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)
16 print_debug(strval); print_debug_hex8(byteval); print_debug("\r\n");
20 static uint8_t ht_lookup_slave_capability(device_t dev)
25 hdr_type = pci_read_config8(dev, PCI_HEADER_TYPE);
29 if ((hdr_type == PCI_HEADER_TYPE_NORMAL) ||
30 (hdr_type == PCI_HEADER_TYPE_BRIDGE)) {
31 pos = PCI_CAPABILITY_LIST;
33 if (pos > PCI_CAP_LIST_NEXT) {
34 pos = pci_read_config8(dev, pos);
36 while(pos != 0) { /* loop through the linked list */
38 cap = pci_read_config8(dev, pos + PCI_CAP_LIST_ID);
39 if (cap == PCI_CAP_ID_HT) {
42 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS);
43 if ((flags >> 13) == 0) {
44 /* Entry is a Slave secondary, success... */
48 pos = pci_read_config8(dev, pos + PCI_CAP_LIST_NEXT);
53 static void ht_collapse_previous_enumeration(uint8_t bus)
58 /* Check if is already collapsed */
59 dev = PCI_DEV(bus, 0, 0);
60 id = pci_read_config32(dev, PCI_VENDOR_ID);
61 if ( ! ( (id == 0xffffffff) || (id == 0x00000000) ||
62 (id == 0x0000ffff) || (id == 0xffff0000) ) ) {
66 /* Spin through the devices and collapse any previous
67 * hypertransport enumeration.
69 for(dev = PCI_DEV(bus, 1, 0); dev <= PCI_DEV(bus, 0x1f, 0x7); dev += PCI_DEV(0, 1, 0)) {
74 id = pci_read_config32(dev, PCI_VENDOR_ID);
75 if ((id == 0xffffffff) || (id == 0x00000000) ||
76 (id == 0x0000ffff) || (id == 0xffff0000)) {
80 #if CK804_DEVN_BASE==0
82 // CK804 UnitID changes not use
83 if(id == 0x005e10de) {
89 pos = ht_lookup_slave_capability(dev);
94 /* Clear the unitid */
95 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS);
97 pci_write_config16(dev, pos + PCI_CAP_FLAGS, flags);
101 static uint16_t ht_read_freq_cap(device_t dev, uint8_t pos)
103 /* Handle bugs in valid hypertransport frequency reporting */
107 freq_cap = pci_read_config16(dev, pos);
108 freq_cap &= ~(1 << HT_FREQ_VENDOR); /* Ignore Vendor HT frequencies */
110 id = pci_read_config32(dev, 0);
112 /* AMD 8131 Errata 48 */
113 if (id == (PCI_VENDOR_ID_AMD | (PCI_DEVICE_ID_AMD_8131_PCIX << 16))) {
114 freq_cap &= ~(1 << HT_FREQ_800Mhz);
118 /* AMD 8151 Errata 23 */
119 if (id == (PCI_VENDOR_ID_AMD | (PCI_DEVICE_ID_AMD_8151_SYSCTRL << 16))) {
120 freq_cap &= ~(1 << HT_FREQ_800Mhz);
124 /* AMD K8 Unsupported 1Ghz? */
125 if (id == (PCI_VENDOR_ID_AMD | (0x1100 << 16))) {
126 freq_cap &= ~(1 << HT_FREQ_1000Mhz);
132 #define LINK_OFFS(WIDTH,FREQ,FREQ_CAP) \
133 (((WIDTH & 0xff) << 16) | ((FREQ & 0xff) << 8) | (FREQ_CAP & 0xFF))
135 #define LINK_WIDTH(OFFS) ((OFFS >> 16) & 0xFF)
136 #define LINK_FREQ(OFFS) ((OFFS >> 8) & 0xFF)
137 #define LINK_FREQ_CAP(OFFS) ((OFFS) & 0xFF)
139 #define PCI_HT_HOST_OFFS LINK_OFFS( \
140 PCI_HT_CAP_HOST_WIDTH, \
141 PCI_HT_CAP_HOST_FREQ, \
142 PCI_HT_CAP_HOST_FREQ_CAP)
144 #define PCI_HT_SLAVE0_OFFS LINK_OFFS( \
145 PCI_HT_CAP_SLAVE_WIDTH0, \
146 PCI_HT_CAP_SLAVE_FREQ0, \
