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 static inline void print_linkn_in (const char *strval, uint8_t byteval)
11 print_debug(strval); print_debug_hex8(byteval); print_debug("\r\n");
14 static uint8_t ht_lookup_slave_capability(device_t dev)
19 hdr_type = pci_read_config8(dev, PCI_HEADER_TYPE);
23 if ((hdr_type == PCI_HEADER_TYPE_NORMAL) ||
24 (hdr_type == PCI_HEADER_TYPE_BRIDGE)) {
25 pos = PCI_CAPABILITY_LIST;
27 if (pos > PCI_CAP_LIST_NEXT) {
28 pos = pci_read_config8(dev, pos);
30 while(pos != 0) { /* loop through the linked list */
32 cap = pci_read_config8(dev, pos + PCI_CAP_LIST_ID);
33 if (cap == PCI_CAP_ID_HT) {
36 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS);
37 if ((flags >> 13) == 0) {
38 /* Entry is a Slave secondary, success... */
42 pos = pci_read_config8(dev, pos + PCI_CAP_LIST_NEXT);
47 static void ht_collapse_previous_enumeration(uint8_t bus)
52 /* Check if is already collapsed */
53 dev = PCI_DEV(bus, 0, 0);
54 id = pci_read_config32(dev, PCI_VENDOR_ID);
55 if ( ! ( (id == 0xffffffff) || (id == 0x00000000) ||
56 (id == 0x0000ffff) || (id == 0xffff0000) ) ) {
60 /* Spin through the devices and collapse any previous
61 * hypertransport enumeration.
63 for(dev = PCI_DEV(bus, 1, 0); dev <= PCI_DEV(bus, 0x1f, 0x7); dev += PCI_DEV(0, 1, 0)) {
68 id = pci_read_config32(dev, PCI_VENDOR_ID);
69 if ((id == 0xffffffff) || (id == 0x00000000) ||
70 (id == 0x0000ffff) || (id == 0xffff0000)) {
74 pos = ht_lookup_slave_capability(dev);
79 /* Clear the unitid */
80 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS);
82 pci_write_config16(dev, pos + PCI_CAP_FLAGS, flags);
86 static uint16_t ht_read_freq_cap(device_t dev, uint8_t pos)
88 /* Handle bugs in valid hypertransport frequency reporting */
92 freq_cap = pci_read_config16(dev, pos);
93 freq_cap &= ~(1 << HT_FREQ_VENDOR); /* Ignore Vendor HT frequencies */
95 id = pci_read_config32(dev, 0);
97 /* AMD 8131 Errata 48 */
98 if (id == (PCI_VENDOR_ID_AMD | (PCI_DEVICE_ID_AMD_8131_PCIX << 16))) {
99 freq_cap &= ~(1 << HT_FREQ_800Mhz);
101 /* AMD 8151 Errata 23 */
102 if (id == (PCI_VENDOR_ID_AMD | (PCI_DEVICE_ID_AMD_8151_SYSCTRL << 16))) {
103 freq_cap &= ~(1 << HT_FREQ_800Mhz);
105 /* AMD K8 Unsupported 1Ghz? */
106 if (id == (PCI_VENDOR_ID_AMD | (0x1100 << 16))) {
107 freq_cap &= ~(1 << HT_FREQ_1000Mhz);
112 #define LINK_OFFS(WIDTH,FREQ,FREQ_CAP) \
113 (((WIDTH & 0xff) << 16) | ((FREQ & 0xff) << 8) | (FREQ_CAP & 0xFF))
115 #define LINK_WIDTH(OFFS) ((OFFS >> 16) & 0xFF)
116 #define LINK_FREQ(OFFS) ((OFFS >> 8) & 0xFF)
117 #define LINK_FREQ_CAP(OFFS) ((OFFS) & 0xFF)
119 #define PCI_HT_HOST_OFFS LINK_OFFS( \
120 PCI_HT_CAP_HOST_WIDTH, \
121 PCI_HT_CAP_HOST_FREQ, \
122 PCI_HT_CAP_HOST_FREQ_CAP)
124 #define PCI_HT_SLAVE0_OFFS LINK_OFFS( \
125 PCI_HT_CAP_SLAVE_WIDTH0, \
126 PCI_HT_CAP_SLAVE_FREQ0, \
127 PCI_HT_CAP_SLAVE_FREQ_CAP0)
129 #define PCI_HT_SLAVE1_OFFS LINK_OFFS( \
