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 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
90 box.x1 = (box.x1 + R300_CLIPRECT_OFFSET) &
92 box.y1 = (box.y1 + R300_CLIPRECT_OFFSET) &
94 box.x2 = (box.x2 + R300_CLIPRECT_OFFSET) &
96 box.y2 = (box.y2 + R300_CLIPRECT_OFFSET) &
100 OUT_RING((box.x1 << R300_CLIPRECT_X_SHIFT) |
101 (box.y1 << R300_CLIPRECT_Y_SHIFT));
102 OUT_RING((box.x2 << R300_CLIPRECT_X_SHIFT) |
103 (box.y2 << R300_CLIPRECT_Y_SHIFT));
107 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, r300_cliprect_cntl[nr - 1]);
109 /* TODO/SECURITY: Force scissors to a safe value, otherwise the
110 * client might be able to trample over memory.
111 * The impact should be very limited, but I'd rather be safe than
114 OUT_RING(CP_PACKET0(R300_RE_SCISSORS_TL, 1));
116 OUT_RING(R300_SCISSORS_X_MASK | R300_SCISSORS_Y_MASK);
119 /* Why we allow zero cliprect rendering:
120 * There are some commands in a command buffer that must be submitted
121 * even when there are no cliprects, e.g. DMA buffer discard
122 * or state setting (though state setting could be avoided by
123 * simulating a loss of context).
125 * Now since the cmdbuf interface is so chaotic right now (and is
126 * bound to remain that way for a bit until things settle down),
127 * it is basically impossible to filter out the commands that are
128 * necessary and those that aren't.
130 * So I choose the safe way and don't do any filtering at all;
131 * instead, I simply set up the engine so that all rendering
132 * can't produce any fragments.
135 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, 0);
142 static u8 r300_reg_flags[0x10000 >> 2];
144 void r300_init_reg_flags(struct drm_device *dev)
147 drm_radeon_private_t *dev_priv = dev->dev_private;
149 memset(r300_reg_flags, 0, 0x10000 >> 2);
150 #define ADD_RANGE_MARK(reg, count,mark) \
151 for(i=((reg)>>2);i<((reg)>>2)+(count);i++)\
152 r300_reg_flags[i]|=(mark);
155 #define MARK_CHECK_OFFSET 2
157 #define ADD_RANGE(reg, count) ADD_RANGE_MARK(reg, count, MARK_SAFE)
159 /* these match cmducs() command in r300_driver/r300/r300_cmdbuf.c */
160 ADD_RANGE(R300_SE_VPORT_XSCALE, 6);
161 ADD_RANGE(R300_VAP_CNTL, 1);
162 ADD_RANGE(R300_SE_VTE_CNTL, 2);
163 ADD_RANGE(0x2134, 2);
164 ADD_RANGE(R300_VAP_CNTL_STATUS, 1);
165 ADD_RANGE(R300_VAP_INPUT_CNTL_0, 2);
166 ADD_RANGE(0x21DC, 1);
167 ADD_RANGE(R300_VAP_UNKNOWN_221C, 1);
168 ADD_RANGE(R300_VAP_CLIP_X_0, 4);
169 ADD_RANGE(R300_VAP_PVS_WAITIDLE, 1);
170 ADD_RANGE(R300_VAP_UNKNOWN_2288, 1);
