xref: /openbmc/linux/arch/mips/pci/pci-octeon.c (revision a09d2831)
1 /*
2  * This file is subject to the terms and conditions of the GNU General Public
3  * License.  See the file "COPYING" in the main directory of this archive
4  * for more details.
5  *
6  * Copyright (C) 2005-2009 Cavium Networks
7  */
8 #include <linux/kernel.h>
9 #include <linux/init.h>
10 #include <linux/pci.h>
11 #include <linux/interrupt.h>
12 #include <linux/time.h>
13 #include <linux/delay.h>
14 
15 #include <asm/time.h>
16 
17 #include <asm/octeon/octeon.h>
18 #include <asm/octeon/cvmx-npi-defs.h>
19 #include <asm/octeon/cvmx-pci-defs.h>
20 #include <asm/octeon/pci-octeon.h>
21 
22 #define USE_OCTEON_INTERNAL_ARBITER
23 
24 /*
25  * Octeon's PCI controller uses did=3, subdid=2 for PCI IO
26  * addresses. Use PCI endian swapping 1 so no address swapping is
27  * necessary. The Linux io routines will endian swap the data.
28  */
29 #define OCTEON_PCI_IOSPACE_BASE     0x80011a0400000000ull
30 #define OCTEON_PCI_IOSPACE_SIZE     (1ull<<32)
31 
32 /* Octeon't PCI controller uses did=3, subdid=3 for PCI memory. */
33 #define OCTEON_PCI_MEMSPACE_OFFSET  (0x00011b0000000000ull)
34 
35 /**
36  * This is the bit decoding used for the Octeon PCI controller addresses
37  */
38 union octeon_pci_address {
39 	uint64_t u64;
40 	struct {
41 		uint64_t upper:2;
42 		uint64_t reserved:13;
43 		uint64_t io:1;
44 		uint64_t did:5;
45 		uint64_t subdid:3;
46 		uint64_t reserved2:4;
47 		uint64_t endian_swap:2;
48 		uint64_t reserved3:10;
49 		uint64_t bus:8;
50 		uint64_t dev:5;
51 		uint64_t func:3;
52 		uint64_t reg:8;
53 	} s;
54 };
55 
56 int __initdata (*octeon_pcibios_map_irq)(const struct pci_dev *dev,
57 					 u8 slot, u8 pin);
58 enum octeon_dma_bar_type octeon_dma_bar_type = OCTEON_DMA_BAR_TYPE_INVALID;
59 
60 /**
61  * Map a PCI device to the appropriate interrupt line
62  *
63  * @dev:    The Linux PCI device structure for the device to map
64  * @slot:   The slot number for this device on __BUS 0__. Linux
65  *               enumerates through all the bridges and figures out the
66  *               slot on Bus 0 where this device eventually hooks to.
67  * @pin:    The PCI interrupt pin read from the device, then swizzled
68  *               as it goes through each bridge.
69  * Returns Interrupt number for the device
70  */
71 int __init pcibios_map_irq(const struct pci_dev *dev, u8 slot, u8 pin)
72 {
73 	if (octeon_pcibios_map_irq)
74 		return octeon_pcibios_map_irq(dev, slot, pin);
75 	else
76 		panic("octeon_pcibios_map_irq not set.");
77 }
78 
79 
80 /*
81  * Called to perform platform specific PCI setup
82  */
83 int pcibios_plat_dev_init(struct pci_dev *dev)
84 {
85 	uint16_t config;
86 	uint32_t dconfig;
87 	int pos;
88 	/*
89 	 * Force the Cache line setting to 64 bytes. The standard
90 	 * Linux bus scan doesn't seem to set it. Octeon really has
91 	 * 128 byte lines, but Intel bridges get really upset if you
92 	 * try and set values above 64 bytes. Value is specified in
93 	 * 32bit words.
