9 files changed, 215 insertions, 392 deletions

diff --git a/arch/sparc64/kernel/pci_iommu.c b/arch/sparc64/kernel/pci_iommu.c
index 425c60cfea1..a11910be101 100644
--- a/arch/sparc64/kernel/pci_iommu.c
+++ b/arch/sparc64/kernel/pci_iommu.c
@@ -49,12 +49,6 @@ static void __iommu_flushall(struct pci_iommu *iommu)
 
 	/* Ensure completion of previous PIO writes. */
 	(void) pci_iommu_read(iommu->write_complete_reg);
-
-	/* Now update everyone's flush point. */
-	for (entry = 0; entry < PBM_NCLUSTERS; entry++) {
-		iommu->alloc_info[entry].flush =
-			iommu->alloc_info[entry].next;
-	}
 }
 
 #define IOPTE_CONSISTENT(CTX) \
@@ -80,120 +74,117 @@ static void inline iopte_make_dummy(struct pci_iommu *iommu, iopte_t *iopte)
 	iopte_val(*iopte) = val;
 }
 
-void pci_iommu_table_init(struct pci_iommu *iommu, int tsbsize)
+/* Based largely upon the ppc64 iommu allocator.  */
+static long pci_arena_alloc(struct pci_iommu *iommu, unsigned long npages)
 {
-	int i;
-
-	tsbsize /= sizeof(iopte_t);
-
-	for (i = 0; i < tsbsize; i++)
-		iopte_make_dummy(iommu, &iommu->page_table[i]);
-}
-
-static iopte_t *alloc_streaming_cluster(struct pci_iommu *iommu, unsigned long npages)
-{
-	iopte_t *iopte, *limit, *first;
-	unsigned long cnum, ent, flush_point;
-
-	cnum = 0;
-	while ((1UL << cnum) < npages)
-		cnum++;
-	iopte  = (iommu->page_table +
-		  (cnum << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS)));
-
-	if (cnum == 0)
-		limit = (iommu->page_table +
-			 iommu->lowest_consistent_map);
-	else
-		limit = (iopte +
-			 (1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS)));
-
-	iopte += ((ent = iommu->alloc_info[cnum].next) << cnum);
-	flush_point = iommu->alloc_info[cnum].flush;
-	
-	first = iopte;
-	for (;;) {
-		if (IOPTE_IS_DUMMY(iommu, iopte)) {
-			if ((iopte + (1 << cnum)) >= limit)
-				ent = 0;
-			else
-				ent = ent + 1;
-			iommu->alloc_info[cnum].next = ent;
-			if (ent == flush_point)
-				__iommu_flushall(iommu);
-			break;
+	struct pci_iommu_arena *arena = &iommu->arena;
+	unsigned long n, i, start, end, limit;
+	int pass;
+
+	limit = arena->limit;
+	start = arena->hint;
+	pass = 0;
+
+again:
+	n = find_next_zero_bit(arena->map, limit, start);
+	end = n + npages;
+	if (unlikely(end >= limit)) {
+		if (likely(pass < 1)) {
+			limit = start;
+			start = 0;
+			__iommu_flushall(iommu);
+			pass++;
+			goto again;
+		} else {
+			/* Scanned the whole thing, give up. */
+			return -1;
 		}
-		iopte += (1 << cnum);
-		ent++;
-		if (iopte >= limit) {
-			iopte = (iommu->page_table +
-				 (cnum <<
-				  (iommu->page_table_sz_bits - PBM_LOGCLUSTERS)));
-			ent = 0;
+	}
+
+	for (i = n; i < end; i++) {
+		if (test_bit(i, arena->map)) {
+			start = i + 1;
+			goto again;
 		}
-		if (ent == flush_point)
-			__iommu_flushall(iommu);
-		if (iopte == first)
-			goto bad;
 	}
 
-	/* I've got your streaming cluster right here buddy boy... */
-	return iopte;
+	for (i = n; i < end; i++)
+		__set_bit(i, arena->map);
 
-bad:
-	printk(KERN_EMERG "pci_iommu: alloc_streaming_cluster of npages(%ld) failed!\n",
-	       npages);
-	return NULL;
+	arena->hint = end;
+
+	return n;
 }
 
-static void free_streaming_cluster(struct pci_iommu *iommu, dma_addr_t base,
-				   unsigned long npages, unsigned long ctx)
+static void pci_arena_free(struct pci_iommu_arena *arena, unsigned long base, unsigned long npages)
 {
-	unsigned long cnum, ent;
+	unsigned long i;
 
-	cnum = 0;
-	while ((1UL << cnum) < npages)
-		cnum++;
+	for (i = base; i < (base + npages); i++)
+		__clear_bit(i, arena->map);
+}
 
