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 "drm_buffer.h"
37 #include "radeon_drm.h"
38 #include "radeon_drv.h"
41 #include <asm/unaligned.h>
43 #define R300_SIMULTANEOUS_CLIPRECTS 4
45 /* Values for R300_RE_CLIPRECT_CNTL depending on the number of cliprects
47 static const int r300_cliprect_cntl[4] = {
55 * Emit up to R300_SIMULTANEOUS_CLIPRECTS cliprects from the given command
56 * buffer, starting with index n.
58 static int r300_emit_cliprects(drm_radeon_private_t *dev_priv,
59 drm_radeon_kcmd_buffer_t *cmdbuf, int n)
61 struct drm_clip_rect box;
66 nr = cmdbuf->nbox - n;
67 if (nr > R300_SIMULTANEOUS_CLIPRECTS)
68 nr = R300_SIMULTANEOUS_CLIPRECTS;
70 DRM_DEBUG("%i cliprects\n", nr);
73 BEGIN_RING(6 + nr * 2);
74 OUT_RING(CP_PACKET0(R300_RE_CLIPRECT_TL_0, nr * 2 - 1));
76 for (i = 0; i < nr; ++i) {
77 if (DRM_COPY_FROM_USER_UNCHECKED
78 (&box, &cmdbuf->boxes[n + i], sizeof(box))) {
79 DRM_ERROR("copy cliprect faulted\n");
83 box.x2--; /* Hardware expects inclusive bottom-right corner */
86 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
96 box.x1 = (box.x1 + R300_CLIPRECT_OFFSET) &
98 box.y1 = (box.y1 + R300_CLIPRECT_OFFSET) &
100 box.x2 = (box.x2 + R300_CLIPRECT_OFFSET) &
102 box.y2 = (box.y2 + R300_CLIPRECT_OFFSET) &
106 OUT_RING((box.x1 << R300_CLIPRECT_X_SHIFT) |
107 (box.y1 << R300_CLIPRECT_Y_SHIFT));
108 OUT_RING((box.x2 << R300_CLIPRECT_X_SHIFT) |
109 (box.y2 << R300_CLIPRECT_Y_SHIFT));
113 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, r300_cliprect_cntl[nr - 1]);
115 /* TODO/SECURITY: Force scissors to a safe value, otherwise the
116 * client might be able to trample over memory.
117 * The impact should be very limited, but I'd rather be safe than
120 OUT_RING(CP_PACKET0(R300_RE_SCISSORS_TL, 1));
122 OUT_RING(R300_SCISSORS_X_MASK | R300_SCISSORS_Y_MASK);
125 /* Why we allow zero cliprect rendering:
126 * There are some commands in a command buffer that must be submitted
127 * even when there are no cliprects, e.g. DMA buffer discard
128 * or state setting (though state setting could be avoided by
129 * simulating a loss of context).
131 * Now since the cmdbuf interface is so chaotic right now (and is
132 * bound to remain that way for a bit until things settle down),
133 * it is basically impossible to filter out the commands that are
134 * necessary and those that aren't.
136 * So I choose the safe way and don't do any filtering at all;
137 * instead, I simply set up the engine so that all rendering
138 * can't produce any fragments.
141 OUT_RING_REG(R300_RE_CLIPRECT_CNTL, 0);
145 /* flus cache and wait idle clean after cliprect change */
147 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
148 OUT_RING(R300_RB3D_DC_FLUSH);
151 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
152 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
155 dev_priv->track_flush |= RADEON_FLUSH_EMITED;
160 static u8 r300_reg_flags[0x10000 >> 2];
162 void r300_init_reg_flags(struct drm_device *dev)
165 drm_radeon_private_t *dev_priv = dev->dev_private;
167 memset(r300_reg_flags, 0, 0x10000 >> 2);
168 #define ADD_RANGE_MARK(reg, count,mark) \
169 for(i=((reg)>>2);i<((reg)>>2)+(count);i++)\
170 r300_reg_flags[i]|=(mark);
173 #define MARK_CHECK_OFFSET 2
175 #define ADD_RANGE(reg, count) ADD_RANGE_MARK(reg, count, MARK_SAFE)
