1 /* r300_cmdbuf.c -- Command buffer emission for R300 -*- linux-c -*-
3 * Copyright (C) The Weather Channel, Inc. 2002.
4 * Copyright (C) 2004 Nicolai Haehnle.
7 * The Weather Channel (TM) funded Tungsten Graphics to develop the
8 * initial release of the Radeon 8500 driver under the XFree86 license.
9 * This notice must be preserved.
11 * Permission is hereby granted, free of charge, to any person obtaining a
12 * copy of this software and associated documentation files (the "Software"),
13 * to deal in the Software without restriction, including without limitation
14 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
15 * and/or sell copies of the Software, and to permit persons to whom the
16 * Software is furnished to do so, subject to the following conditions:
18 * The above copyright notice and this permission notice (including the next
19 * paragraph) shall be included in all copies or substantial portions of the
22 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
23 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
24 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
25 * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
26 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
27 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
28 * DEALINGS IN THE SOFTWARE.
31 * Nicolai Haehnle <prefect_@gmx.net>
36 #include "radeon_drm.h"
37 #include "radeon_drv.h"
40 #define R300_SIMULTANEOUS_CLIPRECTS 4
42 /* Values for R300_RE_CLIPRECT_CNTL depending on the number of cliprects
44 static const int r300_cliprect_cntl[4] = {
52 * Emit up to R300_SIMULTANEOUS_CLIPRECTS cliprects from the given command
53 * buffer, starting with index n.
55 static int r300_emit_cliprects(drm_radeon_private_t *dev_priv,
56 drm_radeon_kcmd_buffer_t *cmdbuf, int n)
58 struct drm_clip_rect box;
63 nr = cmdbuf->nbox - n;
64 if (nr > R300_SIMULTANEOUS_CLIPRECTS)
65 nr = R300_SIMULTANEOUS_CLIPRECTS;
67 DRM_DEBUG("%i cliprects\n", nr);
70 BEGIN_RING(6 + nr * 2);
71 OUT_RING(CP_PACKET0(R300_RE_CLIPRECT_TL_0, nr * 2 - 1));
73 for (i = 0; i < nr; ++i) {
74 if (DRM_COPY_FROM_USER_UNCHECKED
75 (&box, &cmdbuf->boxes[n + i], sizeof(box))) {
76 DRM_ERROR("copy cliprect faulted\n");
80 box.x2--; /* Hardware expects inclusive bottom-right corner */
83 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
93 box.x1 = (box.x1 + R300_CLIPRECT_OFFSET) &
95 box.y1 = (box.y1 + R300_CLIPRECT_OFFSET) &
97 box.x2 = (box.x2 + R300_CLIPRECT_OFFSET) &
99 box.y2 = (box.y2 + R300_CLIPRECT_OFFSET) &
103 OUT_RING((box.x1 << R300_CLIPRECT_X_SHIFT) |
104 (box.y1 << R300_CLIPRECT_Y_SHIFT));
105 OUT_RING((box.x2 << R300_CLIPRECT_X_SHIFT) |
106 (box.y2 << R300_CLIPRECT_Y_SHIFT));
110 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, r300_cliprect_cntl[nr - 1]);
112 /* TODO/SECURITY: Force scissors to a safe value, otherwise the
113 * client might be able to trample over memory.
114 * The impact should be very limited, but I'd rather be safe than
117 OUT_RING(CP_PACKET0(R300_RE_SCISSORS_TL, 1));
119 OUT_RING(R300_SCISSORS_X_MASK | R300_SCISSORS_Y_MASK);
122 /* Why we allow zero cliprect rendering:
123 * There are some commands in a command buffer that must be submitted
124 * even when there are no cliprects, e.g. DMA buffer discard
125 * or state setting (though state setting could be avoided by
126 * simulating a loss of context).
128 * Now since the cmdbuf interface is so chaotic right now (and is
129 * bound to remain that way for a bit until things settle down),
130 * it is basically impossible to filter out the commands that are
131 * necessary and those that aren't.
133 * So I choose the safe way and don't do any filtering at all;
134 * instead, I simply set up the engine so that all rendering
135 * can't produce any fragments.
