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/config.h>
23 #include <linux/types.h>
24 #include <linux/kernel.h>
25 #include <linux/spinlock.h>
26 #include <linux/slab.h>
27 #include <linux/init.h>
30 #include <linux/string.h>
31 #include <linux/pci.h>
33 #include <asm/byteorder.h>
35 #include <asm/dma.h> /* for DMA_CHUNK_SIZE */
37 #include <asm/hardware.h> /* for register_parisc_driver() stuff */
39 #include <linux/proc_fs.h>
40 #include <linux/seq_file.h>
42 #include <asm/runway.h> /* for proc_runway_root */
43 #include <asm/pdc.h> /* for PDC_MODEL_* */
44 #include <asm/pdcpat.h> /* for is_pdc_pat() */
45 #include <asm/parisc-device.h>
48 /* declared in arch/parisc/kernel/setup.c */
49 extern struct proc_dir_entry * proc_mckinley_root;
51 #define MODULE_NAME "SBA"
54 /* depends on proc fs support. But costs CPU performance */
55 #undef SBA_COLLECT_STATS
59 ** The number of debug flags is a clue - this code is fragile.
60 ** Don't even think about messing with it unless you have
61 ** plenty of 710's to sacrifice to the computer gods. :^)
65 #undef DEBUG_SBA_RUN_SG
66 #undef DEBUG_SBA_RESOURCE
67 #undef ASSERT_PDIR_SANITY
68 #undef DEBUG_LARGE_SG_ENTRIES
72 #define DBG_INIT(x...) printk(x)
74 #define DBG_INIT(x...)
78 #define DBG_RUN(x...) printk(x)
83 #ifdef DEBUG_SBA_RUN_SG
84 #define DBG_RUN_SG(x...) printk(x)
86 #define DBG_RUN_SG(x...)
90 #ifdef DEBUG_SBA_RESOURCE
91 #define DBG_RES(x...) printk(x)
96 #if defined(CONFIG_64BIT)
97 /* "low end" PA8800 machines use ZX1 chipset: PAT PDC and only run 64-bit */
101 #define SBA_INLINE __inline__
105 ** The number of pdir entries to "free" before issueing
106 ** a read to PCOM register to flush out PCOM writes.
107 ** Interacts with allocation granularity (ie 4 or 8 entries
108 ** allocated and free'd/purged at a time might make this
109 ** less interesting).
111 #define DELAYED_RESOURCE_CNT 16
113 #define DEFAULT_DMA_HINT_REG 0
115 #define ASTRO_RUNWAY_PORT 0x582
116 #define IKE_MERCED_PORT 0x803
117 #define REO_MERCED_PORT 0x804
118 #define REOG_MERCED_PORT 0x805
119 #define PLUTO_MCKINLEY_PORT 0x880
121 #define SBA_FUNC_ID 0x0000 /* function id */
122 #define SBA_FCLASS 0x0008 /* function class, bist, header, rev... */
124 #define IS_ASTRO(id) ((id)->hversion == ASTRO_RUNWAY_PORT)
125 #define IS_IKE(id) ((id)->hversion == IKE_MERCED_PORT)
126 #define IS_PLUTO(id) ((id)->hversion == PLUTO_MCKINLEY_PORT)
128 #define SBA_FUNC_SIZE 4096 /* SBA configuration function reg set */
130 #define ASTRO_IOC_OFFSET (32 * SBA_FUNC_SIZE)
131 #define PLUTO_IOC_OFFSET (1 * SBA_FUNC_SIZE)
132 /* Ike's IOC's occupy functions 2 and 3 */
133 #define IKE_IOC_OFFSET(p) ((p+2) * SBA_FUNC_SIZE)
135 #define IOC_CTRL 0x8 /* IOC_CTRL offset */
136 #define IOC_CTRL_TC (1 << 0) /* TOC Enable */
137 #define IOC_CTRL_CE (1 << 1) /* Coalesce Enable */
138 #define IOC_CTRL_DE (1 << 2) /* Dillon Enable */
139 #define IOC_CTRL_RM (1 << 8) /* Real Mode */
140 #define IOC_CTRL_NC (1 << 9) /* Non Coherent Mode */
141 #define IOC_CTRL_D4 (1 << 11) /* Disable 4-byte coalescing */
142 #define IOC_CTRL_DD (1 << 13) /* Disable distr. LMMIO range coalescing */
144 #define MAX_IOC 2 /* per Ike. Pluto/Astro only have 1. */
146 #define ROPES_PER_IOC 8 /* per Ike half or Pluto/Astro */
150 ** Offsets into MBIB (Function 0 on Ike and hopefully Astro)
151 ** Firmware programs this stuff. Don't touch it.
153 #define LMMIO_DIRECT0_BASE 0x300
154 #define LMMIO_DIRECT0_MASK 0x308
155 #define LMMIO_DIRECT0_ROUTE 0x310
157 #define LMMIO_DIST_BASE 0x360
158 #define LMMIO_DIST_MASK 0x368
159 #define LMMIO_DIST_ROUTE 0x370
161 #define IOS_DIST_BASE 0x390
162 #define IOS_DIST_MASK 0x398
163 #define IOS_DIST_ROUTE 0x3A0
165 #define IOS_DIRECT_BASE 0x3C0
166 #define IOS_DIRECT_MASK 0x3C8
167 #define IOS_DIRECT_ROUTE 0x3D0
170 ** Offsets into I/O TLB (Function 2 and 3 on Ike)
172 #define ROPE0_CTL 0x200 /* "regbus pci0" */
173 #define ROPE1_CTL 0x208
174 #define ROPE2_CTL 0x210
175 #define ROPE3_CTL 0x218
176 #define ROPE4_CTL 0x220
177 #define ROPE5_CTL 0x228
178 #define ROPE6_CTL 0x230
179 #define ROPE7_CTL 0x238
181 #define IOC_ROPE0_CFG 0x500 /* pluto only */
182 #define IOC_ROPE_AO 0x10 /* Allow "Relaxed Ordering" */
186 #define HF_ENABLE 0x40
189 #define IOC_IBASE 0x300 /* IO TLB */
190 #define IOC_IMASK 0x308
191 #define IOC_PCOM 0x310
192 #define IOC_TCNFG 0x318
193 #define IOC_PDIR_BASE 0x320
195 /* AGP GART driver looks for this */
196 #define SBA_IOMMU_COOKIE 0x0000badbadc0ffeeUL
200 ** IOC supports 4/8/16/64KB page sizes (see TCNFG register)
201 ** It's safer (avoid memory corruption) to keep DMA page mappings
202 ** equivalently sized to VM PAGE_SIZE.
204 ** We really can't avoid generating a new mapping for each
205 ** page since the Virtual Coherence Index has to be generated
206 ** and updated for each page.
208 ** PAGE_SIZE could be greater than IOVP_SIZE. But not the inverse.
210 #define IOVP_SIZE PAGE_SIZE
211 #define IOVP_SHIFT PAGE_SHIFT
212 #define IOVP_MASK PAGE_MASK
214 #define SBA_PERF_CFG 0x708 /* Performance Counter stuff */
215 #define SBA_PERF_MASK1 0x718
216 #define SBA_PERF_MASK2 0x730
220 ** Offsets into PCI Performance Counters (functions 12 and 13)
221 ** Controlled by PERF registers in function 2 & 3 respectively.
223 #define SBA_PERF_CNT1 0x200
224 #define SBA_PERF_CNT2 0x208
225 #define SBA_PERF_CNT3 0x210
229 void __iomem *ioc_hpa; /* I/O MMU base address */
230 char *res_map; /* resource map, bit == pdir entry */
231 u64 *pdir_base; /* physical base address */
232 unsigned long ibase; /* pdir IOV Space base - shared w/lba_pci */
233 unsigned long imask; /* pdir IOV Space mask - shared w/lba_pci */
235 unsigned long iovp_mask; /* help convert IOVA to IOVP */
237 unsigned long *res_hint; /* next avail IOVP - circular search */
239 unsigned int res_bitshift; /* from the LEFT! */
240 unsigned int res_size; /* size of resource map in bytes */
241 #ifdef SBA_HINT_SUPPORT
242 /* FIXME : DMA HINTs not used */
243 unsigned long hint_mask_pdir; /* bits used for DMA hints */
244 unsigned int hint_shift_pdir;
246 #if DELAYED_RESOURCE_CNT > 0
248 struct sba_dma_pair {
251 } saved[DELAYED_RESOURCE_CNT];
254 #ifdef SBA_COLLECT_STATS
255 #define SBA_SEARCH_SAMPLE 0x100
256 unsigned long avg_search[SBA_SEARCH_SAMPLE];
257 unsigned long avg_idx; /* current index into avg_search */
258 unsigned long used_pages;
259 unsigned long msingle_calls;
260 unsigned long msingle_pages;
261 unsigned long msg_calls;
262 unsigned long msg_pages;
263 unsigned long usingle_calls;
264 unsigned long usingle_pages;
265 unsigned long usg_calls;
266 unsigned long usg_pages;
269 /* STUFF We don't need in performance path */
270 unsigned int pdir_size; /* in bytes, determined by IOV Space size */
274 struct sba_device *next; /* list of SBA's in system */
275 struct parisc_device *dev; /* dev found in bus walk */
276 struct parisc_device_id *iodc; /* data about dev from firmware */
278 void __iomem *sba_hpa; /* base address */
280 unsigned int flags; /* state/functionality enabled */
281 unsigned int hw_rev; /* HW revision of chip */
283 struct resource chip_resv; /* MMIO reserved for chip */
284 struct resource iommu_resv; /* MMIO reserved for iommu */
286 unsigned int num_ioc; /* number of on-board IOC's */
287 struct ioc ioc[MAX_IOC];
291 static struct sba_device *sba_list;
293 static unsigned long ioc_needs_fdc = 0;
295 /* global count of IOMMUs in the system */
296 static unsigned int global_ioc_cnt = 0;
298 /* PA8700 (Piranha 2.2) bug workaround */
299 static unsigned long piranha_bad_128k = 0;
301 /* Looks nice and keeps the compiler happy */
302 #define SBA_DEV(d) ((struct sba_device *) (d))
304 #ifdef SBA_AGP_SUPPORT
305 static int reserve_sba_gart = 1;
308 #define ROUNDUP(x,y) ((x + ((y)-1)) & ~((y)-1))
311 /************************************
312 ** SBA register read and write support
314 ** BE WARNED: register writes are posted.