147 PCI_HT_CAP_SLAVE_FREQ_CAP0)
149 #define PCI_HT_SLAVE1_OFFS LINK_OFFS( \
150 PCI_HT_CAP_SLAVE_WIDTH1, \
151 PCI_HT_CAP_SLAVE_FREQ1, \
152 PCI_HT_CAP_SLAVE_FREQ_CAP1)
154 static int ht_optimize_link(
155 device_t dev1, uint8_t pos1, unsigned offs1,
156 device_t dev2, uint8_t pos2, unsigned offs2)
158 static const uint8_t link_width_to_pow2[]= { 3, 4, 0, 5, 1, 2, 0, 0 };
159 static const uint8_t pow2_to_link_width[] = { 0x7, 4, 5, 0, 1, 3 };
160 uint16_t freq_cap1, freq_cap2, freq_cap, freq_mask;
161 uint8_t width_cap1, width_cap2, width_cap, width, old_width, ln_width1, ln_width2;
162 uint8_t freq, old_freq;
164 /* Set link width and frequency */
166 /* Initially assume everything is already optimized and I don't need a reset */
169 /* Get the frequency capabilities */
170 freq_cap1 = ht_read_freq_cap(dev1, pos1 + LINK_FREQ_CAP(offs1));
171 freq_cap2 = ht_read_freq_cap(dev2, pos2 + LINK_FREQ_CAP(offs2));
173 /* Calculate the highest possible frequency */
174 freq = log2(freq_cap1 & freq_cap2);
176 /* See if I am changing the link freqency */
177 old_freq = pci_read_config8(dev1, pos1 + LINK_FREQ(offs1));
178 needs_reset |= old_freq != freq;
179 old_freq = pci_read_config8(dev2, pos2 + LINK_FREQ(offs2));
180 needs_reset |= old_freq != freq;
182 /* Set the Calulcated link frequency */
183 pci_write_config8(dev1, pos1 + LINK_FREQ(offs1), freq);
184 pci_write_config8(dev2, pos2 + LINK_FREQ(offs2), freq);
186 /* Get the width capabilities */
187 width_cap1 = pci_read_config8(dev1, pos1 + LINK_WIDTH(offs1));
188 width_cap2 = pci_read_config8(dev2, pos2 + LINK_WIDTH(offs2));
190 /* Calculate dev1's input width */
191 ln_width1 = link_width_to_pow2[width_cap1 & 7];
192 ln_width2 = link_width_to_pow2[(width_cap2 >> 4) & 7];
193 if (ln_width1 > ln_width2) {
194 ln_width1 = ln_width2;
196 width = pow2_to_link_width[ln_width1];
197 /* Calculate dev1's output width */
198 ln_width1 = link_width_to_pow2[(width_cap1 >> 4) & 7];
199 ln_width2 = link_width_to_pow2[width_cap2 & 7];
200 if (ln_width1 > ln_width2) {
201 ln_width1 = ln_width2;
203 width |= pow2_to_link_width[ln_width1] << 4;
205 /* See if I am changing dev1's width */
206 old_width = pci_read_config8(dev1, pos1 + LINK_WIDTH(offs1) + 1);
207 needs_reset |= old_width != width;
209 /* Set dev1's widths */
210 pci_write_config8(dev1, pos1 + LINK_WIDTH(offs1) + 1, width);
212 /* Calculate dev2's width */
213 width = ((width & 0x70) >> 4) | ((width & 0x7) << 4);
215 /* See if I am changing dev2's width */
216 old_width = pci_read_config8(dev2, pos2 + LINK_WIDTH(offs2) + 1);
217 needs_reset |= old_width != width;
219 /* Set dev2's widths */
220 pci_write_config8(dev2, pos2 + LINK_WIDTH(offs2) + 1, width);
224 static int ht_setup_chain(device_t udev, uint8_t upos)
226 /* Assumption the HT chain that is bus 0 has the HT I/O Hub on it.
