2 ** System Bus Adapter (SBA) I/O MMU manager
4 ** (c) Copyright 2000-2004 Grant Grundler <grundler @ parisc-linux x org>
5 ** (c) Copyright 2004 Naresh Kumar Inna <knaresh at india x hp x com>
6 ** (c) Copyright 2000-2004 Hewlett-Packard Company
8 ** Portions (c) 1999 Dave S. Miller (from sparc64 I/O MMU code)
10 ** This program is free software; you can redistribute it and/or modify
11 ** it under the terms of the GNU General Public License as published by
12 ** the Free Software Foundation; either version 2 of the License, or
13 ** (at your option) any later version.
16 ** This module initializes the IOC (I/O Controller) found on B1000/C3000/
17 ** J5000/J7000/N-class/L-class machines and their successors.
19 ** FIXME: add DMA hint support programming in both sba and lba modules.
22 #include <linux/types.h>
23 #include <linux/kernel.h>
24 #include <linux/spinlock.h>
25 #include <linux/slab.h>
26 #include <linux/init.h>
29 #include <linux/string.h>
30 #include <linux/pci.h>
32 #include <asm/byteorder.h>
34 #include <asm/dma.h> /* for DMA_CHUNK_SIZE */
36 #include <asm/hardware.h> /* for register_parisc_driver() stuff */
38 #include <linux/proc_fs.h>
39 #include <linux/seq_file.h>
41 #include <asm/runway.h> /* for proc_runway_root */
42 #include <asm/pdc.h> /* for PDC_MODEL_* */
43 #include <asm/pdcpat.h> /* for is_pdc_pat() */
44 #include <asm/parisc-device.h>
47 /* declared in arch/parisc/kernel/setup.c */
48 extern struct proc_dir_entry * proc_mckinley_root;
50 #define MODULE_NAME "SBA"
53 /* depends on proc fs support. But costs CPU performance */
54 #undef SBA_COLLECT_STATS
58 ** The number of debug flags is a clue - this code is fragile.
59 ** Don't even think about messing with it unless you have
60 ** plenty of 710's to sacrifice to the computer gods. :^)
64 #undef DEBUG_SBA_RUN_SG
65 #undef DEBUG_SBA_RESOURCE
66 #undef ASSERT_PDIR_SANITY
67 #undef DEBUG_LARGE_SG_ENTRIES
71 #define DBG_INIT(x...) printk(x)
73 #define DBG_INIT(x...)
77 #define DBG_RUN(x...) printk(x)
82 #ifdef DEBUG_SBA_RUN_SG
83 #define DBG_RUN_SG(x...) printk(x)
85 #define DBG_RUN_SG(x...)
89 #ifdef DEBUG_SBA_RESOURCE
90 #define DBG_RES(x...) printk(x)
95 #if defined(CONFIG_64BIT)
96 /* "low end" PA8800 machines use ZX1 chipset: PAT PDC and only run 64-bit */
100 #define SBA_INLINE __inline__
104 ** The number of pdir entries to "free" before issueing
105 ** a read to PCOM register to flush out PCOM writes.
106 ** Interacts with allocation granularity (ie 4 or 8 entries
107 ** allocated and free'd/purged at a time might make this
108 ** less interesting).
110 #define DELAYED_RESOURCE_CNT 16
112 #define DEFAULT_DMA_HINT_REG 0
114 #define ASTRO_RUNWAY_PORT 0x582
115 #define IKE_MERCED_PORT 0x803
116 #define REO_MERCED_PORT 0x804
117 #define REOG_MERCED_PORT 0x805
118 #define PLUTO_MCKINLEY_PORT 0x880
120 #define SBA_FUNC_ID 0x0000 /* function id */
121 #define SBA_FCLASS 0x0008 /* function class, bist, header, rev... */
123 #define IS_ASTRO(id) ((id)->hversion == ASTRO_RUNWAY_PORT)
124 #define IS_IKE(id) ((id)->hversion == IKE_MERCED_PORT)
125 #define IS_PLUTO(id) ((id)->hversion == PLUTO_MCKINLEY_PORT)
127 #define SBA_FUNC_SIZE 4096 /* SBA configuration function reg set */
129 #define ASTRO_IOC_OFFSET (32 * SBA_FUNC_SIZE)
130 #define PLUTO_IOC_OFFSET (1 * SBA_FUNC_SIZE)
131 /* Ike's IOC's occupy functions 2 and 3 */
132 #define IKE_IOC_OFFSET(p) ((p+2) * SBA_FUNC_SIZE)
134 #define IOC_CTRL 0x8 /* IOC_CTRL offset */
135 #define IOC_CTRL_TC (1 << 0) /* TOC Enable */
136 #define IOC_CTRL_CE (1 << 1) /* Coalesce Enable */
137 #define IOC_CTRL_DE (1 << 2) /* Dillon Enable */
138 #define IOC_CTRL_RM (1 << 8) /* Real Mode */
139 #define IOC_CTRL_NC (1 << 9) /* Non Coherent Mode */
140 #define IOC_CTRL_D4 (1 << 11) /* Disable 4-byte coalescing */
141 #define IOC_CTRL_DD (1 << 13) /* Disable distr. LMMIO range coalescing */
143 #define MAX_IOC 2 /* per Ike. Pluto/Astro only have 1. */
145 #define ROPES_PER_IOC 8 /* per Ike half or Pluto/Astro */
149 ** Offsets into MBIB (Function 0 on Ike and hopefully Astro)
150 ** Firmware programs this stuff. Don't touch it.
152 #define LMMIO_DIRECT0_BASE 0x300
153 #define LMMIO_DIRECT0_MASK 0x308
154 #define LMMIO_DIRECT0_ROUTE 0x310
156 #define LMMIO_DIST_BASE 0x360
157 #define LMMIO_DIST_MASK 0x368
158 #define LMMIO_DIST_ROUTE 0x370
160 #define IOS_DIST_BASE 0x390
161 #define IOS_DIST_MASK 0x398
162 #define IOS_DIST_ROUTE 0x3A0
164 #define IOS_DIRECT_BASE 0x3C0
165 #define IOS_DIRECT_MASK 0x3C8
166 #define IOS_DIRECT_ROUTE 0x3D0
169 ** Offsets into I/O TLB (Function 2 and 3 on Ike)
171 #define ROPE0_CTL 0x200 /* "regbus pci0" */
172 #define ROPE1_CTL 0x208
173 #define ROPE2_CTL 0x210
174 #define ROPE3_CTL 0x218
175 #define ROPE4_CTL 0x220
176 #define ROPE5_CTL 0x228
177 #define ROPE6_CTL 0x230
178 #define ROPE7_CTL 0x238
180 #define IOC_ROPE0_CFG 0x500 /* pluto only */
181 #define IOC_ROPE_AO 0x10 /* Allow "Relaxed Ordering" */
185 #define HF_ENABLE 0x40
188 #define IOC_IBASE 0x300 /* IO TLB */
189 #define IOC_IMASK 0x308
190 #define IOC_PCOM 0x310
191 #define IOC_TCNFG 0x318
192 #define IOC_PDIR_BASE 0x320
194 /* AGP GART driver looks for this */
195 #define SBA_IOMMU_COOKIE 0x0000badbadc0ffeeUL
199 ** IOC supports 4/8/16/64KB page sizes (see TCNFG register)
200 ** It's safer (avoid memory corruption) to keep DMA page mappings
201 ** equivalently sized to VM PAGE_SIZE.
203 ** We really can't avoid generating a new mapping for each
204 ** page since the Virtual Coherence Index has to be generated
205 ** and updated for each page.
207 ** PAGE_SIZE could be greater than IOVP_SIZE. But not the inverse.
209 #define IOVP_SIZE PAGE_SIZE
210 #define IOVP_SHIFT PAGE_SHIFT
211 #define IOVP_MASK PAGE_MASK
213 #define SBA_PERF_CFG 0x708 /* Performance Counter stuff */
214 #define SBA_PERF_MASK1 0x718
215 #define SBA_PERF_MASK2 0x730
219 ** Offsets into PCI Performance Counters (functions 12 and 13)
220 ** Controlled by PERF registers in function 2 & 3 respectively.
