xref: /linux/arch/mips/pci/pci-octeon.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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 __initconst (*octeon_pcibios_map_irq)(const struct pci_dev *dev,
63 					 u8 slot, u8 pin);
64 enum octeon_dma_bar_type octeon_dma_bar_type = OCTEON_DMA_BAR_TYPE_INVALID;
65 
66 /**
67  * Map a PCI device to the appropriate interrupt line
68  *
69  * @dev:    The Linux PCI device structure for the device to map
70  * @slot:   The slot number for this device on __BUS 0__. Linux
71  *		 enumerates through all the bridges and figures out the
72  *		 slot on Bus 0 where this device eventually hooks to.
73  * @pin:    The PCI interrupt pin read from the device, then swizzled
74  *		 as it goes through each bridge.
75  * Returns Interrupt number for the device
76  */
77 int __init pcibios_map_irq(const struct pci_dev *dev, u8 slot, u8 pin)
78 {
79 	if (octeon_pcibios_map_irq)
80 		return octeon_pcibios_map_irq(dev, slot, pin);
81 	else
82 		panic("octeon_pcibios_map_irq not set.");
83 }
84 
85 
86 /*
87  * Called to perform platform specific PCI setup
88  */
89 int pcibios_plat_dev_init(struct pci_dev *dev)
90 {
91 	uint16_t config;
92 	uint32_t dconfig;
93 	int pos;
94 	/*
95 	 * Force the Cache line setting to 64 bytes. The standard
96 	 * Linux bus scan doesn't seem to set it. Octeon really has
97 	 * 128 byte lines, but Intel bridges get really upset if you
98 	 * try and set values above 64 bytes. Value is specified in
99 	 * 32bit words.
100 	 */
101 	pci_write_config_byte(dev, PCI_CACHE_LINE_SIZE, 64 / 4);
102 	/* Set latency timers for all devices */
103 	pci_write_config_byte(dev, PCI_LATENCY_TIMER, 64);
104 
105 	/* Enable reporting System errors and parity errors on all devices */
106 	/* Enable parity checking and error reporting */
107 	pci_read_config_word(dev, PCI_COMMAND, &config);
108 	config |= PCI_COMMAND_PARITY | PCI_COMMAND_SERR;
109 	pci_write_config_word(dev, PCI_COMMAND, config);
110 
111 	if (dev->subordinate) {
112 		/* Set latency timers on sub bridges */
113 		pci_write_config_byte(dev, PCI_SEC_LATENCY_TIMER, 64);
114 		/* More bridge error detection */
115 		pci_read_config_word(dev, PCI_BRIDGE_CONTROL, &config);
116 		config |= PCI_BRIDGE_CTL_PARITY | PCI_BRIDGE_CTL_SERR;
117 		pci_write_config_word(dev, PCI_BRIDGE_CONTROL, config);
118 	}
119 
120 	/* Enable the PCIe normal error reporting */
121 	config = PCI_EXP_DEVCTL_CERE; /* Correctable Error Reporting */
122 	config |= PCI_EXP_DEVCTL_NFERE; /* Non-Fatal Error Reporting */
123 	config |= PCI_EXP_DEVCTL_FERE;	/* Fatal Error Reporting */
124 	config |= PCI_EXP_DEVCTL_URRE;	/* Unsupported Request */
125 	pcie_capability_set_word(dev, PCI_EXP_DEVCTL, config);
126 
127 	/* Find the Advanced Error Reporting capability */
128 	pos = pci_find_ext_capability(dev, PCI_EXT_CAP_ID_ERR);
129 	if (pos) {
130 		/* Clear Uncorrectable Error Status */
131 		pci_read_config_dword(dev, pos + PCI_ERR_UNCOR_STATUS,
132 				      &dconfig);
133 		pci_write_config_dword(dev, pos + PCI_ERR_UNCOR_STATUS,
134 				       dconfig);
135 		/* Enable reporting of all uncorrectable errors */
136 		/* Uncorrectable Error Mask - turned on bits disable errors */
137 		pci_write_config_dword(dev, pos + PCI_ERR_UNCOR_MASK, 0);
138 		/*
139 		 * Leave severity at HW default. This only controls if
140 		 * errors are reported as uncorrectable or
141 		 * correctable, not if the error is reported.