171 ADD_RANGE(R300_VAP_OUTPUT_VTX_FMT_0, 2);
172 ADD_RANGE(R300_VAP_PVS_CNTL_1, 3);
173 ADD_RANGE(R300_GB_ENABLE, 1);
174 ADD_RANGE(R300_GB_MSPOS0, 5);
175 ADD_RANGE(R300_TX_CNTL, 1);
176 ADD_RANGE(R300_TX_ENABLE, 1);
177 ADD_RANGE(0x4200, 4);
178 ADD_RANGE(0x4214, 1);
179 ADD_RANGE(R300_RE_POINTSIZE, 1);
180 ADD_RANGE(0x4230, 3);
181 ADD_RANGE(R300_RE_LINE_CNT, 1);
182 ADD_RANGE(R300_RE_UNK4238, 1);
183 ADD_RANGE(0x4260, 3);
184 ADD_RANGE(R300_RE_SHADE, 4);
185 ADD_RANGE(R300_RE_POLYGON_MODE, 5);
186 ADD_RANGE(R300_RE_ZBIAS_CNTL, 1);
187 ADD_RANGE(R300_RE_ZBIAS_T_FACTOR, 4);
188 ADD_RANGE(R300_RE_OCCLUSION_CNTL, 1);
189 ADD_RANGE(R300_RE_CULL_CNTL, 1);
190 ADD_RANGE(0x42C0, 2);
191 ADD_RANGE(R300_RS_CNTL_0, 2);
193 ADD_RANGE(R300_SC_HYPERZ, 2);
194 ADD_RANGE(0x43E8, 1);
196 ADD_RANGE(0x46A4, 5);
198 ADD_RANGE(R300_RE_FOG_STATE, 1);
199 ADD_RANGE(R300_FOG_COLOR_R, 3);
200 ADD_RANGE(R300_PP_ALPHA_TEST, 2);
201 ADD_RANGE(0x4BD8, 1);
202 ADD_RANGE(R300_PFS_PARAM_0_X, 64);
203 ADD_RANGE(0x4E00, 1);
204 ADD_RANGE(R300_RB3D_CBLEND, 2);
205 ADD_RANGE(R300_RB3D_COLORMASK, 1);
206 ADD_RANGE(R300_RB3D_BLEND_COLOR, 3);
207 ADD_RANGE_MARK(R300_RB3D_COLOROFFSET0, 1, MARK_CHECK_OFFSET); /* check offset */
208 ADD_RANGE(R300_RB3D_COLORPITCH0, 1);
209 ADD_RANGE(0x4E50, 9);
210 ADD_RANGE(0x4E88, 1);
211 ADD_RANGE(0x4EA0, 2);
212 ADD_RANGE(R300_ZB_CNTL, 3);
213 ADD_RANGE(R300_ZB_FORMAT, 4);
214 ADD_RANGE_MARK(R300_ZB_DEPTHOFFSET, 1, MARK_CHECK_OFFSET); /* check offset */
215 ADD_RANGE(R300_ZB_DEPTHPITCH, 1);
216 ADD_RANGE(R300_ZB_DEPTHCLEARVALUE, 1);
217 ADD_RANGE(R300_ZB_ZMASK_OFFSET, 13);
219 ADD_RANGE(R300_TX_FILTER_0, 16);
220 ADD_RANGE(R300_TX_FILTER1_0, 16);
221 ADD_RANGE(R300_TX_SIZE_0, 16);
222 ADD_RANGE(R300_TX_FORMAT_0, 16);
223 ADD_RANGE(R300_TX_PITCH_0, 16);
224 /* Texture offset is dangerous and needs more checking */
225 ADD_RANGE_MARK(R300_TX_OFFSET_0, 16, MARK_CHECK_OFFSET);
226 ADD_RANGE(R300_TX_CHROMA_KEY_0, 16);
227 ADD_RANGE(R300_TX_BORDER_COLOR_0, 16);
229 /* Sporadic registers used as primitives are emitted */
230 ADD_RANGE(R300_ZB_ZCACHE_CTLSTAT, 1);
231 ADD_RANGE(R300_RB3D_DSTCACHE_CTLSTAT, 1);
232 ADD_RANGE(R300_VAP_INPUT_ROUTE_0_0, 8);
233 ADD_RANGE(R300_VAP_INPUT_ROUTE_1_0, 8);
235 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
236 ADD_RANGE(R500_VAP_INDEX_OFFSET, 1);
237 ADD_RANGE(R500_US_CONFIG, 2);
238 ADD_RANGE(R500_US_CODE_ADDR, 3);
239 ADD_RANGE(R500_US_FC_CTRL, 1);
240 ADD_RANGE(R500_RS_IP_0, 16);
