xref: /openbmc/linux/arch/mips/pci/pci-octeon.c (revision 791d3ef2)
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 #include <linux/platform_device.h>
15 #include <linux/swiotlb.h>
16 
17 #include <asm/time.h>
18 
19 #include <asm/octeon/octeon.h>
20 #include <asm/octeon/cvmx-npi-defs.h>
21 #include <asm/octeon/cvmx-pci-defs.h>
22 #include <asm/octeon/pci-octeon.h>
23 
24 #include <dma-coherence.h>
25 
26 #define USE_OCTEON_INTERNAL_ARBITER
27 
28 /*
29  * Octeon's PCI controller uses did=3, subdid=2 for PCI IO
30  * addresses. Use PCI endian swapping 1 so no address swapping is
31  * necessary. The Linux io routines will endian swap the data.
32  */
33 #define OCTEON_PCI_IOSPACE_BASE	    0x80011a0400000000ull
34 #define OCTEON_PCI_IOSPACE_SIZE	    (1ull<<32)
35 
36 /* Octeon't PCI controller uses did=3, subdid=3 for PCI memory. */
37 #define OCTEON_PCI_MEMSPACE_OFFSET  (0x00011b0000000000ull)
38 
39 u64 octeon_bar1_pci_phys;
40 
41 /**
42  * This is the bit decoding used for the Octeon PCI controller addresses
43  */
44 union octeon_pci_address {
45 	uint64_t u64;
46 	struct {
47 		uint64_t upper:2;
48 		uint64_t reserved:13;
49 		uint64_t io:1;
50 		uint64_t did:5;
51 		uint64_t subdid:3;
52 		uint64_t reserved2:4;
53 		uint64_t endian_swap:2;
54 		uint64_t reserved3:10;
55 		uint64_t bus:8;
56 		uint64_t dev:5;
57 		uint64_t func:3;
58 		uint64_t reg:8;
59 	} s;
60 };
61 
62 int (*octeon_pcibios_map_irq)(const struct pci_dev *dev, u8 slot, u8 pin);
63 enum octeon_dma_bar_type octeon_dma_bar_type = OCTEON_DMA_BAR_TYPE_INVALID;
64 
65 /**
66  * Map a PCI device to the appropriate interrupt line
67  *
68  * @dev:    The Linux PCI device structure for the device to map
69  * @slot:   The slot number for this device on __BUS 0__. Linux
70  *		 enumerates through all the bridges and figures out the
71  *		 slot on Bus 0 where this device eventually hooks to.
72  * @pin:    The PCI interrupt pin read from the device, then swizzled
73  *		 as it goes through each bridge.
74  * Returns Interrupt number for the device
75  */
76 int pcibios_map_irq(const struct pci_dev *dev, u8 slot, u8 pin)
77 {
78 	if (octeon_pcibios_map_irq)
79 		return octeon_pcibios_map_irq(dev, slot, pin);
80 	else
81 		panic("octeon_pcibios_map_irq not set.");
82 }
83 
84 
85 /*
86  * Called to perform platform specific PCI setup
87  */
88 int pcibios_plat_dev_init(struct pci_dev *dev)
89 {
90 	uint16_t config;
91 	uint32_t dconfig;
92 	int pos;
93 	/*
94 	 * Force the Cache line setting to 64 bytes. The standard
95 	 * Linux bus scan doesn't seem to set it. Octeon really has
96 	 * 128 byte lines, but Intel bridges get really upset if you
97 	 * try and set values above 64 bytes. Value is specified in
98 	 * 32bit words.