94 	 */
95 	pci_write_config_byte(dev, PCI_CACHE_LINE_SIZE, 64 / 4);
96 	/* Set latency timers for all devices */
97 	pci_write_config_byte(dev, PCI_LATENCY_TIMER, 48);
98 
99 	/* Enable reporting System errors and parity errors on all devices */
100 	/* Enable parity checking and error reporting */
101 	pci_read_config_word(dev, PCI_COMMAND, &config);
102 	config |= PCI_COMMAND_PARITY | PCI_COMMAND_SERR;
103 	pci_write_config_word(dev, PCI_COMMAND, config);
104 
105 	if (dev->subordinate) {
106 		/* Set latency timers on sub bridges */
107 		pci_write_config_byte(dev, PCI_SEC_LATENCY_TIMER, 48);
108 		/* More bridge error detection */
109 		pci_read_config_word(dev, PCI_BRIDGE_CONTROL, &config);
110 		config |= PCI_BRIDGE_CTL_PARITY | PCI_BRIDGE_CTL_SERR;
111 		pci_write_config_word(dev, PCI_BRIDGE_CONTROL, config);
112 	}
113 
114 	/* Enable the PCIe normal error reporting */
115 	pos = pci_find_capability(dev, PCI_CAP_ID_EXP);
116 	if (pos) {
117 		/* Update Device Control */
118 		pci_read_config_word(dev, pos + PCI_EXP_DEVCTL, &config);
119 		/* Correctable Error Reporting */
120 		config |= PCI_EXP_DEVCTL_CERE;
121 		/* Non-Fatal Error Reporting */
122 		config |= PCI_EXP_DEVCTL_NFERE;
123 		/* Fatal Error Reporting */
124 		config |= PCI_EXP_DEVCTL_FERE;
125 		/* Unsupported Request */
126 		config |= PCI_EXP_DEVCTL_URRE;
127 		pci_write_config_word(dev, pos + PCI_EXP_DEVCTL, config);
128 	}
129 
130 	/* Find the Advanced Error Reporting capability */
131 	pos = pci_find_ext_capability(dev, PCI_EXT_CAP_ID_ERR);
132 	if (pos) {
133 		/* Clear Uncorrectable Error Status */
134 		pci_read_config_dword(dev, pos + PCI_ERR_UNCOR_STATUS,
135 				      &dconfig);
136 		pci_write_config_dword(dev, pos + PCI_ERR_UNCOR_STATUS,
137 				       dconfig);
138 		/* Enable reporting of all uncorrectable errors */
139 		/* Uncorrectable Error Mask - turned on bits disable errors */
140 		pci_write_config_dword(dev, pos + PCI_ERR_UNCOR_MASK, 0);
141 		/*
142 		 * Leave severity at HW default. This only controls if
143 		 * errors are reported as uncorrectable or
144 		 * correctable, not if the error is reported.
145 		 */
146 		/* PCI_ERR_UNCOR_SEVER - Uncorrectable Error Severity */
147 		/* Clear Correctable Error Status */
148 		pci_read_config_dword(dev, pos + PCI_ERR_COR_STATUS, &dconfig);
149 		pci_write_config_dword(dev, pos + PCI_ERR_COR_STATUS, dconfig);
150 		/* Enable reporting of all correctable errors */
151 		/* Correctable Error Mask - turned on bits disable errors */
152 		pci_write_config_dword(dev, pos + PCI_ERR_COR_MASK, 0);
153 		/* Advanced Error Capabilities */
154 		pci_read_config_dword(dev, pos + PCI_ERR_CAP, &dconfig);
155 		/* ECRC Generation Enable */
156 		if (config & PCI_ERR_CAP_ECRC_GENC)
157 			config |= PCI_ERR_CAP_ECRC_GENE;
158 		/* ECRC Check Enable */
159 		if (config & PCI_ERR_CAP_ECRC_CHKC)
160 			config |= PCI_ERR_CAP_ECRC_CHKE;
161 		pci_write_config_dword(dev, pos + PCI_ERR_CAP, dconfig);
162 		/* PCI_ERR_HEADER_LOG - Header Log Register (16 bytes) */
163 		/* Report all errors to the root complex */
164 		pci_write_config_dword(dev, pos + PCI_ERR_ROOT_COMMAND,
165 				       PCI_ERR_ROOT_CMD_COR_EN |
166 				       PCI_ERR_ROOT_CMD_NONFATAL_EN |
167 				       PCI_ERR_ROOT_CMD_FATAL_EN);
168 		/* Clear the Root status register */
169 		pci_read_config_dword(dev, pos + PCI_ERR_ROOT_STATUS, &dconfig);
170 		pci_write_config_dword(dev, pos + PCI_ERR_ROOT_STATUS, dconfig);
171 	}
172 
173 	return 0;
174 }
175 
176 /**
177  * Return the mapping of PCI device number to IRQ line. Each
178  * character in the return string represents the interrupt
179  * line for the device at that position. Device 1 maps to the
180  * first character, etc. The characters A-D are used for PCI
181  * interrupts.