142 		 */
143 		/* PCI_ERR_UNCOR_SEVER - Uncorrectable Error Severity */
144 		/* Clear Correctable Error Status */
145 		pci_read_config_dword(dev, pos + PCI_ERR_COR_STATUS, &dconfig);
146 		pci_write_config_dword(dev, pos + PCI_ERR_COR_STATUS, dconfig);
147 		/* Enable reporting of all correctable errors */
148 		/* Correctable Error Mask - turned on bits disable errors */
149 		pci_write_config_dword(dev, pos + PCI_ERR_COR_MASK, 0);
150 		/* Advanced Error Capabilities */
151 		pci_read_config_dword(dev, pos + PCI_ERR_CAP, &dconfig);
152 		/* ECRC Generation Enable */
153 		if (config & PCI_ERR_CAP_ECRC_GENC)
154 			config |= PCI_ERR_CAP_ECRC_GENE;
155 		/* ECRC Check Enable */
156 		if (config & PCI_ERR_CAP_ECRC_CHKC)
157 			config |= PCI_ERR_CAP_ECRC_CHKE;
158 		pci_write_config_dword(dev, pos + PCI_ERR_CAP, dconfig);
159 		/* PCI_ERR_HEADER_LOG - Header Log Register (16 bytes) */
160 		/* Report all errors to the root complex */
161 		pci_write_config_dword(dev, pos + PCI_ERR_ROOT_COMMAND,
162 				       PCI_ERR_ROOT_CMD_COR_EN |
163 				       PCI_ERR_ROOT_CMD_NONFATAL_EN |
164 				       PCI_ERR_ROOT_CMD_FATAL_EN);
165 		/* Clear the Root status register */
166 		pci_read_config_dword(dev, pos + PCI_ERR_ROOT_STATUS, &dconfig);
167 		pci_write_config_dword(dev, pos + PCI_ERR_ROOT_STATUS, dconfig);
168 	}
169 
170 	dev->dev.archdata.dma_ops = octeon_pci_dma_map_ops;
171 
172 	return 0;
173 }
174 
175 /**
176  * Return the mapping of PCI device number to IRQ line. Each
177  * character in the return string represents the interrupt
178  * line for the device at that position. Device 1 maps to the
179  * first character, etc. The characters A-D are used for PCI
180  * interrupts.
181  *
182  * Returns PCI interrupt mapping
183  */
184 const char *octeon_get_pci_interrupts(void)
185 {
186 	/*
187 	 * Returning an empty string causes the interrupts to be
188 	 * routed based on the PCI specification. From the PCI spec:
189 	 *
190 	 * INTA# of Device Number 0 is connected to IRQW on the system
191 	 * board.  (Device Number has no significance regarding being
192 	 * located on the system board or in a connector.) INTA# of
193 	 * Device Number 1 is connected to IRQX on the system
194 	 * board. INTA# of Device Number 2 is connected to IRQY on the
195 	 * system board. INTA# of Device Number 3 is connected to IRQZ
196 	 * on the system board. The table below describes how each
197 	 * agent's INTx# lines are connected to the system board
198 	 * interrupt lines. The following equation can be used to
199 	 * determine to which INTx# signal on the system board a given
200 	 * device's INTx# line(s) is connected.
201 	 *
202 	 * MB = (D + I) MOD 4 MB = System board Interrupt (IRQW = 0,
203 	 * IRQX = 1, IRQY = 2, and IRQZ = 3) D = Device Number I =
204 	 * Interrupt Number (INTA# = 0, INTB# = 1, INTC# = 2, and
205 	 * INTD# = 3)
206 	 */
207 	switch (octeon_bootinfo->board_type) {
208 	case CVMX_BOARD_TYPE_NAO38:
209 		/* This is really the NAC38 */
210 		return "AAAAADABAAAAAAAAAAAAAAAAAAAAAAAA";
211 	case CVMX_BOARD_TYPE_EBH3100:
212 	case CVMX_BOARD_TYPE_CN3010_EVB_HS5:
213 	case CVMX_BOARD_TYPE_CN3005_EVB_HS5:
214 		return "AAABAAAAAAAAAAAAAAAAAAAAAAAAAAAA";
215 	case CVMX_BOARD_TYPE_BBGW_REF:
216 		return "AABCD";
217 	case CVMX_BOARD_TYPE_CUST_DSR1000N:
218 		return "CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC";
219 	case CVMX_BOARD_TYPE_THUNDER:
220 	case CVMX_BOARD_TYPE_EBH3000:
221 	default:
222 		return "";
223 	}
224 }
225 
226 /**
227  * Map a PCI device to the appropriate interrupt line
228  *
229  * @dev:    The Linux PCI device structure for the device to map
230  * @slot:   The slot number for this device on __BUS 0__. Linux
231  *		 enumerates through all the bridges and figures out the
232  *		 slot on Bus 0 where this device eventually hooks to.