99 	 */
100 	pci_write_config_byte(dev, PCI_CACHE_LINE_SIZE, 64 / 4);
101 	/* Set latency timers for all devices */
102 	pci_write_config_byte(dev, PCI_LATENCY_TIMER, 64);
103 
104 	/* Enable reporting System errors and parity errors on all devices */
105 	/* Enable parity checking and error reporting */
106 	pci_read_config_word(dev, PCI_COMMAND, &config);
107 	config |= PCI_COMMAND_PARITY | PCI_COMMAND_SERR;
108 	pci_write_config_word(dev, PCI_COMMAND, config);
109 
110 	if (dev->subordinate) {
111 		/* Set latency timers on sub bridges */
112 		pci_write_config_byte(dev, PCI_SEC_LATENCY_TIMER, 64);
113 		/* More bridge error detection */
114 		pci_read_config_word(dev, PCI_BRIDGE_CONTROL, &config);
115 		config |= PCI_BRIDGE_CTL_PARITY | PCI_BRIDGE_CTL_SERR;
116 		pci_write_config_word(dev, PCI_BRIDGE_CONTROL, config);
117 	}
118 
119 	/* Enable the PCIe normal error reporting */
120 	config = PCI_EXP_DEVCTL_CERE; /* Correctable Error Reporting */
121 	config |= PCI_EXP_DEVCTL_NFERE; /* Non-Fatal Error Reporting */
122 	config |= PCI_EXP_DEVCTL_FERE;	/* Fatal Error Reporting */
123 	config |= PCI_EXP_DEVCTL_URRE;	/* Unsupported Request */
124 	pcie_capability_set_word(dev, PCI_EXP_DEVCTL, config);
125 
126 	/* Find the Advanced Error Reporting capability */
127 	pos = pci_find_ext_capability(dev, PCI_EXT_CAP_ID_ERR);
128 	if (pos) {
129 		/* Clear Uncorrectable Error Status */
130 		pci_read_config_dword(dev, pos + PCI_ERR_UNCOR_STATUS,
131 				      &dconfig);
132 		pci_write_config_dword(dev, pos + PCI_ERR_UNCOR_STATUS,
133 				       dconfig);
134 		/* Enable reporting of all uncorrectable errors */
135 		/* Uncorrectable Error Mask - turned on bits disable errors */
136 		pci_write_config_dword(dev, pos + PCI_ERR_UNCOR_MASK, 0);
137 		/*
138 		 * Leave severity at HW default. This only controls if
139 		 * errors are reported as uncorrectable or
140 		 * correctable, not if the error is reported.
141 		 */
142 		/* PCI_ERR_UNCOR_SEVER - Uncorrectable Error Severity */
143 		/* Clear Correctable Error Status */
144 		pci_read_config_dword(dev, pos + PCI_ERR_COR_STATUS, &dconfig);
145 		pci_write_config_dword(dev, pos + PCI_ERR_COR_STATUS, dconfig);
146 		/* Enable reporting of all correctable errors */
147 		/* Correctable Error Mask - turned on bits disable errors */
148 		pci_write_config_dword(dev, pos + PCI_ERR_COR_MASK, 0);
149 		/* Advanced Error Capabilities */
150 		pci_read_config_dword(dev, pos + PCI_ERR_CAP, &dconfig);
151 		/* ECRC Generation Enable */
152 		if (config & PCI_ERR_CAP_ECRC_GENC)
153 			config |= PCI_ERR_CAP_ECRC_GENE;
154 		/* ECRC Check Enable */
155 		if (config & PCI_ERR_CAP_ECRC_CHKC)
156 			config |= PCI_ERR_CAP_ECRC_CHKE;
157 		pci_write_config_dword(dev, pos + PCI_ERR_CAP, dconfig);
158 		/* PCI_ERR_HEADER_LOG - Header Log Register (16 bytes) */
159 		/* Report all errors to the root complex */
160 		pci_write_config_dword(dev, pos + PCI_ERR_ROOT_COMMAND,
161 				       PCI_ERR_ROOT_CMD_COR_EN |
162 				       PCI_ERR_ROOT_CMD_NONFATAL_EN |
163 				       PCI_ERR_ROOT_CMD_FATAL_EN);
164 		/* Clear the Root status register */
165 		pci_read_config_dword(dev, pos + PCI_ERR_ROOT_STATUS, &dconfig);
166 		pci_write_config_dword(dev, pos + PCI_ERR_ROOT_STATUS, dconfig);
167 	}
168 
169 	dev->dev.dma_ops = octeon_pci_dma_map_ops;
170 
171 	return 0;
172 }
173 
174 /**
175  * Return the mapping of PCI device number to IRQ line. Each
176  * character in the return string represents the interrupt
177  * line for the device at that position. Device 1 maps to the
178  * first character, etc. The characters A-D are used for PCI
179  * interrupts.