138 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, 0);
142 /* flus cache and wait idle clean after cliprect change */
144 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
145 OUT_RING(R300_RB3D_DC_FLUSH);
148 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
149 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
152 dev_priv->track_flush |= RADEON_FLUSH_EMITED;
157 static u8 r300_reg_flags[0x10000 >> 2];
159 void r300_init_reg_flags(struct drm_device *dev)
162 drm_radeon_private_t *dev_priv = dev->dev_private;
164 memset(r300_reg_flags, 0, 0x10000 >> 2);
165 #define ADD_RANGE_MARK(reg, count,mark) \
166 for(i=((reg)>>2);i<((reg)>>2)+(count);i++)\
167 r300_reg_flags[i]|=(mark);
170 #define MARK_CHECK_OFFSET 2
172 #define ADD_RANGE(reg, count) ADD_RANGE_MARK(reg, count, MARK_SAFE)
174 /* these match cmducs() command in r300_driver/r300/r300_cmdbuf.c */
175 ADD_RANGE(R300_SE_VPORT_XSCALE, 6);
176 ADD_RANGE(R300_VAP_CNTL, 1);
177 ADD_RANGE(R300_SE_VTE_CNTL, 2);
178 ADD_RANGE(0x2134, 2);
179 ADD_RANGE(R300_VAP_CNTL_STATUS, 1);
180 ADD_RANGE(R300_VAP_INPUT_CNTL_0, 2);
181 ADD_RANGE(0x21DC, 1);
182 ADD_RANGE(R300_VAP_UNKNOWN_221C, 1);
183 ADD_RANGE(R300_VAP_CLIP_X_0, 4);
184 ADD_RANGE(R300_VAP_PVS_STATE_FLUSH_REG, 1);
185 ADD_RANGE(R300_VAP_UNKNOWN_2288, 1);
186 ADD_RANGE(R300_VAP_OUTPUT_VTX_FMT_0, 2);
187 ADD_RANGE(R300_VAP_PVS_CNTL_1, 3);
188 ADD_RANGE(R300_GB_ENABLE, 1);
189 ADD_RANGE(R300_GB_MSPOS0, 5);
190 ADD_RANGE(R300_TX_INVALTAGS, 1);
191 ADD_RANGE(R300_TX_ENABLE, 1);
192 ADD_RANGE(0x4200, 4);
193 ADD_RANGE(0x4214, 1);
194 ADD_RANGE(R300_RE_POINTSIZE, 1);
195 ADD_RANGE(0x4230, 3);
196 ADD_RANGE(R300_RE_LINE_CNT, 1);
197 ADD_RANGE(R300_RE_UNK4238, 1);
198 ADD_RANGE(0x4260, 3);
199 ADD_RANGE(R300_RE_SHADE, 4);
200 ADD_RANGE(R300_RE_POLYGON_MODE, 5);
201 ADD_RANGE(R300_RE_ZBIAS_CNTL, 1);
202 ADD_RANGE(R300_RE_ZBIAS_T_FACTOR, 4);
203 ADD_RANGE(R300_RE_OCCLUSION_CNTL, 1);
204 ADD_RANGE(R300_RE_CULL_CNTL, 1);
205 ADD_RANGE(0x42C0, 2);
206 ADD_RANGE(R300_RS_CNTL_0, 2);
208 ADD_RANGE(R300_SC_HYPERZ, 2);
209 ADD_RANGE(0x43E8, 1);
211 ADD_RANGE(0x46A4, 5);
213 ADD_RANGE(R300_RE_FOG_STATE, 1);
214 ADD_RANGE(R300_FOG_COLOR_R, 3);
215 ADD_RANGE(R300_PP_ALPHA_TEST, 2);
216 ADD_RANGE(0x4BD8, 1);
217 ADD_RANGE(R300_PFS_PARAM_0_X, 64);
218 ADD_RANGE(0x4E00, 1);
219 ADD_RANGE(R300_RB3D_CBLEND, 2);
220 ADD_RANGE(R300_RB3D_COLORMASK, 1);
221 ADD_RANGE(R300_RB3D_BLEND_COLOR, 3);
222 ADD_RANGE_MARK(R300_RB3D_COLOROFFSET0, 1, MARK_CHECK_OFFSET); /* check offset */
223 ADD_RANGE(R300_RB3D_COLORPITCH0, 1);
224 ADD_RANGE(0x4E50, 9);
225 ADD_RANGE(0x4E88, 1);
226 ADD_RANGE(0x4EA0, 2);
227 ADD_RANGE(R300_ZB_CNTL, 3);
228 ADD_RANGE(R300_ZB_FORMAT, 4);
229 ADD_RANGE_MARK(R300_ZB_DEPTHOFFSET, 1, MARK_CHECK_OFFSET); /* check offset */
230 ADD_RANGE(R300_ZB_DEPTHPITCH, 1);
231 ADD_RANGE(R300_ZB_DEPTHCLEARVALUE, 1);
232 ADD_RANGE(R300_ZB_ZMASK_OFFSET, 13);
234 ADD_RANGE(R300_TX_FILTER_0, 16);
235 ADD_RANGE(R300_TX_FILTER1_0, 16);
236 ADD_RANGE(R300_TX_SIZE_0, 16);
237 ADD_RANGE(R300_TX_FORMAT_0, 16);
238 ADD_RANGE(R300_TX_PITCH_0, 16);
239 /* Texture offset is dangerous and needs more checking */
