xref: /linux/drivers/parisc/lba_pci.c (revision f884ab15afdc5514e88105c92a4e2e1e6539869a)
1 /*
2 **
3 **  PCI Lower Bus Adapter (LBA) manager
4 **
5 **	(c) Copyright 1999,2000 Grant Grundler
6 **	(c) Copyright 1999,2000 Hewlett-Packard Company
7 **
8 **	This program is free software; you can redistribute it and/or modify
9 **	it under the terms of the GNU General Public License as published by
10 **      the Free Software Foundation; either version 2 of the License, or
11 **      (at your option) any later version.
12 **
13 **
14 ** This module primarily provides access to PCI bus (config/IOport
15 ** spaces) on platforms with an SBA/LBA chipset. A/B/C/J/L/N-class
16 ** with 4 digit model numbers - eg C3000 (and A400...sigh).
17 **
18 ** LBA driver isn't as simple as the Dino driver because:
19 **   (a) this chip has substantial bug fixes between revisions
20 **       (Only one Dino bug has a software workaround :^(  )
21 **   (b) has more options which we don't (yet) support (DMA hints, OLARD)
22 **   (c) IRQ support lives in the I/O SAPIC driver (not with PCI driver)
23 **   (d) play nicely with both PAT and "Legacy" PA-RISC firmware (PDC).
24 **       (dino only deals with "Legacy" PDC)
25 **
26 ** LBA driver passes the I/O SAPIC HPA to the I/O SAPIC driver.
27 ** (I/O SAPIC is integratd in the LBA chip).
28 **
29 ** FIXME: Add support to SBA and LBA drivers for DMA hint sets
30 ** FIXME: Add support for PCI card hot-plug (OLARD).
31 */
32 
33 #include <linux/delay.h>
34 #include <linux/types.h>
35 #include <linux/kernel.h>
36 #include <linux/spinlock.h>
37 #include <linux/init.h>		/* for __init */
38 #include <linux/pci.h>
39 #include <linux/ioport.h>
40 #include <linux/slab.h>
41 
42 #include <asm/byteorder.h>
43 #include <asm/pdc.h>
44 #include <asm/pdcpat.h>
45 #include <asm/page.h>
46 
47 #include <asm/ropes.h>
48 #include <asm/hardware.h>	/* for register_parisc_driver() stuff */
49 #include <asm/parisc-device.h>
50 #include <asm/io.h>		/* read/write stuff */
51 
52 #undef DEBUG_LBA	/* general stuff */
53 #undef DEBUG_LBA_PORT	/* debug I/O Port access */
54 #undef DEBUG_LBA_CFG	/* debug Config Space Access (ie PCI Bus walk) */
55 #undef DEBUG_LBA_PAT	/* debug PCI Resource Mgt code - PDC PAT only */
56 
57 #undef FBB_SUPPORT	/* Fast Back-Back xfers - NOT READY YET */
58 
59 
60 #ifdef DEBUG_LBA
61 #define DBG(x...)	printk(x)
62 #else
63 #define DBG(x...)
64 #endif
65 
66 #ifdef DEBUG_LBA_PORT
67 #define DBG_PORT(x...)	printk(x)
68 #else
69 #define DBG_PORT(x...)
70 #endif
71 
72 #ifdef DEBUG_LBA_CFG
73 #define DBG_CFG(x...)	printk(x)
74 #else
75 #define DBG_CFG(x...)
76 #endif
77 
78 #ifdef DEBUG_LBA_PAT
79 #define DBG_PAT(x...)	printk(x)
80 #else
81 #define DBG_PAT(x...)
82 #endif
83 
84 
85 /*
86 ** Config accessor functions only pass in the 8-bit bus number and not
87 ** the 8-bit "PCI Segment" number. Each LBA will be assigned a PCI bus
88 ** number based on what firmware wrote into the scratch register.
89 **
90 ** The "secondary" bus number is set to this before calling
91 ** pci_register_ops(). If any PPB's are present, the scan will
92 ** discover them and update the "secondary" and "subordinate"
93 ** fields in the pci_bus structure.
94 **
95 ** Changes in the configuration *may* result in a different
96 ** bus number for each LBA depending on what firmware does.
97 */
98 
99 #define MODULE_NAME "LBA"
100 
101 /* non-postable I/O port space, densely packed */
102 #define LBA_PORT_BASE	(PCI_F_EXTEND | 0xfee00000UL)
103 static void __iomem *astro_iop_base __read_mostly;
104 
105 static u32 lba_t32;
106 
107 /* lba flags */
108 #define LBA_FLAG_SKIP_PROBE	0x10
109 
110 #define LBA_SKIP_PROBE(d) ((d)->flags & LBA_FLAG_SKIP_PROBE)
111 
112 
113 /* Looks nice and keeps the compiler happy */
114 #define LBA_DEV(d) ((struct lba_device *) (d))
115 
116 
117 /*
118 ** Only allow 8 subsidiary busses per LBA
119 ** Problem is the PCI bus numbering is globally shared.
120 */
121 #define LBA_MAX_NUM_BUSES 8
122 
123 /************************************
124  * LBA register read and write support
125  *
126  * BE WARNED: register writes are posted.
127  *  (ie follow writes which must reach HW with a read)
128  */
129 #define READ_U8(addr)  __raw_readb(addr)
130 #define READ_U16(addr) __raw_readw(addr)
131 #define READ_U32(addr) __raw_readl(addr)
132 #define WRITE_U8(value, addr)  __raw_writeb(value, addr)
133 #define WRITE_U16(value, addr) __raw_writew(value, addr)
134 #define WRITE_U32(value, addr) __raw_writel(value, addr)
135 
136 #define READ_REG8(addr)  readb(addr)
137 #define READ_REG16(addr) readw(addr)
138 #define READ_REG32(addr) readl(addr)
139 #define READ_REG64(addr) readq(addr)
140 #define WRITE_REG8(value, addr)  writeb(value, addr)
141 #define WRITE_REG16(value, addr) writew(value, addr)
142 #define WRITE_REG32(value, addr) writel(value, addr)
143 
144 
145 #define LBA_CFG_TOK(bus,dfn) ((u32) ((bus)<<16 | (dfn)<<8))
146 #define LBA_CFG_BUS(tok)  ((u8) ((tok)>>16))
147 #define LBA_CFG_DEV(tok)  ((u8) ((tok)>>11) & 0x1f)
148 #define LBA_CFG_FUNC(tok) ((u8) ((tok)>>8 ) & 0x7)
149 
150 
151 /*
152 ** Extract LBA (Rope) number from HPA
153 ** REVISIT: 16 ropes for Stretch/Ike?
154 */
155 #define ROPES_PER_IOC	8
156 #define LBA_NUM(x)    ((((unsigned long) x) >> 13) & (ROPES_PER_IOC-1))
157 
158 
159 static void
160 lba_dump_res(struct resource *r, int d)
161 {
162 	int i;
163 
164 	if (NULL == r)
165 		return;
166 
167 	printk(KERN_DEBUG "(%p)", r->parent);
168 	for (i = d; i ; --i) printk(" ");
169 	printk(KERN_DEBUG "%p [%lx,%lx]/%lx\n", r,
170 		(long)r->start, (long)r->end, r->flags);
171 	lba_dump_res(r->child, d+2);
172 	lba_dump_res(r->sibling, d);
173 }
174 
175 
176 /*
177 ** LBA rev 2.0, 2.1, 2.2, and 3.0 bus walks require a complex
178 ** workaround for cfg cycles:
179 **	-- preserve  LBA state
180 **	-- prevent any DMA from occurring
181 **	-- turn on smart mode
182 **	-- probe with config writes before doing config reads
183 **	-- check ERROR_STATUS
184 **	-- clear ERROR_STATUS
185 **	-- restore LBA state
186 **
187 ** The workaround is only used for device discovery.
