xref: /openbmc/linux/arch/mips/pci/pci-octeon.c (revision 54cbac81)
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_THUNDER:
218 	case CVMX_BOARD_TYPE_EBH3000:
219 	default:
220 		return "";
221 	}
222 }
223 
224 /**
225  * Map a PCI device to the appropriate interrupt line
226  *
227  * @dev:    The Linux PCI device structure for the device to map
228  * @slot:   The slot number for this device on __BUS 0__. Linux
229  *               enumerates through all the bridges and figures out the
230  *               slot on Bus 0 where this device eventually hooks to.
231  * @pin:    The PCI interrupt pin read from the device, then swizzled
232  *               as it goes through each bridge.
233  * Returns Interrupt number for the device
234  */
235 int __init octeon_pci_pcibios_map_irq(const struct pci_dev *dev,
236 				      u8 slot, u8 pin)
237 {
238 	int irq_num;
239 	const char *interrupts;
240 	int dev_num;
241 
242 	/* Get the board specific interrupt mapping */
243 	interrupts = octeon_get_pci_interrupts();
244 
245 	dev_num = dev->devfn >> 3;
246 	if (dev_num < strlen(interrupts))
247 		irq_num = ((interrupts[dev_num] - 'A' + pin - 1) & 3) +
248 			OCTEON_IRQ_PCI_INT0;
249 	else
250 		irq_num = ((slot + pin - 3) & 3) + OCTEON_IRQ_PCI_INT0;
251 	return irq_num;
252 }
253 
254 
255 /*
256  * Read a value from configuration space
257  */
258 static int octeon_read_config(struct pci_bus *bus, unsigned int devfn,
259 			      int reg, int size, u32 *val)
260 {
261 	union octeon_pci_address pci_addr;
262 
263 	pci_addr.u64 = 0;
264 	pci_addr.s.upper = 2;
265 	pci_addr.s.io = 1;
266 	pci_addr.s.did = 3;
267 	pci_addr.s.subdid = 1;
268 	pci_addr.s.endian_swap = 1;
269 	pci_addr.s.bus = bus->number;
270 	pci_addr.s.dev = devfn >> 3;
271 	pci_addr.s.func = devfn & 0x7;
272 	pci_addr.s.reg = reg;
273 
274 #if PCI_CONFIG_SPACE_DELAY
275 	udelay(PCI_CONFIG_SPACE_DELAY);
276 #endif
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 #if PCI_CONFIG_SPACE_DELAY
312 	udelay(PCI_CONFIG_SPACE_DELAY);
313 #endif
314 	switch (size) {
315 	case 4:
316 		cvmx_write64_uint32(pci_addr.u64, cpu_to_le32(val));
317 		return PCIBIOS_SUCCESSFUL;
318 	case 2:
319 		cvmx_write64_uint16(pci_addr.u64, cpu_to_le16(val));
320 		return PCIBIOS_SUCCESSFUL;
321 	case 1:
322 		cvmx_write64_uint8(pci_addr.u64, val);
323 		return PCIBIOS_SUCCESSFUL;
324 	}
325 	return PCIBIOS_FUNC_NOT_SUPPORTED;
326 }
327 
328 
329 static struct pci_ops octeon_pci_ops = {
330 	octeon_read_config,
331 	octeon_write_config,
332 };
333 
334 static struct resource octeon_pci_mem_resource = {
335 	.start = 0,
336 	.end = 0,
337 	.name = "Octeon PCI MEM",
338 	.flags = IORESOURCE_MEM,
339 };
340 
341 /*
342  * PCI ports must be above 16KB so the ISA bus filtering in the PCI-X to PCI
343  * bridge
344  */
345 static struct resource octeon_pci_io_resource = {
346 	.start = 0x4000,
347 	.end = OCTEON_PCI_IOSPACE_SIZE - 1,
