struct drm_i915_cmd_descriptor *desc,
992 struct drm_i915_cmd_descriptor *default_desc)
996 if (((cmd_header ^ desc->cmd.value) & desc->cmd.mask) == 0)
999 desc = find_cmd_in_table(engine, cmd_header);
1003 mask = engine->get_cmd_length_mask(cmd_header);
1007 default_desc->cmd.value = cmd_header;
1008 default_desc->cmd.mask = ~0u << MIN_OPCODE_SHIFT;
1009 default_desc->length.mask = mask;
1010 default_desc->flags = CMD_DESC_SKIP;
1011 return default_desc;
1014 static const struct drm_i915_reg_descriptor *
1015 __find_reg(const struct drm_i915_reg_descriptor *table, int count, u32 addr)
1017 int start = 0, end = count;
1018 while (start < end) {
1019 int mid = start + (end - start) / 2;
1020 int ret = addr - i915_mmio_reg_offset(table[mid].addr);
1031 static const struct drm_i915_reg_descriptor *
1032 find_reg(const struct intel_engine_cs *engine, bool is_master, u32 addr)
1034 const struct drm_i915_reg_table *table = engine->reg_tables;
1035 int count = engine->reg_table_count;
1037 for (; count > 0; ++table, --count) {
1038 if (!table->master || is_master) {
1039 const struct drm_i915_reg_descriptor *reg;
1041 reg = __find_reg(table->regs, table->num_regs, addr);
1050 /* Returns a vmap'd pointer to dst_obj, which the caller must unmap */
1051 static u32 *copy_batch(struct drm_i915_gem_object *dst_obj,
1052 struct drm_i915_gem_object *src_obj,
1053 u32 batch_start_offset,
1055 bool *needs_clflush_after)
1057 unsigned int src_needs_clflush;
1058 unsigned int dst_needs_clflush;
1062 ret = i915_gem_object_prepare_write(dst_obj, &dst_needs_clflush);
1064 return ERR_PTR(ret);
1066 dst = i915_gem_object_pin_map(dst_obj, I915_MAP_FORCE_WB);
1067 i915_gem_object_finish_access(dst_obj);
1071 ret = i915_gem_object_prepare_read(src_obj, &src_needs_clflush);
1073 i915_gem_object_unpin_map(dst_obj);
1074 return ERR_PTR(ret);
1077 src = ERR_PTR(-ENODEV);
1078 if (src_needs_clflush &&
1079 i915_can_memcpy_from_wc(NULL, batch_start_offset, 0)) {
1080 src = i915_gem_object_pin_map(src_obj, I915_MAP_WC);
1082 i915_memcpy_from_wc(dst,
1083 src + batch_start_offset,
1084 ALIGN(batch_len, 16));
1085 i915_gem_object_unpin_map(src_obj);
1092 offset = offset_in_page(batch_start_offset);
1094 /* We can avoid clflushing partial cachelines before the write
1095 * if we only every write full cache-lines. Since we know that
1096 * both the source and destination are in multiples of
1097 * PAGE_SIZE, we can simply round up to the next cacheline.
1098 * We don't care about copying too much here as we only
1099 * validate up to the end of the batch.