222 #define SBA_PERF_CNT1 0x200
223 #define SBA_PERF_CNT2 0x208
224 #define SBA_PERF_CNT3 0x210
228 void __iomem *ioc_hpa; /* I/O MMU base address */
229 char *res_map; /* resource map, bit == pdir entry */
230 u64 *pdir_base; /* physical base address */
231 unsigned long ibase; /* pdir IOV Space base - shared w/lba_pci */
232 unsigned long imask; /* pdir IOV Space mask - shared w/lba_pci */
234 unsigned long iovp_mask; /* help convert IOVA to IOVP */
236 unsigned long *res_hint; /* next avail IOVP - circular search */
238 unsigned int res_bitshift; /* from the LEFT! */
239 unsigned int res_size; /* size of resource map in bytes */
240 #ifdef SBA_HINT_SUPPORT
241 /* FIXME : DMA HINTs not used */
242 unsigned long hint_mask_pdir; /* bits used for DMA hints */
243 unsigned int hint_shift_pdir;
245 #if DELAYED_RESOURCE_CNT > 0
247 struct sba_dma_pair {
250 } saved[DELAYED_RESOURCE_CNT];
253 #ifdef SBA_COLLECT_STATS
254 #define SBA_SEARCH_SAMPLE 0x100
255 unsigned long avg_search[SBA_SEARCH_SAMPLE];
256 unsigned long avg_idx; /* current index into avg_search */
257 unsigned long used_pages;
258 unsigned long msingle_calls;
259 unsigned long msingle_pages;
260 unsigned long msg_calls;
261 unsigned long msg_pages;
262 unsigned long usingle_calls;
263 unsigned long usingle_pages;
264 unsigned long usg_calls;
265 unsigned long usg_pages;
268 /* STUFF We don't need in performance path */
269 unsigned int pdir_size; /* in bytes, determined by IOV Space size */
273 struct sba_device *next; /* list of SBA's in system */
274 struct parisc_device *dev; /* dev found in bus walk */
275 struct parisc_device_id *iodc; /* data about dev from firmware */
277 void __iomem *sba_hpa; /* base address */
279 unsigned int flags; /* state/functionality enabled */
280 unsigned int hw_rev; /* HW revision of chip */
282 struct resource chip_resv; /* MMIO reserved for chip */
283 struct resource iommu_resv; /* MMIO reserved for iommu */
285 unsigned int num_ioc; /* number of on-board IOC's */
286 struct ioc ioc[MAX_IOC];
290 static struct sba_device *sba_list;
292 static unsigned long ioc_needs_fdc = 0;
294 /* global count of IOMMUs in the system */
295 static unsigned int global_ioc_cnt = 0;
297 /* PA8700 (Piranha 2.2) bug workaround */
298 static unsigned long piranha_bad_128k = 0;
300 /* Looks nice and keeps the compiler happy */
301 #define SBA_DEV(d) ((struct sba_device *) (d))
303 #ifdef SBA_AGP_SUPPORT
304 static int reserve_sba_gart = 1;
307 #define ROUNDUP(x,y) ((x + ((y)-1)) & ~((y)-1))
310 /************************************
311 ** SBA register read and write support
313 ** BE WARNED: register writes are posted.
314 ** (ie follow writes which must reach HW with a read)
316 ** Superdome (in particular, REO) allows only 64-bit CSR accesses.
318 #define READ_REG32(addr) readl(addr)
319 #define READ_REG64(addr) readq(addr)
320 #define WRITE_REG32(val, addr) writel((val), (addr))
321 #define WRITE_REG64(val, addr) writeq((val), (addr))
324 #define READ_REG(addr) READ_REG64(addr)
325 #define WRITE_REG(value, addr) WRITE_REG64(value, addr)
327 #define READ_REG(addr) READ_REG32(addr)
328 #define WRITE_REG(value, addr) WRITE_REG32(value, addr)
331 #ifdef DEBUG_SBA_INIT
333 /* NOTE: When CONFIG_64BIT isn't defined, READ_REG64() is two 32-bit reads */
336 * sba_dump_ranges - debugging only - print ranges assigned to this IOA
337 * @hpa: base address of the sba
339 * Print the MMIO and IO Port address ranges forwarded by an Astro/Ike/RIO
340 * IO Adapter (aka Bus Converter).
343 sba_dump_ranges(void __iomem *hpa)
345 DBG_INIT("SBA at 0x%p\n", hpa);
346 DBG_INIT("IOS_DIST_BASE : %Lx\n", READ_REG64(hpa+IOS_DIST_BASE));
347 DBG_INIT("IOS_DIST_MASK : %Lx\n", READ_REG64(hpa+IOS_DIST_MASK));
348 DBG_INIT("IOS_DIST_ROUTE : %Lx\n", READ_REG64(hpa+IOS_DIST_ROUTE));
350 DBG_INIT("IOS_DIRECT_BASE : %Lx\n", READ_REG64(hpa+IOS_DIRECT_BASE));
351 DBG_INIT("IOS_DIRECT_MASK : %Lx\n", READ_REG64(hpa+IOS_DIRECT_MASK));
352 DBG_INIT("IOS_DIRECT_ROUTE: %Lx\n", READ_REG64(hpa+IOS_DIRECT_ROUTE));
356 * sba_dump_tlb - debugging only - print IOMMU operating parameters
357 * @hpa: base address of the IOMMU
359 * Print the size/location of the IO MMU PDIR.
361 static void sba_dump_tlb(void __iomem *hpa)
363 DBG_INIT("IO TLB at 0x%p\n", hpa);
364 DBG_INIT("IOC_IBASE : 0x%Lx\n", READ_REG64(hpa+IOC_IBASE));
365 DBG_INIT("IOC_IMASK : 0x%Lx\n", READ_REG64(hpa+IOC_IMASK));
366 DBG_INIT("IOC_TCNFG : 0x%Lx\n", READ_REG64(hpa+IOC_TCNFG));
367 DBG_INIT("IOC_PDIR_BASE: 0x%Lx\n", READ_REG64(hpa+IOC_PDIR_BASE));
371 #define sba_dump_ranges(x)
372 #define sba_dump_tlb(x)
373 #endif /* DEBUG_SBA_INIT */
376 #ifdef ASSERT_PDIR_SANITY
379 * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry
380 * @ioc: IO MMU structure which owns the pdir we are interested in.
381 * @msg: text to print ont the output line.
384 * Print one entry of the IO MMU PDIR in human readable form.
387 sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide)
389 /* start printing from lowest pde in rval */
390 u64 *ptr = &(ioc->pdir_base[pide & (~0U * BITS_PER_LONG)]);
391 unsigned long *rptr = (unsigned long *) &(ioc->res_map[(pide >>3) & ~(sizeof(unsigned long) - 1)]);
394 printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n",
396 rptr, pide & (BITS_PER_LONG - 1), *rptr);
399 while (rcnt < BITS_PER_LONG) {
400 printk(KERN_DEBUG "%s %2d %p %016Lx\n",
401 (rcnt == (pide & (BITS_PER_LONG - 1)))
407 printk(KERN_DEBUG "%s", msg);
412 * sba_check_pdir - debugging only - consistency checker
413 * @ioc: IO MMU structure which owns the pdir we are interested in.
414 * @msg: text to print ont the output line.
416 * Verify the resource map and pdir state is consistent
419 sba_check_pdir(struct ioc *ioc, char *msg)
421 u32 *rptr_end = (u32 *) &(ioc->res_map[ioc->res_size]);
422 u32 *rptr = (u32 *) ioc->res_map; /* resource map ptr */
423 u64 *pptr = ioc->pdir_base; /* pdir ptr */
426 while (rptr < rptr_end) {
428 int rcnt = 32; /* number of bits we might check */
431 /* Get last byte and highest bit from that */
432 u32 pde = ((u32) (((char *)pptr)[7])) << 24;
433 if ((rval ^ pde) & 0x80000000)
436 ** BUMMER! -- res_map != pdir --
437 ** Dump rval and matching pdir entries
439 sba_dump_pdir_entry(ioc, msg, pide);
443 rval <<= 1; /* try the next bit */
447 rptr++; /* look at next word of res_map */
449 /* It'd be nice if we always got here :^) */
455 * sba_dump_sg - debugging only - print Scatter-Gather list
456 * @ioc: IO MMU structure which owns the pdir we are interested in.
457 * @startsg: head of the SG list
458 * @nents: number of entries in SG list
460 * print the SG list so we can verify it's correct by hand.
463 sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
465 while (nents-- > 0) {
466 printk(KERN_DEBUG " %d : %08lx/%05x %p/%05x\n",
468 (unsigned long) sg_dma_address(startsg),
470 sg_virt_addr(startsg), startsg->length);
475 #endif /* ASSERT_PDIR_SANITY */
480 /**************************************************************
482 * I/O Pdir Resource Management
484 * Bits set in the resource map are in use.
485 * Each bit can represent a number of pages.
486 * LSbs represent lower addresses (IOVA's).
488 ***************************************************************/
489 #define PAGES_PER_RANGE 1 /* could increase this to 4 or 8 if needed */
491 /* Convert from IOVP to IOVA and vice versa. */
494 /* Pluto (aka ZX1) boxes need to set or clear the ibase bits appropriately */
495 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((ioc->ibase) | (iovp) | (offset))
496 #define SBA_IOVP(ioc,iova) ((iova) & (ioc)->iovp_mask)
498 /* only support Astro and ancestors. Saves a few cycles in key places */
499 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((iovp) | (offset))
500 #define SBA_IOVP(ioc,iova) (iova)
503 #define PDIR_INDEX(iovp) ((iovp)>>IOVP_SHIFT)
505 #define RESMAP_MASK(n) (~0UL << (BITS_PER_LONG - (n)))
506 #define RESMAP_IDX_MASK (sizeof(unsigned long) - 1)
510 * sba_search_bitmap - find free space in IO PDIR resource bitmap
511 * @ioc: IO MMU structure which owns the pdir we are interested in.
512 * @bits_wanted: number of entries we need.
514 * Find consecutive free bits in resource bitmap.
515 * Each bit represents one entry in the IO Pdir.
516 * Cool perf optimization: search for log2(size) bits at a time.
518 static SBA_INLINE unsigned long
519 sba_search_bitmap(struct ioc *ioc, unsigned long bits_wanted)
521 unsigned long *res_ptr = ioc->res_hint;
522 unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]);
523 unsigned long pide = ~0UL;
525 if (bits_wanted > (BITS_PER_LONG/2)) {
526 /* Search word at a time - no mask needed */
527 for(; res_ptr < res_end; ++res_ptr) {
529 *res_ptr = RESMAP_MASK(bits_wanted);
530 pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
531 pide <<= 3; /* convert to bit address */
535 /* point to the next word on next pass */
537 ioc->res_bitshift = 0;
540 ** Search the resource bit map on well-aligned values.
541 ** "o" is the alignment.
542 ** We need the alignment to invalidate I/O TLB using
543 ** SBA HW features in the unmap path.