315 ** (ie follow writes which must reach HW with a read)
317 ** Superdome (in particular, REO) allows only 64-bit CSR accesses.
319 #define READ_REG32(addr) readl(addr)
320 #define READ_REG64(addr) readq(addr)
321 #define WRITE_REG32(val, addr) writel((val), (addr))
322 #define WRITE_REG64(val, addr) writeq((val), (addr))
325 #define READ_REG(addr) READ_REG64(addr)
326 #define WRITE_REG(value, addr) WRITE_REG64(value, addr)
328 #define READ_REG(addr) READ_REG32(addr)
329 #define WRITE_REG(value, addr) WRITE_REG32(value, addr)
332 #ifdef DEBUG_SBA_INIT
334 /* NOTE: When CONFIG_64BIT isn't defined, READ_REG64() is two 32-bit reads */
337 * sba_dump_ranges - debugging only - print ranges assigned to this IOA
338 * @hpa: base address of the sba
340 * Print the MMIO and IO Port address ranges forwarded by an Astro/Ike/RIO
341 * IO Adapter (aka Bus Converter).
344 sba_dump_ranges(void __iomem *hpa)
346 DBG_INIT("SBA at 0x%p\n", hpa);
347 DBG_INIT("IOS_DIST_BASE : %Lx\n", READ_REG64(hpa+IOS_DIST_BASE));
348 DBG_INIT("IOS_DIST_MASK : %Lx\n", READ_REG64(hpa+IOS_DIST_MASK));
349 DBG_INIT("IOS_DIST_ROUTE : %Lx\n", READ_REG64(hpa+IOS_DIST_ROUTE));
351 DBG_INIT("IOS_DIRECT_BASE : %Lx\n", READ_REG64(hpa+IOS_DIRECT_BASE));
352 DBG_INIT("IOS_DIRECT_MASK : %Lx\n", READ_REG64(hpa+IOS_DIRECT_MASK));
353 DBG_INIT("IOS_DIRECT_ROUTE: %Lx\n", READ_REG64(hpa+IOS_DIRECT_ROUTE));
357 * sba_dump_tlb - debugging only - print IOMMU operating parameters
358 * @hpa: base address of the IOMMU
360 * Print the size/location of the IO MMU PDIR.
362 static void sba_dump_tlb(void __iomem *hpa)
364 DBG_INIT("IO TLB at 0x%p\n", hpa);
365 DBG_INIT("IOC_IBASE : 0x%Lx\n", READ_REG64(hpa+IOC_IBASE));
366 DBG_INIT("IOC_IMASK : 0x%Lx\n", READ_REG64(hpa+IOC_IMASK));
367 DBG_INIT("IOC_TCNFG : 0x%Lx\n", READ_REG64(hpa+IOC_TCNFG));
368 DBG_INIT("IOC_PDIR_BASE: 0x%Lx\n", READ_REG64(hpa+IOC_PDIR_BASE));
372 #define sba_dump_ranges(x)
373 #define sba_dump_tlb(x)
374 #endif /* DEBUG_SBA_INIT */
377 #ifdef ASSERT_PDIR_SANITY
380 * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry
381 * @ioc: IO MMU structure which owns the pdir we are interested in.
382 * @msg: text to print ont the output line.
385 * Print one entry of the IO MMU PDIR in human readable form.
388 sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide)
390 /* start printing from lowest pde in rval */
391 u64 *ptr = &(ioc->pdir_base[pide & (~0U * BITS_PER_LONG)]);
392 unsigned long *rptr = (unsigned long *) &(ioc->res_map[(pide >>3) & ~(sizeof(unsigned long) - 1)]);
395 printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n",
397 rptr, pide & (BITS_PER_LONG - 1), *rptr);
400 while (rcnt < BITS_PER_LONG) {
401 printk(KERN_DEBUG "%s %2d %p %016Lx\n",
402 (rcnt == (pide & (BITS_PER_LONG - 1)))
408 printk(KERN_DEBUG "%s", msg);
413 * sba_check_pdir - debugging only - consistency checker
414 * @ioc: IO MMU structure which owns the pdir we are interested in.
415 * @msg: text to print ont the output line.
417 * Verify the resource map and pdir state is consistent
420 sba_check_pdir(struct ioc *ioc, char *msg)
422 u32 *rptr_end = (u32 *) &(ioc->res_map[ioc->res_size]);
423 u32 *rptr = (u32 *) ioc->res_map; /* resource map ptr */
424 u64 *pptr = ioc->pdir_base; /* pdir ptr */
427 while (rptr < rptr_end) {
429 int rcnt = 32; /* number of bits we might check */
432 /* Get last byte and highest bit from that */
433 u32 pde = ((u32) (((char *)pptr)[7])) << 24;
434 if ((rval ^ pde) & 0x80000000)
437 ** BUMMER! -- res_map != pdir --
438 ** Dump rval and matching pdir entries
440 sba_dump_pdir_entry(ioc, msg, pide);
444 rval <<= 1; /* try the next bit */
448 rptr++; /* look at next word of res_map */
450 /* It'd be nice if we always got here :^) */
456 * sba_dump_sg - debugging only - print Scatter-Gather list
457 * @ioc: IO MMU structure which owns the pdir we are interested in.
458 * @startsg: head of the SG list
459 * @nents: number of entries in SG list
461 * print the SG list so we can verify it's correct by hand.
464 sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
466 while (nents-- > 0) {
467 printk(KERN_DEBUG " %d : %08lx/%05x %p/%05x\n",
469 (unsigned long) sg_dma_address(startsg),
471 sg_virt_addr(startsg), startsg->length);
476 #endif /* ASSERT_PDIR_SANITY */
481 /**************************************************************
483 * I/O Pdir Resource Management
485 * Bits set in the resource map are in use.
486 * Each bit can represent a number of pages.
487 * LSbs represent lower addresses (IOVA's).
489 ***************************************************************/
490 #define PAGES_PER_RANGE 1 /* could increase this to 4 or 8 if needed */
492 /* Convert from IOVP to IOVA and vice versa. */
495 /* Pluto (aka ZX1) boxes need to set or clear the ibase bits appropriately */
496 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((ioc->ibase) | (iovp) | (offset))
497 #define SBA_IOVP(ioc,iova) ((iova) & (ioc)->iovp_mask)
499 /* only support Astro and ancestors. Saves a few cycles in key places */
500 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((iovp) | (offset))
501 #define SBA_IOVP(ioc,iova) (iova)
504 #define PDIR_INDEX(iovp) ((iovp)>>IOVP_SHIFT)
506 #define RESMAP_MASK(n) (~0UL << (BITS_PER_LONG - (n)))
507 #define RESMAP_IDX_MASK (sizeof(unsigned long) - 1)
511 * sba_search_bitmap - find free space in IO PDIR resource bitmap
512 * @ioc: IO MMU structure which owns the pdir we are interested in.
513 * @bits_wanted: number of entries we need.
515 * Find consecutive free bits in resource bitmap.
516 * Each bit represents one entry in the IO Pdir.
517 * Cool perf optimization: search for log2(size) bits at a time.
519 static SBA_INLINE unsigned long
520 sba_search_bitmap(struct ioc *ioc, unsigned long bits_wanted)
522 unsigned long *res_ptr = ioc->res_hint;
523 unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]);
524 unsigned long pide = ~0UL;
526 if (bits_wanted > (BITS_PER_LONG/2)) {
527 /* Search word at a time - no mask needed */
528 for(; res_ptr < res_end; ++res_ptr) {
530 *res_ptr = RESMAP_MASK(bits_wanted);
531 pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
532 pide <<= 3; /* convert to bit address */
536 /* point to the next word on next pass */
538 ioc->res_bitshift = 0;
541 ** Search the resource bit map on well-aligned values.
542 ** "o" is the alignment.
543 ** We need the alignment to invalidate I/O TLB using
544 ** SBA HW features in the unmap path.
546 unsigned long o = 1 << get_order(bits_wanted << PAGE_SHIFT);
547 uint bitshiftcnt = ROUNDUP(ioc->res_bitshift, o);
550 if (bitshiftcnt >= BITS_PER_LONG) {
554 mask = RESMAP_MASK(bits_wanted) >> bitshiftcnt;
556 DBG_RES("%s() o %ld %p", __FUNCTION__, o, res_ptr);
557 while(res_ptr < res_end)
559 DBG_RES(" %p %lx %lx\n", res_ptr, mask, *res_ptr);
561 if(((*res_ptr) & mask) == 0) {
562 *res_ptr |= mask; /* mark resources busy! */
563 pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
564 pide <<= 3; /* convert to bit address */
571 mask = RESMAP_MASK(bits_wanted);
576 /* look in the same word on the next pass */
577 ioc->res_bitshift = bitshiftcnt + bits_wanted;
581 if (res_end <= res_ptr) {
582 ioc->res_hint = (unsigned long *) ioc->res_map;
583 ioc->res_bitshift = 0;
585 ioc->res_hint = res_ptr;
592 * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
593 * @ioc: IO MMU structure which owns the pdir we are interested in.