240 ADD_RANGE_MARK(R300_TX_OFFSET_0, 16, MARK_CHECK_OFFSET);
241 ADD_RANGE(R300_TX_CHROMA_KEY_0, 16);
242 ADD_RANGE(R300_TX_BORDER_COLOR_0, 16);
244 /* Sporadic registers used as primitives are emitted */
245 ADD_RANGE(R300_ZB_ZCACHE_CTLSTAT, 1);
246 ADD_RANGE(R300_RB3D_DSTCACHE_CTLSTAT, 1);
247 ADD_RANGE(R300_VAP_INPUT_ROUTE_0_0, 8);
248 ADD_RANGE(R300_VAP_INPUT_ROUTE_1_0, 8);
250 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
251 ADD_RANGE(R500_VAP_INDEX_OFFSET, 1);
252 ADD_RANGE(R500_US_CONFIG, 2);
253 ADD_RANGE(R500_US_CODE_ADDR, 3);
254 ADD_RANGE(R500_US_FC_CTRL, 1);
255 ADD_RANGE(R500_RS_IP_0, 16);
256 ADD_RANGE(R500_RS_INST_0, 16);
257 ADD_RANGE(R500_RB3D_COLOR_CLEAR_VALUE_AR, 2);
258 ADD_RANGE(R500_RB3D_CONSTANT_COLOR_AR, 2);
259 ADD_RANGE(R500_ZB_FIFO_SIZE, 2);
261 ADD_RANGE(R300_PFS_CNTL_0, 3);
262 ADD_RANGE(R300_PFS_NODE_0, 4);
263 ADD_RANGE(R300_PFS_TEXI_0, 64);
264 ADD_RANGE(R300_PFS_INSTR0_0, 64);
265 ADD_RANGE(R300_PFS_INSTR1_0, 64);
266 ADD_RANGE(R300_PFS_INSTR2_0, 64);
267 ADD_RANGE(R300_PFS_INSTR3_0, 64);
268 ADD_RANGE(R300_RS_INTERP_0, 8);
269 ADD_RANGE(R300_RS_ROUTE_0, 8);
274 static __inline__ int r300_check_range(unsigned reg, int count)
279 for (i = (reg >> 2); i < (reg >> 2) + count; i++)
280 if (r300_reg_flags[i] != MARK_SAFE)
285 static __inline__ int r300_emit_carefully_checked_packet0(drm_radeon_private_t *
287 drm_radeon_kcmd_buffer_t
289 drm_r300_cmd_header_t
298 sz = header.packet0.count;
299 reg = (header.packet0.reghi << 8) | header.packet0.reglo;
301 if ((sz > 64) || (sz < 0)) {
303 ("Cannot emit more than 64 values at a time (reg=%04x sz=%d)\n",
307 for (i = 0; i < sz; i++) {
308 values[i] = ((int *)cmdbuf->buf)[i];
309 switch (r300_reg_flags[(reg >> 2) + i]) {
312 case MARK_CHECK_OFFSET:
313 if (!radeon_check_offset(dev_priv, (u32) values[i])) {
315 ("Offset failed range check (reg=%04x sz=%d)\n",
321 DRM_ERROR("Register %04x failed check as flag=%02x\n",
322 reg + i * 4, r300_reg_flags[(reg >> 2) + i]);
328 OUT_RING(CP_PACKET0(reg, sz - 1));
329 OUT_RING_TABLE(values, sz);
332 cmdbuf->buf += sz * 4;
333 cmdbuf->bufsz -= sz * 4;
339 * Emits a packet0 setting arbitrary registers.
340 * Called by r300_do_cp_cmdbuf.
342 * Note that checks are performed on contents and addresses of the registers
344 static __inline__ int r300_emit_packet0(drm_radeon_private_t *dev_priv,
345 drm_radeon_kcmd_buffer_t *cmdbuf,
346 drm_r300_cmd_header_t header)
352 sz = header.packet0.count;
353 reg = (header.packet0.reghi << 8) | header.packet0.reglo;
358 if (sz * 4 > cmdbuf->bufsz)
361 if (reg + sz * 4 >= 0x10000) {
362 DRM_ERROR("No such registers in hardware reg=%04x sz=%d\n", reg,
367 if (r300_check_range(reg, sz)) {
368 /* go and check everything */
369 return r300_emit_carefully_checked_packet0(dev_priv, cmdbuf,
372 /* the rest of the data is safe to emit, whatever the values the user passed */
375 OUT_RING(CP_PACKET0(reg, sz - 1));
376 OUT_RING_TABLE((int *)cmdbuf->buf, sz);
379 cmdbuf->buf += sz * 4;
380 cmdbuf->bufsz -= sz * 4;
386 * Uploads user-supplied vertex program instructions or parameters onto
388 * Called by r300_do_cp_cmdbuf.