348 	.name = "Octeon PCI IO",
349 	.flags = IORESOURCE_IO,
350 };
351 
352 static struct pci_controller octeon_pci_controller = {
353 	.pci_ops = &octeon_pci_ops,
354 	.mem_resource = &octeon_pci_mem_resource,
355 	.mem_offset = OCTEON_PCI_MEMSPACE_OFFSET,
356 	.io_resource = &octeon_pci_io_resource,
357 	.io_offset = 0,
358 	.io_map_base = OCTEON_PCI_IOSPACE_BASE,
359 };
360 
361 
362 /*
363  * Low level initialize the Octeon PCI controller
364  */
365 static void octeon_pci_initialize(void)
366 {
367 	union cvmx_pci_cfg01 cfg01;
368 	union cvmx_npi_ctl_status ctl_status;
369 	union cvmx_pci_ctl_status_2 ctl_status_2;
370 	union cvmx_pci_cfg19 cfg19;
371 	union cvmx_pci_cfg16 cfg16;
372 	union cvmx_pci_cfg22 cfg22;
373 	union cvmx_pci_cfg56 cfg56;
374 
375 	/* Reset the PCI Bus */
376 	cvmx_write_csr(CVMX_CIU_SOFT_PRST, 0x1);
377 	cvmx_read_csr(CVMX_CIU_SOFT_PRST);
378 
379 	udelay(2000);		/* Hold PCI reset for 2 ms */
380 
381 	ctl_status.u64 = 0;	/* cvmx_read_csr(CVMX_NPI_CTL_STATUS); */
382 	ctl_status.s.max_word = 1;
383 	ctl_status.s.timer = 1;
384 	cvmx_write_csr(CVMX_NPI_CTL_STATUS, ctl_status.u64);
385 
386 	/* Deassert PCI reset and advertize PCX Host Mode Device Capability
387 	   (64b) */
388 	cvmx_write_csr(CVMX_CIU_SOFT_PRST, 0x4);
389 	cvmx_read_csr(CVMX_CIU_SOFT_PRST);
390 
391 	udelay(2000);		/* Wait 2 ms after deasserting PCI reset */
392 
393 	ctl_status_2.u32 = 0;
394 	ctl_status_2.s.tsr_hwm = 1;	/* Initializes to 0.  Must be set
395 					   before any PCI reads. */
396 	ctl_status_2.s.bar2pres = 1;	/* Enable BAR2 */
397 	ctl_status_2.s.bar2_enb = 1;
398 	ctl_status_2.s.bar2_cax = 1;	/* Don't use L2 */
399 	ctl_status_2.s.bar2_esx = 1;
400 	ctl_status_2.s.pmo_amod = 1;	/* Round robin priority */
401 	if (octeon_dma_bar_type == OCTEON_DMA_BAR_TYPE_BIG) {
402 		/* BAR1 hole */
403 		ctl_status_2.s.bb1_hole = OCTEON_PCI_BAR1_HOLE_BITS;
404 		ctl_status_2.s.bb1_siz = 1;  /* BAR1 is 2GB */
405 		ctl_status_2.s.bb_ca = 1;    /* Don't use L2 with big bars */
406 		ctl_status_2.s.bb_es = 1;    /* Big bar in byte swap mode */
407 		ctl_status_2.s.bb1 = 1;      /* BAR1 is big */
408 		ctl_status_2.s.bb0 = 1;      /* BAR0 is big */
409 	}
410 
411 	octeon_npi_write32(CVMX_NPI_PCI_CTL_STATUS_2, ctl_status_2.u32);
412 	udelay(2000);		/* Wait 2 ms before doing PCI reads */
413 
414 	ctl_status_2.u32 = octeon_npi_read32(CVMX_NPI_PCI_CTL_STATUS_2);
415 	pr_notice("PCI Status: %s %s-bit\n",
416 		  ctl_status_2.s.ap_pcix ? "PCI-X" : "PCI",
417 		  ctl_status_2.s.ap_64ad ? "64" : "32");
418 
419 	if (OCTEON_IS_MODEL(OCTEON_CN58XX) || OCTEON_IS_MODEL(OCTEON_CN50XX)) {
420 		union cvmx_pci_cnt_reg cnt_reg_start;
421 		union cvmx_pci_cnt_reg cnt_reg_end;
422 		unsigned long cycles, pci_clock;
423 
424 		cnt_reg_start.u64 = cvmx_read_csr(CVMX_NPI_PCI_CNT_REG);
425 		cycles = read_c0_cvmcount();
426 		udelay(1000);
427 		cnt_reg_end.u64 = cvmx_read_csr(CVMX_NPI_PCI_CNT_REG);