1101 if (dst_needs_clflush & CLFLUSH_BEFORE)
1102 batch_len = roundup(batch_len,
1103 boot_cpu_data.x86_clflush_size);
1106 for (n = batch_start_offset >> PAGE_SHIFT; batch_len; n++) {
1107 int len = min_t(int, batch_len, PAGE_SIZE - offset);
1109 src = kmap_atomic(i915_gem_object_get_page(src_obj, n));
1110 if (src_needs_clflush)
1111 drm_clflush_virt_range(src + offset, len);
1112 memcpy(ptr, src + offset, len);
1121 i915_gem_object_finish_access(src_obj);
1123 /* dst_obj is returned with vmap pinned */
1124 *needs_clflush_after = dst_needs_clflush & CLFLUSH_AFTER;
1129 static bool check_cmd(const struct intel_engine_cs *engine,
1130 const struct drm_i915_cmd_descriptor *desc,
1131 const u32 *cmd, u32 length,
1132 const bool is_master)
1134 if (desc->flags & CMD_DESC_SKIP)
1137 if (desc->flags & CMD_DESC_REJECT) {
1138 DRM_DEBUG_DRIVER("CMD: Rejected command: 0x%08X\n", *cmd);
1142 if ((desc->flags & CMD_DESC_MASTER) && !is_master) {
1143 DRM_DEBUG_DRIVER("CMD: Rejected master-only command: 0x%08X\n",
1148 if (desc->flags & CMD_DESC_REGISTER) {
1150 * Get the distance between individual register offset
1151 * fields if the command can perform more than one
1154 const u32 step = desc->reg.step ? desc->reg.step : length;
1157 for (offset = desc->reg.offset; offset < length;
1159 const u32 reg_addr = cmd[offset] & desc->reg.mask;
1160 const struct drm_i915_reg_descriptor *reg =
1161 find_reg(engine, is_master, reg_addr);
1164 DRM_DEBUG_DRIVER("CMD: Rejected register 0x%08X in command: 0x%08X (%s)\n",
1165 reg_addr, *cmd, engine->name);
1170 * Check the value written to the register against the
1171 * allowed mask/value pair given in the whitelist entry.
1174 if (desc->cmd.value == MI_LOAD_REGISTER_MEM) {
1175 DRM_DEBUG_DRIVER("CMD: Rejected LRM to masked register 0x%08X\n",
1180 if (desc->cmd.value == MI_LOAD_REGISTER_REG) {
1181 DRM_DEBUG_DRIVER("CMD: Rejected LRR to masked register 0x%08X\n",
1186 if (desc->cmd.value == MI_LOAD_REGISTER_IMM(1) &&
1187 (offset + 2 > length ||
1188 (cmd[offset + 1] & reg->mask) != reg->value)) {
1189 DRM_DEBUG_DRIVER("CMD: Rejected LRI to masked register 0x%08X\n",
1197 if (desc->flags & CMD_DESC_BITMASK) {
1200 for (i = 0; i < MAX_CMD_DESC_BITMASKS; i++) {
1203 if (desc->bits[i].mask == 0)
1206 if (desc->bits[i].condition_mask != 0) {
1208 desc->bits[i].condition_offset;
1209 u32 condition = cmd[offset] &
1210 desc->bits[i].condition_mask;
1216 if (desc->bits[i].offset >= length) {
1217 DRM_DEBUG_DRIVER("CMD: Rejected command 0x%08X, too short to check bitmask (%s)\n",
1218 *cmd, engine->name);
1222 dword = cmd[desc->bits[i].offset] &
1225 if (dword != desc->bits[i].expected) {
1226 DRM_DEBUG_DRIVER("CMD: Rejected command 0x%08X for bitmask 0x%08X (exp=0x%08X act=0x%08X) (%s)\n",
1229 desc->bits[i].expected,
1230 dword, engine->name);
1239 #define LENGTH_BIAS 2
1242 * i915_parse_cmds() - parse a submitted batch buffer for privilege violations
1243 * @engine: the engine on which the batch is to execute
1244 * @batch_obj: the batch buffer in question
1245 * @shadow_batch_obj: copy of the batch buffer in question
1246 * @batch_start_offset: byte offset in the batch at which execution starts
1247 * @batch_len: length of the commands in batch_obj
1248 * @is_master: is the submitting process the drm master?
1250 * Parses the specified batch buffer looking for privilege violations as
1251 * described in the overview.