188 */
189 
190 static int lba_device_present(u8 bus, u8 dfn, struct lba_device *d)
191 {
192 	u8 first_bus = d->hba.hba_bus->busn_res.start;
193 	u8 last_sub_bus = d->hba.hba_bus->busn_res.end;
194 
195 	if ((bus < first_bus) ||
196 	    (bus > last_sub_bus) ||
197 	    ((bus - first_bus) >= LBA_MAX_NUM_BUSES)) {
198 		return 0;
199 	}
200 
201 	return 1;
202 }
203 
204 
205 
206 #define LBA_CFG_SETUP(d, tok) {				\
207     /* Save contents of error config register.  */			\
208     error_config = READ_REG32(d->hba.base_addr + LBA_ERROR_CONFIG);		\
209 \
210     /* Save contents of status control register.  */			\
211     status_control = READ_REG32(d->hba.base_addr + LBA_STAT_CTL);		\
212 \
213     /* For LBA rev 2.0, 2.1, 2.2, and 3.0, we must disable DMA		\
214     ** arbitration for full bus walks.					\
215     */									\
216 	/* Save contents of arb mask register. */			\
217 	arb_mask = READ_REG32(d->hba.base_addr + LBA_ARB_MASK);		\
218 \
219 	/*								\
220 	 * Turn off all device arbitration bits (i.e. everything	\
221 	 * except arbitration enable bit).				\
222 	 */								\
223 	WRITE_REG32(0x1, d->hba.base_addr + LBA_ARB_MASK);		\
224 \
225     /*									\
226      * Set the smart mode bit so that master aborts don't cause		\
227      * LBA to go into PCI fatal mode (required).			\
228      */									\
229     WRITE_REG32(error_config | LBA_SMART_MODE, d->hba.base_addr + LBA_ERROR_CONFIG);	\
230 }
231 
232 
233 #define LBA_CFG_PROBE(d, tok) {				\
234     /*									\
235      * Setup Vendor ID write and read back the address register		\
236      * to make sure that LBA is the bus master.				\
237      */									\
238     WRITE_REG32(tok | PCI_VENDOR_ID, (d)->hba.base_addr + LBA_PCI_CFG_ADDR);\
239     /*									\
240      * Read address register to ensure that LBA is the bus master,	\
241      * which implies that DMA traffic has stopped when DMA arb is off.	\
242      */									\
243     lba_t32 = READ_REG32((d)->hba.base_addr + LBA_PCI_CFG_ADDR);	\
244     /*									\
245      * Generate a cfg write cycle (will have no affect on		\
246      * Vendor ID register since read-only).				\
247      */									\
248     WRITE_REG32(~0, (d)->hba.base_addr + LBA_PCI_CFG_DATA);		\
249     /*									\
250      * Make sure write has completed before proceeding further,		\
251      * i.e. before setting clear enable.				\
252      */									\
253     lba_t32 = READ_REG32((d)->hba.base_addr + LBA_PCI_CFG_ADDR);	\
254 }
255 
256 
257 /*
258  * HPREVISIT:
259  *   -- Can't tell if config cycle got the error.
260  *
261  *		OV bit is broken until rev 4.0, so can't use OV bit and
262  *		LBA_ERROR_LOG_ADDR to tell if error belongs to config cycle.
263  *
264  *		As of rev 4.0, no longer need the error check.
265  *
266  *   -- Even if we could tell, we still want to return -1
267  *	for **ANY** error (not just master abort).
268  *
269  *   -- Only clear non-fatal errors (we don't want to bring
270  *	LBA out of pci-fatal mode).
271  *
272  *		Actually, there is still a race in which
273  *		we could be clearing a fatal error.  We will
274  *		live with this during our initial bus walk
275  *		until rev 4.0 (no driver activity during
276  *		initial bus walk).  The initial bus walk
277  *		has race conditions concerning the use of
278  *		smart mode as well.
279  */
280 
281 #define LBA_MASTER_ABORT_ERROR 0xc
282 #define LBA_FATAL_ERROR 0x10
283 
284 #define LBA_CFG_MASTER_ABORT_CHECK(d, base, tok, error) {		\
285     u32 error_status = 0;						\
286     /*									\
287      * Set clear enable (CE) bit. Unset by HW when new			\
288      * errors are logged -- LBA HW ERS section 14.3.3).		\
289      */									\
290     WRITE_REG32(status_control | CLEAR_ERRLOG_ENABLE, base + LBA_STAT_CTL); \
291     error_status = READ_REG32(base + LBA_ERROR_STATUS);		\
292     if ((error_status & 0x1f) != 0) {					\
293 	/*								\
294 	 * Fail the config read request.				\
295 	 */								\
296 	error = 1;							\
297 	if ((error_status & LBA_FATAL_ERROR) == 0) {			\
298 	    /*								\
299 	     * Clear error status (if fatal bit not set) by setting	\
300 	     * clear error log bit (CL).				\
301 	     */								\
302 	    WRITE_REG32(status_control | CLEAR_ERRLOG, base + LBA_STAT_CTL); \
303 	}								\
304     }									\
305 }
306 
307 #define LBA_CFG_TR4_ADDR_SETUP(d, addr)					\
308 	WRITE_REG32(((addr) & ~3), (d)->hba.base_addr + LBA_PCI_CFG_ADDR);
309 
310 #define LBA_CFG_ADDR_SETUP(d, addr) {					\
311     WRITE_REG32(((addr) & ~3), (d)->hba.base_addr + LBA_PCI_CFG_ADDR);	\
312     /*									\
313      * Read address register to ensure that LBA is the bus master,	\
314      * which implies that DMA traffic has stopped when DMA arb is off.	\
315      */									\
316     lba_t32 = READ_REG32((d)->hba.base_addr + LBA_PCI_CFG_ADDR);	\
317 }
318 
319 
320 #define LBA_CFG_RESTORE(d, base) {					\
321     /*									\
322      * Restore status control register (turn off clear enable).		\
323      */									\
324     WRITE_REG32(status_control, base + LBA_STAT_CTL);			\
325     /*									\
326      * Restore error config register (turn off smart mode).		\
327      */									\
328     WRITE_REG32(error_config, base + LBA_ERROR_CONFIG);			\
329 	/*								\
330 	 * Restore arb mask register (reenables DMA arbitration).	\
331 	 */								\
332 	WRITE_REG32(arb_mask, base + LBA_ARB_MASK);			\
333 }
334 
335 
336 
337 static unsigned int
338 lba_rd_cfg(struct lba_device *d, u32 tok, u8 reg, u32 size)
339 {
340 	u32 data = ~0U;
341 	int error = 0;
342 	u32 arb_mask = 0;	/* used by LBA_CFG_SETUP/RESTORE */
343 	u32 error_config = 0;	/* used by LBA_CFG_SETUP/RESTORE */
344 	u32 status_control = 0;	/* used by LBA_CFG_SETUP/RESTORE */
345 
346 	LBA_CFG_SETUP(d, tok);
347 	LBA_CFG_PROBE(d, tok);
348 	LBA_CFG_MASTER_ABORT_CHECK(d, d->hba.base_addr, tok, error);
349 	if (!error) {
350 		void __iomem *data_reg = d->hba.base_addr + LBA_PCI_CFG_DATA;
351 
352 		LBA_CFG_ADDR_SETUP(d, tok | reg);
353 		switch (size) {
354 		case 1: data = (u32) READ_REG8(data_reg + (reg & 3)); break;
355 		case 2: data = (u32) READ_REG16(data_reg+ (reg & 2)); break;
356 		case 4: data = READ_REG32(data_reg); break;
357 		}
358 	}
359 	LBA_CFG_RESTORE(d, d->hba.base_addr);
360 	return(data);
361 }
362 
363 
364 static int elroy_cfg_read(struct pci_bus *bus, unsigned int devfn, int pos, int size, u32 *data)
365 {
366 	struct lba_device *d = LBA_DEV(parisc_walk_tree(bus->bridge));
367 	u32 local_bus = (bus->parent == NULL) ? 0 : bus->busn_res.start;
368 	u32 tok = LBA_CFG_TOK(local_bus, devfn);
369 	void __iomem *data_reg = d->hba.base_addr + LBA_PCI_CFG_DATA;
370 
371 	if ((pos > 255) || (devfn > 255))
372 		return -EINVAL;
373 
374 /* FIXME: B2K/C3600 workaround is always use old method... */
375 	/* if (!LBA_SKIP_PROBE(d)) */ {
376 		/* original - Generate config cycle on broken elroy
377 		  with risk we will miss PCI bus errors. */
378 		*data = lba_rd_cfg(d, tok, pos, size);
379 		DBG_CFG("%s(%x+%2x) -> 0x%x (a)\n", __func__, tok, pos, *data);
380 		return 0;
381 	}
382 
383 	if (LBA_SKIP_PROBE(d) && !lba_device_present(bus->busn_res.start, devfn, d)) {
384 		DBG_CFG("%s(%x+%2x) -> -1 (b)\n", __func__, tok, pos);
385 		/* either don't want to look or know device isn't present. */
386 		*data = ~0U;
387 		return(0);
388 	}
389 
390 	/* Basic Algorithm
391 	** Should only get here on fully working LBA rev.
392 	** This is how simple the code should have been.