390 static __inline__ int r300_emit_vpu(drm_radeon_private_t *dev_priv,
391 drm_radeon_kcmd_buffer_t *cmdbuf,
392 drm_r300_cmd_header_t header)
398 sz = header.vpu.count;
399 addr = (header.vpu.adrhi << 8) | header.vpu.adrlo;
403 if (sz * 16 > cmdbuf->bufsz)
406 /* VAP is very sensitive so we purge cache before we program it
407 * and we also flush its state before & after */
409 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
410 OUT_RING(R300_RB3D_DC_FLUSH);
411 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
412 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
413 OUT_RING(CP_PACKET0(R300_VAP_PVS_STATE_FLUSH_REG, 0));
417 dev_priv->track_flush |= RADEON_FLUSH_EMITED;
419 BEGIN_RING(3 + sz * 4);
420 OUT_RING_REG(R300_VAP_PVS_UPLOAD_ADDRESS, addr);
421 OUT_RING(CP_PACKET0_TABLE(R300_VAP_PVS_UPLOAD_DATA, sz * 4 - 1));
422 OUT_RING_TABLE((int *)cmdbuf->buf, sz * 4);
426 OUT_RING(CP_PACKET0(R300_VAP_PVS_STATE_FLUSH_REG, 0));
430 cmdbuf->buf += sz * 16;
431 cmdbuf->bufsz -= sz * 16;
437 * Emit a clear packet from userspace.
438 * Called by r300_emit_packet3.
440 static __inline__ int r300_emit_clear(drm_radeon_private_t *dev_priv,
441 drm_radeon_kcmd_buffer_t *cmdbuf)
445 if (8 * 4 > cmdbuf->bufsz)
449 OUT_RING(CP_PACKET3(R200_3D_DRAW_IMMD_2, 8));
450 OUT_RING(R300_PRIM_TYPE_POINT | R300_PRIM_WALK_RING |
451 (1 << R300_PRIM_NUM_VERTICES_SHIFT));
452 OUT_RING_TABLE((int *)cmdbuf->buf, 8);
456 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
457 OUT_RING(R300_RB3D_DC_FLUSH);
458 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
459 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
462 dev_priv->track_flush |= RADEON_FLUSH_EMITED;
464 cmdbuf->buf += 8 * 4;
465 cmdbuf->bufsz -= 8 * 4;
470 static __inline__ int r300_emit_3d_load_vbpntr(drm_radeon_private_t *dev_priv,
471 drm_radeon_kcmd_buffer_t *cmdbuf,
475 #define MAX_ARRAY_PACKET 64
476 u32 payload[MAX_ARRAY_PACKET];
480 count = (header >> 16) & 0x3fff;
482 if ((count + 1) > MAX_ARRAY_PACKET) {
483 DRM_ERROR("Too large payload in 3D_LOAD_VBPNTR (count=%d)\n",
487 memset(payload, 0, MAX_ARRAY_PACKET * 4);
488 memcpy(payload, cmdbuf->buf + 4, (count + 1) * 4);
490 /* carefully check packet contents */
492 narrays = payload[0];
495 while ((k < narrays) && (i < (count + 1))) {
496 i++; /* skip attribute field */
497 if (!radeon_check_offset(dev_priv, payload[i])) {
499 ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
507 /* have one more to process, they come in pairs */
508 if (!radeon_check_offset(dev_priv, payload[i])) {
510 ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
517 /* do the counts match what we expect ? */
518 if ((k != narrays) || (i != (count + 1))) {
520 ("Malformed 3D_LOAD_VBPNTR packet (k=%d i=%d narrays=%d count+1=%d).\n",
521 k, i, narrays, count + 1);
525 /* all clear, output packet */
527 BEGIN_RING(count + 2);
529 OUT_RING_TABLE(payload, count + 1);
532 cmdbuf->buf += (count + 2) * 4;
533 cmdbuf->bufsz -= (count + 2) * 4;
538 static __inline__ int r300_emit_bitblt_multi(drm_radeon_private_t *dev_priv,
539 drm_radeon_kcmd_buffer_t *cmdbuf)
541 u32 *cmd = (u32 *) cmdbuf->buf;
545 count=(cmd[0]>>16) & 0x3fff;
547 if (cmd[0] & 0x8000) {
550 if (cmd[1] & (RADEON_GMC_SRC_PITCH_OFFSET_CNTL
551 | RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
552 offset = cmd[2] << 10;
553 ret = !radeon_check_offset(dev_priv, offset);
555 DRM_ERROR("Invalid bitblt first offset is %08X\n", offset);
560 if ((cmd[1] & RADEON_GMC_SRC_PITCH_OFFSET_CNTL) &&
561 (cmd[1] & RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
562 offset = cmd[3] << 10;
563 ret = !radeon_check_offset(dev_priv, offset);
565 DRM_ERROR("Invalid bitblt second offset is %08X\n", offset);
