1 // SPDX-License-Identifier: GPL-2.0-or-later
3 ** System Bus Adapter (SBA) I/O MMU manager
5 ** (c) Copyright 2000-2004 Grant Grundler <grundler @ parisc-linux x org>
6 ** (c) Copyright 2004 Naresh Kumar Inna <knaresh at india x hp x com>
7 ** (c) Copyright 2000-2004 Hewlett-Packard Company
9 ** Portions (c) 1999 Dave S. Miller (from sparc64 I/O MMU code)
13 ** This module initializes the IOC (I/O Controller) found on B1000/C3000/
14 ** J5000/J7000/N-class/L-class machines and their successors.
16 ** FIXME: add DMA hint support programming in both sba and lba modules.
19 #include <linux/types.h>
20 #include <linux/kernel.h>
21 #include <linux/spinlock.h>
22 #include <linux/slab.h>
23 #include <linux/init.h>
26 #include <linux/string.h>
27 #include <linux/pci.h>
28 #include <linux/dma-map-ops.h>
29 #include <linux/scatterlist.h>
30 #include <linux/iommu-helper.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>
40 #include <linux/module.h>
42 #include <asm/ropes.h>
43 #include <asm/mckinley.h> /* for proc_mckinley_root */
44 #include <asm/runway.h> /* for proc_runway_root */
45 #include <asm/page.h> /* for PAGE0 */
46 #include <asm/pdc.h> /* for PDC_MODEL_* */
47 #include <asm/pdcpat.h> /* for is_pdc_pat() */
48 #include <asm/parisc-device.h>
52 #define MODULE_NAME "SBA"
55 ** The number of debug flags is a clue - this code is fragile.
56 ** Don't even think about messing with it unless you have
57 ** plenty of 710's to sacrifice to the computer gods. :^)
61 #undef DEBUG_SBA_RUN_SG
62 #undef DEBUG_SBA_RESOURCE
63 #undef ASSERT_PDIR_SANITY
64 #undef DEBUG_LARGE_SG_ENTRIES
68 #define DBG_INIT(x...) printk(x)
70 #define DBG_INIT(x...)
74 #define DBG_RUN(x...) printk(x)
79 #ifdef DEBUG_SBA_RUN_SG
80 #define DBG_RUN_SG(x...) printk(x)
82 #define DBG_RUN_SG(x...)
86 #ifdef DEBUG_SBA_RESOURCE
87 #define DBG_RES(x...) printk(x)
92 #define SBA_INLINE __inline__
94 #define DEFAULT_DMA_HINT_REG 0
96 struct sba_device *sba_list;
97 EXPORT_SYMBOL_GPL(sba_list);
99 static unsigned long ioc_needs_fdc = 0;
101 /* global count of IOMMUs in the system */
102 static unsigned int global_ioc_cnt = 0;
104 /* PA8700 (Piranha 2.2) bug workaround */
105 static unsigned long piranha_bad_128k = 0;
107 /* Looks nice and keeps the compiler happy */
108 #define SBA_DEV(d) ((struct sba_device *) (d))
110 #ifdef CONFIG_AGP_PARISC
111 #define SBA_AGP_SUPPORT
112 #endif /*CONFIG_AGP_PARISC*/
114 #ifdef SBA_AGP_SUPPORT
115 static int sba_reserve_agpgart = 1;
116 module_param(sba_reserve_agpgart, int, 0444);
117 MODULE_PARM_DESC(sba_reserve_agpgart, "Reserve half of IO pdir as AGPGART");
121 /************************************
122 ** SBA register read and write support
124 ** BE WARNED: register writes are posted.
125 ** (ie follow writes which must reach HW with a read)
127 ** Superdome (in particular, REO) allows only 64-bit CSR accesses.
129 #define READ_REG32(addr) readl(addr)
130 #define READ_REG64(addr) readq(addr)
131 #define WRITE_REG32(val, addr) writel((val), (addr))
132 #define WRITE_REG64(val, addr) writeq((val), (addr))
135 #define READ_REG(addr) READ_REG64(addr)
136 #define WRITE_REG(value, addr) WRITE_REG64(value, addr)
138 #define READ_REG(addr) READ_REG32(addr)
139 #define WRITE_REG(value, addr) WRITE_REG32(value, addr)
142 #ifdef DEBUG_SBA_INIT
144 /* NOTE: When CONFIG_64BIT isn't defined, READ_REG64() is two 32-bit reads */
147 * sba_dump_ranges - debugging only - print ranges assigned to this IOA
148 * @hpa: base address of the sba
150 * Print the MMIO and IO Port address ranges forwarded by an Astro/Ike/RIO
151 * IO Adapter (aka Bus Converter).
154 sba_dump_ranges(void __iomem *hpa)
156 DBG_INIT("SBA at 0x%p\n", hpa);
157 DBG_INIT("IOS_DIST_BASE : %Lx\n", READ_REG64(hpa+IOS_DIST_BASE));
158 DBG_INIT("IOS_DIST_MASK : %Lx\n", READ_REG64(hpa+IOS_DIST_MASK));
159 DBG_INIT("IOS_DIST_ROUTE : %Lx\n", READ_REG64(hpa+IOS_DIST_ROUTE));
161 DBG_INIT("IOS_DIRECT_BASE : %Lx\n", READ_REG64(hpa+IOS_DIRECT_BASE));
162 DBG_INIT("IOS_DIRECT_MASK : %Lx\n", READ_REG64(hpa+IOS_DIRECT_MASK));
163 DBG_INIT("IOS_DIRECT_ROUTE: %Lx\n", READ_REG64(hpa+IOS_DIRECT_ROUTE));
167 * sba_dump_tlb - debugging only - print IOMMU operating parameters
168 * @hpa: base address of the IOMMU
170 * Print the size/location of the IO MMU PDIR.
172 static void sba_dump_tlb(void __iomem *hpa)
174 DBG_INIT("IO TLB at 0x%p\n", hpa);
175 DBG_INIT("IOC_IBASE : 0x%Lx\n", READ_REG64(hpa+IOC_IBASE));
176 DBG_INIT("IOC_IMASK : 0x%Lx\n", READ_REG64(hpa+IOC_IMASK));
177 DBG_INIT("IOC_TCNFG : 0x%Lx\n", READ_REG64(hpa+IOC_TCNFG));
178 DBG_INIT("IOC_PDIR_BASE: 0x%Lx\n", READ_REG64(hpa+IOC_PDIR_BASE));
182 #define sba_dump_ranges(x)
183 #define sba_dump_tlb(x)
184 #endif /* DEBUG_SBA_INIT */
187 #ifdef ASSERT_PDIR_SANITY
190 * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry
191 * @ioc: IO MMU structure which owns the pdir we are interested in.
192 * @msg: text to print ont the output line.
195 * Print one entry of the IO MMU PDIR in human readable form.
198 sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide)
200 /* start printing from lowest pde in rval */
201 u64 *ptr = &(ioc->pdir_base[pide & (~0U * BITS_PER_LONG)]);
202 unsigned long *rptr = (unsigned long *) &(ioc->res_map[(pide >>3) & ~(sizeof(unsigned long) - 1)]);
205 printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n",
207 rptr, pide & (BITS_PER_LONG - 1), *rptr);
210 while (rcnt < BITS_PER_LONG) {
211 printk(KERN_DEBUG "%s %2d %p %016Lx\n",
212 (rcnt == (pide & (BITS_PER_LONG - 1)))
218 printk(KERN_DEBUG "%s", msg);
223 * sba_check_pdir - debugging only - consistency checker
224 * @ioc: IO MMU structure which owns the pdir we are interested in.
225 * @msg: text to print ont the output line.
227 * Verify the resource map and pdir state is consistent
230 sba_check_pdir(struct ioc *ioc, char *msg)
232 u32 *rptr_end = (u32 *) &(ioc->res_map[ioc->res_size]);
233 u32 *rptr = (u32 *) ioc->res_map; /* resource map ptr */
234 u64 *pptr = ioc->pdir_base; /* pdir ptr */
237 while (rptr < rptr_end) {
239 int rcnt = 32; /* number of bits we might check */
242 /* Get last byte and highest bit from that */
243 u32 pde = ((u32) (((char *)pptr)[7])) << 24;
244 if ((rval ^ pde) & 0x80000000)
247 ** BUMMER! -- res_map != pdir --
248 ** Dump rval and matching pdir entries
250 sba_dump_pdir_entry(ioc, msg, pide);
254 rval <<= 1; /* try the next bit */
258 rptr++; /* look at next word of res_map */
260 /* It'd be nice if we always got here :^) */
266 * sba_dump_sg - debugging only - print Scatter-Gather list
267 * @ioc: IO MMU structure which owns the pdir we are interested in.