-	ent = (base << (32 - IO_PAGE_SHIFT + PBM_LOGCLUSTERS - iommu->page_table_sz_bits))
-		>> (32 + PBM_LOGCLUSTERS + cnum - iommu->page_table_sz_bits);
+void pci_iommu_table_init(struct pci_iommu *iommu, int tsbsize, u32 dma_offset, u32 dma_addr_mask)
+{
+	unsigned long i, tsbbase, order, sz, num_tsb_entries;
+
+	num_tsb_entries = tsbsize / sizeof(iopte_t);
+
+	/* Setup initial software IOMMU state. */
+	spin_lock_init(&iommu->lock);
+	iommu->ctx_lowest_free = 1;
+	iommu->page_table_map_base = dma_offset;
+	iommu->dma_addr_mask = dma_addr_mask;
+
+	/* Allocate and initialize the free area map.  */
+	sz = num_tsb_entries / 8;
+	sz = (sz + 7UL) & ~7UL;
+	iommu->arena.map = kmalloc(sz, GFP_KERNEL);
+	if (!iommu->arena.map) {
+		prom_printf("PCI_IOMMU: Error, kmalloc(arena.map) failed.\n");
+		prom_halt();
+	}
+	memset(iommu->arena.map, 0, sz);
+	iommu->arena.limit = num_tsb_entries;
 
-	/* If the global flush might not have caught this entry,
-	 * adjust the flush point such that we will flush before
-	 * ever trying to reuse it.
+	/* Allocate and initialize the dummy page which we
+	 * set inactive IO PTEs to point to.
 	 */
-#define between(X,Y,Z)	(((Z) - (Y)) >= ((X) - (Y)))
-	if (between(ent, iommu->alloc_info[cnum].next, iommu->alloc_info[cnum].flush))
-		iommu->alloc_info[cnum].flush = ent;
-#undef between
+	iommu->dummy_page = __get_free_pages(GFP_KERNEL, 0);
+	if (!iommu->dummy_page) {
+		prom_printf("PCI_IOMMU: Error, gfp(dummy_page) failed.\n");
+		prom_halt();
+	}
+	memset((void *)iommu->dummy_page, 0, PAGE_SIZE);
+	iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page);
+
+	/* Now allocate and setup the IOMMU page table itself.  */
+	order = get_order(tsbsize);
+	tsbbase = __get_free_pages(GFP_KERNEL, order);
+	if (!tsbbase) {
+		prom_printf("PCI_IOMMU: Error, gfp(tsb) failed.\n");
+		prom_halt();
+	}
+	iommu->page_table = (iopte_t *)tsbbase;
+
+	for (i = 0; i < num_tsb_entries; i++)
+		iopte_make_dummy(iommu, &iommu->page_table[i]);
 }
 
-/* We allocate consistent mappings from the end of cluster zero. */
-static iopte_t *alloc_consistent_cluster(struct pci_iommu *iommu, unsigned long npages)
+static inline iopte_t *alloc_npages(struct pci_iommu *iommu, unsigned long npages)
 {
-	iopte_t *iopte;
+	long entry;
 
-	iopte = iommu->page_table + (1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS));
-	while (iopte > iommu->page_table) {
-		iopte--;
-		if (IOPTE_IS_DUMMY(iommu, iopte)) {
-			unsigned long tmp = npages;
+	entry = pci_arena_alloc(iommu, npages);
+	if (unlikely(entry < 0))
+		return NULL;
 
-			while (--tmp) {
-				iopte--;
-				if (!IOPTE_IS_DUMMY(iommu, iopte))
-					break;
-			}
-			if (tmp == 0) {
-				u32 entry = (iopte - iommu->page_table);
+	return iommu->page_table + entry;
+}
 
-				if (entry < iommu->lowest_consistent_map)
-					iommu->lowest_consistent_map = entry;
-				return iopte;
-			}
-		}
-	}
-	return NULL;
+static inline void free_npages(struct pci_iommu *iommu, dma_addr_t base, unsigned long npages)
+{
+	pci_arena_free(&iommu->arena, base >> IO_PAGE_SHIFT, npages);
 }
 
 static int iommu_alloc_ctx(struct pci_iommu *iommu)
@@ -233,7 +224,7 @@ void *pci_alloc_consistent(struct pci_dev *pdev, size_t size, dma_addr_t *dma_ad
 	struct pcidev_cookie *pcp;
 	struct pci_iommu *iommu;
 	iopte_t *iopte;
-	unsigned long flags, order, first_page, ctx;
+	unsigned long flags, order, first_page;
 	void *ret;
 	int npages;
 
@@ -251,9 +242,10 @@ void *pci_alloc_consistent(struct pci_dev *pdev, size_t size, dma_addr_t *dma_ad
 	iommu = pcp->pbm->iommu;
 
 	spin_lock_irqsave(&iommu->lock, flags);
-	iopte = alloc_consistent_cluster(iommu, size >> IO_PAGE_SHIFT);
-	if (iopte == NULL) {
-		spin_unlock_irqrestore(&iommu->lock, flags);
+	iopte = alloc_npages(iommu, size >> IO_PAGE_SHIFT);
+	spin_unlock_irqrestore(&iommu->lock, flags);
+
+	if (unlikely(iopte == NULL)) {
 		free_pages(first_page, order);
 		return NULL;
 	}
@@ -262,31 +254,15 @@ void *pci_alloc_consistent(struct pci_dev *pdev, size_t size, dma_addr_t *dma_ad
 		      ((iopte - iommu->page_table) << IO_PAGE_SHIFT));
 	ret = (void *) first_page;
 	npages = size >> IO_PAGE_SHIFT;
-	ctx = 0;
-	if (iommu->iommu_ctxflush)
-		ctx = iommu_alloc_ctx(iommu);
 	first_page = __pa(first_page);
 	while (npages--) {
-		iopte_val(*iopte) = (IOPTE_CONSISTENT(ctx) |
+		iopte_val(*iopte) = (IOPTE_CONSISTENT(0UL) |
 				     IOPTE_WRITE |
 				     (first_page & IOPTE_PAGE));
 		iopte++;
 		first_page += IO_PAGE_SIZE;
 	}
 