241 ADD_RANGE(R500_RS_INST_0, 16);
242 ADD_RANGE(R500_RB3D_COLOR_CLEAR_VALUE_AR, 2);
243 ADD_RANGE(R500_RB3D_CONSTANT_COLOR_AR, 2);
244 ADD_RANGE(R500_ZB_FIFO_SIZE, 2);
246 ADD_RANGE(R300_PFS_CNTL_0, 3);
247 ADD_RANGE(R300_PFS_NODE_0, 4);
248 ADD_RANGE(R300_PFS_TEXI_0, 64);
249 ADD_RANGE(R300_PFS_INSTR0_0, 64);
250 ADD_RANGE(R300_PFS_INSTR1_0, 64);
251 ADD_RANGE(R300_PFS_INSTR2_0, 64);
252 ADD_RANGE(R300_PFS_INSTR3_0, 64);
253 ADD_RANGE(R300_RS_INTERP_0, 8);
254 ADD_RANGE(R300_RS_ROUTE_0, 8);
259 static __inline__ int r300_check_range(unsigned reg, int count)
264 for (i = (reg >> 2); i < (reg >> 2) + count; i++)
265 if (r300_reg_flags[i] != MARK_SAFE)
270 static __inline__ int r300_emit_carefully_checked_packet0(drm_radeon_private_t *
272 drm_radeon_kcmd_buffer_t
274 drm_r300_cmd_header_t
283 sz = header.packet0.count;
284 reg = (header.packet0.reghi << 8) | header.packet0.reglo;
286 if ((sz > 64) || (sz < 0)) {
288 ("Cannot emit more than 64 values at a time (reg=%04x sz=%d)\n",
292 for (i = 0; i < sz; i++) {
293 values[i] = ((int *)cmdbuf->buf)[i];
294 switch (r300_reg_flags[(reg >> 2) + i]) {
297 case MARK_CHECK_OFFSET:
298 if (!radeon_check_offset(dev_priv, (u32) values[i])) {
300 ("Offset failed range check (reg=%04x sz=%d)\n",
306 DRM_ERROR("Register %04x failed check as flag=%02x\n",
307 reg + i * 4, r300_reg_flags[(reg >> 2) + i]);
313 OUT_RING(CP_PACKET0(reg, sz - 1));
314 OUT_RING_TABLE(values, sz);
317 cmdbuf->buf += sz * 4;
318 cmdbuf->bufsz -= sz * 4;
324 * Emits a packet0 setting arbitrary registers.
325 * Called by r300_do_cp_cmdbuf.
327 * Note that checks are performed on contents and addresses of the registers
329 static __inline__ int r300_emit_packet0(drm_radeon_private_t *dev_priv,
330 drm_radeon_kcmd_buffer_t *cmdbuf,
331 drm_r300_cmd_header_t header)
337 sz = header.packet0.count;
338 reg = (header.packet0.reghi << 8) | header.packet0.reglo;
343 if (sz * 4 > cmdbuf->bufsz)
346 if (reg + sz * 4 >= 0x10000) {
347 DRM_ERROR("No such registers in hardware reg=%04x sz=%d\n", reg,
352 if (r300_check_range(reg, sz)) {
353 /* go and check everything */
354 return r300_emit_carefully_checked_packet0(dev_priv, cmdbuf,
357 /* the rest of the data is safe to emit, whatever the values the user passed */
360 OUT_RING(CP_PACKET0(reg, sz - 1));
361 OUT_RING_TABLE((int *)cmdbuf->buf, sz);
364 cmdbuf->buf += sz * 4;
365 cmdbuf->bufsz -= sz * 4;
371 * Uploads user-supplied vertex program instructions or parameters onto
373 * Called by r300_do_cp_cmdbuf.