545 unsigned long o = 1 << get_order(bits_wanted << PAGE_SHIFT);
546 uint bitshiftcnt = ROUNDUP(ioc->res_bitshift, o);
549 if (bitshiftcnt >= BITS_PER_LONG) {
553 mask = RESMAP_MASK(bits_wanted) >> bitshiftcnt;
555 DBG_RES("%s() o %ld %p", __FUNCTION__, o, res_ptr);
556 while(res_ptr < res_end)
558 DBG_RES(" %p %lx %lx\n", res_ptr, mask, *res_ptr);
560 if(((*res_ptr) & mask) == 0) {
561 *res_ptr |= mask; /* mark resources busy! */
562 pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
563 pide <<= 3; /* convert to bit address */
570 mask = RESMAP_MASK(bits_wanted);
575 /* look in the same word on the next pass */
576 ioc->res_bitshift = bitshiftcnt + bits_wanted;
580 if (res_end <= res_ptr) {
581 ioc->res_hint = (unsigned long *) ioc->res_map;
582 ioc->res_bitshift = 0;
584 ioc->res_hint = res_ptr;
591 * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
592 * @ioc: IO MMU structure which owns the pdir we are interested in.
593 * @size: number of bytes to create a mapping for
595 * Given a size, find consecutive unmarked and then mark those bits in the
599 sba_alloc_range(struct ioc *ioc, size_t size)
601 unsigned int pages_needed = size >> IOVP_SHIFT;
602 #ifdef SBA_COLLECT_STATS
603 unsigned long cr_start = mfctl(16);
607 pide = sba_search_bitmap(ioc, pages_needed);
608 if (pide >= (ioc->res_size << 3)) {
609 pide = sba_search_bitmap(ioc, pages_needed);
610 if (pide >= (ioc->res_size << 3))
611 panic("%s: I/O MMU @ %p is out of mapping resources\n",
612 __FILE__, ioc->ioc_hpa);
615 #ifdef ASSERT_PDIR_SANITY
616 /* verify the first enable bit is clear */
617 if(0x00 != ((u8 *) ioc->pdir_base)[pide*sizeof(u64) + 7]) {
618 sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide);
622 DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
623 __FUNCTION__, size, pages_needed, pide,
624 (uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map),
627 #ifdef SBA_COLLECT_STATS
629 unsigned long cr_end = mfctl(16);
630 unsigned long tmp = cr_end - cr_start;
631 /* check for roll over */
632 cr_start = (cr_end < cr_start) ? -(tmp) : (tmp);
634 ioc->avg_search[ioc->avg_idx++] = cr_start;
635 ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;
637 ioc->used_pages += pages_needed;
645 * sba_free_range - unmark bits in IO PDIR resource bitmap
646 * @ioc: IO MMU structure which owns the pdir we are interested in.
647 * @iova: IO virtual address which was previously allocated.
648 * @size: number of bytes to create a mapping for
650 * clear bits in the ioc's resource map
652 static SBA_INLINE void
653 sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size)
655 unsigned long iovp = SBA_IOVP(ioc, iova);
656 unsigned int pide = PDIR_INDEX(iovp);
657 unsigned int ridx = pide >> 3; /* convert bit to byte address */
658 unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]);
660 int bits_not_wanted = size >> IOVP_SHIFT;
662 /* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
663 unsigned long m = RESMAP_MASK(bits_not_wanted) >> (pide & (BITS_PER_LONG - 1));
665 DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n",
666 __FUNCTION__, (uint) iova, size,
667 bits_not_wanted, m, pide, res_ptr, *res_ptr);
669 #ifdef SBA_COLLECT_STATS
670 ioc->used_pages -= bits_not_wanted;
677 /**************************************************************
679 * "Dynamic DMA Mapping" support (aka "Coherent I/O")
681 ***************************************************************/
683 #ifdef SBA_HINT_SUPPORT
684 #define SBA_DMA_HINT(ioc, val) ((val) << (ioc)->hint_shift_pdir)
687 typedef unsigned long space_t;
688 #define KERNEL_SPACE 0
691 * sba_io_pdir_entry - fill in one IO PDIR entry
692 * @pdir_ptr: pointer to IO PDIR entry
693 * @sid: process Space ID - currently only support KERNEL_SPACE
694 * @vba: Virtual CPU address of buffer to map
695 * @hint: DMA hint set to use for this mapping
697 * SBA Mapping Routine
699 * Given a virtual address (vba, arg2) and space id, (sid, arg1)
700 * sba_io_pdir_entry() loads the I/O PDIR entry pointed to by
702 * Using the bass-ackwards HP bit numbering, Each IO Pdir entry
703 * for Astro/Ike looks like:
707 * +-+---------------------+----------------------------------+----+--------+
708 * |V| U | PPN[43:12] | U | VI |
709 * +-+---------------------+----------------------------------+----+--------+
711 * Pluto is basically identical, supports fewer physical address bits:
714 * +-+------------------------+-------------------------------+----+--------+
715 * |V| U | PPN[39:12] | U | VI |
716 * +-+------------------------+-------------------------------+----+--------+
718 * V == Valid Bit (Most Significant Bit is bit 0)
720 * PPN == Physical Page Number
721 * VI == Virtual Index (aka Coherent Index)
723 * LPA instruction output is put into PPN field.
724 * LCI (Load Coherence Index) instruction provides the "VI" bits.
726 * We pre-swap the bytes since PCX-W is Big Endian and the
727 * IOMMU uses little endian for the pdir.
731 sba_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba,
734 u64 pa; /* physical address */
735 register unsigned ci; /* coherent index */
737 pa = virt_to_phys(vba);
741 asm("lci 0(%%sr1, %1), %0" : "=r" (ci) : "r" (vba));
742 pa |= (ci >> 12) & 0xff; /* move CI (8 bits) into lowest byte */
744 pa |= 0x8000000000000000ULL; /* set "valid" bit */
745 *pdir_ptr = cpu_to_le64(pa); /* swap and store into I/O Pdir */
748 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
749 * (bit #61, big endian), we have to flush and sync every time
750 * IO-PDIR is changed in Ike/Astro.
753 asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
758 * sba_mark_invalid - invalidate one or more IO PDIR entries
759 * @ioc: IO MMU structure which owns the pdir we are interested in.
760 * @iova: IO Virtual Address mapped earlier
761 * @byte_cnt: number of bytes this mapping covers.
763 * Marking the IO PDIR entry(ies) as Invalid and invalidate
764 * corresponding IO TLB entry. The Ike PCOM (Purge Command Register)
765 * is to purge stale entries in the IO TLB when unmapping entries.
767 * The PCOM register supports purging of multiple pages, with a minium
768 * of 1 page and a maximum of 2GB. Hardware requires the address be
769 * aligned to the size of the range being purged. The size of the range
770 * must be a power of 2. The "Cool perf optimization" in the
771 * allocation routine helps keep that true.
773 static SBA_INLINE void
774 sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
776 u32 iovp = (u32) SBA_IOVP(ioc,iova);
777 u64 *pdir_ptr = &ioc->pdir_base[PDIR_INDEX(iovp)];
779 #ifdef ASSERT_PDIR_SANITY
780 /* Assert first pdir entry is set.
782 ** Even though this is a big-endian machine, the entries
783 ** in the iopdir are little endian. That's why we look at
784 ** the byte at +7 instead of at +0.
786 if (0x80 != (((u8 *) pdir_ptr)[7])) {
787 sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp));
791 if (byte_cnt > IOVP_SIZE)
794 unsigned long entries_per_cacheline = ioc_needs_fdc ?
795 L1_CACHE_ALIGN(((unsigned long) pdir_ptr))
796 - (unsigned long) pdir_ptr;
800 /* set "size" field for PCOM */
801 iovp |= get_order(byte_cnt) + PAGE_SHIFT;
804 /* clear I/O Pdir entry "valid" bit first */
805 ((u8 *) pdir_ptr)[7] = 0;
807 asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
809 entries_per_cacheline = L1_CACHE_SHIFT - 3;
813 byte_cnt -= IOVP_SIZE;
814 } while (byte_cnt > IOVP_SIZE);
816 iovp |= IOVP_SHIFT; /* set "size" field for PCOM */
819 ** clear I/O PDIR entry "valid" bit.
820 ** We have to R/M/W the cacheline regardless how much of the
821 ** pdir entry that we clobber.
822 ** The rest of the entry would be useful for debugging if we
823 ** could dump core on HPMC.
825 ((u8 *) pdir_ptr)[7] = 0;
827 asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
829 WRITE_REG( SBA_IOVA(ioc, iovp, 0, 0), ioc->ioc_hpa+IOC_PCOM);
833 * sba_dma_supported - PCI driver can query DMA support
834 * @dev: instance of PCI owned by the driver that's asking
835 * @mask: number of address bits this PCI device can handle
837 * See Documentation/DMA-mapping.txt
839 static int sba_dma_supported( struct device *dev, u64 mask)
844 printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n");
849 /* Documentation/DMA-mapping.txt tells drivers to try 64-bit first,
850 * then fall back to 32-bit if that fails.
851 * We are just "encouraging" 32-bit DMA masks here since we can
852 * never allow IOMMU bypass unless we add special support for ZX1.
860 * check if mask is >= than the current max IO Virt Address
861 * The max IO Virt address will *always* < 30 bits.
863 return((int)(mask >= (ioc->ibase - 1 +
864 (ioc->pdir_size / sizeof(u64) * IOVP_SIZE) )));
869 * sba_map_single - map one buffer and return IOVA for DMA
870 * @dev: instance of PCI owned by the driver that's asking.
871 * @addr: driver buffer to map.
872 * @size: number of bytes to map in driver buffer.
873 * @direction: R/W or both.
