USB: ohci: quirk AMD prefetch for USB 1.1 ISO transfer
[safe/jmp/linux-2.6] / drivers / pci / iova.c
index 717fafa..7914951 100644 (file)
@@ -1,27 +1,37 @@
 /*
- * Copyright (c) 2006, Intel Corporation.
+ * Copyright © 2006-2009, Intel Corporation.
  *
- * This file is released under the GPLv2.
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
  *
- * Copyright (C) 2006 Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
+ * Place - Suite 330, Boston, MA 02111-1307 USA.
+ *
+ * Author: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
  */
 
-#include "iova.h"
+#include <linux/iova.h>
 
 void
-init_iova_domain(struct iova_domain *iovad)
+init_iova_domain(struct iova_domain *iovad, unsigned long pfn_32bit)
 {
-       spin_lock_init(&iovad->iova_alloc_lock);
        spin_lock_init(&iovad->iova_rbtree_lock);
        iovad->rbroot = RB_ROOT;
        iovad->cached32_node = NULL;
-
+       iovad->dma_32bit_pfn = pfn_32bit;
 }
 
 static struct rb_node *
 __get_cached_rbnode(struct iova_domain *iovad, unsigned long *limit_pfn)
 {
-       if ((*limit_pfn != DMA_32BIT_PFN) ||
+       if ((*limit_pfn != iovad->dma_32bit_pfn) ||
                (iovad->cached32_node == NULL))
                return rb_last(&iovad->rbroot);
        else {
@@ -37,7 +47,7 @@ static void
 __cached_rbnode_insert_update(struct iova_domain *iovad,
        unsigned long limit_pfn, struct iova *new)
 {
-       if (limit_pfn != DMA_32BIT_PFN)
+       if (limit_pfn != iovad->dma_32bit_pfn)
                return;
        iovad->cached32_node = &new->node;
 }
@@ -57,39 +67,103 @@ __cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
                iovad->cached32_node = rb_next(&free->node);
 }
 
-static int __alloc_iova_range(struct iova_domain *iovad,
-       unsigned long size, unsigned long limit_pfn, struct iova *new)
+/* Computes the padding size required, to make the
+ * the start address naturally aligned on its size
+ */
+static int
+iova_get_pad_size(int size, unsigned int limit_pfn)
 {
-       struct rb_node *curr = NULL;
+       unsigned int pad_size = 0;
+       unsigned int order = ilog2(size);
+
+       if (order)
+               pad_size = (limit_pfn + 1) % (1 << order);
+
+       return pad_size;
+}
+
+static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
+               unsigned long size, unsigned long limit_pfn,
+                       struct iova *new, bool size_aligned)
+{
+       struct rb_node *prev, *curr = NULL;
        unsigned long flags;
        unsigned long saved_pfn;
+       unsigned int pad_size = 0;
 
        /* Walk the tree backwards */
        spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
        saved_pfn = limit_pfn;
        curr = __get_cached_rbnode(iovad, &limit_pfn);
+       prev = curr;
        while (curr) {
                struct iova *curr_iova = container_of(curr, struct iova, node);
+
                if (limit_pfn < curr_iova->pfn_lo)
                        goto move_left;
-               if (limit_pfn < curr_iova->pfn_hi)
+               else if (limit_pfn < curr_iova->pfn_hi)
                        goto adjust_limit_pfn;
-               if ((curr_iova->pfn_hi + size) <= limit_pfn)
-                       break;  /* found a free slot */
+               else {
+                       if (size_aligned)
+                               pad_size = iova_get_pad_size(size, limit_pfn);
+                       if ((curr_iova->pfn_hi + size + pad_size) <= limit_pfn)
+                               break;  /* found a free slot */
+               }
 adjust_limit_pfn:
                limit_pfn = curr_iova->pfn_lo - 1;
 move_left:
+               prev = curr;
                curr = rb_prev(curr);
        }
 
-       if ((!curr) && !(IOVA_START_PFN + size <= limit_pfn)) {
-               spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
-               return -ENOMEM;
+       if (!curr) {
+               if (size_aligned)
+                       pad_size = iova_get_pad_size(size, limit_pfn);
+               if ((IOVA_START_PFN + size + pad_size) > limit_pfn) {
+                       spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
+                       return -ENOMEM;
+               }
        }
-       new->pfn_hi = limit_pfn;
-       new->pfn_lo = limit_pfn - size + 1;
+
+       /* pfn_lo will point to size aligned address if size_aligned is set */
+       new->pfn_lo = limit_pfn - (size + pad_size) + 1;
+       new->pfn_hi = new->pfn_lo + size - 1;
+
+       /* Insert the new_iova into domain rbtree by holding writer lock */
+       /* Add new node and rebalance tree. */
+       {
+               struct rb_node **entry, *parent = NULL;
+
+               /* If we have 'prev', it's a valid place to start the
+                  insertion. Otherwise, start from the root. */
+               if (prev)
+                       entry = &prev;
+               else
+                       entry = &iovad->rbroot.rb_node;
+
+               /* Figure out where to put new node */
+               while (*entry) {
+                       struct iova *this = container_of(*entry,
+                                                       struct iova, node);
+                       parent = *entry;
+
+                       if (new->pfn_lo < this->pfn_lo)
+                               entry = &((*entry)->rb_left);
+                       else if (new->pfn_lo > this->pfn_lo)
+                               entry = &((*entry)->rb_right);
+                       else
+                               BUG(); /* this should not happen */
+               }
+
+               /* Add new node and rebalance tree. */
+               rb_link_node(&new->node, parent, entry);
+               rb_insert_color(&new->node, &iovad->rbroot);
+       }
+       __cached_rbnode_insert_update(iovad, saved_pfn, new);
 
        spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
+
+
        return 0;
 }
 
@@ -119,14 +193,17 @@ iova_insert_rbtree(struct rb_root *root, struct iova *iova)
  * @iovad - iova domain in question
  * @size - size of page frames to allocate
  * @limit_pfn - max limit address
+ * @size_aligned - set if size_aligned address range is required
  * This function allocates an iova in the range limit_pfn to IOVA_START_PFN
- * looking from limit_pfn instead from IOVA_START_PFN.
+ * looking from limit_pfn instead from IOVA_START_PFN. If the size_aligned
+ * flag is set then the allocated address iova->pfn_lo will be naturally
+ * aligned on roundup_power_of_two(size).
  */
 struct iova *
 alloc_iova(struct iova_domain *iovad, unsigned long size,
-       unsigned long limit_pfn)
+       unsigned long limit_pfn,
+       bool size_aligned)
 {
-       unsigned long flags;
        struct iova *new_iova;
        int ret;
 
@@ -134,23 +211,20 @@ alloc_iova(struct iova_domain *iovad, unsigned long size,
        if (!new_iova)
                return NULL;
 
-       spin_lock_irqsave(&iovad->iova_alloc_lock, flags);
-       ret = __alloc_iova_range(iovad, size, limit_pfn, new_iova);
+       /* If size aligned is set then round the size to
+        * to next power of two.
+        */
+       if (size_aligned)
+               size = __roundup_pow_of_two(size);
+
+       ret = __alloc_and_insert_iova_range(iovad, size, limit_pfn,
+                       new_iova, size_aligned);
 
        if (ret) {
-               spin_unlock_irqrestore(&iovad->iova_alloc_lock, flags);
                free_iova_mem(new_iova);
                return NULL;
        }
 
-       /* Insert the new_iova into domain rbtree by holding writer lock */
-       spin_lock(&iovad->iova_rbtree_lock);
-       iova_insert_rbtree(&iovad->rbroot, new_iova);
-       __cached_rbnode_insert_update(iovad, limit_pfn, new_iova);
-       spin_unlock(&iovad->iova_rbtree_lock);
-
-       spin_unlock_irqrestore(&iovad->iova_alloc_lock, flags);
-
        return new_iova;
 }
 
@@ -303,8 +377,7 @@ reserve_iova(struct iova_domain *iovad,
        struct iova *iova;
        unsigned int overlap = 0;
 
-       spin_lock_irqsave(&iovad->iova_alloc_lock, flags);
-       spin_lock(&iovad->iova_rbtree_lock);
+       spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
        for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
                if (__is_range_overlap(node, pfn_lo, pfn_hi)) {
                        iova = container_of(node, struct iova, node);
@@ -324,8 +397,7 @@ reserve_iova(struct iova_domain *iovad,
        iova = __insert_new_range(iovad, pfn_lo, pfn_hi);
 finish:
 
-       spin_unlock(&iovad->iova_rbtree_lock);
-       spin_unlock_irqrestore(&iovad->iova_alloc_lock, flags);
+       spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
        return iova;
 }
 
@@ -342,8 +414,7 @@ copy_reserved_iova(struct iova_domain *from, struct iova_domain *to)
        unsigned long flags;
        struct rb_node *node;
 
-       spin_lock_irqsave(&from->iova_alloc_lock, flags);
-       spin_lock(&from->iova_rbtree_lock);
+       spin_lock_irqsave(&from->iova_rbtree_lock, flags);
        for (node = rb_first(&from->rbroot); node; node = rb_next(node)) {
                struct iova *iova = container_of(node, struct iova, node);
                struct iova *new_iova;
@@ -352,6 +423,5 @@ copy_reserved_iova(struct iova_domain *from, struct iova_domain *to)
                        printk(KERN_ERR "Reserve iova range %lx@%lx failed\n",
                                iova->pfn_lo, iova->pfn_lo);
        }
-       spin_unlock(&from->iova_rbtree_lock);
-       spin_unlock_irqrestore(&from->iova_alloc_lock, flags);
+       spin_unlock_irqrestore(&from->iova_rbtree_lock, flags);
 }