393 	*/
394 	LBA_CFG_ADDR_SETUP(d, tok | pos);
395 	switch(size) {
396 	case 1: *data = READ_REG8 (data_reg + (pos & 3)); break;
397 	case 2: *data = READ_REG16(data_reg + (pos & 2)); break;
398 	case 4: *data = READ_REG32(data_reg); break;
399 	}
400 	DBG_CFG("%s(%x+%2x) -> 0x%x (c)\n", __func__, tok, pos, *data);
401 	return 0;
402 }
403 
404 
405 static void
406 lba_wr_cfg(struct lba_device *d, u32 tok, u8 reg, u32 data, u32 size)
407 {
408 	int error = 0;
409 	u32 arb_mask = 0;
410 	u32 error_config = 0;
411 	u32 status_control = 0;
412 	void __iomem *data_reg = d->hba.base_addr + LBA_PCI_CFG_DATA;
413 
414 	LBA_CFG_SETUP(d, tok);
415 	LBA_CFG_ADDR_SETUP(d, tok | reg);
416 	switch (size) {
417 	case 1: WRITE_REG8 (data, data_reg + (reg & 3)); break;
418 	case 2: WRITE_REG16(data, data_reg + (reg & 2)); break;
419 	case 4: WRITE_REG32(data, data_reg);             break;
420 	}
421 	LBA_CFG_MASTER_ABORT_CHECK(d, d->hba.base_addr, tok, error);
422 	LBA_CFG_RESTORE(d, d->hba.base_addr);
423 }
424 
425 
426 /*
427  * LBA 4.0 config write code implements non-postable semantics
428  * by doing a read of CONFIG ADDR after the write.
429  */
430 
431 static int elroy_cfg_write(struct pci_bus *bus, unsigned int devfn, int pos, int size, u32 data)
432 {
433 	struct lba_device *d = LBA_DEV(parisc_walk_tree(bus->bridge));
434 	u32 local_bus = (bus->parent == NULL) ? 0 : bus->busn_res.start;
435 	u32 tok = LBA_CFG_TOK(local_bus,devfn);
436 
437 	if ((pos > 255) || (devfn > 255))
438 		return -EINVAL;
439 
440 	if (!LBA_SKIP_PROBE(d)) {
441 		/* Original Workaround */
442 		lba_wr_cfg(d, tok, pos, (u32) data, size);
443 		DBG_CFG("%s(%x+%2x) = 0x%x (a)\n", __func__, tok, pos,data);
444 		return 0;
445 	}
446 
447 	if (LBA_SKIP_PROBE(d) && (!lba_device_present(bus->busn_res.start, devfn, d))) {
448 		DBG_CFG("%s(%x+%2x) = 0x%x (b)\n", __func__, tok, pos,data);
449 		return 1; /* New Workaround */
450 	}
451 
452 	DBG_CFG("%s(%x+%2x) = 0x%x (c)\n", __func__, tok, pos, data);
453 
454 	/* Basic Algorithm */
455 	LBA_CFG_ADDR_SETUP(d, tok | pos);
456 	switch(size) {
457 	case 1: WRITE_REG8 (data, d->hba.base_addr + LBA_PCI_CFG_DATA + (pos & 3));
458 		   break;
459 	case 2: WRITE_REG16(data, d->hba.base_addr + LBA_PCI_CFG_DATA + (pos & 2));
460 		   break;
461 	case 4: WRITE_REG32(data, d->hba.base_addr + LBA_PCI_CFG_DATA);
462 		   break;
463 	}
464 	/* flush posted write */
465 	lba_t32 = READ_REG32(d->hba.base_addr + LBA_PCI_CFG_ADDR);
466 	return 0;
467 }
468 
469 
470 static struct pci_ops elroy_cfg_ops = {
471 	.read =		elroy_cfg_read,
472 	.write =	elroy_cfg_write,
473 };
474 
475 /*
476  * The mercury_cfg_ops are slightly misnamed; they're also used for Elroy
477  * TR4.0 as no additional bugs were found in this areea between Elroy and
478  * Mercury
479  */
480 
481 static int mercury_cfg_read(struct pci_bus *bus, unsigned int devfn, int pos, int size, u32 *data)
482 {
483 	struct lba_device *d = LBA_DEV(parisc_walk_tree(bus->bridge));
484 	u32 local_bus = (bus->parent == NULL) ? 0 : bus->busn_res.start;
485 	u32 tok = LBA_CFG_TOK(local_bus, devfn);
486 	void __iomem *data_reg = d->hba.base_addr + LBA_PCI_CFG_DATA;
487 
488 	if ((pos > 255) || (devfn > 255))
489 		return -EINVAL;
490 
491 	LBA_CFG_TR4_ADDR_SETUP(d, tok | pos);
492 	switch(size) {
493 	case 1:
494 		*data = READ_REG8(data_reg + (pos & 3));
495 		break;
496 	case 2:
497 		*data = READ_REG16(data_reg + (pos & 2));
498 		break;
499 	case 4:
500 		*data = READ_REG32(data_reg);             break;
501 		break;
502 	}
503 
504 	DBG_CFG("mercury_cfg_read(%x+%2x) -> 0x%x\n", tok, pos, *data);
505 	return 0;
506 }
507 
508 /*
509  * LBA 4.0 config write code implements non-postable semantics
510  * by doing a read of CONFIG ADDR after the write.
511  */
512 
513 static int mercury_cfg_write(struct pci_bus *bus, unsigned int devfn, int pos, int size, u32 data)
514 {
515 	struct lba_device *d = LBA_DEV(parisc_walk_tree(bus->bridge));
516 	void __iomem *data_reg = d->hba.base_addr + LBA_PCI_CFG_DATA;
517 	u32 local_bus = (bus->parent == NULL) ? 0 : bus->busn_res.start;
518 	u32 tok = LBA_CFG_TOK(local_bus,devfn);
519 
520 	if ((pos > 255) || (devfn > 255))
521 		return -EINVAL;
522 
523 	DBG_CFG("%s(%x+%2x) <- 0x%x (c)\n", __func__, tok, pos, data);
524 
525 	LBA_CFG_TR4_ADDR_SETUP(d, tok | pos);
526 	switch(size) {
527 	case 1:
528 		WRITE_REG8 (data, data_reg + (pos & 3));
529 		break;
530 	case 2:
531 		WRITE_REG16(data, data_reg + (pos & 2));
532 		break;
533 	case 4:
534 		WRITE_REG32(data, data_reg);
535 		break;
536 	}
537 
538 	/* flush posted write */
539 	lba_t32 = READ_U32(d->hba.base_addr + LBA_PCI_CFG_ADDR);
540 	return 0;
541 }
542 
543 static struct pci_ops mercury_cfg_ops = {
544 	.read =		mercury_cfg_read,
545 	.write =	mercury_cfg_write,
546 };
547 
548 
549 static void
550 lba_bios_init(void)
551 {
552 	DBG(MODULE_NAME ": lba_bios_init\n");
553 }
554 
555 
556 #ifdef CONFIG_64BIT
557 
558 /*
559  * truncate_pat_collision:  Deal with overlaps or outright collisions
560  *			between PAT PDC reported ranges.
561  *
562  *   Broken PA8800 firmware will report lmmio range that
563  *   overlaps with CPU HPA. Just truncate the lmmio range.
564  *
565  *   BEWARE: conflicts with this lmmio range may be an
566  *   elmmio range which is pointing down another rope.
567  *
568  *  FIXME: only deals with one collision per range...theoretically we
569  *  could have several. Supporting more than one collision will get messy.
570  */
571 static unsigned long
572 truncate_pat_collision(struct resource *root, struct resource *new)
573 {
574 	unsigned long start = new->start;
575 	unsigned long end = new->end;
576 	struct resource *tmp = root->child;
577 
578 	if (end <= start || start < root->start || !tmp)
579 		return 0;
580 
581 	/* find first overlap */
582 	while (tmp && tmp->end < start)
583 		tmp = tmp->sibling;
584 
585 	/* no entries overlap */
586 	if (!tmp)  return 0;
587 
588 	/* found one that starts behind the new one
589 	** Don't need to do anything.
590 	*/
591 	if (tmp->start >= end) return 0;
592 
593 	if (tmp->start <= start) {
594 		/* "front" of new one overlaps */
595 		new->start = tmp->end + 1;
596 
597 		if (tmp->end >= end) {
598 			/* AACCKK! totally overlaps! drop this range. */
599 			return 1;
600 		}
601 	}
602 
603 	if (tmp->end < end ) {
604 		/* "end" of new one overlaps */
605 		new->end = tmp->start - 1;
606 	}
607 
608 	printk(KERN_WARNING "LBA: Truncating lmmio_space [%lx/%lx] "
609 					"to [%lx,%lx]\n",
610 			start, end,
611 			(long)new->start, (long)new->end );
612 
613 	return 0;	/* truncation successful */
614 }
615 
616 #else
617 #define truncate_pat_collision(r,n)  (0)
618 #endif
619 
620 /*
621 ** The algorithm is generic code.
622 ** But it needs to access local data structures to get the IRQ base.
623 ** Could make this a "pci_fixup_irq(bus, region)" but not sure
624 ** it's worth it.
625 **
626 ** Called by do_pci_scan_bus() immediately after each PCI bus is walked.
627 ** Resources aren't allocated until recursive buswalk below HBA is completed.
628 */
629 static void
630 lba_fixup_bus(struct pci_bus *bus)
631 {
632 	struct pci_dev *dev;
633 #ifdef FBB_SUPPORT
634 	u16 status;
635 #endif
636 	struct lba_device *ldev = LBA_DEV(parisc_walk_tree(bus->bridge));
637 
638 	DBG("lba_fixup_bus(0x%p) bus %d platform_data 0x%p\n",
639 		bus, (int)bus->busn_res.start, bus->bridge->platform_data);
640 
641 	/*
642 	** Properly Setup MMIO resources for this bus.
