Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / gpu / drm / ttm / ttm_bo_vm.c
diff --git a/kernel/drivers/gpu/drm/ttm/ttm_bo_vm.c b/kernel/drivers/gpu/drm/ttm/ttm_bo_vm.c
new file mode 100644 (file)
index 0000000..8fb7213
--- /dev/null
@@ -0,0 +1,364 @@
+/**************************************************************************
+ *
+ * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * 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 NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS 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.
+ *
+ **************************************************************************/
+/*
+ * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
+ */
+
+#define pr_fmt(fmt) "[TTM] " fmt
+
+#include <ttm/ttm_module.h>
+#include <ttm/ttm_bo_driver.h>
+#include <ttm/ttm_placement.h>
+#include <drm/drm_vma_manager.h>
+#include <linux/mm.h>
+#include <linux/rbtree.h>
+#include <linux/module.h>
+#include <linux/uaccess.h>
+
+#define TTM_BO_VM_NUM_PREFAULT 16
+
+static int ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
+                               struct vm_area_struct *vma,
+                               struct vm_fault *vmf)
+{
+       int ret = 0;
+
+       if (likely(!test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)))
+               goto out_unlock;
+
+       /*
+        * Quick non-stalling check for idle.
+        */
+       ret = ttm_bo_wait(bo, false, false, true);
+       if (likely(ret == 0))
+               goto out_unlock;
+
+       /*
+        * If possible, avoid waiting for GPU with mmap_sem
+        * held.
+        */
+       if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
+               ret = VM_FAULT_RETRY;
+               if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT)
+                       goto out_unlock;
+
+               up_read(&vma->vm_mm->mmap_sem);
+               (void) ttm_bo_wait(bo, false, true, false);
+               goto out_unlock;
+       }
+
+       /*
+        * Ordinary wait.
+        */
+       ret = ttm_bo_wait(bo, false, true, false);
+       if (unlikely(ret != 0))
+               ret = (ret != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
+                       VM_FAULT_NOPAGE;
+
+out_unlock:
+       return ret;
+}
+
+static int ttm_bo_vm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
+{
+       struct ttm_buffer_object *bo = (struct ttm_buffer_object *)
+           vma->vm_private_data;
+       struct ttm_bo_device *bdev = bo->bdev;
+       unsigned long page_offset;
+       unsigned long page_last;
+       unsigned long pfn;
+       struct ttm_tt *ttm = NULL;
+       struct page *page;
+       int ret;
+       int i;
+       unsigned long address = (unsigned long)vmf->virtual_address;
+       int retval = VM_FAULT_NOPAGE;
+       struct ttm_mem_type_manager *man =
+               &bdev->man[bo->mem.mem_type];
+       struct vm_area_struct cvma;
+
+       /*
+        * Work around locking order reversal in fault / nopfn
+        * between mmap_sem and bo_reserve: Perform a trylock operation
+        * for reserve, and if it fails, retry the fault after waiting
+        * for the buffer to become unreserved.
+        */
+       ret = ttm_bo_reserve(bo, true, true, false, NULL);
+       if (unlikely(ret != 0)) {
+               if (ret != -EBUSY)
+                       return VM_FAULT_NOPAGE;
+
+               if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) {
+                       if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
+                               up_read(&vma->vm_mm->mmap_sem);
+                               (void) ttm_bo_wait_unreserved(bo);
+                       }
+
+                       return VM_FAULT_RETRY;
+               }
+
+               /*
+                * If we'd want to change locking order to
+                * mmap_sem -> bo::reserve, we'd use a blocking reserve here
+                * instead of retrying the fault...
+                */
+               return VM_FAULT_NOPAGE;
+       }
+
+       /*
+        * Refuse to fault imported pages. This should be handled
+        * (if at all) by redirecting mmap to the exporter.
+        */
+       if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
+               retval = VM_FAULT_SIGBUS;
+               goto out_unlock;
+       }
+
+       if (bdev->driver->fault_reserve_notify) {
+               ret = bdev->driver->fault_reserve_notify(bo);
+               switch (ret) {
+               case 0:
+                       break;
+               case -EBUSY:
+               case -ERESTARTSYS:
+                       retval = VM_FAULT_NOPAGE;
+                       goto out_unlock;
+               default:
+                       retval = VM_FAULT_SIGBUS;
+                       goto out_unlock;
+               }
+       }
+
+       /*
+        * Wait for buffer data in transit, due to a pipelined
+        * move.
+        */
+       ret = ttm_bo_vm_fault_idle(bo, vma, vmf);
+       if (unlikely(ret != 0)) {
+               retval = ret;
+               goto out_unlock;
+       }
+
+       ret = ttm_mem_io_lock(man, true);
+       if (unlikely(ret != 0)) {
+               retval = VM_FAULT_NOPAGE;
+               goto out_unlock;
+       }
+       ret = ttm_mem_io_reserve_vm(bo);
+       if (unlikely(ret != 0)) {
+               retval = VM_FAULT_SIGBUS;
+               goto out_io_unlock;
+       }
+
+       page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) +
+               vma->vm_pgoff - drm_vma_node_start(&bo->vma_node);
+       page_last = vma_pages(vma) + vma->vm_pgoff -
+               drm_vma_node_start(&bo->vma_node);
+
+       if (unlikely(page_offset >= bo->num_pages)) {
+               retval = VM_FAULT_SIGBUS;
+               goto out_io_unlock;
+       }
+
+       /*
+        * Make a local vma copy to modify the page_prot member
+        * and vm_flags if necessary. The vma parameter is protected
+        * by mmap_sem in write mode.