428 		cycles = read_c0_cvmcount() - cycles;
429 		pci_clock = (cnt_reg_end.s.pcicnt - cnt_reg_start.s.pcicnt) /
430 			    (cycles / (mips_hpt_frequency / 1000000));
431 		pr_notice("PCI Clock: %lu MHz\n", pci_clock);
432 	}
433 
434 	/*
435 	 * TDOMC must be set to one in PCI mode. TDOMC should be set to 4
436 	 * in PCI-X mode to allow four outstanding splits. Otherwise,
437 	 * should not change from its reset value. Don't write PCI_CFG19
438 	 * in PCI mode (0x82000001 reset value), write it to 0x82000004
439 	 * after PCI-X mode is known. MRBCI,MDWE,MDRE -> must be zero.
440 	 * MRBCM -> must be one.
441 	 */
442 	if (ctl_status_2.s.ap_pcix) {
443 		cfg19.u32 = 0;
444 		/*
445 		 * Target Delayed/Split request outstanding maximum
446 		 * count. [1..31] and 0=32.  NOTE: If the user
447 		 * programs these bits beyond the Designed Maximum
448 		 * outstanding count, then the designed maximum table
449 		 * depth will be used instead.  No additional
450 		 * Deferred/Split transactions will be accepted if
451 		 * this outstanding maximum count is
452 		 * reached. Furthermore, no additional deferred/split
453 		 * transactions will be accepted if the I/O delay/ I/O
454 		 * Split Request outstanding maximum is reached.
455 		 */
456 		cfg19.s.tdomc = 4;
457 		/*
458 		 * Master Deferred Read Request Outstanding Max Count
459 		 * (PCI only).  CR4C[26:24] Max SAC cycles MAX DAC
460 		 * cycles 000 8 4 001 1 0 010 2 1 011 3 1 100 4 2 101
461 		 * 5 2 110 6 3 111 7 3 For example, if these bits are
462 		 * programmed to 100, the core can support 2 DAC
463 		 * cycles, 4 SAC cycles or a combination of 1 DAC and
464 		 * 2 SAC cycles. NOTE: For the PCI-X maximum
465 		 * outstanding split transactions, refer to
466 		 * CRE0[22:20].
467 		 */
468 		cfg19.s.mdrrmc = 2;
469 		/*
470 		 * Master Request (Memory Read) Byte Count/Byte Enable
471 		 * select. 0 = Byte Enables valid. In PCI mode, a
472 		 * burst transaction cannot be performed using Memory
473 		 * Read command=4?h6. 1 = DWORD Byte Count valid
474 		 * (default). In PCI Mode, the memory read byte
475 		 * enables are automatically generated by the
476 		 * core. Note: N3 Master Request transaction sizes are
477 		 * always determined through the
478 		 * am_attr[<35:32>|<7:0>] field.
479 		 */
480 		cfg19.s.mrbcm = 1;
481 		octeon_npi_write32(CVMX_NPI_PCI_CFG19, cfg19.u32);
482 	}
483 
484 
485 	cfg01.u32 = 0;
486 	cfg01.s.msae = 1;	/* Memory Space Access Enable */
487 	cfg01.s.me = 1;		/* Master Enable */
488 	cfg01.s.pee = 1;	/* PERR# Enable */
489 	cfg01.s.see = 1;	/* System Error Enable */
490 	cfg01.s.fbbe = 1;	/* Fast Back to Back Transaction Enable */
491 
492 	octeon_npi_write32(CVMX_NPI_PCI_CFG01, cfg01.u32);
493 
494 #ifdef USE_OCTEON_INTERNAL_ARBITER
495 	/*
496 	 * When OCTEON is a PCI host, most systems will use OCTEON's
497 	 * internal arbiter, so must enable it before any PCI/PCI-X
498 	 * traffic can occur.