574 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
577 cmdbuf->buf += (count+2)*4;
578 cmdbuf->bufsz -= (count+2)*4;
583 static __inline__ int r300_emit_draw_indx_2(drm_radeon_private_t *dev_priv,
584 drm_radeon_kcmd_buffer_t *cmdbuf)
591 cmd = (u32 *) cmdbuf->buf;
592 count = (cmd[0]>>16) & 0x3fff;
593 expected_count = cmd[1] >> 16;
594 if (!(cmd[1] & R300_VAP_VF_CNTL__INDEX_SIZE_32bit))
595 expected_count = (expected_count+1)/2;
597 if (count && count != expected_count) {
598 DRM_ERROR("3D_DRAW_INDX_2: packet size %i, expected %i\n",
599 count, expected_count);
605 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
608 cmdbuf->buf += (count+2)*4;
609 cmdbuf->bufsz -= (count+2)*4;
612 drm_r300_cmd_header_t header;
614 if (cmdbuf->bufsz < 4*4 + sizeof(header)) {
615 DRM_ERROR("3D_DRAW_INDX_2: expect subsequent INDX_BUFFER, but stream is too short.\n");
619 header.u = *(unsigned int *)cmdbuf->buf;
621 cmdbuf->buf += sizeof(header);
622 cmdbuf->bufsz -= sizeof(header);
623 cmd = (u32 *) cmdbuf->buf;
625 if (header.header.cmd_type != R300_CMD_PACKET3 ||
626 header.packet3.packet != R300_CMD_PACKET3_RAW ||
627 cmd[0] != CP_PACKET3(RADEON_CP_INDX_BUFFER, 2)) {
628 DRM_ERROR("3D_DRAW_INDX_2: expect subsequent INDX_BUFFER.\n");
632 if ((cmd[1] & 0x8000ffff) != 0x80000810) {
633 DRM_ERROR("Invalid indx_buffer reg address %08X\n", cmd[1]);
636 if (!radeon_check_offset(dev_priv, cmd[2])) {
637 DRM_ERROR("Invalid indx_buffer offset is %08X\n", cmd[2]);
640 if (cmd[3] != expected_count) {
641 DRM_ERROR("INDX_BUFFER: buffer size %i, expected %i\n",
642 cmd[3], expected_count);
648 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), 3);
652 cmdbuf->bufsz -= 4*4;
658 static __inline__ int r300_emit_raw_packet3(drm_radeon_private_t *dev_priv,
659 drm_radeon_kcmd_buffer_t *cmdbuf)
665 if (4 > cmdbuf->bufsz)
668 /* Fixme !! This simply emits a packet without much checking.
669 We need to be smarter. */
671 /* obtain first word - actual packet3 header */
672 header = *(u32 *) cmdbuf->buf;
674 /* Is it packet 3 ? */
675 if ((header >> 30) != 0x3) {
676 DRM_ERROR("Not a packet3 header (0x%08x)\n", header);
680 count = (header >> 16) & 0x3fff;
682 /* Check again now that we know how much data to expect */
683 if ((count + 2) * 4 > cmdbuf->bufsz) {
685 ("Expected packet3 of length %d but have only %d bytes left\n",
686 (count + 2) * 4, cmdbuf->bufsz);
690 /* Is it a packet type we know about ? */
691 switch (header & 0xff00) {
692 case RADEON_3D_LOAD_VBPNTR: /* load vertex array pointers */
693 return r300_emit_3d_load_vbpntr(dev_priv, cmdbuf, header);
695 case RADEON_CNTL_BITBLT_MULTI:
696 return r300_emit_bitblt_multi(dev_priv, cmdbuf);
698 case RADEON_CP_INDX_BUFFER:
699 DRM_ERROR("packet3 INDX_BUFFER without preceding 3D_DRAW_INDX_2 is illegal.\n");
701 case RADEON_CP_3D_DRAW_IMMD_2:
702 /* triggers drawing using in-packet vertex data */
703 case RADEON_CP_3D_DRAW_VBUF_2:
704 /* triggers drawing of vertex buffers setup elsewhere */
705 dev_priv->track_flush &= ~(RADEON_FLUSH_EMITED |
706 RADEON_PURGE_EMITED);
708 case RADEON_CP_3D_DRAW_INDX_2:
709 /* triggers drawing using indices to vertex buffer */
710 /* whenever we send vertex we clear flush & purge */
711 dev_priv->track_flush &= ~(RADEON_FLUSH_EMITED |
712 RADEON_PURGE_EMITED);
713 return r300_emit_draw_indx_2(dev_priv, cmdbuf);
714 case RADEON_WAIT_FOR_IDLE:
716 /* these packets are safe */
719 DRM_ERROR("Unknown packet3 header (0x%08x)\n", header);
723 BEGIN_RING(count + 2);
725 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
728 cmdbuf->buf += (count + 2) * 4;
729 cmdbuf->bufsz -= (count + 2) * 4;
735 * Emit a rendering packet3 from userspace.
