2 This should be done by Eric
3 2004.12 yhlu add multi ht chain dynamically support
5 #include <device/pci_def.h>
6 #include <device/pci_ids.h>
7 #include <device/hypertransport_def.h>
9 #ifndef K8_HT_FREQ_1G_SUPPORT
10 #define K8_HT_FREQ_1G_SUPPORT 0
13 static inline void print_linkn_in (const char *strval, uint8_t byteval)
17 printk_debug("%s%02x\r\n", strval, byteval);
19 print_debug(strval); print_debug_hex8(byteval); print_debug("\r\n");
24 static uint8_t ht_lookup_slave_capability(device_t dev)
29 hdr_type = pci_read_config8(dev, PCI_HEADER_TYPE);
33 if ((hdr_type == PCI_HEADER_TYPE_NORMAL) ||
34 (hdr_type == PCI_HEADER_TYPE_BRIDGE)) {
35 pos = PCI_CAPABILITY_LIST;
37 if (pos > PCI_CAP_LIST_NEXT) {
38 pos = pci_read_config8(dev, pos);
40 while(pos != 0) { /* loop through the linked list */
42 cap = pci_read_config8(dev, pos + PCI_CAP_LIST_ID);
43 if (cap == PCI_CAP_ID_HT) {
46 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS);
47 if ((flags >> 13) == 0) {
48 /* Entry is a Slave secondary, success... */
52 pos = pci_read_config8(dev, pos + PCI_CAP_LIST_NEXT);
57 static void ht_collapse_previous_enumeration(uint8_t bus)
62 /* Check if is already collapsed */
63 dev = PCI_DEV(bus, 0, 0);
64 id = pci_read_config32(dev, PCI_VENDOR_ID);
65 if ( ! ( (id == 0xffffffff) || (id == 0x00000000) ||
66 (id == 0x0000ffff) || (id == 0xffff0000) ) ) {
70 /* Spin through the devices and collapse any previous
71 * hypertransport enumeration.
73 for(dev = PCI_DEV(bus, 1, 0); dev <= PCI_DEV(bus, 0x1f, 0x7); dev += PCI_DEV(0, 1, 0)) {
78 id = pci_read_config32(dev, PCI_VENDOR_ID);
79 if ((id == 0xffffffff) || (id == 0x00000000) ||
80 (id == 0x0000ffff) || (id == 0xffff0000)) {
84 #if CK804_DEVN_BASE==0
86 // CK804 UnitID changes not use
87 if(id == 0x005e10de) {
93 pos = ht_lookup_slave_capability(dev);
98 /* Clear the unitid */
99 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS);
101 pci_write_config16(dev, pos + PCI_CAP_FLAGS, flags);
105 static uint16_t ht_read_freq_cap(device_t dev, uint8_t pos)
107 /* Handle bugs in valid hypertransport frequency reporting */
111 freq_cap = pci_read_config16(dev, pos);
112 freq_cap &= ~(1 << HT_FREQ_VENDOR); /* Ignore Vendor HT frequencies */
114 id = pci_read_config32(dev, 0);
116 /* AMD 8131 Errata 48 */
117 if (id == (PCI_VENDOR_ID_AMD | (PCI_DEVICE_ID_AMD_8131_PCIX << 16))) {
118 freq_cap &= ~(1 << HT_FREQ_800Mhz);
122 /* AMD 8151 Errata 23 */
123 if (id == (PCI_VENDOR_ID_AMD | (PCI_DEVICE_ID_AMD_8151_SYSCTRL << 16))) {
124 freq_cap &= ~(1 << HT_FREQ_800Mhz);
128 /* AMD K8 Unsupported 1Ghz? */
129 if (id == (PCI_VENDOR_ID_AMD | (0x1100 << 16))) {
130 #if K8_HT_FREQ_1G_SUPPORT == 1
131 if (is_cpu_pre_e0()) // CK804 support 1G?