182  *
183  * Returns PCI interrupt mapping
184  */
185 const char *octeon_get_pci_interrupts(void)
186 {
187 	/*
188 	 * Returning an empty string causes the interrupts to be
189 	 * routed based on the PCI specification. From the PCI spec:
190 	 *
191 	 * INTA# of Device Number 0 is connected to IRQW on the system
192 	 * board.  (Device Number has no significance regarding being
193 	 * located on the system board or in a connector.) INTA# of
194 	 * Device Number 1 is connected to IRQX on the system
195 	 * board. INTA# of Device Number 2 is connected to IRQY on the
196 	 * system board. INTA# of Device Number 3 is connected to IRQZ
197 	 * on the system board. The table below describes how each
198 	 * agent's INTx# lines are connected to the system board
199 	 * interrupt lines. The following equation can be used to
200 	 * determine to which INTx# signal on the system board a given
201 	 * device's INTx# line(s) is connected.
202 	 *
203 	 * MB = (D + I) MOD 4 MB = System board Interrupt (IRQW = 0,
204 	 * IRQX = 1, IRQY = 2, and IRQZ = 3) D = Device Number I =
205 	 * Interrupt Number (INTA# = 0, INTB# = 1, INTC# = 2, and
206 	 * INTD# = 3)
207 	 */
208 	switch (octeon_bootinfo->board_type) {
209 	case CVMX_BOARD_TYPE_NAO38:
210 		/* This is really the NAC38 */
211 		return "AAAAADABAAAAAAAAAAAAAAAAAAAAAAAA";
212 	case CVMX_BOARD_TYPE_THUNDER:
213 		return "";
214 	case CVMX_BOARD_TYPE_EBH3000:
215 		return "";
216 	case CVMX_BOARD_TYPE_EBH3100:
217 	case CVMX_BOARD_TYPE_CN3010_EVB_HS5:
218 	case CVMX_BOARD_TYPE_CN3005_EVB_HS5:
219 		return "AAABAAAAAAAAAAAAAAAAAAAAAAAAAAAA";
220 	case CVMX_BOARD_TYPE_BBGW_REF:
221 		return "AABCD";
222 	default:
223 		return "";
224 	}
225 }
226 
227 /**
228  * Map a PCI device to the appropriate interrupt line
229  *
230  * @dev:    The Linux PCI device structure for the device to map
231  * @slot:   The slot number for this device on __BUS 0__. Linux
232  *               enumerates through all the bridges and figures out the
233  *               slot on Bus 0 where this device eventually hooks to.
234  * @pin:    The PCI interrupt pin read from the device, then swizzled
235  *               as it goes through each bridge.