643 	** pci_alloc_primary_bus() mangles this.
644 	*/
645 	if (bus->parent) {
646 		int i;
647 		/* PCI-PCI Bridge */
648 		pci_read_bridge_bases(bus);
649 		for (i = PCI_BRIDGE_RESOURCES; i < PCI_NUM_RESOURCES; i++) {
650 			pci_claim_resource(bus->self, i);
651 		}
652 	} else {
653 		/* Host-PCI Bridge */
654 		int err;
655 
656 		DBG("lba_fixup_bus() %s [%lx/%lx]/%lx\n",
657 			ldev->hba.io_space.name,
658 			ldev->hba.io_space.start, ldev->hba.io_space.end,
659 			ldev->hba.io_space.flags);
660 		DBG("lba_fixup_bus() %s [%lx/%lx]/%lx\n",
661 			ldev->hba.lmmio_space.name,
662 			ldev->hba.lmmio_space.start, ldev->hba.lmmio_space.end,
663 			ldev->hba.lmmio_space.flags);
664 
665 		err = request_resource(&ioport_resource, &(ldev->hba.io_space));
666 		if (err < 0) {
667 			lba_dump_res(&ioport_resource, 2);
668 			BUG();
669 		}
670 
671 		if (ldev->hba.elmmio_space.start) {
672 			err = request_resource(&iomem_resource,
673 					&(ldev->hba.elmmio_space));
674 			if (err < 0) {
675 
676 				printk("FAILED: lba_fixup_bus() request for "
677 						"elmmio_space [%lx/%lx]\n",
678 						(long)ldev->hba.elmmio_space.start,
679 						(long)ldev->hba.elmmio_space.end);
680 
681 				/* lba_dump_res(&iomem_resource, 2); */
682 				/* BUG(); */
683 			}
684 		}
685 
686 		if (ldev->hba.lmmio_space.flags) {
687 			err = request_resource(&iomem_resource, &(ldev->hba.lmmio_space));
688 			if (err < 0) {
689 				printk(KERN_ERR "FAILED: lba_fixup_bus() request for "
690 					"lmmio_space [%lx/%lx]\n",
691 					(long)ldev->hba.lmmio_space.start,
692 					(long)ldev->hba.lmmio_space.end);
693 			}
694 		}
695 
696 #ifdef CONFIG_64BIT
697 		/* GMMIO is  distributed range. Every LBA/Rope gets part it. */
698 		if (ldev->hba.gmmio_space.flags) {
699 			err = request_resource(&iomem_resource, &(ldev->hba.gmmio_space));
700 			if (err < 0) {
701 				printk("FAILED: lba_fixup_bus() request for "
702 					"gmmio_space [%lx/%lx]\n",
703 					(long)ldev->hba.gmmio_space.start,
704 					(long)ldev->hba.gmmio_space.end);
705 				lba_dump_res(&iomem_resource, 2);
706 				BUG();
707 			}
708 		}
709 #endif
710 
711 	}
712 
713 	list_for_each_entry(dev, &bus->devices, bus_list) {
714 		int i;
715 
716 		DBG("lba_fixup_bus() %s\n", pci_name(dev));
717 
718 		/* Virtualize Device/Bridge Resources. */
719 		for (i = 0; i < PCI_BRIDGE_RESOURCES; i++) {
720 			struct resource *res = &dev->resource[i];
721 
722 			/* If resource not allocated - skip it */
723 			if (!res->start)
724 				continue;
725 
726 			/*
727 			** FIXME: this will result in whinging for devices
728 			** that share expansion ROMs (think quad tulip), but
729 			** isn't harmful.
730 			*/
731 			pci_claim_resource(dev, i);
732 		}
733 
734 #ifdef FBB_SUPPORT
735 		/*
736 		** If one device does not support FBB transfers,
737 		** No one on the bus can be allowed to use them.
738 		*/
739 		(void) pci_read_config_word(dev, PCI_STATUS, &status);
740 		bus->bridge_ctl &= ~(status & PCI_STATUS_FAST_BACK);
741 #endif
742 
743                 /*
744 		** P2PB's have no IRQs. ignore them.
745 		*/
746 		if ((dev->class >> 8) == PCI_CLASS_BRIDGE_PCI)
747 			continue;
748 
749 		/* Adjust INTERRUPT_LINE for this dev */
750 		iosapic_fixup_irq(ldev->iosapic_obj, dev);
751 	}
752 
753 #ifdef FBB_SUPPORT
754 /* FIXME/REVISIT - finish figuring out to set FBB on both
755 ** pci_setup_bridge() clobbers PCI_BRIDGE_CONTROL.
756 ** Can't fixup here anyway....garr...
757 */
758 	if (fbb_enable) {
759 		if (bus->parent) {
760 			u8 control;
761 			/* enable on PPB */
762 			(void) pci_read_config_byte(bus->self, PCI_BRIDGE_CONTROL, &control);
763 			(void) pci_write_config_byte(bus->self, PCI_BRIDGE_CONTROL, control | PCI_STATUS_FAST_BACK);
764 
765 		} else {
766 			/* enable on LBA */
767 		}
768 		fbb_enable = PCI_COMMAND_FAST_BACK;
769 	}
770 
771 	/* Lastly enable FBB/PERR/SERR on all devices too */
772 	list_for_each_entry(dev, &bus->devices, bus_list) {
773 		(void) pci_read_config_word(dev, PCI_COMMAND, &status);
774 		status |= PCI_COMMAND_PARITY | PCI_COMMAND_SERR | fbb_enable;
775 		(void) pci_write_config_word(dev, PCI_COMMAND, status);
776 	}
777 #endif
778 }
779 
780 
781 static struct pci_bios_ops lba_bios_ops = {
782 	.init =		lba_bios_init,
783 	.fixup_bus =	lba_fixup_bus,
784 };
785 
786 
787 
788 
789 /*******************************************************
790 **
791 ** LBA Sprockets "I/O Port" Space Accessor Functions
792 **
793 ** This set of accessor functions is intended for use with
794 ** "legacy firmware" (ie Sprockets on Allegro/Forte boxes).
795 **
796 ** Many PCI devices don't require use of I/O port space (eg Tulip,
797 ** NCR720) since they export the same registers to both MMIO and
798 ** I/O port space. In general I/O port space is slower than
799 ** MMIO since drivers are designed so PIO writes can be posted.
800 **
801 ********************************************************/
802 
803 #define LBA_PORT_IN(size, mask) \
804 static u##size lba_astro_in##size (struct pci_hba_data *d, u16 addr) \
805 { \
806 	u##size t; \
807 	t = READ_REG##size(astro_iop_base + addr); \
808 	DBG_PORT(" 0x%x\n", t); \
809 	return (t); \
810 }
811 
812 LBA_PORT_IN( 8, 3)
813 LBA_PORT_IN(16, 2)
814 LBA_PORT_IN(32, 0)
815 
816 
817 
818 /*
819 ** BUG X4107:  Ordering broken - DMA RD return can bypass PIO WR
820 **
821 ** Fixed in Elroy 2.2. The READ_U32(..., LBA_FUNC_ID) below is
822 ** guarantee non-postable completion semantics - not avoid X4107.
823 ** The READ_U32 only guarantees the write data gets to elroy but
824 ** out to the PCI bus. We can't read stuff from I/O port space
825 ** since we don't know what has side-effects. Attempting to read
826 ** from configuration space would be suicidal given the number of
827 ** bugs in that elroy functionality.
828 **
829 **      Description:
830 **          DMA read results can improperly pass PIO writes (X4107).  The
831 **          result of this bug is that if a processor modifies a location in
832 **          memory after having issued PIO writes, the PIO writes are not
833 **          guaranteed to be completed before a PCI device is allowed to see
834 **          the modified data in a DMA read.
835 **
836 **          Note that IKE bug X3719 in TR1 IKEs will result in the same
837 **          symptom.
838 **
839 **      Workaround:
840 **          The workaround for this bug is to always follow a PIO write with
841 **          a PIO read to the same bus before starting DMA on that PCI bus.