+        */
+       cvma = *vma;
+       cvma.vm_page_prot = vm_get_page_prot(cvma.vm_flags);
+
+       if (bo->mem.bus.is_iomem) {
+               cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
+                                               cvma.vm_page_prot);
+       } else {
+               ttm = bo->ttm;
+               cvma.vm_page_prot = ttm_io_prot(bo->mem.placement,
+                                               cvma.vm_page_prot);
+
+               /* Allocate all page at once, most common usage */
+               if (ttm->bdev->driver->ttm_tt_populate(ttm)) {
+                       retval = VM_FAULT_OOM;
+                       goto out_io_unlock;
+               }
+       }
+
+       /*
+        * Speculatively prefault a number of pages. Only error on
+        * first page.
+        */
+       for (i = 0; i < TTM_BO_VM_NUM_PREFAULT; ++i) {
+               if (bo->mem.bus.is_iomem)
+                       pfn = ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT) + page_offset;
+               else {
+                       page = ttm->pages[page_offset];
+                       if (unlikely(!page && i == 0)) {
+                               retval = VM_FAULT_OOM;
+                               goto out_io_unlock;
+                       } else if (unlikely(!page)) {
+                               break;
+                       }
+                       page->mapping = vma->vm_file->f_mapping;
+                       page->index = drm_vma_node_start(&bo->vma_node) +
+                               page_offset;
+                       pfn = page_to_pfn(page);
+               }
+
+               if (vma->vm_flags & VM_MIXEDMAP)
+                       ret = vm_insert_mixed(&cvma, address, pfn);
+               else
+                       ret = vm_insert_pfn(&cvma, address, pfn);
+
+               /*
+                * Somebody beat us to this PTE or prefaulting to
+                * an already populated PTE, or prefaulting error.
+                */
+
+               if (unlikely((ret == -EBUSY) || (ret != 0 && i > 0)))
+                       break;
+               else if (unlikely(ret != 0)) {
+                       retval =
+                           (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
+                       goto out_io_unlock;
+               }
+
+               address += PAGE_SIZE;
+               if (unlikely(++page_offset >= page_last))
+                       break;
+       }
+out_io_unlock:
+       ttm_mem_io_unlock(man);
+out_unlock:
+       ttm_bo_unreserve(bo);
+       return retval;
+}
+
+static void ttm_bo_vm_open(struct vm_area_struct *vma)
+{
+       struct ttm_buffer_object *bo =
+           (struct ttm_buffer_object *)vma->vm_private_data;
+
+       WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping);
+
+       (void)ttm_bo_reference(bo);
+}
+
+static void ttm_bo_vm_close(struct vm_area_struct *vma)
+{
+       struct ttm_buffer_object *bo = (struct ttm_buffer_object *)vma->vm_private_data;
+
+       ttm_bo_unref(&bo);
+       vma->vm_private_data = NULL;
+}
+
+static const struct vm_operations_struct ttm_bo_vm_ops = {
+       .fault = ttm_bo_vm_fault,
+       .open = ttm_bo_vm_open,
+       .close = ttm_bo_vm_close
+};
+
+static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
+                                                 unsigned long offset,
+                                                 unsigned long pages)
+{
+       struct drm_vma_offset_node *node;
+       struct ttm_buffer_object *bo = NULL;
+
+       drm_vma_offset_lock_lookup(&bdev->vma_manager);
+
+       node = drm_vma_offset_lookup_locked(&bdev->vma_manager, offset, pages);
+       if (likely(node)) {
+               bo = container_of(node, struct ttm_buffer_object, vma_node);
+               if (!kref_get_unless_zero(&bo->kref))
+                       bo = NULL;
+       }
+
+       drm_vma_offset_unlock_lookup(&bdev->vma_manager);
+
+       if (!bo)
+               pr_err("Could not find buffer object to map\n");
+
+       return bo;
+}
+
+int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
+               struct ttm_bo_device *bdev)
+{
+       struct ttm_bo_driver *driver;
+       struct ttm_buffer_object *bo;
+       int ret;
+
+       bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
+       if (unlikely(!bo))
+               return -EINVAL;
+
+       driver = bo->bdev->driver;
+       if (unlikely(!driver->verify_access)) {
+               ret = -EPERM;
+               goto out_unref;
+       }
+       ret = driver->verify_access(bo, filp);
+       if (unlikely(ret != 0))
+               goto out_unref;
+
+       vma->vm_ops = &ttm_bo_vm_ops;
+
+       /*
+        * Note: We're transferring the bo reference to
+        * vma->vm_private_data here.
+        */
+
+       vma->vm_private_data = bo;
+
+       /*
+        * We'd like to use VM_PFNMAP on shared mappings, where
+        * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
+        * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
+        * bad for performance. Until that has been sorted out, use
+        * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
+        */
+       vma->vm_flags |= VM_MIXEDMAP;
+       vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
+       return 0;
+out_unref:
+       ttm_bo_unref(&bo);
+       return ret;
+}
+EXPORT_SYMBOL(ttm_bo_mmap);
+
+int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo)
+{
+       if (vma->vm_pgoff != 0)
+               return -EACCES;
+
+       vma->vm_ops = &ttm_bo_vm_ops;
+       vma->vm_private_data = ttm_bo_reference(bo);
+       vma->vm_flags |= VM_MIXEDMAP;
+       vma->vm_flags |= VM_IO | VM_DONTEXPAND;
+       return 0;
+}
+EXPORT_SYMBOL(ttm_fbdev_mmap);