130 PCI_HT_CAP_SLAVE_WIDTH1, \
131 PCI_HT_CAP_SLAVE_FREQ1, \
132 PCI_HT_CAP_SLAVE_FREQ_CAP1)
134 static int ht_optimize_link(
135 device_t dev1, uint8_t pos1, unsigned offs1,
136 device_t dev2, uint8_t pos2, unsigned offs2)
138 static const uint8_t link_width_to_pow2[]= { 3, 4, 0, 5, 1, 2, 0, 0 };
139 static const uint8_t pow2_to_link_width[] = { 0x7, 4, 5, 0, 1, 3 };
140 uint16_t freq_cap1, freq_cap2, freq_cap, freq_mask;
141 uint8_t width_cap1, width_cap2, width_cap, width, old_width, ln_width1, ln_width2;
142 uint8_t freq, old_freq;
144 /* Set link width and frequency */
146 /* Initially assume everything is already optimized and I don't need a reset */
149 /* Get the frequency capabilities */
150 freq_cap1 = ht_read_freq_cap(dev1, pos1 + LINK_FREQ_CAP(offs1));
151 freq_cap2 = ht_read_freq_cap(dev2, pos2 + LINK_FREQ_CAP(offs2));
153 /* Calculate the highest possible frequency */
154 freq = log2(freq_cap1 & freq_cap2);
156 /* See if I am changing the link freqency */
157 old_freq = pci_read_config8(dev1, pos1 + LINK_FREQ(offs1));
158 needs_reset |= old_freq != freq;
159 old_freq = pci_read_config8(dev2, pos2 + LINK_FREQ(offs2));
160 needs_reset |= old_freq != freq;
162 /* Set the Calulcated link frequency */
163 pci_write_config8(dev1, pos1 + LINK_FREQ(offs1), freq);
164 pci_write_config8(dev2, pos2 + LINK_FREQ(offs2), freq);
166 /* Get the width capabilities */
167 width_cap1 = pci_read_config8(dev1, pos1 + LINK_WIDTH(offs1));
168 width_cap2 = pci_read_config8(dev2, pos2 + LINK_WIDTH(offs2));
170 /* Calculate dev1's input width */
171 ln_width1 = link_width_to_pow2[width_cap1 & 7];
172 ln_width2 = link_width_to_pow2[(width_cap2 >> 4) & 7];
173 if (ln_width1 > ln_width2) {
174 ln_width1 = ln_width2;
176 width = pow2_to_link_width[ln_width1];
177 /* Calculate dev1's output width */
178 ln_width1 = link_width_to_pow2[(width_cap1 >> 4) & 7];
179 ln_width2 = link_width_to_pow2[width_cap2 & 7];
180 if (ln_width1 > ln_width2) {
181 ln_width1 = ln_width2;
183 width |= pow2_to_link_width[ln_width1] << 4;
185 /* See if I am changing dev1's width */
186 old_width = pci_read_config8(dev1, pos1 + LINK_WIDTH(offs1) + 1);
187 needs_reset |= old_width != width;
189 /* Set dev1's widths */
190 pci_write_config8(dev1, pos1 + LINK_WIDTH(offs1) + 1, width);
192 /* Calculate dev2's width */
193 width = ((width & 0x70) >> 4) | ((width & 0x7) << 4);
195 /* See if I am changing dev2's width */
196 old_width = pci_read_config8(dev2, pos2 + LINK_WIDTH(offs2) + 1);
197 needs_reset |= old_width != width;
199 /* Set dev2's widths */
200 pci_write_config8(dev2, pos2 + LINK_WIDTH(offs2) + 1, width);
204 static int ht_setup_chain(device_t udev, uint8_t upos)
206 /* Assumption the HT chain that is bus 0 has the HT I/O Hub on it.
207 * On most boards this just happens. If a cpu has multiple
208 * non Coherent links the appropriate bus registers for the
209 * links needs to be programed to point at bus 0.