268 * @startsg: head of the SG list
269 * @nents: number of entries in SG list
271 * print the SG list so we can verify it's correct by hand.
274 sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
276 while (nents-- > 0) {
277 printk(KERN_DEBUG " %d : %08lx/%05x %p/%05x\n",
279 (unsigned long) sg_dma_address(startsg),
281 sg_virt(startsg), startsg->length);
286 #endif /* ASSERT_PDIR_SANITY */
291 /**************************************************************
293 * I/O Pdir Resource Management
295 * Bits set in the resource map are in use.
296 * Each bit can represent a number of pages.
297 * LSbs represent lower addresses (IOVA's).
299 ***************************************************************/
300 #define PAGES_PER_RANGE 1 /* could increase this to 4 or 8 if needed */
302 /* Convert from IOVP to IOVA and vice versa. */
305 /* Pluto (aka ZX1) boxes need to set or clear the ibase bits appropriately */
306 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((ioc->ibase) | (iovp) | (offset))
307 #define SBA_IOVP(ioc,iova) ((iova) & (ioc)->iovp_mask)
309 /* only support Astro and ancestors. Saves a few cycles in key places */
310 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((iovp) | (offset))
311 #define SBA_IOVP(ioc,iova) (iova)
314 #define PDIR_INDEX(iovp) ((iovp)>>IOVP_SHIFT)
316 #define RESMAP_MASK(n) (~0UL << (BITS_PER_LONG - (n)))
317 #define RESMAP_IDX_MASK (sizeof(unsigned long) - 1)
319 static unsigned long ptr_to_pide(struct ioc *ioc, unsigned long *res_ptr,
320 unsigned int bitshiftcnt)
322 return (((unsigned long)res_ptr - (unsigned long)ioc->res_map) << 3)
327 * sba_search_bitmap - find free space in IO PDIR resource bitmap
328 * @ioc: IO MMU structure which owns the pdir we are interested in.
329 * @bits_wanted: number of entries we need.
331 * Find consecutive free bits in resource bitmap.
332 * Each bit represents one entry in the IO Pdir.
333 * Cool perf optimization: search for log2(size) bits at a time.
335 static SBA_INLINE unsigned long
336 sba_search_bitmap(struct ioc *ioc, struct device *dev,
337 unsigned long bits_wanted)
339 unsigned long *res_ptr = ioc->res_hint;
340 unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]);
341 unsigned long pide = ~0UL, tpide;
342 unsigned long boundary_size;
346 boundary_size = dma_get_seg_boundary_nr_pages(dev, IOVP_SHIFT);
348 #if defined(ZX1_SUPPORT)
349 BUG_ON(ioc->ibase & ~IOVP_MASK);
350 shift = ioc->ibase >> IOVP_SHIFT;
355 if (bits_wanted > (BITS_PER_LONG/2)) {
356 /* Search word at a time - no mask needed */
357 for(; res_ptr < res_end; ++res_ptr) {
358 tpide = ptr_to_pide(ioc, res_ptr, 0);
359 ret = iommu_is_span_boundary(tpide, bits_wanted,
362 if ((*res_ptr == 0) && !ret) {
363 *res_ptr = RESMAP_MASK(bits_wanted);
368 /* point to the next word on next pass */
370 ioc->res_bitshift = 0;
373 ** Search the resource bit map on well-aligned values.
374 ** "o" is the alignment.
375 ** We need the alignment to invalidate I/O TLB using
376 ** SBA HW features in the unmap path.
378 unsigned long o = 1 << get_order(bits_wanted << PAGE_SHIFT);
379 uint bitshiftcnt = ALIGN(ioc->res_bitshift, o);
382 if (bitshiftcnt >= BITS_PER_LONG) {
386 mask = RESMAP_MASK(bits_wanted) >> bitshiftcnt;
388 DBG_RES("%s() o %ld %p", __func__, o, res_ptr);
389 while(res_ptr < res_end)
391 DBG_RES(" %p %lx %lx\n", res_ptr, mask, *res_ptr);
393 tpide = ptr_to_pide(ioc, res_ptr, bitshiftcnt);
394 ret = iommu_is_span_boundary(tpide, bits_wanted,
397 if ((((*res_ptr) & mask) == 0) && !ret) {
398 *res_ptr |= mask; /* mark resources busy! */
405 mask = RESMAP_MASK(bits_wanted);
410 /* look in the same word on the next pass */
411 ioc->res_bitshift = bitshiftcnt + bits_wanted;
415 if (res_end <= res_ptr) {
416 ioc->res_hint = (unsigned long *) ioc->res_map;
417 ioc->res_bitshift = 0;
419 ioc->res_hint = res_ptr;
426 * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
427 * @ioc: IO MMU structure which owns the pdir we are interested in.
428 * @size: number of bytes to create a mapping for
430 * Given a size, find consecutive unmarked and then mark those bits in the
434 sba_alloc_range(struct ioc *ioc, struct device *dev, size_t size)
436 unsigned int pages_needed = size >> IOVP_SHIFT;
437 #ifdef SBA_COLLECT_STATS
438 unsigned long cr_start = mfctl(16);
442 pide = sba_search_bitmap(ioc, dev, pages_needed);
443 if (pide >= (ioc->res_size << 3)) {
444 pide = sba_search_bitmap(ioc, dev, pages_needed);
445 if (pide >= (ioc->res_size << 3))
446 panic("%s: I/O MMU @ %p is out of mapping resources\n",
447 __FILE__, ioc->ioc_hpa);
450 #ifdef ASSERT_PDIR_SANITY
451 /* verify the first enable bit is clear */
452 if(0x00 != ((u8 *) ioc->pdir_base)[pide*sizeof(u64) + 7]) {
453 sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide);
457 DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
458 __func__, size, pages_needed, pide,
459 (uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map),
462 #ifdef SBA_COLLECT_STATS
464 unsigned long cr_end = mfctl(16);
465 unsigned long tmp = cr_end - cr_start;
466 /* check for roll over */
467 cr_start = (cr_end < cr_start) ? -(tmp) : (tmp);
469 ioc->avg_search[ioc->avg_idx++] = cr_start;
470 ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;
472 ioc->used_pages += pages_needed;
480 * sba_free_range - unmark bits in IO PDIR resource bitmap
481 * @ioc: IO MMU structure which owns the pdir we are interested in.
482 * @iova: IO virtual address which was previously allocated.
483 * @size: number of bytes to create a mapping for
485 * clear bits in the ioc's resource map
487 static SBA_INLINE void
488 sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size)
490 unsigned long iovp = SBA_IOVP(ioc, iova);
491 unsigned int pide = PDIR_INDEX(iovp);
492 unsigned int ridx = pide >> 3; /* convert bit to byte address */
493 unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]);
495 int bits_not_wanted = size >> IOVP_SHIFT;
497 /* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
498 unsigned long m = RESMAP_MASK(bits_not_wanted) >> (pide & (BITS_PER_LONG - 1));
500 DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n",
501 __func__, (uint) iova, size,
502 bits_not_wanted, m, pide, res_ptr, *res_ptr);
504 #ifdef SBA_COLLECT_STATS
505 ioc->used_pages -= bits_not_wanted;
512 /**************************************************************
514 * "Dynamic DMA Mapping" support (aka "Coherent I/O")
516 ***************************************************************/
518 #ifdef SBA_HINT_SUPPORT
519 #define SBA_DMA_HINT(ioc, val) ((val) << (ioc)->hint_shift_pdir)
522 typedef unsigned long space_t;
523 #define KERNEL_SPACE 0
526 * sba_io_pdir_entry - fill in one IO PDIR entry
527 * @pdir_ptr: pointer to IO PDIR entry
528 * @sid: process Space ID - currently only support KERNEL_SPACE
529 * @vba: Virtual CPU address of buffer to map
530 * @hint: DMA hint set to use for this mapping
532 * SBA Mapping Routine
534 * Given a virtual address (vba, arg2) and space id, (sid, arg1)
535 * sba_io_pdir_entry() loads the I/O PDIR entry pointed to by
537 * Using the bass-ackwards HP bit numbering, Each IO Pdir entry
538 * for Astro/Ike looks like:
542 * +-+---------------------+----------------------------------+----+--------+
543 * |V| U | PPN[43:12] | U | VI |
544 * +-+---------------------+----------------------------------+----+--------+
546 * Pluto is basically identical, supports fewer physical address bits:
549 * +-+------------------------+-------------------------------+----+--------+
550 * |V| U | PPN[39:12] | U | VI |
551 * +-+------------------------+-------------------------------+----+--------+
553 * V == Valid Bit (Most Significant Bit is bit 0)
555 * PPN == Physical Page Number
556 * VI == Virtual Index (aka Coherent Index)
558 * LPA instruction output is put into PPN field.