236  * Returns Interrupt number for the device
237  */
238 int __init octeon_pci_pcibios_map_irq(const struct pci_dev *dev,
239 				      u8 slot, u8 pin)
240 {
241 	int irq_num;
242 	const char *interrupts;
243 	int dev_num;
244 
245 	/* Get the board specific interrupt mapping */
246 	interrupts = octeon_get_pci_interrupts();
247 
248 	dev_num = dev->devfn >> 3;
249 	if (dev_num < strlen(interrupts))
250 		irq_num = ((interrupts[dev_num] - 'A' + pin - 1) & 3) +
251 			OCTEON_IRQ_PCI_INT0;
252 	else
253 		irq_num = ((slot + pin - 3) & 3) + OCTEON_IRQ_PCI_INT0;
254 	return irq_num;
255 }
256 
257 
258 /*
259  * Read a value from configuration space
260  */
261 static int octeon_read_config(struct pci_bus *bus, unsigned int devfn,
262 			      int reg, int size, u32 *val)
263 {
264 	union octeon_pci_address pci_addr;
265 
266 	pci_addr.u64 = 0;
267 	pci_addr.s.upper = 2;
268 	pci_addr.s.io = 1;
269 	pci_addr.s.did = 3;
270 	pci_addr.s.subdid = 1;
271 	pci_addr.s.endian_swap = 1;
272 	pci_addr.s.bus = bus->number;
273 	pci_addr.s.dev = devfn >> 3;
274 	pci_addr.s.func = devfn & 0x7;
275 	pci_addr.s.reg = reg;
276 
277 #if PCI_CONFIG_SPACE_DELAY
278 	udelay(PCI_CONFIG_SPACE_DELAY);
279 #endif
280 	switch (size) {
281 	case 4:
282 		*val = le32_to_cpu(cvmx_read64_uint32(pci_addr.u64));
283 		return PCIBIOS_SUCCESSFUL;
284 	case 2:
285 		*val = le16_to_cpu(cvmx_read64_uint16(pci_addr.u64));
286 		return PCIBIOS_SUCCESSFUL;
287 	case 1:
288 		*val = cvmx_read64_uint8(pci_addr.u64);
289 		return PCIBIOS_SUCCESSFUL;
290 	}
291 	return PCIBIOS_FUNC_NOT_SUPPORTED;
292 }
293 
294 
295 /*
296  * Write a value to PCI configuration space
297  */
298 static int octeon_write_config(struct pci_bus *bus, unsigned int devfn,
299 			       int reg, int size, u32 val)
300 {
301 	union octeon_pci_address pci_addr;
302 
303 	pci_addr.u64 = 0;
304 	pci_addr.s.upper = 2;
305 	pci_addr.s.io = 1;
306 	pci_addr.s.did = 3;
307 	pci_addr.s.subdid = 1;
308 	pci_addr.s.endian_swap = 1;
309 	pci_addr.s.bus = bus->number;
310 	pci_addr.s.dev = devfn >> 3;
311 	pci_addr.s.func = devfn & 0x7;
312 	pci_addr.s.reg = reg;
313 
314 #if PCI_CONFIG_SPACE_DELAY
315 	udelay(PCI_CONFIG_SPACE_DELAY);
316 #endif
317 	switch (size) {
318 	case 4:
319 		cvmx_write64_uint32(pci_addr.u64, cpu_to_le32(val));
320 		return PCIBIOS_SUCCESSFUL;
321 	case 2:
322 		cvmx_write64_uint16(pci_addr.u64, cpu_to_le16(val));
323 		return PCIBIOS_SUCCESSFUL;
324 	case 1:
325 		cvmx_write64_uint8(pci_addr.u64, val);
326 		return PCIBIOS_SUCCESSFUL;
327 	}
328 	return PCIBIOS_FUNC_NOT_SUPPORTED;
329 }
330 
331 
332 static struct pci_ops octeon_pci_ops = {
333 	octeon_read_config,
334 	octeon_write_config,
335 };
336 
337 static struct resource octeon_pci_mem_resource = {
338 	.start = 0,
339 	.end = 0,
340 	.name = "Octeon PCI MEM",
341 	.flags = IORESOURCE_MEM,
342 };
343 
344 /*
345  * PCI ports must be above 16KB so the ISA bus filtering in the PCI-X to PCI
346  * bridge
347  */
348 static struct resource octeon_pci_io_resource = {
349 	.start = 0x4000,
350 	.end = OCTEON_PCI_IOSPACE_SIZE - 1,
351 	.name = "Octeon PCI IO",
352 	.flags = IORESOURCE_IO,
353 };
354 
355 static struct pci_controller octeon_pci_controller = {
356 	.pci_ops = &octeon_pci_ops,
357 	.mem_resource = &octeon_pci_mem_resource,