133 freq_cap &= ~(1 << HT_FREQ_1000Mhz);
139 #define LINK_OFFS(WIDTH,FREQ,FREQ_CAP) \
140 (((WIDTH & 0xff) << 16) | ((FREQ & 0xff) << 8) | (FREQ_CAP & 0xFF))
142 #define LINK_WIDTH(OFFS) ((OFFS >> 16) & 0xFF)
143 #define LINK_FREQ(OFFS) ((OFFS >> 8) & 0xFF)
144 #define LINK_FREQ_CAP(OFFS) ((OFFS) & 0xFF)
146 #define PCI_HT_HOST_OFFS LINK_OFFS( \
147 PCI_HT_CAP_HOST_WIDTH, \
148 PCI_HT_CAP_HOST_FREQ, \
149 PCI_HT_CAP_HOST_FREQ_CAP)
151 #define PCI_HT_SLAVE0_OFFS LINK_OFFS( \
152 PCI_HT_CAP_SLAVE_WIDTH0, \
153 PCI_HT_CAP_SLAVE_FREQ0, \
154 PCI_HT_CAP_SLAVE_FREQ_CAP0)
156 #define PCI_HT_SLAVE1_OFFS LINK_OFFS( \
157 PCI_HT_CAP_SLAVE_WIDTH1, \
158 PCI_HT_CAP_SLAVE_FREQ1, \
159 PCI_HT_CAP_SLAVE_FREQ_CAP1)
161 static int ht_optimize_link(
162 device_t dev1, uint8_t pos1, unsigned offs1,
163 device_t dev2, uint8_t pos2, unsigned offs2)
165 static const uint8_t link_width_to_pow2[]= { 3, 4, 0, 5, 1, 2, 0, 0 };
166 static const uint8_t pow2_to_link_width[] = { 0x7, 4, 5, 0, 1, 3 };
167 uint16_t freq_cap1, freq_cap2, freq_cap, freq_mask;
168 uint8_t width_cap1, width_cap2, width_cap, width, old_width, ln_width1, ln_width2;
169 uint8_t freq, old_freq;
171 /* Set link width and frequency */
173 /* Initially assume everything is already optimized and I don't need a reset */
176 /* Get the frequency capabilities */
177 freq_cap1 = ht_read_freq_cap(dev1, pos1 + LINK_FREQ_CAP(offs1));
178 freq_cap2 = ht_read_freq_cap(dev2, pos2 + LINK_FREQ_CAP(offs2));
180 /* Calculate the highest possible frequency */
181 freq = log2(freq_cap1 & freq_cap2);
183 /* See if I am changing the link freqency */
184 old_freq = pci_read_config8(dev1, pos1 + LINK_FREQ(offs1));
185 needs_reset |= old_freq != freq;
186 old_freq = pci_read_config8(dev2, pos2 + LINK_FREQ(offs2));
187 needs_reset |= old_freq != freq;
189 /* Set the Calulcated link frequency */
190 pci_write_config8(dev1, pos1 + LINK_FREQ(offs1), freq);
191 pci_write_config8(dev2, pos2 + LINK_FREQ(offs2), freq);
193 /* Get the width capabilities */
194 width_cap1 = pci_read_config8(dev1, pos1 + LINK_WIDTH(offs1));
195 width_cap2 = pci_read_config8(dev2, pos2 + LINK_WIDTH(offs2));
197 /* Calculate dev1's input width */
198 ln_width1 = link_width_to_pow2[width_cap1 & 7];
199 ln_width2 = link_width_to_pow2[(width_cap2 >> 4) & 7];
200 if (ln_width1 > ln_width2) {
201 ln_width1 = ln_width2;
203 width = pow2_to_link_width[ln_width1];
204 /* Calculate dev1's output width */
205 ln_width1 = link_width_to_pow2[(width_cap1 >> 4) & 7];
206 ln_width2 = link_width_to_pow2[width_cap2 & 7];
207 if (ln_width1 > ln_width2) {
208 ln_width1 = ln_width2;
210 width |= pow2_to_link_width[ln_width1] << 4;
212 /* See if I am changing dev1's width */
213 old_width = pci_read_config8(dev1, pos1 + LINK_WIDTH(offs1) + 1);
214 needs_reset |= old_width != width;
216 /* Set dev1's widths */
217 pci_write_config8(dev1, pos1 + LINK_WIDTH(offs1) + 1, width);
219 /* Calculate dev2's width */
220 width = ((width & 0x70) >> 4) | ((width & 0x7) << 4);
222 /* See if I am changing dev2's width */
223 old_width = pci_read_config8(dev2, pos2 + LINK_WIDTH(offs2) + 1);
224 needs_reset |= old_width != width;
226 /* Set dev2's widths */
227 pci_write_config8(dev2, pos2 + LINK_WIDTH(offs2) + 1, width);
231 static int ht_setup_chain(device_t udev, uint8_t upos)
233 /* Assumption the HT chain that is bus 0 has the HT I/O Hub on it.
