2 ** IA64 System Bus Adapter (SBA) I/O MMU manager
4 ** (c) Copyright 2002-2005 Alex Williamson
5 ** (c) Copyright 2002-2003 Grant Grundler
6 ** (c) Copyright 2002-2005 Hewlett-Packard Company
8 ** Portions (c) 2000 Grant Grundler (from parisc I/O MMU code)
9 ** Portions (c) 1999 Dave S. Miller (from sparc64 I/O MMU code)
11 ** This program is free software; you can redistribute it and/or modify
12 ** it under the terms of the GNU General Public License as published by
13 ** the Free Software Foundation; either version 2 of the License, or
14 ** (at your option) any later version.
17 ** This module initializes the IOC (I/O Controller) found on HP
18 ** McKinley machines and their successors.
22 #include <linux/types.h>
23 #include <linux/kernel.h>
24 #include <linux/module.h>
25 #include <linux/spinlock.h>
26 #include <linux/slab.h>
27 #include <linux/init.h>
29 #include <linux/string.h>
30 #include <linux/pci.h>
31 #include <linux/proc_fs.h>
32 #include <linux/seq_file.h>
33 #include <linux/acpi.h>
34 #include <linux/efi.h>
35 #include <linux/nodemask.h>
36 #include <linux/bitops.h> /* hweight64() */
37 #include <linux/crash_dump.h>
38 #include <linux/iommu-helper.h>
40 #include <asm/delay.h> /* ia64_get_itc() */
42 #include <asm/page.h> /* PAGE_OFFSET */
44 #include <asm/system.h> /* wmb() */
46 #include <asm/acpi-ext.h>
48 extern int swiotlb_late_init_with_default_size (size_t size);
53 ** Enabling timing search of the pdir resource map. Output in /proc.
54 ** Disabled by default to optimize performance.
56 #undef PDIR_SEARCH_TIMING
59 ** This option allows cards capable of 64bit DMA to bypass the IOMMU. If
60 ** not defined, all DMA will be 32bit and go through the TLB.
61 ** There's potentially a conflict in the bio merge code with us
62 ** advertising an iommu, but then bypassing it. Since I/O MMU bypassing
63 ** appears to give more performance than bio-level virtual merging, we'll
64 ** do the former for now. NOTE: BYPASS_SG also needs to be undef'd to
65 ** completely restrict DMA to the IOMMU.
67 #define ALLOW_IOV_BYPASS
70 ** This option specifically allows/disallows bypassing scatterlists with
71 ** multiple entries. Coalescing these entries can allow better DMA streaming
72 ** and in some cases shows better performance than entirely bypassing the
73 ** IOMMU. Performance increase on the order of 1-2% sequential output/input
74 ** using bonnie++ on a RAID0 MD device (sym2 & mpt).
76 #undef ALLOW_IOV_BYPASS_SG
79 ** If a device prefetches beyond the end of a valid pdir entry, it will cause
80 ** a hard failure, ie. MCA. Version 3.0 and later of the zx1 LBA should
81 ** disconnect on 4k boundaries and prevent such issues. If the device is
82 ** particularly aggressive, this option will keep the entire pdir valid such
83 ** that prefetching will hit a valid address. This could severely impact
84 ** error containment, and is therefore off by default. The page that is
85 ** used for spill-over is poisoned, so that should help debugging somewhat.
87 #undef FULL_VALID_PDIR
89 #define ENABLE_MARK_CLEAN
92 ** The number of debug flags is a clue - this code is fragile. NOTE: since
93 ** tightening the use of res_lock the resource bitmap and actual pdir are no
94 ** longer guaranteed to stay in sync. The sanity checking code isn't going to
99 #undef DEBUG_SBA_RUN_SG
100 #undef DEBUG_SBA_RESOURCE
101 #undef ASSERT_PDIR_SANITY
102 #undef DEBUG_LARGE_SG_ENTRIES
105 #if defined(FULL_VALID_PDIR) && defined(ASSERT_PDIR_SANITY)
106 #error FULL_VALID_PDIR and ASSERT_PDIR_SANITY are mutually exclusive
109 #define SBA_INLINE __inline__
110 /* #define SBA_INLINE */
112 #ifdef DEBUG_SBA_INIT
113 #define DBG_INIT(x...) printk(x)
115 #define DBG_INIT(x...)
119 #define DBG_RUN(x...) printk(x)
121 #define DBG_RUN(x...)
124 #ifdef DEBUG_SBA_RUN_SG
125 #define DBG_RUN_SG(x...) printk(x)
127 #define DBG_RUN_SG(x...)
131 #ifdef DEBUG_SBA_RESOURCE
132 #define DBG_RES(x...) printk(x)
134 #define DBG_RES(x...)
138 #define DBG_BYPASS(x...) printk(x)
140 #define DBG_BYPASS(x...)
143 #ifdef ASSERT_PDIR_SANITY
144 #define ASSERT(expr) \
146 printk( "\n" __FILE__ ":%d: Assertion " #expr " failed!\n",__LINE__); \
154 ** The number of pdir entries to "free" before issuing
155 ** a read to PCOM register to flush out PCOM writes.
156 ** Interacts with allocation granularity (ie 4 or 8 entries
157 ** allocated and free'd/purged at a time might make this
158 ** less interesting).
160 #define DELAYED_RESOURCE_CNT 64
162 #define PCI_DEVICE_ID_HP_SX2000_IOC 0x12ec
164 #define ZX1_IOC_ID ((PCI_DEVICE_ID_HP_ZX1_IOC << 16) | PCI_VENDOR_ID_HP)
165 #define ZX2_IOC_ID ((PCI_DEVICE_ID_HP_ZX2_IOC << 16) | PCI_VENDOR_ID_HP)
166 #define REO_IOC_ID ((PCI_DEVICE_ID_HP_REO_IOC << 16) | PCI_VENDOR_ID_HP)
167 #define SX1000_IOC_ID ((PCI_DEVICE_ID_HP_SX1000_IOC << 16) | PCI_VENDOR_ID_HP)
168 #define SX2000_IOC_ID ((PCI_DEVICE_ID_HP_SX2000_IOC << 16) | PCI_VENDOR_ID_HP)
170 #define ZX1_IOC_OFFSET 0x1000 /* ACPI reports SBA, we want IOC */
172 #define IOC_FUNC_ID 0x000
173 #define IOC_FCLASS 0x008 /* function class, bist, header, rev... */
174 #define IOC_IBASE 0x300 /* IO TLB */
175 #define IOC_IMASK 0x308
176 #define IOC_PCOM 0x310
177 #define IOC_TCNFG 0x318
178 #define IOC_PDIR_BASE 0x320
180 #define IOC_ROPE0_CFG 0x500
181 #define IOC_ROPE_AO 0x10 /* Allow "Relaxed Ordering" */
184 /* AGP GART driver looks for this */
185 #define ZX1_SBA_IOMMU_COOKIE 0x0000badbadc0ffeeUL
188 ** The zx1 IOC supports 4/8/16/64KB page sizes (see TCNFG register)
190 ** Some IOCs (sx1000) can run at the above pages sizes, but are
191 ** really only supported using the IOC at a 4k page size.
193 ** iovp_size could only be greater than PAGE_SIZE if we are
194 ** confident the drivers really only touch the next physical
195 ** page iff that driver instance owns it.