736 * Called by r300_do_cp_cmdbuf.
738 static __inline__ int r300_emit_packet3(drm_radeon_private_t *dev_priv,
739 drm_radeon_kcmd_buffer_t *cmdbuf,
740 drm_r300_cmd_header_t header)
744 char *orig_buf = cmdbuf->buf;
745 int orig_bufsz = cmdbuf->bufsz;
747 /* This is a do-while-loop so that we run the interior at least once,
748 * even if cmdbuf->nbox is 0. Compare r300_emit_cliprects for rationale.
752 if (cmdbuf->nbox > R300_SIMULTANEOUS_CLIPRECTS) {
753 ret = r300_emit_cliprects(dev_priv, cmdbuf, n);
757 cmdbuf->buf = orig_buf;
758 cmdbuf->bufsz = orig_bufsz;
761 switch (header.packet3.packet) {
762 case R300_CMD_PACKET3_CLEAR:
763 DRM_DEBUG("R300_CMD_PACKET3_CLEAR\n");
764 ret = r300_emit_clear(dev_priv, cmdbuf);
766 DRM_ERROR("r300_emit_clear failed\n");
771 case R300_CMD_PACKET3_RAW:
772 DRM_DEBUG("R300_CMD_PACKET3_RAW\n");
773 ret = r300_emit_raw_packet3(dev_priv, cmdbuf);
775 DRM_ERROR("r300_emit_raw_packet3 failed\n");
781 DRM_ERROR("bad packet3 type %i at %p\n",
782 header.packet3.packet,
783 cmdbuf->buf - sizeof(header));
787 n += R300_SIMULTANEOUS_CLIPRECTS;
788 } while (n < cmdbuf->nbox);
793 /* Some of the R300 chips seem to be extremely touchy about the two registers
794 * that are configured in r300_pacify.
795 * Among the worst offenders seems to be the R300 ND (0x4E44): When userspace
796 * sends a command buffer that contains only state setting commands and a
797 * vertex program/parameter upload sequence, this will eventually lead to a
798 * lockup, unless the sequence is bracketed by calls to r300_pacify.
799 * So we should take great care to *always* call r300_pacify before
800 * *anything* 3D related, and again afterwards. This is what the
801 * call bracket in r300_do_cp_cmdbuf is for.
805 * Emit the sequence to pacify R300.
807 static __inline__ void r300_pacify(drm_radeon_private_t *dev_priv)
809 uint32_t cache_z, cache_3d, cache_2d;
812 cache_z = R300_ZC_FLUSH;
813 cache_2d = R300_RB2D_DC_FLUSH;
814 cache_3d = R300_RB3D_DC_FLUSH;
815 if (!(dev_priv->track_flush & RADEON_PURGE_EMITED)) {
816 /* we can purge, primitive where draw since last purge */
817 cache_z |= R300_ZC_FREE;
818 cache_2d |= R300_RB2D_DC_FREE;
819 cache_3d |= R300_RB3D_DC_FREE;
822 /* flush & purge zbuffer */
824 OUT_RING(CP_PACKET0(R300_ZB_ZCACHE_CTLSTAT, 0));
827 /* flush & purge 3d */
829 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
832 /* flush & purge texture */
834 OUT_RING(CP_PACKET0(R300_TX_INVALTAGS, 0));
837 /* FIXME: is this one really needed ? */
839 OUT_RING(CP_PACKET0(R300_RB3D_AARESOLVE_CTL, 0));
843 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
844 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
846 /* flush & purge 2d through E2 as RB2D will trigger lockup */
848 OUT_RING(CP_PACKET0(R300_DSTCACHE_CTLSTAT, 0));
850 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
851 OUT_RING(RADEON_WAIT_2D_IDLECLEAN |
852 RADEON_WAIT_HOST_IDLECLEAN);
854 /* set flush & purge flags */
855 dev_priv->track_flush |= RADEON_FLUSH_EMITED | RADEON_PURGE_EMITED;
859 * Called by r300_do_cp_cmdbuf to update the internal buffer age and state.
860 * The actual age emit is done by r300_do_cp_cmdbuf, which is why you must
861 * be careful about how this function is called.