358 	.mem_offset = OCTEON_PCI_MEMSPACE_OFFSET,
359 	.io_resource = &octeon_pci_io_resource,
360 	.io_offset = 0,
361 	.io_map_base = OCTEON_PCI_IOSPACE_BASE,
362 };
363 
364 
365 /*
366  * Low level initialize the Octeon PCI controller
367  */
368 static void octeon_pci_initialize(void)
369 {
370 	union cvmx_pci_cfg01 cfg01;
371 	union cvmx_npi_ctl_status ctl_status;
372 	union cvmx_pci_ctl_status_2 ctl_status_2;
373 	union cvmx_pci_cfg19 cfg19;
374 	union cvmx_pci_cfg16 cfg16;
375 	union cvmx_pci_cfg22 cfg22;
376 	union cvmx_pci_cfg56 cfg56;
377 
378 	/* Reset the PCI Bus */
379 	cvmx_write_csr(CVMX_CIU_SOFT_PRST, 0x1);
380 	cvmx_read_csr(CVMX_CIU_SOFT_PRST);
381 
382 	udelay(2000);		/* Hold PCI reset for 2 ms */
383 
384 	ctl_status.u64 = 0;	/* cvmx_read_csr(CVMX_NPI_CTL_STATUS); */
385 	ctl_status.s.max_word = 1;
386 	ctl_status.s.timer = 1;
387 	cvmx_write_csr(CVMX_NPI_CTL_STATUS, ctl_status.u64);
388 
389 	/* Deassert PCI reset and advertize PCX Host Mode Device Capability
390 	   (64b) */
391 	cvmx_write_csr(CVMX_CIU_SOFT_PRST, 0x4);
392 	cvmx_read_csr(CVMX_CIU_SOFT_PRST);
393 
394 	udelay(2000);		/* Wait 2 ms after deasserting PCI reset */
395 
396 	ctl_status_2.u32 = 0;
397 	ctl_status_2.s.tsr_hwm = 1;	/* Initializes to 0.  Must be set
398 					   before any PCI reads. */
399 	ctl_status_2.s.bar2pres = 1;	/* Enable BAR2 */
400 	ctl_status_2.s.bar2_enb = 1;
401 	ctl_status_2.s.bar2_cax = 1;	/* Don't use L2 */
402 	ctl_status_2.s.bar2_esx = 1;
403 	ctl_status_2.s.pmo_amod = 1;	/* Round robin priority */
404 	if (octeon_dma_bar_type == OCTEON_DMA_BAR_TYPE_BIG) {
405 		/* BAR1 hole */
406 		ctl_status_2.s.bb1_hole = OCTEON_PCI_BAR1_HOLE_BITS;
407 		ctl_status_2.s.bb1_siz = 1;  /* BAR1 is 2GB */
408 		ctl_status_2.s.bb_ca = 1;    /* Don't use L2 with big bars */
409 		ctl_status_2.s.bb_es = 1;    /* Big bar in byte swap mode */
410 		ctl_status_2.s.bb1 = 1;      /* BAR1 is big */
411 		ctl_status_2.s.bb0 = 1;      /* BAR0 is big */
412 	}
413 
414 	octeon_npi_write32(CVMX_NPI_PCI_CTL_STATUS_2, ctl_status_2.u32);
415 	udelay(2000);		/* Wait 2 ms before doing PCI reads */
416 
417 	ctl_status_2.u32 = octeon_npi_read32(CVMX_NPI_PCI_CTL_STATUS_2);
418 	pr_notice("PCI Status: %s %s-bit\n",
419 		  ctl_status_2.s.ap_pcix ? "PCI-X" : "PCI",
420 		  ctl_status_2.s.ap_64ad ? "64" : "32");
421 
422 	if (OCTEON_IS_MODEL(OCTEON_CN58XX) || OCTEON_IS_MODEL(OCTEON_CN50XX)) {
423 		union cvmx_pci_cnt_reg cnt_reg_start;
424 		union cvmx_pci_cnt_reg cnt_reg_end;
425 		unsigned long cycles, pci_clock;
426 
427 		cnt_reg_start.u64 = cvmx_read_csr(CVMX_NPI_PCI_CNT_REG);
428 		cycles = read_c0_cvmcount();
429 		udelay(1000);
430 		cnt_reg_end.u64 = cvmx_read_csr(CVMX_NPI_PCI_CNT_REG);
431 		cycles = read_c0_cvmcount() - cycles;
432 		pci_clock = (cnt_reg_end.s.pcicnt - cnt_reg_start.s.pcicnt) /
433 			    (cycles / (mips_hpt_frequency / 1000000));
434 		pr_notice("PCI Clock: %lu MHz\n", pci_clock);
435 	}
436 
437 	/*
438 	 * TDOMC must be set to one in PCI mode. TDOMC should be set to 4
439 	 * in PCI-X mode to allow four oustanding splits. Otherwise,
440 	 * should not change from its reset value. Don't write PCI_CFG19
441 	 * in PCI mode (0x82000001 reset value), write it to 0x82000004
442 	 * after PCI-X mode is known. MRBCI,MDWE,MDRE -> must be zero.