177 /* these match cmducs() command in r300_driver/r300/r300_cmdbuf.c */
178 ADD_RANGE(R300_SE_VPORT_XSCALE, 6);
179 ADD_RANGE(R300_VAP_CNTL, 1);
180 ADD_RANGE(R300_SE_VTE_CNTL, 2);
181 ADD_RANGE(0x2134, 2);
182 ADD_RANGE(R300_VAP_CNTL_STATUS, 1);
183 ADD_RANGE(R300_VAP_INPUT_CNTL_0, 2);
184 ADD_RANGE(0x21DC, 1);
185 ADD_RANGE(R300_VAP_UNKNOWN_221C, 1);
186 ADD_RANGE(R300_VAP_CLIP_X_0, 4);
187 ADD_RANGE(R300_VAP_PVS_STATE_FLUSH_REG, 1);
188 ADD_RANGE(R300_VAP_UNKNOWN_2288, 1);
189 ADD_RANGE(R300_VAP_OUTPUT_VTX_FMT_0, 2);
190 ADD_RANGE(R300_VAP_PVS_CNTL_1, 3);
191 ADD_RANGE(R300_GB_ENABLE, 1);
192 ADD_RANGE(R300_GB_MSPOS0, 5);
193 ADD_RANGE(R300_TX_INVALTAGS, 1);
194 ADD_RANGE(R300_TX_ENABLE, 1);
195 ADD_RANGE(0x4200, 4);
196 ADD_RANGE(0x4214, 1);
197 ADD_RANGE(R300_RE_POINTSIZE, 1);
198 ADD_RANGE(0x4230, 3);
199 ADD_RANGE(R300_RE_LINE_CNT, 1);
200 ADD_RANGE(R300_RE_UNK4238, 1);
201 ADD_RANGE(0x4260, 3);
202 ADD_RANGE(R300_RE_SHADE, 4);
203 ADD_RANGE(R300_RE_POLYGON_MODE, 5);
204 ADD_RANGE(R300_RE_ZBIAS_CNTL, 1);
205 ADD_RANGE(R300_RE_ZBIAS_T_FACTOR, 4);
206 ADD_RANGE(R300_RE_OCCLUSION_CNTL, 1);
207 ADD_RANGE(R300_RE_CULL_CNTL, 1);
208 ADD_RANGE(0x42C0, 2);
209 ADD_RANGE(R300_RS_CNTL_0, 2);
211 ADD_RANGE(R300_SU_REG_DEST, 1);
212 if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV530)
213 ADD_RANGE(RV530_FG_ZBREG_DEST, 1);
215 ADD_RANGE(R300_SC_HYPERZ, 2);
216 ADD_RANGE(0x43E8, 1);
218 ADD_RANGE(0x46A4, 5);
220 ADD_RANGE(R300_RE_FOG_STATE, 1);
221 ADD_RANGE(R300_FOG_COLOR_R, 3);
222 ADD_RANGE(R300_PP_ALPHA_TEST, 2);
223 ADD_RANGE(0x4BD8, 1);
224 ADD_RANGE(R300_PFS_PARAM_0_X, 64);
225 ADD_RANGE(0x4E00, 1);
226 ADD_RANGE(R300_RB3D_CBLEND, 2);
227 ADD_RANGE(R300_RB3D_COLORMASK, 1);
228 ADD_RANGE(R300_RB3D_BLEND_COLOR, 3);
229 ADD_RANGE_MARK(R300_RB3D_COLOROFFSET0, 1, MARK_CHECK_OFFSET); /* check offset */
230 ADD_RANGE(R300_RB3D_COLORPITCH0, 1);
231 ADD_RANGE(0x4E50, 9);
232 ADD_RANGE(0x4E88, 1);
233 ADD_RANGE(0x4EA0, 2);
234 ADD_RANGE(R300_ZB_CNTL, 3);
235 ADD_RANGE(R300_ZB_FORMAT, 4);
236 ADD_RANGE_MARK(R300_ZB_DEPTHOFFSET, 1, MARK_CHECK_OFFSET); /* check offset */
237 ADD_RANGE(R300_ZB_DEPTHPITCH, 1);
238 ADD_RANGE(R300_ZB_DEPTHCLEARVALUE, 1);
239 ADD_RANGE(R300_ZB_ZMASK_OFFSET, 13);
240 ADD_RANGE(R300_ZB_ZPASS_DATA, 2); /* ZB_ZPASS_DATA, ZB_ZPASS_ADDR */
242 ADD_RANGE(R300_TX_FILTER_0, 16);
243 ADD_RANGE(R300_TX_FILTER1_0, 16);
244 ADD_RANGE(R300_TX_SIZE_0, 16);
245 ADD_RANGE(R300_TX_FORMAT_0, 16);
246 ADD_RANGE(R300_TX_PITCH_0, 16);
247 /* Texture offset is dangerous and needs more checking */
248 ADD_RANGE_MARK(R300_TX_OFFSET_0, 16, MARK_CHECK_OFFSET);
249 ADD_RANGE(R300_TX_CHROMA_KEY_0, 16);
250 ADD_RANGE(R300_TX_BORDER_COLOR_0, 16);
252 /* Sporadic registers used as primitives are emitted */
253 ADD_RANGE(R300_ZB_ZCACHE_CTLSTAT, 1);
254 ADD_RANGE(R300_RB3D_DSTCACHE_CTLSTAT, 1);
255 ADD_RANGE(R300_VAP_INPUT_ROUTE_0_0, 8);
256 ADD_RANGE(R300_VAP_INPUT_ROUTE_1_0, 8);
258 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
259 ADD_RANGE(R500_VAP_INDEX_OFFSET, 1);
260 ADD_RANGE(R500_US_CONFIG, 2);
261 ADD_RANGE(R500_US_CODE_ADDR, 3);
262 ADD_RANGE(R500_US_FC_CTRL, 1);
263 ADD_RANGE(R500_RS_IP_0, 16);
264 ADD_RANGE(R500_RS_INST_0, 16);
265 ADD_RANGE(R500_RB3D_COLOR_CLEAR_VALUE_AR, 2);
266 ADD_RANGE(R500_RB3D_CONSTANT_COLOR_AR, 2);
267 ADD_RANGE(R500_ZB_FIFO_SIZE, 2);