233  * @pin:    The PCI interrupt pin read from the device, then swizzled
234  *		 as it goes through each bridge.
235  * Returns Interrupt number for the device
236  */
237 int __init octeon_pci_pcibios_map_irq(const struct pci_dev *dev,
238 				      u8 slot, u8 pin)
239 {
240 	int irq_num;
241 	const char *interrupts;
242 	int dev_num;
243 
244 	/* Get the board specific interrupt mapping */
245 	interrupts = octeon_get_pci_interrupts();
246 
247 	dev_num = dev->devfn >> 3;
248 	if (dev_num < strlen(interrupts))
249 		irq_num = ((interrupts[dev_num] - 'A' + pin - 1) & 3) +
250 			OCTEON_IRQ_PCI_INT0;
251 	else
252 		irq_num = ((slot + pin - 3) & 3) + OCTEON_IRQ_PCI_INT0;
253 	return irq_num;
254 }
255 
256 
257 /*
258  * Read a value from configuration space
259  */
260 static int octeon_read_config(struct pci_bus *bus, unsigned int devfn,
261 			      int reg, int size, u32 *val)
262 {
263 	union octeon_pci_address pci_addr;
264 
265 	pci_addr.u64 = 0;
266 	pci_addr.s.upper = 2;
267 	pci_addr.s.io = 1;
268 	pci_addr.s.did = 3;
269 	pci_addr.s.subdid = 1;
270 	pci_addr.s.endian_swap = 1;
271 	pci_addr.s.bus = bus->number;
272 	pci_addr.s.dev = devfn >> 3;
273 	pci_addr.s.func = devfn & 0x7;
274 	pci_addr.s.reg = reg;
275 
276 	switch (size) {
277 	case 4:
278 		*val = le32_to_cpu(cvmx_read64_uint32(pci_addr.u64));
279 		return PCIBIOS_SUCCESSFUL;
280 	case 2:
281 		*val = le16_to_cpu(cvmx_read64_uint16(pci_addr.u64));
282 		return PCIBIOS_SUCCESSFUL;
283 	case 1:
284 		*val = cvmx_read64_uint8(pci_addr.u64);
285 		return PCIBIOS_SUCCESSFUL;
286 	}
287 	return PCIBIOS_FUNC_NOT_SUPPORTED;
288 }
289 
290 
291 /*
292  * Write a value to PCI configuration space
293  */
294 static int octeon_write_config(struct pci_bus *bus, unsigned int devfn,
295 			       int reg, int size, u32 val)
296 {
297 	union octeon_pci_address pci_addr;
298 
299 	pci_addr.u64 = 0;
300 	pci_addr.s.upper = 2;
301 	pci_addr.s.io = 1;
302 	pci_addr.s.did = 3;
303 	pci_addr.s.subdid = 1;
304 	pci_addr.s.endian_swap = 1;
305 	pci_addr.s.bus = bus->number;
306 	pci_addr.s.dev = devfn >> 3;
307 	pci_addr.s.func = devfn & 0x7;
308 	pci_addr.s.reg = reg;
309 
310 	switch (size) {
311 	case 4:
312 		cvmx_write64_uint32(pci_addr.u64, cpu_to_le32(val));
313 		return PCIBIOS_SUCCESSFUL;
314 	case 2:
315 		cvmx_write64_uint16(pci_addr.u64, cpu_to_le16(val));
316 		return PCIBIOS_SUCCESSFUL;
317 	case 1:
318 		cvmx_write64_uint8(pci_addr.u64, val);
319 		return PCIBIOS_SUCCESSFUL;
320 	}
321 	return PCIBIOS_FUNC_NOT_SUPPORTED;
322 }
323 
324 
325 static struct pci_ops octeon_pci_ops = {
326 	.read	= octeon_read_config,
327 	.write	= octeon_write_config,
328 };
329 
330 static struct resource octeon_pci_mem_resource = {
331 	.start = 0,
332 	.end = 0,