559 * LCI (Load Coherence Index) instruction provides the "VI" bits.
561 * We pre-swap the bytes since PCX-W is Big Endian and the
562 * IOMMU uses little endian for the pdir.
565 static void SBA_INLINE
566 sba_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba,
569 u64 pa; /* physical address */
570 register unsigned ci; /* coherent index */
575 asm("lci 0(%1), %0" : "=r" (ci) : "r" (vba));
576 pa |= (ci >> PAGE_SHIFT) & 0xff; /* move CI (8 bits) into lowest byte */
578 pa |= SBA_PDIR_VALID_BIT; /* set "valid" bit */
579 *pdir_ptr = cpu_to_le64(pa); /* swap and store into I/O Pdir */
582 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
583 * (bit #61, big endian), we have to flush and sync every time
584 * IO-PDIR is changed in Ike/Astro.
586 asm_io_fdc(pdir_ptr);
591 * sba_mark_invalid - invalidate one or more IO PDIR entries
592 * @ioc: IO MMU structure which owns the pdir we are interested in.
593 * @iova: IO Virtual Address mapped earlier
594 * @byte_cnt: number of bytes this mapping covers.
596 * Marking the IO PDIR entry(ies) as Invalid and invalidate
597 * corresponding IO TLB entry. The Ike PCOM (Purge Command Register)
598 * is to purge stale entries in the IO TLB when unmapping entries.
600 * The PCOM register supports purging of multiple pages, with a minium
601 * of 1 page and a maximum of 2GB. Hardware requires the address be
602 * aligned to the size of the range being purged. The size of the range
603 * must be a power of 2. The "Cool perf optimization" in the
604 * allocation routine helps keep that true.
606 static SBA_INLINE void
607 sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
609 u32 iovp = (u32) SBA_IOVP(ioc,iova);
610 u64 *pdir_ptr = &ioc->pdir_base[PDIR_INDEX(iovp)];
612 #ifdef ASSERT_PDIR_SANITY
613 /* Assert first pdir entry is set.
615 ** Even though this is a big-endian machine, the entries
616 ** in the iopdir are little endian. That's why we look at
617 ** the byte at +7 instead of at +0.
619 if (0x80 != (((u8 *) pdir_ptr)[7])) {
620 sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp));
624 if (byte_cnt > IOVP_SIZE)
627 unsigned long entries_per_cacheline = ioc_needs_fdc ?
628 L1_CACHE_ALIGN(((unsigned long) pdir_ptr))
629 - (unsigned long) pdir_ptr;
633 /* set "size" field for PCOM */
634 iovp |= get_order(byte_cnt) + PAGE_SHIFT;
637 /* clear I/O Pdir entry "valid" bit first */
638 ((u8 *) pdir_ptr)[7] = 0;
639 asm_io_fdc(pdir_ptr);
642 entries_per_cacheline = L1_CACHE_SHIFT - 3;
646 byte_cnt -= IOVP_SIZE;
647 } while (byte_cnt > IOVP_SIZE);
649 iovp |= IOVP_SHIFT; /* set "size" field for PCOM */
652 ** clear I/O PDIR entry "valid" bit.
653 ** We have to R/M/W the cacheline regardless how much of the
654 ** pdir entry that we clobber.
655 ** The rest of the entry would be useful for debugging if we
656 ** could dump core on HPMC.
658 ((u8 *) pdir_ptr)[7] = 0;
659 asm_io_fdc(pdir_ptr);
661 WRITE_REG( SBA_IOVA(ioc, iovp, 0, 0), ioc->ioc_hpa+IOC_PCOM);
665 * sba_dma_supported - PCI driver can query DMA support
666 * @dev: instance of PCI owned by the driver that's asking
667 * @mask: number of address bits this PCI device can handle
669 * See Documentation/core-api/dma-api-howto.rst
671 static int sba_dma_supported( struct device *dev, u64 mask)
676 printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n");
686 * check if mask is >= than the current max IO Virt Address
687 * The max IO Virt address will *always* < 30 bits.
689 return((int)(mask >= (ioc->ibase - 1 +
690 (ioc->pdir_size / sizeof(u64) * IOVP_SIZE) )));
695 * sba_map_single - map one buffer and return IOVA for DMA
696 * @dev: instance of PCI owned by the driver that's asking.
697 * @addr: driver buffer to map.
698 * @size: number of bytes to map in driver buffer.
699 * @direction: R/W or both.
701 * See Documentation/core-api/dma-api-howto.rst
704 sba_map_single(struct device *dev, void *addr, size_t size,
705 enum dma_data_direction direction)
716 return DMA_MAPPING_ERROR;
718 /* save offset bits */
719 offset = ((dma_addr_t) (long) addr) & ~IOVP_MASK;
721 /* round up to nearest IOVP_SIZE */
722 size = (size + offset + ~IOVP_MASK) & IOVP_MASK;
724 spin_lock_irqsave(&ioc->res_lock, flags);
725 #ifdef ASSERT_PDIR_SANITY
726 sba_check_pdir(ioc,"Check before sba_map_single()");
729 #ifdef SBA_COLLECT_STATS
730 ioc->msingle_calls++;
731 ioc->msingle_pages += size >> IOVP_SHIFT;
733 pide = sba_alloc_range(ioc, dev, size);
734 iovp = (dma_addr_t) pide << IOVP_SHIFT;
736 DBG_RUN("%s() 0x%p -> 0x%lx\n",
737 __func__, addr, (long) iovp | offset);
739 pdir_start = &(ioc->pdir_base[pide]);
742 sba_io_pdir_entry(pdir_start, KERNEL_SPACE, (unsigned long) addr, 0);
744 DBG_RUN(" pdir 0x%p %02x%02x%02x%02x%02x%02x%02x%02x\n",
746 (u8) (((u8 *) pdir_start)[7]),
747 (u8) (((u8 *) pdir_start)[6]),
748 (u8) (((u8 *) pdir_start)[5]),
749 (u8) (((u8 *) pdir_start)[4]),
750 (u8) (((u8 *) pdir_start)[3]),
751 (u8) (((u8 *) pdir_start)[2]),
752 (u8) (((u8 *) pdir_start)[1]),
753 (u8) (((u8 *) pdir_start)[0])
761 /* force FDC ops in io_pdir_entry() to be visible to IOMMU */
764 #ifdef ASSERT_PDIR_SANITY
765 sba_check_pdir(ioc,"Check after sba_map_single()");
767 spin_unlock_irqrestore(&ioc->res_lock, flags);
769 /* form complete address */
770 return SBA_IOVA(ioc, iovp, offset, DEFAULT_DMA_HINT_REG);
775 sba_map_page(struct device *dev, struct page *page, unsigned long offset,
776 size_t size, enum dma_data_direction direction,
779 return sba_map_single(dev, page_address(page) + offset, size,
785 * sba_unmap_page - unmap one IOVA and free resources
786 * @dev: instance of PCI owned by the driver that's asking.
787 * @iova: IOVA of driver buffer previously mapped.
788 * @size: number of bytes mapped in driver buffer.
789 * @direction: R/W or both.
791 * See Documentation/core-api/dma-api-howto.rst
794 sba_unmap_page(struct device *dev, dma_addr_t iova, size_t size,
795 enum dma_data_direction direction, unsigned long attrs)
798 #if DELAYED_RESOURCE_CNT > 0
799 struct sba_dma_pair *d;
804 DBG_RUN("%s() iovp 0x%lx/%x\n", __func__, (long) iova, size);
811 offset = iova & ~IOVP_MASK;
812 iova ^= offset; /* clear offset bits */
814 size = ALIGN(size, IOVP_SIZE);
816 spin_lock_irqsave(&ioc->res_lock, flags);
818 #ifdef SBA_COLLECT_STATS
819 ioc->usingle_calls++;
820 ioc->usingle_pages += size >> IOVP_SHIFT;
823 sba_mark_invalid(ioc, iova, size);
825 #if DELAYED_RESOURCE_CNT > 0
826 /* Delaying when we re-use a IO Pdir entry reduces the number
827 * of MMIO reads needed to flush writes to the PCOM register.
829 d = &(ioc->saved[ioc->saved_cnt]);
832 if (++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT) {
833 int cnt = ioc->saved_cnt;
835 sba_free_range(ioc, d->iova, d->size);
840 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */
842 #else /* DELAYED_RESOURCE_CNT == 0 */
843 sba_free_range(ioc, iova, size);
845 /* If fdc's were issued, force fdc's to be visible now */
848 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */
849 #endif /* DELAYED_RESOURCE_CNT == 0 */
851 spin_unlock_irqrestore(&ioc->res_lock, flags);
853 /* XXX REVISIT for 2.5 Linux - need syncdma for zero-copy support.