211 unsigned next_unitid, last_unitid;
215 /* Make certain the HT bus is not enumerated */
216 ht_collapse_previous_enumeration(0);
219 uoffs = PCI_HT_HOST_OFFS;
224 uint16_t flags, count;
227 device_t dev = PCI_DEV(0, 0, 0);
228 last_unitid = next_unitid;
230 id = pci_read_config32(dev, PCI_VENDOR_ID);
231 /* If the chain is enumerated quit */
232 if (((id & 0xffff) == 0x0000) || ((id & 0xffff) == 0xffff) ||
233 (((id >> 16) & 0xffff) == 0xffff) ||
234 (((id >> 16) & 0xffff) == 0x0000)) {
238 pos = ht_lookup_slave_capability(dev);
240 print_err("HT link capability not found\r\n");
244 /* get ht direction */
245 offs = (pci_read_config16(dev, pos + PCI_CAP_FLAGS) & (1<<10)) ? PCI_HT_SLAVE1_OFFS : PCI_HT_SLAVE0_OFFS;
247 /* Setup the Hypertransport link */
248 reset_needed |= ht_optimize_link(udev, upos, uoffs, dev, pos, offs);
250 /* Update the Unitid of the current device */
251 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS);
252 flags &= ~0x1f; /* mask out the bse Unit ID */
253 flags |= next_unitid & 0x1f;
254 pci_write_config16(dev, pos + PCI_CAP_FLAGS, flags);
256 /* Remeber the location of the last device */
257 udev = PCI_DEV(0, next_unitid, 0);
259 uoffs = (offs != PCI_HT_SLAVE0_OFFS) ? PCI_HT_SLAVE0_OFFS : PCI_HT_SLAVE1_OFFS;
261 /* Compute the number of unitids consumed */
262 count = (flags >> 5) & 0x1f;
263 next_unitid += count;
265 } while((last_unitid != next_unitid) && (next_unitid <= 0x1f));
269 static int ht_setup_chainx(device_t udev, uint8_t upos, uint8_t bus)
271 unsigned next_unitid, last_unitid;
275 uoffs = PCI_HT_HOST_OFFS;
281 uint16_t flags, count;
284 device_t dev = PCI_DEV(bus, 0, 0);
285 last_unitid = next_unitid;
287 id = pci_read_config32(dev, PCI_VENDOR_ID);
288 /* If the chain is enumerated quit */
289 if (((id & 0xffff) == 0x0000) || ((id & 0xffff) == 0xffff) ||
290 (((id >> 16) & 0xffff) == 0xffff) ||
291 (((id >> 16) & 0xffff) == 0x0000)) {
295 pos = ht_lookup_slave_capability(dev);
297 print_err("HT link capability not found\r\n");
301 /* get ht direction */
302 offs = (pci_read_config16(dev, pos + PCI_CAP_FLAGS) & (1<<10)) ? PCI_HT_SLAVE1_OFFS : PCI_HT_SLAVE0_OFFS;
304 /* Setup the Hypertransport link */
305 reset_needed |= ht_optimize_link(udev, upos, uoffs, dev, pos, offs);
307 /* Update the Unitid of the current device */
308 flags = pci_read_config16(dev, pos + PCI_CAP_FLAGS);
309 flags &= ~0x1f; /* mask out the bse Unit ID */
310 flags |= next_unitid & 0x1f;
311 pci_write_config16(dev, pos + PCI_CAP_FLAGS, flags);
313 /* Remeber the location of the last device */
314 udev = PCI_DEV(bus, next_unitid, 0);
316 uoffs = ( offs != PCI_HT_SLAVE0_OFFS ) ? PCI_HT_SLAVE0_OFFS : PCI_HT_SLAVE1_OFFS;
318 /* Compute the number of unitids consumed */
319 count = (flags >> 5) & 0x1f;
320 next_unitid += count;
322 } while((last_unitid != next_unitid) && (next_unitid <= 0x1f));
326 static int optimize_link_read_pointer(uint8_t node, uint8_t linkn, uint8_t linkt, uint8_t val)
328 uint32_t dword, dword_old;
331 /* This works on an Athlon64 because unimplemented links return 0 */
332 dword = pci_read_config32(PCI_DEV(0,0x18+node,0), 0x98 + (linkn * 0x20));
333 link_type = dword & 0xff;
335 dword_old = dword = pci_read_config32(PCI_DEV(0,0x18+node,3), 0xdc);
337 if ( (link_type & 7) == linkt ) { /* Coherent Link only linkt = 3, ncoherent = 7*/
338 dword &= ~( 0xff<<(linkn *8) );
339 dword |= val << (linkn *8);
342 if (dword != dword_old) {
343 pci_write_config32(PCI_DEV(0,0x18+node,3), 0xdc, dword);
350 static int optimize_link_in_coherent(uint8_t ht_c_num)
357 for (i = 0; i < ht_c_num; i++) {
359 uint8_t nodeid, linkn;
363 reg = pci_read_config32(PCI_DEV(0,0x18,1), 0xe0 + i * 4);
365 nodeid = ((reg & 0xf0)>>4); // nodeid
366 linkn = ((reg & 0xf00)>>8); // link n
367 busn = (reg & 0xff0000)>>16; //busn
369 reg = pci_read_config32( PCI_DEV(busn, 1, 0), PCI_VENDOR_ID);
370 if ( (reg & 0xffff) == PCI_VENDOR_ID_AMD) {
372 } else if ( (reg & 0xffff) == 0x10de ) {
378 reset_needed |= optimize_link_read_pointer(nodeid, linkn, 0x07, val);
385 static int ht_setup_chains(uint8_t ht_c_num)
387 /* Assumption the HT chain that is bus 0 has the HT I/O Hub on it.
388 * On most boards this just happens. If a cpu has multiple
389 * non Coherent links the appropriate bus registers for the
390 * links needs to be programed to point at bus 0.
