Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / infiniband / hw / mthca / mthca_memfree.c
diff --git a/kernel/drivers/infiniband/hw/mthca/mthca_memfree.c b/kernel/drivers/infiniband/hw/mthca/mthca_memfree.c
new file mode 100644 (file)
index 0000000..7d2e42d
--- /dev/null
@@ -0,0 +1,760 @@
+/*
+ * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  All rights reserved.
+ * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/mm.h>
+#include <linux/scatterlist.h>
+#include <linux/sched.h>
+#include <linux/slab.h>
+
+#include <asm/page.h>
+
+#include "mthca_memfree.h"
+#include "mthca_dev.h"
+#include "mthca_cmd.h"
+
+/*
+ * We allocate in as big chunks as we can, up to a maximum of 256 KB
+ * per chunk.
+ */
+enum {
+       MTHCA_ICM_ALLOC_SIZE   = 1 << 18,
+       MTHCA_TABLE_CHUNK_SIZE = 1 << 18
+};
+
+struct mthca_user_db_table {
+       struct mutex mutex;
+       struct {
+               u64                uvirt;
+               struct scatterlist mem;
+               int                refcount;
+       }                page[0];
+};
+
+static void mthca_free_icm_pages(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
+{
+       int i;
+
+       if (chunk->nsg > 0)
+               pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
+                            PCI_DMA_BIDIRECTIONAL);
+
+       for (i = 0; i < chunk->npages; ++i)
+               __free_pages(sg_page(&chunk->mem[i]),
+                            get_order(chunk->mem[i].length));
+}
+
+static void mthca_free_icm_coherent(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
+{
+       int i;
+
+       for (i = 0; i < chunk->npages; ++i) {
+               dma_free_coherent(&dev->pdev->dev, chunk->mem[i].length,
+                                 lowmem_page_address(sg_page(&chunk->mem[i])),
+                                 sg_dma_address(&chunk->mem[i]));
+       }
+}
+
+void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm, int coherent)
+{
+       struct mthca_icm_chunk *chunk, *tmp;
+
+       if (!icm)
+               return;
+
+       list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
+               if (coherent)
+                       mthca_free_icm_coherent(dev, chunk);
+               else
+                       mthca_free_icm_pages(dev, chunk);
+
+               kfree(chunk);
+       }
+
+       kfree(icm);
+}
+
+static int mthca_alloc_icm_pages(struct scatterlist *mem, int order, gfp_t gfp_mask)
+{
+       struct page *page;
+
+       /*
+        * Use __GFP_ZERO because buggy firmware assumes ICM pages are
+        * cleared, and subtle failures are seen if they aren't.
+        */
+       page = alloc_pages(gfp_mask | __GFP_ZERO, order);
+       if (!page)
+               return -ENOMEM;
+
+       sg_set_page(mem, page, PAGE_SIZE << order, 0);
+       return 0;
+}
+
+static int mthca_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
+                                   int order, gfp_t gfp_mask)
+{
+       void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order, &sg_dma_address(mem),
+                                      gfp_mask);
+       if (!buf)
+               return -ENOMEM;
+
+       sg_set_buf(mem, buf, PAGE_SIZE << order);
+       BUG_ON(mem->offset);
+       sg_dma_len(mem) = PAGE_SIZE << order;
+       return 0;
+}
+
+struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
+                                 gfp_t gfp_mask, int coherent)
+{
+       struct mthca_icm *icm;
+       struct mthca_icm_chunk *chunk = NULL;
+       int cur_order;
+       int ret;
+
+       /* We use sg_set_buf for coherent allocs, which assumes low memory */
+       BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
+
+       icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
+       if (!icm)
+               return icm;
+
+       icm->refcount = 0;
+       INIT_LIST_HEAD(&icm->chunk_list);
+
+       cur_order = get_order(MTHCA_ICM_ALLOC_SIZE);
+
+       while (npages > 0) {
+               if (!chunk) {
+                       chunk = kmalloc(sizeof *chunk,
+                                       gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
+                       if (!chunk)
+                               goto fail;
+
+                       sg_init_table(chunk->mem, MTHCA_ICM_CHUNK_LEN);
+                       chunk->npages = 0;
+                       chunk->nsg    = 0;
+                       list_add_tail(&chunk->list, &icm->chunk_list);
+               }
+
+               while (1 << cur_order > npages)
+                       --cur_order;
+
+               if (coherent)
+                       ret = mthca_alloc_icm_coherent(&dev->pdev->dev,