227 * On most boards this just happens. If a cpu has multiple
228 * non Coherent links the appropriate bus registers for the
229 * links needs to be programed to point at bus 0.
231 uint8_t next_unitid, last_unitid;
235 /* Make certain the HT bus is not enumerated */
236 ht_collapse_previous_enumeration(0);
239 uoffs = PCI_HT_HOST_OFFS;
248 device_t dev = PCI_DEV(0, 0, 0);
249 last_unitid = next_unitid;
251 id = pci_read_config32(dev, PCI_VENDOR_ID);
252 /* If the chain is enumerated quit */
253 if (((id & 0xffff) == 0x0000) || ((id & 0xffff) == 0xffff) ||
254 (((id >> 16) & 0xffff) == 0xffff) ||
255 (((id >> 16) & 0xffff) == 0x0000)) {
259 pos = ht_lookup_slave_capability(dev);
261 print_err("HT link capability not found\r\n");
264 #if CK804_DEVN_BASE==0
266 // CK804 UnitID changes not use
267 id = pci_read_config32(dev, PCI_VENDOR_ID);
268 if(id != 0x005e10de) {
271 /* Update the Unitid of the current device */
272 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS);
273 flags &= ~0x1f; /* mask out the bse Unit ID */
274 flags |= next_unitid & 0x1f;
275 pci_write_config16(dev, pos + PCI_CAP_FLAGS, flags);
277 dev = PCI_DEV(0, next_unitid, 0);
278 #if CK804_DEVN_BASE==0
281 dev = PCI_DEV(0, 0, 0);
285 /* Compute the number of unitids consumed */
286 count = (flags >> 5) & 0x1f;
287 next_unitid += count;
289 /* get ht direction */
290 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS); // double read ??
292 offs = ((flags>>10) & 1) ? PCI_HT_SLAVE1_OFFS : PCI_HT_SLAVE0_OFFS;
294 /* Setup the Hypertransport link */
295 reset_needed |= ht_optimize_link(udev, upos, uoffs, dev, pos, offs);
297 #if CK804_DEVN_BASE==0
298 if(id == 0x005e10de) {
303 /* Remeber the location of the last device */
306 uoffs = (offs != PCI_HT_SLAVE0_OFFS) ? PCI_HT_SLAVE0_OFFS : PCI_HT_SLAVE1_OFFS;
308 } while((last_unitid != next_unitid) && (next_unitid <= 0x1f));
312 static int ht_setup_chainx(device_t udev, uint8_t upos, uint8_t bus)
314 uint8_t next_unitid, last_unitid;
318 uoffs = PCI_HT_HOST_OFFS;
328 device_t dev = PCI_DEV(bus, 0, 0);
329 last_unitid = next_unitid;
331 id = pci_read_config32(dev, PCI_VENDOR_ID);
333 /* If the chain is enumerated quit */
334 if (((id & 0xffff) == 0x0000) || ((id & 0xffff) == 0xffff) ||
335 (((id >> 16) & 0xffff) == 0xffff) ||
336 (((id >> 16) & 0xffff) == 0x0000)) {
340 pos = ht_lookup_slave_capability(dev);
342 print_err(" HT link capability not found\r\n");
346 #if CK804_DEVN_BASE==0
348 // CK804 UnitID changes not use
349 id = pci_read_config32(dev, PCI_VENDOR_ID);
350 if(id != 0x005e10de) {
353 /* Update the Unitid of the current device */
354 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS);
355 flags &= ~0x1f; /* mask out the bse Unit ID */
356 flags |= next_unitid & 0x1f;
357 pci_write_config16(dev, pos + PCI_CAP_FLAGS, flags);
359 dev = PCI_DEV(bus, next_unitid, 0);
360 #if CK804_DEVN_BASE==0
363 dev = PCI_DEV(bus, 0, 0);
367 /* Compute the number of unitids consumed */
368 count = (flags >> 5) & 0x1f;
369 next_unitid += count;
371 /* get ht direction */
372 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS); // double read ??