180  *
181  * Returns PCI interrupt mapping
182  */
183 const char *octeon_get_pci_interrupts(void)
184 {
185 	/*
186 	 * Returning an empty string causes the interrupts to be
187 	 * routed based on the PCI specification. From the PCI spec:
188 	 *
189 	 * INTA# of Device Number 0 is connected to IRQW on the system
190 	 * board.  (Device Number has no significance regarding being
191 	 * located on the system board or in a connector.) INTA# of
192 	 * Device Number 1 is connected to IRQX on the system
193 	 * board. INTA# of Device Number 2 is connected to IRQY on the
194 	 * system board. INTA# of Device Number 3 is connected to IRQZ
195 	 * on the system board. The table below describes how each
196 	 * agent's INTx# lines are connected to the system board
197 	 * interrupt lines. The following equation can be used to
198 	 * determine to which INTx# signal on the system board a given
199 	 * device's INTx# line(s) is connected.
200 	 *
201 	 * MB = (D + I) MOD 4 MB = System board Interrupt (IRQW = 0,
202 	 * IRQX = 1, IRQY = 2, and IRQZ = 3) D = Device Number I =
203 	 * Interrupt Number (INTA# = 0, INTB# = 1, INTC# = 2, and
204 	 * INTD# = 3)
205 	 */
206 	if (of_machine_is_compatible("dlink,dsr-500n"))
207 		return "CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC";
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_EBH3100:
213 	case CVMX_BOARD_TYPE_CN3010_EVB_HS5:
214 	case CVMX_BOARD_TYPE_CN3005_EVB_HS5:
215 		return "AAABAAAAAAAAAAAAAAAAAAAAAAAAAAAA";
216 	case CVMX_BOARD_TYPE_BBGW_REF:
217 		return "AABCD";
218 	case CVMX_BOARD_TYPE_CUST_DSR1000N:
219 		return "CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC";
220 	case CVMX_BOARD_TYPE_THUNDER:
221 	case CVMX_BOARD_TYPE_EBH3000:
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 	switch (size) {
278 	case 4:
279 		*val = le32_to_cpu(cvmx_read64_uint32(pci_addr.u64));
280 		return PCIBIOS_SUCCESSFUL;
281 	case 2:
282 		*val = le16_to_cpu(cvmx_read64_uint16(pci_addr.u64));
283 		return PCIBIOS_SUCCESSFUL;
284 	case 1:
285 		*val = cvmx_read64_uint8(pci_addr.u64);
286 		return PCIBIOS_SUCCESSFUL;
287 	}
288 	return PCIBIOS_FUNC_NOT_SUPPORTED;
289 }
290 
291 
292 /*
293  * Write a value to PCI configuration space
294  */
295 static int octeon_write_config(struct pci_bus *bus, unsigned int devfn,
296 			       int reg, int size, u32 val)
297 {
298 	union octeon_pci_address pci_addr;
299 
300 	pci_addr.u64 = 0;
301 	pci_addr.s.upper = 2;
302 	pci_addr.s.io = 1;
303 	pci_addr.s.did = 3;
304 	pci_addr.s.subdid = 1;
305 	pci_addr.s.endian_swap = 1;
306 	pci_addr.s.bus = bus->number;
307 	pci_addr.s.dev = devfn >> 3;
308 	pci_addr.s.func = devfn & 0x7;
309 	pci_addr.s.reg = reg;
310 
311 	switch (size) {
312 	case 4:
313 		cvmx_write64_uint32(pci_addr.u64, cpu_to_le32(val));
314 		return PCIBIOS_SUCCESSFUL;
315 	case 2:
316 		cvmx_write64_uint16(pci_addr.u64, cpu_to_le16(val));
317 		return PCIBIOS_SUCCESSFUL;