875 * See Documentation/DMA-mapping.txt
878 sba_map_single(struct device *dev, void *addr, size_t size,
879 enum dma_data_direction direction)
890 /* save offset bits */
891 offset = ((dma_addr_t) (long) addr) & ~IOVP_MASK;
893 /* round up to nearest IOVP_SIZE */
894 size = (size + offset + ~IOVP_MASK) & IOVP_MASK;
896 spin_lock_irqsave(&ioc->res_lock, flags);
897 #ifdef ASSERT_PDIR_SANITY
898 sba_check_pdir(ioc,"Check before sba_map_single()");
901 #ifdef SBA_COLLECT_STATS
902 ioc->msingle_calls++;
903 ioc->msingle_pages += size >> IOVP_SHIFT;
905 pide = sba_alloc_range(ioc, size);
906 iovp = (dma_addr_t) pide << IOVP_SHIFT;
908 DBG_RUN("%s() 0x%p -> 0x%lx\n",
909 __FUNCTION__, addr, (long) iovp | offset);
911 pdir_start = &(ioc->pdir_base[pide]);
914 sba_io_pdir_entry(pdir_start, KERNEL_SPACE, (unsigned long) addr, 0);
916 DBG_RUN(" pdir 0x%p %02x%02x%02x%02x%02x%02x%02x%02x\n",
918 (u8) (((u8 *) pdir_start)[7]),
919 (u8) (((u8 *) pdir_start)[6]),
920 (u8) (((u8 *) pdir_start)[5]),
921 (u8) (((u8 *) pdir_start)[4]),
922 (u8) (((u8 *) pdir_start)[3]),
923 (u8) (((u8 *) pdir_start)[2]),
924 (u8) (((u8 *) pdir_start)[1]),
925 (u8) (((u8 *) pdir_start)[0])
933 /* force FDC ops in io_pdir_entry() to be visible to IOMMU */
935 asm volatile("sync" : : );
937 #ifdef ASSERT_PDIR_SANITY
938 sba_check_pdir(ioc,"Check after sba_map_single()");
940 spin_unlock_irqrestore(&ioc->res_lock, flags);
942 /* form complete address */
943 return SBA_IOVA(ioc, iovp, offset, DEFAULT_DMA_HINT_REG);
948 * sba_unmap_single - unmap one IOVA and free resources
949 * @dev: instance of PCI owned by the driver that's asking.
950 * @iova: IOVA of driver buffer previously mapped.
951 * @size: number of bytes mapped in driver buffer.
952 * @direction: R/W or both.
954 * See Documentation/DMA-mapping.txt
957 sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size,
958 enum dma_data_direction direction)
961 #if DELAYED_RESOURCE_CNT > 0
962 struct sba_dma_pair *d;
967 DBG_RUN("%s() iovp 0x%lx/%x\n", __FUNCTION__, (long) iova, size);
970 offset = iova & ~IOVP_MASK;
971 iova ^= offset; /* clear offset bits */
973 size = ROUNDUP(size, IOVP_SIZE);
975 spin_lock_irqsave(&ioc->res_lock, flags);
977 #ifdef SBA_COLLECT_STATS
978 ioc->usingle_calls++;
979 ioc->usingle_pages += size >> IOVP_SHIFT;
982 sba_mark_invalid(ioc, iova, size);
984 #if DELAYED_RESOURCE_CNT > 0
985 /* Delaying when we re-use a IO Pdir entry reduces the number
986 * of MMIO reads needed to flush writes to the PCOM register.
988 d = &(ioc->saved[ioc->saved_cnt]);
991 if (++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT) {
992 int cnt = ioc->saved_cnt;
994 sba_free_range(ioc, d->iova, d->size);
999 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */
1001 #else /* DELAYED_RESOURCE_CNT == 0 */
1002 sba_free_range(ioc, iova, size);
1004 /* If fdc's were issued, force fdc's to be visible now */
1006 asm volatile("sync" : : );
1008 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */
1009 #endif /* DELAYED_RESOURCE_CNT == 0 */
1011 spin_unlock_irqrestore(&ioc->res_lock, flags);
1013 /* XXX REVISIT for 2.5 Linux - need syncdma for zero-copy support.
1014 ** For Astro based systems this isn't a big deal WRT performance.
1015 ** As long as 2.4 kernels copyin/copyout data from/to userspace,
1016 ** we don't need the syncdma. The issue here is I/O MMU cachelines
1017 ** are *not* coherent in all cases. May be hwrev dependent.
1018 ** Need to investigate more.
1019 asm volatile("syncdma");
1025 * sba_alloc_consistent - allocate/map shared mem for DMA
1026 * @hwdev: instance of PCI owned by the driver that's asking.
1027 * @size: number of bytes mapped in driver buffer.
1028 * @dma_handle: IOVA of new buffer.
1030 * See Documentation/DMA-mapping.txt
1032 static void *sba_alloc_consistent(struct device *hwdev, size_t size,
1033 dma_addr_t *dma_handle, gfp_t gfp)
1038 /* only support PCI */
1043 ret = (void *) __get_free_pages(gfp, get_order(size));
1046 memset(ret, 0, size);
1047 *dma_handle = sba_map_single(hwdev, ret, size, 0);
1055 * sba_free_consistent - free/unmap shared mem for DMA
1056 * @hwdev: instance of PCI owned by the driver that's asking.
1057 * @size: number of bytes mapped in driver buffer.
1058 * @vaddr: virtual address IOVA of "consistent" buffer.
1059 * @dma_handler: IO virtual address of "consistent" buffer.
1061 * See Documentation/DMA-mapping.txt
1064 sba_free_consistent(struct device *hwdev, size_t size, void *vaddr,
1065 dma_addr_t dma_handle)
1067 sba_unmap_single(hwdev, dma_handle, size, 0);
1068 free_pages((unsigned long) vaddr, get_order(size));
1073 ** Since 0 is a valid pdir_base index value, can't use that
1074 ** to determine if a value is valid or not. Use a flag to indicate
1075 ** the SG list entry contains a valid pdir index.
1077 #define PIDE_FLAG 0x80000000UL
1079 #ifdef SBA_COLLECT_STATS
1080 #define IOMMU_MAP_STATS
1082 #include "iommu-helpers.h"
1084 #ifdef DEBUG_LARGE_SG_ENTRIES
1085 int dump_run_sg = 0;
1090 * sba_map_sg - map Scatter/Gather list
1091 * @dev: instance of PCI owned by the driver that's asking.
1092 * @sglist: array of buffer/length pairs
1093 * @nents: number of entries in list
1094 * @direction: R/W or both.
1096 * See Documentation/DMA-mapping.txt
1099 sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
1100 enum dma_data_direction direction)
1103 int coalesced, filled = 0;
1104 unsigned long flags;
1106 DBG_RUN_SG("%s() START %d entries\n", __FUNCTION__, nents);
1110 /* Fast path single entry scatterlists. */
1112 sg_dma_address(sglist) = sba_map_single(dev,
1113 (void *)sg_virt_addr(sglist),
1114 sglist->length, direction);
1115 sg_dma_len(sglist) = sglist->length;
1119 spin_lock_irqsave(&ioc->res_lock, flags);
1121 #ifdef ASSERT_PDIR_SANITY
1122 if (sba_check_pdir(ioc,"Check before sba_map_sg()"))
1124 sba_dump_sg(ioc, sglist, nents);
1125 panic("Check before sba_map_sg()");
1129 #ifdef SBA_COLLECT_STATS
1134 ** First coalesce the chunks and allocate I/O pdir space
1136 ** If this is one DMA stream, we can properly map using the
1137 ** correct virtual address associated with each DMA page.
1138 ** w/o this association, we wouldn't have coherent DMA!
1139 ** Access to the virtual address is what forces a two pass algorithm.
1141 coalesced = iommu_coalesce_chunks(ioc, sglist, nents, sba_alloc_range);
1144 ** Program the I/O Pdir
1146 ** map the virtual addresses to the I/O Pdir
1147 ** o dma_address will contain the pdir index
1148 ** o dma_len will contain the number of bytes to map
1149 ** o address contains the virtual address.
1151 filled = iommu_fill_pdir(ioc, sglist, nents, 0, sba_io_pdir_entry);
1153 /* force FDC ops in io_pdir_entry() to be visible to IOMMU */
1155 asm volatile("sync" : : );
1157 #ifdef ASSERT_PDIR_SANITY
1158 if (sba_check_pdir(ioc,"Check after sba_map_sg()"))
1160 sba_dump_sg(ioc, sglist, nents);
1161 panic("Check after sba_map_sg()\n");
1165 spin_unlock_irqrestore(&ioc->res_lock, flags);
1167 DBG_RUN_SG("%s() DONE %d mappings\n", __FUNCTION__, filled);
1174 * sba_unmap_sg - unmap Scatter/Gather list
1175 * @dev: instance of PCI owned by the driver that's asking.
1176 * @sglist: array of buffer/length pairs
1177 * @nents: number of entries in list
1178 * @direction: R/W or both.
