[PATCH] IB uverbs: memory pinning implementation
authorRoland Dreier <rolandd@cisco.com>
Fri, 8 Jul 2005 00:57:14 +0000 (17:57 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Fri, 8 Jul 2005 01:23:48 +0000 (18:23 -0700)
Add support for pinning userspace memory regions and returning a list of pages
in the region.  This includes tracking pinned memory against vm_locked and
preventing unprivileged users from exceeding RLIMIT_MEMLOCK.

Signed-off-by: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
drivers/infiniband/core/uverbs_mem.c [new file with mode: 0644]

diff --git a/drivers/infiniband/core/uverbs_mem.c b/drivers/infiniband/core/uverbs_mem.c
new file mode 100644 (file)
index 0000000..ed550f6
--- /dev/null
@@ -0,0 +1,221 @@
+/*
+ * Copyright (c) 2005 Topspin Communications.  All rights reserved.
+ * Copyright (c) 2005 Cisco Systems.  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.
+ *
+ * $Id: uverbs_mem.c 2743 2005-06-28 22:27:59Z roland $
+ */
+
+#include <linux/mm.h>
+#include <linux/dma-mapping.h>
+
+#include "uverbs.h"
+
+struct ib_umem_account_work {
+       struct work_struct work;
+       struct mm_struct  *mm;
+       unsigned long      diff;
+};
+
+
+static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
+{
+       struct ib_umem_chunk *chunk, *tmp;
+       int i;
+
+       list_for_each_entry_safe(chunk, tmp, &umem->chunk_list, list) {
+               dma_unmap_sg(dev->dma_device, chunk->page_list,
+                            chunk->nents, DMA_BIDIRECTIONAL);
+               for (i = 0; i < chunk->nents; ++i) {
+                       if (umem->writable && dirty)
+                               set_page_dirty_lock(chunk->page_list[i].page);
+                       put_page(chunk->page_list[i].page);
+               }
+
+               kfree(chunk);
+       }
+}
+
+int ib_umem_get(struct ib_device *dev, struct ib_umem *mem,
+               void *addr, size_t size, int write)
+{
+       struct page **page_list;
+       struct ib_umem_chunk *chunk;
+       unsigned long locked;
+       unsigned long lock_limit;
+       unsigned long cur_base;
+       unsigned long npages;
+       int ret = 0;
+       int off;
+       int i;
+
+       if (!can_do_mlock())
+               return -EPERM;
+
+       page_list = (struct page **) __get_free_page(GFP_KERNEL);
+       if (!page_list)
+               return -ENOMEM;
+
+       mem->user_base = (unsigned long) addr;
+       mem->length    = size;
+       mem->offset    = (unsigned long) addr & ~PAGE_MASK;
+       mem->page_size = PAGE_SIZE;
+       mem->writable  = write;
+
+       INIT_LIST_HEAD(&mem->chunk_list);
+
+       npages = PAGE_ALIGN(size + mem->offset) >> PAGE_SHIFT;
+
+       down_write(&current->mm->mmap_sem);
+
+       locked     = npages + current->mm->locked_vm;
+       lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
+
+       if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
+               ret = -ENOMEM;
+               goto out;
+       }
+
+       cur_base = (unsigned long) addr & PAGE_MASK;
+
+       while (npages) {
+               ret = get_user_pages(current, current->mm, cur_base,
+                                    min_t(int, npages,
+                                          PAGE_SIZE / sizeof (struct page *)),
+                                    1, !write, page_list, NULL);
+
+               if (ret < 0)
+                       goto out;
+
+               cur_base += ret * PAGE_SIZE;
+               npages   -= ret;
+
+               off = 0;
+
+               while (ret) {
+                       chunk = kmalloc(sizeof *chunk + sizeof (struct scatterlist) *
+                                       min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK),
+                                       GFP_KERNEL);
+                       if (!chunk) {
+                               ret = -ENOMEM;
+                               goto out;
+                       }
+
+                       chunk->nents = min_t(int, ret, IB_UMEM_MAX_PAGE_CHUNK);
+                       for (i = 0; i < chunk->nents; ++i) {
+                               chunk->page_list[i].page   = page_list[i + off];
+                               chunk->page_list[i].offset = 0;
+                               chunk->page_list[i].length = PAGE_SIZE;
+                       }
+
+                       chunk->nmap = dma_map_sg(dev->dma_device,
+                                                &chunk->page_list[0],
+                                                chunk->nents,
+                                                DMA_BIDIRECTIONAL);
+                       if (chunk->nmap <= 0) {
+                               for (i = 0; i < chunk->nents; ++i)
+                                       put_page(chunk->page_list[i].page);
+                               kfree(chunk);
+
+                               ret = -ENOMEM;
+                               goto out;
+                       }
+
+                       ret -= chunk->nents;
+                       off += chunk->nents;
+                       list_add_tail(&chunk->list, &mem->chunk_list);
+               }
+
+               ret = 0;
+       }
+
+out:
+       if (ret < 0)
+               __ib_umem_release(dev, mem, 0);
+       else
+               current->mm->locked_vm = locked;
+
+       up_write(&current->mm->mmap_sem);
+       free_page((unsigned long) page_list);
+
+       return ret;
+}
+
+void ib_umem_release(struct ib_device *dev, struct ib_umem *umem)
+{
+       __ib_umem_release(dev, umem, 1);
+
+       down_write(&current->mm->mmap_sem);
+       current->mm->locked_vm -=
+               PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
+       up_write(&current->mm->mmap_sem);
+}
+
+static void ib_umem_account(void *work_ptr)
+{
+       struct ib_umem_account_work *work = work_ptr;
+
+       down_write(&work->mm->mmap_sem);
+       work->mm->locked_vm -= work->diff;
+       up_write(&work->mm->mmap_sem);
+       mmput(work->mm);
+       kfree(work);
+}
+
+void ib_umem_release_on_close(struct ib_device *dev, struct ib_umem *umem)
+{
+       struct ib_umem_account_work *work;
+       struct mm_struct *mm;
+
+       __ib_umem_release(dev, umem, 1);
+
+       mm = get_task_mm(current);
+       if (!mm)
+               return;
+
+       /*
+        * We may be called with the mm's mmap_sem already held.  This
+        * can happen when a userspace munmap() is the call that drops
+        * the last reference to our file and calls our release
+        * method.  If there are memory regions to destroy, we'll end
+        * up here and not be able to take the mmap_sem.  Therefore we
+        * defer the vm_locked accounting to the system workqueue.
+        */
+
+       work = kmalloc(sizeof *work, GFP_KERNEL);
+       if (!work)
+               return;
+
+       INIT_WORK(&work->work, ib_umem_account, work);
+       work->mm   = mm;
+       work->diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT;
+
+       schedule_work(&work->work);
+}