333 	.name = "Octeon PCI MEM",
334 	.flags = IORESOURCE_MEM,
335 };
336 
337 /*
338  * PCI ports must be above 16KB so the ISA bus filtering in the PCI-X to PCI
339  * bridge
340  */
341 static struct resource octeon_pci_io_resource = {
342 	.start = 0x4000,
343 	.end = OCTEON_PCI_IOSPACE_SIZE - 1,
344 	.name = "Octeon PCI IO",
345 	.flags = IORESOURCE_IO,
346 };
347 
348 static struct pci_controller octeon_pci_controller = {
349 	.pci_ops = &octeon_pci_ops,
350 	.mem_resource = &octeon_pci_mem_resource,
351 	.mem_offset = OCTEON_PCI_MEMSPACE_OFFSET,
352 	.io_resource = &octeon_pci_io_resource,
353 	.io_offset = 0,
354 	.io_map_base = OCTEON_PCI_IOSPACE_BASE,
355 };
356 
357 
358 /*
359  * Low level initialize the Octeon PCI controller
360  */
361 static void octeon_pci_initialize(void)
362 {
363 	union cvmx_pci_cfg01 cfg01;
364 	union cvmx_npi_ctl_status ctl_status;
365 	union cvmx_pci_ctl_status_2 ctl_status_2;
366 	union cvmx_pci_cfg19 cfg19;
367 	union cvmx_pci_cfg16 cfg16;
368 	union cvmx_pci_cfg22 cfg22;
369 	union cvmx_pci_cfg56 cfg56;
370 
371 	/* Reset the PCI Bus */
372 	cvmx_write_csr(CVMX_CIU_SOFT_PRST, 0x1);
373 	cvmx_read_csr(CVMX_CIU_SOFT_PRST);
374 
375 	udelay(2000);		/* Hold PCI reset for 2 ms */
376 
377 	ctl_status.u64 = 0;	/* cvmx_read_csr(CVMX_NPI_CTL_STATUS); */
378 	ctl_status.s.max_word = 1;
379 	ctl_status.s.timer = 1;
380 	cvmx_write_csr(CVMX_NPI_CTL_STATUS, ctl_status.u64);
381 
382 	/* Deassert PCI reset and advertize PCX Host Mode Device Capability
383 	   (64b) */
384 	cvmx_write_csr(CVMX_CIU_SOFT_PRST, 0x4);
385 	cvmx_read_csr(CVMX_CIU_SOFT_PRST);
386 
387 	udelay(2000);		/* Wait 2 ms after deasserting PCI reset */
388 
389 	ctl_status_2.u32 = 0;
390 	ctl_status_2.s.tsr_hwm = 1;	/* Initializes to 0.  Must be set
391 					   before any PCI reads. */
392 	ctl_status_2.s.bar2pres = 1;	/* Enable BAR2 */
393 	ctl_status_2.s.bar2_enb = 1;
394 	ctl_status_2.s.bar2_cax = 1;	/* Don't use L2 */
395 	ctl_status_2.s.bar2_esx = 1;
396 	ctl_status_2.s.pmo_amod = 1;	/* Round robin priority */
397 	if (octeon_dma_bar_type == OCTEON_DMA_BAR_TYPE_BIG) {
398 		/* BAR1 hole */
399 		ctl_status_2.s.bb1_hole = OCTEON_PCI_BAR1_HOLE_BITS;
400 		ctl_status_2.s.bb1_siz = 1;  /* BAR1 is 2GB */
401 		ctl_status_2.s.bb_ca = 1;    /* Don't use L2 with big bars */
402 		ctl_status_2.s.bb_es = 1;    /* Big bar in byte swap mode */
403 		ctl_status_2.s.bb1 = 1;	     /* BAR1 is big */
404 		ctl_status_2.s.bb0 = 1;	     /* BAR0 is big */
405 	}
406 
407 	octeon_npi_write32(CVMX_NPI_PCI_CTL_STATUS_2, ctl_status_2.u32);
408 	udelay(2000);		/* Wait 2 ms before doing PCI reads */
409 
410 	ctl_status_2.u32 = octeon_npi_read32(CVMX_NPI_PCI_CTL_STATUS_2);
411 	pr_notice("PCI Status: %s %s-bit\n",