-	{
-		int i;
-		u32 daddr = *dma_addrp;
-
-		npages = size >> IO_PAGE_SHIFT;
-		for (i = 0; i < npages; i++) {
-			pci_iommu_write(iommu->iommu_flush, daddr);
-			daddr += IO_PAGE_SIZE;
-		}
-	}
-
-	spin_unlock_irqrestore(&iommu->lock, flags);
-
 	return ret;
 }
 
@@ -296,7 +272,7 @@ void pci_free_consistent(struct pci_dev *pdev, size_t size, void *cpu, dma_addr_
 	struct pcidev_cookie *pcp;
 	struct pci_iommu *iommu;
 	iopte_t *iopte;
-	unsigned long flags, order, npages, i, ctx;
+	unsigned long flags, order, npages;
 
 	npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT;
 	pcp = pdev->sysdata;
@@ -306,46 +282,7 @@ void pci_free_consistent(struct pci_dev *pdev, size_t size, void *cpu, dma_addr_
 
 	spin_lock_irqsave(&iommu->lock, flags);
 
-	if ((iopte - iommu->page_table) ==
-	    iommu->lowest_consistent_map) {
-		iopte_t *walk = iopte + npages;
-		iopte_t *limit;
-
-		limit = (iommu->page_table +
-			 (1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS)));
-		while (walk < limit) {
-			if (!IOPTE_IS_DUMMY(iommu, walk))
-				break;
-			walk++;
-		}
-		iommu->lowest_consistent_map =
-			(walk - iommu->page_table);
-	}
-
-	/* Data for consistent mappings cannot enter the streaming
-	 * buffers, so we only need to update the TSB.  We flush
-	 * the IOMMU here as well to prevent conflicts with the
-	 * streaming mapping deferred tlb flush scheme.
-	 */
-
-	ctx = 0;
-	if (iommu->iommu_ctxflush)
-		ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL;
-
-	for (i = 0; i < npages; i++, iopte++)
-		iopte_make_dummy(iommu, iopte);
-
-	if (iommu->iommu_ctxflush) {
-		pci_iommu_write(iommu->iommu_ctxflush, ctx);
-	} else {
-		for (i = 0; i < npages; i++) {
-			u32 daddr = dvma + (i << IO_PAGE_SHIFT);
-
-			pci_iommu_write(iommu->iommu_flush, daddr);
-		}
-	}
-
-	iommu_free_ctx(iommu, ctx);
+	free_npages(iommu, dvma, npages);
 
 	spin_unlock_irqrestore(&iommu->lock, flags);
 
@@ -372,25 +309,27 @@ dma_addr_t pci_map_single(struct pci_dev *pdev, void *ptr, size_t sz, int direct
 	iommu = pcp->pbm->iommu;
 	strbuf = &pcp->pbm->stc;
 
-	if (direction == PCI_DMA_NONE)
-		BUG();
+	if (unlikely(direction == PCI_DMA_NONE))
+		goto bad_no_ctx;
 
 	oaddr = (unsigned long)ptr;
 	npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK);
 	npages >>= IO_PAGE_SHIFT;
 
 	spin_lock_irqsave(&iommu->lock, flags);
+	base = alloc_npages(iommu, npages);
+	ctx = 0;
+	if (iommu->iommu_ctxflush)
+		ctx = iommu_alloc_ctx(iommu);
+	spin_unlock_irqrestore(&iommu->lock, flags);
 
-	base = alloc_streaming_cluster(iommu, npages);
-	if (base == NULL)
+	if (unlikely(!base))
 		goto bad;
+
 	bus_addr = (iommu->page_table_map_base +
 		    ((base - iommu->page_table) << IO_PAGE_SHIFT));
 	ret = bus_addr | (oaddr & ~IO_PAGE_MASK);
 	base_paddr = __pa(oaddr & IO_PAGE_MASK);
-	ctx = 0;
-	if (iommu->iommu_ctxflush)
-		ctx = iommu_alloc_ctx(iommu);
 	if (strbuf->strbuf_enabled)
 		iopte_protection = IOPTE_STREAMING(ctx);
 	else
@@ -401,12 +340,13 @@ dma_addr_t pci_map_single(struct pci_dev *pdev, void *ptr, size_t sz, int direct
 	for (i = 0; i < npages; i++, base++, base_paddr += IO_PAGE_SIZE)
 		iopte_val(*base) = iopte_protection | base_paddr;
 
-	spin_unlock_irqrestore(&iommu->lock, flags);
-
 	return ret;
 
 bad:
-	spin_unlock_irqrestore(&iommu->lock, flags);
+	iommu_free_ctx(iommu, ctx);
+bad_no_ctx:
+	if (printk_ratelimit())
+		WARN_ON(1);
 	return PCI_DMA_ERROR_CODE;
 }
 
