/*
 * Copyright © 2008-2015 Intel Corporation
 *
 * 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, sublicense,
 * 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 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.
 *
 * Authors:
 *    Eric Anholt <eric@anholt.net>
 *
 */

#include <drm/drmP.h>
#include <drm/drm_vma_manager.h>
#include <drm/i915_drm.h>
#include "i915_drv.h"
#include "i915_gem_clflush.h"
#include "i915_vgpu.h"
#include "i915_trace.h"
#include "intel_drv.h"
#include "intel_frontbuffer.h"
#include "intel_mocs.h"
#include "i915_gemfs.h"
#include <linux/dma-fence-array.h>
#include <linux/kthread.h>
#include <linux/reservation.h>
#include <linux/shmem_fs.h>
#include <linux/slab.h>
#include <linux/stop_machine.h>
#include <linux/swap.h>
#include <linux/pci.h>
#include <linux/dma-buf.h>

static void i915_gem_flush_free_objects(struct drm_i915_private *i915);

static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj)
{
        if (obj->cache_dirty)
                return false;

        if (!(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE))
                return true;

        return obj->pin_global; /* currently in use by HW, keep flushed */
}

static int
insert_mappable_node(struct i915_ggtt *ggtt,
                     struct drm_mm_node *node, u32 size)
{
        memset(node, 0, sizeof(*node));
        return drm_mm_insert_node_in_range(&ggtt->base.mm, node,
                                           size, 0, I915_COLOR_UNEVICTABLE,
                                           0, ggtt->mappable_end,
                                           DRM_MM_INSERT_LOW);
}

static void
remove_mappable_node(struct drm_mm_node *node)
{
        drm_mm_remove_node(node);
}

/* some bookkeeping */
static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
                                  u64 size)
{
        spin_lock(&dev_priv->mm.object_stat_lock);
        dev_priv->mm.object_count++;
        dev_priv->mm.object_memory += size;
        spin_unlock(&dev_priv->mm.object_stat_lock);
}

static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
                                     u64 size)
{
        spin_lock(&dev_priv->mm.object_stat_lock);
        dev_priv->mm.object_count--;
        dev_priv->mm.object_memory -= size;
        spin_unlock(&dev_priv->mm.object_stat_lock);
}

static int
i915_gem_wait_for_error(struct i915_gpu_error *error)
{
        int ret;

        might_sleep();

        /*
         * Only wait 10 seconds for the gpu reset to complete to avoid hanging
         * userspace. If it takes that long something really bad is going on and
         * we should simply try to bail out and fail as gracefully as possible.
         */
        ret = wait_event_interruptible_timeout(error->reset_queue,
                                               !i915_reset_backoff(error),
                                               I915_RESET_TIMEOUT);
        if (ret == 0) {
                DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
                return -EIO;
        } else if (ret < 0) {
                return ret;
        } else {
                return 0;
        }
}

int i915_mutex_lock_interruptible(struct drm_device *dev)
{
        struct drm_i915_private *dev_priv = to_i915(dev);
        int ret;

        ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
        if (ret)
                return ret;

        ret = mutex_lock_interruptible(&dev->struct_mutex);
        if (ret)
                return ret;

        return 0;
}

int
i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
                            struct drm_file *file)
{
        struct drm_i915_private *dev_priv = to_i915(dev);
        struct i915_ggtt *ggtt = &dev_priv->ggtt;
        struct drm_i915_gem_get_aperture *args = data;
        struct i915_vma *vma;
        u64 pinned;

        pinned = ggtt->base.reserved;
        mutex_lock(&dev->struct_mutex);
        list_for_each_entry(vma, &ggtt->base.active_list, vm_link)
                if (i915_vma_is_pinned(vma))
                        pinned += vma->node.size;
        list_for_each_entry(vma, &ggtt->base.inactive_list, vm_link)
                if (i915_vma_is_pinned(vma))
                        pinned += vma->node.size;
        mutex_unlock(&dev->struct_mutex);

        args->aper_size = ggtt->base.total;
        args->aper_available_size = args->aper_size - pinned;

        return 0;
}

static int i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj)
{
        struct address_space *mapping = obj->base.filp->f_mapping;
        drm_dma_handle_t *phys;
        struct sg_table *st;
        struct scatterlist *sg;
        char *vaddr;
        int i;
        int err;

        if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj)))
                return -EINVAL;