318 	case 1:
319 		cvmx_write64_uint8(pci_addr.u64, val);
320 		return PCIBIOS_SUCCESSFUL;
321 	}
322 	return PCIBIOS_FUNC_NOT_SUPPORTED;
323 }
324 
325 
326 static struct pci_ops octeon_pci_ops = {
327 	.read	= octeon_read_config,
328 	.write	= octeon_write_config,
329 };
330 
331 static struct resource octeon_pci_mem_resource = {
332 	.start = 0,
333 	.end = 0,
334 	.name = "Octeon PCI MEM",
335 	.flags = IORESOURCE_MEM,
336 };
337 
338 /*
339  * PCI ports must be above 16KB so the ISA bus filtering in the PCI-X to PCI
340  * bridge
341  */
342 static struct resource octeon_pci_io_resource = {
343 	.start = 0x4000,
344 	.end = OCTEON_PCI_IOSPACE_SIZE - 1,
345 	.name = "Octeon PCI IO",
346 	.flags = IORESOURCE_IO,
347 };
348 
349 static struct pci_controller octeon_pci_controller = {
350 	.pci_ops = &octeon_pci_ops,
351 	.mem_resource = &octeon_pci_mem_resource,
352 	.mem_offset = OCTEON_PCI_MEMSPACE_OFFSET,
353 	.io_resource = &octeon_pci_io_resource,
354 	.io_offset = 0,
355 	.io_map_base = OCTEON_PCI_IOSPACE_BASE,
356 };
357 
358 
359 /*
360  * Low level initialize the Octeon PCI controller
361  */
362 static void octeon_pci_initialize(void)
363 {
364 	union cvmx_pci_cfg01 cfg01;
365 	union cvmx_npi_ctl_status ctl_status;
366 	union cvmx_pci_ctl_status_2 ctl_status_2;
367 	union cvmx_pci_cfg19 cfg19;
368 	union cvmx_pci_cfg16 cfg16;
369 	union cvmx_pci_cfg22 cfg22;
370 	union cvmx_pci_cfg56 cfg56;
371 
372 	/* Reset the PCI Bus */
373 	cvmx_write_csr(CVMX_CIU_SOFT_PRST, 0x1);
374 	cvmx_read_csr(CVMX_CIU_SOFT_PRST);
375 
376 	udelay(2000);		/* Hold PCI reset for 2 ms */
377 
378 	ctl_status.u64 = 0;	/* cvmx_read_csr(CVMX_NPI_CTL_STATUS); */
379 	ctl_status.s.max_word = 1;
380 	ctl_status.s.timer = 1;
381 	cvmx_write_csr(CVMX_NPI_CTL_STATUS, ctl_status.u64);
382 
383 	/* Deassert PCI reset and advertize PCX Host Mode Device Capability
384 	   (64b) */
385 	cvmx_write_csr(CVMX_CIU_SOFT_PRST, 0x4);
386 	cvmx_read_csr(CVMX_CIU_SOFT_PRST);
387 
388 	udelay(2000);		/* Wait 2 ms after deasserting PCI reset */
389 
390 	ctl_status_2.u32 = 0;
391 	ctl_status_2.s.tsr_hwm = 1;	/* Initializes to 0.  Must be set
392 					   before any PCI reads. */
393 	ctl_status_2.s.bar2pres = 1;	/* Enable BAR2 */
394 	ctl_status_2.s.bar2_enb = 1;
395 	ctl_status_2.s.bar2_cax = 1;	/* Don't use L2 */
396 	ctl_status_2.s.bar2_esx = 1;
397 	ctl_status_2.s.pmo_amod = 1;	/* Round robin priority */
398 	if (octeon_dma_bar_type == OCTEON_DMA_BAR_TYPE_BIG) {
399 		/* BAR1 hole */
400 		ctl_status_2.s.bb1_hole = OCTEON_PCI_BAR1_HOLE_BITS;
401 		ctl_status_2.s.bb1_siz = 1;  /* BAR1 is 2GB */
402 		ctl_status_2.s.bb_ca = 1;    /* Don't use L2 with big bars */
403 		ctl_status_2.s.bb_es = 1;    /* Big bar in byte swap mode */
404 		ctl_status_2.s.bb1 = 1;	     /* BAR1 is big */
405 		ctl_status_2.s.bb0 = 1;	     /* BAR0 is big */
406 	}
407 
408 	octeon_npi_write32(CVMX_NPI_PCI_CTL_STATUS_2, ctl_status_2.u32);