197 static unsigned long iovp_size;
198 static unsigned long iovp_shift;
199 static unsigned long iovp_mask;
202 void __iomem *ioc_hpa; /* I/O MMU base address */
203 char *res_map; /* resource map, bit == pdir entry */
204 u64 *pdir_base; /* physical base address */
205 unsigned long ibase; /* pdir IOV Space base */
206 unsigned long imask; /* pdir IOV Space mask */
208 unsigned long *res_hint; /* next avail IOVP - circular search */
209 unsigned long dma_mask;
210 spinlock_t res_lock; /* protects the resource bitmap, but must be held when */
211 /* clearing pdir to prevent races with allocations. */
212 unsigned int res_bitshift; /* from the RIGHT! */
213 unsigned int res_size; /* size of resource map in bytes */
215 unsigned int node; /* node where this IOC lives */
217 #if DELAYED_RESOURCE_CNT > 0
218 spinlock_t saved_lock; /* may want to try to get this on a separate cacheline */
219 /* than res_lock for bigger systems. */
221 struct sba_dma_pair {
224 } saved[DELAYED_RESOURCE_CNT];
227 #ifdef PDIR_SEARCH_TIMING
228 #define SBA_SEARCH_SAMPLE 0x100
229 unsigned long avg_search[SBA_SEARCH_SAMPLE];
230 unsigned long avg_idx; /* current index into avg_search */
233 /* Stuff we don't need in performance path */
234 struct ioc *next; /* list of IOC's in system */
235 acpi_handle handle; /* for multiple IOC's */
237 unsigned int func_id;
238 unsigned int rev; /* HW revision of chip */
240 unsigned int pdir_size; /* in bytes, determined by IOV Space size */
241 struct pci_dev *sac_only_dev;
244 static struct ioc *ioc_list;
245 static int reserve_sba_gart = 1;
247 static SBA_INLINE void sba_mark_invalid(struct ioc *, dma_addr_t, size_t);
248 static SBA_INLINE void sba_free_range(struct ioc *, dma_addr_t, size_t);
250 #define sba_sg_address(sg) sg_virt((sg))
252 #ifdef FULL_VALID_PDIR
253 static u64 prefetch_spill_page;
257 # define GET_IOC(dev) (((dev)->bus == &pci_bus_type) \
258 ? ((struct ioc *) PCI_CONTROLLER(to_pci_dev(dev))->iommu) : NULL)
260 # define GET_IOC(dev) NULL
264 ** DMA_CHUNK_SIZE is used by the SCSI mid-layer to break up
265 ** (or rather not merge) DMAs into manageable chunks.
266 ** On parisc, this is more of the software/tuning constraint
267 ** rather than the HW. I/O MMU allocation algorithms can be
268 ** faster with smaller sizes (to some degree).
270 #define DMA_CHUNK_SIZE (BITS_PER_LONG*iovp_size)
272 #define ROUNDUP(x,y) ((x + ((y)-1)) & ~((y)-1))
274 /************************************
275 ** SBA register read and write support
277 ** BE WARNED: register writes are posted.
278 ** (ie follow writes which must reach HW with a read)
281 #define READ_REG(addr) __raw_readq(addr)
282 #define WRITE_REG(val, addr) __raw_writeq(val, addr)
284 #ifdef DEBUG_SBA_INIT
287 * sba_dump_tlb - debugging only - print IOMMU operating parameters
288 * @hpa: base address of the IOMMU
290 * Print the size/location of the IO MMU PDIR.
293 sba_dump_tlb(char *hpa)
295 DBG_INIT("IO TLB at 0x%p\n", (void *)hpa);
296 DBG_INIT("IOC_IBASE : %016lx\n", READ_REG(hpa+IOC_IBASE));
297 DBG_INIT("IOC_IMASK : %016lx\n", READ_REG(hpa+IOC_IMASK));
298 DBG_INIT("IOC_TCNFG : %016lx\n", READ_REG(hpa+IOC_TCNFG));
299 DBG_INIT("IOC_PDIR_BASE: %016lx\n", READ_REG(hpa+IOC_PDIR_BASE));
305 #ifdef ASSERT_PDIR_SANITY
308 * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry
309 * @ioc: IO MMU structure which owns the pdir we are interested in.
310 * @msg: text to print ont the output line.
313 * Print one entry of the IO MMU PDIR in human readable form.
316 sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide)
318 /* start printing from lowest pde in rval */
319 u64 *ptr = &ioc->pdir_base[pide & ~(BITS_PER_LONG - 1)];
320 unsigned long *rptr = (unsigned long *) &ioc->res_map[(pide >>3) & -sizeof(unsigned long)];
323 printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n",
324 msg, rptr, pide & (BITS_PER_LONG - 1), *rptr);
327 while (rcnt < BITS_PER_LONG) {
328 printk(KERN_DEBUG "%s %2d %p %016Lx\n",
329 (rcnt == (pide & (BITS_PER_LONG - 1)))
331 rcnt, ptr, (unsigned long long) *ptr );
335 printk(KERN_DEBUG "%s", msg);
340 * sba_check_pdir - debugging only - consistency checker
341 * @ioc: IO MMU structure which owns the pdir we are interested in.
342 * @msg: text to print ont the output line.
344 * Verify the resource map and pdir state is consistent
347 sba_check_pdir(struct ioc *ioc, char *msg)
349 u64 *rptr_end = (u64 *) &(ioc->res_map[ioc->res_size]);
350 u64 *rptr = (u64 *) ioc->res_map; /* resource map ptr */
351 u64 *pptr = ioc->pdir_base; /* pdir ptr */
354 while (rptr < rptr_end) {
356 int rcnt; /* number of bits we might check */
362 /* Get last byte and highest bit from that */
363 u32 pde = ((u32)((*pptr >> (63)) & 0x1));
364 if ((rval & 0x1) ^ pde)
367 ** BUMMER! -- res_map != pdir --
368 ** Dump rval and matching pdir entries
370 sba_dump_pdir_entry(ioc, msg, pide);
374 rval >>= 1; /* try the next bit */
378 rptr++; /* look at next word of res_map */
380 /* It'd be nice if we always got here :^) */
386 * sba_dump_sg - debugging only - print Scatter-Gather list
387 * @ioc: IO MMU structure which owns the pdir we are interested in.
388 * @startsg: head of the SG list
389 * @nents: number of entries in SG list
391 * print the SG list so we can verify it's correct by hand.
394 sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
396 while (nents-- > 0) {
397 printk(KERN_DEBUG " %d : DMA %08lx/%05x CPU %p\n", nents,
398 startsg->dma_address, startsg->dma_length,
399 sba_sg_address(startsg));
400 startsg = sg_next(startsg);
405 sba_check_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
407 struct scatterlist *the_sg = startsg;
408 int the_nents = nents;
410 while (the_nents-- > 0) {
411 if (sba_sg_address(the_sg) == 0x0UL)
412 sba_dump_sg(NULL, startsg, nents);
413 the_sg = sg_next(the_sg);
417 #endif /* ASSERT_PDIR_SANITY */
422 /**************************************************************
424 * I/O Pdir Resource Management
426 * Bits set in the resource map are in use.
427 * Each bit can represent a number of pages.
428 * LSbs represent lower addresses (IOVA's).
430 ***************************************************************/
431 #define PAGES_PER_RANGE 1 /* could increase this to 4 or 8 if needed */
433 /* Convert from IOVP to IOVA and vice versa. */
434 #define SBA_IOVA(ioc,iovp,offset) ((ioc->ibase) | (iovp) | (offset))
435 #define SBA_IOVP(ioc,iova) ((iova) & ~(ioc->ibase))
437 #define PDIR_ENTRY_SIZE sizeof(u64)
439 #define PDIR_INDEX(iovp) ((iovp)>>iovp_shift)
441 #define RESMAP_MASK(n) ~(~0UL << (n))
442 #define RESMAP_IDX_MASK (sizeof(unsigned long) - 1)
446 * For most cases the normal get_order is sufficient, however it limits us
447 * to PAGE_SIZE being the minimum mapping alignment and TC flush granularity.
448 * It only incurs about 1 clock cycle to use this one with the static variable
449 * and makes the code more intuitive.
451 static SBA_INLINE int
452 get_iovp_order (unsigned long size)
454 long double d = size - 1;
457 order = ia64_getf_exp(d);
458 order = order - iovp_shift - 0xffff + 1;
464 static unsigned long ptr_to_pide(struct ioc *ioc, unsigned long *res_ptr,
465 unsigned int bitshiftcnt)
467 return (((unsigned long)res_ptr - (unsigned long)ioc->res_map) << 3)
472 * sba_search_bitmap - find free space in IO PDIR resource bitmap
473 * @ioc: IO MMU structure which owns the pdir we are interested in.
474 * @bits_wanted: number of entries we need.
475 * @use_hint: use res_hint to indicate where to start looking
477 * Find consecutive free bits in resource bitmap.
478 * Each bit represents one entry in the IO Pdir.
479 * Cool perf optimization: search for log2(size) bits at a time.
481 static SBA_INLINE unsigned long
482 sba_search_bitmap(struct ioc *ioc, struct device *dev,
483 unsigned long bits_wanted, int use_hint)
485 unsigned long *res_ptr;
486 unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]);
487 unsigned long flags, pide = ~0UL, tpide;
488 unsigned long boundary_size;
492 ASSERT(((unsigned long) ioc->res_hint & (sizeof(unsigned long) - 1UL)) == 0);
493 ASSERT(res_ptr < res_end);
495 boundary_size = (unsigned long long)dma_get_seg_boundary(dev) + 1;
496 boundary_size = ALIGN(boundary_size, 1ULL << iovp_shift) >> iovp_shift;
498 BUG_ON(ioc->ibase & ~iovp_mask);
499 shift = ioc->ibase >> iovp_shift;
501 spin_lock_irqsave(&ioc->res_lock, flags);
503 /* Allow caller to force a search through the entire resource space */
504 if (likely(use_hint)) {
505 res_ptr = ioc->res_hint;
507 res_ptr = (ulong *)ioc->res_map;
508 ioc->res_bitshift = 0;
512 * N.B. REO/Grande defect AR2305 can cause TLB fetch timeouts
513 * if a TLB entry is purged while in use. sba_mark_invalid()
514 * purges IOTLB entries in power-of-two sizes, so we also
515 * allocate IOVA space in power-of-two sizes.