+                                                      &chunk->mem[chunk->npages],
+                                                      cur_order, gfp_mask);
+               else
+                       ret = mthca_alloc_icm_pages(&chunk->mem[chunk->npages],
+                                                   cur_order, gfp_mask);
+
+               if (!ret) {
+                       ++chunk->npages;
+
+                       if (coherent)
+                               ++chunk->nsg;
+                       else if (chunk->npages == MTHCA_ICM_CHUNK_LEN) {
+                               chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
+                                                       chunk->npages,
+                                                       PCI_DMA_BIDIRECTIONAL);
+
+                               if (chunk->nsg <= 0)
+                                       goto fail;
+                       }
+
+                       if (chunk->npages == MTHCA_ICM_CHUNK_LEN)
+                               chunk = NULL;
+
+                       npages -= 1 << cur_order;
+               } else {
+                       --cur_order;
+                       if (cur_order < 0)
+                               goto fail;
+               }
+       }
+
+       if (!coherent && chunk) {
+               chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
+                                       chunk->npages,
+                                       PCI_DMA_BIDIRECTIONAL);
+
+               if (chunk->nsg <= 0)
+                       goto fail;
+       }
+
+       return icm;
+
+fail:
+       mthca_free_icm(dev, icm, coherent);
+       return NULL;
+}
+
+int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
+{
+       int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
+       int ret = 0;
+
+       mutex_lock(&table->mutex);
+
+       if (table->icm[i]) {
+               ++table->icm[i]->refcount;
+               goto out;
+       }
+
+       table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
+                                       (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
+                                       __GFP_NOWARN, table->coherent);
+       if (!table->icm[i]) {
+               ret = -ENOMEM;
+               goto out;
+       }
+
+       if (mthca_MAP_ICM(dev, table->icm[i],
+                         table->virt + i * MTHCA_TABLE_CHUNK_SIZE)) {
+               mthca_free_icm(dev, table->icm[i], table->coherent);
+               table->icm[i] = NULL;
+               ret = -ENOMEM;
+               goto out;
+       }
+
+       ++table->icm[i]->refcount;
+
+out:
+       mutex_unlock(&table->mutex);
+       return ret;
+}
+
+void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
+{
+       int i;
+
+       if (!mthca_is_memfree(dev))
+               return;
+
+       i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
+
+       mutex_lock(&table->mutex);
+
+       if (--table->icm[i]->refcount == 0) {
+               mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
+                               MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE);
+               mthca_free_icm(dev, table->icm[i], table->coherent);
+               table->icm[i] = NULL;
+       }
+
+       mutex_unlock(&table->mutex);
+}
+
+void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle)
+{
+       int idx, offset, dma_offset, i;
+       struct mthca_icm_chunk *chunk;
+       struct mthca_icm *icm;
+       struct page *page = NULL;
+
+       if (!table->lowmem)
+               return NULL;
+
+       mutex_lock(&table->mutex);
+
+       idx = (obj & (table->num_obj - 1)) * table->obj_size;
+       icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
+       dma_offset = offset = idx % MTHCA_TABLE_CHUNK_SIZE;
+
+       if (!icm)
+               goto out;
+
+       list_for_each_entry(chunk, &icm->chunk_list, list) {
+               for (i = 0; i < chunk->npages; ++i) {
+                       if (dma_handle && dma_offset >= 0) {
+                               if (sg_dma_len(&chunk->mem[i]) > dma_offset)
+                                       *dma_handle = sg_dma_address(&chunk->mem[i]) +
+                                               dma_offset;
+                               dma_offset -= sg_dma_len(&chunk->mem[i]);
+                       }
+                       /* DMA mapping can merge pages but not split them,
+                        * so if we found the page, dma_handle has already
+                        * been assigned to. */
+                       if (chunk->mem[i].length > offset) {