234 * On most boards this just happens. If a cpu has multiple
235 * non Coherent links the appropriate bus registers for the
236 * links needs to be programed to point at bus 0.
238 uint8_t next_unitid, last_unitid;
242 /* Make certain the HT bus is not enumerated */
243 ht_collapse_previous_enumeration(0);
246 uoffs = PCI_HT_HOST_OFFS;
255 device_t dev = PCI_DEV(0, 0, 0);
256 last_unitid = next_unitid;
258 id = pci_read_config32(dev, PCI_VENDOR_ID);
259 /* If the chain is enumerated quit */
260 if (((id & 0xffff) == 0x0000) || ((id & 0xffff) == 0xffff) ||
261 (((id >> 16) & 0xffff) == 0xffff) ||
262 (((id >> 16) & 0xffff) == 0x0000)) {
266 pos = ht_lookup_slave_capability(dev);
268 print_err("HT link capability not found\r\n");
271 #if CK804_DEVN_BASE==0
273 // CK804 UnitID changes not use
274 id = pci_read_config32(dev, PCI_VENDOR_ID);
275 if(id != 0x005e10de) {
278 /* Update the Unitid of the current device */
279 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS);
280 flags &= ~0x1f; /* mask out the bse Unit ID */
281 flags |= next_unitid & 0x1f;
282 pci_write_config16(dev, pos + PCI_CAP_FLAGS, flags);
284 dev = PCI_DEV(0, next_unitid, 0);
285 #if CK804_DEVN_BASE==0
288 dev = PCI_DEV(0, 0, 0);
292 /* Compute the number of unitids consumed */
293 count = (flags >> 5) & 0x1f;
294 next_unitid += count;
296 /* get ht direction */
297 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS); // double read ??
299 offs = ((flags>>10) & 1) ? PCI_HT_SLAVE1_OFFS : PCI_HT_SLAVE0_OFFS;
301 /* Setup the Hypertransport link */
302 reset_needed |= ht_optimize_link(udev, upos, uoffs, dev, pos, offs);
304 #if CK804_DEVN_BASE==0
305 if(id == 0x005e10de) {
310 /* Remeber the location of the last device */
313 uoffs = (offs != PCI_HT_SLAVE0_OFFS) ? PCI_HT_SLAVE0_OFFS : PCI_HT_SLAVE1_OFFS;
315 } while((last_unitid != next_unitid) && (next_unitid <= 0x1f));
319 static int ht_setup_chainx(device_t udev, uint8_t upos, uint8_t bus)
321 uint8_t next_unitid, last_unitid;
325 uoffs = PCI_HT_HOST_OFFS;
335 device_t dev = PCI_DEV(bus, 0, 0);
336 last_unitid = next_unitid;
338 id = pci_read_config32(dev, PCI_VENDOR_ID);
340 /* If the chain is enumerated quit */
341 if (((id & 0xffff) == 0x0000) || ((id & 0xffff) == 0xffff) ||
342 (((id >> 16) & 0xffff) == 0xffff) ||
343 (((id >> 16) & 0xffff) == 0x0000)) {
347 pos = ht_lookup_slave_capability(dev);
349 print_err(" HT link capability not found\r\n");
353 #if CK804_DEVN_BASE==0
355 // CK804 UnitID changes not use
356 id = pci_read_config32(dev, PCI_VENDOR_ID);
357 if(id != 0x005e10de) {
360 /* Update the Unitid of the current device */
361 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS);
362 flags &= ~0x1f; /* mask out the bse Unit ID */
363 flags |= next_unitid & 0x1f;
364 pci_write_config16(dev, pos + PCI_CAP_FLAGS, flags);
366 dev = PCI_DEV(bus, next_unitid, 0);
367 #if CK804_DEVN_BASE==0
370 dev = PCI_DEV(bus, 0, 0);
374 /* Compute the number of unitids consumed */
375 count = (flags >> 5) & 0x1f;
376 next_unitid += count;
378 /* get ht direction */
379 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS); // double read ??