374 offs = ((flags>>10) & 1) ? PCI_HT_SLAVE1_OFFS : PCI_HT_SLAVE0_OFFS;
376 /* Setup the Hypertransport link */
377 reset_needed |= ht_optimize_link(udev, upos, uoffs, dev, pos, offs);
379 #if CK804_DEVN_BASE==0
380 if(id == 0x005e10de) {
385 /* Remeber the location of the last device */
388 uoffs = ( offs != PCI_HT_SLAVE0_OFFS ) ? PCI_HT_SLAVE0_OFFS : PCI_HT_SLAVE1_OFFS;
390 } while((last_unitid != next_unitid) && (next_unitid <= 0x1f));
394 static int optimize_link_read_pointer(uint8_t node, uint8_t linkn, uint8_t linkt, uint8_t val)
396 uint32_t dword, dword_old;
399 /* This works on an Athlon64 because unimplemented links return 0 */
400 dword = pci_read_config32(PCI_DEV(0,0x18+node,0), 0x98 + (linkn * 0x20));
401 link_type = dword & 0xff;
403 dword_old = dword = pci_read_config32(PCI_DEV(0,0x18+node,3), 0xdc);
405 if ( (link_type & 7) == linkt ) { /* Coherent Link only linkt = 3, ncoherent = 7*/
406 dword &= ~( 0xff<<(linkn *8) );
407 dword |= val << (linkn *8);
410 if (dword != dword_old) {
411 pci_write_config32(PCI_DEV(0,0x18+node,3), 0xdc, dword);
418 static int optimize_link_in_coherent(uint8_t ht_c_num)
425 for (i = 0; i < ht_c_num; i++) {
427 uint8_t nodeid, linkn;
431 reg = pci_read_config32(PCI_DEV(0,0x18,1), 0xe0 + i * 4);
433 nodeid = ((reg & 0xf0)>>4); // nodeid
434 linkn = ((reg & 0xf00)>>8); // link n
435 busn = (reg & 0xff0000)>>16; //busn
437 reg = pci_read_config32( PCI_DEV(busn, 1, 0), PCI_VENDOR_ID);
438 if ( (reg & 0xffff) == PCI_VENDOR_ID_AMD) {
440 } else if ( (reg & 0xffff) == 0x10de ) {
446 reset_needed |= optimize_link_read_pointer(nodeid, linkn, 0x07, val);
453 static int ht_setup_chains(uint8_t ht_c_num)
455 /* Assumption the HT chain that is bus 0 has the HT I/O Hub on it.
456 * On most boards this just happens. If a cpu has multiple
457 * non Coherent links the appropriate bus registers for the
458 * links needs to be programed to point at bus 0.