412 		  ctl_status_2.s.ap_pcix ? "PCI-X" : "PCI",
413 		  ctl_status_2.s.ap_64ad ? "64" : "32");
414 
415 	if (OCTEON_IS_MODEL(OCTEON_CN58XX) || OCTEON_IS_MODEL(OCTEON_CN50XX)) {
416 		union cvmx_pci_cnt_reg cnt_reg_start;
417 		union cvmx_pci_cnt_reg cnt_reg_end;
418 		unsigned long cycles, pci_clock;
419 
420 		cnt_reg_start.u64 = cvmx_read_csr(CVMX_NPI_PCI_CNT_REG);
421 		cycles = read_c0_cvmcount();
422 		udelay(1000);
423 		cnt_reg_end.u64 = cvmx_read_csr(CVMX_NPI_PCI_CNT_REG);
424 		cycles = read_c0_cvmcount() - cycles;
425 		pci_clock = (cnt_reg_end.s.pcicnt - cnt_reg_start.s.pcicnt) /
426 			    (cycles / (mips_hpt_frequency / 1000000));
427 		pr_notice("PCI Clock: %lu MHz\n", pci_clock);
428 	}
429 
430 	/*
431 	 * TDOMC must be set to one in PCI mode. TDOMC should be set to 4
432 	 * in PCI-X mode to allow four outstanding splits. Otherwise,
433 	 * should not change from its reset value. Don't write PCI_CFG19
434 	 * in PCI mode (0x82000001 reset value), write it to 0x82000004
435 	 * after PCI-X mode is known. MRBCI,MDWE,MDRE -> must be zero.
436 	 * MRBCM -> must be one.
437 	 */
438 	if (ctl_status_2.s.ap_pcix) {
439 		cfg19.u32 = 0;
440 		/*
441 		 * Target Delayed/Split request outstanding maximum
442 		 * count. [1..31] and 0=32.  NOTE: If the user
443 		 * programs these bits beyond the Designed Maximum
444 		 * outstanding count, then the designed maximum table
445 		 * depth will be used instead.	No additional
446 		 * Deferred/Split transactions will be accepted if
447 		 * this outstanding maximum count is
448 		 * reached. Furthermore, no additional deferred/split
449 		 * transactions will be accepted if the I/O delay/ I/O
450 		 * Split Request outstanding maximum is reached.
451 		 */
452 		cfg19.s.tdomc = 4;
453 		/*
454 		 * Master Deferred Read Request Outstanding Max Count
455 		 * (PCI only).	CR4C[26:24] Max SAC cycles MAX DAC
456 		 * cycles 000 8 4 001 1 0 010 2 1 011 3 1 100 4 2 101
457 		 * 5 2 110 6 3 111 7 3 For example, if these bits are
458 		 * programmed to 100, the core can support 2 DAC
459 		 * cycles, 4 SAC cycles or a combination of 1 DAC and
460 		 * 2 SAC cycles. NOTE: For the PCI-X maximum
461 		 * outstanding split transactions, refer to
462 		 * CRE0[22:20].
463 		 */
464 		cfg19.s.mdrrmc = 2;
465 		/*
466 		 * Master Request (Memory Read) Byte Count/Byte Enable
467 		 * select. 0 = Byte Enables valid. In PCI mode, a
468 		 * burst transaction cannot be performed using Memory
469 		 * Read command=4?h6. 1 = DWORD Byte Count valid
470 		 * (default). In PCI Mode, the memory read byte
471 		 * enables are automatically generated by the
472 		 * core. Note: N3 Master Request transaction sizes are
473 		 * always determined through the
474 		 * am_attr[<35:32>|<7:0>] field.