443 	 * MRBCM -> must be one.
444 	 */
445 	if (ctl_status_2.s.ap_pcix) {
446 		cfg19.u32 = 0;
447 		/*
448 		 * Target Delayed/Split request outstanding maximum
449 		 * count. [1..31] and 0=32.  NOTE: If the user
450 		 * programs these bits beyond the Designed Maximum
451 		 * outstanding count, then the designed maximum table
452 		 * depth will be used instead.  No additional
453 		 * Deferred/Split transactions will be accepted if
454 		 * this outstanding maximum count is
455 		 * reached. Furthermore, no additional deferred/split
456 		 * transactions will be accepted if the I/O delay/ I/O
457 		 * Split Request outstanding maximum is reached.
458 		 */
459 		cfg19.s.tdomc = 4;
460 		/*
461 		 * Master Deferred Read Request Outstanding Max Count
462 		 * (PCI only).  CR4C[26:24] Max SAC cycles MAX DAC
463 		 * cycles 000 8 4 001 1 0 010 2 1 011 3 1 100 4 2 101
464 		 * 5 2 110 6 3 111 7 3 For example, if these bits are
465 		 * programmed to 100, the core can support 2 DAC
466 		 * cycles, 4 SAC cycles or a combination of 1 DAC and
467 		 * 2 SAC cycles. NOTE: For the PCI-X maximum
468 		 * outstanding split transactions, refer to
469 		 * CRE0[22:20].
470 		 */
471 		cfg19.s.mdrrmc = 2;
472 		/*
473 		 * Master Request (Memory Read) Byte Count/Byte Enable
474 		 * select. 0 = Byte Enables valid. In PCI mode, a
475 		 * burst transaction cannot be performed using Memory
476 		 * Read command=4?h6. 1 = DWORD Byte Count valid
477 		 * (default). In PCI Mode, the memory read byte
478 		 * enables are automatically generated by the
479 		 * core. Note: N3 Master Request transaction sizes are
480 		 * always determined through the
481 		 * am_attr[<35:32>|<7:0>] field.
482 		 */
483 		cfg19.s.mrbcm = 1;
484 		octeon_npi_write32(CVMX_NPI_PCI_CFG19, cfg19.u32);
485 	}
486 
487 
488 	cfg01.u32 = 0;
489 	cfg01.s.msae = 1;	/* Memory Space Access Enable */
490 	cfg01.s.me = 1;		/* Master Enable */
491 	cfg01.s.pee = 1;	/* PERR# Enable */
492 	cfg01.s.see = 1;	/* System Error Enable */
493 	cfg01.s.fbbe = 1;	/* Fast Back to Back Transaction Enable */
494 
495 	octeon_npi_write32(CVMX_NPI_PCI_CFG01, cfg01.u32);
496 
497 #ifdef USE_OCTEON_INTERNAL_ARBITER
498 	/*
499 	 * When OCTEON is a PCI host, most systems will use OCTEON's
500 	 * internal arbiter, so must enable it before any PCI/PCI-X
501 	 * traffic can occur.
502 	 */
503 	{
504 		union cvmx_npi_pci_int_arb_cfg pci_int_arb_cfg;
505 
506 		pci_int_arb_cfg.u64 = 0;
507 		pci_int_arb_cfg.s.en = 1;	/* Internal arbiter enable */
508 		cvmx_write_csr(CVMX_NPI_PCI_INT_ARB_CFG, pci_int_arb_cfg.u64);
509 	}
510 #endif	/* USE_OCTEON_INTERNAL_ARBITER */
511 
512 	/*
513 	 * Preferrably written to 1 to set MLTD. [RDSATI,TRTAE,
514 	 * TWTAE,TMAE,DPPMR -> must be zero. TILT -> must not be set to
515 	 * 1..7.