269 ADD_RANGE(R300_PFS_CNTL_0, 3);
270 ADD_RANGE(R300_PFS_NODE_0, 4);
271 ADD_RANGE(R300_PFS_TEXI_0, 64);
272 ADD_RANGE(R300_PFS_INSTR0_0, 64);
273 ADD_RANGE(R300_PFS_INSTR1_0, 64);
274 ADD_RANGE(R300_PFS_INSTR2_0, 64);
275 ADD_RANGE(R300_PFS_INSTR3_0, 64);
276 ADD_RANGE(R300_RS_INTERP_0, 8);
277 ADD_RANGE(R300_RS_ROUTE_0, 8);
282 static __inline__ int r300_check_range(unsigned reg, int count)
287 for (i = (reg >> 2); i < (reg >> 2) + count; i++)
288 if (r300_reg_flags[i] != MARK_SAFE)
293 static __inline__ int r300_emit_carefully_checked_packet0(drm_radeon_private_t *
295 drm_radeon_kcmd_buffer_t
297 drm_r300_cmd_header_t
306 sz = header.packet0.count;
307 reg = (header.packet0.reghi << 8) | header.packet0.reglo;
309 if ((sz > 64) || (sz < 0)) {
310 DRM_ERROR("Cannot emit more than 64 values at a time (reg=%04x sz=%d)\n",
315 for (i = 0; i < sz; i++) {
316 switch (r300_reg_flags[(reg >> 2) + i]) {
319 case MARK_CHECK_OFFSET:
320 value = drm_buffer_pointer_to_dword(cmdbuf->buffer, i);
321 if (!radeon_check_offset(dev_priv, *value)) {
322 DRM_ERROR("Offset failed range check (reg=%04x sz=%d)\n",
328 DRM_ERROR("Register %04x failed check as flag=%02x\n",
329 reg + i * 4, r300_reg_flags[(reg >> 2) + i]);
335 OUT_RING(CP_PACKET0(reg, sz - 1));
336 OUT_RING_DRM_BUFFER(cmdbuf->buffer, sz);
343 * Emits a packet0 setting arbitrary registers.
344 * Called by r300_do_cp_cmdbuf.
346 * Note that checks are performed on contents and addresses of the registers
348 static __inline__ int r300_emit_packet0(drm_radeon_private_t *dev_priv,
349 drm_radeon_kcmd_buffer_t *cmdbuf,
350 drm_r300_cmd_header_t header)
356 sz = header.packet0.count;
357 reg = (header.packet0.reghi << 8) | header.packet0.reglo;
362 if (sz * 4 > drm_buffer_unprocessed(cmdbuf->buffer))
365 if (reg + sz * 4 >= 0x10000) {
366 DRM_ERROR("No such registers in hardware reg=%04x sz=%d\n", reg,
371 if (r300_check_range(reg, sz)) {
372 /* go and check everything */
373 return r300_emit_carefully_checked_packet0(dev_priv, cmdbuf,
376 /* the rest of the data is safe to emit, whatever the values the user passed */
379 OUT_RING(CP_PACKET0(reg, sz - 1));
380 OUT_RING_DRM_BUFFER(cmdbuf->buffer, sz);
387 * Uploads user-supplied vertex program instructions or parameters onto
389 * Called by r300_do_cp_cmdbuf.
391 static __inline__ int r300_emit_vpu(drm_radeon_private_t *dev_priv,
392 drm_radeon_kcmd_buffer_t *cmdbuf,
393 drm_r300_cmd_header_t header)
399 sz = header.vpu.count;
400 addr = (header.vpu.adrhi << 8) | header.vpu.adrlo;
404 if (sz * 16 > drm_buffer_unprocessed(cmdbuf->buffer))
407 /* VAP is very sensitive so we purge cache before we program it
408 * and we also flush its state before & after */
410 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
411 OUT_RING(R300_RB3D_DC_FLUSH);
412 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
413 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
414 OUT_RING(CP_PACKET0(R300_VAP_PVS_STATE_FLUSH_REG, 0));
418 dev_priv->track_flush |= RADEON_FLUSH_EMITED;
420 BEGIN_RING(3 + sz * 4);
421 OUT_RING_REG(R300_VAP_PVS_UPLOAD_ADDRESS, addr);
422 OUT_RING(CP_PACKET0_TABLE(R300_VAP_PVS_UPLOAD_DATA, sz * 4 - 1));
423 OUT_RING_DRM_BUFFER(cmdbuf->buffer, sz * 4);
427 OUT_RING(CP_PACKET0(R300_VAP_PVS_STATE_FLUSH_REG, 0));
435 * Emit a clear packet from userspace.