467 for (i = 0; i < ht_c_num; i++) {
474 reg = pci_read_config32(PCI_DEV(0,0x18,1), 0xe0 + i * 4);
476 //We need setup 0x94, 0xb4, and 0xd4 according to the reg
477 devpos = ((reg & 0xf0)>>4)+0x18; // nodeid; it will decide 0x18 or 0x19
478 regpos = ((reg & 0xf00)>>8) * 0x20 + 0x94; // link n; it will decide 0x94 or 0xb4, 0x0xd4;
479 busn = (reg & 0xff0000)>>16;
481 dword = pci_read_config32( PCI_DEV(0, devpos, 0), regpos) ;
482 dword &= ~(0xffff<<8);
483 dword |= (reg & 0xffff0000)>>8;
484 pci_write_config32( PCI_DEV(0, devpos,0), regpos , dword);
486 /* Make certain the HT bus is not enumerated */
487 ht_collapse_previous_enumeration(busn);
489 upos = ((reg & 0xf00)>>8) * 0x20 + 0x80;
490 udev = PCI_DEV(0, devpos, 0);
492 reset_needed |= ht_setup_chainx(udev,upos,busn);
497 reset_needed |= optimize_link_in_coherent(ht_c_num);
502 static int ht_setup_chains_x(void)
511 /* read PCI_DEV(0,0x18,0) 0x64 bit [8:9] to find out SbLink m */
512 reg = pci_read_config32(PCI_DEV(0, 0x18, 0), 0x64);
513 /* update PCI_DEV(0, 0x18, 1) 0xe0 to 0x05000m03, and next_busn=5+1 */
514 print_linkn_in("SBLink=", ((reg>>8) & 3) );
515 tempreg = 3 | ( 0<<4) | (((reg>>8) & 3)<<8) | (0<<16)| (5<<24);
516 pci_write_config32(PCI_DEV(0, 0x18, 1), 0xe0, tempreg);
518 next_busn=5+1; /* 0 will be used ht chain with SB we need to keep SB in bus0 in auto stage*/
520 for(ht_c_num=1;ht_c_num<4; ht_c_num++) {
521 pci_write_config32(PCI_DEV(0, 0x18, 1), 0xe0 + ht_c_num * 4, 0);
524 nodes = ((pci_read_config32(PCI_DEV(0, 0x18, 0), 0x60)>>4) & 7) + 1;
526 for(nodeid=0; nodeid<nodes; nodeid++) {
529 dev = PCI_DEV(0, 0x18+nodeid,0);
530 for(linkn = 0; linkn<3; linkn++) {
532 regpos = 0x98 + 0x20 * linkn;
533 reg = pci_read_config32(dev, regpos);
534 if ((reg & 0x17) != 7) continue; /* it is not non conherent or not connected*/
535 print_linkn_in("NC node|link=", ((nodeid & 0xf)<<4)|(linkn & 0xf));
536 tempreg = 3 | (nodeid <<4) | (linkn<<8);
537 /*compare (temp & 0xffff), with (PCI(0, 0x18, 1) 0xe0 to 0xec & 0xfffff) */
538 for(ht_c_num=0;ht_c_num<4; ht_c_num++) {
539 reg = pci_read_config32(PCI_DEV(0, 0x18, 1), 0xe0 + ht_c_num * 4);
540 if(((reg & 0xffff) == (tempreg & 0xffff)) || ((reg & 0xffff) == 0x0000)) { /*we got it*/
544 if(ht_c_num == 4) break; /*used up onle 4 non conherent allowed*/
545 /*update to 0xe0...*/
546 if((reg & 0xf) == 3) continue; /*SbLink so don't touch it */
547 print_linkn_in("\tbusn=", next_busn);
548 tempreg |= (next_busn<<16)|((next_busn+5)<<24);
549 pci_write_config32(PCI_DEV(0, 0x18, 1), 0xe0 + ht_c_num * 4, tempreg);
553 /*update 0xe0, 0xe4, 0xe8, 0xec from PCI_DEV(0, 0x18,1) to PCI_DEV(0, 0x19,1) to PCI_DEV(0, 0x1f,1);*/
555 for(nodeid = 1; nodeid<nodes; nodeid++) {
558 dev = PCI_DEV(0, 0x18+nodeid,1);
559 for(i = 0; i< 4; i++) {
561 regpos = 0xe0 + i * 4;
562 reg = pci_read_config32(PCI_DEV(0, 0x18, 1), regpos);
563 pci_write_config32(dev, regpos, reg);
568 /* recount ht_c_num*/
570 for(ht_c_num=0;ht_c_num<4; ht_c_num++) {
571 reg = pci_read_config32(PCI_DEV(0, 0x18, 1), 0xe0 + ht_c_num * 4);
572 if(((reg & 0xf) != 0x0)) {
577 return ht_setup_chains(i);