594 * @size: number of bytes to create a mapping for
596 * Given a size, find consecutive unmarked and then mark those bits in the
600 sba_alloc_range(struct ioc *ioc, size_t size)
602 unsigned int pages_needed = size >> IOVP_SHIFT;
603 #ifdef SBA_COLLECT_STATS
604 unsigned long cr_start = mfctl(16);
608 pide = sba_search_bitmap(ioc, pages_needed);
609 if (pide >= (ioc->res_size << 3)) {
610 pide = sba_search_bitmap(ioc, pages_needed);
611 if (pide >= (ioc->res_size << 3))
612 panic("%s: I/O MMU @ %p is out of mapping resources\n",
613 __FILE__, ioc->ioc_hpa);
616 #ifdef ASSERT_PDIR_SANITY
617 /* verify the first enable bit is clear */
618 if(0x00 != ((u8 *) ioc->pdir_base)[pide*sizeof(u64) + 7]) {
619 sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide);
623 DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
624 __FUNCTION__, size, pages_needed, pide,
625 (uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map),
628 #ifdef SBA_COLLECT_STATS
630 unsigned long cr_end = mfctl(16);
631 unsigned long tmp = cr_end - cr_start;
632 /* check for roll over */
633 cr_start = (cr_end < cr_start) ? -(tmp) : (tmp);
635 ioc->avg_search[ioc->avg_idx++] = cr_start;
636 ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;
638 ioc->used_pages += pages_needed;
646 * sba_free_range - unmark bits in IO PDIR resource bitmap
647 * @ioc: IO MMU structure which owns the pdir we are interested in.
648 * @iova: IO virtual address which was previously allocated.
649 * @size: number of bytes to create a mapping for
651 * clear bits in the ioc's resource map
653 static SBA_INLINE void
654 sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size)
656 unsigned long iovp = SBA_IOVP(ioc, iova);
657 unsigned int pide = PDIR_INDEX(iovp);
658 unsigned int ridx = pide >> 3; /* convert bit to byte address */
659 unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]);
661 int bits_not_wanted = size >> IOVP_SHIFT;
663 /* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
664 unsigned long m = RESMAP_MASK(bits_not_wanted) >> (pide & (BITS_PER_LONG - 1));
666 DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n",
667 __FUNCTION__, (uint) iova, size,
668 bits_not_wanted, m, pide, res_ptr, *res_ptr);
670 #ifdef SBA_COLLECT_STATS
671 ioc->used_pages -= bits_not_wanted;
678 /**************************************************************
680 * "Dynamic DMA Mapping" support (aka "Coherent I/O")
682 ***************************************************************/
684 #ifdef SBA_HINT_SUPPORT
685 #define SBA_DMA_HINT(ioc, val) ((val) << (ioc)->hint_shift_pdir)
688 typedef unsigned long space_t;
689 #define KERNEL_SPACE 0
692 * sba_io_pdir_entry - fill in one IO PDIR entry
693 * @pdir_ptr: pointer to IO PDIR entry
694 * @sid: process Space ID - currently only support KERNEL_SPACE
695 * @vba: Virtual CPU address of buffer to map
696 * @hint: DMA hint set to use for this mapping
698 * SBA Mapping Routine
700 * Given a virtual address (vba, arg2) and space id, (sid, arg1)
701 * sba_io_pdir_entry() loads the I/O PDIR entry pointed to by
703 * Using the bass-ackwards HP bit numbering, Each IO Pdir entry
704 * for Astro/Ike looks like:
708 * +-+---------------------+----------------------------------+----+--------+
709 * |V| U | PPN[43:12] | U | VI |
710 * +-+---------------------+----------------------------------+----+--------+
712 * Pluto is basically identical, supports fewer physical address bits:
715 * +-+------------------------+-------------------------------+----+--------+
716 * |V| U | PPN[39:12] | U | VI |
717 * +-+------------------------+-------------------------------+----+--------+
719 * V == Valid Bit (Most Significant Bit is bit 0)
721 * PPN == Physical Page Number
722 * VI == Virtual Index (aka Coherent Index)
724 * LPA instruction output is put into PPN field.
725 * LCI (Load Coherence Index) instruction provides the "VI" bits.
727 * We pre-swap the bytes since PCX-W is Big Endian and the
728 * IOMMU uses little endian for the pdir.
732 sba_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba,
735 u64 pa; /* physical address */
736 register unsigned ci; /* coherent index */
738 pa = virt_to_phys(vba);
742 asm("lci 0(%%sr1, %1), %0" : "=r" (ci) : "r" (vba));
743 pa |= (ci >> 12) & 0xff; /* move CI (8 bits) into lowest byte */
745 pa |= 0x8000000000000000ULL; /* set "valid" bit */
746 *pdir_ptr = cpu_to_le64(pa); /* swap and store into I/O Pdir */
749 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
750 * (bit #61, big endian), we have to flush and sync every time
751 * IO-PDIR is changed in Ike/Astro.
754 asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
759 * sba_mark_invalid - invalidate one or more IO PDIR entries
760 * @ioc: IO MMU structure which owns the pdir we are interested in.
761 * @iova: IO Virtual Address mapped earlier
762 * @byte_cnt: number of bytes this mapping covers.
764 * Marking the IO PDIR entry(ies) as Invalid and invalidate
765 * corresponding IO TLB entry. The Ike PCOM (Purge Command Register)
766 * is to purge stale entries in the IO TLB when unmapping entries.
768 * The PCOM register supports purging of multiple pages, with a minium
769 * of 1 page and a maximum of 2GB. Hardware requires the address be
770 * aligned to the size of the range being purged. The size of the range
771 * must be a power of 2. The "Cool perf optimization" in the
772 * allocation routine helps keep that true.
774 static SBA_INLINE void
775 sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
777 u32 iovp = (u32) SBA_IOVP(ioc,iova);
778 u64 *pdir_ptr = &ioc->pdir_base[PDIR_INDEX(iovp)];
780 #ifdef ASSERT_PDIR_SANITY
781 /* Assert first pdir entry is set.
783 ** Even though this is a big-endian machine, the entries
784 ** in the iopdir are little endian. That's why we look at
785 ** the byte at +7 instead of at +0.
787 if (0x80 != (((u8 *) pdir_ptr)[7])) {
788 sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp));
792 if (byte_cnt > IOVP_SIZE)
795 unsigned long entries_per_cacheline = ioc_needs_fdc ?
796 L1_CACHE_ALIGN(((unsigned long) pdir_ptr))
797 - (unsigned long) pdir_ptr;
801 /* set "size" field for PCOM */
802 iovp |= get_order(byte_cnt) + PAGE_SHIFT;
805 /* clear I/O Pdir entry "valid" bit first */
806 ((u8 *) pdir_ptr)[7] = 0;
808 asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
810 entries_per_cacheline = L1_CACHE_SHIFT - 3;
814 byte_cnt -= IOVP_SIZE;
815 } while (byte_cnt > IOVP_SIZE);
817 iovp |= IOVP_SHIFT; /* set "size" field for PCOM */
820 ** clear I/O PDIR entry "valid" bit.
821 ** We have to R/M/W the cacheline regardless how much of the
822 ** pdir entry that we clobber.
823 ** The rest of the entry would be useful for debugging if we
824 ** could dump core on HPMC.
826 ((u8 *) pdir_ptr)[7] = 0;
828 asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
830 WRITE_REG( SBA_IOVA(ioc, iovp, 0, 0), ioc->ioc_hpa+IOC_PCOM);
834 * sba_dma_supported - PCI driver can query DMA support
835 * @dev: instance of PCI owned by the driver that's asking
836 * @mask: number of address bits this PCI device can handle
838 * See Documentation/DMA-mapping.txt
840 static int sba_dma_supported( struct device *dev, u64 mask)
845 printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n");
850 /* Documentation/DMA-mapping.txt tells drivers to try 64-bit first,
851 * then fall back to 32-bit if that fails.
852 * We are just "encouraging" 32-bit DMA masks here since we can
853 * never allow IOMMU bypass unless we add special support for ZX1.
861 * check if mask is >= than the current max IO Virt Address
862 * The max IO Virt address will *always* < 30 bits.
864 return((int)(mask >= (ioc->ibase - 1 +
865 (ioc->pdir_size / sizeof(u64) * IOVP_SIZE) )));
870 * sba_map_single - map one buffer and return IOVA for DMA
871 * @dev: instance of PCI owned by the driver that's asking.
872 * @addr: driver buffer to map.
873 * @size: number of bytes to map in driver buffer.
874 * @direction: R/W or both.