375 static __inline__ int r300_emit_vpu(drm_radeon_private_t *dev_priv,
376 drm_radeon_kcmd_buffer_t *cmdbuf,
377 drm_r300_cmd_header_t header)
383 sz = header.vpu.count;
384 addr = (header.vpu.adrhi << 8) | header.vpu.adrlo;
388 if (sz * 16 > cmdbuf->bufsz)
391 BEGIN_RING(5 + sz * 4);
392 /* Wait for VAP to come to senses.. */
393 /* there is no need to emit it multiple times, (only once before VAP is programmed,
394 but this optimization is for later */
395 OUT_RING_REG(R300_VAP_PVS_WAITIDLE, 0);
396 OUT_RING_REG(R300_VAP_PVS_UPLOAD_ADDRESS, addr);
397 OUT_RING(CP_PACKET0_TABLE(R300_VAP_PVS_UPLOAD_DATA, sz * 4 - 1));
398 OUT_RING_TABLE((int *)cmdbuf->buf, sz * 4);
402 cmdbuf->buf += sz * 16;
403 cmdbuf->bufsz -= sz * 16;
409 * Emit a clear packet from userspace.
410 * Called by r300_emit_packet3.
412 static __inline__ int r300_emit_clear(drm_radeon_private_t *dev_priv,
413 drm_radeon_kcmd_buffer_t *cmdbuf)
417 if (8 * 4 > cmdbuf->bufsz)
421 OUT_RING(CP_PACKET3(R200_3D_DRAW_IMMD_2, 8));
422 OUT_RING(R300_PRIM_TYPE_POINT | R300_PRIM_WALK_RING |
423 (1 << R300_PRIM_NUM_VERTICES_SHIFT));
424 OUT_RING_TABLE((int *)cmdbuf->buf, 8);
427 cmdbuf->buf += 8 * 4;
428 cmdbuf->bufsz -= 8 * 4;
433 static __inline__ int r300_emit_3d_load_vbpntr(drm_radeon_private_t *dev_priv,
434 drm_radeon_kcmd_buffer_t *cmdbuf,
438 #define MAX_ARRAY_PACKET 64
439 u32 payload[MAX_ARRAY_PACKET];
443 count = (header >> 16) & 0x3fff;
445 if ((count + 1) > MAX_ARRAY_PACKET) {
446 DRM_ERROR("Too large payload in 3D_LOAD_VBPNTR (count=%d)\n",
450 memset(payload, 0, MAX_ARRAY_PACKET * 4);
451 memcpy(payload, cmdbuf->buf + 4, (count + 1) * 4);
453 /* carefully check packet contents */
455 narrays = payload[0];
458 while ((k < narrays) && (i < (count + 1))) {
459 i++; /* skip attribute field */
460 if (!radeon_check_offset(dev_priv, payload[i])) {
462 ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
470 /* have one more to process, they come in pairs */
471 if (!radeon_check_offset(dev_priv, payload[i])) {
473 ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
480 /* do the counts match what we expect ? */
481 if ((k != narrays) || (i != (count + 1))) {
483 ("Malformed 3D_LOAD_VBPNTR packet (k=%d i=%d narrays=%d count+1=%d).\n",
484 k, i, narrays, count + 1);
488 /* all clear, output packet */
490 BEGIN_RING(count + 2);
492 OUT_RING_TABLE(payload, count + 1);
495 cmdbuf->buf += (count + 2) * 4;
496 cmdbuf->bufsz -= (count + 2) * 4;
501 static __inline__ int r300_emit_bitblt_multi(drm_radeon_private_t *dev_priv,
502 drm_radeon_kcmd_buffer_t *cmdbuf)
504 u32 *cmd = (u32 *) cmdbuf->buf;
508 count=(cmd[0]>>16) & 0x3fff;
510 if (cmd[0] & 0x8000) {
513 if (cmd[1] & (RADEON_GMC_SRC_PITCH_OFFSET_CNTL
514 | RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
515 offset = cmd[2] << 10;
516 ret = !radeon_check_offset(dev_priv, offset);
518 DRM_ERROR("Invalid bitblt first offset is %08X\n", offset);
523 if ((cmd[1] & RADEON_GMC_SRC_PITCH_OFFSET_CNTL) &&