854 ** For Astro based systems this isn't a big deal WRT performance.
855 ** As long as 2.4 kernels copyin/copyout data from/to userspace,
856 ** we don't need the syncdma. The issue here is I/O MMU cachelines
857 ** are *not* coherent in all cases. May be hwrev dependent.
858 ** Need to investigate more.
859 asm volatile("syncdma");
865 * sba_alloc - allocate/map shared mem for DMA
866 * @hwdev: instance of PCI owned by the driver that's asking.
867 * @size: number of bytes mapped in driver buffer.
868 * @dma_handle: IOVA of new buffer.
870 * See Documentation/core-api/dma-api-howto.rst
872 static void *sba_alloc(struct device *hwdev, size_t size, dma_addr_t *dma_handle,
873 gfp_t gfp, unsigned long attrs)
878 /* only support PCI */
883 ret = (void *) __get_free_pages(gfp, get_order(size));
886 memset(ret, 0, size);
887 *dma_handle = sba_map_single(hwdev, ret, size, 0);
895 * sba_free - free/unmap shared mem for DMA
896 * @hwdev: instance of PCI owned by the driver that's asking.
897 * @size: number of bytes mapped in driver buffer.
898 * @vaddr: virtual address IOVA of "consistent" buffer.
899 * @dma_handler: IO virtual address of "consistent" buffer.
901 * See Documentation/core-api/dma-api-howto.rst
904 sba_free(struct device *hwdev, size_t size, void *vaddr,
905 dma_addr_t dma_handle, unsigned long attrs)
907 sba_unmap_page(hwdev, dma_handle, size, 0, 0);
908 free_pages((unsigned long) vaddr, get_order(size));
913 ** Since 0 is a valid pdir_base index value, can't use that
914 ** to determine if a value is valid or not. Use a flag to indicate
915 ** the SG list entry contains a valid pdir index.
917 #define PIDE_FLAG 0x80000000UL
919 #ifdef SBA_COLLECT_STATS
920 #define IOMMU_MAP_STATS
922 #include "iommu-helpers.h"
924 #ifdef DEBUG_LARGE_SG_ENTRIES
930 * sba_map_sg - map Scatter/Gather list
931 * @dev: instance of PCI owned by the driver that's asking.
932 * @sglist: array of buffer/length pairs
933 * @nents: number of entries in list
934 * @direction: R/W or both.
936 * See Documentation/core-api/dma-api-howto.rst
939 sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
940 enum dma_data_direction direction, unsigned long attrs)
943 int coalesced, filled = 0;
946 DBG_RUN_SG("%s() START %d entries\n", __func__, nents);
952 /* Fast path single entry scatterlists. */
954 sg_dma_address(sglist) = sba_map_single(dev, sg_virt(sglist),
955 sglist->length, direction);
956 sg_dma_len(sglist) = sglist->length;
960 spin_lock_irqsave(&ioc->res_lock, flags);
962 #ifdef ASSERT_PDIR_SANITY
963 if (sba_check_pdir(ioc,"Check before sba_map_sg()"))
965 sba_dump_sg(ioc, sglist, nents);
966 panic("Check before sba_map_sg()");
970 #ifdef SBA_COLLECT_STATS
975 ** First coalesce the chunks and allocate I/O pdir space
977 ** If this is one DMA stream, we can properly map using the
978 ** correct virtual address associated with each DMA page.
979 ** w/o this association, we wouldn't have coherent DMA!
980 ** Access to the virtual address is what forces a two pass algorithm.
982 coalesced = iommu_coalesce_chunks(ioc, dev, sglist, nents, sba_alloc_range);
985 ** Program the I/O Pdir
987 ** map the virtual addresses to the I/O Pdir
988 ** o dma_address will contain the pdir index
989 ** o dma_len will contain the number of bytes to map
990 ** o address contains the virtual address.
992 filled = iommu_fill_pdir(ioc, sglist, nents, 0, sba_io_pdir_entry);
994 /* force FDC ops in io_pdir_entry() to be visible to IOMMU */
997 #ifdef ASSERT_PDIR_SANITY
998 if (sba_check_pdir(ioc,"Check after sba_map_sg()"))
1000 sba_dump_sg(ioc, sglist, nents);
1001 panic("Check after sba_map_sg()\n");
1005 spin_unlock_irqrestore(&ioc->res_lock, flags);
1007 DBG_RUN_SG("%s() DONE %d mappings\n", __func__, filled);
1014 * sba_unmap_sg - unmap Scatter/Gather list
1015 * @dev: instance of PCI owned by the driver that's asking.
1016 * @sglist: array of buffer/length pairs
1017 * @nents: number of entries in list
1018 * @direction: R/W or both.
1020 * See Documentation/core-api/dma-api-howto.rst
1023 sba_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents,
1024 enum dma_data_direction direction, unsigned long attrs)
1027 #ifdef ASSERT_PDIR_SANITY
1028 unsigned long flags;
1031 DBG_RUN_SG("%s() START %d entries, %p,%x\n",
1032 __func__, nents, sg_virt(sglist), sglist->length);
1040 #ifdef SBA_COLLECT_STATS
1044 #ifdef ASSERT_PDIR_SANITY
1045 spin_lock_irqsave(&ioc->res_lock, flags);
1046 sba_check_pdir(ioc,"Check before sba_unmap_sg()");
1047 spin_unlock_irqrestore(&ioc->res_lock, flags);
1050 while (nents && sg_dma_len(sglist)) {
1052 sba_unmap_page(dev, sg_dma_address(sglist), sg_dma_len(sglist),
1054 #ifdef SBA_COLLECT_STATS
1055 ioc->usg_pages += ((sg_dma_address(sglist) & ~IOVP_MASK) + sg_dma_len(sglist) + IOVP_SIZE - 1) >> PAGE_SHIFT;
1056 ioc->usingle_calls--; /* kluge since call is unmap_sg() */
1062 DBG_RUN_SG("%s() DONE (nents %d)\n", __func__, nents);
1064 #ifdef ASSERT_PDIR_SANITY
1065 spin_lock_irqsave(&ioc->res_lock, flags);
1066 sba_check_pdir(ioc,"Check after sba_unmap_sg()");
1067 spin_unlock_irqrestore(&ioc->res_lock, flags);
1072 static const struct dma_map_ops sba_ops = {
1073 .dma_supported = sba_dma_supported,
1076 .map_page = sba_map_page,
1077 .unmap_page = sba_unmap_page,
1078 .map_sg = sba_map_sg,
1079 .unmap_sg = sba_unmap_sg,
1080 .get_sgtable = dma_common_get_sgtable,
1081 .alloc_pages = dma_common_alloc_pages,
1082 .free_pages = dma_common_free_pages,
1086 /**************************************************************************
1088 ** SBA PAT PDC support
1090 ** o call pdc_pat_cell_module()
1091 ** o store ranges in PCI "resource" structures
1093 **************************************************************************/
1096 sba_get_pat_resources(struct sba_device *sba_dev)
1100 ** TODO/REVISIT/FIXME: support for directed ranges requires calls to
1101 ** PAT PDC to program the SBA/LBA directed range registers...this
1102 ** burden may fall on the LBA code since it directly supports the
1103 ** PCI subsystem. It's not clear yet. - ggg
1105 PAT_MOD(mod)->mod_info.mod_pages = PAT_GET_MOD_PAGES(temp);
1107 PAT_MOD(mod)->mod_info.dvi = PAT_GET_DVI(temp);
1108 Tells where the dvi bits are located in the address.
1109 PAT_MOD(mod)->mod_info.ioc = PAT_GET_IOC(temp);
1115 /**************************************************************
1117 * Initialization and claim
1119 ***************************************************************/
1120 #define PIRANHA_ADDR_MASK 0x00160000UL /* bit 17,18,20 */
1121 #define PIRANHA_ADDR_VAL 0x00060000UL /* bit 17,18 on */
1123 sba_alloc_pdir(unsigned int pdir_size)
1125 unsigned long pdir_base;
1126 unsigned long pdir_order = get_order(pdir_size);
1128 pdir_base = __get_free_pages(GFP_KERNEL, pdir_order);
1129 if (NULL == (void *) pdir_base) {
1130 panic("%s() could not allocate I/O Page Table\n",
1134 /* If this is not PA8700 (PCX-W2)
1135 ** OR newer than ver 2.2
1136 ** OR in a system that doesn't need VINDEX bits from SBA,
1138 ** then we aren't exposed to the HW bug.