436 * Called by r300_emit_packet3.
438 static __inline__ int r300_emit_clear(drm_radeon_private_t *dev_priv,
439 drm_radeon_kcmd_buffer_t *cmdbuf)
443 if (8 * 4 > drm_buffer_unprocessed(cmdbuf->buffer))
447 OUT_RING(CP_PACKET3(R200_3D_DRAW_IMMD_2, 8));
448 OUT_RING(R300_PRIM_TYPE_POINT | R300_PRIM_WALK_RING |
449 (1 << R300_PRIM_NUM_VERTICES_SHIFT));
450 OUT_RING_DRM_BUFFER(cmdbuf->buffer, 8);
454 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
455 OUT_RING(R300_RB3D_DC_FLUSH);
456 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
457 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
460 dev_priv->track_flush |= RADEON_FLUSH_EMITED;
465 static __inline__ int r300_emit_3d_load_vbpntr(drm_radeon_private_t *dev_priv,
466 drm_radeon_kcmd_buffer_t *cmdbuf,
470 #define MAX_ARRAY_PACKET 64
475 count = (header & RADEON_CP_PACKET_COUNT_MASK) >> 16;
477 if ((count + 1) > MAX_ARRAY_PACKET) {
478 DRM_ERROR("Too large payload in 3D_LOAD_VBPNTR (count=%d)\n",
482 /* carefully check packet contents */
484 /* We have already read the header so advance the buffer. */
485 drm_buffer_advance(cmdbuf->buffer, 4);
487 narrays = *(u32 *)drm_buffer_pointer_to_dword(cmdbuf->buffer, 0);
490 while ((k < narrays) && (i < (count + 1))) {
491 i++; /* skip attribute field */
492 data = drm_buffer_pointer_to_dword(cmdbuf->buffer, i);
493 if (!radeon_check_offset(dev_priv, *data)) {
495 ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
503 /* have one more to process, they come in pairs */
504 data = drm_buffer_pointer_to_dword(cmdbuf->buffer, i);
505 if (!radeon_check_offset(dev_priv, *data)) {
507 ("Offset failed range check (k=%d i=%d) while processing 3D_LOAD_VBPNTR packet.\n",
514 /* do the counts match what we expect ? */
515 if ((k != narrays) || (i != (count + 1))) {
517 ("Malformed 3D_LOAD_VBPNTR packet (k=%d i=%d narrays=%d count+1=%d).\n",
518 k, i, narrays, count + 1);
522 /* all clear, output packet */
524 BEGIN_RING(count + 2);
526 OUT_RING_DRM_BUFFER(cmdbuf->buffer, count + 1);
532 static __inline__ int r300_emit_bitblt_multi(drm_radeon_private_t *dev_priv,
533 drm_radeon_kcmd_buffer_t *cmdbuf)
535 u32 *cmd = drm_buffer_pointer_to_dword(cmdbuf->buffer, 0);
540 count = (*cmd & RADEON_CP_PACKET_COUNT_MASK) >> 16;
544 u32 *cmd1 = drm_buffer_pointer_to_dword(cmdbuf->buffer, 1);
545 if (*cmd1 & (RADEON_GMC_SRC_PITCH_OFFSET_CNTL
546 | RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
548 u32 *cmd2 = drm_buffer_pointer_to_dword(cmdbuf->buffer, 2);
549 offset = *cmd2 << 10;
550 ret = !radeon_check_offset(dev_priv, offset);
552 DRM_ERROR("Invalid bitblt first offset is %08X\n", offset);
557 if ((*cmd1 & RADEON_GMC_SRC_PITCH_OFFSET_CNTL) &&
558 (*cmd1 & RADEON_GMC_DST_PITCH_OFFSET_CNTL)) {
559 u32 *cmd3 = drm_buffer_pointer_to_dword(cmdbuf->buffer, 3);
560 offset = *cmd3 << 10;
561 ret = !radeon_check_offset(dev_priv, offset);
563 DRM_ERROR("Invalid bitblt second offset is %08X\n", offset);
571 OUT_RING_DRM_BUFFER(cmdbuf->buffer, count + 2);
577 static __inline__ int r300_emit_draw_indx_2(drm_radeon_private_t *dev_priv,
578 drm_radeon_kcmd_buffer_t *cmdbuf)
580 u32 *cmd = drm_buffer_pointer_to_dword(cmdbuf->buffer, 0);
581 u32 *cmd1 = drm_buffer_pointer_to_dword(cmdbuf->buffer, 1);
586 count = (*cmd & RADEON_CP_PACKET_COUNT_MASK) >> 16;
588 expected_count = *cmd1 >> 16;
589 if (!(*cmd1 & R300_VAP_VF_CNTL__INDEX_SIZE_32bit))