842 **
843 */
844 #define LBA_PORT_OUT(size, mask) \
845 static void lba_astro_out##size (struct pci_hba_data *d, u16 addr, u##size val) \
846 { \
847 	DBG_PORT("%s(0x%p, 0x%x, 0x%x)\n", __func__, d, addr, val); \
848 	WRITE_REG##size(val, astro_iop_base + addr); \
849 	if (LBA_DEV(d)->hw_rev < 3) \
850 		lba_t32 = READ_U32(d->base_addr + LBA_FUNC_ID); \
851 }
852 
853 LBA_PORT_OUT( 8, 3)
854 LBA_PORT_OUT(16, 2)
855 LBA_PORT_OUT(32, 0)
856 
857 
858 static struct pci_port_ops lba_astro_port_ops = {
859 	.inb =	lba_astro_in8,
860 	.inw =	lba_astro_in16,
861 	.inl =	lba_astro_in32,
862 	.outb =	lba_astro_out8,
863 	.outw =	lba_astro_out16,
864 	.outl =	lba_astro_out32
865 };
866 
867 
868 #ifdef CONFIG_64BIT
869 #define PIOP_TO_GMMIO(lba, addr) \
870 	((lba)->iop_base + (((addr)&0xFFFC)<<10) + ((addr)&3))
871 
872 /*******************************************************
873 **
874 ** LBA PAT "I/O Port" Space Accessor Functions
875 **
876 ** This set of accessor functions is intended for use with
877 ** "PAT PDC" firmware (ie Prelude/Rhapsody/Piranha boxes).
878 **
879 ** This uses the PIOP space located in the first 64MB of GMMIO.
880 ** Each rope gets a full 64*KB* (ie 4 bytes per page) this way.
881 ** bits 1:0 stay the same.  bits 15:2 become 25:12.
882 ** Then add the base and we can generate an I/O Port cycle.
883 ********************************************************/
884 #undef LBA_PORT_IN
885 #define LBA_PORT_IN(size, mask) \
886 static u##size lba_pat_in##size (struct pci_hba_data *l, u16 addr) \
887 { \
888 	u##size t; \
889 	DBG_PORT("%s(0x%p, 0x%x) ->", __func__, l, addr); \
890 	t = READ_REG##size(PIOP_TO_GMMIO(LBA_DEV(l), addr)); \
891 	DBG_PORT(" 0x%x\n", t); \
892 	return (t); \
893 }
894 
895 LBA_PORT_IN( 8, 3)
896 LBA_PORT_IN(16, 2)
897 LBA_PORT_IN(32, 0)
898 
899 
900 #undef LBA_PORT_OUT
901 #define LBA_PORT_OUT(size, mask) \
902 static void lba_pat_out##size (struct pci_hba_data *l, u16 addr, u##size val) \
903 { \
904 	void __iomem *where = PIOP_TO_GMMIO(LBA_DEV(l), addr); \
905 	DBG_PORT("%s(0x%p, 0x%x, 0x%x)\n", __func__, l, addr, val); \
906 	WRITE_REG##size(val, where); \
907 	/* flush the I/O down to the elroy at least */ \
908 	lba_t32 = READ_U32(l->base_addr + LBA_FUNC_ID); \
909 }
910 
911 LBA_PORT_OUT( 8, 3)
912 LBA_PORT_OUT(16, 2)
913 LBA_PORT_OUT(32, 0)
914 
915 
916 static struct pci_port_ops lba_pat_port_ops = {
917 	.inb =	lba_pat_in8,
918 	.inw =	lba_pat_in16,
919 	.inl =	lba_pat_in32,
920 	.outb =	lba_pat_out8,
921 	.outw =	lba_pat_out16,
922 	.outl =	lba_pat_out32
923 };
924 
925 
926 
927 /*
928 ** make range information from PDC available to PCI subsystem.
929 ** We make the PDC call here in order to get the PCI bus range
930 ** numbers. The rest will get forwarded in pcibios_fixup_bus().
931 ** We don't have a struct pci_bus assigned to us yet.
932 */
933 static void
934 lba_pat_resources(struct parisc_device *pa_dev, struct lba_device *lba_dev)
935 {
936 	unsigned long bytecnt;
937 	long io_count;
938 	long status;	/* PDC return status */
939 	long pa_count;
940 	pdc_pat_cell_mod_maddr_block_t *pa_pdc_cell;	/* PA_VIEW */
941 	pdc_pat_cell_mod_maddr_block_t *io_pdc_cell;	/* IO_VIEW */
942 	int i;
943 
944 	pa_pdc_cell = kzalloc(sizeof(pdc_pat_cell_mod_maddr_block_t), GFP_KERNEL);
945 	if (!pa_pdc_cell)
946 		return;
947 
948 	io_pdc_cell = kzalloc(sizeof(pdc_pat_cell_mod_maddr_block_t), GFP_KERNEL);
949 	if (!io_pdc_cell) {
950 		kfree(pa_pdc_cell);
951 		return;
952 	}
953 
954 	/* return cell module (IO view) */
955 	status = pdc_pat_cell_module(&bytecnt, pa_dev->pcell_loc, pa_dev->mod_index,
956 				PA_VIEW, pa_pdc_cell);
957 	pa_count = pa_pdc_cell->mod[1];
958 
959 	status |= pdc_pat_cell_module(&bytecnt, pa_dev->pcell_loc, pa_dev->mod_index,
960 				IO_VIEW, io_pdc_cell);
961 	io_count = io_pdc_cell->mod[1];
962 
963 	/* We've already done this once for device discovery...*/
964 	if (status != PDC_OK) {
965 		panic("pdc_pat_cell_module() call failed for LBA!\n");
966 	}
967 
968 	if (PAT_GET_ENTITY(pa_pdc_cell->mod_info) != PAT_ENTITY_LBA) {
969 		panic("pdc_pat_cell_module() entity returned != PAT_ENTITY_LBA!\n");
970 	}
971 
972 	/*
973 	** Inspect the resources PAT tells us about
974 	*/
975 	for (i = 0; i < pa_count; i++) {
976 		struct {
977 			unsigned long type;
978 			unsigned long start;
979 			unsigned long end;	/* aka finish */
980 		} *p, *io;
981 		struct resource *r;
982 
983 		p = (void *) &(pa_pdc_cell->mod[2+i*3]);
984 		io = (void *) &(io_pdc_cell->mod[2+i*3]);
985 
986 		/* Convert the PAT range data to PCI "struct resource" */
987 		switch(p->type & 0xff) {
988 		case PAT_PBNUM:
989 			lba_dev->hba.bus_num.start = p->start;
990 			lba_dev->hba.bus_num.end   = p->end;
991 			lba_dev->hba.bus_num.flags = IORESOURCE_BUS;
992 			break;
993 
994 		case PAT_LMMIO:
995 			/* used to fix up pre-initialized MEM BARs */
996 			if (!lba_dev->hba.lmmio_space.start) {
997 				sprintf(lba_dev->hba.lmmio_name,
998 						"PCI%02x LMMIO",
999 						(int)lba_dev->hba.bus_num.start);
1000 				lba_dev->hba.lmmio_space_offset = p->start -
1001 					io->start;
1002 				r = &lba_dev->hba.lmmio_space;
1003 				r->name = lba_dev->hba.lmmio_name;
1004 			} else if (!lba_dev->hba.elmmio_space.start) {
1005 				sprintf(lba_dev->hba.elmmio_name,
1006 						"PCI%02x ELMMIO",
1007 						(int)lba_dev->hba.bus_num.start);
1008 				r = &lba_dev->hba.elmmio_space;
1009 				r->name = lba_dev->hba.elmmio_name;
1010 			} else {
1011 				printk(KERN_WARNING MODULE_NAME
1012 					" only supports 2 LMMIO resources!\n");
1013 				break;
1014 			}
1015 
1016 			r->start  = p->start;
1017 			r->end    = p->end;
1018 			r->flags  = IORESOURCE_MEM;
1019 			r->parent = r->sibling = r->child = NULL;
1020 			break;
1021 
1022 		case PAT_GMMIO:
1023 			/* MMIO space > 4GB phys addr; for 64-bit BAR */
1024 			sprintf(lba_dev->hba.gmmio_name, "PCI%02x GMMIO",
1025 					(int)lba_dev->hba.bus_num.start);
1026 			r = &lba_dev->hba.gmmio_space;
1027 			r->name  = lba_dev->hba.gmmio_name;
1028 			r->start  = p->start;
1029 			r->end    = p->end;
1030 			r->flags  = IORESOURCE_MEM;
1031 			r->parent = r->sibling = r->child = NULL;
1032 			break;
1033 
1034 		case PAT_NPIOP:
1035 			printk(KERN_WARNING MODULE_NAME
1036 				" range[%d] : ignoring NPIOP (0x%lx)\n",
1037 				i, p->start);
1038 			break;
1039 
1040 		case PAT_PIOP:
1041 			/*
1042 			** Postable I/O port space is per PCI host adapter.