863 static void r300_discard_buffer(struct drm_device *dev, struct drm_master *master, struct drm_buf *buf)
865 drm_radeon_buf_priv_t *buf_priv = buf->dev_private;
866 struct drm_radeon_master_private *master_priv = master->driver_priv;
868 buf_priv->age = ++master_priv->sarea_priv->last_dispatch;
873 static void r300_cmd_wait(drm_radeon_private_t * dev_priv,
874 drm_r300_cmd_header_t header)
879 if (!header.wait.flags)
884 switch(header.wait.flags) {
886 wait_until = RADEON_WAIT_2D_IDLE;
889 wait_until = RADEON_WAIT_3D_IDLE;
891 case R300_NEW_WAIT_2D_3D:
892 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_3D_IDLE;
894 case R300_NEW_WAIT_2D_2D_CLEAN:
895 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_2D_IDLECLEAN;
897 case R300_NEW_WAIT_3D_3D_CLEAN:
898 wait_until = RADEON_WAIT_3D_IDLE|RADEON_WAIT_3D_IDLECLEAN;
900 case R300_NEW_WAIT_2D_2D_CLEAN_3D_3D_CLEAN:
901 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_2D_IDLECLEAN;
902 wait_until |= RADEON_WAIT_3D_IDLE|RADEON_WAIT_3D_IDLECLEAN;
909 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
910 OUT_RING(wait_until);
914 static int r300_scratch(drm_radeon_private_t *dev_priv,
915 drm_radeon_kcmd_buffer_t *cmdbuf,
916 drm_r300_cmd_header_t header)
919 u32 i, buf_idx, h_pending;
923 (sizeof(u64) + header.scratch.n_bufs * sizeof(buf_idx))) {
927 if (header.scratch.reg >= 5) {
931 dev_priv->scratch_ages[header.scratch.reg]++;
933 ref_age_base = (u32 *)(unsigned long)*((uint64_t *)cmdbuf->buf);
935 cmdbuf->buf += sizeof(u64);
936 cmdbuf->bufsz -= sizeof(u64);
938 for (i=0; i < header.scratch.n_bufs; i++) {
939 buf_idx = *(u32 *)cmdbuf->buf;
940 buf_idx *= 2; /* 8 bytes per buf */
942 if (DRM_COPY_TO_USER(ref_age_base + buf_idx, &dev_priv->scratch_ages[header.scratch.reg], sizeof(u32))) {
946 if (DRM_COPY_FROM_USER(&h_pending, ref_age_base + buf_idx + 1, sizeof(u32))) {
950 if (h_pending == 0) {
956 if (DRM_COPY_TO_USER(ref_age_base + buf_idx + 1, &h_pending, sizeof(u32))) {
960 cmdbuf->buf += sizeof(buf_idx);
961 cmdbuf->bufsz -= sizeof(buf_idx);
965 OUT_RING( CP_PACKET0( RADEON_SCRATCH_REG0 + header.scratch.reg * 4, 0 ) );
966 OUT_RING( dev_priv->scratch_ages[header.scratch.reg] );
973 * Uploads user-supplied vertex program instructions or parameters onto
975 * Called by r300_do_cp_cmdbuf.
977 static inline int r300_emit_r500fp(drm_radeon_private_t *dev_priv,
978 drm_radeon_kcmd_buffer_t *cmdbuf,
979 drm_r300_cmd_header_t header)
988 sz = header.r500fp.count;
989 /* address is 9 bits 0 - 8, bit 1 of flags is part of address */
990 addr = ((header.r500fp.adrhi_flags & 1) << 8) | header.r500fp.adrlo;
992 type = !!(header.r500fp.adrhi_flags & R500FP_CONSTANT_TYPE);
993 clamp = !!(header.r500fp.adrhi_flags & R500FP_CONSTANT_CLAMP);
995 addr |= (type << 16);
996 addr |= (clamp << 17);
998 stride = type ? 4 : 6;
1000 DRM_DEBUG("r500fp %d %d type: %d\n", sz, addr, type);
1003 if (sz * stride * 4 > cmdbuf->bufsz)
1006 BEGIN_RING(3 + sz * stride);
1007 OUT_RING_REG(R500_GA_US_VECTOR_INDEX, addr);
1008 OUT_RING(CP_PACKET0_TABLE(R500_GA_US_VECTOR_DATA, sz * stride - 1));
1009 OUT_RING_TABLE((int *)cmdbuf->buf, sz * stride);
1013 cmdbuf->buf += sz * stride * 4;
1014 cmdbuf->bufsz -= sz * stride * 4;
1021 * Parses and validates a user-supplied command buffer and emits appropriate
1022 * commands on the DMA ring buffer.
1023 * Called by the ioctl handler function radeon_cp_cmdbuf.