499 	 */
500 	{
501 		union cvmx_npi_pci_int_arb_cfg pci_int_arb_cfg;
502 
503 		pci_int_arb_cfg.u64 = 0;
504 		pci_int_arb_cfg.s.en = 1;	/* Internal arbiter enable */
505 		cvmx_write_csr(CVMX_NPI_PCI_INT_ARB_CFG, pci_int_arb_cfg.u64);
506 	}
507 #endif	/* USE_OCTEON_INTERNAL_ARBITER */
508 
509 	/*
510 	 * Preferably written to 1 to set MLTD. [RDSATI,TRTAE,
511 	 * TWTAE,TMAE,DPPMR -> must be zero. TILT -> must not be set to
512 	 * 1..7.
513 	 */
514 	cfg16.u32 = 0;
515 	cfg16.s.mltd = 1;	/* Master Latency Timer Disable */
516 	octeon_npi_write32(CVMX_NPI_PCI_CFG16, cfg16.u32);
517 
518 	/*
519 	 * Should be written to 0x4ff00. MTTV -> must be zero.
520 	 * FLUSH -> must be 1. MRV -> should be 0xFF.
521 	 */
522 	cfg22.u32 = 0;
523 	/* Master Retry Value [1..255] and 0=infinite */
524 	cfg22.s.mrv = 0xff;
525 	/*
526 	 * AM_DO_FLUSH_I control NOTE: This bit MUST BE ONE for proper
527 	 * N3K operation.
528 	 */
529 	cfg22.s.flush = 1;
530 	octeon_npi_write32(CVMX_NPI_PCI_CFG22, cfg22.u32);
531 
532 	/*
533 	 * MOST Indicates the maximum number of outstanding splits (in -1
534 	 * notation) when OCTEON is in PCI-X mode.  PCI-X performance is
535 	 * affected by the MOST selection.  Should generally be written
536 	 * with one of 0x3be807, 0x2be807, 0x1be807, or 0x0be807,
537 	 * depending on the desired MOST of 3, 2, 1, or 0, respectively.
538 	 */
539 	cfg56.u32 = 0;
540 	cfg56.s.pxcid = 7;	/* RO - PCI-X Capability ID */
541 	cfg56.s.ncp = 0xe8;	/* RO - Next Capability Pointer */
542 	cfg56.s.dpere = 1;	/* Data Parity Error Recovery Enable */
543 	cfg56.s.roe = 1;	/* Relaxed Ordering Enable */
544 	cfg56.s.mmbc = 1;	/* Maximum Memory Byte Count
545 				   [0=512B,1=1024B,2=2048B,3=4096B] */
546 	cfg56.s.most = 3;	/* Maximum outstanding Split transactions [0=1
547 				   .. 7=32] */
548 
549 	octeon_npi_write32(CVMX_NPI_PCI_CFG56, cfg56.u32);
550 
551 	/*
552 	 * Affects PCI performance when OCTEON services reads to its
553 	 * BAR1/BAR2. Refer to Section 10.6.1.  The recommended values are
554 	 * 0x22, 0x33, and 0x33 for PCI_READ_CMD_6, PCI_READ_CMD_C, and
555 	 * PCI_READ_CMD_E, respectively. Unfortunately due to errata DDR-700,
556 	 * these values need to be changed so they won't possibly prefetch off
557 	 * of the end of memory if PCI is DMAing a buffer at the end of
558 	 * memory. Note that these values differ from their reset values.