1253 * Return: non-zero if the parser finds violations or otherwise fails; -EACCES
1254 * if the batch appears legal but should use hardware parsing
1256 int intel_engine_cmd_parser(struct intel_engine_cs *engine,
1257 struct drm_i915_gem_object *batch_obj,
1258 struct drm_i915_gem_object *shadow_batch_obj,
1259 u32 batch_start_offset,
1263 u32 *cmd, *batch_end;
1264 struct drm_i915_cmd_descriptor default_desc = noop_desc;
1265 const struct drm_i915_cmd_descriptor *desc = &default_desc;
1266 bool needs_clflush_after = false;
1269 cmd = copy_batch(shadow_batch_obj, batch_obj,
1270 batch_start_offset, batch_len,
1271 &needs_clflush_after);
1273 DRM_DEBUG_DRIVER("CMD: Failed to copy batch\n");
1274 return PTR_ERR(cmd);
1278 * We use the batch length as size because the shadow object is as
1279 * large or larger and copy_batch() will write MI_NOPs to the extra
1280 * space. Parsing should be faster in some cases this way.
1282 batch_end = cmd + (batch_len / sizeof(*batch_end));
1286 if (*cmd == MI_BATCH_BUFFER_END) {
1287 if (needs_clflush_after) {
1288 void *ptr = page_mask_bits(shadow_batch_obj->mm.mapping);
1289 drm_clflush_virt_range(ptr,
1290 (void *)(cmd + 1) - ptr);
1295 desc = find_cmd(engine, *cmd, desc, &default_desc);
1297 DRM_DEBUG_DRIVER("CMD: Unrecognized command: 0x%08X\n",
1304 * If the batch buffer contains a chained batch, return an
1305 * error that tells the caller to abort and dispatch the
1306 * workload as a non-secure batch.
1308 if (desc->cmd.value == MI_BATCH_BUFFER_START) {
1313 if (desc->flags & CMD_DESC_FIXED)
1314 length = desc->length.fixed;
1316 length = ((*cmd & desc->length.mask) + LENGTH_BIAS);
1318 if ((batch_end - cmd) < length) {
1319 DRM_DEBUG_DRIVER("CMD: Command length exceeds batch length: 0x%08X length=%u batchlen=%td\n",
1327 if (!check_cmd(engine, desc, cmd, length, is_master)) {
1333 if (cmd >= batch_end) {
1334 DRM_DEBUG_DRIVER("CMD: Got to the end of the buffer w/o a BBE cmd!\n");
1340 i915_gem_object_unpin_map(shadow_batch_obj);
1345 * i915_cmd_parser_get_version() - get the cmd parser version number
1346 * @dev_priv: i915 device private
1348 * The cmd parser maintains a simple increasing integer version number suitable
1349 * for passing to userspace clients to determine what operations are permitted.
1351 * Return: the current version number of the cmd parser
1353 int i915_cmd_parser_get_version(struct drm_i915_private *dev_priv)
1355 struct intel_engine_cs *engine;
1356 bool active = false;
1358 /* If the command parser is not enabled, report 0 - unsupported */
1359 for_each_uabi_engine(engine, dev_priv) {
1360 if (intel_engine_needs_cmd_parser(engine)) {
1369 * Command parser version history
1371 * 1. Initial version. Checks batches and reports violations, but leaves
1372 * hardware parsing enabled (so does not allow new use cases).
1373 * 2. Allow access to the MI_PREDICATE_SRC0 and
1374 * MI_PREDICATE_SRC1 registers.
1375 * 3. Allow access to the GPGPU_THREADS_DISPATCHED register.
1376 * 4. L3 atomic chicken bits of HSW_SCRATCH1 and HSW_ROW_CHICKEN3.
1377 * 5. GPGPU dispatch compute indirect registers.
1378 * 6. TIMESTAMP register and Haswell CS GPR registers
1379 * 7. Allow MI_LOAD_REGISTER_REG between whitelisted registers.
1380 * 8. Don't report cmd_check() failures as EINVAL errors to userspace;
1381 * rely on the HW to NOOP disallowed commands as it would without
1382 * the parser enabled.
1383 * 9. Don't whitelist or handle oacontrol specially, as ownership
1384 * for oacontrol state is moving to i915-perf.