@@ -481,10 +421,13 @@ void pci_unmap_single(struct pci_dev *pdev, dma_addr_t bus_addr, size_t sz, int
 	struct pci_iommu *iommu;
 	struct pci_strbuf *strbuf;
 	iopte_t *base;
-	unsigned long flags, npages, ctx;
+	unsigned long flags, npages, ctx, i;
 
-	if (direction == PCI_DMA_NONE)
-		BUG();
+	if (unlikely(direction == PCI_DMA_NONE)) {
+		if (printk_ratelimit())
+			WARN_ON(1);
+		return;
+	}
 
 	pcp = pdev->sysdata;
 	iommu = pcp->pbm->iommu;
@@ -510,13 +453,14 @@ void pci_unmap_single(struct pci_dev *pdev, dma_addr_t bus_addr, size_t sz, int
 
 	/* Step 1: Kick data out of streaming buffers if necessary. */
 	if (strbuf->strbuf_enabled)
-		pci_strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
+		pci_strbuf_flush(strbuf, iommu, bus_addr, ctx,
+				 npages, direction);
 
-	/* Step 2: Clear out first TSB entry. */
-	iopte_make_dummy(iommu, base);
+	/* Step 2: Clear out TSB entries. */
+	for (i = 0; i < npages; i++)
+		iopte_make_dummy(iommu, base + i);
 
-	free_streaming_cluster(iommu, bus_addr - iommu->page_table_map_base,
-			       npages, ctx);
+	free_npages(iommu, bus_addr - iommu->page_table_map_base, npages);
 
 	iommu_free_ctx(iommu, ctx);
 
@@ -621,6 +565,8 @@ int pci_map_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int
 			pci_map_single(pdev,
 				       (page_address(sglist->page) + sglist->offset),
 				       sglist->length, direction);
+		if (unlikely(sglist->dma_address == PCI_DMA_ERROR_CODE))
+			return 0;
 		sglist->dma_length = sglist->length;
 		return 1;
 	}
@@ -629,21 +575,29 @@ int pci_map_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int
 	iommu = pcp->pbm->iommu;
 	strbuf = &pcp->pbm->stc;
 	
-	if (direction == PCI_DMA_NONE)
-		BUG();
+	if (unlikely(direction == PCI_DMA_NONE))
+		goto bad_no_ctx;
 
 	/* Step 1: Prepare scatter list. */
 
 	npages = prepare_sg(sglist, nelems);
 
-	/* Step 2: Allocate a cluster. */
+	/* Step 2: Allocate a cluster and context, if necessary. */
 
 	spin_lock_irqsave(&iommu->lock, flags);
 
-	base = alloc_streaming_cluster(iommu, npages);
+	base = alloc_npages(iommu, npages);
+	ctx = 0;
+	if (iommu->iommu_ctxflush)
+		ctx = iommu_alloc_ctx(iommu);
+
+	spin_unlock_irqrestore(&iommu->lock, flags);
+
 	if (base == NULL)
 		goto bad;
-	dma_base = iommu->page_table_map_base + ((base - iommu->page_table) << IO_PAGE_SHIFT);
+
+	dma_base = iommu->page_table_map_base +
+		((base - iommu->page_table) << IO_PAGE_SHIFT);
 
 	/* Step 3: Normalize DMA addresses. */
 	used = nelems;
@@ -656,30 +610,28 @@ int pci_map_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int
 	}
 	used = nelems - used;
 
-	/* Step 4: Choose a context if necessary. */
-	ctx = 0;
-	if (iommu->iommu_ctxflush)
-		ctx = iommu_alloc_ctx(iommu);
-
-	/* Step 5: Create the mappings. */
+	/* Step 4: Create the mappings. */
 	if (strbuf->strbuf_enabled)
 		iopte_protection = IOPTE_STREAMING(ctx);
 	else
 		iopte_protection = IOPTE_CONSISTENT(ctx);
 	if (direction != PCI_DMA_TODEVICE)
 		iopte_protection |= IOPTE_WRITE;
-	fill_sg (base, sglist, used, nelems, iopte_protection);
+
+	fill_sg(base, sglist, used, nelems, iopte_protection);
+
 #ifdef VERIFY_SG
 	verify_sglist(sglist, nelems, base, npages);
 #endif
 
-	spin_unlock_irqrestore(&iommu->lock, flags);
-
 	return used;
 
 bad:
-	spin_unlock_irqrestore(&iommu->lock, flags);
-	return PCI_DMA_ERROR_CODE;
+	iommu_free_ctx(iommu, ctx);
+bad_no_ctx:
+	if (printk_ratelimit())
+		WARN_ON(1);
+	return 0;
 }
 
 /* Unmap a set of streaming mode DMA translations. */
@@ -692,8 +644,10 @@ void pci_unmap_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems,
 	unsigned long flags, ctx, i, npages;
 	u32 bus_addr;
 
-	if (direction == PCI_DMA_NONE)
-		BUG();
+	if (unlikely(direction == PCI_DMA_NONE)) {
+		if (printk_ratelimit())
+			WARN_ON(1);
+	}
 
 	pcp = pdev->sysdata;
 	iommu = pcp->pbm->iommu;
@@ -705,7 +659,8 @@ void pci_unmap_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems,
 		if (sglist[i].dma_length == 0)
 			break;
 	i--;
-	npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length) - bus_addr) >> IO_PAGE_SHIFT;
+	npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length) -
+		  bus_addr) >> IO_PAGE_SHIFT;
 
 	base = iommu->page_table +
 		((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
@@ -726,11 +681,11 @@ void pci_unmap_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems,
 	if (strbuf->strbuf_enabled)
 		pci_strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
 