524 (cmd[1] & RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
525 offset = cmd[3] << 10;
526 ret = !radeon_check_offset(dev_priv, offset);
528 DRM_ERROR("Invalid bitblt second offset is %08X\n", offset);
537 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
540 cmdbuf->buf += (count+2)*4;
541 cmdbuf->bufsz -= (count+2)*4;
546 static __inline__ int r300_emit_indx_buffer(drm_radeon_private_t *dev_priv,
547 drm_radeon_kcmd_buffer_t *cmdbuf)
549 u32 *cmd = (u32 *) cmdbuf->buf;
553 count=(cmd[0]>>16) & 0x3fff;
555 if ((cmd[1] & 0x8000ffff) != 0x80000810) {
556 DRM_ERROR("Invalid indx_buffer reg address %08X\n", cmd[1]);
559 ret = !radeon_check_offset(dev_priv, cmd[2]);
561 DRM_ERROR("Invalid indx_buffer offset is %08X\n", cmd[2]);
567 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
570 cmdbuf->buf += (count+2)*4;
571 cmdbuf->bufsz -= (count+2)*4;
576 static __inline__ int r300_emit_raw_packet3(drm_radeon_private_t *dev_priv,
577 drm_radeon_kcmd_buffer_t *cmdbuf)
583 if (4 > cmdbuf->bufsz)
586 /* Fixme !! This simply emits a packet without much checking.
587 We need to be smarter. */
589 /* obtain first word - actual packet3 header */
590 header = *(u32 *) cmdbuf->buf;
592 /* Is it packet 3 ? */
593 if ((header >> 30) != 0x3) {
594 DRM_ERROR("Not a packet3 header (0x%08x)\n", header);
598 count = (header >> 16) & 0x3fff;
600 /* Check again now that we know how much data to expect */
601 if ((count + 2) * 4 > cmdbuf->bufsz) {
603 ("Expected packet3 of length %d but have only %d bytes left\n",
604 (count + 2) * 4, cmdbuf->bufsz);
608 /* Is it a packet type we know about ? */
609 switch (header & 0xff00) {
610 case RADEON_3D_LOAD_VBPNTR: /* load vertex array pointers */
611 return r300_emit_3d_load_vbpntr(dev_priv, cmdbuf, header);
613 case RADEON_CNTL_BITBLT_MULTI:
614 return r300_emit_bitblt_multi(dev_priv, cmdbuf);
616 case RADEON_CP_INDX_BUFFER: /* DRAW_INDX_2 without INDX_BUFFER seems to lock up the gpu */
617 return r300_emit_indx_buffer(dev_priv, cmdbuf);
618 case RADEON_CP_3D_DRAW_IMMD_2: /* triggers drawing using in-packet vertex data */
619 case RADEON_CP_3D_DRAW_VBUF_2: /* triggers drawing of vertex buffers setup elsewhere */
620 case RADEON_CP_3D_DRAW_INDX_2: /* triggers drawing using indices to vertex buffer */
621 case RADEON_WAIT_FOR_IDLE:
623 /* these packets are safe */
626 DRM_ERROR("Unknown packet3 header (0x%08x)\n", header);
630 BEGIN_RING(count + 2);
632 OUT_RING_TABLE((int *)(cmdbuf->buf + 4), count + 1);
635 cmdbuf->buf += (count + 2) * 4;
636 cmdbuf->bufsz -= (count + 2) * 4;
642 * Emit a rendering packet3 from userspace.
643 * Called by r300_do_cp_cmdbuf.
645 static __inline__ int r300_emit_packet3(drm_radeon_private_t *dev_priv,
646 drm_radeon_kcmd_buffer_t *cmdbuf,
647 drm_r300_cmd_header_t header)
651 char *orig_buf = cmdbuf->buf;
652 int orig_bufsz = cmdbuf->bufsz;
654 /* This is a do-while-loop so that we run the interior at least once,
655 * even if cmdbuf->nbox is 0. Compare r300_emit_cliprects for rationale.