559 	 */
560 	octeon_npi_write32(CVMX_NPI_PCI_READ_CMD_6, 0x21);
561 	octeon_npi_write32(CVMX_NPI_PCI_READ_CMD_C, 0x31);
562 	octeon_npi_write32(CVMX_NPI_PCI_READ_CMD_E, 0x31);
563 }
564 
565 
566 /*
567  * Initialize the Octeon PCI controller
568  */
569 static int __init octeon_pci_setup(void)
570 {
571 	union cvmx_npi_mem_access_subidx mem_access;
572 	int index;
573 
574 	/* Only these chips have PCI */
575 	if (octeon_has_feature(OCTEON_FEATURE_PCIE))
576 		return 0;
577 
578 	/* Point pcibios_map_irq() to the PCI version of it */
579 	octeon_pcibios_map_irq = octeon_pci_pcibios_map_irq;
580 
581 	/* Only use the big bars on chips that support it */
582 	if (OCTEON_IS_MODEL(OCTEON_CN31XX) ||
583 	    OCTEON_IS_MODEL(OCTEON_CN38XX_PASS2) ||
584 	    OCTEON_IS_MODEL(OCTEON_CN38XX_PASS1))
585 		octeon_dma_bar_type = OCTEON_DMA_BAR_TYPE_SMALL;
586 	else
587 		octeon_dma_bar_type = OCTEON_DMA_BAR_TYPE_BIG;
588 
589 	/* PCI I/O and PCI MEM values */
590 	set_io_port_base(OCTEON_PCI_IOSPACE_BASE);
591 	ioport_resource.start = 0;
592 	ioport_resource.end = OCTEON_PCI_IOSPACE_SIZE - 1;
593 	if (!octeon_is_pci_host()) {
594 		pr_notice("Not in host mode, PCI Controller not initialized\n");
595 		return 0;
596 	}
597 
598 	pr_notice("%s Octeon big bar support\n",
599 		  (octeon_dma_bar_type ==
600 		  OCTEON_DMA_BAR_TYPE_BIG) ? "Enabling" : "Disabling");
601 
602 	octeon_pci_initialize();
603 
604 	mem_access.u64 = 0;
605 	mem_access.s.esr = 1;	/* Endian-Swap on read. */
606 	mem_access.s.esw = 1;	/* Endian-Swap on write. */
607 	mem_access.s.nsr = 0;	/* No-Snoop on read. */
608 	mem_access.s.nsw = 0;	/* No-Snoop on write. */
609 	mem_access.s.ror = 0;	/* Relax Read on read. */
610 	mem_access.s.row = 0;	/* Relax Order on write. */
611 	mem_access.s.ba = 0;	/* PCI Address bits [63:36]. */
612 	cvmx_write_csr(CVMX_NPI_MEM_ACCESS_SUBID3, mem_access.u64);
613 
614 	/*
615 	 * Remap the Octeon BAR 2 above all 32 bit devices
616 	 * (0x8000000000ul).  This is done here so it is remapped
617 	 * before the readl()'s below. We don't want BAR2 overlapping
618 	 * with BAR0/BAR1 during these reads.
619 	 */
620 	octeon_npi_write32(CVMX_NPI_PCI_CFG08,
621 			   (u32)(OCTEON_BAR2_PCI_ADDRESS & 0xffffffffull));
622 	octeon_npi_write32(CVMX_NPI_PCI_CFG09,
623 			   (u32)(OCTEON_BAR2_PCI_ADDRESS >> 32));
624 
625 	if (octeon_dma_bar_type == OCTEON_DMA_BAR_TYPE_BIG) {
626 		/* Remap the Octeon BAR 0 to 0-2GB */
627 		octeon_npi_write32(CVMX_NPI_PCI_CFG04, 0);
628 		octeon_npi_write32(CVMX_NPI_PCI_CFG05, 0);
629 
630 		/*
631 		 * Remap the Octeon BAR 1 to map 2GB-4GB (minus the
632 		 * BAR 1 hole).