-	/* Step 2: Clear out first TSB entry. */
-	iopte_make_dummy(iommu, base);
+	/* Step 2: Clear out the TSB entries. */
+	for (i = 0; i < npages; i++)
+		iopte_make_dummy(iommu, base + i);
 
-	free_streaming_cluster(iommu, bus_addr - iommu->page_table_map_base,
-			       npages, ctx);
+	free_npages(iommu, bus_addr - iommu->page_table_map_base, npages);
 
 	iommu_free_ctx(iommu, ctx);
 
diff --git a/arch/sparc64/kernel/pci_psycho.c b/arch/sparc64/kernel/pci_psycho.c
index 6ed1ef25e0a..c03ed5f49d3 100644
--- a/arch/sparc64/kernel/pci_psycho.c
+++ b/arch/sparc64/kernel/pci_psycho.c
@@ -1207,13 +1207,9 @@ static void psycho_scan_bus(struct pci_controller_info *p)
 static void psycho_iommu_init(struct pci_controller_info *p)
 {
 	struct pci_iommu *iommu = p->pbm_A.iommu;
-	unsigned long tsbbase, i;
+	unsigned long i;
 	u64 control;
 
-	/* Setup initial software IOMMU state. */
-	spin_lock_init(&iommu->lock);
-	iommu->ctx_lowest_free = 1;
-
 	/* Register addresses. */
 	iommu->iommu_control  = p->pbm_A.controller_regs + PSYCHO_IOMMU_CONTROL;
 	iommu->iommu_tsbbase  = p->pbm_A.controller_regs + PSYCHO_IOMMU_TSBBASE;
@@ -1240,40 +1236,10 @@ static void psycho_iommu_init(struct pci_controller_info *p)
 	/* Leave diag mode enabled for full-flushing done
 	 * in pci_iommu.c
 	 */
+	pci_iommu_table_init(iommu, IO_TSB_SIZE, 0xc0000000, 0xffffffff);
 
-	iommu->dummy_page = __get_free_pages(GFP_KERNEL, 0);
-	if (!iommu->dummy_page) {
-		prom_printf("PSYCHO_IOMMU: Error, gfp(dummy_page) failed.\n");
-		prom_halt();
-	}
-	memset((void *)iommu->dummy_page, 0, PAGE_SIZE);
-	iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page);
-
-	/* Using assumed page size 8K with 128K entries we need 1MB iommu page
-	 * table (128K ioptes * 8 bytes per iopte).  This is
-	 * page order 7 on UltraSparc.
-	 */
-	tsbbase = __get_free_pages(GFP_KERNEL, get_order(IO_TSB_SIZE));
-	if (!tsbbase) {
-		prom_printf("PSYCHO_IOMMU: Error, gfp(tsb) failed.\n");
-		prom_halt();
-	}
-	iommu->page_table = (iopte_t *)tsbbase;
-	iommu->page_table_sz_bits = 17;
-	iommu->page_table_map_base = 0xc0000000;
-	iommu->dma_addr_mask = 0xffffffff;
-	pci_iommu_table_init(iommu, IO_TSB_SIZE);
-
-	/* We start with no consistent mappings. */
-	iommu->lowest_consistent_map =
-		1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS);
-
-	for (i = 0; i < PBM_NCLUSTERS; i++) {
-		iommu->alloc_info[i].flush = 0;
-		iommu->alloc_info[i].next = 0;
-	}
-
-	psycho_write(p->pbm_A.controller_regs + PSYCHO_IOMMU_TSBBASE, __pa(tsbbase));
+	psycho_write(p->pbm_A.controller_regs + PSYCHO_IOMMU_TSBBASE,
+		     __pa(iommu->page_table));
 
 	control = psycho_read(p->pbm_A.controller_regs + PSYCHO_IOMMU_CONTROL);
 	control &= ~(PSYCHO_IOMMU_CTRL_TSBSZ | PSYCHO_IOMMU_CTRL_TBWSZ);
@@ -1281,7 +1247,7 @@ static void psycho_iommu_init(struct pci_controller_info *p)
 	psycho_write(p->pbm_A.controller_regs + PSYCHO_IOMMU_CONTROL, control);
 
 	/* If necessary, hook us up for starfire IRQ translations. */
-	if(this_is_starfire)
+	if (this_is_starfire)
 		p->starfire_cookie = starfire_hookup(p->pbm_A.portid);
 	else
 		p->starfire_cookie = NULL;
diff --git a/arch/sparc64/kernel/pci_sabre.c b/arch/sparc64/kernel/pci_sabre.c
index 0ee6bd5b9ac..da8e1364194 100644
--- a/arch/sparc64/kernel/pci_sabre.c
+++ b/arch/sparc64/kernel/pci_sabre.c
@@ -1267,13 +1267,9 @@ static void sabre_iommu_init(struct pci_controller_info *p,
 			     u32 dma_mask)
 {
 	struct pci_iommu *iommu = p->pbm_A.iommu;
-	unsigned long tsbbase, i, order;
+	unsigned long i;
 	u64 control;
 