1180 * See Documentation/DMA-mapping.txt
1183 sba_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents,
1184 enum dma_data_direction direction)
1187 #ifdef ASSERT_PDIR_SANITY
1188 unsigned long flags;
1191 DBG_RUN_SG("%s() START %d entries, %p,%x\n",
1192 __FUNCTION__, nents, sg_virt_addr(sglist), sglist->length);
1196 #ifdef SBA_COLLECT_STATS
1200 #ifdef ASSERT_PDIR_SANITY
1201 spin_lock_irqsave(&ioc->res_lock, flags);
1202 sba_check_pdir(ioc,"Check before sba_unmap_sg()");
1203 spin_unlock_irqrestore(&ioc->res_lock, flags);
1206 while (sg_dma_len(sglist) && nents--) {
1208 sba_unmap_single(dev, sg_dma_address(sglist), sg_dma_len(sglist), direction);
1209 #ifdef SBA_COLLECT_STATS
1210 ioc->usg_pages += ((sg_dma_address(sglist) & ~IOVP_MASK) + sg_dma_len(sglist) + IOVP_SIZE - 1) >> PAGE_SHIFT;
1211 ioc->usingle_calls--; /* kluge since call is unmap_sg() */
1216 DBG_RUN_SG("%s() DONE (nents %d)\n", __FUNCTION__, nents);
1218 #ifdef ASSERT_PDIR_SANITY
1219 spin_lock_irqsave(&ioc->res_lock, flags);
1220 sba_check_pdir(ioc,"Check after sba_unmap_sg()");
1221 spin_unlock_irqrestore(&ioc->res_lock, flags);
1226 static struct hppa_dma_ops sba_ops = {
1227 .dma_supported = sba_dma_supported,
1228 .alloc_consistent = sba_alloc_consistent,
1229 .alloc_noncoherent = sba_alloc_consistent,
1230 .free_consistent = sba_free_consistent,
1231 .map_single = sba_map_single,
1232 .unmap_single = sba_unmap_single,
1233 .map_sg = sba_map_sg,
1234 .unmap_sg = sba_unmap_sg,
1235 .dma_sync_single_for_cpu = NULL,
1236 .dma_sync_single_for_device = NULL,
1237 .dma_sync_sg_for_cpu = NULL,
1238 .dma_sync_sg_for_device = NULL,
1242 /**************************************************************************
1244 ** SBA PAT PDC support
1246 ** o call pdc_pat_cell_module()
1247 ** o store ranges in PCI "resource" structures
1249 **************************************************************************/
1252 sba_get_pat_resources(struct sba_device *sba_dev)
1256 ** TODO/REVISIT/FIXME: support for directed ranges requires calls to
1257 ** PAT PDC to program the SBA/LBA directed range registers...this
1258 ** burden may fall on the LBA code since it directly supports the
1259 ** PCI subsystem. It's not clear yet. - ggg
1261 PAT_MOD(mod)->mod_info.mod_pages = PAT_GET_MOD_PAGES(temp);
1263 PAT_MOD(mod)->mod_info.dvi = PAT_GET_DVI(temp);
1264 Tells where the dvi bits are located in the address.
1265 PAT_MOD(mod)->mod_info.ioc = PAT_GET_IOC(temp);
1271 /**************************************************************
1273 * Initialization and claim
1275 ***************************************************************/
1276 #define PIRANHA_ADDR_MASK 0x00160000UL /* bit 17,18,20 */
1277 #define PIRANHA_ADDR_VAL 0x00060000UL /* bit 17,18 on */
1279 sba_alloc_pdir(unsigned int pdir_size)
1281 unsigned long pdir_base;
1282 unsigned long pdir_order = get_order(pdir_size);
1284 pdir_base = __get_free_pages(GFP_KERNEL, pdir_order);
1285 if (NULL == (void *) pdir_base) {
1286 panic("%s() could not allocate I/O Page Table\n",
1290 /* If this is not PA8700 (PCX-W2)
1291 ** OR newer than ver 2.2
1292 ** OR in a system that doesn't need VINDEX bits from SBA,
1294 ** then we aren't exposed to the HW bug.
1296 if ( ((boot_cpu_data.pdc.cpuid >> 5) & 0x7f) != 0x13
1297 || (boot_cpu_data.pdc.versions > 0x202)
1298 || (boot_cpu_data.pdc.capabilities & 0x08L) )
1299 return (void *) pdir_base;
1302 * PA8700 (PCX-W2, aka piranha) silent data corruption fix
1304 * An interaction between PA8700 CPU (Ver 2.2 or older) and
1305 * Ike/Astro can cause silent data corruption. This is only
1306 * a problem if the I/O PDIR is located in memory such that
1307 * (little-endian) bits 17 and 18 are on and bit 20 is off.
1309 * Since the max IO Pdir size is 2MB, by cleverly allocating the
1310 * right physical address, we can either avoid (IOPDIR <= 1MB)
1311 * or minimize (2MB IO Pdir) the problem if we restrict the
1312 * IO Pdir to a maximum size of 2MB-128K (1902K).
1314 * Because we always allocate 2^N sized IO pdirs, either of the
1315 * "bad" regions will be the last 128K if at all. That's easy
1319 if (pdir_order <= (19-12)) {
1320 if (((virt_to_phys(pdir_base)+pdir_size-1) & PIRANHA_ADDR_MASK) == PIRANHA_ADDR_VAL) {
1321 /* allocate a new one on 512k alignment */
1322 unsigned long new_pdir = __get_free_pages(GFP_KERNEL, (19-12));
1323 /* release original */
1324 free_pages(pdir_base, pdir_order);
1326 pdir_base = new_pdir;
1328 /* release excess */
1329 while (pdir_order < (19-12)) {
1330 new_pdir += pdir_size;
1331 free_pages(new_pdir, pdir_order);
1339 ** Needs to be aligned on an "odd" 1MB boundary.
1341 unsigned long new_pdir = __get_free_pages(GFP_KERNEL, pdir_order+1); /* 2 or 4MB */
1343 /* release original */
1344 free_pages( pdir_base, pdir_order);
1346 /* release first 1MB */
1347 free_pages(new_pdir, 20-12);
1349 pdir_base = new_pdir + 1024*1024;
1351 if (pdir_order > (20-12)) {
1355 ** Flag tells init_bitmap() to mark bad 128k as used
1356 ** and to reduce the size by 128k.
1358 piranha_bad_128k = 1;
1360 new_pdir += 3*1024*1024;
1361 /* release last 1MB */
1362 free_pages(new_pdir, 20-12);
1364 /* release unusable 128KB */
1365 free_pages(new_pdir - 128*1024 , 17-12);
1367 pdir_size -= 128*1024;
1371 memset((void *) pdir_base, 0, pdir_size);
1372 return (void *) pdir_base;
1375 static struct device *next_device(struct klist_iter *i)
1377 struct klist_node * n = klist_next(i);
1378 return n ? container_of(n, struct device, knode_parent) : NULL;
1381 /* setup Mercury or Elroy IBASE/IMASK registers. */
1383 setup_ibase_imask(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1385 /* lba_set_iregs() is in drivers/parisc/lba_pci.c */
1386 extern void lba_set_iregs(struct parisc_device *, u32, u32);
1388 struct klist_iter i;
1390 klist_iter_init(&sba->dev.klist_children, &i);
1391 while ((dev = next_device(&i))) {
1392 struct parisc_device *lba = to_parisc_device(dev);
1393 int rope_num = (lba->hpa.start >> 13) & 0xf;
1394 if (rope_num >> 3 == ioc_num)
1395 lba_set_iregs(lba, ioc->ibase, ioc->imask);
1397 klist_iter_exit(&i);
1401 sba_ioc_init_pluto(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1403 u32 iova_space_mask;
1404 u32 iova_space_size;
1405 int iov_order, tcnfg;
1406 #ifdef SBA_AGP_SUPPORT
1410 ** Firmware programs the base and size of a "safe IOVA space"
1411 ** (one that doesn't overlap memory or LMMIO space) in the
1412 ** IBASE and IMASK registers.
1414 ioc->ibase = READ_REG(ioc->ioc_hpa + IOC_IBASE);
1415 iova_space_size = ~(READ_REG(ioc->ioc_hpa + IOC_IMASK) & 0xFFFFFFFFUL) + 1;
1417 if ((ioc->ibase < 0xfed00000UL) && ((ioc->ibase + iova_space_size) > 0xfee00000UL)) {
1418 printk("WARNING: IOV space overlaps local config and interrupt message, truncating\n");
1419 iova_space_size /= 2;
1423 ** iov_order is always based on a 1GB IOVA space since we want to
1424 ** turn on the other half for AGP GART.
1426 iov_order = get_order(iova_space_size >> (IOVP_SHIFT - PAGE_SHIFT));
1427 ioc->pdir_size = (iova_space_size / IOVP_SIZE) * sizeof(u64);
1429 DBG_INIT("%s() hpa 0x%p IOV %dMB (%d bits)\n",
1430 __FUNCTION__, ioc->ioc_hpa, iova_space_size >> 20,
1431 iov_order + PAGE_SHIFT);
1433 ioc->pdir_base = (void *) __get_free_pages(GFP_KERNEL,
1434 get_order(ioc->pdir_size));
1435 if (!ioc->pdir_base)
1436 panic("Couldn't allocate I/O Page Table\n");
1438 memset(ioc->pdir_base, 0, ioc->pdir_size);
1440 DBG_INIT("%s() pdir %p size %x\n",
1441 __FUNCTION__, ioc->pdir_base, ioc->pdir_size);
1443 #ifdef SBA_HINT_SUPPORT
1444 ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1445 ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1447 DBG_INIT(" hint_shift_pdir %x hint_mask_pdir %lx\n",
1448 ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1451 WARN_ON((((unsigned long) ioc->pdir_base) & PAGE_MASK) != (unsigned long) ioc->pdir_base);
1452 WRITE_REG(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1454 /* build IMASK for IOC and Elroy */
1455 iova_space_mask = 0xffffffff;
1456 iova_space_mask <<= (iov_order + PAGE_SHIFT);
1457 ioc->imask = iova_space_mask;
1459 ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1461 sba_dump_tlb(ioc->ioc_hpa);
1463 setup_ibase_imask(sba, ioc, ioc_num);
1465 WRITE_REG(ioc->imask, ioc->ioc_hpa + IOC_IMASK);
1469 ** Setting the upper bits makes checking for bypass addresses
1470 ** a little faster later on.