516 	 */
517 	cfg16.u32 = 0;
518 	cfg16.s.mltd = 1;	/* Master Latency Timer Disable */
519 	octeon_npi_write32(CVMX_NPI_PCI_CFG16, cfg16.u32);
520 
521 	/*
522 	 * Should be written to 0x4ff00. MTTV -> must be zero.
523 	 * FLUSH -> must be 1. MRV -> should be 0xFF.
524 	 */
525 	cfg22.u32 = 0;
526 	/* Master Retry Value [1..255] and 0=infinite */
527 	cfg22.s.mrv = 0xff;
528 	/*
529 	 * AM_DO_FLUSH_I control NOTE: This bit MUST BE ONE for proper
530 	 * N3K operation.
531 	 */
532 	cfg22.s.flush = 1;
533 	octeon_npi_write32(CVMX_NPI_PCI_CFG22, cfg22.u32);
534 
535 	/*
536 	 * MOST Indicates the maximum number of outstanding splits (in -1
537 	 * notation) when OCTEON is in PCI-X mode.  PCI-X performance is
538 	 * affected by the MOST selection.  Should generally be written
539 	 * with one of 0x3be807, 0x2be807, 0x1be807, or 0x0be807,
540 	 * depending on the desired MOST of 3, 2, 1, or 0, respectively.
541 	 */
542 	cfg56.u32 = 0;
543 	cfg56.s.pxcid = 7;	/* RO - PCI-X Capability ID */
544 	cfg56.s.ncp = 0xe8;	/* RO - Next Capability Pointer */
545 	cfg56.s.dpere = 1;	/* Data Parity Error Recovery Enable */
546 	cfg56.s.roe = 1;	/* Relaxed Ordering Enable */
547 	cfg56.s.mmbc = 1;	/* Maximum Memory Byte Count
548 				   [0=512B,1=1024B,2=2048B,3=4096B] */
549 	cfg56.s.most = 3;	/* Maximum outstanding Split transactions [0=1
550 				   .. 7=32] */
551 
552 	octeon_npi_write32(CVMX_NPI_PCI_CFG56, cfg56.u32);
553 
554 	/*
555 	 * Affects PCI performance when OCTEON services reads to its
556 	 * BAR1/BAR2. Refer to Section 10.6.1.  The recommended values are
557 	 * 0x22, 0x33, and 0x33 for PCI_READ_CMD_6, PCI_READ_CMD_C, and
558 	 * PCI_READ_CMD_E, respectively. Unfortunately due to errata DDR-700,
559 	 * these values need to be changed so they won't possibly prefetch off
560 	 * of the end of memory if PCI is DMAing a buffer at the end of
561 	 * memory. Note that these values differ from their reset values.
562 	 */
563 	octeon_npi_write32(CVMX_NPI_PCI_READ_CMD_6, 0x21);
564 	octeon_npi_write32(CVMX_NPI_PCI_READ_CMD_C, 0x31);
565 	octeon_npi_write32(CVMX_NPI_PCI_READ_CMD_E, 0x31);
566 }
567 
568 
569 /*
570  * Initialize the Octeon PCI controller
571  */
572 static int __init octeon_pci_setup(void)
573 {
574 	union cvmx_npi_mem_access_subidx mem_access;
575 	int index;
576 
577 	/* Only these chips have PCI */
578 	if (octeon_has_feature(OCTEON_FEATURE_PCIE))
579 		return 0;
580 
581 	/* Point pcibios_map_irq() to the PCI version of it */
582 	octeon_pcibios_map_irq = octeon_pci_pcibios_map_irq;
583 
584 	/* Only use the big bars on chips that support it */
585 	if (OCTEON_IS_MODEL(OCTEON_CN31XX) ||
586 	    OCTEON_IS_MODEL(OCTEON_CN38XX_PASS2) ||
587 	    OCTEON_IS_MODEL(OCTEON_CN38XX_PASS1))
588 		octeon_dma_bar_type = OCTEON_DMA_BAR_TYPE_SMALL;
589 	else
590 		octeon_dma_bar_type = OCTEON_DMA_BAR_TYPE_BIG;
591 
592 	/* PCI I/O and PCI MEM values */
593 	set_io_port_base(OCTEON_PCI_IOSPACE_BASE);