1140 if ( ((boot_cpu_data.pdc.cpuid >> 5) & 0x7f) != 0x13
1141 || (boot_cpu_data.pdc.versions > 0x202)
1142 || (boot_cpu_data.pdc.capabilities & 0x08L) )
1143 return (void *) pdir_base;
1146 * PA8700 (PCX-W2, aka piranha) silent data corruption fix
1148 * An interaction between PA8700 CPU (Ver 2.2 or older) and
1149 * Ike/Astro can cause silent data corruption. This is only
1150 * a problem if the I/O PDIR is located in memory such that
1151 * (little-endian) bits 17 and 18 are on and bit 20 is off.
1153 * Since the max IO Pdir size is 2MB, by cleverly allocating the
1154 * right physical address, we can either avoid (IOPDIR <= 1MB)
1155 * or minimize (2MB IO Pdir) the problem if we restrict the
1156 * IO Pdir to a maximum size of 2MB-128K (1902K).
1158 * Because we always allocate 2^N sized IO pdirs, either of the
1159 * "bad" regions will be the last 128K if at all. That's easy
1163 if (pdir_order <= (19-12)) {
1164 if (((virt_to_phys(pdir_base)+pdir_size-1) & PIRANHA_ADDR_MASK) == PIRANHA_ADDR_VAL) {
1165 /* allocate a new one on 512k alignment */
1166 unsigned long new_pdir = __get_free_pages(GFP_KERNEL, (19-12));
1167 /* release original */
1168 free_pages(pdir_base, pdir_order);
1170 pdir_base = new_pdir;
1172 /* release excess */
1173 while (pdir_order < (19-12)) {
1174 new_pdir += pdir_size;
1175 free_pages(new_pdir, pdir_order);
1183 ** Needs to be aligned on an "odd" 1MB boundary.
1185 unsigned long new_pdir = __get_free_pages(GFP_KERNEL, pdir_order+1); /* 2 or 4MB */
1187 /* release original */
1188 free_pages( pdir_base, pdir_order);
1190 /* release first 1MB */
1191 free_pages(new_pdir, 20-12);
1193 pdir_base = new_pdir + 1024*1024;
1195 if (pdir_order > (20-12)) {
1199 ** Flag tells init_bitmap() to mark bad 128k as used
1200 ** and to reduce the size by 128k.
1202 piranha_bad_128k = 1;
1204 new_pdir += 3*1024*1024;
1205 /* release last 1MB */
1206 free_pages(new_pdir, 20-12);
1208 /* release unusable 128KB */
1209 free_pages(new_pdir - 128*1024 , 17-12);
1211 pdir_size -= 128*1024;
1215 memset((void *) pdir_base, 0, pdir_size);
1216 return (void *) pdir_base;
1219 struct ibase_data_struct {
1224 static int setup_ibase_imask_callback(struct device *dev, void *data)
1226 /* lba_set_iregs() is in drivers/parisc/lba_pci.c */
1227 extern void lba_set_iregs(struct parisc_device *, u32, u32);
1228 struct parisc_device *lba = to_parisc_device(dev);
1229 struct ibase_data_struct *ibd = data;
1230 int rope_num = (lba->hpa.start >> 13) & 0xf;
1231 if (rope_num >> 3 == ibd->ioc_num)
1232 lba_set_iregs(lba, ibd->ioc->ibase, ibd->ioc->imask);
1236 /* setup Mercury or Elroy IBASE/IMASK registers. */
1238 setup_ibase_imask(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1240 struct ibase_data_struct ibase_data = {
1245 device_for_each_child(&sba->dev, &ibase_data,
1246 setup_ibase_imask_callback);
1249 #ifdef SBA_AGP_SUPPORT
1251 sba_ioc_find_quicksilver(struct device *dev, void *data)
1253 int *agp_found = data;
1254 struct parisc_device *lba = to_parisc_device(dev);
1256 if (IS_QUICKSILVER(lba))
1263 sba_ioc_init_pluto(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1265 u32 iova_space_mask;
1266 u32 iova_space_size;
1267 int iov_order, tcnfg;
1268 #ifdef SBA_AGP_SUPPORT
1272 ** Firmware programs the base and size of a "safe IOVA space"
1273 ** (one that doesn't overlap memory or LMMIO space) in the
1274 ** IBASE and IMASK registers.
1276 ioc->ibase = READ_REG(ioc->ioc_hpa + IOC_IBASE) & ~0x1fffffULL;
1277 iova_space_size = ~(READ_REG(ioc->ioc_hpa + IOC_IMASK) & 0xFFFFFFFFUL) + 1;
1279 if ((ioc->ibase < 0xfed00000UL) && ((ioc->ibase + iova_space_size) > 0xfee00000UL)) {
1280 printk("WARNING: IOV space overlaps local config and interrupt message, truncating\n");
1281 iova_space_size /= 2;
1285 ** iov_order is always based on a 1GB IOVA space since we want to
1286 ** turn on the other half for AGP GART.
1288 iov_order = get_order(iova_space_size >> (IOVP_SHIFT - PAGE_SHIFT));
1289 ioc->pdir_size = (iova_space_size / IOVP_SIZE) * sizeof(u64);
1291 DBG_INIT("%s() hpa 0x%p IOV %dMB (%d bits)\n",
1292 __func__, ioc->ioc_hpa, iova_space_size >> 20,
1293 iov_order + PAGE_SHIFT);
1295 ioc->pdir_base = (void *) __get_free_pages(GFP_KERNEL,
1296 get_order(ioc->pdir_size));
1297 if (!ioc->pdir_base)
1298 panic("Couldn't allocate I/O Page Table\n");
1300 memset(ioc->pdir_base, 0, ioc->pdir_size);
1302 DBG_INIT("%s() pdir %p size %x\n",
1303 __func__, ioc->pdir_base, ioc->pdir_size);
1305 #ifdef SBA_HINT_SUPPORT
1306 ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1307 ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1309 DBG_INIT(" hint_shift_pdir %x hint_mask_pdir %lx\n",
1310 ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1313 WARN_ON((((unsigned long) ioc->pdir_base) & PAGE_MASK) != (unsigned long) ioc->pdir_base);
1314 WRITE_REG(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1316 /* build IMASK for IOC and Elroy */
1317 iova_space_mask = 0xffffffff;
1318 iova_space_mask <<= (iov_order + PAGE_SHIFT);
1319 ioc->imask = iova_space_mask;
1321 ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1323 sba_dump_tlb(ioc->ioc_hpa);
1325 setup_ibase_imask(sba, ioc, ioc_num);
1327 WRITE_REG(ioc->imask, ioc->ioc_hpa + IOC_IMASK);
1331 ** Setting the upper bits makes checking for bypass addresses
1332 ** a little faster later on.
1334 ioc->imask |= 0xFFFFFFFF00000000UL;
1337 /* Set I/O PDIR Page size to system page size */
1338 switch (PAGE_SHIFT) {
1339 case 12: tcnfg = 0; break; /* 4K */
1340 case 13: tcnfg = 1; break; /* 8K */
1341 case 14: tcnfg = 2; break; /* 16K */
1342 case 16: tcnfg = 3; break; /* 64K */
1344 panic(__FILE__ "Unsupported system page size %d",
1348 WRITE_REG(tcnfg, ioc->ioc_hpa + IOC_TCNFG);
1351 ** Program the IOC's ibase and enable IOVA translation
1352 ** Bit zero == enable bit.
1354 WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa + IOC_IBASE);
1357 ** Clear I/O TLB of any possible entries.
1358 ** (Yes. This is a bit paranoid...but so what)
1360 WRITE_REG(ioc->ibase | 31, ioc->ioc_hpa + IOC_PCOM);
1362 #ifdef SBA_AGP_SUPPORT
1365 ** If an AGP device is present, only use half of the IOV space
1366 ** for PCI DMA. Unfortunately we can't know ahead of time
1367 ** whether GART support will actually be used, for now we
1368 ** can just key on any AGP device found in the system.
1369 ** We program the next pdir index after we stop w/ a key for
1370 ** the GART code to handshake on.
1372 device_for_each_child(&sba->dev, &agp_found, sba_ioc_find_quicksilver);
1374 if (agp_found && sba_reserve_agpgart) {
1375 printk(KERN_INFO "%s: reserving %dMb of IOVA space for agpgart\n",
1376 __func__, (iova_space_size/2) >> 20);
1377 ioc->pdir_size /= 2;
1378 ioc->pdir_base[PDIR_INDEX(iova_space_size/2)] = SBA_AGPGART_COOKIE;
1380 #endif /*SBA_AGP_SUPPORT*/
1384 sba_ioc_init(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1386 u32 iova_space_size, iova_space_mask;
1387 unsigned int pdir_size, iov_order, tcnfg;
1390 ** Determine IOVA Space size from memory size.