659 if (cmdbuf->nbox > R300_SIMULTANEOUS_CLIPRECTS) {
660 ret = r300_emit_cliprects(dev_priv, cmdbuf, n);
664 cmdbuf->buf = orig_buf;
665 cmdbuf->bufsz = orig_bufsz;
668 switch (header.packet3.packet) {
669 case R300_CMD_PACKET3_CLEAR:
670 DRM_DEBUG("R300_CMD_PACKET3_CLEAR\n");
671 ret = r300_emit_clear(dev_priv, cmdbuf);
673 DRM_ERROR("r300_emit_clear failed\n");
678 case R300_CMD_PACKET3_RAW:
679 DRM_DEBUG("R300_CMD_PACKET3_RAW\n");
680 ret = r300_emit_raw_packet3(dev_priv, cmdbuf);
682 DRM_ERROR("r300_emit_raw_packet3 failed\n");
688 DRM_ERROR("bad packet3 type %i at %p\n",
689 header.packet3.packet,
690 cmdbuf->buf - sizeof(header));
694 n += R300_SIMULTANEOUS_CLIPRECTS;
695 } while (n < cmdbuf->nbox);
700 /* Some of the R300 chips seem to be extremely touchy about the two registers
701 * that are configured in r300_pacify.
702 * Among the worst offenders seems to be the R300 ND (0x4E44): When userspace
703 * sends a command buffer that contains only state setting commands and a
704 * vertex program/parameter upload sequence, this will eventually lead to a
705 * lockup, unless the sequence is bracketed by calls to r300_pacify.
706 * So we should take great care to *always* call r300_pacify before
707 * *anything* 3D related, and again afterwards. This is what the
708 * call bracket in r300_do_cp_cmdbuf is for.
712 * Emit the sequence to pacify R300.
714 static __inline__ void r300_pacify(drm_radeon_private_t *dev_priv)
719 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
720 OUT_RING(R300_RB3D_DSTCACHE_UNKNOWN_0A);
721 OUT_RING(CP_PACKET0(R300_ZB_ZCACHE_CTLSTAT, 0));
722 OUT_RING(R300_ZB_ZCACHE_CTLSTAT_ZC_FLUSH_FLUSH_AND_FREE|
723 R300_ZB_ZCACHE_CTLSTAT_ZC_FREE_FREE);
724 OUT_RING(CP_PACKET3(RADEON_CP_NOP, 0));
730 * Called by r300_do_cp_cmdbuf to update the internal buffer age and state.
731 * The actual age emit is done by r300_do_cp_cmdbuf, which is why you must
732 * be careful about how this function is called.
734 static void r300_discard_buffer(struct drm_device * dev, struct drm_buf * buf)
736 drm_radeon_private_t *dev_priv = dev->dev_private;
737 drm_radeon_buf_priv_t *buf_priv = buf->dev_private;
739 buf_priv->age = ++dev_priv->sarea_priv->last_dispatch;
744 static void r300_cmd_wait(drm_radeon_private_t * dev_priv,
745 drm_r300_cmd_header_t header)
750 if (!header.wait.flags)
755 switch(header.wait.flags) {
757 wait_until = RADEON_WAIT_2D_IDLE;
760 wait_until = RADEON_WAIT_3D_IDLE;
762 case R300_NEW_WAIT_2D_3D:
763 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_3D_IDLE;
765 case R300_NEW_WAIT_2D_2D_CLEAN:
766 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_2D_IDLECLEAN;