409 	udelay(2000);		/* Wait 2 ms before doing PCI reads */
410 
411 	ctl_status_2.u32 = octeon_npi_read32(CVMX_NPI_PCI_CTL_STATUS_2);
412 	pr_notice("PCI Status: %s %s-bit\n",
413 		  ctl_status_2.s.ap_pcix ? "PCI-X" : "PCI",
414 		  ctl_status_2.s.ap_64ad ? "64" : "32");
415 
416 	if (OCTEON_IS_MODEL(OCTEON_CN58XX) || OCTEON_IS_MODEL(OCTEON_CN50XX)) {
417 		union cvmx_pci_cnt_reg cnt_reg_start;
418 		union cvmx_pci_cnt_reg cnt_reg_end;
419 		unsigned long cycles, pci_clock;
420 
421 		cnt_reg_start.u64 = cvmx_read_csr(CVMX_NPI_PCI_CNT_REG);
422 		cycles = read_c0_cvmcount();
423 		udelay(1000);
424 		cnt_reg_end.u64 = cvmx_read_csr(CVMX_NPI_PCI_CNT_REG);
425 		cycles = read_c0_cvmcount() - cycles;
426 		pci_clock = (cnt_reg_end.s.pcicnt - cnt_reg_start.s.pcicnt) /
427 			    (cycles / (mips_hpt_frequency / 1000000));
428 		pr_notice("PCI Clock: %lu MHz\n", pci_clock);
429 	}
430 
431 	/*
432 	 * TDOMC must be set to one in PCI mode. TDOMC should be set to 4
433 	 * in PCI-X mode to allow four outstanding splits. Otherwise,
434 	 * should not change from its reset value. Don't write PCI_CFG19
435 	 * in PCI mode (0x82000001 reset value), write it to 0x82000004
436 	 * after PCI-X mode is known. MRBCI,MDWE,MDRE -> must be zero.
437 	 * MRBCM -> must be one.
438 	 */
439 	if (ctl_status_2.s.ap_pcix) {
440 		cfg19.u32 = 0;
441 		/*
442 		 * Target Delayed/Split request outstanding maximum
443 		 * count. [1..31] and 0=32.  NOTE: If the user
444 		 * programs these bits beyond the Designed Maximum
445 		 * outstanding count, then the designed maximum table
446 		 * depth will be used instead.	No additional
447 		 * Deferred/Split transactions will be accepted if
448 		 * this outstanding maximum count is
449 		 * reached. Furthermore, no additional deferred/split
450 		 * transactions will be accepted if the I/O delay/ I/O
451 		 * Split Request outstanding maximum is reached.
452 		 */
453 		cfg19.s.tdomc = 4;
454 		/*
455 		 * Master Deferred Read Request Outstanding Max Count
456 		 * (PCI only).	CR4C[26:24] Max SAC cycles MAX DAC
457 		 * cycles 000 8 4 001 1 0 010 2 1 011 3 1 100 4 2 101
458 		 * 5 2 110 6 3 111 7 3 For example, if these bits are
459 		 * programmed to 100, the core can support 2 DAC
460 		 * cycles, 4 SAC cycles or a combination of 1 DAC and
461 		 * 2 SAC cycles. NOTE: For the PCI-X maximum
462 		 * outstanding split transactions, refer to
463 		 * CRE0[22:20].
464 		 */
465 		cfg19.s.mdrrmc = 2;
466 		/*
467 		 * Master Request (Memory Read) Byte Count/Byte Enable
468 		 * select. 0 = Byte Enables valid. In PCI mode, a
469 		 * burst transaction cannot be performed using Memory
470 		 * Read command=4?h6. 1 = DWORD Byte Count valid
471 		 * (default). In PCI Mode, the memory read byte
472 		 * enables are automatically generated by the
473 		 * core. Note: N3 Master Request transaction sizes are
474 		 * always determined through the
475 		 * am_attr[<35:32>|<7:0>] field.