594 	ioport_resource.start = 0;
595 	ioport_resource.end = OCTEON_PCI_IOSPACE_SIZE - 1;
596 	if (!octeon_is_pci_host()) {
597 		pr_notice("Not in host mode, PCI Controller not initialized\n");
598 		return 0;
599 	}
600 
601 	pr_notice("%s Octeon big bar support\n",
602 		  (octeon_dma_bar_type ==
603 		  OCTEON_DMA_BAR_TYPE_BIG) ? "Enabling" : "Disabling");
604 
605 	octeon_pci_initialize();
606 
607 	mem_access.u64 = 0;
608 	mem_access.s.esr = 1;	/* Endian-Swap on read. */
609 	mem_access.s.esw = 1;	/* Endian-Swap on write. */
610 	mem_access.s.nsr = 0;	/* No-Snoop on read. */
611 	mem_access.s.nsw = 0;	/* No-Snoop on write. */
612 	mem_access.s.ror = 0;	/* Relax Read on read. */
613 	mem_access.s.row = 0;	/* Relax Order on write. */
614 	mem_access.s.ba = 0;	/* PCI Address bits [63:36]. */
615 	cvmx_write_csr(CVMX_NPI_MEM_ACCESS_SUBID3, mem_access.u64);
616 
617 	/*
618 	 * Remap the Octeon BAR 2 above all 32 bit devices
619 	 * (0x8000000000ul).  This is done here so it is remapped
620 	 * before the readl()'s below. We don't want BAR2 overlapping
621 	 * with BAR0/BAR1 during these reads.
622 	 */
623 	octeon_npi_write32(CVMX_NPI_PCI_CFG08, 0);
624 	octeon_npi_write32(CVMX_NPI_PCI_CFG09, 0x80);
625 
626 	/* Disable the BAR1 movable mappings */
627 	for (index = 0; index < 32; index++)
628 		octeon_npi_write32(CVMX_NPI_PCI_BAR1_INDEXX(index), 0);
629 
630 	if (octeon_dma_bar_type == OCTEON_DMA_BAR_TYPE_BIG) {
631 		/* Remap the Octeon BAR 0 to 0-2GB */
632 		octeon_npi_write32(CVMX_NPI_PCI_CFG04, 0);
633 		octeon_npi_write32(CVMX_NPI_PCI_CFG05, 0);
634 
635 		/*
636 		 * Remap the Octeon BAR 1 to map 2GB-4GB (minus the
637 		 * BAR 1 hole).
638 		 */
639 		octeon_npi_write32(CVMX_NPI_PCI_CFG06, 2ul << 30);
640 		octeon_npi_write32(CVMX_NPI_PCI_CFG07, 0);
641 
642 		/* Devices go after BAR1 */
643 		octeon_pci_mem_resource.start =
644 			OCTEON_PCI_MEMSPACE_OFFSET + (4ul << 30) -
645 			(OCTEON_PCI_BAR1_HOLE_SIZE << 20);
646 		octeon_pci_mem_resource.end =
647 			octeon_pci_mem_resource.start + (1ul << 30);
648 	} else {
649 		/* Remap the Octeon BAR 0 to map 128MB-(128MB+4KB) */
650 		octeon_npi_write32(CVMX_NPI_PCI_CFG04, 128ul << 20);
651 		octeon_npi_write32(CVMX_NPI_PCI_CFG05, 0);
652 
653 		/* Remap the Octeon BAR 1 to map 0-128MB */
654 		octeon_npi_write32(CVMX_NPI_PCI_CFG06, 0);
655 		octeon_npi_write32(CVMX_NPI_PCI_CFG07, 0);
656 
657 		/* Devices go after BAR0 */
658 		octeon_pci_mem_resource.start =
659 			OCTEON_PCI_MEMSPACE_OFFSET + (128ul << 20) +
660 			(4ul << 10);
661 		octeon_pci_mem_resource.end =
662 			octeon_pci_mem_resource.start + (1ul << 30);
663 	}
664 
665 	register_pci_controller(&octeon_pci_controller);
666 
667 	/*
668 	 * Clear any errors that might be pending from before the bus
669 	 * was setup properly.
670 	 */
671 	cvmx_write_csr(CVMX_NPI_PCI_INT_SUM2, -1);
672 	return 0;
673 }
674 
675 arch_initcall(octeon_pci_setup);
676