-	/* Setup initial software IOMMU state. */
-	spin_lock_init(&iommu->lock);
-	iommu->ctx_lowest_free = 1;
-
 	/* Register addresses. */
 	iommu->iommu_control  = p->pbm_A.controller_regs + SABRE_IOMMU_CONTROL;
 	iommu->iommu_tsbbase  = p->pbm_A.controller_regs + SABRE_IOMMU_TSBBASE;
@@ -1295,26 +1291,10 @@ static void sabre_iommu_init(struct pci_controller_info *p,
 	/* Leave diag mode enabled for full-flushing done
 	 * in pci_iommu.c
 	 */
+	pci_iommu_table_init(iommu, tsbsize * 1024 * 8, dvma_offset, dma_mask);
 
-	iommu->dummy_page = __get_free_pages(GFP_KERNEL, 0);
-	if (!iommu->dummy_page) {
-		prom_printf("PSYCHO_IOMMU: Error, gfp(dummy_page) failed.\n");
-		prom_halt();
-	}
-	memset((void *)iommu->dummy_page, 0, PAGE_SIZE);
-	iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page);
-
-	tsbbase = __get_free_pages(GFP_KERNEL, order = get_order(tsbsize * 1024 * 8));
-	if (!tsbbase) {
-		prom_printf("SABRE_IOMMU: Error, gfp(tsb) failed.\n");
-		prom_halt();
-	}
-	iommu->page_table = (iopte_t *)tsbbase;
-	iommu->page_table_map_base = dvma_offset;
-	iommu->dma_addr_mask = dma_mask;
-	pci_iommu_table_init(iommu, PAGE_SIZE << order);
-
-	sabre_write(p->pbm_A.controller_regs + SABRE_IOMMU_TSBBASE, __pa(tsbbase));
+	sabre_write(p->pbm_A.controller_regs + SABRE_IOMMU_TSBBASE,
+		    __pa(iommu->page_table));
 
 	control = sabre_read(p->pbm_A.controller_regs + SABRE_IOMMU_CONTROL);
 	control &= ~(SABRE_IOMMUCTRL_TSBSZ | SABRE_IOMMUCTRL_TBWSZ);
@@ -1322,11 +1302,9 @@ static void sabre_iommu_init(struct pci_controller_info *p,
 	switch(tsbsize) {
 	case 64:
 		control |= SABRE_IOMMU_TSBSZ_64K;
-		iommu->page_table_sz_bits = 16;
 		break;
 	case 128:
 		control |= SABRE_IOMMU_TSBSZ_128K;
-		iommu->page_table_sz_bits = 17;
 		break;
 	default:
 		prom_printf("iommu_init: Illegal TSB size %d\n", tsbsize);
@@ -1334,15 +1312,6 @@ static void sabre_iommu_init(struct pci_controller_info *p,
 		break;
 	}
 	sabre_write(p->pbm_A.controller_regs + SABRE_IOMMU_CONTROL, control);
-
-	/* We start with no consistent mappings. */
-	iommu->lowest_consistent_map =
-		1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS);
-
-	for (i = 0; i < PBM_NCLUSTERS; i++) {
-		iommu->alloc_info[i].flush = 0;
-		iommu->alloc_info[i].next = 0;
-	}
 }
 
 static void pbm_register_toplevel_resources(struct pci_controller_info *p,
diff --git a/arch/sparc64/kernel/pci_schizo.c b/arch/sparc64/kernel/pci_schizo.c
index cae5b61fe2f..d8c4e0919b4 100644
--- a/arch/sparc64/kernel/pci_schizo.c
+++ b/arch/sparc64/kernel/pci_schizo.c
@@ -1765,7 +1765,7 @@ static void schizo_pbm_strbuf_init(struct pci_pbm_info *pbm)
 static void schizo_pbm_iommu_init(struct pci_pbm_info *pbm)
 {
 	struct pci_iommu *iommu = pbm->iommu;
-	unsigned long tsbbase, i, tagbase, database, order;
+	unsigned long i, tagbase, database;
 	u32 vdma[2], dma_mask;
 	u64 control;
 	int err, tsbsize;
@@ -1800,10 +1800,6 @@ static void schizo_pbm_iommu_init(struct pci_pbm_info *pbm)
 			prom_halt();
 	};
 
-	/* Setup initial software IOMMU state. */
-	spin_lock_init(&iommu->lock);
-	iommu->ctx_lowest_free = 1;
-
 	/* Register addresses, SCHIZO has iommu ctx flushing. */
 	iommu->iommu_control  = pbm->pbm_regs + SCHIZO_IOMMU_CONTROL;
 	iommu->iommu_tsbbase  = pbm->pbm_regs + SCHIZO_IOMMU_TSBBASE;
@@ -1832,56 +1828,9 @@ static void schizo_pbm_iommu_init(struct pci_pbm_info *pbm)
 	/* Leave diag mode enabled for full-flushing done
 	 * in pci_iommu.c
 	 */
+	pci_iommu_table_init(iommu, tsbsize * 8 * 1024, vdma[0], dma_mask);
 