517 bits_wanted = 1UL << get_iovp_order(bits_wanted << iovp_shift);
519 if (likely(bits_wanted == 1)) {
520 unsigned int bitshiftcnt;
521 for(; res_ptr < res_end ; res_ptr++) {
522 if (likely(*res_ptr != ~0UL)) {
523 bitshiftcnt = ffz(*res_ptr);
524 *res_ptr |= (1UL << bitshiftcnt);
525 pide = ptr_to_pide(ioc, res_ptr, bitshiftcnt);
526 ioc->res_bitshift = bitshiftcnt + bits_wanted;
534 if (likely(bits_wanted <= BITS_PER_LONG/2)) {
536 ** Search the resource bit map on well-aligned values.
537 ** "o" is the alignment.
538 ** We need the alignment to invalidate I/O TLB using
539 ** SBA HW features in the unmap path.
541 unsigned long o = 1 << get_iovp_order(bits_wanted << iovp_shift);
542 uint bitshiftcnt = ROUNDUP(ioc->res_bitshift, o);
543 unsigned long mask, base_mask;
545 base_mask = RESMAP_MASK(bits_wanted);
546 mask = base_mask << bitshiftcnt;
548 DBG_RES("%s() o %ld %p", __func__, o, res_ptr);
549 for(; res_ptr < res_end ; res_ptr++)
551 DBG_RES(" %p %lx %lx\n", res_ptr, mask, *res_ptr);
553 for (; mask ; mask <<= o, bitshiftcnt += o) {
554 tpide = ptr_to_pide(ioc, res_ptr, bitshiftcnt);
555 ret = iommu_is_span_boundary(tpide, bits_wanted,
558 if ((0 == ((*res_ptr) & mask)) && !ret) {
559 *res_ptr |= mask; /* mark resources busy! */
561 ioc->res_bitshift = bitshiftcnt + bits_wanted;
575 qwords = bits_wanted >> 6; /* /64 */
576 bits = bits_wanted - (qwords * BITS_PER_LONG);
578 end = res_end - qwords;
580 for (; res_ptr < end; res_ptr++) {
581 tpide = ptr_to_pide(ioc, res_ptr, 0);
582 ret = iommu_is_span_boundary(tpide, bits_wanted,
583 shift, boundary_size);
586 for (i = 0 ; i < qwords ; i++) {
590 if (bits && res_ptr[i] && (__ffs(res_ptr[i]) < bits))
593 /* Found it, mark it */
594 for (i = 0 ; i < qwords ; i++)
596 res_ptr[i] |= RESMAP_MASK(bits);
600 ioc->res_bitshift = bits;
608 prefetch(ioc->res_map);
609 ioc->res_hint = (unsigned long *) ioc->res_map;
610 ioc->res_bitshift = 0;
611 spin_unlock_irqrestore(&ioc->res_lock, flags);
615 ioc->res_hint = res_ptr;
616 spin_unlock_irqrestore(&ioc->res_lock, flags);
622 * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
623 * @ioc: IO MMU structure which owns the pdir we are interested in.
624 * @size: number of bytes to create a mapping for
626 * Given a size, find consecutive unmarked and then mark those bits in the
630 sba_alloc_range(struct ioc *ioc, struct device *dev, size_t size)
632 unsigned int pages_needed = size >> iovp_shift;
633 #ifdef PDIR_SEARCH_TIMING
634 unsigned long itc_start;
638 ASSERT(pages_needed);
639 ASSERT(0 == (size & ~iovp_mask));
641 #ifdef PDIR_SEARCH_TIMING
642 itc_start = ia64_get_itc();
645 ** "seek and ye shall find"...praying never hurts either...
647 pide = sba_search_bitmap(ioc, dev, pages_needed, 1);
648 if (unlikely(pide >= (ioc->res_size << 3))) {
649 pide = sba_search_bitmap(ioc, dev, pages_needed, 0);
650 if (unlikely(pide >= (ioc->res_size << 3))) {
651 #if DELAYED_RESOURCE_CNT > 0
655 ** With delayed resource freeing, we can give this one more shot. We're
656 ** getting close to being in trouble here, so do what we can to make this
659 spin_lock_irqsave(&ioc->saved_lock, flags);
660 if (ioc->saved_cnt > 0) {
661 struct sba_dma_pair *d;
662 int cnt = ioc->saved_cnt;
664 d = &(ioc->saved[ioc->saved_cnt - 1]);
666 spin_lock(&ioc->res_lock);
668 sba_mark_invalid(ioc, d->iova, d->size);
669 sba_free_range(ioc, d->iova, d->size);
673 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */
674 spin_unlock(&ioc->res_lock);
676 spin_unlock_irqrestore(&ioc->saved_lock, flags);
678 pide = sba_search_bitmap(ioc, dev, pages_needed, 0);
679 if (unlikely(pide >= (ioc->res_size << 3)))
680 panic(__FILE__ ": I/O MMU @ %p is out of mapping resources\n",
683 panic(__FILE__ ": I/O MMU @ %p is out of mapping resources\n",
689 #ifdef PDIR_SEARCH_TIMING
690 ioc->avg_search[ioc->avg_idx++] = (ia64_get_itc() - itc_start) / pages_needed;
691 ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;
694 prefetchw(&(ioc->pdir_base[pide]));
696 #ifdef ASSERT_PDIR_SANITY
697 /* verify the first enable bit is clear */
698 if(0x00 != ((u8 *) ioc->pdir_base)[pide*PDIR_ENTRY_SIZE + 7]) {
699 sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide);
703 DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
704 __func__, size, pages_needed, pide,
705 (uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map),
713 * sba_free_range - unmark bits in IO PDIR resource bitmap
714 * @ioc: IO MMU structure which owns the pdir we are interested in.
715 * @iova: IO virtual address which was previously allocated.
716 * @size: number of bytes to create a mapping for
718 * clear bits in the ioc's resource map
720 static SBA_INLINE void
721 sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size)
723 unsigned long iovp = SBA_IOVP(ioc, iova);
724 unsigned int pide = PDIR_INDEX(iovp);
725 unsigned int ridx = pide >> 3; /* convert bit to byte address */
726 unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]);
727 int bits_not_wanted = size >> iovp_shift;
730 /* Round up to power-of-two size: see AR2305 note above */
731 bits_not_wanted = 1UL << get_iovp_order(bits_not_wanted << iovp_shift);
732 for (; bits_not_wanted > 0 ; res_ptr++) {
734 if (unlikely(bits_not_wanted > BITS_PER_LONG)) {
736 /* these mappings start 64bit aligned */
738 bits_not_wanted -= BITS_PER_LONG;
739 pide += BITS_PER_LONG;
743 /* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
744 m = RESMAP_MASK(bits_not_wanted) << (pide & (BITS_PER_LONG - 1));
747 DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n", __func__, (uint) iova, size,
748 bits_not_wanted, m, pide, res_ptr, *res_ptr);
751 ASSERT(bits_not_wanted);
752 ASSERT((*res_ptr & m) == m); /* verify same bits are set */
759 /**************************************************************
761 * "Dynamic DMA Mapping" support (aka "Coherent I/O")
763 ***************************************************************/
766 * sba_io_pdir_entry - fill in one IO PDIR entry
767 * @pdir_ptr: pointer to IO PDIR entry
768 * @vba: Virtual CPU address of buffer to map
770 * SBA Mapping Routine
772 * Given a virtual address (vba, arg1) sba_io_pdir_entry()
773 * loads the I/O PDIR entry pointed to by pdir_ptr (arg0).