876 * See Documentation/DMA-mapping.txt
879 sba_map_single(struct device *dev, void *addr, size_t size,
880 enum dma_data_direction direction)
891 /* save offset bits */
892 offset = ((dma_addr_t) (long) addr) & ~IOVP_MASK;
894 /* round up to nearest IOVP_SIZE */
895 size = (size + offset + ~IOVP_MASK) & IOVP_MASK;
897 spin_lock_irqsave(&ioc->res_lock, flags);
898 #ifdef ASSERT_PDIR_SANITY
899 sba_check_pdir(ioc,"Check before sba_map_single()");
902 #ifdef SBA_COLLECT_STATS
903 ioc->msingle_calls++;
904 ioc->msingle_pages += size >> IOVP_SHIFT;
906 pide = sba_alloc_range(ioc, size);
907 iovp = (dma_addr_t) pide << IOVP_SHIFT;
909 DBG_RUN("%s() 0x%p -> 0x%lx\n",
910 __FUNCTION__, addr, (long) iovp | offset);
912 pdir_start = &(ioc->pdir_base[pide]);
915 sba_io_pdir_entry(pdir_start, KERNEL_SPACE, (unsigned long) addr, 0);
917 DBG_RUN(" pdir 0x%p %02x%02x%02x%02x%02x%02x%02x%02x\n",
919 (u8) (((u8 *) pdir_start)[7]),
920 (u8) (((u8 *) pdir_start)[6]),
921 (u8) (((u8 *) pdir_start)[5]),
922 (u8) (((u8 *) pdir_start)[4]),
923 (u8) (((u8 *) pdir_start)[3]),
924 (u8) (((u8 *) pdir_start)[2]),
925 (u8) (((u8 *) pdir_start)[1]),
926 (u8) (((u8 *) pdir_start)[0])
934 /* force FDC ops in io_pdir_entry() to be visible to IOMMU */
936 asm volatile("sync" : : );
938 #ifdef ASSERT_PDIR_SANITY
939 sba_check_pdir(ioc,"Check after sba_map_single()");
941 spin_unlock_irqrestore(&ioc->res_lock, flags);
943 /* form complete address */
944 return SBA_IOVA(ioc, iovp, offset, DEFAULT_DMA_HINT_REG);
949 * sba_unmap_single - unmap one IOVA and free resources
950 * @dev: instance of PCI owned by the driver that's asking.
951 * @iova: IOVA of driver buffer previously mapped.
952 * @size: number of bytes mapped in driver buffer.
953 * @direction: R/W or both.
955 * See Documentation/DMA-mapping.txt
958 sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size,
959 enum dma_data_direction direction)
962 #if DELAYED_RESOURCE_CNT > 0
963 struct sba_dma_pair *d;
968 DBG_RUN("%s() iovp 0x%lx/%x\n", __FUNCTION__, (long) iova, size);
971 offset = iova & ~IOVP_MASK;
972 iova ^= offset; /* clear offset bits */
974 size = ROUNDUP(size, IOVP_SIZE);
976 spin_lock_irqsave(&ioc->res_lock, flags);
978 #ifdef SBA_COLLECT_STATS
979 ioc->usingle_calls++;
980 ioc->usingle_pages += size >> IOVP_SHIFT;
983 sba_mark_invalid(ioc, iova, size);
985 #if DELAYED_RESOURCE_CNT > 0
986 /* Delaying when we re-use a IO Pdir entry reduces the number
987 * of MMIO reads needed to flush writes to the PCOM register.
989 d = &(ioc->saved[ioc->saved_cnt]);
992 if (++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT) {
993 int cnt = ioc->saved_cnt;
995 sba_free_range(ioc, d->iova, d->size);
1000 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */
1002 #else /* DELAYED_RESOURCE_CNT == 0 */
1003 sba_free_range(ioc, iova, size);
1005 /* If fdc's were issued, force fdc's to be visible now */
1007 asm volatile("sync" : : );
1009 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */
1010 #endif /* DELAYED_RESOURCE_CNT == 0 */
1012 spin_unlock_irqrestore(&ioc->res_lock, flags);
1014 /* XXX REVISIT for 2.5 Linux - need syncdma for zero-copy support.
1015 ** For Astro based systems this isn't a big deal WRT performance.
1016 ** As long as 2.4 kernels copyin/copyout data from/to userspace,
1017 ** we don't need the syncdma. The issue here is I/O MMU cachelines
1018 ** are *not* coherent in all cases. May be hwrev dependent.
1019 ** Need to investigate more.
1020 asm volatile("syncdma");
1026 * sba_alloc_consistent - allocate/map shared mem for DMA
1027 * @hwdev: instance of PCI owned by the driver that's asking.
1028 * @size: number of bytes mapped in driver buffer.
1029 * @dma_handle: IOVA of new buffer.
1031 * See Documentation/DMA-mapping.txt
1033 static void *sba_alloc_consistent(struct device *hwdev, size_t size,
1034 dma_addr_t *dma_handle, gfp_t gfp)
1039 /* only support PCI */
1044 ret = (void *) __get_free_pages(gfp, get_order(size));
1047 memset(ret, 0, size);
1048 *dma_handle = sba_map_single(hwdev, ret, size, 0);
1056 * sba_free_consistent - free/unmap shared mem for DMA
1057 * @hwdev: instance of PCI owned by the driver that's asking.
1058 * @size: number of bytes mapped in driver buffer.
1059 * @vaddr: virtual address IOVA of "consistent" buffer.
1060 * @dma_handler: IO virtual address of "consistent" buffer.
1062 * See Documentation/DMA-mapping.txt
1065 sba_free_consistent(struct device *hwdev, size_t size, void *vaddr,
1066 dma_addr_t dma_handle)
1068 sba_unmap_single(hwdev, dma_handle, size, 0);
1069 free_pages((unsigned long) vaddr, get_order(size));
1074 ** Since 0 is a valid pdir_base index value, can't use that
1075 ** to determine if a value is valid or not. Use a flag to indicate
1076 ** the SG list entry contains a valid pdir index.
1078 #define PIDE_FLAG 0x80000000UL
1080 #ifdef SBA_COLLECT_STATS
1081 #define IOMMU_MAP_STATS
1083 #include "iommu-helpers.h"
1085 #ifdef DEBUG_LARGE_SG_ENTRIES
1086 int dump_run_sg = 0;
1091 * sba_map_sg - map Scatter/Gather list
1092 * @dev: instance of PCI owned by the driver that's asking.
1093 * @sglist: array of buffer/length pairs
1094 * @nents: number of entries in list
1095 * @direction: R/W or both.
1097 * See Documentation/DMA-mapping.txt
1100 sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
1101 enum dma_data_direction direction)
1104 int coalesced, filled = 0;
1105 unsigned long flags;
1107 DBG_RUN_SG("%s() START %d entries\n", __FUNCTION__, nents);
1111 /* Fast path single entry scatterlists. */
1113 sg_dma_address(sglist) = sba_map_single(dev,
1114 (void *)sg_virt_addr(sglist),
1115 sglist->length, direction);
1116 sg_dma_len(sglist) = sglist->length;
1120 spin_lock_irqsave(&ioc->res_lock, flags);
1122 #ifdef ASSERT_PDIR_SANITY
1123 if (sba_check_pdir(ioc,"Check before sba_map_sg()"))
1125 sba_dump_sg(ioc, sglist, nents);
1126 panic("Check before sba_map_sg()");
1130 #ifdef SBA_COLLECT_STATS
1135 ** First coalesce the chunks and allocate I/O pdir space
1137 ** If this is one DMA stream, we can properly map using the
1138 ** correct virtual address associated with each DMA page.
1139 ** w/o this association, we wouldn't have coherent DMA!
1140 ** Access to the virtual address is what forces a two pass algorithm.
1142 coalesced = iommu_coalesce_chunks(ioc, sglist, nents, sba_alloc_range);
1145 ** Program the I/O Pdir
1147 ** map the virtual addresses to the I/O Pdir
1148 ** o dma_address will contain the pdir index
1149 ** o dma_len will contain the number of bytes to map
1150 ** o address contains the virtual address.
1152 filled = iommu_fill_pdir(ioc, sglist, nents, 0, sba_io_pdir_entry);
1154 /* force FDC ops in io_pdir_entry() to be visible to IOMMU */
1156 asm volatile("sync" : : );
1158 #ifdef ASSERT_PDIR_SANITY
1159 if (sba_check_pdir(ioc,"Check after sba_map_sg()"))
1161 sba_dump_sg(ioc, sglist, nents);
1162 panic("Check after sba_map_sg()\n");
1166 spin_unlock_irqrestore(&ioc->res_lock, flags);
1168 DBG_RUN_SG("%s() DONE %d mappings\n", __FUNCTION__, filled);
1175 * sba_unmap_sg - unmap Scatter/Gather list
1176 * @dev: instance of PCI owned by the driver that's asking.
1177 * @sglist: array of buffer/length pairs
1178 * @nents: number of entries in list
1179 * @direction: R/W or both.