381 offs = ((flags>>10) & 1) ? PCI_HT_SLAVE1_OFFS : PCI_HT_SLAVE0_OFFS;
383 /* Setup the Hypertransport link */
384 reset_needed |= ht_optimize_link(udev, upos, uoffs, dev, pos, offs);
386 #if CK804_DEVN_BASE==0
387 if(id == 0x005e10de) {
392 /* Remeber the location of the last device */
395 uoffs = ( offs != PCI_HT_SLAVE0_OFFS ) ? PCI_HT_SLAVE0_OFFS : PCI_HT_SLAVE1_OFFS;
397 } while((last_unitid != next_unitid) && (next_unitid <= 0x1f));
401 static int optimize_link_read_pointer(uint8_t node, uint8_t linkn, uint8_t linkt, uint8_t val)
403 uint32_t dword, dword_old;
406 /* This works on an Athlon64 because unimplemented links return 0 */
407 dword = pci_read_config32(PCI_DEV(0,0x18+node,0), 0x98 + (linkn * 0x20));
408 link_type = dword & 0xff;
410 dword_old = dword = pci_read_config32(PCI_DEV(0,0x18+node,3), 0xdc);
412 if ( (link_type & 7) == linkt ) { /* Coherent Link only linkt = 3, ncoherent = 7*/
413 dword &= ~( 0xff<<(linkn *8) );
414 dword |= val << (linkn *8);
417 if (dword != dword_old) {
418 pci_write_config32(PCI_DEV(0,0x18+node,3), 0xdc, dword);
425 static int optimize_link_in_coherent(uint8_t ht_c_num)
432 for (i = 0; i < ht_c_num; i++) {
434 uint8_t nodeid, linkn;
438 reg = pci_read_config32(PCI_DEV(0,0x18,1), 0xe0 + i * 4);
440 nodeid = ((reg & 0xf0)>>4); // nodeid
441 linkn = ((reg & 0xf00)>>8); // link n
442 busn = (reg & 0xff0000)>>16; //busn
444 reg = pci_read_config32( PCI_DEV(busn, 1, 0), PCI_VENDOR_ID);
445 if ( (reg & 0xffff) == PCI_VENDOR_ID_AMD) {
447 } else if ( (reg & 0xffff) == 0x10de ) {
453 reset_needed |= optimize_link_read_pointer(nodeid, linkn, 0x07, val);
460 static int ht_setup_chains(uint8_t ht_c_num)
462 /* Assumption the HT chain that is bus 0 has the HT I/O Hub on it.
463 * On most boards this just happens. If a cpu has multiple
464 * non Coherent links the appropriate bus registers for the
465 * links needs to be programed to point at bus 0.