476 		 */
477 		cfg19.s.mrbcm = 1;
478 		octeon_npi_write32(CVMX_NPI_PCI_CFG19, cfg19.u32);
479 	}
480 
481 
482 	cfg01.u32 = 0;
483 	cfg01.s.msae = 1;	/* Memory Space Access Enable */
484 	cfg01.s.me = 1;		/* Master Enable */
485 	cfg01.s.pee = 1;	/* PERR# Enable */
486 	cfg01.s.see = 1;	/* System Error Enable */
487 	cfg01.s.fbbe = 1;	/* Fast Back to Back Transaction Enable */
488 
489 	octeon_npi_write32(CVMX_NPI_PCI_CFG01, cfg01.u32);
490 
491 #ifdef USE_OCTEON_INTERNAL_ARBITER
492 	/*
493 	 * When OCTEON is a PCI host, most systems will use OCTEON's
494 	 * internal arbiter, so must enable it before any PCI/PCI-X
495 	 * traffic can occur.
496 	 */
497 	{
498 		union cvmx_npi_pci_int_arb_cfg pci_int_arb_cfg;
499 
500 		pci_int_arb_cfg.u64 = 0;
501 		pci_int_arb_cfg.s.en = 1;	/* Internal arbiter enable */
502 		cvmx_write_csr(CVMX_NPI_PCI_INT_ARB_CFG, pci_int_arb_cfg.u64);
503 	}
504 #endif	/* USE_OCTEON_INTERNAL_ARBITER */
505 
506 	/*
507 	 * Preferably written to 1 to set MLTD. [RDSATI,TRTAE,
508 	 * TWTAE,TMAE,DPPMR -> must be zero. TILT -> must not be set to
509 	 * 1..7.
510 	 */
511 	cfg16.u32 = 0;
512 	cfg16.s.mltd = 1;	/* Master Latency Timer Disable */
513 	octeon_npi_write32(CVMX_NPI_PCI_CFG16, cfg16.u32);
514 
515 	/*
516 	 * Should be written to 0x4ff00. MTTV -> must be zero.
517 	 * FLUSH -> must be 1. MRV -> should be 0xFF.
518 	 */
519 	cfg22.u32 = 0;
520 	/* Master Retry Value [1..255] and 0=infinite */
521 	cfg22.s.mrv = 0xff;
522 	/*
523 	 * AM_DO_FLUSH_I control NOTE: This bit MUST BE ONE for proper
524 	 * N3K operation.
525 	 */
526 	cfg22.s.flush = 1;
527 	octeon_npi_write32(CVMX_NPI_PCI_CFG22, cfg22.u32);
528 
529 	/*
530 	 * MOST Indicates the maximum number of outstanding splits (in -1
531 	 * notation) when OCTEON is in PCI-X mode.  PCI-X performance is
532 	 * affected by the MOST selection.  Should generally be written
533 	 * with one of 0x3be807, 0x2be807, 0x1be807, or 0x0be807,
534 	 * depending on the desired MOST of 3, 2, 1, or 0, respectively.
535 	 */
536 	cfg56.u32 = 0;
537 	cfg56.s.pxcid = 7;	/* RO - PCI-X Capability ID */
538 	cfg56.s.ncp = 0xe8;	/* RO - Next Capability Pointer */
539 	cfg56.s.dpere = 1;	/* Data Parity Error Recovery Enable */
540 	cfg56.s.roe = 1;	/* Relaxed Ordering Enable */
541 	cfg56.s.mmbc = 1;	/* Maximum Memory Byte Count
542 				   [0=512B,1=1024B,2=2048B,3=4096B] */
543 	cfg56.s.most = 3;	/* Maximum outstanding Split transactions [0=1
544 				   .. 7=32] */
545 
546 	octeon_npi_write32(CVMX_NPI_PCI_CFG56, cfg56.u32);
547 
548 	/*
549 	 * Affects PCI performance when OCTEON services reads to its
550 	 * BAR1/BAR2. Refer to Section 10.6.1.	The recommended values are
551 	 * 0x22, 0x33, and 0x33 for PCI_READ_CMD_6, PCI_READ_CMD_C, and
552 	 * PCI_READ_CMD_E, respectively. Unfortunately due to errata DDR-700,
553 	 * these values need to be changed so they won't possibly prefetch off
554 	 * of the end of memory if PCI is DMAing a buffer at the end of
555 	 * memory. Note that these values differ from their reset values.