774 * Each IO Pdir entry consists of 8 bytes as shown below
778 * +-+---------------------+----------------------------------+----+--------+
779 * |V| U | PPN[39:12] | U | FF |
780 * +-+---------------------+----------------------------------+----+--------+
784 * PPN == Physical Page Number
786 * The physical address fields are filled with the results of virt_to_phys()
791 #define sba_io_pdir_entry(pdir_ptr, vba) *pdir_ptr = ((vba & ~0xE000000000000FFFULL) \
792 | 0x8000000000000000ULL)
795 sba_io_pdir_entry(u64 *pdir_ptr, unsigned long vba)
797 *pdir_ptr = ((vba & ~0xE000000000000FFFULL) | 0x80000000000000FFULL);
801 #ifdef ENABLE_MARK_CLEAN
803 * Since DMA is i-cache coherent, any (complete) pages that were written via
804 * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to
805 * flush them when they get mapped into an executable vm-area.
808 mark_clean (void *addr, size_t size)
810 unsigned long pg_addr, end;
812 pg_addr = PAGE_ALIGN((unsigned long) addr);
813 end = (unsigned long) addr + size;
814 while (pg_addr + PAGE_SIZE <= end) {
815 struct page *page = virt_to_page((void *)pg_addr);
816 set_bit(PG_arch_1, &page->flags);
817 pg_addr += PAGE_SIZE;
823 * sba_mark_invalid - invalidate one or more IO PDIR entries
824 * @ioc: IO MMU structure which owns the pdir we are interested in.
825 * @iova: IO Virtual Address mapped earlier
826 * @byte_cnt: number of bytes this mapping covers.
828 * Marking the IO PDIR entry(ies) as Invalid and invalidate
829 * corresponding IO TLB entry. The PCOM (Purge Command Register)
830 * is to purge stale entries in the IO TLB when unmapping entries.
832 * The PCOM register supports purging of multiple pages, with a minium
833 * of 1 page and a maximum of 2GB. Hardware requires the address be
834 * aligned to the size of the range being purged. The size of the range
835 * must be a power of 2. The "Cool perf optimization" in the
836 * allocation routine helps keep that true.
838 static SBA_INLINE void
839 sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
841 u32 iovp = (u32) SBA_IOVP(ioc,iova);
843 int off = PDIR_INDEX(iovp);
845 /* Must be non-zero and rounded up */
846 ASSERT(byte_cnt > 0);
847 ASSERT(0 == (byte_cnt & ~iovp_mask));
849 #ifdef ASSERT_PDIR_SANITY
850 /* Assert first pdir entry is set */
851 if (!(ioc->pdir_base[off] >> 60)) {
852 sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp));
856 if (byte_cnt <= iovp_size)
858 ASSERT(off < ioc->pdir_size);
860 iovp |= iovp_shift; /* set "size" field for PCOM */
862 #ifndef FULL_VALID_PDIR
864 ** clear I/O PDIR entry "valid" bit
865 ** Do NOT clear the rest - save it for debugging.
866 ** We should only clear bits that have previously
869 ioc->pdir_base[off] &= ~(0x80000000000000FFULL);
872 ** If we want to maintain the PDIR as valid, put in
873 ** the spill page so devices prefetching won't
874 ** cause a hard fail.
876 ioc->pdir_base[off] = (0x80000000000000FFULL | prefetch_spill_page);
879 u32 t = get_iovp_order(byte_cnt) + iovp_shift;
882 ASSERT(t <= 31); /* 2GB! Max value of "size" field */
885 /* verify this pdir entry is enabled */
886 ASSERT(ioc->pdir_base[off] >> 63);
887 #ifndef FULL_VALID_PDIR
888 /* clear I/O Pdir entry "valid" bit first */
889 ioc->pdir_base[off] &= ~(0x80000000000000FFULL);
891 ioc->pdir_base[off] = (0x80000000000000FFULL | prefetch_spill_page);
894 byte_cnt -= iovp_size;
895 } while (byte_cnt > 0);
898 WRITE_REG(iovp | ioc->ibase, ioc->ioc_hpa+IOC_PCOM);
902 * sba_map_single - map one buffer and return IOVA for DMA
903 * @dev: instance of PCI owned by the driver that's asking.
904 * @addr: driver buffer to map.
905 * @size: number of bytes to map in driver buffer.
908 * See Documentation/DMA-mapping.txt
911 sba_map_single(struct device *dev, void *addr, size_t size, int dir)
918 #ifdef ASSERT_PDIR_SANITY
921 #ifdef ALLOW_IOV_BYPASS
922 unsigned long pci_addr = virt_to_phys(addr);
925 #ifdef ALLOW_IOV_BYPASS
926 ASSERT(to_pci_dev(dev)->dma_mask);
928 ** Check if the PCI device can DMA to ptr... if so, just return ptr
930 if (likely((pci_addr & ~to_pci_dev(dev)->dma_mask) == 0)) {
932 ** Device is bit capable of DMA'ing to the buffer...
933 ** just return the PCI address of ptr
935 DBG_BYPASS("sba_map_single() bypass mask/addr: 0x%lx/0x%lx\n",
936 to_pci_dev(dev)->dma_mask, pci_addr);
943 prefetch(ioc->res_hint);
946 ASSERT(size <= DMA_CHUNK_SIZE);
948 /* save offset bits */
949 offset = ((dma_addr_t) (long) addr) & ~iovp_mask;
951 /* round up to nearest iovp_size */
952 size = (size + offset + ~iovp_mask) & iovp_mask;
954 #ifdef ASSERT_PDIR_SANITY
955 spin_lock_irqsave(&ioc->res_lock, flags);
956 if (sba_check_pdir(ioc,"Check before sba_map_single()"))
957 panic("Sanity check failed");
958 spin_unlock_irqrestore(&ioc->res_lock, flags);
961 pide = sba_alloc_range(ioc, dev, size);
963 iovp = (dma_addr_t) pide << iovp_shift;
965 DBG_RUN("%s() 0x%p -> 0x%lx\n", __func__, addr, (long) iovp | offset);
967 pdir_start = &(ioc->pdir_base[pide]);
970 ASSERT(((u8 *)pdir_start)[7] == 0); /* verify availability */
971 sba_io_pdir_entry(pdir_start, (unsigned long) addr);
973 DBG_RUN(" pdir 0x%p %lx\n", pdir_start, *pdir_start);
979 /* force pdir update */
982 /* form complete address */
983 #ifdef ASSERT_PDIR_SANITY
984 spin_lock_irqsave(&ioc->res_lock, flags);
985 sba_check_pdir(ioc,"Check after sba_map_single()");
986 spin_unlock_irqrestore(&ioc->res_lock, flags);
988 return SBA_IOVA(ioc, iovp, offset);
991 #ifdef ENABLE_MARK_CLEAN
992 static SBA_INLINE void
993 sba_mark_clean(struct ioc *ioc, dma_addr_t iova, size_t size)
995 u32 iovp = (u32) SBA_IOVP(ioc,iova);
996 int off = PDIR_INDEX(iovp);
999 if (size <= iovp_size) {
1000 addr = phys_to_virt(ioc->pdir_base[off] &
1001 ~0xE000000000000FFFULL);
1002 mark_clean(addr, size);
1005 addr = phys_to_virt(ioc->pdir_base[off] &
1006 ~0xE000000000000FFFULL);
1007 mark_clean(addr, min(size, iovp_size));
1016 * sba_unmap_single - unmap one IOVA and free resources
1017 * @dev: instance of PCI owned by the driver that's asking.
1018 * @iova: IOVA of driver buffer previously mapped.
1019 * @size: number of bytes mapped in driver buffer.
1020 * @dir: R/W or both.
1022 * See Documentation/DMA-mapping.txt
1024 void sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size, int dir)
1027 #if DELAYED_RESOURCE_CNT > 0
1028 struct sba_dma_pair *d;
1030 unsigned long flags;
1036 #ifdef ALLOW_IOV_BYPASS
1037 if (likely((iova & ioc->imask) != ioc->ibase)) {
1039 ** Address does not fall w/in IOVA, must be bypassing
1041 DBG_BYPASS("sba_unmap_single() bypass addr: 0x%lx\n", iova);
1043 #ifdef ENABLE_MARK_CLEAN
1044 if (dir == DMA_FROM_DEVICE) {
1045 mark_clean(phys_to_virt(iova), size);
1051 offset = iova & ~iovp_mask;
1053 DBG_RUN("%s() iovp 0x%lx/%x\n", __func__, (long) iova, size);
1055 iova ^= offset; /* clear offset bits */
1057 size = ROUNDUP(size, iovp_size);
1059 #ifdef ENABLE_MARK_CLEAN
1060 if (dir == DMA_FROM_DEVICE)
1061 sba_mark_clean(ioc, iova, size);
1064 #if DELAYED_RESOURCE_CNT > 0
1065 spin_lock_irqsave(&ioc->saved_lock, flags);
1066 d = &(ioc->saved[ioc->saved_cnt]);
1069 if (unlikely(++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT)) {
1070 int cnt = ioc->saved_cnt;
1071 spin_lock(&ioc->res_lock);
1073 sba_mark_invalid(ioc, d->iova, d->size);
1074 sba_free_range(ioc, d->iova, d->size);
1078 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */
1079 spin_unlock(&ioc->res_lock);
1081 spin_unlock_irqrestore(&ioc->saved_lock, flags);
1082 #else /* DELAYED_RESOURCE_CNT == 0 */
1083 spin_lock_irqsave(&ioc->res_lock, flags);
1084 sba_mark_invalid(ioc, iova, size);
1085 sba_free_range(ioc, iova, size);
1086 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */
1087 spin_unlock_irqrestore(&ioc->res_lock, flags);
1088 #endif /* DELAYED_RESOURCE_CNT == 0 */
1093 * sba_alloc_coherent - allocate/map shared mem for DMA
1094 * @dev: instance of PCI owned by the driver that's asking.