+                               page = sg_page(&chunk->mem[i]);
+                               goto out;
+                       }
+                       offset -= chunk->mem[i].length;
+               }
+       }
+
+out:
+       mutex_unlock(&table->mutex);
+       return page ? lowmem_page_address(page) + offset : NULL;
+}
+
+int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
+                         int start, int end)
+{
+       int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
+       int i, err;
+
+       for (i = start; i <= end; i += inc) {
+               err = mthca_table_get(dev, table, i);
+               if (err)
+                       goto fail;
+       }
+
+       return 0;
+
+fail:
+       while (i > start) {
+               i -= inc;
+               mthca_table_put(dev, table, i);
+       }
+
+       return err;
+}
+
+void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
+                          int start, int end)
+{
+       int i;
+
+       if (!mthca_is_memfree(dev))
+               return;
+
+       for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
+               mthca_table_put(dev, table, i);
+}
+
+struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
+                                             u64 virt, int obj_size,
+                                             int nobj, int reserved,
+                                             int use_lowmem, int use_coherent)
+{
+       struct mthca_icm_table *table;
+       int obj_per_chunk;
+       int num_icm;
+       unsigned chunk_size;
+       int i;
+
+       obj_per_chunk = MTHCA_TABLE_CHUNK_SIZE / obj_size;
+       num_icm = DIV_ROUND_UP(nobj, obj_per_chunk);
+
+       table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
+       if (!table)
+               return NULL;
+
+       table->virt     = virt;
+       table->num_icm  = num_icm;
+       table->num_obj  = nobj;
+       table->obj_size = obj_size;
+       table->lowmem   = use_lowmem;
+       table->coherent = use_coherent;
+       mutex_init(&table->mutex);
+
+       for (i = 0; i < num_icm; ++i)
+               table->icm[i] = NULL;
+
+       for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
+               chunk_size = MTHCA_TABLE_CHUNK_SIZE;
+               if ((i + 1) * MTHCA_TABLE_CHUNK_SIZE > nobj * obj_size)
+                       chunk_size = nobj * obj_size - i * MTHCA_TABLE_CHUNK_SIZE;
+
+               table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
+                                               (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
+                                               __GFP_NOWARN, use_coherent);
+               if (!table->icm[i])
+                       goto err;
+               if (mthca_MAP_ICM(dev, table->icm[i],
+                                 virt + i * MTHCA_TABLE_CHUNK_SIZE)) {
+                       mthca_free_icm(dev, table->icm[i], table->coherent);
+                       table->icm[i] = NULL;
+                       goto err;
+               }
+
+               /*
+                * Add a reference to this ICM chunk so that it never
+                * gets freed (since it contains reserved firmware objects).
+                */
+               ++table->icm[i]->refcount;
+       }
+
+       return table;
+
+err:
+       for (i = 0; i < num_icm; ++i)
+               if (table->icm[i]) {
+                       mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
+                                       MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE);
+                       mthca_free_icm(dev, table->icm[i], table->coherent);
+               }
+
+       kfree(table);
+
+       return NULL;
+}
+
+void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
+{
+       int i;
+
+       for (i = 0; i < table->num_icm; ++i)
+               if (table->icm[i]) {
+                       mthca_UNMAP_ICM(dev,
+                                       table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
+                                       MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE);
+                       mthca_free_icm(dev, table->icm[i], table->coherent);
+               }
+
+       kfree(table);
+}
+
+static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
+{
+       return dev->uar_table.uarc_base +
+               uar->index * dev->uar_table.uarc_size +
+               page * MTHCA_ICM_PAGE_SIZE;
+}
+
+int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
+                     struct mthca_user_db_table *db_tab, int index, u64 uaddr)
+{
+       struct page *pages[1];