556 	 */
557 	octeon_npi_write32(CVMX_NPI_PCI_READ_CMD_6, 0x21);
558 	octeon_npi_write32(CVMX_NPI_PCI_READ_CMD_C, 0x31);
559 	octeon_npi_write32(CVMX_NPI_PCI_READ_CMD_E, 0x31);
560 }
561 
562 
563 /*
564  * Initialize the Octeon PCI controller
565  */
566 static int __init octeon_pci_setup(void)
567 {
568 	union cvmx_npi_mem_access_subidx mem_access;
569 	int index;
570 
571 	/* Only these chips have PCI */
572 	if (octeon_has_feature(OCTEON_FEATURE_PCIE))
573 		return 0;
574 
575 	/* Point pcibios_map_irq() to the PCI version of it */
576 	octeon_pcibios_map_irq = octeon_pci_pcibios_map_irq;
577 
578 	/* Only use the big bars on chips that support it */
579 	if (OCTEON_IS_MODEL(OCTEON_CN31XX) ||
580 	    OCTEON_IS_MODEL(OCTEON_CN38XX_PASS2) ||
581 	    OCTEON_IS_MODEL(OCTEON_CN38XX_PASS1))
582 		octeon_dma_bar_type = OCTEON_DMA_BAR_TYPE_SMALL;
583 	else
584 		octeon_dma_bar_type = OCTEON_DMA_BAR_TYPE_BIG;
585 
586 	if (!octeon_is_pci_host()) {
587 		pr_notice("Not in host mode, PCI Controller not initialized\n");
588 		return 0;
589 	}
590 
591 	/* PCI I/O and PCI MEM values */
592 	set_io_port_base(OCTEON_PCI_IOSPACE_BASE);
593 	ioport_resource.start = 0;
594 	ioport_resource.end = OCTEON_PCI_IOSPACE_SIZE - 1;
595 
596 	pr_notice("%s Octeon big bar support\n",
597 		  (octeon_dma_bar_type ==
598 		  OCTEON_DMA_BAR_TYPE_BIG) ? "Enabling" : "Disabling");
599 
600 	octeon_pci_initialize();
601 
602 	mem_access.u64 = 0;
603 	mem_access.s.esr = 1;	/* Endian-Swap on read. */
604 	mem_access.s.esw = 1;	/* Endian-Swap on write. */
605 	mem_access.s.nsr = 0;	/* No-Snoop on read. */
606 	mem_access.s.nsw = 0;	/* No-Snoop on write. */
607 	mem_access.s.ror = 0;	/* Relax Read on read. */
608 	mem_access.s.row = 0;	/* Relax Order on write. */
609 	mem_access.s.ba = 0;	/* PCI Address bits [63:36]. */
610 	cvmx_write_csr(CVMX_NPI_MEM_ACCESS_SUBID3, mem_access.u64);
611 
612 	/*
613 	 * Remap the Octeon BAR 2 above all 32 bit devices
614 	 * (0x8000000000ul).  This is done here so it is remapped
615 	 * before the readl()'s below. We don't want BAR2 overlapping
616 	 * with BAR0/BAR1 during these reads.
617 	 */
618 	octeon_npi_write32(CVMX_NPI_PCI_CFG08,
619 			   (u32)(OCTEON_BAR2_PCI_ADDRESS & 0xffffffffull));
620 	octeon_npi_write32(CVMX_NPI_PCI_CFG09,
621 			   (u32)(OCTEON_BAR2_PCI_ADDRESS >> 32));
622 
623 	if (octeon_dma_bar_type == OCTEON_DMA_BAR_TYPE_BIG) {
624 		/* Remap the Octeon BAR 0 to 0-2GB */
625 		octeon_npi_write32(CVMX_NPI_PCI_CFG04, 0);
626 		octeon_npi_write32(CVMX_NPI_PCI_CFG05, 0);
627 
628 		/*
629 		 * Remap the Octeon BAR 1 to map 2GB-4GB (minus the
630 		 * BAR 1 hole).