1095 * @size: number of bytes mapped in driver buffer.
1096 * @dma_handle: IOVA of new buffer.
1098 * See Documentation/DMA-mapping.txt
1101 sba_alloc_coherent (struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flags)
1112 page = alloc_pages_node(ioc->node == MAX_NUMNODES ?
1113 numa_node_id() : ioc->node, flags,
1116 if (unlikely(!page))
1119 addr = page_address(page);
1122 addr = (void *) __get_free_pages(flags, get_order(size));
1124 if (unlikely(!addr))
1127 memset(addr, 0, size);
1128 *dma_handle = virt_to_phys(addr);
1130 #ifdef ALLOW_IOV_BYPASS
1131 ASSERT(dev->coherent_dma_mask);
1133 ** Check if the PCI device can DMA to ptr... if so, just return ptr
1135 if (likely((*dma_handle & ~dev->coherent_dma_mask) == 0)) {
1136 DBG_BYPASS("sba_alloc_coherent() bypass mask/addr: 0x%lx/0x%lx\n",
1137 dev->coherent_dma_mask, *dma_handle);
1144 * If device can't bypass or bypass is disabled, pass the 32bit fake
1145 * device to map single to get an iova mapping.
1147 *dma_handle = sba_map_single(&ioc->sac_only_dev->dev, addr, size, 0);
1154 * sba_free_coherent - free/unmap shared mem for DMA
1155 * @dev: instance of PCI owned by the driver that's asking.
1156 * @size: number of bytes mapped in driver buffer.
1157 * @vaddr: virtual address IOVA of "consistent" buffer.
1158 * @dma_handler: IO virtual address of "consistent" buffer.
1160 * See Documentation/DMA-mapping.txt
1162 void sba_free_coherent (struct device *dev, size_t size, void *vaddr, dma_addr_t dma_handle)
1164 sba_unmap_single(dev, dma_handle, size, 0);
1165 free_pages((unsigned long) vaddr, get_order(size));
1170 ** Since 0 is a valid pdir_base index value, can't use that
1171 ** to determine if a value is valid or not. Use a flag to indicate
1172 ** the SG list entry contains a valid pdir index.
1174 #define PIDE_FLAG 0x1UL
1176 #ifdef DEBUG_LARGE_SG_ENTRIES
1177 int dump_run_sg = 0;
1182 * sba_fill_pdir - write allocated SG entries into IO PDIR
1183 * @ioc: IO MMU structure which owns the pdir we are interested in.
1184 * @startsg: list of IOVA/size pairs
1185 * @nents: number of entries in startsg list
1187 * Take preprocessed SG list and write corresponding entries
1191 static SBA_INLINE int
1194 struct scatterlist *startsg,
1197 struct scatterlist *dma_sg = startsg; /* pointer to current DMA */
1200 unsigned long dma_offset = 0;
1202 while (nents-- > 0) {
1203 int cnt = startsg->dma_length;
1204 startsg->dma_length = 0;
1206 #ifdef DEBUG_LARGE_SG_ENTRIES
1208 printk(" %2d : %08lx/%05x %p\n",
1209 nents, startsg->dma_address, cnt,
1210 sba_sg_address(startsg));
1212 DBG_RUN_SG(" %d : %08lx/%05x %p\n",
1213 nents, startsg->dma_address, cnt,
1214 sba_sg_address(startsg));
1217 ** Look for the start of a new DMA stream
1219 if (startsg->dma_address & PIDE_FLAG) {
1220 u32 pide = startsg->dma_address & ~PIDE_FLAG;
1221 dma_offset = (unsigned long) pide & ~iovp_mask;
1222 startsg->dma_address = 0;
1224 dma_sg = sg_next(dma_sg);
1225 dma_sg->dma_address = pide | ioc->ibase;
1226 pdirp = &(ioc->pdir_base[pide >> iovp_shift]);
1231 ** Look for a VCONTIG chunk
1234 unsigned long vaddr = (unsigned long) sba_sg_address(startsg);
1237 /* Since multiple Vcontig blocks could make up
1238 ** one DMA stream, *add* cnt to dma_len.
1240 dma_sg->dma_length += cnt;
1242 dma_offset=0; /* only want offset on first chunk */
1243 cnt = ROUNDUP(cnt, iovp_size);
1245 sba_io_pdir_entry(pdirp, vaddr);
1251 startsg = sg_next(startsg);
1253 /* force pdir update */
1256 #ifdef DEBUG_LARGE_SG_ENTRIES
1264 ** Two address ranges are DMA contiguous *iff* "end of prev" and
1265 ** "start of next" are both on an IOV page boundary.
1267 ** (shift left is a quick trick to mask off upper bits)
1269 #define DMA_CONTIG(__X, __Y) \
1270 (((((unsigned long) __X) | ((unsigned long) __Y)) << (BITS_PER_LONG - iovp_shift)) == 0UL)
1274 * sba_coalesce_chunks - preprocess the SG list
1275 * @ioc: IO MMU structure which owns the pdir we are interested in.
1276 * @startsg: list of IOVA/size pairs
1277 * @nents: number of entries in startsg list
1279 * First pass is to walk the SG list and determine where the breaks are
1280 * in the DMA stream. Allocates PDIR entries but does not fill them.
1281 * Returns the number of DMA chunks.
1283 * Doing the fill separate from the coalescing/allocation keeps the
1284 * code simpler. Future enhancement could make one pass through
1285 * the sglist do both.
1287 static SBA_INLINE int
1288 sba_coalesce_chunks(struct ioc *ioc, struct device *dev,
1289 struct scatterlist *startsg,
1292 struct scatterlist *vcontig_sg; /* VCONTIG chunk head */
1293 unsigned long vcontig_len; /* len of VCONTIG chunk */
1294 unsigned long vcontig_end;
1295 struct scatterlist *dma_sg; /* next DMA stream head */
1296 unsigned long dma_offset, dma_len; /* start/len of DMA stream */
1298 unsigned int max_seg_size = dma_get_max_seg_size(dev);
1301 unsigned long vaddr = (unsigned long) sba_sg_address(startsg);
1304 ** Prepare for first/next DMA stream
1306 dma_sg = vcontig_sg = startsg;
1307 dma_len = vcontig_len = vcontig_end = startsg->length;
1308 vcontig_end += vaddr;
1309 dma_offset = vaddr & ~iovp_mask;
1311 /* PARANOID: clear entries */
1312 startsg->dma_address = startsg->dma_length = 0;
1315 ** This loop terminates one iteration "early" since
1316 ** it's always looking one "ahead".
1318 while (--nents > 0) {
1319 unsigned long vaddr; /* tmp */
1321 startsg = sg_next(startsg);
1324 startsg->dma_address = startsg->dma_length = 0;
1326 /* catch brokenness in SCSI layer */
1327 ASSERT(startsg->length <= DMA_CHUNK_SIZE);
1330 ** First make sure current dma stream won't
1331 ** exceed DMA_CHUNK_SIZE if we coalesce the
1334 if (((dma_len + dma_offset + startsg->length + ~iovp_mask) & iovp_mask)
1338 if (dma_len + startsg->length > max_seg_size)
1342 ** Then look for virtually contiguous blocks.
1344 ** append the next transaction?
1346 vaddr = (unsigned long) sba_sg_address(startsg);
1347 if (vcontig_end == vaddr)
1349 vcontig_len += startsg->length;
1350 vcontig_end += startsg->length;
1351 dma_len += startsg->length;
1355 #ifdef DEBUG_LARGE_SG_ENTRIES
1356 dump_run_sg = (vcontig_len > iovp_size);
1360 ** Not virtually contigous.