590 expected_count = (expected_count+1)/2;
592 if (count && count != expected_count) {
593 DRM_ERROR("3D_DRAW_INDX_2: packet size %i, expected %i\n",
594 count, expected_count);
599 OUT_RING_DRM_BUFFER(cmdbuf->buffer, count + 2);
603 drm_r300_cmd_header_t stack_header, *header;
604 u32 *cmd1, *cmd2, *cmd3;
606 if (drm_buffer_unprocessed(cmdbuf->buffer)
607 < 4*4 + sizeof(stack_header)) {
608 DRM_ERROR("3D_DRAW_INDX_2: expect subsequent INDX_BUFFER, but stream is too short.\n");
612 header = drm_buffer_read_object(cmdbuf->buffer,
613 sizeof(stack_header), &stack_header);
615 cmd = drm_buffer_pointer_to_dword(cmdbuf->buffer, 0);
616 cmd1 = drm_buffer_pointer_to_dword(cmdbuf->buffer, 1);
617 cmd2 = drm_buffer_pointer_to_dword(cmdbuf->buffer, 2);
618 cmd3 = drm_buffer_pointer_to_dword(cmdbuf->buffer, 3);
620 if (header->header.cmd_type != R300_CMD_PACKET3 ||
621 header->packet3.packet != R300_CMD_PACKET3_RAW ||
622 *cmd != CP_PACKET3(RADEON_CP_INDX_BUFFER, 2)) {
623 DRM_ERROR("3D_DRAW_INDX_2: expect subsequent INDX_BUFFER.\n");
627 if ((*cmd1 & 0x8000ffff) != 0x80000810) {
628 DRM_ERROR("Invalid indx_buffer reg address %08X\n",
632 if (!radeon_check_offset(dev_priv, *cmd2)) {
633 DRM_ERROR("Invalid indx_buffer offset is %08X\n",
637 if (*cmd3 != expected_count) {
638 DRM_ERROR("INDX_BUFFER: buffer size %i, expected %i\n",
639 *cmd3, expected_count);
644 OUT_RING_DRM_BUFFER(cmdbuf->buffer, 4);
651 static __inline__ int r300_emit_raw_packet3(drm_radeon_private_t *dev_priv,
652 drm_radeon_kcmd_buffer_t *cmdbuf)
658 if (4 > drm_buffer_unprocessed(cmdbuf->buffer))
661 /* Fixme !! This simply emits a packet without much checking.
662 We need to be smarter. */
664 /* obtain first word - actual packet3 header */
665 header = drm_buffer_pointer_to_dword(cmdbuf->buffer, 0);
667 /* Is it packet 3 ? */
668 if ((*header >> 30) != 0x3) {
669 DRM_ERROR("Not a packet3 header (0x%08x)\n", *header);
673 count = (*header >> 16) & 0x3fff;
675 /* Check again now that we know how much data to expect */
676 if ((count + 2) * 4 > drm_buffer_unprocessed(cmdbuf->buffer)) {
678 ("Expected packet3 of length %d but have only %d bytes left\n",
679 (count + 2) * 4, drm_buffer_unprocessed(cmdbuf->buffer));
683 /* Is it a packet type we know about ? */
684 switch (*header & 0xff00) {
685 case RADEON_3D_LOAD_VBPNTR: /* load vertex array pointers */
686 return r300_emit_3d_load_vbpntr(dev_priv, cmdbuf, *header);
688 case RADEON_CNTL_BITBLT_MULTI:
689 return r300_emit_bitblt_multi(dev_priv, cmdbuf);
691 case RADEON_CP_INDX_BUFFER:
692 DRM_ERROR("packet3 INDX_BUFFER without preceding 3D_DRAW_INDX_2 is illegal.\n");
694 case RADEON_CP_3D_DRAW_IMMD_2:
695 /* triggers drawing using in-packet vertex data */
696 case RADEON_CP_3D_DRAW_VBUF_2:
697 /* triggers drawing of vertex buffers setup elsewhere */
698 dev_priv->track_flush &= ~(RADEON_FLUSH_EMITED |
699 RADEON_PURGE_EMITED);
701 case RADEON_CP_3D_DRAW_INDX_2:
702 /* triggers drawing using indices to vertex buffer */
703 /* whenever we send vertex we clear flush & purge */
704 dev_priv->track_flush &= ~(RADEON_FLUSH_EMITED |
705 RADEON_PURGE_EMITED);
706 return r300_emit_draw_indx_2(dev_priv, cmdbuf);
707 case RADEON_WAIT_FOR_IDLE:
709 /* these packets are safe */
712 DRM_ERROR("Unknown packet3 header (0x%08x)\n", *header);
716 BEGIN_RING(count + 2);
717 OUT_RING_DRM_BUFFER(cmdbuf->buffer, count + 2);
724 * Emit a rendering packet3 from userspace.