1043 			** base of 64MB PIOP region
1044 			*/
1045 			lba_dev->iop_base = ioremap_nocache(p->start, 64 * 1024 * 1024);
1046 
1047 			sprintf(lba_dev->hba.io_name, "PCI%02x Ports",
1048 					(int)lba_dev->hba.bus_num.start);
1049 			r = &lba_dev->hba.io_space;
1050 			r->name  = lba_dev->hba.io_name;
1051 			r->start  = HBA_PORT_BASE(lba_dev->hba.hba_num);
1052 			r->end    = r->start + HBA_PORT_SPACE_SIZE - 1;
1053 			r->flags  = IORESOURCE_IO;
1054 			r->parent = r->sibling = r->child = NULL;
1055 			break;
1056 
1057 		default:
1058 			printk(KERN_WARNING MODULE_NAME
1059 				" range[%d] : unknown pat range type (0x%lx)\n",
1060 				i, p->type & 0xff);
1061 			break;
1062 		}
1063 	}
1064 
1065 	kfree(pa_pdc_cell);
1066 	kfree(io_pdc_cell);
1067 }
1068 #else
1069 /* keep compiler from complaining about missing declarations */
1070 #define lba_pat_port_ops lba_astro_port_ops
1071 #define lba_pat_resources(pa_dev, lba_dev)
1072 #endif	/* CONFIG_64BIT */
1073 
1074 
1075 extern void sba_distributed_lmmio(struct parisc_device *, struct resource *);
1076 extern void sba_directed_lmmio(struct parisc_device *, struct resource *);
1077 
1078 
1079 static void
1080 lba_legacy_resources(struct parisc_device *pa_dev, struct lba_device *lba_dev)
1081 {
1082 	struct resource *r;
1083 	int lba_num;
1084 
1085 	lba_dev->hba.lmmio_space_offset = PCI_F_EXTEND;
1086 
1087 	/*
1088 	** With "legacy" firmware, the lowest byte of FW_SCRATCH
1089 	** represents bus->secondary and the second byte represents
1090 	** bus->subsidiary (i.e. highest PPB programmed by firmware).
1091 	** PCI bus walk *should* end up with the same result.
1092 	** FIXME: But we don't have sanity checks in PCI or LBA.
1093 	*/
1094 	lba_num = READ_REG32(lba_dev->hba.base_addr + LBA_FW_SCRATCH);
1095 	r = &(lba_dev->hba.bus_num);
1096 	r->name = "LBA PCI Busses";
1097 	r->start = lba_num & 0xff;
1098 	r->end = (lba_num>>8) & 0xff;
1099 
1100 	/* Set up local PCI Bus resources - we don't need them for
1101 	** Legacy boxes but it's nice to see in /proc/iomem.
1102 	*/
1103 	r = &(lba_dev->hba.lmmio_space);
1104 	sprintf(lba_dev->hba.lmmio_name, "PCI%02x LMMIO",
1105 					(int)lba_dev->hba.bus_num.start);
1106 	r->name  = lba_dev->hba.lmmio_name;
1107 
1108 #if 1
1109 	/* We want the CPU -> IO routing of addresses.
1110 	 * The SBA BASE/MASK registers control CPU -> IO routing.
1111 	 * Ask SBA what is routed to this rope/LBA.
1112 	 */
1113 	sba_distributed_lmmio(pa_dev, r);
1114 #else
1115 	/*
1116 	 * The LBA BASE/MASK registers control IO -> System routing.
1117 	 *
1118 	 * The following code works but doesn't get us what we want.
1119 	 * Well, only because firmware (v5.0) on C3000 doesn't program
1120 	 * the LBA BASE/MASE registers to be the exact inverse of
1121 	 * the corresponding SBA registers. Other Astro/Pluto
1122 	 * based platform firmware may do it right.
1123 	 *
1124 	 * Should someone want to mess with MSI, they may need to
1125 	 * reprogram LBA BASE/MASK registers. Thus preserve the code
1126 	 * below until MSI is known to work on C3000/A500/N4000/RP3440.
1127 	 *
1128 	 * Using the code below, /proc/iomem shows:
1129 	 * ...
1130 	 * f0000000-f0ffffff : PCI00 LMMIO
1131 	 *   f05d0000-f05d0000 : lcd_data
1132 	 *   f05d0008-f05d0008 : lcd_cmd
1133 	 * f1000000-f1ffffff : PCI01 LMMIO
1134 	 * f4000000-f4ffffff : PCI02 LMMIO
1135 	 *   f4000000-f4001fff : sym53c8xx
1136 	 *   f4002000-f4003fff : sym53c8xx
1137 	 *   f4004000-f40043ff : sym53c8xx
1138 	 *   f4005000-f40053ff : sym53c8xx
1139 	 *   f4007000-f4007fff : ohci_hcd
1140 	 *   f4008000-f40083ff : tulip
1141 	 * f6000000-f6ffffff : PCI03 LMMIO
1142 	 * f8000000-fbffffff : PCI00 ELMMIO
1143 	 *   fa100000-fa4fffff : stifb mmio
1144 	 *   fb000000-fb1fffff : stifb fb
1145 	 *
1146 	 * But everything listed under PCI02 actually lives under PCI00.
1147 	 * This is clearly wrong.
1148 	 *
1149 	 * Asking SBA how things are routed tells the correct story:
1150 	 * LMMIO_BASE/MASK/ROUTE f4000001 fc000000 00000000
1151 	 * DIR0_BASE/MASK/ROUTE fa000001 fe000000 00000006
1152 	 * DIR1_BASE/MASK/ROUTE f9000001 ff000000 00000004
1153 	 * DIR2_BASE/MASK/ROUTE f0000000 fc000000 00000000
1154 	 * DIR3_BASE/MASK/ROUTE f0000000 fc000000 00000000
1155 	 *
1156 	 * Which looks like this in /proc/iomem:
1157 	 * f4000000-f47fffff : PCI00 LMMIO
1158 	 *   f4000000-f4001fff : sym53c8xx
1159 	 *   ...[deteled core devices - same as above]...
1160 	 *   f4008000-f40083ff : tulip
1161 	 * f4800000-f4ffffff : PCI01 LMMIO
1162 	 * f6000000-f67fffff : PCI02 LMMIO
1163 	 * f7000000-f77fffff : PCI03 LMMIO
1164 	 * f9000000-f9ffffff : PCI02 ELMMIO
1165 	 * fa000000-fbffffff : PCI03 ELMMIO
1166 	 *   fa100000-fa4fffff : stifb mmio
1167 	 *   fb000000-fb1fffff : stifb fb
1168 	 *
1169 	 * ie all Built-in core are under now correctly under PCI00.
1170 	 * The "PCI02 ELMMIO" directed range is for:
1171 	 *  +-[02]---03.0  3Dfx Interactive, Inc. Voodoo 2
1172 	 *
1173 	 * All is well now.
1174 	 */
1175 	r->start = READ_REG32(lba_dev->hba.base_addr + LBA_LMMIO_BASE);
1176 	if (r->start & 1) {
1177 		unsigned long rsize;
1178 
1179 		r->flags = IORESOURCE_MEM;
1180 		/* mmio_mask also clears Enable bit */
1181 		r->start &= mmio_mask;
1182 		r->start = PCI_HOST_ADDR(HBA_DATA(lba_dev), r->start);
1183 		rsize = ~ READ_REG32(lba_dev->hba.base_addr + LBA_LMMIO_MASK);
1184 
1185 		/*
1186 		** Each rope only gets part of the distributed range.
1187 		** Adjust "window" for this rope.
1188 		*/
1189 		rsize /= ROPES_PER_IOC;
1190 		r->start += (rsize + 1) * LBA_NUM(pa_dev->hpa.start);
1191 		r->end = r->start + rsize;
1192 	} else {
1193 		r->end = r->start = 0;	/* Not enabled. */
1194 	}
1195 #endif
1196 
1197 	/*
1198 	** "Directed" ranges are used when the "distributed range" isn't
1199 	** sufficient for all devices below a given LBA.  Typically devices
1200 	** like graphics cards or X25 may need a directed range when the
1201 	** bus has multiple slots (ie multiple devices) or the device
1202 	** needs more than the typical 4 or 8MB a distributed range offers.
1203 	**
1204 	** The main reason for ignoring it now frigging complications.
1205 	** Directed ranges may overlap (and have precedence) over
1206 	** distributed ranges. Or a distributed range assigned to a unused
1207 	** rope may be used by a directed range on a different rope.
1208 	** Support for graphics devices may require fixing this
1209 	** since they may be assigned a directed range which overlaps
1210 	** an existing (but unused portion of) distributed range.