1361 ** Terminate prev chunk.
1362 ** Start a new chunk.
1364 ** Once we start a new VCONTIG chunk, dma_offset
1365 ** can't change. And we need the offset from the first
1366 ** chunk - not the last one. Ergo Successive chunks
1367 ** must start on page boundaries and dove tail
1368 ** with it's predecessor.
1370 vcontig_sg->dma_length = vcontig_len;
1372 vcontig_sg = startsg;
1373 vcontig_len = startsg->length;
1376 ** 3) do the entries end/start on page boundaries?
1377 ** Don't update vcontig_end until we've checked.
1379 if (DMA_CONTIG(vcontig_end, vaddr))
1381 vcontig_end = vcontig_len + vaddr;
1382 dma_len += vcontig_len;
1390 ** End of DMA Stream
1391 ** Terminate last VCONTIG block.
1392 ** Allocate space for DMA stream.
1394 vcontig_sg->dma_length = vcontig_len;
1395 dma_len = (dma_len + dma_offset + ~iovp_mask) & iovp_mask;
1396 ASSERT(dma_len <= DMA_CHUNK_SIZE);
1397 dma_sg->dma_address = (dma_addr_t) (PIDE_FLAG
1398 | (sba_alloc_range(ioc, dev, dma_len) << iovp_shift)
1408 * sba_map_sg - map Scatter/Gather list
1409 * @dev: instance of PCI owned by the driver that's asking.
1410 * @sglist: array of buffer/length pairs
1411 * @nents: number of entries in list
1412 * @dir: R/W or both.
1414 * See Documentation/DMA-mapping.txt
1416 int sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents, int dir)
1419 int coalesced, filled = 0;
1420 #ifdef ASSERT_PDIR_SANITY
1421 unsigned long flags;
1423 #ifdef ALLOW_IOV_BYPASS_SG
1424 struct scatterlist *sg;
1427 DBG_RUN_SG("%s() START %d entries\n", __func__, nents);
1431 #ifdef ALLOW_IOV_BYPASS_SG
1432 ASSERT(to_pci_dev(dev)->dma_mask);
1433 if (likely((ioc->dma_mask & ~to_pci_dev(dev)->dma_mask) == 0)) {
1434 for_each_sg(sglist, sg, nents, filled) {
1435 sg->dma_length = sg->length;
1436 sg->dma_address = virt_to_phys(sba_sg_address(sg));
1441 /* Fast path single entry scatterlists. */
1443 sglist->dma_length = sglist->length;
1444 sglist->dma_address = sba_map_single(dev, sba_sg_address(sglist), sglist->length, dir);
1448 #ifdef ASSERT_PDIR_SANITY
1449 spin_lock_irqsave(&ioc->res_lock, flags);
1450 if (sba_check_pdir(ioc,"Check before sba_map_sg()"))
1452 sba_dump_sg(ioc, sglist, nents);
1453 panic("Check before sba_map_sg()");
1455 spin_unlock_irqrestore(&ioc->res_lock, flags);
1458 prefetch(ioc->res_hint);
1461 ** First coalesce the chunks and allocate I/O pdir space
1463 ** If this is one DMA stream, we can properly map using the
1464 ** correct virtual address associated with each DMA page.
1465 ** w/o this association, we wouldn't have coherent DMA!
1466 ** Access to the virtual address is what forces a two pass algorithm.
1468 coalesced = sba_coalesce_chunks(ioc, dev, sglist, nents);
1471 ** Program the I/O Pdir
1473 ** map the virtual addresses to the I/O Pdir
1474 ** o dma_address will contain the pdir index
1475 ** o dma_len will contain the number of bytes to map
1476 ** o address contains the virtual address.
1478 filled = sba_fill_pdir(ioc, sglist, nents);
1480 #ifdef ASSERT_PDIR_SANITY
1481 spin_lock_irqsave(&ioc->res_lock, flags);
1482 if (sba_check_pdir(ioc,"Check after sba_map_sg()"))
1484 sba_dump_sg(ioc, sglist, nents);
1485 panic("Check after sba_map_sg()\n");
1487 spin_unlock_irqrestore(&ioc->res_lock, flags);
1490 ASSERT(coalesced == filled);
1491 DBG_RUN_SG("%s() DONE %d mappings\n", __func__, filled);
1498 * sba_unmap_sg - unmap Scatter/Gather list
1499 * @dev: instance of PCI owned by the driver that's asking.
1500 * @sglist: array of buffer/length pairs
1501 * @nents: number of entries in list
1502 * @dir: R/W or both.
1504 * See Documentation/DMA-mapping.txt
1506 void sba_unmap_sg (struct device *dev, struct scatterlist *sglist, int nents, int dir)
1508 #ifdef ASSERT_PDIR_SANITY
1510 unsigned long flags;
1513 DBG_RUN_SG("%s() START %d entries, %p,%x\n",
1514 __func__, nents, sba_sg_address(sglist), sglist->length);
1516 #ifdef ASSERT_PDIR_SANITY
1520 spin_lock_irqsave(&ioc->res_lock, flags);
1521 sba_check_pdir(ioc,"Check before sba_unmap_sg()");
1522 spin_unlock_irqrestore(&ioc->res_lock, flags);
1525 while (nents && sglist->dma_length) {
1527 sba_unmap_single(dev, sglist->dma_address, sglist->dma_length, dir);
1528 sglist = sg_next(sglist);
1532 DBG_RUN_SG("%s() DONE (nents %d)\n", __func__, nents);
1534 #ifdef ASSERT_PDIR_SANITY
1535 spin_lock_irqsave(&ioc->res_lock, flags);
1536 sba_check_pdir(ioc,"Check after sba_unmap_sg()");
1537 spin_unlock_irqrestore(&ioc->res_lock, flags);
1542 /**************************************************************
1544 * Initialization and claim
1546 ***************************************************************/
1549 ioc_iova_init(struct ioc *ioc)
1553 struct pci_dev *device = NULL;
1554 #ifdef FULL_VALID_PDIR
1555 unsigned long index;
1559 ** Firmware programs the base and size of a "safe IOVA space"
1560 ** (one that doesn't overlap memory or LMMIO space) in the
1561 ** IBASE and IMASK registers.
1563 ioc->ibase = READ_REG(ioc->ioc_hpa + IOC_IBASE) & ~0x1UL;
1564 ioc->imask = READ_REG(ioc->ioc_hpa + IOC_IMASK) | 0xFFFFFFFF00000000UL;
1566 ioc->iov_size = ~ioc->imask + 1;
1568 DBG_INIT("%s() hpa %p IOV base 0x%lx mask 0x%lx (%dMB)\n",
1569 __func__, ioc->ioc_hpa, ioc->ibase, ioc->imask,
1570 ioc->iov_size >> 20);
1572 switch (iovp_size) {
1573 case 4*1024: tcnfg = 0; break;
1574 case 8*1024: tcnfg = 1; break;
1575 case 16*1024: tcnfg = 2; break;
1576 case 64*1024: tcnfg = 3; break;
1578 panic(PFX "Unsupported IOTLB page size %ldK",
1582 WRITE_REG(tcnfg, ioc->ioc_hpa + IOC_TCNFG);
1584 ioc->pdir_size = (ioc->iov_size / iovp_size) * PDIR_ENTRY_SIZE;
1585 ioc->pdir_base = (void *) __get_free_pages(GFP_KERNEL,
1586 get_order(ioc->pdir_size));
1587 if (!ioc->pdir_base)
1588 panic(PFX "Couldn't allocate I/O Page Table\n");
1590 memset(ioc->pdir_base, 0, ioc->pdir_size);
1592 DBG_INIT("%s() IOV page size %ldK pdir %p size %x\n", __func__,
1593 iovp_size >> 10, ioc->pdir_base, ioc->pdir_size);
1595 ASSERT(ALIGN((unsigned long) ioc->pdir_base, 4*1024) == (unsigned long) ioc->pdir_base);
1596 WRITE_REG(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1599 ** If an AGP device is present, only use half of the IOV space
1600 ** for PCI DMA. Unfortunately we can't know ahead of time
1601 ** whether GART support will actually be used, for now we
1602 ** can just key on an AGP device found in the system.
1603 ** We program the next pdir index after we stop w/ a key for
1604 ** the GART code to handshake on.