1472 ioc->imask |= 0xFFFFFFFF00000000UL;
1475 /* Set I/O PDIR Page size to system page size */
1476 switch (PAGE_SHIFT) {
1477 case 12: tcnfg = 0; break; /* 4K */
1478 case 13: tcnfg = 1; break; /* 8K */
1479 case 14: tcnfg = 2; break; /* 16K */
1480 case 16: tcnfg = 3; break; /* 64K */
1482 panic(__FILE__ "Unsupported system page size %d",
1486 WRITE_REG(tcnfg, ioc->ioc_hpa + IOC_TCNFG);
1489 ** Program the IOC's ibase and enable IOVA translation
1490 ** Bit zero == enable bit.
1492 WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa + IOC_IBASE);
1495 ** Clear I/O TLB of any possible entries.
1496 ** (Yes. This is a bit paranoid...but so what)
1498 WRITE_REG(ioc->ibase | 31, ioc->ioc_hpa + IOC_PCOM);
1500 #ifdef SBA_AGP_SUPPORT
1502 ** If an AGP device is present, only use half of the IOV space
1503 ** for PCI DMA. Unfortunately we can't know ahead of time
1504 ** whether GART support will actually be used, for now we
1505 ** can just key on any AGP device found in the system.
1506 ** We program the next pdir index after we stop w/ a key for
1507 ** the GART code to handshake on.
1510 for (lba = sba->child; lba; lba = lba->sibling) {
1511 if (IS_QUICKSILVER(lba))
1516 DBG_INIT("%s: Reserving half of IOVA space for AGP GART support\n", __FUNCTION__);
1517 ioc->pdir_size /= 2;
1518 ((u64 *)ioc->pdir_base)[PDIR_INDEX(iova_space_size/2)] = SBA_IOMMU_COOKIE;
1520 DBG_INIT("%s: No GART needed - no AGP controller found\n", __FUNCTION__);
1527 sba_ioc_init(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1529 u32 iova_space_size, iova_space_mask;
1530 unsigned int pdir_size, iov_order;
1533 ** Determine IOVA Space size from memory size.
1535 ** Ideally, PCI drivers would register the maximum number
1536 ** of DMA they can have outstanding for each device they
1537 ** own. Next best thing would be to guess how much DMA
1538 ** can be outstanding based on PCI Class/sub-class. Both
1539 ** methods still require some "extra" to support PCI
1540 ** Hot-Plug/Removal of PCI cards. (aka PCI OLARD).
1542 ** While we have 32-bits "IOVA" space, top two 2 bits are used
1543 ** for DMA hints - ergo only 30 bits max.
1546 iova_space_size = (u32) (num_physpages/global_ioc_cnt);
1548 /* limit IOVA space size to 1MB-1GB */
1549 if (iova_space_size < (1 << (20 - PAGE_SHIFT))) {
1550 iova_space_size = 1 << (20 - PAGE_SHIFT);
1552 else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) {
1553 iova_space_size = 1 << (30 - PAGE_SHIFT);
1557 ** iova space must be log2() in size.
1558 ** thus, pdir/res_map will also be log2().
1559 ** PIRANHA BUG: Exception is when IO Pdir is 2MB (gets reduced)
1561 iov_order = get_order(iova_space_size << PAGE_SHIFT);
1563 /* iova_space_size is now bytes, not pages */
1564 iova_space_size = 1 << (iov_order + PAGE_SHIFT);
1566 ioc->pdir_size = pdir_size = (iova_space_size/IOVP_SIZE) * sizeof(u64);
1568 DBG_INIT("%s() hpa 0x%lx mem %ldMB IOV %dMB (%d bits)\n",
1571 (unsigned long) num_physpages >> (20 - PAGE_SHIFT),
1572 iova_space_size>>20,
1573 iov_order + PAGE_SHIFT);
1575 ioc->pdir_base = sba_alloc_pdir(pdir_size);
1577 DBG_INIT("%s() pdir %p size %x\n",
1578 __FUNCTION__, ioc->pdir_base, pdir_size);
1580 #ifdef SBA_HINT_SUPPORT
1581 /* FIXME : DMA HINTs not used */
1582 ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1583 ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1585 DBG_INIT(" hint_shift_pdir %x hint_mask_pdir %lx\n",
1586 ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1589 WRITE_REG64(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1591 /* build IMASK for IOC and Elroy */
1592 iova_space_mask = 0xffffffff;
1593 iova_space_mask <<= (iov_order + PAGE_SHIFT);
1596 ** On C3000 w/512MB mem, HP-UX 10.20 reports:
1597 ** ibase=0, imask=0xFE000000, size=0x2000000.
1600 ioc->imask = iova_space_mask; /* save it */
1602 ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1605 DBG_INIT("%s() IOV base 0x%lx mask 0x%0lx\n",
1606 __FUNCTION__, ioc->ibase, ioc->imask);
1609 ** FIXME: Hint registers are programmed with default hint
1610 ** values during boot, so hints should be sane even if we
1611 ** can't reprogram them the way drivers want.
1614 setup_ibase_imask(sba, ioc, ioc_num);
1617 ** Program the IOC's ibase and enable IOVA translation
1619 WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa+IOC_IBASE);
1620 WRITE_REG(ioc->imask, ioc->ioc_hpa+IOC_IMASK);
1622 /* Set I/O PDIR Page size to 4K */
1623 WRITE_REG(0, ioc->ioc_hpa+IOC_TCNFG);
1626 ** Clear I/O TLB of any possible entries.
1627 ** (Yes. This is a bit paranoid...but so what)
1629 WRITE_REG(0 | 31, ioc->ioc_hpa+IOC_PCOM);
1631 ioc->ibase = 0; /* used by SBA_IOVA and related macros */
1633 DBG_INIT("%s() DONE\n", __FUNCTION__);
1638 /**************************************************************************
1640 ** SBA initialization code (HW and SW)
1642 ** o identify SBA chip itself
1643 ** o initialize SBA chip modes (HardFail)
1644 ** o initialize SBA chip modes (HardFail)
1645 ** o FIXME: initialize DMA hints for reasonable defaults
1647 **************************************************************************/
1649 static void __iomem *ioc_remap(struct sba_device *sba_dev, unsigned int offset)
1651 return ioremap_nocache(sba_dev->dev->hpa.start + offset, SBA_FUNC_SIZE);
1654 static void sba_hw_init(struct sba_device *sba_dev)
1660 if (!is_pdc_pat()) {
1661 /* Shutdown the USB controller on Astro-based workstations.
1662 ** Once we reprogram the IOMMU, the next DMA performed by
1663 ** USB will HPMC the box. USB is only enabled if a
1664 ** keyboard is present and found.
1666 ** With serial console, j6k v5.0 firmware says:
1667 ** mem_kbd hpa 0xfee003f8 sba 0x0 pad 0x0 cl_class 0x7
1669 ** FIXME: Using GFX+USB console at power up but direct
1670 ** linux to serial console is still broken.
1671 ** USB could generate DMA so we must reset USB.
1672 ** The proper sequence would be:
1673 ** o block console output
1674 ** o reset USB device
1675 ** o reprogram serial port
1676 ** o unblock console output
1678 if (PAGE0->mem_kbd.cl_class == CL_KEYBD) {
1679 pdc_io_reset_devices();
1686 printk("sba_hw_init(): mem_boot 0x%x 0x%x 0x%x 0x%x\n", PAGE0->mem_boot.hpa,
1687 PAGE0->mem_boot.spa, PAGE0->mem_boot.pad, PAGE0->mem_boot.cl_class);
1690 ** Need to deal with DMA from LAN.
1691 ** Maybe use page zero boot device as a handle to talk
1692 ** to PDC about which device to shutdown.
1694 ** Netbooting, j6k v5.0 firmware says:
1695 ** mem_boot hpa 0xf4008000 sba 0x0 pad 0x0 cl_class 0x1002
1696 ** ARGH! invalid class.
1698 if ((PAGE0->mem_boot.cl_class != CL_RANDOM)
1699 && (PAGE0->mem_boot.cl_class != CL_SEQU)) {
1704 if (!IS_PLUTO(sba_dev->iodc)) {
1705 ioc_ctl = READ_REG(sba_dev->sba_hpa+IOC_CTRL);
1706 DBG_INIT("%s() hpa 0x%lx ioc_ctl 0x%Lx ->",
1707 __FUNCTION__, sba_dev->sba_hpa, ioc_ctl);
1708 ioc_ctl &= ~(IOC_CTRL_RM | IOC_CTRL_NC | IOC_CTRL_CE);
1709 ioc_ctl |= IOC_CTRL_DD | IOC_CTRL_D4 | IOC_CTRL_TC;
1710 /* j6700 v1.6 firmware sets 0x294f */
1711 /* A500 firmware sets 0x4d */
1713 WRITE_REG(ioc_ctl, sba_dev->sba_hpa+IOC_CTRL);
1715 #ifdef DEBUG_SBA_INIT
1716 ioc_ctl = READ_REG64(sba_dev->sba_hpa+IOC_CTRL);
1717 DBG_INIT(" 0x%Lx\n", ioc_ctl);
1721 if (IS_ASTRO(sba_dev->iodc)) {
1723 /* PAT_PDC (L-class) also reports the same goofy base */
1724 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, ASTRO_IOC_OFFSET);
1727 sba_dev->chip_resv.name = "Astro Intr Ack";
1728 sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfef00000UL;
1729 sba_dev->chip_resv.end = PCI_F_EXTEND | (0xff000000UL - 1) ;
1730 err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1733 } else if (IS_PLUTO(sba_dev->iodc)) {
1736 /* We use a negative value for IOC HPA so it gets
1737 * corrected when we add it with IKE's IOC offset.
1738 * Doesnt look clean, but fewer code.