1392 ** Ideally, PCI drivers would register the maximum number
1393 ** of DMA they can have outstanding for each device they
1394 ** own. Next best thing would be to guess how much DMA
1395 ** can be outstanding based on PCI Class/sub-class. Both
1396 ** methods still require some "extra" to support PCI
1397 ** Hot-Plug/Removal of PCI cards. (aka PCI OLARD).
1399 ** While we have 32-bits "IOVA" space, top two 2 bits are used
1400 ** for DMA hints - ergo only 30 bits max.
1403 iova_space_size = (u32) (totalram_pages()/global_ioc_cnt);
1405 /* limit IOVA space size to 1MB-1GB */
1406 if (iova_space_size < (1 << (20 - PAGE_SHIFT))) {
1407 iova_space_size = 1 << (20 - PAGE_SHIFT);
1409 else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) {
1410 iova_space_size = 1 << (30 - PAGE_SHIFT);
1414 ** iova space must be log2() in size.
1415 ** thus, pdir/res_map will also be log2().
1416 ** PIRANHA BUG: Exception is when IO Pdir is 2MB (gets reduced)
1418 iov_order = get_order(iova_space_size << PAGE_SHIFT);
1420 /* iova_space_size is now bytes, not pages */
1421 iova_space_size = 1 << (iov_order + PAGE_SHIFT);
1423 ioc->pdir_size = pdir_size = (iova_space_size/IOVP_SIZE) * sizeof(u64);
1425 DBG_INIT("%s() hpa 0x%lx mem %ldMB IOV %dMB (%d bits)\n",
1428 (unsigned long) totalram_pages() >> (20 - PAGE_SHIFT),
1429 iova_space_size>>20,
1430 iov_order + PAGE_SHIFT);
1432 ioc->pdir_base = sba_alloc_pdir(pdir_size);
1434 DBG_INIT("%s() pdir %p size %x\n",
1435 __func__, ioc->pdir_base, pdir_size);
1437 #ifdef SBA_HINT_SUPPORT
1438 /* FIXME : DMA HINTs not used */
1439 ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1440 ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1442 DBG_INIT(" hint_shift_pdir %x hint_mask_pdir %lx\n",
1443 ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1446 WRITE_REG64(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1448 /* build IMASK for IOC and Elroy */
1449 iova_space_mask = 0xffffffff;
1450 iova_space_mask <<= (iov_order + PAGE_SHIFT);
1453 ** On C3000 w/512MB mem, HP-UX 10.20 reports:
1454 ** ibase=0, imask=0xFE000000, size=0x2000000.
1457 ioc->imask = iova_space_mask; /* save it */
1459 ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1462 DBG_INIT("%s() IOV base 0x%lx mask 0x%0lx\n",
1463 __func__, ioc->ibase, ioc->imask);
1466 ** FIXME: Hint registers are programmed with default hint
1467 ** values during boot, so hints should be sane even if we
1468 ** can't reprogram them the way drivers want.
1471 setup_ibase_imask(sba, ioc, ioc_num);
1474 ** Program the IOC's ibase and enable IOVA translation
1476 WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa+IOC_IBASE);
1477 WRITE_REG(ioc->imask, ioc->ioc_hpa+IOC_IMASK);
1479 /* Set I/O PDIR Page size to system page size */
1480 switch (PAGE_SHIFT) {
1481 case 12: tcnfg = 0; break; /* 4K */
1482 case 13: tcnfg = 1; break; /* 8K */
1483 case 14: tcnfg = 2; break; /* 16K */
1484 case 16: tcnfg = 3; break; /* 64K */
1486 panic(__FILE__ "Unsupported system page size %d",
1490 /* Set I/O PDIR Page size to PAGE_SIZE (4k/16k/...) */
1491 WRITE_REG(tcnfg, ioc->ioc_hpa+IOC_TCNFG);
1494 ** Clear I/O TLB of any possible entries.
1495 ** (Yes. This is a bit paranoid...but so what)
1497 WRITE_REG(0 | 31, ioc->ioc_hpa+IOC_PCOM);
1499 ioc->ibase = 0; /* used by SBA_IOVA and related macros */
1501 DBG_INIT("%s() DONE\n", __func__);
1506 /**************************************************************************
1508 ** SBA initialization code (HW and SW)
1510 ** o identify SBA chip itself
1511 ** o initialize SBA chip modes (HardFail)
1512 ** o initialize SBA chip modes (HardFail)
1513 ** o FIXME: initialize DMA hints for reasonable defaults
1515 **************************************************************************/
1517 static void __iomem *ioc_remap(struct sba_device *sba_dev, unsigned int offset)
1519 return ioremap(sba_dev->dev->hpa.start + offset, SBA_FUNC_SIZE);
1522 static void sba_hw_init(struct sba_device *sba_dev)
1528 if (!is_pdc_pat()) {
1529 /* Shutdown the USB controller on Astro-based workstations.
1530 ** Once we reprogram the IOMMU, the next DMA performed by
1531 ** USB will HPMC the box. USB is only enabled if a
1532 ** keyboard is present and found.
1534 ** With serial console, j6k v5.0 firmware says:
1535 ** mem_kbd hpa 0xfee003f8 sba 0x0 pad 0x0 cl_class 0x7
1537 ** FIXME: Using GFX+USB console at power up but direct
1538 ** linux to serial console is still broken.
1539 ** USB could generate DMA so we must reset USB.
1540 ** The proper sequence would be:
1541 ** o block console output
1542 ** o reset USB device
1543 ** o reprogram serial port
1544 ** o unblock console output
1546 if (PAGE0->mem_kbd.cl_class == CL_KEYBD) {
1547 pdc_io_reset_devices();
1554 printk("sba_hw_init(): mem_boot 0x%x 0x%x 0x%x 0x%x\n", PAGE0->mem_boot.hpa,
1555 PAGE0->mem_boot.spa, PAGE0->mem_boot.pad, PAGE0->mem_boot.cl_class);
1558 ** Need to deal with DMA from LAN.
1559 ** Maybe use page zero boot device as a handle to talk
1560 ** to PDC about which device to shutdown.
1562 ** Netbooting, j6k v5.0 firmware says:
1563 ** mem_boot hpa 0xf4008000 sba 0x0 pad 0x0 cl_class 0x1002
1564 ** ARGH! invalid class.
1566 if ((PAGE0->mem_boot.cl_class != CL_RANDOM)
1567 && (PAGE0->mem_boot.cl_class != CL_SEQU)) {
1572 if (!IS_PLUTO(sba_dev->dev)) {
1573 ioc_ctl = READ_REG(sba_dev->sba_hpa+IOC_CTRL);
1574 DBG_INIT("%s() hpa 0x%lx ioc_ctl 0x%Lx ->",
1575 __func__, sba_dev->sba_hpa, ioc_ctl);
1576 ioc_ctl &= ~(IOC_CTRL_RM | IOC_CTRL_NC | IOC_CTRL_CE);
1577 ioc_ctl |= IOC_CTRL_DD | IOC_CTRL_D4 | IOC_CTRL_TC;
1578 /* j6700 v1.6 firmware sets 0x294f */
1579 /* A500 firmware sets 0x4d */
1581 WRITE_REG(ioc_ctl, sba_dev->sba_hpa+IOC_CTRL);
1583 #ifdef DEBUG_SBA_INIT
1584 ioc_ctl = READ_REG64(sba_dev->sba_hpa+IOC_CTRL);
1585 DBG_INIT(" 0x%Lx\n", ioc_ctl);
1589 if (IS_ASTRO(sba_dev->dev)) {
1591 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, ASTRO_IOC_OFFSET);
1594 sba_dev->chip_resv.name = "Astro Intr Ack";
1595 sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfef00000UL;
1596 sba_dev->chip_resv.end = PCI_F_EXTEND | (0xff000000UL - 1) ;
1597 err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1600 } else if (IS_PLUTO(sba_dev->dev)) {
1603 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, PLUTO_IOC_OFFSET);
1606 sba_dev->chip_resv.name = "Pluto Intr/PIOP/VGA";
1607 sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfee00000UL;
1608 sba_dev->chip_resv.end = PCI_F_EXTEND | (0xff200000UL - 1);
1609 err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1612 sba_dev->iommu_resv.name = "IOVA Space";
1613 sba_dev->iommu_resv.start = 0x40000000UL;
1614 sba_dev->iommu_resv.end = 0x50000000UL - 1;
1615 err = request_resource(&iomem_resource, &(sba_dev->iommu_resv));
1619 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(0));
1620 sba_dev->ioc[1].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(1));
1623 /* TODO - LOOKUP Ike/Stretch chipset mem map */
1625 /* XXX: What about Reo Grande? */
1627 sba_dev->num_ioc = num_ioc;
1628 for (i = 0; i < num_ioc; i++) {
1629 void __iomem *ioc_hpa = sba_dev->ioc[i].ioc_hpa;
1632 for (j=0; j < sizeof(u64) * ROPES_PER_IOC; j+=sizeof(u64)) {
1635 * Clear ROPE(N)_CONFIG AO bit.