631 		 */
632 		octeon_npi_write32(CVMX_NPI_PCI_CFG06, 2ul << 30);
633 		octeon_npi_write32(CVMX_NPI_PCI_CFG07, 0);
634 
635 		/* BAR1 movable mappings set for identity mapping */
636 		octeon_bar1_pci_phys = 0x80000000ull;
637 		for (index = 0; index < 32; index++) {
638 			union cvmx_pci_bar1_indexx bar1_index;
639 
640 			bar1_index.u32 = 0;
641 			/* Address bits[35:22] sent to L2C */
642 			bar1_index.s.addr_idx =
643 				(octeon_bar1_pci_phys >> 22) + index;
644 			/* Don't put PCI accesses in L2. */
645 			bar1_index.s.ca = 1;
646 			/* Endian Swap Mode */
647 			bar1_index.s.end_swp = 1;
648 			/* Set '1' when the selected address range is valid. */
649 			bar1_index.s.addr_v = 1;
650 			octeon_npi_write32(CVMX_NPI_PCI_BAR1_INDEXX(index),
651 					   bar1_index.u32);
652 		}
653 
654 		/* Devices go after BAR1 */
655 		octeon_pci_mem_resource.start =
656 			OCTEON_PCI_MEMSPACE_OFFSET + (4ul << 30) -
657 			(OCTEON_PCI_BAR1_HOLE_SIZE << 20);
658 		octeon_pci_mem_resource.end =
659 			octeon_pci_mem_resource.start + (1ul << 30);
660 	} else {
661 		/* Remap the Octeon BAR 0 to map 128MB-(128MB+4KB) */
662 		octeon_npi_write32(CVMX_NPI_PCI_CFG04, 128ul << 20);
663 		octeon_npi_write32(CVMX_NPI_PCI_CFG05, 0);
664 
665 		/* Remap the Octeon BAR 1 to map 0-128MB */
666 		octeon_npi_write32(CVMX_NPI_PCI_CFG06, 0);
667 		octeon_npi_write32(CVMX_NPI_PCI_CFG07, 0);
668 
669 		/* BAR1 movable regions contiguous to cover the swiotlb */
670 		octeon_bar1_pci_phys =
671 			virt_to_phys(octeon_swiotlb) & ~((1ull << 22) - 1);
672 
673 		for (index = 0; index < 32; index++) {
674 			union cvmx_pci_bar1_indexx bar1_index;
675 
676 			bar1_index.u32 = 0;
677 			/* Address bits[35:22] sent to L2C */
678 			bar1_index.s.addr_idx =
679 				(octeon_bar1_pci_phys >> 22) + index;
680 			/* Don't put PCI accesses in L2. */
681 			bar1_index.s.ca = 1;
682 			/* Endian Swap Mode */
683 			bar1_index.s.end_swp = 1;
684 			/* Set '1' when the selected address range is valid. */
685 			bar1_index.s.addr_v = 1;
686 			octeon_npi_write32(CVMX_NPI_PCI_BAR1_INDEXX(index),
687 					   bar1_index.u32);
688 		}
689 
690 		/* Devices go after BAR0 */
691 		octeon_pci_mem_resource.start =
692 			OCTEON_PCI_MEMSPACE_OFFSET + (128ul << 20) +
693 			(4ul << 10);
694 		octeon_pci_mem_resource.end =
695 			octeon_pci_mem_resource.start + (1ul << 30);
696 	}
697 
698 	register_pci_controller(&octeon_pci_controller);
699 
700 	/*
701 	 * Clear any errors that might be pending from before the bus
702 	 * was setup properly.
703 	 */
704 	cvmx_write_csr(CVMX_NPI_PCI_INT_SUM2, -1);
705 
706 	if (IS_ERR(platform_device_register_simple("octeon_pci_edac",
707 						   -1, NULL, 0)))
708 		pr_err("Registration of co_pci_edac failed!\n");
709 
710 	octeon_pci_dma_init();
711 
712 	return 0;
713 }
714 
715 arch_initcall(octeon_pci_setup);
716