725 * Called by r300_do_cp_cmdbuf.
727 static __inline__ int r300_emit_packet3(drm_radeon_private_t *dev_priv,
728 drm_radeon_kcmd_buffer_t *cmdbuf,
729 drm_r300_cmd_header_t header)
733 int orig_iter = cmdbuf->buffer->iterator;
735 /* This is a do-while-loop so that we run the interior at least once,
736 * even if cmdbuf->nbox is 0. Compare r300_emit_cliprects for rationale.
740 if (cmdbuf->nbox > R300_SIMULTANEOUS_CLIPRECTS) {
741 ret = r300_emit_cliprects(dev_priv, cmdbuf, n);
745 cmdbuf->buffer->iterator = orig_iter;
748 switch (header.packet3.packet) {
749 case R300_CMD_PACKET3_CLEAR:
750 DRM_DEBUG("R300_CMD_PACKET3_CLEAR\n");
751 ret = r300_emit_clear(dev_priv, cmdbuf);
753 DRM_ERROR("r300_emit_clear failed\n");
758 case R300_CMD_PACKET3_RAW:
759 DRM_DEBUG("R300_CMD_PACKET3_RAW\n");
760 ret = r300_emit_raw_packet3(dev_priv, cmdbuf);
762 DRM_ERROR("r300_emit_raw_packet3 failed\n");
768 DRM_ERROR("bad packet3 type %i at byte %d\n",
769 header.packet3.packet,
770 cmdbuf->buffer->iterator - (int)sizeof(header));
774 n += R300_SIMULTANEOUS_CLIPRECTS;
775 } while (n < cmdbuf->nbox);
780 /* Some of the R300 chips seem to be extremely touchy about the two registers
781 * that are configured in r300_pacify.
782 * Among the worst offenders seems to be the R300 ND (0x4E44): When userspace
783 * sends a command buffer that contains only state setting commands and a
784 * vertex program/parameter upload sequence, this will eventually lead to a
785 * lockup, unless the sequence is bracketed by calls to r300_pacify.
786 * So we should take great care to *always* call r300_pacify before
787 * *anything* 3D related, and again afterwards. This is what the
788 * call bracket in r300_do_cp_cmdbuf is for.
792 * Emit the sequence to pacify R300.
794 static void r300_pacify(drm_radeon_private_t *dev_priv)
796 uint32_t cache_z, cache_3d, cache_2d;
799 cache_z = R300_ZC_FLUSH;
800 cache_2d = R300_RB2D_DC_FLUSH;
801 cache_3d = R300_RB3D_DC_FLUSH;
802 if (!(dev_priv->track_flush & RADEON_PURGE_EMITED)) {
803 /* we can purge, primitive where draw since last purge */
804 cache_z |= R300_ZC_FREE;
805 cache_2d |= R300_RB2D_DC_FREE;
806 cache_3d |= R300_RB3D_DC_FREE;
809 /* flush & purge zbuffer */
811 OUT_RING(CP_PACKET0(R300_ZB_ZCACHE_CTLSTAT, 0));
814 /* flush & purge 3d */
816 OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
819 /* flush & purge texture */
821 OUT_RING(CP_PACKET0(R300_TX_INVALTAGS, 0));
824 /* FIXME: is this one really needed ? */
826 OUT_RING(CP_PACKET0(R300_RB3D_AARESOLVE_CTL, 0));
830 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
831 OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
833 /* flush & purge 2d through E2 as RB2D will trigger lockup */
835 OUT_RING(CP_PACKET0(R300_DSTCACHE_CTLSTAT, 0));
837 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
838 OUT_RING(RADEON_WAIT_2D_IDLECLEAN |
839 RADEON_WAIT_HOST_IDLECLEAN);
841 /* set flush & purge flags */
842 dev_priv->track_flush |= RADEON_FLUSH_EMITED | RADEON_PURGE_EMITED;
846 * Called by r300_do_cp_cmdbuf to update the internal buffer age and state.
847 * The actual age emit is done by r300_do_cp_cmdbuf, which is why you must
848 * be careful about how this function is called.