1181 * See Documentation/DMA-mapping.txt
1184 sba_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents,
1185 enum dma_data_direction direction)
1188 #ifdef ASSERT_PDIR_SANITY
1189 unsigned long flags;
1192 DBG_RUN_SG("%s() START %d entries, %p,%x\n",
1193 __FUNCTION__, nents, sg_virt_addr(sglist), sglist->length);
1197 #ifdef SBA_COLLECT_STATS
1201 #ifdef ASSERT_PDIR_SANITY
1202 spin_lock_irqsave(&ioc->res_lock, flags);
1203 sba_check_pdir(ioc,"Check before sba_unmap_sg()");
1204 spin_unlock_irqrestore(&ioc->res_lock, flags);
1207 while (sg_dma_len(sglist) && nents--) {
1209 sba_unmap_single(dev, sg_dma_address(sglist), sg_dma_len(sglist), direction);
1210 #ifdef SBA_COLLECT_STATS
1211 ioc->usg_pages += ((sg_dma_address(sglist) & ~IOVP_MASK) + sg_dma_len(sglist) + IOVP_SIZE - 1) >> PAGE_SHIFT;
1212 ioc->usingle_calls--; /* kluge since call is unmap_sg() */
1217 DBG_RUN_SG("%s() DONE (nents %d)\n", __FUNCTION__, nents);
1219 #ifdef ASSERT_PDIR_SANITY
1220 spin_lock_irqsave(&ioc->res_lock, flags);
1221 sba_check_pdir(ioc,"Check after sba_unmap_sg()");
1222 spin_unlock_irqrestore(&ioc->res_lock, flags);
1227 static struct hppa_dma_ops sba_ops = {
1228 .dma_supported = sba_dma_supported,
1229 .alloc_consistent = sba_alloc_consistent,
1230 .alloc_noncoherent = sba_alloc_consistent,
1231 .free_consistent = sba_free_consistent,
1232 .map_single = sba_map_single,
1233 .unmap_single = sba_unmap_single,
1234 .map_sg = sba_map_sg,
1235 .unmap_sg = sba_unmap_sg,
1236 .dma_sync_single_for_cpu = NULL,
1237 .dma_sync_single_for_device = NULL,
1238 .dma_sync_sg_for_cpu = NULL,
1239 .dma_sync_sg_for_device = NULL,
1243 /**************************************************************************
1245 ** SBA PAT PDC support
1247 ** o call pdc_pat_cell_module()
1248 ** o store ranges in PCI "resource" structures
1250 **************************************************************************/
1253 sba_get_pat_resources(struct sba_device *sba_dev)
1257 ** TODO/REVISIT/FIXME: support for directed ranges requires calls to
1258 ** PAT PDC to program the SBA/LBA directed range registers...this
1259 ** burden may fall on the LBA code since it directly supports the
1260 ** PCI subsystem. It's not clear yet. - ggg
1262 PAT_MOD(mod)->mod_info.mod_pages = PAT_GET_MOD_PAGES(temp);
1264 PAT_MOD(mod)->mod_info.dvi = PAT_GET_DVI(temp);
1265 Tells where the dvi bits are located in the address.
1266 PAT_MOD(mod)->mod_info.ioc = PAT_GET_IOC(temp);
1272 /**************************************************************
1274 * Initialization and claim
1276 ***************************************************************/
1277 #define PIRANHA_ADDR_MASK 0x00160000UL /* bit 17,18,20 */
1278 #define PIRANHA_ADDR_VAL 0x00060000UL /* bit 17,18 on */
1280 sba_alloc_pdir(unsigned int pdir_size)
1282 unsigned long pdir_base;
1283 unsigned long pdir_order = get_order(pdir_size);
1285 pdir_base = __get_free_pages(GFP_KERNEL, pdir_order);
1286 if (NULL == (void *) pdir_base) {
1287 panic("%s() could not allocate I/O Page Table\n",
1291 /* If this is not PA8700 (PCX-W2)
1292 ** OR newer than ver 2.2
1293 ** OR in a system that doesn't need VINDEX bits from SBA,
1295 ** then we aren't exposed to the HW bug.
1297 if ( ((boot_cpu_data.pdc.cpuid >> 5) & 0x7f) != 0x13
1298 || (boot_cpu_data.pdc.versions > 0x202)
1299 || (boot_cpu_data.pdc.capabilities & 0x08L) )
1300 return (void *) pdir_base;
1303 * PA8700 (PCX-W2, aka piranha) silent data corruption fix
1305 * An interaction between PA8700 CPU (Ver 2.2 or older) and
1306 * Ike/Astro can cause silent data corruption. This is only
1307 * a problem if the I/O PDIR is located in memory such that
1308 * (little-endian) bits 17 and 18 are on and bit 20 is off.
1310 * Since the max IO Pdir size is 2MB, by cleverly allocating the
1311 * right physical address, we can either avoid (IOPDIR <= 1MB)
1312 * or minimize (2MB IO Pdir) the problem if we restrict the
1313 * IO Pdir to a maximum size of 2MB-128K (1902K).
1315 * Because we always allocate 2^N sized IO pdirs, either of the
1316 * "bad" regions will be the last 128K if at all. That's easy
1320 if (pdir_order <= (19-12)) {
1321 if (((virt_to_phys(pdir_base)+pdir_size-1) & PIRANHA_ADDR_MASK) == PIRANHA_ADDR_VAL) {
1322 /* allocate a new one on 512k alignment */
1323 unsigned long new_pdir = __get_free_pages(GFP_KERNEL, (19-12));
1324 /* release original */
1325 free_pages(pdir_base, pdir_order);
1327 pdir_base = new_pdir;
1329 /* release excess */
1330 while (pdir_order < (19-12)) {
1331 new_pdir += pdir_size;
1332 free_pages(new_pdir, pdir_order);
1340 ** Needs to be aligned on an "odd" 1MB boundary.
1342 unsigned long new_pdir = __get_free_pages(GFP_KERNEL, pdir_order+1); /* 2 or 4MB */
1344 /* release original */
1345 free_pages( pdir_base, pdir_order);
1347 /* release first 1MB */
1348 free_pages(new_pdir, 20-12);
1350 pdir_base = new_pdir + 1024*1024;
1352 if (pdir_order > (20-12)) {
1356 ** Flag tells init_bitmap() to mark bad 128k as used
1357 ** and to reduce the size by 128k.
1359 piranha_bad_128k = 1;
1361 new_pdir += 3*1024*1024;
1362 /* release last 1MB */
1363 free_pages(new_pdir, 20-12);
1365 /* release unusable 128KB */
1366 free_pages(new_pdir - 128*1024 , 17-12);
1368 pdir_size -= 128*1024;
1372 memset((void *) pdir_base, 0, pdir_size);
1373 return (void *) pdir_base;
1376 static struct device *next_device(struct klist_iter *i)
1378 struct klist_node * n = klist_next(i);
1379 return n ? container_of(n, struct device, knode_parent) : NULL;
1382 /* setup Mercury or Elroy IBASE/IMASK registers. */
1384 setup_ibase_imask(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1386 /* lba_set_iregs() is in drivers/parisc/lba_pci.c */
1387 extern void lba_set_iregs(struct parisc_device *, u32, u32);
1389 struct klist_iter i;
1391 klist_iter_init(&sba->dev.klist_children, &i);
1392 while ((dev = next_device(&i))) {
1393 struct parisc_device *lba = to_parisc_device(dev);
1394 int rope_num = (lba->hpa.start >> 13) & 0xf;
1395 if (rope_num >> 3 == ioc_num)
1396 lba_set_iregs(lba, ioc->ibase, ioc->imask);
1398 klist_iter_exit(&i);
1402 sba_ioc_init_pluto(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1404 u32 iova_space_mask;
1405 u32 iova_space_size;
1406 int iov_order, tcnfg;
1407 #ifdef SBA_AGP_SUPPORT
1411 ** Firmware programs the base and size of a "safe IOVA space"
1412 ** (one that doesn't overlap memory or LMMIO space) in the
1413 ** IBASE and IMASK registers.
1415 ioc->ibase = READ_REG(ioc->ioc_hpa + IOC_IBASE);
1416 iova_space_size = ~(READ_REG(ioc->ioc_hpa + IOC_IMASK) & 0xFFFFFFFFUL) + 1;
1418 if ((ioc->ibase < 0xfed00000UL) && ((ioc->ibase + iova_space_size) > 0xfee00000UL)) {
1419 printk("WARNING: IOV space overlaps local config and interrupt message, truncating\n");
1420 iova_space_size /= 2;
1424 ** iov_order is always based on a 1GB IOVA space since we want to
1425 ** turn on the other half for AGP GART.
1427 iov_order = get_order(iova_space_size >> (IOVP_SHIFT - PAGE_SHIFT));
1428 ioc->pdir_size = (iova_space_size / IOVP_SIZE) * sizeof(u64);
1430 DBG_INIT("%s() hpa 0x%p IOV %dMB (%d bits)\n",
1431 __FUNCTION__, ioc->ioc_hpa, iova_space_size >> 20,
1432 iov_order + PAGE_SHIFT);
1434 ioc->pdir_base = (void *) __get_free_pages(GFP_KERNEL,
1435 get_order(ioc->pdir_size));
1436 if (!ioc->pdir_base)
1437 panic("Couldn't allocate I/O Page Table\n");
1439 memset(ioc->pdir_base, 0, ioc->pdir_size);
1441 DBG_INIT("%s() pdir %p size %x\n",
1442 __FUNCTION__, ioc->pdir_base, ioc->pdir_size);
1444 #ifdef SBA_HINT_SUPPORT
1445 ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1446 ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1448 DBG_INIT(" hint_shift_pdir %x hint_mask_pdir %lx\n",
1449 ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1452 WARN_ON((((unsigned long) ioc->pdir_base) & PAGE_MASK) != (unsigned long) ioc->pdir_base);
1453 WRITE_REG(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1455 /* build IMASK for IOC and Elroy */
1456 iova_space_mask = 0xffffffff;
1457 iova_space_mask <<= (iov_order + PAGE_SHIFT);
1458 ioc->imask = iova_space_mask;
1460 ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1462 sba_dump_tlb(ioc->ioc_hpa);
1464 setup_ibase_imask(sba, ioc, ioc_num);
1466 WRITE_REG(ioc->imask, ioc->ioc_hpa + IOC_IMASK);
1470 ** Setting the upper bits makes checking for bypass addresses
1471 ** a little faster later on.
1473 ioc->imask |= 0xFFFFFFFF00000000UL;
1476 /* Set I/O PDIR Page size to system page size */
1477 switch (PAGE_SHIFT) {
1478 case 12: tcnfg = 0; break; /* 4K */
1479 case 13: tcnfg = 1; break; /* 8K */
1480 case 14: tcnfg = 2; break; /* 16K */
1481 case 16: tcnfg = 3; break; /* 64K */
1483 panic(__FILE__ "Unsupported system page size %d",
1487 WRITE_REG(tcnfg, ioc->ioc_hpa + IOC_TCNFG);
1490 ** Program the IOC's ibase and enable IOVA translation
1491 ** Bit zero == enable bit.