768 case R300_NEW_WAIT_3D_3D_CLEAN:
769 wait_until = RADEON_WAIT_3D_IDLE|RADEON_WAIT_3D_IDLECLEAN;
771 case R300_NEW_WAIT_2D_2D_CLEAN_3D_3D_CLEAN:
772 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_2D_IDLECLEAN;
773 wait_until |= RADEON_WAIT_3D_IDLE|RADEON_WAIT_3D_IDLECLEAN;
780 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
781 OUT_RING(wait_until);
785 static int r300_scratch(drm_radeon_private_t *dev_priv,
786 drm_radeon_kcmd_buffer_t *cmdbuf,
787 drm_r300_cmd_header_t header)
790 u32 i, buf_idx, h_pending;
794 (sizeof(u64) + header.scratch.n_bufs * sizeof(buf_idx))) {
798 if (header.scratch.reg >= 5) {
802 dev_priv->scratch_ages[header.scratch.reg]++;
804 ref_age_base = (u32 *)(unsigned long)*((uint64_t *)cmdbuf->buf);
806 cmdbuf->buf += sizeof(u64);
807 cmdbuf->bufsz -= sizeof(u64);
809 for (i=0; i < header.scratch.n_bufs; i++) {
810 buf_idx = *(u32 *)cmdbuf->buf;
811 buf_idx *= 2; /* 8 bytes per buf */
813 if (DRM_COPY_TO_USER(ref_age_base + buf_idx, &dev_priv->scratch_ages[header.scratch.reg], sizeof(u32))) {
817 if (DRM_COPY_FROM_USER(&h_pending, ref_age_base + buf_idx + 1, sizeof(u32))) {
821 if (h_pending == 0) {
827 if (DRM_COPY_TO_USER(ref_age_base + buf_idx + 1, &h_pending, sizeof(u32))) {
831 cmdbuf->buf += sizeof(buf_idx);
832 cmdbuf->bufsz -= sizeof(buf_idx);
836 OUT_RING( CP_PACKET0( RADEON_SCRATCH_REG0 + header.scratch.reg * 4, 0 ) );
837 OUT_RING( dev_priv->scratch_ages[header.scratch.reg] );
844 * Uploads user-supplied vertex program instructions or parameters onto
846 * Called by r300_do_cp_cmdbuf.
848 static inline int r300_emit_r500fp(drm_radeon_private_t *dev_priv,
849 drm_radeon_kcmd_buffer_t *cmdbuf,
850 drm_r300_cmd_header_t header)
859 sz = header.r500fp.count;
860 /* address is 9 bits 0 - 8, bit 1 of flags is part of address */
861 addr = ((header.r500fp.adrhi_flags & 1) << 8) | header.r500fp.adrlo;
863 type = !!(header.r500fp.adrhi_flags & R500FP_CONSTANT_TYPE);
864 clamp = !!(header.r500fp.adrhi_flags & R500FP_CONSTANT_CLAMP);
866 addr |= (type << 16);
867 addr |= (clamp << 17);
869 stride = type ? 4 : 6;
871 DRM_DEBUG("r500fp %d %d type: %d\n", sz, addr, type);
874 if (sz * stride * 4 > cmdbuf->bufsz)
877 BEGIN_RING(3 + sz * stride);
878 OUT_RING_REG(R500_GA_US_VECTOR_INDEX, addr);
879 OUT_RING(CP_PACKET0_TABLE(R500_GA_US_VECTOR_DATA, sz * stride - 1));
880 OUT_RING_TABLE((int *)cmdbuf->buf, sz * stride);
884 cmdbuf->buf += sz * stride * 4;
885 cmdbuf->bufsz -= sz * stride * 4;
892 * Parses and validates a user-supplied command buffer and emits appropriate
893 * commands on the DMA ring buffer.
894 * Called by the ioctl handler function radeon_cp_cmdbuf.