1606 for_each_pci_dev(device)
1607 agp_found |= pci_find_capability(device, PCI_CAP_ID_AGP);
1609 if (agp_found && reserve_sba_gart) {
1610 printk(KERN_INFO PFX "reserving %dMb of IOVA space at 0x%lx for agpgart\n",
1611 ioc->iov_size/2 >> 20, ioc->ibase + ioc->iov_size/2);
1612 ioc->pdir_size /= 2;
1613 ((u64 *)ioc->pdir_base)[PDIR_INDEX(ioc->iov_size/2)] = ZX1_SBA_IOMMU_COOKIE;
1615 #ifdef FULL_VALID_PDIR
1617 ** Check to see if the spill page has been allocated, we don't need more than
1618 ** one across multiple SBAs.
1620 if (!prefetch_spill_page) {
1621 char *spill_poison = "SBAIOMMU POISON";
1622 int poison_size = 16;
1623 void *poison_addr, *addr;
1625 addr = (void *)__get_free_pages(GFP_KERNEL, get_order(iovp_size));
1627 panic(PFX "Couldn't allocate PDIR spill page\n");
1630 for ( ; (u64) poison_addr < addr + iovp_size; poison_addr += poison_size)
1631 memcpy(poison_addr, spill_poison, poison_size);
1633 prefetch_spill_page = virt_to_phys(addr);
1635 DBG_INIT("%s() prefetch spill addr: 0x%lx\n", __func__, prefetch_spill_page);
1638 ** Set all the PDIR entries valid w/ the spill page as the target
1640 for (index = 0 ; index < (ioc->pdir_size / PDIR_ENTRY_SIZE) ; index++)
1641 ((u64 *)ioc->pdir_base)[index] = (0x80000000000000FF | prefetch_spill_page);
1644 /* Clear I/O TLB of any possible entries */
1645 WRITE_REG(ioc->ibase | (get_iovp_order(ioc->iov_size) + iovp_shift), ioc->ioc_hpa + IOC_PCOM);
1646 READ_REG(ioc->ioc_hpa + IOC_PCOM);
1648 /* Enable IOVA translation */
1649 WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa + IOC_IBASE);
1650 READ_REG(ioc->ioc_hpa + IOC_IBASE);
1654 ioc_resource_init(struct ioc *ioc)
1656 spin_lock_init(&ioc->res_lock);
1657 #if DELAYED_RESOURCE_CNT > 0
1658 spin_lock_init(&ioc->saved_lock);
1661 /* resource map size dictated by pdir_size */
1662 ioc->res_size = ioc->pdir_size / PDIR_ENTRY_SIZE; /* entries */
1663 ioc->res_size >>= 3; /* convert bit count to byte count */
1664 DBG_INIT("%s() res_size 0x%x\n", __func__, ioc->res_size);
1666 ioc->res_map = (char *) __get_free_pages(GFP_KERNEL,
1667 get_order(ioc->res_size));
1669 panic(PFX "Couldn't allocate resource map\n");
1671 memset(ioc->res_map, 0, ioc->res_size);
1672 /* next available IOVP - circular search */
1673 ioc->res_hint = (unsigned long *) ioc->res_map;
1675 #ifdef ASSERT_PDIR_SANITY
1676 /* Mark first bit busy - ie no IOVA 0 */
1677 ioc->res_map[0] = 0x1;
1678 ioc->pdir_base[0] = 0x8000000000000000ULL | ZX1_SBA_IOMMU_COOKIE;
1680 #ifdef FULL_VALID_PDIR
1681 /* Mark the last resource used so we don't prefetch beyond IOVA space */
1682 ioc->res_map[ioc->res_size - 1] |= 0x80UL; /* res_map is chars */
1683 ioc->pdir_base[(ioc->pdir_size / PDIR_ENTRY_SIZE) - 1] = (0x80000000000000FF
1684 | prefetch_spill_page);
1687 DBG_INIT("%s() res_map %x %p\n", __func__,
1688 ioc->res_size, (void *) ioc->res_map);
1692 ioc_sac_init(struct ioc *ioc)
1694 struct pci_dev *sac = NULL;
1695 struct pci_controller *controller = NULL;
1698 * pci_alloc_coherent() must return a DMA address which is
1699 * SAC (single address cycle) addressable, so allocate a
1700 * pseudo-device to enforce that.
1702 sac = kzalloc(sizeof(*sac), GFP_KERNEL);
1704 panic(PFX "Couldn't allocate struct pci_dev");
1706 controller = kzalloc(sizeof(*controller), GFP_KERNEL);
1708 panic(PFX "Couldn't allocate struct pci_controller");
1710 controller->iommu = ioc;
1711 sac->sysdata = controller;
1712 sac->dma_mask = 0xFFFFFFFFUL;
1714 sac->dev.bus = &pci_bus_type;
1716 ioc->sac_only_dev = sac;
1720 ioc_zx1_init(struct ioc *ioc)
1722 unsigned long rope_config;
1725 if (ioc->rev < 0x20)
1726 panic(PFX "IOC 2.0 or later required for IOMMU support\n");
1728 /* 38 bit memory controller + extra bit for range displaced by MMIO */
1729 ioc->dma_mask = (0x1UL << 39) - 1;
1732 ** Clear ROPE(N)_CONFIG AO bit.
1733 ** Disables "NT Ordering" (~= !"Relaxed Ordering")
1734 ** Overrides bit 1 in DMA Hint Sets.
1735 ** Improves netperf UDP_STREAM by ~10% for tg3 on bcm5701.
1737 for (i=0; i<(8*8); i+=8) {
1738 rope_config = READ_REG(ioc->ioc_hpa + IOC_ROPE0_CFG + i);
1739 rope_config &= ~IOC_ROPE_AO;
1740 WRITE_REG(rope_config, ioc->ioc_hpa + IOC_ROPE0_CFG + i);
1744 typedef void (initfunc)(struct ioc *);
1752 static struct ioc_iommu ioc_iommu_info[] __initdata = {
1753 { ZX1_IOC_ID, "zx1", ioc_zx1_init },
1754 { ZX2_IOC_ID, "zx2", NULL },
1755 { SX1000_IOC_ID, "sx1000", NULL },
1756 { SX2000_IOC_ID, "sx2000", NULL },
1759 static struct ioc * __init
1760 ioc_init(u64 hpa, void *handle)
1763 struct ioc_iommu *info;
1765 ioc = kzalloc(sizeof(*ioc), GFP_KERNEL);
1769 ioc->next = ioc_list;
1772 ioc->handle = handle;
1773 ioc->ioc_hpa = ioremap(hpa, 0x1000);
1775 ioc->func_id = READ_REG(ioc->ioc_hpa + IOC_FUNC_ID);
1776 ioc->rev = READ_REG(ioc->ioc_hpa + IOC_FCLASS) & 0xFFUL;
1777 ioc->dma_mask = 0xFFFFFFFFFFFFFFFFUL; /* conservative */
1779 for (info = ioc_iommu_info; info < ioc_iommu_info + ARRAY_SIZE(ioc_iommu_info); info++) {
1780 if (ioc->func_id == info->func_id) {
1781 ioc->name = info->name;
1787 iovp_size = (1 << iovp_shift);
1788 iovp_mask = ~(iovp_size - 1);
1790 DBG_INIT("%s: PAGE_SIZE %ldK, iovp_size %ldK\n", __func__,
1791 PAGE_SIZE >> 10, iovp_size >> 10);
1794 ioc->name = kmalloc(24, GFP_KERNEL);
1796 sprintf((char *) ioc->name, "Unknown (%04x:%04x)",
1797 ioc->func_id & 0xFFFF, (ioc->func_id >> 16) & 0xFFFF);
1799 ioc->name = "Unknown";
1803 ioc_resource_init(ioc);
1806 if ((long) ~iovp_mask > (long) ia64_max_iommu_merge_mask)
1807 ia64_max_iommu_merge_mask = ~iovp_mask;
1809 printk(KERN_INFO PFX
1810 "%s %d.%d HPA 0x%lx IOVA space %dMb at 0x%lx\n",
1811 ioc->name, (ioc->rev >> 4) & 0xF, ioc->rev & 0xF,
1812 hpa, ioc->iov_size >> 20, ioc->ibase);
1819 /**************************************************************************
1821 ** SBA initialization code (HW and SW)
1823 ** o identify SBA chip itself
1824 ** o FIXME: initialize DMA hints for reasonable defaults
1826 **************************************************************************/
1828 #ifdef CONFIG_PROC_FS
1830 ioc_start(struct seq_file *s, loff_t *pos)
1835 for (ioc = ioc_list; ioc; ioc = ioc->next)
1843 ioc_next(struct seq_file *s, void *v, loff_t *pos)
1845 struct ioc *ioc = v;
1852 ioc_stop(struct seq_file *s, void *v)
1857 ioc_show(struct seq_file *s, void *v)