1636 * Disables "NT Ordering" (~= !"Relaxed Ordering")
1637 * Overrides bit 1 in DMA Hint Sets.
1638 * Improves netperf UDP_STREAM by ~10% for bcm5701.
1640 if (IS_PLUTO(sba_dev->dev)) {
1641 void __iomem *rope_cfg;
1642 unsigned long cfg_val;
1644 rope_cfg = ioc_hpa + IOC_ROPE0_CFG + j;
1645 cfg_val = READ_REG(rope_cfg);
1646 cfg_val &= ~IOC_ROPE_AO;
1647 WRITE_REG(cfg_val, rope_cfg);
1651 ** Make sure the box crashes on rope errors.
1653 WRITE_REG(HF_ENABLE, ioc_hpa + ROPE0_CTL + j);
1656 /* flush out the last writes */
1657 READ_REG(sba_dev->ioc[i].ioc_hpa + ROPE7_CTL);
1659 DBG_INIT(" ioc[%d] ROPE_CFG 0x%Lx ROPE_DBG 0x%Lx\n",
1661 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x40),
1662 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x50)
1664 DBG_INIT(" STATUS_CONTROL 0x%Lx FLUSH_CTRL 0x%Lx\n",
1665 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x108),
1666 READ_REG(sba_dev->ioc[i].ioc_hpa + 0x400)
1669 if (IS_PLUTO(sba_dev->dev)) {
1670 sba_ioc_init_pluto(sba_dev->dev, &(sba_dev->ioc[i]), i);
1672 sba_ioc_init(sba_dev->dev, &(sba_dev->ioc[i]), i);
1678 sba_common_init(struct sba_device *sba_dev)
1682 /* add this one to the head of the list (order doesn't matter)
1683 ** This will be useful for debugging - especially if we get coredumps
1685 sba_dev->next = sba_list;
1688 for(i=0; i< sba_dev->num_ioc; i++) {
1690 #ifdef DEBUG_DMB_TRAP
1691 extern void iterate_pages(unsigned long , unsigned long ,
1692 void (*)(pte_t * , unsigned long),
1694 void set_data_memory_break(pte_t * , unsigned long);
1696 /* resource map size dictated by pdir_size */
1697 res_size = sba_dev->ioc[i].pdir_size/sizeof(u64); /* entries */
1699 /* Second part of PIRANHA BUG */
1700 if (piranha_bad_128k) {
1701 res_size -= (128*1024)/sizeof(u64);
1704 res_size >>= 3; /* convert bit count to byte count */
1705 DBG_INIT("%s() res_size 0x%x\n",
1706 __func__, res_size);
1708 sba_dev->ioc[i].res_size = res_size;
1709 sba_dev->ioc[i].res_map = (char *) __get_free_pages(GFP_KERNEL, get_order(res_size));
1711 #ifdef DEBUG_DMB_TRAP
1712 iterate_pages( sba_dev->ioc[i].res_map, res_size,
1713 set_data_memory_break, 0);
1716 if (NULL == sba_dev->ioc[i].res_map)
1718 panic("%s:%s() could not allocate resource map\n",
1719 __FILE__, __func__ );
1722 memset(sba_dev->ioc[i].res_map, 0, res_size);
1723 /* next available IOVP - circular search */
1724 sba_dev->ioc[i].res_hint = (unsigned long *)
1725 &(sba_dev->ioc[i].res_map[L1_CACHE_BYTES]);
1727 #ifdef ASSERT_PDIR_SANITY
1728 /* Mark first bit busy - ie no IOVA 0 */
1729 sba_dev->ioc[i].res_map[0] = 0x80;
1730 sba_dev->ioc[i].pdir_base[0] = 0xeeffc0addbba0080ULL;
1733 /* Third (and last) part of PIRANHA BUG */
1734 if (piranha_bad_128k) {
1735 /* region from +1408K to +1536 is un-usable. */
1737 int idx_start = (1408*1024/sizeof(u64)) >> 3;
1738 int idx_end = (1536*1024/sizeof(u64)) >> 3;
1739 long *p_start = (long *) &(sba_dev->ioc[i].res_map[idx_start]);
1740 long *p_end = (long *) &(sba_dev->ioc[i].res_map[idx_end]);
1742 /* mark that part of the io pdir busy */
1743 while (p_start < p_end)
1748 #ifdef DEBUG_DMB_TRAP
1749 iterate_pages( sba_dev->ioc[i].res_map, res_size,
1750 set_data_memory_break, 0);
1751 iterate_pages( sba_dev->ioc[i].pdir_base, sba_dev->ioc[i].pdir_size,
1752 set_data_memory_break, 0);
1755 DBG_INIT("%s() %d res_map %x %p\n",
1756 __func__, i, res_size, sba_dev->ioc[i].res_map);
1759 spin_lock_init(&sba_dev->sba_lock);
1760 ioc_needs_fdc = boot_cpu_data.pdc.capabilities & PDC_MODEL_IOPDIR_FDC;
1762 #ifdef DEBUG_SBA_INIT
1764 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
1765 * (bit #61, big endian), we have to flush and sync every time
1766 * IO-PDIR is changed in Ike/Astro.
1768 if (ioc_needs_fdc) {
1769 printk(KERN_INFO MODULE_NAME " FDC/SYNC required.\n");
1771 printk(KERN_INFO MODULE_NAME " IOC has cache coherent PDIR.\n");
1776 #ifdef CONFIG_PROC_FS
1777 static int sba_proc_info(struct seq_file *m, void *p)
1779 struct sba_device *sba_dev = sba_list;
1780 struct ioc *ioc = &sba_dev->ioc[0]; /* FIXME: Multi-IOC support! */
1781 int total_pages = (int) (ioc->res_size << 3); /* 8 bits per byte */
1782 #ifdef SBA_COLLECT_STATS
1783 unsigned long avg = 0, min, max;
1787 seq_printf(m, "%s rev %d.%d\n",
1789 (sba_dev->hw_rev & 0x7) + 1,
1790 (sba_dev->hw_rev & 0x18) >> 3);
1791 seq_printf(m, "IO PDIR size : %d bytes (%d entries)\n",
1792 (int)((ioc->res_size << 3) * sizeof(u64)), /* 8 bits/byte */
1795 seq_printf(m, "Resource bitmap : %d bytes (%d pages)\n",
1796 ioc->res_size, ioc->res_size << 3); /* 8 bits per byte */
1798 seq_printf(m, "LMMIO_BASE/MASK/ROUTE %08x %08x %08x\n",
1799 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_BASE),
1800 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_MASK),
1801 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_ROUTE));
1804 seq_printf(m, "DIR%d_BASE/MASK/ROUTE %08x %08x %08x\n",
1806 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_BASE + i*0x18),
1807 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_MASK + i*0x18),
1808 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_ROUTE + i*0x18));
1810 #ifdef SBA_COLLECT_STATS
1811 seq_printf(m, "IO PDIR entries : %ld free %ld used (%d%%)\n",
1812 total_pages - ioc->used_pages, ioc->used_pages,
1813 (int)(ioc->used_pages * 100 / total_pages));
1815 min = max = ioc->avg_search[0];
1816 for (i = 0; i < SBA_SEARCH_SAMPLE; i++) {
1817 avg += ioc->avg_search[i];
1818 if (ioc->avg_search[i] > max) max = ioc->avg_search[i];
1819 if (ioc->avg_search[i] < min) min = ioc->avg_search[i];
1821 avg /= SBA_SEARCH_SAMPLE;
1822 seq_printf(m, " Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n",
1825 seq_printf(m, "pci_map_single(): %12ld calls %12ld pages (avg %d/1000)\n",
1826 ioc->msingle_calls, ioc->msingle_pages,
1827 (int)((ioc->msingle_pages * 1000)/ioc->msingle_calls));
1829 /* KLUGE - unmap_sg calls unmap_single for each mapped page */
1830 min = ioc->usingle_calls;
1831 max = ioc->usingle_pages - ioc->usg_pages;
1832 seq_printf(m, "pci_unmap_single: %12ld calls %12ld pages (avg %d/1000)\n",
1833 min, max, (int)((max * 1000)/min));
1835 seq_printf(m, "pci_map_sg() : %12ld calls %12ld pages (avg %d/1000)\n",
1836 ioc->msg_calls, ioc->msg_pages,
1837 (int)((ioc->msg_pages * 1000)/ioc->msg_calls));
1839 seq_printf(m, "pci_unmap_sg() : %12ld calls %12ld pages (avg %d/1000)\n",
1840 ioc->usg_calls, ioc->usg_pages,
1841 (int)((ioc->usg_pages * 1000)/ioc->usg_calls));
1848 sba_proc_bitmap_info(struct seq_file *m, void *p)
1850 struct sba_device *sba_dev = sba_list;
1851 struct ioc *ioc = &sba_dev->ioc[0]; /* FIXME: Multi-IOC support! */
1853 seq_hex_dump(m, " ", DUMP_PREFIX_NONE, 32, 4, ioc->res_map,
1854 ioc->res_size, false);
1859 #endif /* CONFIG_PROC_FS */
1861 static const struct parisc_device_id sba_tbl[] __initconst = {
1862 { HPHW_IOA, HVERSION_REV_ANY_ID, ASTRO_RUNWAY_PORT, 0xb },
1863 { HPHW_BCPORT, HVERSION_REV_ANY_ID, IKE_MERCED_PORT, 0xc },
1864 { HPHW_BCPORT, HVERSION_REV_ANY_ID, REO_MERCED_PORT, 0xc },
1865 { HPHW_BCPORT, HVERSION_REV_ANY_ID, REOG_MERCED_PORT, 0xc },
1866 { HPHW_IOA, HVERSION_REV_ANY_ID, PLUTO_MCKINLEY_PORT, 0xc },
1870 static int sba_driver_callback(struct parisc_device *);
1872 static struct parisc_driver sba_driver __refdata = {
1873 .name = MODULE_NAME,
1874 .id_table = sba_tbl,
1875 .probe = sba_driver_callback,
1879 ** Determine if sba should claim this chip (return 0) or not (return 1).