475 		 */
476 		cfg19.s.mrbcm = 1;
477 		octeon_npi_write32(CVMX_NPI_PCI_CFG19, cfg19.u32);
478 	}
479 
480 
481 	cfg01.u32 = 0;
482 	cfg01.s.msae = 1;	/* Memory Space Access Enable */
483 	cfg01.s.me = 1;		/* Master Enable */
484 	cfg01.s.pee = 1;	/* PERR# Enable */
485 	cfg01.s.see = 1;	/* System Error Enable */
486 	cfg01.s.fbbe = 1;	/* Fast Back to Back Transaction Enable */
487 
488 	octeon_npi_write32(CVMX_NPI_PCI_CFG01, cfg01.u32);
489 
490 #ifdef USE_OCTEON_INTERNAL_ARBITER
491 	/*
492 	 * When OCTEON is a PCI host, most systems will use OCTEON's
493 	 * internal arbiter, so must enable it before any PCI/PCI-X
494 	 * traffic can occur.
495 	 */
496 	{
497 		union cvmx_npi_pci_int_arb_cfg pci_int_arb_cfg;
498 
499 		pci_int_arb_cfg.u64 = 0;
500 		pci_int_arb_cfg.s.en = 1;	/* Internal arbiter enable */
501 		cvmx_write_csr(CVMX_NPI_PCI_INT_ARB_CFG, pci_int_arb_cfg.u64);
502 	}
503 #endif	/* USE_OCTEON_INTERNAL_ARBITER */
504 
505 	/*
506 	 * Preferably written to 1 to set MLTD. [RDSATI,TRTAE,
507 	 * TWTAE,TMAE,DPPMR -> must be zero. TILT -> must not be set to
508 	 * 1..7.
509 	 */
510 	cfg16.u32 = 0;
511 	cfg16.s.mltd = 1;	/* Master Latency Timer Disable */
512 	octeon_npi_write32(CVMX_NPI_PCI_CFG16, cfg16.u32);
513 
514 	/*
515 	 * Should be written to 0x4ff00. MTTV -> must be zero.
516 	 * FLUSH -> must be 1. MRV -> should be 0xFF.
517 	 */
518 	cfg22.u32 = 0;
519 	/* Master Retry Value [1..255] and 0=infinite */
520 	cfg22.s.mrv = 0xff;
521 	/*
522 	 * AM_DO_FLUSH_I control NOTE: This bit MUST BE ONE for proper
523 	 * N3K operation.
524 	 */
525 	cfg22.s.flush = 1;
526 	octeon_npi_write32(CVMX_NPI_PCI_CFG22, cfg22.u32);
527 
528 	/*
529 	 * MOST Indicates the maximum number of outstanding splits (in -1
530 	 * notation) when OCTEON is in PCI-X mode.  PCI-X performance is
531 	 * affected by the MOST selection.  Should generally be written
532 	 * with one of 0x3be807, 0x2be807, 0x1be807, or 0x0be807,
533 	 * depending on the desired MOST of 3, 2, 1, or 0, respectively.
534 	 */
535 	cfg56.u32 = 0;
536 	cfg56.s.pxcid = 7;	/* RO - PCI-X Capability ID */
537 	cfg56.s.ncp = 0xe8;	/* RO - Next Capability Pointer */
538 	cfg56.s.dpere = 1;	/* Data Parity Error Recovery Enable */
539 	cfg56.s.roe = 1;	/* Relaxed Ordering Enable */
540 	cfg56.s.mmbc = 1;	/* Maximum Memory Byte Count
541 				   [0=512B,1=1024B,2=2048B,3=4096B] */
542 	cfg56.s.most = 3;	/* Maximum outstanding Split transactions [0=1
543 				   .. 7=32] */
544 
545 	octeon_npi_write32(CVMX_NPI_PCI_CFG56, cfg56.u32);
546 
547 	/*
548 	 * Affects PCI performance when OCTEON services reads to its
549 	 * BAR1/BAR2. Refer to Section 10.6.1.	The recommended values are
550 	 * 0x22, 0x33, and 0x33 for PCI_READ_CMD_6, PCI_READ_CMD_C, and
551 	 * PCI_READ_CMD_E, respectively. Unfortunately due to errata DDR-700,
552 	 * these values need to be changed so they won't possibly prefetch off
553 	 * of the end of memory if PCI is DMAing a buffer at the end of
554 	 * memory. Note that these values differ from their reset values.