1493 WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa + IOC_IBASE);
1496 ** Clear I/O TLB of any possible entries.
1497 ** (Yes. This is a bit paranoid...but so what)
1499 WRITE_REG(ioc->ibase | 31, ioc->ioc_hpa + IOC_PCOM);
1501 #ifdef SBA_AGP_SUPPORT
1503 ** If an AGP device is present, only use half of the IOV space
1504 ** for PCI DMA. Unfortunately we can't know ahead of time
1505 ** whether GART support will actually be used, for now we
1506 ** can just key on any AGP device found in the system.
1507 ** We program the next pdir index after we stop w/ a key for
1508 ** the GART code to handshake on.
1511 for (lba = sba->child; lba; lba = lba->sibling) {
1512 if (IS_QUICKSILVER(lba))
1517 DBG_INIT("%s: Reserving half of IOVA space for AGP GART support\n", __FUNCTION__);
1518 ioc->pdir_size /= 2;
1519 ((u64 *)ioc->pdir_base)[PDIR_INDEX(iova_space_size/2)] = SBA_IOMMU_COOKIE;
1521 DBG_INIT("%s: No GART needed - no AGP controller found\n", __FUNCTION__);
1528 sba_ioc_init(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1530 u32 iova_space_size, iova_space_mask;
1531 unsigned int pdir_size, iov_order;
1534 ** Determine IOVA Space size from memory size.
1536 ** Ideally, PCI drivers would register the maximum number
1537 ** of DMA they can have outstanding for each device they
1538 ** own. Next best thing would be to guess how much DMA
1539 ** can be outstanding based on PCI Class/sub-class. Both
1540 ** methods still require some "extra" to support PCI
1541 ** Hot-Plug/Removal of PCI cards. (aka PCI OLARD).
1543 ** While we have 32-bits "IOVA" space, top two 2 bits are used
1544 ** for DMA hints - ergo only 30 bits max.
1547 iova_space_size = (u32) (num_physpages/global_ioc_cnt);
1549 /* limit IOVA space size to 1MB-1GB */
1550 if (iova_space_size < (1 << (20 - PAGE_SHIFT))) {
1551 iova_space_size = 1 << (20 - PAGE_SHIFT);
1553 else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) {
1554 iova_space_size = 1 << (30 - PAGE_SHIFT);
1558 ** iova space must be log2() in size.
1559 ** thus, pdir/res_map will also be log2().
1560 ** PIRANHA BUG: Exception is when IO Pdir is 2MB (gets reduced)
1562 iov_order = get_order(iova_space_size << PAGE_SHIFT);
1564 /* iova_space_size is now bytes, not pages */
1565 iova_space_size = 1 << (iov_order + PAGE_SHIFT);
1567 ioc->pdir_size = pdir_size = (iova_space_size/IOVP_SIZE) * sizeof(u64);
1569 DBG_INIT("%s() hpa 0x%lx mem %ldMB IOV %dMB (%d bits)\n",
1572 (unsigned long) num_physpages >> (20 - PAGE_SHIFT),
1573 iova_space_size>>20,
1574 iov_order + PAGE_SHIFT);
1576 ioc->pdir_base = sba_alloc_pdir(pdir_size);
1578 DBG_INIT("%s() pdir %p size %x\n",
1579 __FUNCTION__, ioc->pdir_base, pdir_size);
1581 #ifdef SBA_HINT_SUPPORT
1582 /* FIXME : DMA HINTs not used */
1583 ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1584 ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1586 DBG_INIT(" hint_shift_pdir %x hint_mask_pdir %lx\n",
1587 ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1590 WRITE_REG64(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1592 /* build IMASK for IOC and Elroy */
1593 iova_space_mask = 0xffffffff;
1594 iova_space_mask <<= (iov_order + PAGE_SHIFT);
1597 ** On C3000 w/512MB mem, HP-UX 10.20 reports:
1598 ** ibase=0, imask=0xFE000000, size=0x2000000.
1601 ioc->imask = iova_space_mask; /* save it */
1603 ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1606 DBG_INIT("%s() IOV base 0x%lx mask 0x%0lx\n",
1607 __FUNCTION__, ioc->ibase, ioc->imask);
1610 ** FIXME: Hint registers are programmed with default hint
1611 ** values during boot, so hints should be sane even if we
1612 ** can't reprogram them the way drivers want.
1615 setup_ibase_imask(sba, ioc, ioc_num);
1618 ** Program the IOC's ibase and enable IOVA translation
1620 WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa+IOC_IBASE);
1621 WRITE_REG(ioc->imask, ioc->ioc_hpa+IOC_IMASK);
1623 /* Set I/O PDIR Page size to 4K */
1624 WRITE_REG(0, ioc->ioc_hpa+IOC_TCNFG);
1627 ** Clear I/O TLB of any possible entries.
1628 ** (Yes. This is a bit paranoid...but so what)
1630 WRITE_REG(0 | 31, ioc->ioc_hpa+IOC_PCOM);
1632 ioc->ibase = 0; /* used by SBA_IOVA and related macros */
1634 DBG_INIT("%s() DONE\n", __FUNCTION__);
1639 /**************************************************************************
1641 ** SBA initialization code (HW and SW)
1643 ** o identify SBA chip itself
1644 ** o initialize SBA chip modes (HardFail)
1645 ** o initialize SBA chip modes (HardFail)
1646 ** o FIXME: initialize DMA hints for reasonable defaults
1648 **************************************************************************/
1650 static void __iomem *ioc_remap(struct sba_device *sba_dev, unsigned int offset)
1652 return ioremap_nocache(sba_dev->dev->hpa.start + offset, SBA_FUNC_SIZE);
1655 static void sba_hw_init(struct sba_device *sba_dev)
1661 if (!is_pdc_pat()) {
1662 /* Shutdown the USB controller on Astro-based workstations.
1663 ** Once we reprogram the IOMMU, the next DMA performed by
1664 ** USB will HPMC the box. USB is only enabled if a
1665 ** keyboard is present and found.
1667 ** With serial console, j6k v5.0 firmware says:
1668 ** mem_kbd hpa 0xfee003f8 sba 0x0 pad 0x0 cl_class 0x7
1670 ** FIXME: Using GFX+USB console at power up but direct
1671 ** linux to serial console is still broken.
1672 ** USB could generate DMA so we must reset USB.
1673 ** The proper sequence would be:
1674 ** o block console output
1675 ** o reset USB device
1676 ** o reprogram serial port
1677 ** o unblock console output
1679 if (PAGE0->mem_kbd.cl_class == CL_KEYBD) {
1680 pdc_io_reset_devices();
1687 printk("sba_hw_init(): mem_boot 0x%x 0x%x 0x%x 0x%x\n", PAGE0->mem_boot.hpa,
1688 PAGE0->mem_boot.spa, PAGE0->mem_boot.pad, PAGE0->mem_boot.cl_class);
1691 ** Need to deal with DMA from LAN.
1692 ** Maybe use page zero boot device as a handle to talk
1693 ** to PDC about which device to shutdown.
1695 ** Netbooting, j6k v5.0 firmware says:
1696 ** mem_boot hpa 0xf4008000 sba 0x0 pad 0x0 cl_class 0x1002
1697 ** ARGH! invalid class.
1699 if ((PAGE0->mem_boot.cl_class != CL_RANDOM)
1700 && (PAGE0->mem_boot.cl_class != CL_SEQU)) {
1705 if (!IS_PLUTO(sba_dev->iodc)) {
1706 ioc_ctl = READ_REG(sba_dev->sba_hpa+IOC_CTRL);
1707 DBG_INIT("%s() hpa 0x%lx ioc_ctl 0x%Lx ->",
1708 __FUNCTION__, sba_dev->sba_hpa, ioc_ctl);
1709 ioc_ctl &= ~(IOC_CTRL_RM | IOC_CTRL_NC | IOC_CTRL_CE);
1710 ioc_ctl |= IOC_CTRL_DD | IOC_CTRL_D4 | IOC_CTRL_TC;
1711 /* j6700 v1.6 firmware sets 0x294f */
1712 /* A500 firmware sets 0x4d */
1714 WRITE_REG(ioc_ctl, sba_dev->sba_hpa+IOC_CTRL);
1716 #ifdef DEBUG_SBA_INIT
1717 ioc_ctl = READ_REG64(sba_dev->sba_hpa+IOC_CTRL);
1718 DBG_INIT(" 0x%Lx\n", ioc_ctl);
1722 if (IS_ASTRO(sba_dev->iodc)) {
1724 /* PAT_PDC (L-class) also reports the same goofy base */
1725 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, ASTRO_IOC_OFFSET);
1728 sba_dev->chip_resv.name = "Astro Intr Ack";
1729 sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfef00000UL;
1730 sba_dev->chip_resv.end = PCI_F_EXTEND | (0xff000000UL - 1) ;
1731 err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1734 } else if (IS_PLUTO(sba_dev->iodc)) {
1737 /* We use a negative value for IOC HPA so it gets
1738 * corrected when we add it with IKE's IOC offset.
1739 * Doesnt look clean, but fewer code.