+       int ret = 0;
+       int i;
+
+       if (!mthca_is_memfree(dev))
+               return 0;
+
+       if (index < 0 || index > dev->uar_table.uarc_size / 8)
+               return -EINVAL;
+
+       mutex_lock(&db_tab->mutex);
+
+       i = index / MTHCA_DB_REC_PER_PAGE;
+
+       if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE)       ||
+           (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
+           (uaddr & 4095)) {
+               ret = -EINVAL;
+               goto out;
+       }
+
+       if (db_tab->page[i].refcount) {
+               ++db_tab->page[i].refcount;
+               goto out;
+       }
+
+       ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
+                            pages, NULL);
+       if (ret < 0)
+               goto out;
+
+       sg_set_page(&db_tab->page[i].mem, pages[0], MTHCA_ICM_PAGE_SIZE,
+                       uaddr & ~PAGE_MASK);
+
+       ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
+       if (ret < 0) {
+               put_page(pages[0]);
+               goto out;
+       }
+
+       ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
+                                mthca_uarc_virt(dev, uar, i));
+       if (ret) {
+               pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
+               put_page(sg_page(&db_tab->page[i].mem));
+               goto out;
+       }
+
+       db_tab->page[i].uvirt    = uaddr;
+       db_tab->page[i].refcount = 1;
+
+out:
+       mutex_unlock(&db_tab->mutex);
+       return ret;
+}
+
+void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
+                        struct mthca_user_db_table *db_tab, int index)
+{
+       if (!mthca_is_memfree(dev))
+               return;
+
+       /*
+        * To make our bookkeeping simpler, we don't unmap DB
+        * pages until we clean up the whole db table.
+        */
+
+       mutex_lock(&db_tab->mutex);
+
+       --db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
+
+       mutex_unlock(&db_tab->mutex);
+}
+
+struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
+{
+       struct mthca_user_db_table *db_tab;
+       int npages;
+       int i;
+
+       if (!mthca_is_memfree(dev))
+               return NULL;
+
+       npages = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
+       db_tab = kmalloc(sizeof *db_tab + npages * sizeof *db_tab->page, GFP_KERNEL);
+       if (!db_tab)
+               return ERR_PTR(-ENOMEM);
+
+       mutex_init(&db_tab->mutex);
+       for (i = 0; i < npages; ++i) {
+               db_tab->page[i].refcount = 0;
+               db_tab->page[i].uvirt    = 0;
+               sg_init_table(&db_tab->page[i].mem, 1);
+       }
+
+       return db_tab;
+}
+
+void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
+                              struct mthca_user_db_table *db_tab)
+{
+       int i;
+
+       if (!mthca_is_memfree(dev))
+               return;
+
+       for (i = 0; i < dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE; ++i) {
+               if (db_tab->page[i].uvirt) {
+                       mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1);
+                       pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
+                       put_page(sg_page(&db_tab->page[i].mem));
+               }
+       }
+
+       kfree(db_tab);
+}
+
+int mthca_alloc_db(struct mthca_dev *dev, enum mthca_db_type type,
+                  u32 qn, __be32 **db)
+{
+       int group;
+       int start, end, dir;
+       int i, j;
+       struct mthca_db_page *page;
+       int ret = 0;
+
+       mutex_lock(&dev->db_tab->mutex);
+
+       switch (type) {
+       case MTHCA_DB_TYPE_CQ_ARM:
+       case MTHCA_DB_TYPE_SQ:
+               group = 0;
+               start = 0;
+               end   = dev->db_tab->max_group1;
+               dir   = 1;
+               break;
+
+       case MTHCA_DB_TYPE_CQ_SET_CI:
+       case MTHCA_DB_TYPE_RQ:
+       case MTHCA_DB_TYPE_SRQ:
+               group = 1;
+               start = dev->db_tab->npages - 1;
+               end   = dev->db_tab->min_group2;
+               dir   = -1;
+               break;
+
+       default:
+               ret = -EINVAL;
+               goto out;
+       }
+
+       for (i = start; i != end; i += dir)
+               if (dev->db_tab->page[i].db_rec &&
+                   !bitmap_full(dev->db_tab->page[i].used,
+                                MTHCA_DB_REC_PER_PAGE)) {
+                       page = dev->db_tab->page + i;
+                       goto found;
+               }
+
+       for (i = start; i != end; i += dir)
+               if (!dev->db_tab->page[i].db_rec) {