555 	 */
556 	octeon_npi_write32(CVMX_NPI_PCI_READ_CMD_6, 0x21);
557 	octeon_npi_write32(CVMX_NPI_PCI_READ_CMD_C, 0x31);
558 	octeon_npi_write32(CVMX_NPI_PCI_READ_CMD_E, 0x31);
559 }
560 
561 
562 /*
563  * Initialize the Octeon PCI controller
564  */
565 static int __init octeon_pci_setup(void)
566 {
567 	union cvmx_npi_mem_access_subidx mem_access;
568 	int index;
569 
570 	/* Only these chips have PCI */
571 	if (octeon_has_feature(OCTEON_FEATURE_PCIE))
572 		return 0;
573 
574 	/* Point pcibios_map_irq() to the PCI version of it */
575 	octeon_pcibios_map_irq = octeon_pci_pcibios_map_irq;
576 
577 	/* Only use the big bars on chips that support it */
578 	if (OCTEON_IS_MODEL(OCTEON_CN31XX) ||
579 	    OCTEON_IS_MODEL(OCTEON_CN38XX_PASS2) ||
580 	    OCTEON_IS_MODEL(OCTEON_CN38XX_PASS1))
581 		octeon_dma_bar_type = OCTEON_DMA_BAR_TYPE_SMALL;
582 	else
583 		octeon_dma_bar_type = OCTEON_DMA_BAR_TYPE_BIG;
584 
585 	if (!octeon_is_pci_host()) {
586 		pr_notice("Not in host mode, PCI Controller not initialized\n");
587 		return 0;
588 	}
589 
590 	/* PCI I/O and PCI MEM values */
591 	set_io_port_base(OCTEON_PCI_IOSPACE_BASE);
592 	ioport_resource.start = 0;
593 	ioport_resource.end = OCTEON_PCI_IOSPACE_SIZE - 1;
594 
595 	pr_notice("%s Octeon big bar support\n",
596 		  (octeon_dma_bar_type ==
597 		  OCTEON_DMA_BAR_TYPE_BIG) ? "Enabling" : "Disabling");
598 
599 	octeon_pci_initialize();
600 
601 	mem_access.u64 = 0;
602 	mem_access.s.esr = 1;	/* Endian-Swap on read. */
603 	mem_access.s.esw = 1;	/* Endian-Swap on write. */
604 	mem_access.s.nsr = 0;	/* No-Snoop on read. */
605 	mem_access.s.nsw = 0;	/* No-Snoop on write. */
606 	mem_access.s.ror = 0;	/* Relax Read on read. */
607 	mem_access.s.row = 0;	/* Relax Order on write. */
608 	mem_access.s.ba = 0;	/* PCI Address bits [63:36]. */
609 	cvmx_write_csr(CVMX_NPI_MEM_ACCESS_SUBID3, mem_access.u64);
610 
611 	/*
612 	 * Remap the Octeon BAR 2 above all 32 bit devices
613 	 * (0x8000000000ul).  This is done here so it is remapped
614 	 * before the readl()'s below. We don't want BAR2 overlapping
615 	 * with BAR0/BAR1 during these reads.
616 	 */
617 	octeon_npi_write32(CVMX_NPI_PCI_CFG08,
618 			   (u32)(OCTEON_BAR2_PCI_ADDRESS & 0xffffffffull));
619 	octeon_npi_write32(CVMX_NPI_PCI_CFG09,
620 			   (u32)(OCTEON_BAR2_PCI_ADDRESS >> 32));
621 
622 	if (octeon_dma_bar_type == OCTEON_DMA_BAR_TYPE_BIG) {
623 		/* Remap the Octeon BAR 0 to 0-2GB */
624 		octeon_npi_write32(CVMX_NPI_PCI_CFG04, 0);
625 		octeon_npi_write32(CVMX_NPI_PCI_CFG05, 0);
626 
627 		/*
628 		 * Remap the Octeon BAR 1 to map 2GB-4GB (minus the
629 		 * BAR 1 hole).