1859 struct ioc *ioc = v;
1860 unsigned long *res_ptr = (unsigned long *)ioc->res_map;
1863 seq_printf(s, "Hewlett Packard %s IOC rev %d.%d\n",
1864 ioc->name, ((ioc->rev >> 4) & 0xF), (ioc->rev & 0xF));
1866 if (ioc->node != MAX_NUMNODES)
1867 seq_printf(s, "NUMA node : %d\n", ioc->node);
1869 seq_printf(s, "IOVA size : %ld MB\n", ((ioc->pdir_size >> 3) * iovp_size)/(1024*1024));
1870 seq_printf(s, "IOVA page size : %ld kb\n", iovp_size/1024);
1872 for (i = 0; i < (ioc->res_size / sizeof(unsigned long)); ++i, ++res_ptr)
1873 used += hweight64(*res_ptr);
1875 seq_printf(s, "PDIR size : %d entries\n", ioc->pdir_size >> 3);
1876 seq_printf(s, "PDIR used : %d entries\n", used);
1878 #ifdef PDIR_SEARCH_TIMING
1880 unsigned long i = 0, avg = 0, min, max;
1881 min = max = ioc->avg_search[0];
1882 for (i = 0; i < SBA_SEARCH_SAMPLE; i++) {
1883 avg += ioc->avg_search[i];
1884 if (ioc->avg_search[i] > max) max = ioc->avg_search[i];
1885 if (ioc->avg_search[i] < min) min = ioc->avg_search[i];
1887 avg /= SBA_SEARCH_SAMPLE;
1888 seq_printf(s, "Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles/IOVA page)\n",
1892 #ifndef ALLOW_IOV_BYPASS
1893 seq_printf(s, "IOVA bypass disabled\n");
1898 static const struct seq_operations ioc_seq_ops = {
1906 ioc_open(struct inode *inode, struct file *file)
1908 return seq_open(file, &ioc_seq_ops);
1911 static const struct file_operations ioc_fops = {
1914 .llseek = seq_lseek,
1915 .release = seq_release
1921 struct proc_dir_entry *dir, *entry;
1923 dir = proc_mkdir("bus/mckinley", NULL);
1927 entry = create_proc_entry(ioc_list->name, 0, dir);
1929 entry->proc_fops = &ioc_fops;
1934 sba_connect_bus(struct pci_bus *bus)
1936 acpi_handle handle, parent;
1940 if (!PCI_CONTROLLER(bus))
1941 panic(PFX "no sysdata on bus %d!\n", bus->number);
1943 if (PCI_CONTROLLER(bus)->iommu)
1946 handle = PCI_CONTROLLER(bus)->acpi_handle;
1951 * The IOC scope encloses PCI root bridges in the ACPI
1952 * namespace, so work our way out until we find an IOC we
1953 * claimed previously.
1956 for (ioc = ioc_list; ioc; ioc = ioc->next)
1957 if (ioc->handle == handle) {
1958 PCI_CONTROLLER(bus)->iommu = ioc;
1962 status = acpi_get_parent(handle, &parent);
1964 } while (ACPI_SUCCESS(status));
1966 printk(KERN_WARNING "No IOC for PCI Bus %04x:%02x in ACPI\n", pci_domain_nr(bus), bus->number);
1971 sba_map_ioc_to_node(struct ioc *ioc, acpi_handle handle)
1976 ioc->node = MAX_NUMNODES;
1978 pxm = acpi_get_pxm(handle);
1983 node = pxm_to_node(pxm);
1985 if (node >= MAX_NUMNODES || !node_online(node))
1992 #define sba_map_ioc_to_node(ioc, handle)
1996 acpi_sba_ioc_add(struct acpi_device *device)
2001 struct acpi_buffer buffer;
2002 struct acpi_device_info *dev_info;
2004 status = hp_acpi_csr_space(device->handle, &hpa, &length);
2005 if (ACPI_FAILURE(status))
2008 buffer.length = ACPI_ALLOCATE_LOCAL_BUFFER;
2009 status = acpi_get_object_info(device->handle, &buffer);
2010 if (ACPI_FAILURE(status))
2012 dev_info = buffer.pointer;
2015 * For HWP0001, only SBA appears in ACPI namespace. It encloses the PCI
2016 * root bridges, and its CSR space includes the IOC function.
2018 if (strncmp("HWP0001", dev_info->hardware_id.value, 7) == 0) {
2019 hpa += ZX1_IOC_OFFSET;
2020 /* zx1 based systems default to kernel page size iommu pages */
2022 iovp_shift = min(PAGE_SHIFT, 16);
2027 * default anything not caught above or specified on cmdline to 4k
2033 ioc = ioc_init(hpa, device->handle);
2037 /* setup NUMA node association */
2038 sba_map_ioc_to_node(ioc, device->handle);
2042 static const struct acpi_device_id hp_ioc_iommu_device_ids[] = {
2047 static struct acpi_driver acpi_sba_ioc_driver = {
2048 .name = "IOC IOMMU Driver",
2049 .ids = hp_ioc_iommu_device_ids,
2051 .add = acpi_sba_ioc_add,
2058 if (!ia64_platform_is("hpzx1") && !ia64_platform_is("hpzx1_swiotlb"))
2061 #if defined(CONFIG_IA64_GENERIC) && defined(CONFIG_CRASH_DUMP) && \
2062 defined(CONFIG_PROC_FS)
2063 /* If we are booting a kdump kernel, the sba_iommu will
2064 * cause devices that were not shutdown properly to MCA
2065 * as soon as they are turned back on. Our only option for
2066 * a successful kdump kernel boot is to use the swiotlb.
2068 if (elfcorehdr_addr < ELFCORE_ADDR_MAX) {
2069 if (swiotlb_late_init_with_default_size(64 * (1<<20)) != 0)
2070 panic("Unable to initialize software I/O TLB:"
2071 " Try machvec=dig boot option");
2072 machvec_init("dig");
2077 acpi_bus_register_driver(&acpi_sba_ioc_driver);
2079 #ifdef CONFIG_IA64_GENERIC
2081 * If we didn't find something sba_iommu can claim, we
2082 * need to setup the swiotlb and switch to the dig machvec.
2084 if (swiotlb_late_init_with_default_size(64 * (1<<20)) != 0)
2085 panic("Unable to find SBA IOMMU or initialize "
2086 "software I/O TLB: Try machvec=dig boot option");
2087 machvec_init("dig");
2089 panic("Unable to find SBA IOMMU: Try a generic or DIG kernel");
2094 #if defined(CONFIG_IA64_GENERIC) || defined(CONFIG_IA64_HP_ZX1_SWIOTLB)
2096 * hpzx1_swiotlb needs to have a fairly small swiotlb bounce
2097 * buffer setup to support devices with smaller DMA masks than
2098 * sba_iommu can handle.
2100 if (ia64_platform_is("hpzx1_swiotlb")) {
2101 extern void hwsw_init(void);
2109 struct pci_bus *b = NULL;
2110 while ((b = pci_find_next_bus(b)) != NULL)
2115 #ifdef CONFIG_PROC_FS
2121 subsys_initcall(sba_init); /* must be initialized after ACPI etc., but before any drivers... */
2124 nosbagart(char *str)
2126 reserve_sba_gart = 0;
2131 sba_dma_supported (struct device *dev, u64 mask)
2133 /* make sure it's at least 32bit capable */
2134 return ((mask & 0xFFFFFFFFUL) == 0xFFFFFFFFUL);
2138 sba_dma_mapping_error (dma_addr_t dma_addr)
2143 __setup("nosbagart", nosbagart);
2146 sba_page_override(char *str)
2148 unsigned long page_size;
2150 page_size = memparse(str, &str);
2151 switch (page_size) {
2156 iovp_shift = ffs(page_size) - 1;
2159 printk("%s: unknown/unsupported iommu page size %ld\n",
2160 __func__, page_size);
2166 __setup("sbapagesize=",sba_page_override);
2168 EXPORT_SYMBOL(sba_dma_mapping_error);
2169 EXPORT_SYMBOL(sba_map_single);
2170 EXPORT_SYMBOL(sba_unmap_single);
2171 EXPORT_SYMBOL(sba_map_sg);
2172 EXPORT_SYMBOL(sba_unmap_sg);
2173 EXPORT_SYMBOL(sba_dma_supported);
2174 EXPORT_SYMBOL(sba_alloc_coherent);
2175 EXPORT_SYMBOL(sba_free_coherent);