+                       page = dev->db_tab->page + i;
+                       goto alloc;
+               }
+
+       if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
+               ret = -ENOMEM;
+               goto out;
+       }
+
+       if (group == 0)
+               ++dev->db_tab->max_group1;
+       else
+               --dev->db_tab->min_group2;
+
+       page = dev->db_tab->page + end;
+
+alloc:
+       page->db_rec = dma_alloc_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
+                                         &page->mapping, GFP_KERNEL);
+       if (!page->db_rec) {
+               ret = -ENOMEM;
+               goto out;
+       }
+       memset(page->db_rec, 0, MTHCA_ICM_PAGE_SIZE);
+
+       ret = mthca_MAP_ICM_page(dev, page->mapping,
+                                mthca_uarc_virt(dev, &dev->driver_uar, i));
+       if (ret) {
+               dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
+                                 page->db_rec, page->mapping);
+               goto out;
+       }
+
+       bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
+
+found:
+       j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
+       set_bit(j, page->used);
+
+       if (group == 1)
+               j = MTHCA_DB_REC_PER_PAGE - 1 - j;
+
+       ret = i * MTHCA_DB_REC_PER_PAGE + j;
+
+       page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
+
+       *db = (__be32 *) &page->db_rec[j];
+
+out:
+       mutex_unlock(&dev->db_tab->mutex);
+
+       return ret;
+}
+
+void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
+{
+       int i, j;
+       struct mthca_db_page *page;
+
+       i = db_index / MTHCA_DB_REC_PER_PAGE;
+       j = db_index % MTHCA_DB_REC_PER_PAGE;
+
+       page = dev->db_tab->page + i;
+
+       mutex_lock(&dev->db_tab->mutex);
+
+       page->db_rec[j] = 0;
+       if (i >= dev->db_tab->min_group2)
+               j = MTHCA_DB_REC_PER_PAGE - 1 - j;
+       clear_bit(j, page->used);
+
+       if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
+           i >= dev->db_tab->max_group1 - 1) {
+               mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1);
+
+               dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
+                                 page->db_rec, page->mapping);
+               page->db_rec = NULL;
+
+               if (i == dev->db_tab->max_group1) {
+                       --dev->db_tab->max_group1;
+                       /* XXX may be able to unmap more pages now */
+               }
+               if (i == dev->db_tab->min_group2)
+                       ++dev->db_tab->min_group2;
+       }
+
+       mutex_unlock(&dev->db_tab->mutex);
+}
+
+int mthca_init_db_tab(struct mthca_dev *dev)
+{
+       int i;
+
+       if (!mthca_is_memfree(dev))
+               return 0;
+
+       dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
+       if (!dev->db_tab)
+               return -ENOMEM;
+
+       mutex_init(&dev->db_tab->mutex);
+
+       dev->db_tab->npages     = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
+       dev->db_tab->max_group1 = 0;
+       dev->db_tab->min_group2 = dev->db_tab->npages - 1;
+
+       dev->db_tab->page = kmalloc(dev->db_tab->npages *
+                                   sizeof *dev->db_tab->page,
+                                   GFP_KERNEL);
+       if (!dev->db_tab->page) {
+               kfree(dev->db_tab);
+               return -ENOMEM;
+       }
+
+       for (i = 0; i < dev->db_tab->npages; ++i)
+               dev->db_tab->page[i].db_rec = NULL;
+
+       return 0;
+}
+
+void mthca_cleanup_db_tab(struct mthca_dev *dev)
+{
+       int i;
+
+       if (!mthca_is_memfree(dev))
+               return;
+
+       /*
+        * Because we don't always free our UARC pages when they
+        * become empty to make mthca_free_db() simpler we need to
+        * make a sweep through the doorbell pages and free any
+        * leftover pages now.
+        */
+       for (i = 0; i < dev->db_tab->npages; ++i) {
+               if (!dev->db_tab->page[i].db_rec)
+                       continue;
+
+               if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
+                       mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
+
+               mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1);
+
+               dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
+                                 dev->db_tab->page[i].db_rec,
+                                 dev->db_tab->page[i].mapping);
+       }
+
+       kfree(dev->db_tab->page);
+       kfree(dev->db_tab);
+}