1740 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, PLUTO_IOC_OFFSET);
1743 sba_dev->chip_resv.name = "Pluto Intr/PIOP/VGA";
1744 sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfee00000UL;
1745 sba_dev->chip_resv.end = PCI_F_EXTEND | (0xff200000UL - 1);
1746 err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1749 sba_dev->iommu_resv.name = "IOVA Space";
1750 sba_dev->iommu_resv.start = 0x40000000UL;
1751 sba_dev->iommu_resv.end = 0x50000000UL - 1;
1752 err = request_resource(&iomem_resource, &(sba_dev->iommu_resv));
1755 /* IS_IKE (ie N-class, L3000, L1500) */
1756 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(0));
1757 sba_dev->ioc[1].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(1));
1760 /* TODO - LOOKUP Ike/Stretch chipset mem map */
1762 /* XXX: What about Reo? */
1764 sba_dev->num_ioc = num_ioc;
1765 for (i = 0; i < num_ioc; i++) {
1766 void __iomem *ioc_hpa = sba_dev->ioc[i].ioc_hpa;
1769 for (j=0; j < sizeof(u64) * ROPES_PER_IOC; j+=sizeof(u64)) {
1772 * Clear ROPE(N)_CONFIG AO bit.
1773 * Disables "NT Ordering" (~= !"Relaxed Ordering")
1774 * Overrides bit 1 in DMA Hint Sets.
1775 * Improves netperf UDP_STREAM by ~10% for bcm5701.
1777 if (IS_PLUTO(sba_dev->iodc)) {
1778 void __iomem *rope_cfg;
1779 unsigned long cfg_val;
1781 rope_cfg = ioc_hpa + IOC_ROPE0_CFG + j;
1782 cfg_val = READ_REG(rope_cfg);
1783 cfg_val &= ~IOC_ROPE_AO;
1784 WRITE_REG(cfg_val, rope_cfg);
1788 ** Make sure the box crashes on rope errors.
1790 WRITE_REG(HF_ENABLE, ioc_hpa + ROPE0_CTL + j);
1793 /* flush out the last writes */
1794 READ_REG(sba_dev->ioc[i].ioc_hpa + ROPE7_CTL);
1796 DBG_INIT(" ioc[%d] ROPE_CFG 0x%Lx ROPE_DBG 0x%Lx\n",
1798 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x40),
1799 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x50)
1801 DBG_INIT(" STATUS_CONTROL 0x%Lx FLUSH_CTRL 0x%Lx\n",
1802 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x108),
1803 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x400)
1806 if (IS_PLUTO(sba_dev->iodc)) {
1807 sba_ioc_init_pluto(sba_dev->dev, &(sba_dev->ioc[i]), i);
1809 sba_ioc_init(sba_dev->dev, &(sba_dev->ioc[i]), i);
1815 sba_common_init(struct sba_device *sba_dev)
1819 /* add this one to the head of the list (order doesn't matter)
1820 ** This will be useful for debugging - especially if we get coredumps
1822 sba_dev->next = sba_list;
1825 for(i=0; i< sba_dev->num_ioc; i++) {
1827 #ifdef DEBUG_DMB_TRAP
1828 extern void iterate_pages(unsigned long , unsigned long ,
1829 void (*)(pte_t * , unsigned long),
1831 void set_data_memory_break(pte_t * , unsigned long);
1833 /* resource map size dictated by pdir_size */
1834 res_size = sba_dev->ioc[i].pdir_size/sizeof(u64); /* entries */
1836 /* Second part of PIRANHA BUG */
1837 if (piranha_bad_128k) {
1838 res_size -= (128*1024)/sizeof(u64);
1841 res_size >>= 3; /* convert bit count to byte count */
1842 DBG_INIT("%s() res_size 0x%x\n",
1843 __FUNCTION__, res_size);
1845 sba_dev->ioc[i].res_size = res_size;
1846 sba_dev->ioc[i].res_map = (char *) __get_free_pages(GFP_KERNEL, get_order(res_size));
1848 #ifdef DEBUG_DMB_TRAP
1849 iterate_pages( sba_dev->ioc[i].res_map, res_size,
1850 set_data_memory_break, 0);
1853 if (NULL == sba_dev->ioc[i].res_map)
1855 panic("%s:%s() could not allocate resource map\n",
1856 __FILE__, __FUNCTION__ );
1859 memset(sba_dev->ioc[i].res_map, 0, res_size);
1860 /* next available IOVP - circular search */
1861 sba_dev->ioc[i].res_hint = (unsigned long *)
1862 &(sba_dev->ioc[i].res_map[L1_CACHE_BYTES]);
1864 #ifdef ASSERT_PDIR_SANITY
1865 /* Mark first bit busy - ie no IOVA 0 */
1866 sba_dev->ioc[i].res_map[0] = 0x80;
1867 sba_dev->ioc[i].pdir_base[0] = 0xeeffc0addbba0080ULL;
1870 /* Third (and last) part of PIRANHA BUG */
1871 if (piranha_bad_128k) {
1872 /* region from +1408K to +1536 is un-usable. */
1874 int idx_start = (1408*1024/sizeof(u64)) >> 3;
1875 int idx_end = (1536*1024/sizeof(u64)) >> 3;
1876 long *p_start = (long *) &(sba_dev->ioc[i].res_map[idx_start]);
1877 long *p_end = (long *) &(sba_dev->ioc[i].res_map[idx_end]);
1879 /* mark that part of the io pdir busy */
1880 while (p_start < p_end)
1885 #ifdef DEBUG_DMB_TRAP
1886 iterate_pages( sba_dev->ioc[i].res_map, res_size,
1887 set_data_memory_break, 0);
1888 iterate_pages( sba_dev->ioc[i].pdir_base, sba_dev->ioc[i].pdir_size,
1889 set_data_memory_break, 0);
1892 DBG_INIT("%s() %d res_map %x %p\n",
1893 __FUNCTION__, i, res_size, sba_dev->ioc[i].res_map);
1896 spin_lock_init(&sba_dev->sba_lock);
1897 ioc_needs_fdc = boot_cpu_data.pdc.capabilities & PDC_MODEL_IOPDIR_FDC;
1899 #ifdef DEBUG_SBA_INIT
1901 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
1902 * (bit #61, big endian), we have to flush and sync every time
1903 * IO-PDIR is changed in Ike/Astro.
1905 if (ioc_needs_fdc) {
1906 printk(KERN_INFO MODULE_NAME " FDC/SYNC required.\n");
1908 printk(KERN_INFO MODULE_NAME " IOC has cache coherent PDIR.\n");
1913 #ifdef CONFIG_PROC_FS
1914 static int sba_proc_info(struct seq_file *m, void *p)
1916 struct sba_device *sba_dev = sba_list;
1917 struct ioc *ioc = &sba_dev->ioc[0]; /* FIXME: Multi-IOC support! */
1918 int total_pages = (int) (ioc->res_size << 3); /* 8 bits per byte */
1919 #ifdef SBA_COLLECT_STATS
1920 unsigned long avg = 0, min, max;
1924 len += seq_printf(m, "%s rev %d.%d\n",
1926 (sba_dev->hw_rev & 0x7) + 1,
1927 (sba_dev->hw_rev & 0x18) >> 3
1929 len += seq_printf(m, "IO PDIR size : %d bytes (%d entries)\n",
1930 (int) ((ioc->res_size << 3) * sizeof(u64)), /* 8 bits/byte */
1933 len += seq_printf(m, "Resource bitmap : %d bytes (%d pages)\n",
1934 ioc->res_size, ioc->res_size << 3); /* 8 bits per byte */
1936 len += seq_printf(m, "LMMIO_BASE/MASK/ROUTE %08x %08x %08x\n",
1937 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_BASE),
1938 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_MASK),
1939 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_ROUTE)
1943 len += seq_printf(m, "DIR%d_BASE/MASK/ROUTE %08x %08x %08x\n", i,
1944 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_BASE + i*0x18),
1945 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_MASK + i*0x18),
1946 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_ROUTE + i*0x18)
1949 #ifdef SBA_COLLECT_STATS
1950 len += seq_printf(m, "IO PDIR entries : %ld free %ld used (%d%%)\n",
1951 total_pages - ioc->used_pages, ioc->used_pages,
1952 (int) (ioc->used_pages * 100 / total_pages));
1954 min = max = ioc->avg_search[0];
1955 for (i = 0; i < SBA_SEARCH_SAMPLE; i++) {
1956 avg += ioc->avg_search[i];
1957 if (ioc->avg_search[i] > max) max = ioc->avg_search[i];
1958 if (ioc->avg_search[i] < min) min = ioc->avg_search[i];
1960 avg /= SBA_SEARCH_SAMPLE;
1961 len += seq_printf(m, " Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n",
1964 len += seq_printf(m, "pci_map_single(): %12ld calls %12ld pages (avg %d/1000)\n",
1965 ioc->msingle_calls, ioc->msingle_pages,
1966 (int) ((ioc->msingle_pages * 1000)/ioc->msingle_calls));
1968 /* KLUGE - unmap_sg calls unmap_single for each mapped page */
1969 min = ioc->usingle_calls;
1970 max = ioc->usingle_pages - ioc->usg_pages;
1971 len += seq_printf(m, "pci_unmap_single: %12ld calls %12ld pages (avg %d/1000)\n",
1972 min, max, (int) ((max * 1000)/min));
1974 len += seq_printf(m, "pci_map_sg() : %12ld calls %12ld pages (avg %d/1000)\n",
1975 ioc->msg_calls, ioc->msg_pages,
1976 (int) ((ioc->msg_pages * 1000)/ioc->msg_calls));
1978 len += seq_printf(m, "pci_unmap_sg() : %12ld calls %12ld pages (avg %d/1000)\n",
1979 ioc->usg_calls, ioc->usg_pages,
1980 (int) ((ioc->usg_pages * 1000)/ioc->usg_calls));
1987 sba_proc_open(struct inode *i, struct file *f)
1989 return single_open(f, &sba_proc_info, NULL);
1992 static struct file_operations sba_proc_fops = {
1993 .owner = THIS_MODULE,
1994 .open = sba_proc_open,
1996 .llseek = seq_lseek,
1997 .release = single_release,
2001 sba_proc_bitmap_info(struct seq_file *m, void *p)
2003 struct sba_device *sba_dev = sba_list;
2004 struct ioc *ioc = &sba_dev->ioc[0]; /* FIXME: Multi-IOC support! */
2005 unsigned int *res_ptr = (unsigned int *)ioc->res_map;
2008 for (i = 0; i < (ioc->res_size/sizeof(unsigned int)); ++i, ++res_ptr) {
2010 len += seq_printf(m, "\n ");
2011 len += seq_printf(m, " %08x", *res_ptr);
2013 len += seq_printf(m, "\n");
2019 sba_proc_bitmap_open(struct inode *i, struct file *f)
2021 return single_open(f, &sba_proc_bitmap_info, NULL);
2024 static struct file_operations sba_proc_bitmap_fops = {
2025 .owner = THIS_MODULE,
2026 .open = sba_proc_bitmap_open,
2028 .llseek = seq_lseek,
2029 .release = single_release,
2031 #endif /* CONFIG_PROC_FS */
2033 static struct parisc_device_id sba_tbl[] = {
2034 { HPHW_IOA, HVERSION_REV_ANY_ID, ASTRO_RUNWAY_PORT, 0xb },
2035 { HPHW_BCPORT, HVERSION_REV_ANY_ID, IKE_MERCED_PORT, 0xc },
2036 { HPHW_BCPORT, HVERSION_REV_ANY_ID, REO_MERCED_PORT, 0xc },
2037 { HPHW_BCPORT, HVERSION_REV_ANY_ID, REOG_MERCED_PORT, 0xc },
2038 { HPHW_IOA, HVERSION_REV_ANY_ID, PLUTO_MCKINLEY_PORT, 0xc },
2042 int sba_driver_callback(struct parisc_device *);
2044 static struct parisc_driver sba_driver = {
2045 .name = MODULE_NAME,
2046 .id_table = sba_tbl,
2047 .probe = sba_driver_callback,
2051 ** Determine if sba should claim this chip (return 0) or not (return 1).