850 static void r300_discard_buffer(struct drm_device *dev, struct drm_master *master, struct drm_buf *buf)
852 drm_radeon_buf_priv_t *buf_priv = buf->dev_private;
853 struct drm_radeon_master_private *master_priv = master->driver_priv;
855 buf_priv->age = ++master_priv->sarea_priv->last_dispatch;
860 static void r300_cmd_wait(drm_radeon_private_t * dev_priv,
861 drm_r300_cmd_header_t header)
866 if (!header.wait.flags)
871 switch(header.wait.flags) {
873 wait_until = RADEON_WAIT_2D_IDLE;
876 wait_until = RADEON_WAIT_3D_IDLE;
878 case R300_NEW_WAIT_2D_3D:
879 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_3D_IDLE;
881 case R300_NEW_WAIT_2D_2D_CLEAN:
882 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_2D_IDLECLEAN;
884 case R300_NEW_WAIT_3D_3D_CLEAN:
885 wait_until = RADEON_WAIT_3D_IDLE|RADEON_WAIT_3D_IDLECLEAN;
887 case R300_NEW_WAIT_2D_2D_CLEAN_3D_3D_CLEAN:
888 wait_until = RADEON_WAIT_2D_IDLE|RADEON_WAIT_2D_IDLECLEAN;
889 wait_until |= RADEON_WAIT_3D_IDLE|RADEON_WAIT_3D_IDLECLEAN;
896 OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
897 OUT_RING(wait_until);
901 static int r300_scratch(drm_radeon_private_t *dev_priv,
902 drm_radeon_kcmd_buffer_t *cmdbuf,
903 drm_r300_cmd_header_t header)
906 u32 i, *buf_idx, h_pending;
911 if (drm_buffer_unprocessed(cmdbuf->buffer) <
912 (sizeof(u64) + header.scratch.n_bufs * sizeof(*buf_idx))) {
916 if (header.scratch.reg >= 5) {
920 dev_priv->scratch_ages[header.scratch.reg]++;
922 ptr_addr = drm_buffer_read_object(cmdbuf->buffer,
923 sizeof(stack_ptr_addr), &stack_ptr_addr);
924 ref_age_base = (u32 *)(unsigned long)get_unaligned(ptr_addr);
926 for (i=0; i < header.scratch.n_bufs; i++) {
927 buf_idx = drm_buffer_pointer_to_dword(cmdbuf->buffer, 0);
928 *buf_idx *= 2; /* 8 bytes per buf */
930 if (DRM_COPY_TO_USER(ref_age_base + *buf_idx,
931 &dev_priv->scratch_ages[header.scratch.reg],
935 if (DRM_COPY_FROM_USER(&h_pending,
936 ref_age_base + *buf_idx + 1,
945 if (DRM_COPY_TO_USER(ref_age_base + *buf_idx + 1,
950 drm_buffer_advance(cmdbuf->buffer, sizeof(*buf_idx));
954 OUT_RING( CP_PACKET0( RADEON_SCRATCH_REG0 + header.scratch.reg * 4, 0 ) );
955 OUT_RING( dev_priv->scratch_ages[header.scratch.reg] );
962 * Uploads user-supplied vertex program instructions or parameters onto
964 * Called by r300_do_cp_cmdbuf.
966 static inline int r300_emit_r500fp(drm_radeon_private_t *dev_priv,
967 drm_radeon_kcmd_buffer_t *cmdbuf,
968 drm_r300_cmd_header_t header)
977 sz = header.r500fp.count;
978 /* address is 9 bits 0 - 8, bit 1 of flags is part of address */
979 addr = ((header.r500fp.adrhi_flags & 1) << 8) | header.r500fp.adrlo;
981 type = !!(header.r500fp.adrhi_flags & R500FP_CONSTANT_TYPE);
982 isclamp = !!(header.r500fp.adrhi_flags & R500FP_CONSTANT_CLAMP);
984 addr |= (type << 16);
985 addr |= (isclamp << 17);
987 stride = type ? 4 : 6;
989 DRM_DEBUG("r500fp %d %d type: %d\n", sz, addr, type);
992 if (sz * stride * 4 > drm_buffer_unprocessed(cmdbuf->buffer))
995 BEGIN_RING(3 + sz * stride);
996 OUT_RING_REG(R500_GA_US_VECTOR_INDEX, addr);
997 OUT_RING(CP_PACKET0_TABLE(R500_GA_US_VECTOR_DATA, sz * stride - 1));
998 OUT_RING_DRM_BUFFER(cmdbuf->buffer, sz * stride);
1007 * Parses and validates a user-supplied command buffer and emits appropriate
1008 * commands on the DMA ring buffer.
1009 * Called by the ioctl handler function radeon_cp_cmdbuf.