1880 ** If so, initialize the chip and tell other partners in crime they
1883 static int __init sba_driver_callback(struct parisc_device *dev)
1885 struct sba_device *sba_dev;
1889 void __iomem *sba_addr = ioremap(dev->hpa.start, SBA_FUNC_SIZE);
1890 #ifdef CONFIG_PROC_FS
1891 struct proc_dir_entry *root;
1894 sba_dump_ranges(sba_addr);
1896 /* Read HW Rev First */
1897 func_class = READ_REG(sba_addr + SBA_FCLASS);
1899 if (IS_ASTRO(dev)) {
1900 unsigned long fclass;
1901 static char astro_rev[]="Astro ?.?";
1903 /* Astro is broken...Read HW Rev First */
1904 fclass = READ_REG(sba_addr);
1906 astro_rev[6] = '1' + (char) (fclass & 0x7);
1907 astro_rev[8] = '0' + (char) ((fclass & 0x18) >> 3);
1908 version = astro_rev;
1910 } else if (IS_IKE(dev)) {
1911 static char ike_rev[] = "Ike rev ?";
1912 ike_rev[8] = '0' + (char) (func_class & 0xff);
1914 } else if (IS_PLUTO(dev)) {
1915 static char pluto_rev[]="Pluto ?.?";
1916 pluto_rev[6] = '0' + (char) ((func_class & 0xf0) >> 4);
1917 pluto_rev[8] = '0' + (char) (func_class & 0x0f);
1918 version = pluto_rev;
1920 static char reo_rev[] = "REO rev ?";
1921 reo_rev[8] = '0' + (char) (func_class & 0xff);
1925 if (!global_ioc_cnt) {
1926 global_ioc_cnt = count_parisc_driver(&sba_driver);
1928 /* Astro and Pluto have one IOC per SBA */
1929 if ((!IS_ASTRO(dev)) || (!IS_PLUTO(dev)))
1930 global_ioc_cnt *= 2;
1933 printk(KERN_INFO "%s found %s at 0x%llx\n",
1934 MODULE_NAME, version, (unsigned long long)dev->hpa.start);
1936 sba_dev = kzalloc(sizeof(struct sba_device), GFP_KERNEL);
1938 printk(KERN_ERR MODULE_NAME " - couldn't alloc sba_device\n");
1942 parisc_set_drvdata(dev, sba_dev);
1944 for(i=0; i<MAX_IOC; i++)
1945 spin_lock_init(&(sba_dev->ioc[i].res_lock));
1948 sba_dev->hw_rev = func_class;
1949 sba_dev->name = dev->name;
1950 sba_dev->sba_hpa = sba_addr;
1952 sba_get_pat_resources(sba_dev);
1953 sba_hw_init(sba_dev);
1954 sba_common_init(sba_dev);
1956 hppa_dma_ops = &sba_ops;
1958 #ifdef CONFIG_PROC_FS
1959 switch (dev->id.hversion) {
1960 case PLUTO_MCKINLEY_PORT:
1961 root = proc_mckinley_root;
1963 case ASTRO_RUNWAY_PORT:
1964 case IKE_MERCED_PORT:
1966 root = proc_runway_root;
1970 proc_create_single("sba_iommu", 0, root, sba_proc_info);
1971 proc_create_single("sba_iommu-bitmap", 0, root, sba_proc_bitmap_info);
1977 ** One time initialization to let the world know the SBA was found.
1978 ** This is the only routine which is NOT static.
1979 ** Must be called exactly once before pci_init().
1981 void __init sba_init(void)
1983 register_parisc_driver(&sba_driver);
1988 * sba_get_iommu - Assign the iommu pointer for the pci bus controller.
1989 * @dev: The parisc device.
1991 * Returns the appropriate IOMMU data for the given parisc PCI controller.
1992 * This is cached and used later for PCI DMA Mapping.
1994 void * sba_get_iommu(struct parisc_device *pci_hba)
1996 struct parisc_device *sba_dev = parisc_parent(pci_hba);
1997 struct sba_device *sba = dev_get_drvdata(&sba_dev->dev);
1998 char t = sba_dev->id.hw_type;
1999 int iocnum = (pci_hba->hw_path >> 3); /* rope # */
2001 WARN_ON((t != HPHW_IOA) && (t != HPHW_BCPORT));
2003 return &(sba->ioc[iocnum]);
2008 * sba_directed_lmmio - return first directed LMMIO range routed to rope
2009 * @pa_dev: The parisc device.
2010 * @r: resource PCI host controller wants start/end fields assigned.
2012 * For the given parisc PCI controller, determine if any direct ranges
2013 * are routed down the corresponding rope.
2015 void sba_directed_lmmio(struct parisc_device *pci_hba, struct resource *r)
2017 struct parisc_device *sba_dev = parisc_parent(pci_hba);
2018 struct sba_device *sba = dev_get_drvdata(&sba_dev->dev);
2019 char t = sba_dev->id.hw_type;
2021 int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1)); /* rope # */
2023 BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2025 r->start = r->end = 0;
2027 /* Astro has 4 directed ranges. Not sure about Ike/Pluto/et al */
2028 for (i=0; i<4; i++) {
2030 void __iomem *reg = sba->sba_hpa + i*0x18;
2032 base = READ_REG32(reg + LMMIO_DIRECT0_BASE);
2033 if ((base & 1) == 0)
2034 continue; /* not enabled */
2036 size = READ_REG32(reg + LMMIO_DIRECT0_ROUTE);
2038 if ((size & (ROPES_PER_IOC-1)) != rope)
2039 continue; /* directed down different rope */
2041 r->start = (base & ~1UL) | PCI_F_EXTEND;
2042 size = ~ READ_REG32(reg + LMMIO_DIRECT0_MASK);
2043 r->end = r->start + size;
2044 r->flags = IORESOURCE_MEM;
2050 * sba_distributed_lmmio - return portion of distributed LMMIO range
2051 * @pa_dev: The parisc device.
2052 * @r: resource PCI host controller wants start/end fields assigned.
2054 * For the given parisc PCI controller, return portion of distributed LMMIO
2055 * range. The distributed LMMIO is always present and it's just a question
2056 * of the base address and size of the range.
2058 void sba_distributed_lmmio(struct parisc_device *pci_hba, struct resource *r )
2060 struct parisc_device *sba_dev = parisc_parent(pci_hba);
2061 struct sba_device *sba = dev_get_drvdata(&sba_dev->dev);
2062 char t = sba_dev->id.hw_type;
2064 int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1)); /* rope # */
2066 BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2068 r->start = r->end = 0;
2070 base = READ_REG32(sba->sba_hpa + LMMIO_DIST_BASE);
2071 if ((base & 1) == 0) {
2072 BUG(); /* Gah! Distr Range wasn't enabled! */
2076 r->start = (base & ~1UL) | PCI_F_EXTEND;
2078 size = (~READ_REG32(sba->sba_hpa + LMMIO_DIST_MASK)) / ROPES_PER_IOC;
2079 r->start += rope * (size + 1); /* adjust base for this rope */
2080 r->end = r->start + size;
2081 r->flags = IORESOURCE_MEM;