1211 	*/
1212 	r = &(lba_dev->hba.elmmio_space);
1213 	sprintf(lba_dev->hba.elmmio_name, "PCI%02x ELMMIO",
1214 					(int)lba_dev->hba.bus_num.start);
1215 	r->name  = lba_dev->hba.elmmio_name;
1216 
1217 #if 1
1218 	/* See comment which precedes call to sba_directed_lmmio() */
1219 	sba_directed_lmmio(pa_dev, r);
1220 #else
1221 	r->start = READ_REG32(lba_dev->hba.base_addr + LBA_ELMMIO_BASE);
1222 
1223 	if (r->start & 1) {
1224 		unsigned long rsize;
1225 		r->flags = IORESOURCE_MEM;
1226 		/* mmio_mask also clears Enable bit */
1227 		r->start &= mmio_mask;
1228 		r->start = PCI_HOST_ADDR(HBA_DATA(lba_dev), r->start);
1229 		rsize = READ_REG32(lba_dev->hba.base_addr + LBA_ELMMIO_MASK);
1230 		r->end = r->start + ~rsize;
1231 	}
1232 #endif
1233 
1234 	r = &(lba_dev->hba.io_space);
1235 	sprintf(lba_dev->hba.io_name, "PCI%02x Ports",
1236 					(int)lba_dev->hba.bus_num.start);
1237 	r->name  = lba_dev->hba.io_name;
1238 	r->flags = IORESOURCE_IO;
1239 	r->start = READ_REG32(lba_dev->hba.base_addr + LBA_IOS_BASE) & ~1L;
1240 	r->end   = r->start + (READ_REG32(lba_dev->hba.base_addr + LBA_IOS_MASK) ^ (HBA_PORT_SPACE_SIZE - 1));
1241 
1242 	/* Virtualize the I/O Port space ranges */
1243 	lba_num = HBA_PORT_BASE(lba_dev->hba.hba_num);
1244 	r->start |= lba_num;
1245 	r->end   |= lba_num;
1246 }
1247 
1248 
1249 /**************************************************************************
1250 **
1251 **   LBA initialization code (HW and SW)
1252 **
1253 **   o identify LBA chip itself
1254 **   o initialize LBA chip modes (HardFail)
1255 **   o FIXME: initialize DMA hints for reasonable defaults
1256 **   o enable configuration functions
1257 **   o call pci_register_ops() to discover devs (fixup/fixup_bus get invoked)
1258 **
1259 **************************************************************************/
1260 
1261 static int __init
1262 lba_hw_init(struct lba_device *d)
1263 {
1264 	u32 stat;
1265 	u32 bus_reset;	/* PDC_PAT_BUG */
1266 
1267 #if 0
1268 	printk(KERN_DEBUG "LBA %lx  STAT_CTL %Lx  ERROR_CFG %Lx  STATUS %Lx DMA_CTL %Lx\n",
1269 		d->hba.base_addr,
1270 		READ_REG64(d->hba.base_addr + LBA_STAT_CTL),
1271 		READ_REG64(d->hba.base_addr + LBA_ERROR_CONFIG),
1272 		READ_REG64(d->hba.base_addr + LBA_ERROR_STATUS),
1273 		READ_REG64(d->hba.base_addr + LBA_DMA_CTL) );
1274 	printk(KERN_DEBUG "	ARB mask %Lx  pri %Lx  mode %Lx  mtlt %Lx\n",
1275 		READ_REG64(d->hba.base_addr + LBA_ARB_MASK),
1276 		READ_REG64(d->hba.base_addr + LBA_ARB_PRI),
1277 		READ_REG64(d->hba.base_addr + LBA_ARB_MODE),
1278 		READ_REG64(d->hba.base_addr + LBA_ARB_MTLT) );
1279 	printk(KERN_DEBUG "	HINT cfg 0x%Lx\n",
1280 		READ_REG64(d->hba.base_addr + LBA_HINT_CFG));
1281 	printk(KERN_DEBUG "	HINT reg ");
1282 	{ int i;
1283 	for (i=LBA_HINT_BASE; i< (14*8 + LBA_HINT_BASE); i+=8)
1284 		printk(" %Lx", READ_REG64(d->hba.base_addr + i));
1285 	}
1286 	printk("\n");
1287 #endif	/* DEBUG_LBA_PAT */
1288 
1289 #ifdef CONFIG_64BIT
1290 /*
1291  * FIXME add support for PDC_PAT_IO "Get slot status" - OLAR support
1292  * Only N-Class and up can really make use of Get slot status.
1293  * maybe L-class too but I've never played with it there.
1294  */
1295 #endif
1296 
1297 	/* PDC_PAT_BUG: exhibited in rev 40.48  on L2000 */
1298 	bus_reset = READ_REG32(d->hba.base_addr + LBA_STAT_CTL + 4) & 1;
1299 	if (bus_reset) {
1300 		printk(KERN_DEBUG "NOTICE: PCI bus reset still asserted! (clearing)\n");
1301 	}
1302 
1303 	stat = READ_REG32(d->hba.base_addr + LBA_ERROR_CONFIG);
1304 	if (stat & LBA_SMART_MODE) {
1305 		printk(KERN_DEBUG "NOTICE: LBA in SMART mode! (cleared)\n");
1306 		stat &= ~LBA_SMART_MODE;
1307 		WRITE_REG32(stat, d->hba.base_addr + LBA_ERROR_CONFIG);
1308 	}
1309 
1310 	/* Set HF mode as the default (vs. -1 mode). */
1311         stat = READ_REG32(d->hba.base_addr + LBA_STAT_CTL);
1312 	WRITE_REG32(stat | HF_ENABLE, d->hba.base_addr + LBA_STAT_CTL);
1313 
1314 	/*
1315 	** Writing a zero to STAT_CTL.rf (bit 0) will clear reset signal
1316 	** if it's not already set. If we just cleared the PCI Bus Reset
1317 	** signal, wait a bit for the PCI devices to recover and setup.
1318 	*/
1319 	if (bus_reset)
1320 		mdelay(pci_post_reset_delay);
1321 
1322 	if (0 == READ_REG32(d->hba.base_addr + LBA_ARB_MASK)) {
1323 		/*
1324 		** PDC_PAT_BUG: PDC rev 40.48 on L2000.
1325 		** B2000/C3600/J6000 also have this problem?
1326 		**
1327 		** Elroys with hot pluggable slots don't get configured
1328 		** correctly if the slot is empty.  ARB_MASK is set to 0
1329 		** and we can't master transactions on the bus if it's
1330 		** not at least one. 0x3 enables elroy and first slot.
1331 		*/
1332 		printk(KERN_DEBUG "NOTICE: Enabling PCI Arbitration\n");
1333 		WRITE_REG32(0x3, d->hba.base_addr + LBA_ARB_MASK);
1334 	}
1335 
1336 	/*
1337 	** FIXME: Hint registers are programmed with default hint
1338 	** values by firmware. Hints should be sane even if we
1339 	** can't reprogram them the way drivers want.
1340 	*/
1341 	return 0;
1342 }
1343 
1344 /*
1345  * Unfortunately, when firmware numbers busses, it doesn't take into account
1346  * Cardbus bridges.  So we have to renumber the busses to suit ourselves.
1347  * Elroy/Mercury don't actually know what bus number they're attached to;
1348  * we use bus 0 to indicate the directly attached bus and any other bus
1349  * number will be taken care of by the PCI-PCI bridge.
1350  */
1351 static unsigned int lba_next_bus = 0;
1352 
1353 /*
1354  * Determine if lba should claim this chip (return 0) or not (return 1).
1355  * If so, initialize the chip and tell other partners in crime they
1356  * have work to do.
1357  */
1358 static int __init
1359 lba_driver_probe(struct parisc_device *dev)
1360 {
1361 	struct lba_device *lba_dev;
1362 	LIST_HEAD(resources);
1363 	struct pci_bus *lba_bus;
1364 	struct pci_ops *cfg_ops;
1365 	u32 func_class;
1366 	void *tmp_obj;
1367 	char *version;
1368 	void __iomem *addr = ioremap_nocache(dev->hpa.start, 4096);
1369 	int max;
1370 
1371 	/* Read HW Rev First */
1372 	func_class = READ_REG32(addr + LBA_FCLASS);
1373 
1374 	if (IS_ELROY(dev)) {
1375 		func_class &= 0xf;
1376 		switch (func_class) {
1377 		case 0:	version = "TR1.0"; break;
1378 		case 1:	version = "TR2.0"; break;
1379 		case 2:	version = "TR2.1"; break;
1380 		case 3:	version = "TR2.2"; break;
1381 		case 4:	version = "TR3.0"; break;
1382 		case 5:	version = "TR4.0"; break;
1383 		default: version = "TR4+";
1384 		}
1385 
1386 		printk(KERN_INFO "Elroy version %s (0x%x) found at 0x%lx\n",
1387 		       version, func_class & 0xf, (long)dev->hpa.start);
1388 
1389 		if (func_class < 2) {
1390 			printk(KERN_WARNING "Can't support LBA older than "
1391 				"TR2.1 - continuing under adversity.\n");
1392 		}
1393 
1394 #if 0
1395 /* Elroy TR4.0 should work with simple algorithm.
1396    But it doesn't.  Still missing something. *sigh*
1397 */
1398 		if (func_class > 4) {
1399 			cfg_ops = &mercury_cfg_ops;
1400 		} else
1401 #endif
1402 		{
1403 			cfg_ops = &elroy_cfg_ops;
1404 		}
1405 
1406 	} else if (IS_MERCURY(dev) || IS_QUICKSILVER(dev)) {
1407 		int major, minor;
1408 
1409 		func_class &= 0xff;
1410 		major = func_class >> 4, minor = func_class & 0xf;
1411 
1412 		/* We could use one printk for both Elroy and Mercury,
1413                  * but for the mask for func_class.