1011 int r300_do_cp_cmdbuf(struct drm_device *dev,
1012 struct drm_file *file_priv,
1013 drm_radeon_kcmd_buffer_t *cmdbuf)
1015 drm_radeon_private_t *dev_priv = dev->dev_private;
1016 struct drm_radeon_master_private *master_priv = file_priv->master->driver_priv;
1017 struct drm_device_dma *dma = dev->dma;
1018 struct drm_buf *buf = NULL;
1019 int emit_dispatch_age = 0;
1025 r300_pacify(dev_priv);
1027 if (cmdbuf->nbox <= R300_SIMULTANEOUS_CLIPRECTS) {
1028 ret = r300_emit_cliprects(dev_priv, cmdbuf, 0);
1033 while (drm_buffer_unprocessed(cmdbuf->buffer)
1034 >= sizeof(drm_r300_cmd_header_t)) {
1036 drm_r300_cmd_header_t *header, stack_header;
1038 header = drm_buffer_read_object(cmdbuf->buffer,
1039 sizeof(stack_header), &stack_header);
1041 switch (header->header.cmd_type) {
1042 case R300_CMD_PACKET0:
1043 DRM_DEBUG("R300_CMD_PACKET0\n");
1044 ret = r300_emit_packet0(dev_priv, cmdbuf, *header);
1046 DRM_ERROR("r300_emit_packet0 failed\n");
1052 DRM_DEBUG("R300_CMD_VPU\n");
1053 ret = r300_emit_vpu(dev_priv, cmdbuf, *header);
1055 DRM_ERROR("r300_emit_vpu failed\n");
1060 case R300_CMD_PACKET3:
1061 DRM_DEBUG("R300_CMD_PACKET3\n");
1062 ret = r300_emit_packet3(dev_priv, cmdbuf, *header);
1064 DRM_ERROR("r300_emit_packet3 failed\n");
1069 case R300_CMD_END3D:
1070 DRM_DEBUG("R300_CMD_END3D\n");
1072 Ideally userspace driver should not need to issue this call,
1073 i.e. the drm driver should issue it automatically and prevent
1076 In practice, we do not understand why this call is needed and what
1077 it does (except for some vague guesses that it has to do with cache
1078 coherence) and so the user space driver does it.
1080 Once we are sure which uses prevent lockups the code could be moved
1081 into the kernel and the userspace driver will not
1082 need to use this command.
1084 Note that issuing this command does not hurt anything
1085 except, possibly, performance */
1086 r300_pacify(dev_priv);
1089 case R300_CMD_CP_DELAY:
1090 /* simple enough, we can do it here */
1091 DRM_DEBUG("R300_CMD_CP_DELAY\n");
1096 BEGIN_RING(header->delay.count);
1097 for (i = 0; i < header->delay.count; i++)
1098 OUT_RING(RADEON_CP_PACKET2);
1103 case R300_CMD_DMA_DISCARD:
1104 DRM_DEBUG("RADEON_CMD_DMA_DISCARD\n");
1105 idx = header->dma.buf_idx;
1106 if (idx < 0 || idx >= dma->buf_count) {
1107 DRM_ERROR("buffer index %d (of %d max)\n",
1108 idx, dma->buf_count - 1);
1113 buf = dma->buflist[idx];
1114 if (buf->file_priv != file_priv || buf->pending) {
1115 DRM_ERROR("bad buffer %p %p %d\n",
1116 buf->file_priv, file_priv,
1122 emit_dispatch_age = 1;
1123 r300_discard_buffer(dev, file_priv->master, buf);
1127 DRM_DEBUG("R300_CMD_WAIT\n");
1128 r300_cmd_wait(dev_priv, *header);
1131 case R300_CMD_SCRATCH:
1132 DRM_DEBUG("R300_CMD_SCRATCH\n");
1133 ret = r300_scratch(dev_priv, cmdbuf, *header);
1135 DRM_ERROR("r300_scratch failed\n");
1140 case R300_CMD_R500FP:
1141 if ((dev_priv->flags & RADEON_FAMILY_MASK) < CHIP_RV515) {
1142 DRM_ERROR("Calling r500 command on r300 card\n");
1146 DRM_DEBUG("R300_CMD_R500FP\n");
1147 ret = r300_emit_r500fp(dev_priv, cmdbuf, *header);
1149 DRM_ERROR("r300_emit_r500fp failed\n");
1154 DRM_ERROR("bad cmd_type %i at byte %d\n",
1155 header->header.cmd_type,
1156 cmdbuf->buffer->iterator - (int)sizeof(*header));
1165 r300_pacify(dev_priv);
1167 /* We emit the vertex buffer age here, outside the pacifier "brackets"
1169 * (1) This may coalesce multiple age emissions into a single one and
1170 * (2) more importantly, some chips lock up hard when scratch registers
1171 * are written inside the pacifier bracket.
1173 if (emit_dispatch_age) {
1176 /* Emit the vertex buffer age */
1178 RADEON_DISPATCH_AGE(master_priv->sarea_priv->last_dispatch);