630 		 */
631 		octeon_npi_write32(CVMX_NPI_PCI_CFG06, 2ul << 30);
632 		octeon_npi_write32(CVMX_NPI_PCI_CFG07, 0);
633 
634 		/* BAR1 movable mappings set for identity mapping */
635 		octeon_bar1_pci_phys = 0x80000000ull;
636 		for (index = 0; index < 32; index++) {
637 			union cvmx_pci_bar1_indexx bar1_index;
638 
639 			bar1_index.u32 = 0;
640 			/* Address bits[35:22] sent to L2C */
641 			bar1_index.s.addr_idx =
642 				(octeon_bar1_pci_phys >> 22) + index;
643 			/* Don't put PCI accesses in L2. */
644 			bar1_index.s.ca = 1;
645 			/* Endian Swap Mode */
646 			bar1_index.s.end_swp = 1;
647 			/* Set '1' when the selected address range is valid. */
648 			bar1_index.s.addr_v = 1;
649 			octeon_npi_write32(CVMX_NPI_PCI_BAR1_INDEXX(index),
650 					   bar1_index.u32);
651 		}
652 
653 		/* Devices go after BAR1 */
654 		octeon_pci_mem_resource.start =
655 			OCTEON_PCI_MEMSPACE_OFFSET + (4ul << 30) -
656 			(OCTEON_PCI_BAR1_HOLE_SIZE << 20);
657 		octeon_pci_mem_resource.end =
658 			octeon_pci_mem_resource.start + (1ul << 30);
659 	} else {
660 		/* Remap the Octeon BAR 0 to map 128MB-(128MB+4KB) */
661 		octeon_npi_write32(CVMX_NPI_PCI_CFG04, 128ul << 20);
662 		octeon_npi_write32(CVMX_NPI_PCI_CFG05, 0);
663 
664 		/* Remap the Octeon BAR 1 to map 0-128MB */
665 		octeon_npi_write32(CVMX_NPI_PCI_CFG06, 0);
666 		octeon_npi_write32(CVMX_NPI_PCI_CFG07, 0);
667 
668 		/* BAR1 movable regions contiguous to cover the swiotlb */
669 		octeon_bar1_pci_phys =
670 			virt_to_phys(octeon_swiotlb) & ~((1ull << 22) - 1);
671 
672 		for (index = 0; index < 32; index++) {
673 			union cvmx_pci_bar1_indexx bar1_index;
674 
675 			bar1_index.u32 = 0;
676 			/* Address bits[35:22] sent to L2C */
677 			bar1_index.s.addr_idx =
678 				(octeon_bar1_pci_phys >> 22) + index;
679 			/* Don't put PCI accesses in L2. */
680 			bar1_index.s.ca = 1;
681 			/* Endian Swap Mode */
682 			bar1_index.s.end_swp = 1;
683 			/* Set '1' when the selected address range is valid. */
684 			bar1_index.s.addr_v = 1;
685 			octeon_npi_write32(CVMX_NPI_PCI_BAR1_INDEXX(index),
686 					   bar1_index.u32);
687 		}
688 
689 		/* Devices go after BAR0 */
690 		octeon_pci_mem_resource.start =
691 			OCTEON_PCI_MEMSPACE_OFFSET + (128ul << 20) +
692 			(4ul << 10);
693 		octeon_pci_mem_resource.end =
694 			octeon_pci_mem_resource.start + (1ul << 30);
695 	}
696 
697 	register_pci_controller(&octeon_pci_controller);
698 
699 	/*
700 	 * Clear any errors that might be pending from before the bus
701 	 * was setup properly.
702 	 */
703 	cvmx_write_csr(CVMX_NPI_PCI_INT_SUM2, -1);
704 
705 	if (IS_ERR(platform_device_register_simple("octeon_pci_edac",
706 						   -1, NULL, 0)))
707 		pr_err("Registration of co_pci_edac failed!\n");
708 
709 	octeon_pci_dma_init();
710 
711 	return 0;
712 }
713 
714 arch_initcall(octeon_pci_setup);
715