-	iommu->dummy_page = __get_free_pages(GFP_KERNEL, 0);
-	if (!iommu->dummy_page) {
-		prom_printf("PSYCHO_IOMMU: Error, gfp(dummy_page) failed.\n");
-		prom_halt();
-	}
-	memset((void *)iommu->dummy_page, 0, PAGE_SIZE);
-	iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page);
-
-	/* Using assumed page size 8K with 128K entries we need 1MB iommu page
-	 * table (128K ioptes * 8 bytes per iopte).  This is
-	 * page order 7 on UltraSparc.
-	 */
-	order = get_order(tsbsize * 8 * 1024);
-	tsbbase = __get_free_pages(GFP_KERNEL, order);
-	if (!tsbbase) {
-		prom_printf("%s: Error, gfp(tsb) failed.\n", pbm->name);
-		prom_halt();
-	}
-
-	iommu->page_table = (iopte_t *)tsbbase;
-	iommu->page_table_map_base = vdma[0];
-	iommu->dma_addr_mask = dma_mask;
-	pci_iommu_table_init(iommu, PAGE_SIZE << order);
-
-	switch (tsbsize) {
-	case 64:
-		iommu->page_table_sz_bits = 16;
-		break;
-
-	case 128:
-		iommu->page_table_sz_bits = 17;
-		break;
-
-	default:
-		prom_printf("iommu_init: Illegal TSB size %d\n", tsbsize);
-		prom_halt();
-		break;
-	};
-
-	/* We start with no consistent mappings. */
-	iommu->lowest_consistent_map =
-		1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS);
-
-	for (i = 0; i < PBM_NCLUSTERS; i++) {
-		iommu->alloc_info[i].flush = 0;
-		iommu->alloc_info[i].next = 0;
-	}
-
-	schizo_write(iommu->iommu_tsbbase, __pa(tsbbase));
+	schizo_write(iommu->iommu_tsbbase, __pa(iommu->page_table));
 
 	control = schizo_read(iommu->iommu_control);
 	control &= ~(SCHIZO_IOMMU_CTRL_TSBSZ | SCHIZO_IOMMU_CTRL_TBWSZ);
diff --git a/arch/sparc64/kernel/smp.c b/arch/sparc64/kernel/smp.c
index 590df5a16f5..b137fd63f5e 100644
--- a/arch/sparc64/kernel/smp.c
+++ b/arch/sparc64/kernel/smp.c
@@ -1001,13 +1001,6 @@ void smp_penguin_jailcell(int irq, struct pt_regs *regs)
 	preempt_enable();
 }
 
-extern unsigned long xcall_promstop;
-
-void smp_promstop_others(void)
-{
-	smp_cross_call(&xcall_promstop, 0, 0, 0);
-}
-
 #define prof_multiplier(__cpu)		cpu_data(__cpu).multiplier
 #define prof_counter(__cpu)		cpu_data(__cpu).counter
 
diff --git a/arch/sparc64/mm/ultra.S b/arch/sparc64/mm/ultra.S
index 058b8126c1a..e4c9151fa11 100644
--- a/arch/sparc64/mm/ultra.S
+++ b/arch/sparc64/mm/ultra.S
@@ -453,22 +453,6 @@ xcall_flush_dcache_page_spitfire: /* %g1 == physical page address
 	nop
 	nop
 
-	.globl		xcall_promstop
-xcall_promstop:
-	rdpr		%pstate, %g2
-	wrpr		%g2, PSTATE_IG | PSTATE_AG, %pstate
-	rdpr		%pil, %g2
-	wrpr		%g0, 15, %pil
-	sethi		%hi(109f), %g7
-	b,pt		%xcc, etrap_irq
-109:	 or		%g7, %lo(109b), %g7
-	flushw
-	call		prom_stopself
-	 nop
-	/* We should not return, just spin if we do... */
-1:	b,a,pt		%xcc, 1b
-	nop
-
 	.data
 
 errata32_hwbug:
diff --git a/arch/sparc64/prom/misc.c b/arch/sparc64/prom/misc.c
index 9b895faf077..87f5cfce23b 100644
--- a/arch/sparc64/prom/misc.c
+++ b/arch/sparc64/prom/misc.c
@@ -68,19 +68,11 @@ void prom_cmdline(void)
 	local_irq_restore(flags);
 }
 
-#ifdef CONFIG_SMP
-extern void smp_promstop_others(void);
-#endif
-
 /* Drop into the prom, but completely terminate the program.
  * No chance of continuing.
  */
 void prom_halt(void)
 {
-#ifdef CONFIG_SMP
-	smp_promstop_others();
-	udelay(8000);
-#endif
 again:
 	p1275_cmd("exit", P1275_INOUT(0, 0));
 	goto again; /* PROM is out to get me -DaveM */
@@ -88,10 +80,6 @@ again:
 
 void prom_halt_power_off(void)
 {
-#ifdef CONFIG_SMP
-	smp_promstop_others();
-	udelay(8000);
-#endif
 	p1275_cmd("SUNW,power-off", P1275_INOUT(0, 0));
 
 	/* if nothing else helps, we just halt */
diff --git a/drivers/scsi/qlogicpti.c b/drivers/scsi/qlogicpti.c
index a917ab7475a..1fd5fc6d0fe 100644
--- a/drivers/scsi/qlogicpti.c
+++ b/drivers/scsi/qlogicpti.c
@@ -1119,6 +1119,36 @@ static inline void update_can_queue(struct Scsi_Host *host, u_int in_ptr, u_int
 	host->sg_tablesize = QLOGICPTI_MAX_SG(num_free);
 }
 