399 for (i = 0; i < ht_c_num; i++) {
406 reg = pci_read_config32(PCI_DEV(0,0x18,1), 0xe0 + i * 4);
408 //We need setup 0x94, 0xb4, and 0xd4 according to the reg
409 devpos = ((reg & 0xf0)>>4)+0x18; // nodeid; it will decide 0x18 or 0x19
410 regpos = ((reg & 0xf00)>>8) * 0x20 + 0x94; // link n; it will decide 0x94 or 0xb4, 0x0xd4;
411 busn = (reg & 0xff0000)>>16;
413 dword = pci_read_config32( PCI_DEV(0, devpos, 0), regpos) ;
414 dword &= ~(0xffff<<8);
415 dword |= (reg & 0xffff0000)>>8;
416 pci_write_config32( PCI_DEV(0, devpos,0), regpos , dword);
418 /* Make certain the HT bus is not enumerated */
419 ht_collapse_previous_enumeration(busn);
421 upos = ((reg & 0xf00)>>8) * 0x20 + 0x80;
422 udev = PCI_DEV(0, devpos, 0);
424 reset_needed |= ht_setup_chainx(udev,upos,busn);
429 reset_needed |= optimize_link_in_coherent(ht_c_num);
434 static int ht_setup_chains_x(void)
443 /* read PCI_DEV(0,0x18,0) 0x64 bit [8:9] to find out SbLink m */
444 reg = pci_read_config32(PCI_DEV(0, 0x18, 0), 0x64);
445 /* update PCI_DEV(0, 0x18, 1) 0xe0 to 0x05000m03, and next_busn=5+1 */
446 print_linkn_in("SBLink=", ((reg>>8) & 3) );
447 tempreg = 3 | ( 0<<4) | (((reg>>8) & 3)<<8) | (0<<16)| (5<<24);
448 pci_write_config32(PCI_DEV(0, 0x18, 1), 0xe0, tempreg);
450 next_busn=5+1; /* 0 will be used ht chain with SB we need to keep SB in bus0 in auto stage*/
452 for(ht_c_num=1;ht_c_num<4; ht_c_num++) {
453 pci_write_config32(PCI_DEV(0, 0x18, 1), 0xe0 + ht_c_num * 4, 0);
456 nodes = ((pci_read_config32(PCI_DEV(0, 0x18, 0), 0x60)>>4) & 7) + 1;
458 for(nodeid=0; nodeid<nodes; nodeid++) {
461 dev = PCI_DEV(0, 0x18+nodeid,0);
462 for(linkn = 0; linkn<3; linkn++) {
464 regpos = 0x98 + 0x20 * linkn;
465 reg = pci_read_config32(dev, regpos);
466 if ((reg & 0x17) != 7) continue; /* it is not non conherent or not connected*/
467 print_linkn_in("NC node|link=", ((nodeid & 0xf)<<4)|(linkn & 0xf));
468 tempreg = 3 | (nodeid <<4) | (linkn<<8);
469 /*compare (temp & 0xffff), with (PCI(0, 0x18, 1) 0xe0 to 0xec & 0xfffff) */
470 for(ht_c_num=0;ht_c_num<4; ht_c_num++) {
471 reg = pci_read_config32(PCI_DEV(0, 0x18, 1), 0xe0 + ht_c_num * 4);
472 if(((reg & 0xffff) == (tempreg & 0xffff)) || ((reg & 0xffff) == 0x0000)) { /*we got it*/
476 if(ht_c_num == 4) break; /*used up onle 4 non conherent allowed*/
477 /*update to 0xe0...*/
478 if((reg & 0xf) == 3) continue; /*SbLink so don't touch it */
479 print_linkn_in("\tbusn=", next_busn);
480 tempreg |= (next_busn<<16)|((next_busn+5)<<24);
481 pci_write_config32(PCI_DEV(0, 0x18, 1), 0xe0 + ht_c_num * 4, tempreg);
485 /*update 0xe0, 0xe4, 0xe8, 0xec from PCI_DEV(0, 0x18,1) to PCI_DEV(0, 0x19,1) to PCI_DEV(0, 0x1f,1);*/
487 for(nodeid = 1; nodeid<nodes; nodeid++) {
490 dev = PCI_DEV(0, 0x18+nodeid,1);
491 for(i = 0; i< 4; i++) {
493 regpos = 0xe0 + i * 4;
494 reg = pci_read_config32(PCI_DEV(0, 0x18, 1), regpos);
495 pci_write_config32(dev, regpos, reg);
500 /* recount ht_c_num*/
502 for(ht_c_num=0;ht_c_num<4; ht_c_num++) {
503 reg = pci_read_config32(PCI_DEV(0, 0x18, 1), 0xe0 + ht_c_num * 4);
504 if(((reg & 0xf) != 0x0)) {
509 return ht_setup_chains(i);