896 int r300_do_cp_cmdbuf(struct drm_device *dev,
897 struct drm_file *file_priv,
898 drm_radeon_kcmd_buffer_t *cmdbuf)
900 drm_radeon_private_t *dev_priv = dev->dev_private;
901 struct drm_device_dma *dma = dev->dma;
902 struct drm_buf *buf = NULL;
903 int emit_dispatch_age = 0;
908 /* See the comment above r300_emit_begin3d for why this call must be here,
909 * and what the cleanup gotos are for. */
910 r300_pacify(dev_priv);
912 if (cmdbuf->nbox <= R300_SIMULTANEOUS_CLIPRECTS) {
913 ret = r300_emit_cliprects(dev_priv, cmdbuf, 0);
918 while (cmdbuf->bufsz >= sizeof(drm_r300_cmd_header_t)) {
920 drm_r300_cmd_header_t header;
922 header.u = *(unsigned int *)cmdbuf->buf;
924 cmdbuf->buf += sizeof(header);
925 cmdbuf->bufsz -= sizeof(header);
927 switch (header.header.cmd_type) {
928 case R300_CMD_PACKET0:
929 DRM_DEBUG("R300_CMD_PACKET0\n");
930 ret = r300_emit_packet0(dev_priv, cmdbuf, header);
932 DRM_ERROR("r300_emit_packet0 failed\n");
938 DRM_DEBUG("R300_CMD_VPU\n");
939 ret = r300_emit_vpu(dev_priv, cmdbuf, header);
941 DRM_ERROR("r300_emit_vpu failed\n");
946 case R300_CMD_PACKET3:
947 DRM_DEBUG("R300_CMD_PACKET3\n");
948 ret = r300_emit_packet3(dev_priv, cmdbuf, header);
950 DRM_ERROR("r300_emit_packet3 failed\n");
956 DRM_DEBUG("R300_CMD_END3D\n");
958 Ideally userspace driver should not need to issue this call,
959 i.e. the drm driver should issue it automatically and prevent
962 In practice, we do not understand why this call is needed and what
963 it does (except for some vague guesses that it has to do with cache
964 coherence) and so the user space driver does it.
966 Once we are sure which uses prevent lockups the code could be moved
967 into the kernel and the userspace driver will not
968 need to use this command.
970 Note that issuing this command does not hurt anything
971 except, possibly, performance */
972 r300_pacify(dev_priv);
975 case R300_CMD_CP_DELAY:
976 /* simple enough, we can do it here */
977 DRM_DEBUG("R300_CMD_CP_DELAY\n");
982 BEGIN_RING(header.delay.count);
983 for (i = 0; i < header.delay.count; i++)
984 OUT_RING(RADEON_CP_PACKET2);
989 case R300_CMD_DMA_DISCARD:
990 DRM_DEBUG("RADEON_CMD_DMA_DISCARD\n");
991 idx = header.dma.buf_idx;
992 if (idx < 0 || idx >= dma->buf_count) {
993 DRM_ERROR("buffer index %d (of %d max)\n",
994 idx, dma->buf_count - 1);
999 buf = dma->buflist[idx];
1000 if (buf->file_priv != file_priv || buf->pending) {
1001 DRM_ERROR("bad buffer %p %p %d\n",
1002 buf->file_priv, file_priv,
1008 emit_dispatch_age = 1;
1009 r300_discard_buffer(dev, buf);
1013 DRM_DEBUG("R300_CMD_WAIT\n");
1014 r300_cmd_wait(dev_priv, header);
1017 case R300_CMD_SCRATCH:
1018 DRM_DEBUG("R300_CMD_SCRATCH\n");
1019 ret = r300_scratch(dev_priv, cmdbuf, header);
1021 DRM_ERROR("r300_scratch failed\n");
1026 case R300_CMD_R500FP:
1027 if ((dev_priv->flags & RADEON_FAMILY_MASK) < CHIP_RV515) {
1028 DRM_ERROR("Calling r500 command on r300 card\n");
1032 DRM_DEBUG("R300_CMD_R500FP\n");
1033 ret = r300_emit_r500fp(dev_priv, cmdbuf, header);
1035 DRM_ERROR("r300_emit_r500fp failed\n");
1040 DRM_ERROR("bad cmd_type %i at %p\n",
1041 header.header.cmd_type,
1042 cmdbuf->buf - sizeof(header));
1051 r300_pacify(dev_priv);
1053 /* We emit the vertex buffer age here, outside the pacifier "brackets"
1055 * (1) This may coalesce multiple age emissions into a single one and
1056 * (2) more importantly, some chips lock up hard when scratch registers
1057 * are written inside the pacifier bracket.
1059 if (emit_dispatch_age) {
1062 /* Emit the vertex buffer age */
1064 RADEON_DISPATCH_AGE(dev_priv->sarea_priv->last_dispatch);