+static unsigned int scsi_rbuf_get(struct scsi_cmnd *cmd, unsigned char **buf_out)
+{
+	unsigned char *buf;
+	unsigned int buflen;
+
+	if (cmd->use_sg) {
+		struct scatterlist *sg;
+
+		sg = (struct scatterlist *) cmd->request_buffer;
+		buf = kmap_atomic(sg->page, KM_IRQ0) + sg->offset;
+		buflen = sg->length;
+	} else {
+		buf = cmd->request_buffer;
+		buflen = cmd->request_bufflen;
+	}
+
+	*buf_out = buf;
+	return buflen;
+}
+
+static void scsi_rbuf_put(struct scsi_cmnd *cmd, unsigned char *buf)
+{
+	if (cmd->use_sg) {
+		struct scatterlist *sg;
+
+		sg = (struct scatterlist *) cmd->request_buffer;
+		kunmap_atomic(buf - sg->offset, KM_IRQ0);
+	}
+}
+
 /*
  * Until we scan the entire bus with inquiries, go throught this fella...
  */
@@ -1145,11 +1175,9 @@ static void ourdone(struct scsi_cmnd *Cmnd)
 		int ok = host_byte(Cmnd->result) == DID_OK;
 		if (Cmnd->cmnd[0] == 0x12 && ok) {
 			unsigned char *iqd;
+			unsigned int iqd_len;
 
-			if (Cmnd->use_sg != 0)
-				BUG();
-
-			iqd = ((unsigned char *)Cmnd->buffer);
+			iqd_len = scsi_rbuf_get(Cmnd, &iqd);
 
 			/* tags handled in midlayer */
 			/* enable sync mode? */
@@ -1163,6 +1191,9 @@ static void ourdone(struct scsi_cmnd *Cmnd)
 			if (iqd[7] & 0x20) {
 				qpti->dev_param[tgt].device_flags |= 0x20;
 			}
+
+			scsi_rbuf_put(Cmnd, iqd);
+
 			qpti->sbits |= (1 << tgt);
 		} else if (!ok) {
 			qpti->sbits |= (1 << tgt);
diff --git a/include/asm-sparc64/pbm.h b/include/asm-sparc64/pbm.h
index 38bbbccb406..dd35a2c7798 100644
--- a/include/asm-sparc64/pbm.h
+++ b/include/asm-sparc64/pbm.h
@@ -27,23 +27,27 @@
  * PCI bus.
  */
 
-#define PBM_LOGCLUSTERS 3
-#define PBM_NCLUSTERS (1 << PBM_LOGCLUSTERS)
-
 struct pci_controller_info;
 
 /* This contains the software state necessary to drive a PCI
  * controller's IOMMU.
  */
+struct pci_iommu_arena {
+	unsigned long	*map;
+	unsigned int	hint;
+	unsigned int	limit;
+};
+
 struct pci_iommu {
 	/* This protects the controller's IOMMU and all
 	 * streaming buffers underneath.
 	 */
 	spinlock_t	lock;
 
+	struct pci_iommu_arena arena;
+
 	/* IOMMU page table, a linear array of ioptes. */
 	iopte_t		*page_table;		/* The page table itself. */
-	int		page_table_sz_bits;	/* log2 of ow many pages does it map? */
 
 	/* Base PCI memory space address where IOMMU mappings
 	 * begin.
@@ -62,12 +66,6 @@ struct pci_iommu {
 	 */
 	unsigned long	write_complete_reg;
 
-	/* The lowest used consistent mapping entry.  Since
-	 * we allocate consistent maps out of cluster 0 this
-	 * is relative to the beginning of closter 0.
-	 */
-	u32		lowest_consistent_map;
-
 	/* In order to deal with some buggy third-party PCI bridges that
 	 * do wrong prefetching, we never mark valid mappings as invalid.
 	 * Instead we point them at this dummy page.
@@ -75,16 +73,6 @@ struct pci_iommu {
 	unsigned long	dummy_page;
 	unsigned long	dummy_page_pa;
 
-	/* If PBM_NCLUSTERS is ever decreased to 4 or lower,
-	 * or if largest supported page_table_sz * 8K goes above
-	 * 2GB, you must increase the size of the type of
-	 * these counters.  You have been duly warned. -DaveM
-	 */
-	struct {
-		u16	next;
-		u16	flush;
-	} alloc_info[PBM_NCLUSTERS];
-
 	/* CTX allocation. */
 	unsigned long ctx_lowest_free;
 	unsigned long ctx_bitmap[IOMMU_NUM_CTXS / (sizeof(unsigned long) * 8)];
@@ -102,7 +90,7 @@ struct pci_iommu {
 	u32 dma_addr_mask;
 };
 
-extern void pci_iommu_table_init(struct pci_iommu *, int);
+extern void pci_iommu_table_init(struct pci_iommu *iommu, int tsbsize, u32 dma_offset, u32 dma_addr_mask);
 
 /* This describes a PCI bus module's streaming buffer. */
 struct pci_strbuf {