474 for (i = 0; i < ht_c_num; i++) {
481 reg = pci_read_config32(PCI_DEV(0,0x18,1), 0xe0 + i * 4);
483 //We need setup 0x94, 0xb4, and 0xd4 according to the reg
484 devpos = ((reg & 0xf0)>>4)+0x18; // nodeid; it will decide 0x18 or 0x19
485 regpos = ((reg & 0xf00)>>8) * 0x20 + 0x94; // link n; it will decide 0x94 or 0xb4, 0x0xd4;
486 busn = (reg & 0xff0000)>>16;
488 dword = pci_read_config32( PCI_DEV(0, devpos, 0), regpos) ;
489 dword &= ~(0xffff<<8);
490 dword |= (reg & 0xffff0000)>>8;
491 pci_write_config32( PCI_DEV(0, devpos,0), regpos , dword);
493 /* Make certain the HT bus is not enumerated */
494 ht_collapse_previous_enumeration(busn);
496 upos = ((reg & 0xf00)>>8) * 0x20 + 0x80;
497 udev = PCI_DEV(0, devpos, 0);
499 reset_needed |= ht_setup_chainx(udev,upos,busn);
504 reset_needed |= optimize_link_in_coherent(ht_c_num);
509 static int ht_setup_chains_x(void)
518 /* read PCI_DEV(0,0x18,0) 0x64 bit [8:9] to find out SbLink m */
519 reg = pci_read_config32(PCI_DEV(0, 0x18, 0), 0x64);
520 /* update PCI_DEV(0, 0x18, 1) 0xe0 to 0x05000m03, and next_busn=5+1 */
521 print_linkn_in("SBLink=", ((reg>>8) & 3) );
522 tempreg = 3 | ( 0<<4) | (((reg>>8) & 3)<<8) | (0<<16)| (5<<24);
523 pci_write_config32(PCI_DEV(0, 0x18, 1), 0xe0, tempreg);
525 next_busn=5+1; /* 0 will be used ht chain with SB we need to keep SB in bus0 in auto stage*/
527 for(ht_c_num=1;ht_c_num<4; ht_c_num++) {
528 pci_write_config32(PCI_DEV(0, 0x18, 1), 0xe0 + ht_c_num * 4, 0);
531 nodes = ((pci_read_config32(PCI_DEV(0, 0x18, 0), 0x60)>>4) & 7) + 1;
533 for(nodeid=0; nodeid<nodes; nodeid++) {
536 dev = PCI_DEV(0, 0x18+nodeid,0);
537 for(linkn = 0; linkn<3; linkn++) {
539 regpos = 0x98 + 0x20 * linkn;
540 reg = pci_read_config32(dev, regpos);
541 if ((reg & 0x17) != 7) continue; /* it is not non conherent or not connected*/
542 print_linkn_in("NC node|link=", ((nodeid & 0xf)<<4)|(linkn & 0xf));
543 tempreg = 3 | (nodeid <<4) | (linkn<<8);
544 /*compare (temp & 0xffff), with (PCI(0, 0x18, 1) 0xe0 to 0xec & 0xfffff) */
545 for(ht_c_num=0;ht_c_num<4; ht_c_num++) {
546 reg = pci_read_config32(PCI_DEV(0, 0x18, 1), 0xe0 + ht_c_num * 4);
547 if(((reg & 0xffff) == (tempreg & 0xffff)) || ((reg & 0xffff) == 0x0000)) { /*we got it*/
551 if(ht_c_num == 4) break; /*used up onle 4 non conherent allowed*/
552 /*update to 0xe0...*/
553 if((reg & 0xf) == 3) continue; /*SbLink so don't touch it */
554 print_linkn_in("\tbusn=", next_busn);
555 tempreg |= (next_busn<<16)|((next_busn+5)<<24);
556 pci_write_config32(PCI_DEV(0, 0x18, 1), 0xe0 + ht_c_num * 4, tempreg);
560 /*update 0xe0, 0xe4, 0xe8, 0xec from PCI_DEV(0, 0x18,1) to PCI_DEV(0, 0x19,1) to PCI_DEV(0, 0x1f,1);*/
562 for(nodeid = 1; nodeid<nodes; nodeid++) {
565 dev = PCI_DEV(0, 0x18+nodeid,1);
566 for(i = 0; i< 4; i++) {
568 regpos = 0xe0 + i * 4;
569 reg = pci_read_config32(PCI_DEV(0, 0x18, 1), regpos);
570 pci_write_config32(dev, regpos, reg);
575 /* recount ht_c_num*/
577 for(ht_c_num=0;ht_c_num<4; ht_c_num++) {
578 reg = pci_read_config32(PCI_DEV(0, 0x18, 1), 0xe0 + ht_c_num * 4);
579 if(((reg & 0xf) != 0x0)) {
584 return ht_setup_chains(i);