2052 ** If so, initialize the chip and tell other partners in crime they
2056 sba_driver_callback(struct parisc_device *dev)
2058 struct sba_device *sba_dev;
2062 void __iomem *sba_addr = ioremap_nocache(dev->hpa.start, SBA_FUNC_SIZE);
2063 struct proc_dir_entry *info_entry, *bitmap_entry, *root;
2065 sba_dump_ranges(sba_addr);
2067 /* Read HW Rev First */
2068 func_class = READ_REG(sba_addr + SBA_FCLASS);
2070 if (IS_ASTRO(&dev->id)) {
2071 unsigned long fclass;
2072 static char astro_rev[]="Astro ?.?";
2074 /* Astro is broken...Read HW Rev First */
2075 fclass = READ_REG(sba_addr);
2077 astro_rev[6] = '1' + (char) (fclass & 0x7);
2078 astro_rev[8] = '0' + (char) ((fclass & 0x18) >> 3);
2079 version = astro_rev;
2081 } else if (IS_IKE(&dev->id)) {
2082 static char ike_rev[] = "Ike rev ?";
2083 ike_rev[8] = '0' + (char) (func_class & 0xff);
2085 } else if (IS_PLUTO(&dev->id)) {
2086 static char pluto_rev[]="Pluto ?.?";
2087 pluto_rev[6] = '0' + (char) ((func_class & 0xf0) >> 4);
2088 pluto_rev[8] = '0' + (char) (func_class & 0x0f);
2089 version = pluto_rev;
2091 static char reo_rev[] = "REO rev ?";
2092 reo_rev[8] = '0' + (char) (func_class & 0xff);
2096 if (!global_ioc_cnt) {
2097 global_ioc_cnt = count_parisc_driver(&sba_driver);
2099 /* Astro and Pluto have one IOC per SBA */
2100 if ((!IS_ASTRO(&dev->id)) || (!IS_PLUTO(&dev->id)))
2101 global_ioc_cnt *= 2;
2104 printk(KERN_INFO "%s found %s at 0x%lx\n",
2105 MODULE_NAME, version, dev->hpa.start);
2107 sba_dev = kzalloc(sizeof(struct sba_device), GFP_KERNEL);
2109 printk(KERN_ERR MODULE_NAME " - couldn't alloc sba_device\n");
2113 parisc_set_drvdata(dev, sba_dev);
2115 for(i=0; i<MAX_IOC; i++)
2116 spin_lock_init(&(sba_dev->ioc[i].res_lock));
2119 sba_dev->hw_rev = func_class;
2120 sba_dev->iodc = &dev->id;
2121 sba_dev->name = dev->name;
2122 sba_dev->sba_hpa = sba_addr;
2124 sba_get_pat_resources(sba_dev);
2125 sba_hw_init(sba_dev);
2126 sba_common_init(sba_dev);
2128 hppa_dma_ops = &sba_ops;
2130 #ifdef CONFIG_PROC_FS
2131 switch (dev->id.hversion) {
2132 case PLUTO_MCKINLEY_PORT:
2133 root = proc_mckinley_root;
2135 case ASTRO_RUNWAY_PORT:
2136 case IKE_MERCED_PORT:
2138 root = proc_runway_root;
2142 info_entry = create_proc_entry("sba_iommu", 0, root);
2143 bitmap_entry = create_proc_entry("sba_iommu-bitmap", 0, root);
2146 info_entry->proc_fops = &sba_proc_fops;
2149 bitmap_entry->proc_fops = &sba_proc_bitmap_fops;
2152 parisc_vmerge_boundary = IOVP_SIZE;
2153 parisc_vmerge_max_size = IOVP_SIZE * BITS_PER_LONG;
2159 ** One time initialization to let the world know the SBA was found.
2160 ** This is the only routine which is NOT static.
2161 ** Must be called exactly once before pci_init().
2163 void __init sba_init(void)
2165 register_parisc_driver(&sba_driver);
2170 * sba_get_iommu - Assign the iommu pointer for the pci bus controller.
2171 * @dev: The parisc device.
2173 * Returns the appropriate IOMMU data for the given parisc PCI controller.
2174 * This is cached and used later for PCI DMA Mapping.
2176 void * sba_get_iommu(struct parisc_device *pci_hba)
2178 struct parisc_device *sba_dev = parisc_parent(pci_hba);
2179 struct sba_device *sba = sba_dev->dev.driver_data;
2180 char t = sba_dev->id.hw_type;
2181 int iocnum = (pci_hba->hw_path >> 3); /* rope # */
2183 WARN_ON((t != HPHW_IOA) && (t != HPHW_BCPORT));
2185 return &(sba->ioc[iocnum]);
2190 * sba_directed_lmmio - return first directed LMMIO range routed to rope
2191 * @pa_dev: The parisc device.
2192 * @r: resource PCI host controller wants start/end fields assigned.
2194 * For the given parisc PCI controller, determine if any direct ranges
2195 * are routed down the corresponding rope.
2197 void sba_directed_lmmio(struct parisc_device *pci_hba, struct resource *r)
2199 struct parisc_device *sba_dev = parisc_parent(pci_hba);
2200 struct sba_device *sba = sba_dev->dev.driver_data;
2201 char t = sba_dev->id.hw_type;
2203 int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1)); /* rope # */
2205 BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2207 r->start = r->end = 0;
2209 /* Astro has 4 directed ranges. Not sure about Ike/Pluto/et al */
2210 for (i=0; i<4; i++) {
2212 void __iomem *reg = sba->sba_hpa + i*0x18;
2214 base = READ_REG32(reg + LMMIO_DIRECT0_BASE);
2215 if ((base & 1) == 0)
2216 continue; /* not enabled */
2218 size = READ_REG32(reg + LMMIO_DIRECT0_ROUTE);
2220 if ((size & (ROPES_PER_IOC-1)) != rope)
2221 continue; /* directed down different rope */
2223 r->start = (base & ~1UL) | PCI_F_EXTEND;
2224 size = ~ READ_REG32(reg + LMMIO_DIRECT0_MASK);
2225 r->end = r->start + size;
2231 * sba_distributed_lmmio - return portion of distributed LMMIO range
2232 * @pa_dev: The parisc device.
2233 * @r: resource PCI host controller wants start/end fields assigned.
2235 * For the given parisc PCI controller, return portion of distributed LMMIO
2236 * range. The distributed LMMIO is always present and it's just a question
2237 * of the base address and size of the range.
2239 void sba_distributed_lmmio(struct parisc_device *pci_hba, struct resource *r )
2241 struct parisc_device *sba_dev = parisc_parent(pci_hba);
2242 struct sba_device *sba = sba_dev->dev.driver_data;
2243 char t = sba_dev->id.hw_type;
2245 int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1)); /* rope # */
2247 BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2249 r->start = r->end = 0;
2251 base = READ_REG32(sba->sba_hpa + LMMIO_DIST_BASE);
2252 if ((base & 1) == 0) {
2253 BUG(); /* Gah! Distr Range wasn't enabled! */
2257 r->start = (base & ~1UL) | PCI_F_EXTEND;
2259 size = (~READ_REG32(sba->sba_hpa + LMMIO_DIST_MASK)) / ROPES_PER_IOC;
2260 r->start += rope * (size + 1); /* adjust base for this rope */
2261 r->end = r->start + size;