1741 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, PLUTO_IOC_OFFSET);
1744 sba_dev->chip_resv.name = "Pluto Intr/PIOP/VGA";
1745 sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfee00000UL;
1746 sba_dev->chip_resv.end = PCI_F_EXTEND | (0xff200000UL - 1);
1747 err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1750 sba_dev->iommu_resv.name = "IOVA Space";
1751 sba_dev->iommu_resv.start = 0x40000000UL;
1752 sba_dev->iommu_resv.end = 0x50000000UL - 1;
1753 err = request_resource(&iomem_resource, &(sba_dev->iommu_resv));
1756 /* IS_IKE (ie N-class, L3000, L1500) */
1757 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(0));
1758 sba_dev->ioc[1].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(1));
1761 /* TODO - LOOKUP Ike/Stretch chipset mem map */
1763 /* XXX: What about Reo? */
1765 sba_dev->num_ioc = num_ioc;
1766 for (i = 0; i < num_ioc; i++) {
1767 void __iomem *ioc_hpa = sba_dev->ioc[i].ioc_hpa;
1770 for (j=0; j < sizeof(u64) * ROPES_PER_IOC; j+=sizeof(u64)) {
1773 * Clear ROPE(N)_CONFIG AO bit.
1774 * Disables "NT Ordering" (~= !"Relaxed Ordering")
1775 * Overrides bit 1 in DMA Hint Sets.
1776 * Improves netperf UDP_STREAM by ~10% for bcm5701.
1778 if (IS_PLUTO(sba_dev->iodc)) {
1779 void __iomem *rope_cfg;
1780 unsigned long cfg_val;
1782 rope_cfg = ioc_hpa + IOC_ROPE0_CFG + j;
1783 cfg_val = READ_REG(rope_cfg);
1784 cfg_val &= ~IOC_ROPE_AO;
1785 WRITE_REG(cfg_val, rope_cfg);
1789 ** Make sure the box crashes on rope errors.
1791 WRITE_REG(HF_ENABLE, ioc_hpa + ROPE0_CTL + j);
1794 /* flush out the last writes */
1795 READ_REG(sba_dev->ioc[i].ioc_hpa + ROPE7_CTL);
1797 DBG_INIT(" ioc[%d] ROPE_CFG 0x%Lx ROPE_DBG 0x%Lx\n",
1799 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x40),
1800 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x50)
1802 DBG_INIT(" STATUS_CONTROL 0x%Lx FLUSH_CTRL 0x%Lx\n",
1803 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x108),
1804 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x400)
1807 if (IS_PLUTO(sba_dev->iodc)) {
1808 sba_ioc_init_pluto(sba_dev->dev, &(sba_dev->ioc[i]), i);
1810 sba_ioc_init(sba_dev->dev, &(sba_dev->ioc[i]), i);
1816 sba_common_init(struct sba_device *sba_dev)
1820 /* add this one to the head of the list (order doesn't matter)
1821 ** This will be useful for debugging - especially if we get coredumps
1823 sba_dev->next = sba_list;
1826 for(i=0; i< sba_dev->num_ioc; i++) {
1828 #ifdef DEBUG_DMB_TRAP
1829 extern void iterate_pages(unsigned long , unsigned long ,
1830 void (*)(pte_t * , unsigned long),
1832 void set_data_memory_break(pte_t * , unsigned long);
1834 /* resource map size dictated by pdir_size */
1835 res_size = sba_dev->ioc[i].pdir_size/sizeof(u64); /* entries */
1837 /* Second part of PIRANHA BUG */
1838 if (piranha_bad_128k) {
1839 res_size -= (128*1024)/sizeof(u64);
1842 res_size >>= 3; /* convert bit count to byte count */
1843 DBG_INIT("%s() res_size 0x%x\n",
1844 __FUNCTION__, res_size);
1846 sba_dev->ioc[i].res_size = res_size;
1847 sba_dev->ioc[i].res_map = (char *) __get_free_pages(GFP_KERNEL, get_order(res_size));
1849 #ifdef DEBUG_DMB_TRAP
1850 iterate_pages( sba_dev->ioc[i].res_map, res_size,
1851 set_data_memory_break, 0);
1854 if (NULL == sba_dev->ioc[i].res_map)
1856 panic("%s:%s() could not allocate resource map\n",
1857 __FILE__, __FUNCTION__ );
1860 memset(sba_dev->ioc[i].res_map, 0, res_size);
1861 /* next available IOVP - circular search */
1862 sba_dev->ioc[i].res_hint = (unsigned long *)
1863 &(sba_dev->ioc[i].res_map[L1_CACHE_BYTES]);
1865 #ifdef ASSERT_PDIR_SANITY
1866 /* Mark first bit busy - ie no IOVA 0 */
1867 sba_dev->ioc[i].res_map[0] = 0x80;
1868 sba_dev->ioc[i].pdir_base[0] = 0xeeffc0addbba0080ULL;
1871 /* Third (and last) part of PIRANHA BUG */
1872 if (piranha_bad_128k) {
1873 /* region from +1408K to +1536 is un-usable. */
1875 int idx_start = (1408*1024/sizeof(u64)) >> 3;
1876 int idx_end = (1536*1024/sizeof(u64)) >> 3;
1877 long *p_start = (long *) &(sba_dev->ioc[i].res_map[idx_start]);
1878 long *p_end = (long *) &(sba_dev->ioc[i].res_map[idx_end]);
1880 /* mark that part of the io pdir busy */
1881 while (p_start < p_end)
1886 #ifdef DEBUG_DMB_TRAP
1887 iterate_pages( sba_dev->ioc[i].res_map, res_size,
1888 set_data_memory_break, 0);
1889 iterate_pages( sba_dev->ioc[i].pdir_base, sba_dev->ioc[i].pdir_size,
1890 set_data_memory_break, 0);
1893 DBG_INIT("%s() %d res_map %x %p\n",
1894 __FUNCTION__, i, res_size, sba_dev->ioc[i].res_map);
1897 spin_lock_init(&sba_dev->sba_lock);
1898 ioc_needs_fdc = boot_cpu_data.pdc.capabilities & PDC_MODEL_IOPDIR_FDC;
1900 #ifdef DEBUG_SBA_INIT
1902 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
1903 * (bit #61, big endian), we have to flush and sync every time
1904 * IO-PDIR is changed in Ike/Astro.
1906 if (ioc_needs_fdc) {
1907 printk(KERN_INFO MODULE_NAME " FDC/SYNC required.\n");
1909 printk(KERN_INFO MODULE_NAME " IOC has cache coherent PDIR.\n");
1914 #ifdef CONFIG_PROC_FS
1915 static int sba_proc_info(struct seq_file *m, void *p)
1917 struct sba_device *sba_dev = sba_list;
1918 struct ioc *ioc = &sba_dev->ioc[0]; /* FIXME: Multi-IOC support! */
1919 int total_pages = (int) (ioc->res_size << 3); /* 8 bits per byte */
1920 #ifdef SBA_COLLECT_STATS
1921 unsigned long avg = 0, min, max;
1925 len += seq_printf(m, "%s rev %d.%d\n",
1927 (sba_dev->hw_rev & 0x7) + 1,
1928 (sba_dev->hw_rev & 0x18) >> 3
1930 len += seq_printf(m, "IO PDIR size : %d bytes (%d entries)\n",
1931 (int) ((ioc->res_size << 3) * sizeof(u64)), /* 8 bits/byte */
1934 len += seq_printf(m, "Resource bitmap : %d bytes (%d pages)\n",
1935 ioc->res_size, ioc->res_size << 3); /* 8 bits per byte */
1937 len += seq_printf(m, "LMMIO_BASE/MASK/ROUTE %08x %08x %08x\n",
1938 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_BASE),
1939 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_MASK),
1940 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_ROUTE)
1944 len += seq_printf(m, "DIR%d_BASE/MASK/ROUTE %08x %08x %08x\n", i,
1945 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_BASE + i*0x18),
1946 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_MASK + i*0x18),
1947 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_ROUTE + i*0x18)
1950 #ifdef SBA_COLLECT_STATS
1951 len += seq_printf(m, "IO PDIR entries : %ld free %ld used (%d%%)\n",
1952 total_pages - ioc->used_pages, ioc->used_pages,
1953 (int) (ioc->used_pages * 100 / total_pages));
1955 min = max = ioc->avg_search[0];
1956 for (i = 0; i < SBA_SEARCH_SAMPLE; i++) {
1957 avg += ioc->avg_search[i];
1958 if (ioc->avg_search[i] > max) max = ioc->avg_search[i];
1959 if (ioc->avg_search[i] < min) min = ioc->avg_search[i];
1961 avg /= SBA_SEARCH_SAMPLE;
1962 len += seq_printf(m, " Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n",
1965 len += seq_printf(m, "pci_map_single(): %12ld calls %12ld pages (avg %d/1000)\n",
1966 ioc->msingle_calls, ioc->msingle_pages,
1967 (int) ((ioc->msingle_pages * 1000)/ioc->msingle_calls));
1969 /* KLUGE - unmap_sg calls unmap_single for each mapped page */
1970 min = ioc->usingle_calls;
1971 max = ioc->usingle_pages - ioc->usg_pages;
1972 len += seq_printf(m, "pci_unmap_single: %12ld calls %12ld pages (avg %d/1000)\n",
1973 min, max, (int) ((max * 1000)/min));
1975 len += seq_printf(m, "pci_map_sg() : %12ld calls %12ld pages (avg %d/1000)\n",
1976 ioc->msg_calls, ioc->msg_pages,
1977 (int) ((ioc->msg_pages * 1000)/ioc->msg_calls));
1979 len += seq_printf(m, "pci_unmap_sg() : %12ld calls %12ld pages (avg %d/1000)\n",
1980 ioc->usg_calls, ioc->usg_pages,
1981 (int) ((ioc->usg_pages * 1000)/ioc->usg_calls));
1988 sba_proc_open(struct inode *i, struct file *f)
1990 return single_open(f, &sba_proc_info, NULL);
1993 static struct file_operations sba_proc_fops = {
1994 .owner = THIS_MODULE,
1995 .open = sba_proc_open,
1997 .llseek = seq_lseek,
1998 .release = single_release,
2002 sba_proc_bitmap_info(struct seq_file *m, void *p)
2004 struct sba_device *sba_dev = sba_list;
2005 struct ioc *ioc = &sba_dev->ioc[0]; /* FIXME: Multi-IOC support! */
2006 unsigned int *res_ptr = (unsigned int *)ioc->res_map;
2009 for (i = 0; i < (ioc->res_size/sizeof(unsigned int)); ++i, ++res_ptr) {
2011 len += seq_printf(m, "\n ");
2012 len += seq_printf(m, " %08x", *res_ptr);
2014 len += seq_printf(m, "\n");
2020 sba_proc_bitmap_open(struct inode *i, struct file *f)
2022 return single_open(f, &sba_proc_bitmap_info, NULL);
2025 static struct file_operations sba_proc_bitmap_fops = {
2026 .owner = THIS_MODULE,
2027 .open = sba_proc_bitmap_open,
2029 .llseek = seq_lseek,
2030 .release = single_release,
2032 #endif /* CONFIG_PROC_FS */
2034 static struct parisc_device_id sba_tbl[] = {
2035 { HPHW_IOA, HVERSION_REV_ANY_ID, ASTRO_RUNWAY_PORT, 0xb },
2036 { HPHW_BCPORT, HVERSION_REV_ANY_ID, IKE_MERCED_PORT, 0xc },
2037 { HPHW_BCPORT, HVERSION_REV_ANY_ID, REO_MERCED_PORT, 0xc },
2038 { HPHW_BCPORT, HVERSION_REV_ANY_ID, REOG_MERCED_PORT, 0xc },
2039 { HPHW_IOA, HVERSION_REV_ANY_ID, PLUTO_MCKINLEY_PORT, 0xc },
2043 int sba_driver_callback(struct parisc_device *);
2045 static struct parisc_driver sba_driver = {
2046 .name = MODULE_NAME,
2047 .id_table = sba_tbl,
2048 .probe = sba_driver_callback,
2052 ** Determine if sba should claim this chip (return 0) or not (return 1).