1025 int r300_do_cp_cmdbuf(struct drm_device *dev,
1026 struct drm_file *file_priv,
1027 drm_radeon_kcmd_buffer_t *cmdbuf)
1029 drm_radeon_private_t *dev_priv = dev->dev_private;
1030 struct drm_radeon_master_private *master_priv = file_priv->master->driver_priv;
1031 struct drm_device_dma *dma = dev->dma;
1032 struct drm_buf *buf = NULL;
1033 int emit_dispatch_age = 0;
1039 r300_pacify(dev_priv);
1041 if (cmdbuf->nbox <= R300_SIMULTANEOUS_CLIPRECTS) {
1042 ret = r300_emit_cliprects(dev_priv, cmdbuf, 0);
1047 while (cmdbuf->bufsz >= sizeof(drm_r300_cmd_header_t)) {
1049 drm_r300_cmd_header_t header;
1051 header.u = *(unsigned int *)cmdbuf->buf;
1053 cmdbuf->buf += sizeof(header);
1054 cmdbuf->bufsz -= sizeof(header);
1056 switch (header.header.cmd_type) {
1057 case R300_CMD_PACKET0:
1058 DRM_DEBUG("R300_CMD_PACKET0\n");
1059 ret = r300_emit_packet0(dev_priv, cmdbuf, header);
1061 DRM_ERROR("r300_emit_packet0 failed\n");
1067 DRM_DEBUG("R300_CMD_VPU\n");
1068 ret = r300_emit_vpu(dev_priv, cmdbuf, header);
1070 DRM_ERROR("r300_emit_vpu failed\n");
1075 case R300_CMD_PACKET3:
1076 DRM_DEBUG("R300_CMD_PACKET3\n");
1077 ret = r300_emit_packet3(dev_priv, cmdbuf, header);
1079 DRM_ERROR("r300_emit_packet3 failed\n");
1084 case R300_CMD_END3D:
1085 DRM_DEBUG("R300_CMD_END3D\n");
1087 Ideally userspace driver should not need to issue this call,
1088 i.e. the drm driver should issue it automatically and prevent
1091 In practice, we do not understand why this call is needed and what
1092 it does (except for some vague guesses that it has to do with cache
1093 coherence) and so the user space driver does it.
1095 Once we are sure which uses prevent lockups the code could be moved
1096 into the kernel and the userspace driver will not
1097 need to use this command.
1099 Note that issuing this command does not hurt anything
1100 except, possibly, performance */
1101 r300_pacify(dev_priv);
1104 case R300_CMD_CP_DELAY:
1105 /* simple enough, we can do it here */
1106 DRM_DEBUG("R300_CMD_CP_DELAY\n");
1111 BEGIN_RING(header.delay.count);
1112 for (i = 0; i < header.delay.count; i++)
1113 OUT_RING(RADEON_CP_PACKET2);
1118 case R300_CMD_DMA_DISCARD:
1119 DRM_DEBUG("RADEON_CMD_DMA_DISCARD\n");
1120 idx = header.dma.buf_idx;
1121 if (idx < 0 || idx >= dma->buf_count) {
1122 DRM_ERROR("buffer index %d (of %d max)\n",
1123 idx, dma->buf_count - 1);
1128 buf = dma->buflist[idx];
1129 if (buf->file_priv != file_priv || buf->pending) {
1130 DRM_ERROR("bad buffer %p %p %d\n",
1131 buf->file_priv, file_priv,
1137 emit_dispatch_age = 1;
1138 r300_discard_buffer(dev, file_priv->master, buf);
1142 DRM_DEBUG("R300_CMD_WAIT\n");
1143 r300_cmd_wait(dev_priv, header);
1146 case R300_CMD_SCRATCH:
1147 DRM_DEBUG("R300_CMD_SCRATCH\n");
1148 ret = r300_scratch(dev_priv, cmdbuf, header);
1150 DRM_ERROR("r300_scratch failed\n");
1155 case R300_CMD_R500FP:
1156 if ((dev_priv->flags & RADEON_FAMILY_MASK) < CHIP_RV515) {
1157 DRM_ERROR("Calling r500 command on r300 card\n");
1161 DRM_DEBUG("R300_CMD_R500FP\n");
1162 ret = r300_emit_r500fp(dev_priv, cmdbuf, header);
1164 DRM_ERROR("r300_emit_r500fp failed\n");
1169 DRM_ERROR("bad cmd_type %i at %p\n",
1170 header.header.cmd_type,
1171 cmdbuf->buf - sizeof(header));
1180 r300_pacify(dev_priv);
1182 /* We emit the vertex buffer age here, outside the pacifier "brackets"
1184 * (1) This may coalesce multiple age emissions into a single one and
1185 * (2) more importantly, some chips lock up hard when scratch registers
1186 * are written inside the pacifier bracket.
1188 if (emit_dispatch_age) {
1191 /* Emit the vertex buffer age */
1193 RADEON_DISPATCH_AGE(master_priv->sarea_priv->last_dispatch);