633 		 */
634 		octeon_npi_write32(CVMX_NPI_PCI_CFG06, 2ul << 30);
635 		octeon_npi_write32(CVMX_NPI_PCI_CFG07, 0);
636 
637 		/* BAR1 movable mappings set for identity mapping */
638 		octeon_bar1_pci_phys = 0x80000000ull;
639 		for (index = 0; index < 32; index++) {
640 			union cvmx_pci_bar1_indexx bar1_index;
641 
642 			bar1_index.u32 = 0;
643 			/* Address bits[35:22] sent to L2C */
644 			bar1_index.s.addr_idx =
645 				(octeon_bar1_pci_phys >> 22) + index;
646 			/* Don't put PCI accesses in L2. */
647 			bar1_index.s.ca = 1;
648 			/* Endian Swap Mode */
649 			bar1_index.s.end_swp = 1;
650 			/* Set '1' when the selected address range is valid. */
651 			bar1_index.s.addr_v = 1;
652 			octeon_npi_write32(CVMX_NPI_PCI_BAR1_INDEXX(index),
653 					   bar1_index.u32);
654 		}
655 
656 		/* Devices go after BAR1 */
657 		octeon_pci_mem_resource.start =
658 			OCTEON_PCI_MEMSPACE_OFFSET + (4ul << 30) -
659 			(OCTEON_PCI_BAR1_HOLE_SIZE << 20);
660 		octeon_pci_mem_resource.end =
661 			octeon_pci_mem_resource.start + (1ul << 30);
662 	} else {
663 		/* Remap the Octeon BAR 0 to map 128MB-(128MB+4KB) */
664 		octeon_npi_write32(CVMX_NPI_PCI_CFG04, 128ul << 20);
665 		octeon_npi_write32(CVMX_NPI_PCI_CFG05, 0);
666 
667 		/* Remap the Octeon BAR 1 to map 0-128MB */
668 		octeon_npi_write32(CVMX_NPI_PCI_CFG06, 0);
669 		octeon_npi_write32(CVMX_NPI_PCI_CFG07, 0);
670 
671 		/* BAR1 movable regions contiguous to cover the swiotlb */
672 		octeon_bar1_pci_phys =
673 			virt_to_phys(octeon_swiotlb) & ~((1ull << 22) - 1);
674 
675 		for (index = 0; index < 32; index++) {
676 			union cvmx_pci_bar1_indexx bar1_index;
677 
678 			bar1_index.u32 = 0;
679 			/* Address bits[35:22] sent to L2C */
680 			bar1_index.s.addr_idx =
681 				(octeon_bar1_pci_phys >> 22) + index;
682 			/* Don't put PCI accesses in L2. */
683 			bar1_index.s.ca = 1;
684 			/* Endian Swap Mode */
685 			bar1_index.s.end_swp = 1;
686 			/* Set '1' when the selected address range is valid. */
687 			bar1_index.s.addr_v = 1;
688 			octeon_npi_write32(CVMX_NPI_PCI_BAR1_INDEXX(index),
689 					   bar1_index.u32);
690 		}
691 
692 		/* Devices go after BAR0 */
693 		octeon_pci_mem_resource.start =
694 			OCTEON_PCI_MEMSPACE_OFFSET + (128ul << 20) +
695 			(4ul << 10);
696 		octeon_pci_mem_resource.end =
697 			octeon_pci_mem_resource.start + (1ul << 30);
698 	}
699 
700 	register_pci_controller(&octeon_pci_controller);
701 
702 	/*
703 	 * Clear any errors that might be pending from before the bus
704 	 * was setup properly.
705 	 */
706 	cvmx_write_csr(CVMX_NPI_PCI_INT_SUM2, -1);
707 
708 	if (IS_ERR(platform_device_register_simple("octeon_pci_edac",
709 						   -1, NULL, 0)))
710 		pr_err("Registation of co_pci_edac failed!\n");
711 
712 	octeon_pci_dma_init();
713 
714 	return 0;
715 }
716 
717 arch_initcall(octeon_pci_setup);
718