2053 ** If so, initialize the chip and tell other partners in crime they
2057 sba_driver_callback(struct parisc_device *dev)
2059 struct sba_device *sba_dev;
2063 void __iomem *sba_addr = ioremap_nocache(dev->hpa.start, SBA_FUNC_SIZE);
2064 struct proc_dir_entry *info_entry, *bitmap_entry, *root;
2066 sba_dump_ranges(sba_addr);
2068 /* Read HW Rev First */
2069 func_class = READ_REG(sba_addr + SBA_FCLASS);
2071 if (IS_ASTRO(&dev->id)) {
2072 unsigned long fclass;
2073 static char astro_rev[]="Astro ?.?";
2075 /* Astro is broken...Read HW Rev First */
2076 fclass = READ_REG(sba_addr);
2078 astro_rev[6] = '1' + (char) (fclass & 0x7);
2079 astro_rev[8] = '0' + (char) ((fclass & 0x18) >> 3);
2080 version = astro_rev;
2082 } else if (IS_IKE(&dev->id)) {
2083 static char ike_rev[] = "Ike rev ?";
2084 ike_rev[8] = '0' + (char) (func_class & 0xff);
2086 } else if (IS_PLUTO(&dev->id)) {
2087 static char pluto_rev[]="Pluto ?.?";
2088 pluto_rev[6] = '0' + (char) ((func_class & 0xf0) >> 4);
2089 pluto_rev[8] = '0' + (char) (func_class & 0x0f);
2090 version = pluto_rev;
2092 static char reo_rev[] = "REO rev ?";
2093 reo_rev[8] = '0' + (char) (func_class & 0xff);
2097 if (!global_ioc_cnt) {
2098 global_ioc_cnt = count_parisc_driver(&sba_driver);
2100 /* Astro and Pluto have one IOC per SBA */
2101 if ((!IS_ASTRO(&dev->id)) || (!IS_PLUTO(&dev->id)))
2102 global_ioc_cnt *= 2;
2105 printk(KERN_INFO "%s found %s at 0x%lx\n",
2106 MODULE_NAME, version, dev->hpa.start);
2108 sba_dev = kzalloc(sizeof(struct sba_device), GFP_KERNEL);
2110 printk(KERN_ERR MODULE_NAME " - couldn't alloc sba_device\n");
2114 parisc_set_drvdata(dev, sba_dev);
2116 for(i=0; i<MAX_IOC; i++)
2117 spin_lock_init(&(sba_dev->ioc[i].res_lock));
2120 sba_dev->hw_rev = func_class;
2121 sba_dev->iodc = &dev->id;
2122 sba_dev->name = dev->name;
2123 sba_dev->sba_hpa = sba_addr;
2125 sba_get_pat_resources(sba_dev);
2126 sba_hw_init(sba_dev);
2127 sba_common_init(sba_dev);
2129 hppa_dma_ops = &sba_ops;
2131 #ifdef CONFIG_PROC_FS
2132 switch (dev->id.hversion) {
2133 case PLUTO_MCKINLEY_PORT:
2134 root = proc_mckinley_root;
2136 case ASTRO_RUNWAY_PORT:
2137 case IKE_MERCED_PORT:
2139 root = proc_runway_root;
2143 info_entry = create_proc_entry("sba_iommu", 0, root);
2144 bitmap_entry = create_proc_entry("sba_iommu-bitmap", 0, root);
2147 info_entry->proc_fops = &sba_proc_fops;
2150 bitmap_entry->proc_fops = &sba_proc_bitmap_fops;
2153 parisc_vmerge_boundary = IOVP_SIZE;
2154 parisc_vmerge_max_size = IOVP_SIZE * BITS_PER_LONG;
2160 ** One time initialization to let the world know the SBA was found.
2161 ** This is the only routine which is NOT static.
2162 ** Must be called exactly once before pci_init().
2164 void __init sba_init(void)
2166 register_parisc_driver(&sba_driver);
2171 * sba_get_iommu - Assign the iommu pointer for the pci bus controller.
2172 * @dev: The parisc device.
2174 * Returns the appropriate IOMMU data for the given parisc PCI controller.
2175 * This is cached and used later for PCI DMA Mapping.
2177 void * sba_get_iommu(struct parisc_device *pci_hba)
2179 struct parisc_device *sba_dev = parisc_parent(pci_hba);
2180 struct sba_device *sba = sba_dev->dev.driver_data;
2181 char t = sba_dev->id.hw_type;
2182 int iocnum = (pci_hba->hw_path >> 3); /* rope # */
2184 WARN_ON((t != HPHW_IOA) && (t != HPHW_BCPORT));
2186 return &(sba->ioc[iocnum]);
2191 * sba_directed_lmmio - return first directed LMMIO range routed to rope
2192 * @pa_dev: The parisc device.
2193 * @r: resource PCI host controller wants start/end fields assigned.
2195 * For the given parisc PCI controller, determine if any direct ranges
2196 * are routed down the corresponding rope.
2198 void sba_directed_lmmio(struct parisc_device *pci_hba, struct resource *r)
2200 struct parisc_device *sba_dev = parisc_parent(pci_hba);
2201 struct sba_device *sba = sba_dev->dev.driver_data;
2202 char t = sba_dev->id.hw_type;
2204 int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1)); /* rope # */
2206 BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2208 r->start = r->end = 0;
2210 /* Astro has 4 directed ranges. Not sure about Ike/Pluto/et al */
2211 for (i=0; i<4; i++) {
2213 void __iomem *reg = sba->sba_hpa + i*0x18;
2215 base = READ_REG32(reg + LMMIO_DIRECT0_BASE);
2216 if ((base & 1) == 0)
2217 continue; /* not enabled */
2219 size = READ_REG32(reg + LMMIO_DIRECT0_ROUTE);
2221 if ((size & (ROPES_PER_IOC-1)) != rope)
2222 continue; /* directed down different rope */
2224 r->start = (base & ~1UL) | PCI_F_EXTEND;
2225 size = ~ READ_REG32(reg + LMMIO_DIRECT0_MASK);
2226 r->end = r->start + size;
2232 * sba_distributed_lmmio - return portion of distributed LMMIO range
2233 * @pa_dev: The parisc device.
2234 * @r: resource PCI host controller wants start/end fields assigned.
2236 * For the given parisc PCI controller, return portion of distributed LMMIO
2237 * range. The distributed LMMIO is always present and it's just a question
2238 * of the base address and size of the range.
2240 void sba_distributed_lmmio(struct parisc_device *pci_hba, struct resource *r )
2242 struct parisc_device *sba_dev = parisc_parent(pci_hba);
2243 struct sba_device *sba = sba_dev->dev.driver_data;
2244 char t = sba_dev->id.hw_type;
2246 int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1)); /* rope # */
2248 BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2250 r->start = r->end = 0;
2252 base = READ_REG32(sba->sba_hpa + LMMIO_DIST_BASE);
2253 if ((base & 1) == 0) {
2254 BUG(); /* Gah! Distr Range wasn't enabled! */
2258 r->start = (base & ~1UL) | PCI_F_EXTEND;
2260 size = (~READ_REG32(sba->sba_hpa + LMMIO_DIST_MASK)) / ROPES_PER_IOC;
2261 r->start += rope * (size + 1); /* adjust base for this rope */
2262 r->end = r->start + size;