1414                  */
1415 		printk(KERN_INFO "%s version TR%d.%d (0x%x) found at 0x%lx\n",
1416 		       IS_MERCURY(dev) ? "Mercury" : "Quicksilver", major,
1417 		       minor, func_class, (long)dev->hpa.start);
1418 
1419 		cfg_ops = &mercury_cfg_ops;
1420 	} else {
1421 		printk(KERN_ERR "Unknown LBA found at 0x%lx\n",
1422 			(long)dev->hpa.start);
1423 		return -ENODEV;
1424 	}
1425 
1426 	/* Tell I/O SAPIC driver we have a IRQ handler/region. */
1427 	tmp_obj = iosapic_register(dev->hpa.start + LBA_IOSAPIC_BASE);
1428 
1429 	/* NOTE: PCI devices (e.g. 103c:1005 graphics card) which don't
1430 	**	have an IRT entry will get NULL back from iosapic code.
1431 	*/
1432 
1433 	lba_dev = kzalloc(sizeof(struct lba_device), GFP_KERNEL);
1434 	if (!lba_dev) {
1435 		printk(KERN_ERR "lba_init_chip - couldn't alloc lba_device\n");
1436 		return(1);
1437 	}
1438 
1439 
1440 	/* ---------- First : initialize data we already have --------- */
1441 
1442 	lba_dev->hw_rev = func_class;
1443 	lba_dev->hba.base_addr = addr;
1444 	lba_dev->hba.dev = dev;
1445 	lba_dev->iosapic_obj = tmp_obj;  /* save interrupt handle */
1446 	lba_dev->hba.iommu = sba_get_iommu(dev);  /* get iommu data */
1447 	parisc_set_drvdata(dev, lba_dev);
1448 
1449 	/* ------------ Second : initialize common stuff ---------- */
1450 	pci_bios = &lba_bios_ops;
1451 	pcibios_register_hba(HBA_DATA(lba_dev));
1452 	spin_lock_init(&lba_dev->lba_lock);
1453 
1454 	if (lba_hw_init(lba_dev))
1455 		return(1);
1456 
1457 	/* ---------- Third : setup I/O Port and MMIO resources  --------- */
1458 
1459 	if (is_pdc_pat()) {
1460 		/* PDC PAT firmware uses PIOP region of GMMIO space. */
1461 		pci_port = &lba_pat_port_ops;
1462 		/* Go ask PDC PAT what resources this LBA has */
1463 		lba_pat_resources(dev, lba_dev);
1464 	} else {
1465 		if (!astro_iop_base) {
1466 			/* Sprockets PDC uses NPIOP region */
1467 			astro_iop_base = ioremap_nocache(LBA_PORT_BASE, 64 * 1024);
1468 			pci_port = &lba_astro_port_ops;
1469 		}
1470 
1471 		/* Poke the chip a bit for /proc output */
1472 		lba_legacy_resources(dev, lba_dev);
1473 	}
1474 
1475 	if (lba_dev->hba.bus_num.start < lba_next_bus)
1476 		lba_dev->hba.bus_num.start = lba_next_bus;
1477 
1478 	/*   Overlaps with elmmio can (and should) fail here.
1479 	 *   We will prune (or ignore) the distributed range.
1480 	 *
1481 	 *   FIXME: SBA code should register all elmmio ranges first.
1482 	 *      that would take care of elmmio ranges routed
1483 	 *	to a different rope (already discovered) from
1484 	 *	getting registered *after* LBA code has already
1485 	 *	registered it's distributed lmmio range.
1486 	 */
1487 	if (truncate_pat_collision(&iomem_resource,
1488 				   &(lba_dev->hba.lmmio_space))) {
1489 		printk(KERN_WARNING "LBA: lmmio_space [%lx/%lx] duplicate!\n",
1490 				(long)lba_dev->hba.lmmio_space.start,
1491 				(long)lba_dev->hba.lmmio_space.end);
1492 		lba_dev->hba.lmmio_space.flags = 0;
1493 	}
1494 
1495 	pci_add_resource_offset(&resources, &lba_dev->hba.io_space,
1496 				HBA_PORT_BASE(lba_dev->hba.hba_num));
1497 	if (lba_dev->hba.elmmio_space.start)
1498 		pci_add_resource_offset(&resources, &lba_dev->hba.elmmio_space,
1499 					lba_dev->hba.lmmio_space_offset);
1500 	if (lba_dev->hba.lmmio_space.flags)
1501 		pci_add_resource_offset(&resources, &lba_dev->hba.lmmio_space,
1502 					lba_dev->hba.lmmio_space_offset);
1503 	if (lba_dev->hba.gmmio_space.flags)
1504 		pci_add_resource(&resources, &lba_dev->hba.gmmio_space);
1505 
1506 	pci_add_resource(&resources, &lba_dev->hba.bus_num);
1507 
1508 	dev->dev.platform_data = lba_dev;
1509 	lba_bus = lba_dev->hba.hba_bus =
1510 		pci_create_root_bus(&dev->dev, lba_dev->hba.bus_num.start,
1511 				    cfg_ops, NULL, &resources);
1512 	if (!lba_bus) {
1513 		pci_free_resource_list(&resources);
1514 		return 0;
1515 	}
1516 
1517 	max = pci_scan_child_bus(lba_bus);
1518 
1519 	/* This is in lieu of calling pci_assign_unassigned_resources() */
1520 	if (is_pdc_pat()) {
1521 		/* assign resources to un-initialized devices */
1522 
1523 		DBG_PAT("LBA pci_bus_size_bridges()\n");
1524 		pci_bus_size_bridges(lba_bus);
1525 
1526 		DBG_PAT("LBA pci_bus_assign_resources()\n");
1527 		pci_bus_assign_resources(lba_bus);
1528 
1529 #ifdef DEBUG_LBA_PAT
1530 		DBG_PAT("\nLBA PIOP resource tree\n");
1531 		lba_dump_res(&lba_dev->hba.io_space, 2);
1532 		DBG_PAT("\nLBA LMMIO resource tree\n");
1533 		lba_dump_res(&lba_dev->hba.lmmio_space, 2);
1534 #endif
1535 	}
1536 	pci_enable_bridges(lba_bus);
1537 
1538 	/*
1539 	** Once PCI register ops has walked the bus, access to config
1540 	** space is restricted. Avoids master aborts on config cycles.
1541 	** Early LBA revs go fatal on *any* master abort.
1542 	*/
1543 	if (cfg_ops == &elroy_cfg_ops) {
1544 		lba_dev->flags |= LBA_FLAG_SKIP_PROBE;
1545 	}
1546 
1547 	lba_next_bus = max + 1;
1548 	pci_bus_add_devices(lba_bus);
1549 
1550 	/* Whew! Finally done! Tell services we got this one covered. */
1551 	return 0;
1552 }
1553 
1554 static struct parisc_device_id lba_tbl[] = {
1555 	{ HPHW_BRIDGE, HVERSION_REV_ANY_ID, ELROY_HVERS, 0xa },
1556 	{ HPHW_BRIDGE, HVERSION_REV_ANY_ID, MERCURY_HVERS, 0xa },
1557 	{ HPHW_BRIDGE, HVERSION_REV_ANY_ID, QUICKSILVER_HVERS, 0xa },
1558 	{ 0, }
1559 };
1560 
1561 static struct parisc_driver lba_driver = {
1562 	.name =		MODULE_NAME,
1563 	.id_table =	lba_tbl,
1564 	.probe =	lba_driver_probe,
1565 };
1566 
1567 /*
1568 ** One time initialization to let the world know the LBA was found.
1569 ** Must be called exactly once before pci_init().
1570 */
1571 void __init lba_init(void)
1572 {
1573 	register_parisc_driver(&lba_driver);
1574 }
1575 
1576 /*
1577 ** Initialize the IBASE/IMASK registers for LBA (Elroy).
1578 ** Only called from sba_iommu.c in order to route ranges (MMIO vs DMA).
1579 ** sba_iommu is responsible for locking (none needed at init time).
1580 */
1581 void lba_set_iregs(struct parisc_device *lba, u32 ibase, u32 imask)
1582 {
1583 	void __iomem * base_addr = ioremap_nocache(lba->hpa.start, 4096);
1584 
1585 	imask <<= 2;	/* adjust for hints - 2 more bits */
1586 
1587 	/* Make sure we aren't trying to set bits that aren't writeable. */
1588 	WARN_ON((ibase & 0x001fffff) != 0);
1589 	WARN_ON((imask & 0x001fffff) != 0);
1590 
1591 	DBG("%s() ibase 0x%x imask 0x%x\n", __func__, ibase, imask);
1592 	WRITE_REG32( imask, base_addr + LBA_IMASK);
1593 	WRITE_REG32( ibase, base_addr + LBA_IBASE);
1594 	iounmap(base_addr);
1595 }
1596 
1597