        /* Always aligning to the object size, allows a single allocation
         * to handle all possible callers, and given typical object sizes,
         * the alignment of the buddy allocation will naturally match.
         */
        phys = drm_pci_alloc(obj->base.dev,
                             roundup_pow_of_two(obj->base.size),
                             roundup_pow_of_two(obj->base.size));
        if (!phys)
                return -ENOMEM;

        vaddr = phys->vaddr;
        for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
                struct page *page;
                char *src;

                page = shmem_read_mapping_page(mapping, i);
                if (IS_ERR(page)) {
                        err = PTR_ERR(page);
                        goto err_phys;
                }

                src = kmap_atomic(page);
                memcpy(vaddr, src, PAGE_SIZE);
                drm_clflush_virt_range(vaddr, PAGE_SIZE);
                kunmap_atomic(src);

                put_page(page);
                vaddr += PAGE_SIZE;
        }

        i915_gem_chipset_flush(to_i915(obj->base.dev));

        st = kmalloc(sizeof(*st), GFP_KERNEL);
        if (!st) {
                err = -ENOMEM;
                goto err_phys;
        }

        if (sg_alloc_table(st, 1, GFP_KERNEL)) {
                kfree(st);
                err = -ENOMEM;
                goto err_phys;
        }

        sg = st->sgl;
        sg->offset = 0;
        sg->length = obj->base.size;

        sg_dma_address(sg) = phys->busaddr;
        sg_dma_len(sg) = obj->base.size;

        obj->phys_handle = phys;

        __i915_gem_object_set_pages(obj, st, sg->length);

        return 0;

err_phys:
        drm_pci_free(obj->base.dev, phys);

        return err;
}

static void __start_cpu_write(struct drm_i915_gem_object *obj)
{
        obj->read_domains = I915_GEM_DOMAIN_CPU;
        obj->write_domain = I915_GEM_DOMAIN_CPU;
        if (cpu_write_needs_clflush(obj))
                obj->cache_dirty = true;
}

static void
__i915_gem_object_release_shmem(struct drm_i915_gem_object *obj,
                                struct sg_table *pages,
                                bool needs_clflush)
{
        GEM_BUG_ON(obj->mm.madv == __I915_MADV_PURGED);

        if (obj->mm.madv == I915_MADV_DONTNEED)
                obj->mm.dirty = false;

        if (needs_clflush &&
            (obj->read_domains & I915_GEM_DOMAIN_CPU) == 0 &&
            !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ))
                drm_clflush_sg(pages);

        __start_cpu_write(obj);
}

static void
i915_gem_object_put_pages_phys(struct drm_i915_gem_object *obj,
                               struct sg_table *pages)
{
        __i915_gem_object_release_shmem(obj, pages, false);

        if (obj->mm.dirty) {
                struct address_space *mapping = obj->base.filp->f_mapping;
                char *vaddr = obj->phys_handle->vaddr;
                int i;

                for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
                        struct page *page;
                        char *dst;

                        page = shmem_read_mapping_page(mapping, i);
                        if (IS_ERR(page))
                                continue;

                        dst = kmap_atomic(page);
                        drm_clflush_virt_range(vaddr, PAGE_SIZE);
                        memcpy(dst, vaddr, PAGE_SIZE);
                        kunmap_atomic(dst);

                        set_page_dirty(page);
                        if (obj->mm.madv == I915_MADV_WILLNEED)
                                mark_page_accessed(page);
                        put_page(page);
                        vaddr += PAGE_SIZE;
                }
                obj->mm.dirty = false;
        }

        sg_free_table(pages);
        kfree(pages);

        drm_pci_free(obj->base.dev, obj->phys_handle);
}

static void
i915_gem_object_release_phys(struct drm_i915_gem_object *obj)
{
        i915_gem_object_unpin_pages(obj);
}

static const struct drm_i915_gem_object_ops i915_gem_phys_ops = {
        .get_pages = i915_gem_object_get_pages_phys,
        .put_pages = i915_gem_object_put_pages_phys,
        .release = i915_gem_object_release_phys,
};

static const struct drm_i915_gem_object_ops i915_gem_object_ops;

int i915_gem_object_unbind(struct drm_i915_gem_object *obj)
{
        struct i915_vma *vma;
        LIST_HEAD(still_in_list);
        int ret;

        lockdep_assert_held(&obj->base.dev->struct_mutex);

        /* Closed vma are removed from the obj->vma_list - but they may
         * still have an active binding on the object. To remove those we
         * must wait for all rendering to complete to the object (as unbinding
         * must anyway), and retire the requests.
         */
        ret = i915_gem_object_set_to_cpu_domain(obj, false);
        if (ret)
                return ret;

        while ((vma = list_first_entry_or_null(&obj->vma_list,
                                               struct i915_vma,
                                               obj_link))) {
                list_move_tail(&vma->obj_link, &still_in_list);
                ret = i915_vma_unbind(vma);
                if (ret)
                        break;
        }
        list_splice(&still_in_list, &obj->vma_list);

        return ret;
}

static long
i915_gem_object_wait_fence(struct dma_fence *fence,
                           unsigned int flags,
                           long timeout,
                           struct intel_rps_client *rps_client)
{
        struct i915_request *rq;

        BUILD_BUG_ON(I915_WAIT_INTERRUPTIBLE != 0x1);

        if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
                return timeout;

        if (!dma_fence_is_i915(fence))
                return dma_fence_wait_timeout(fence,
                                              flags & I915_WAIT_INTERRUPTIBLE,
                                              timeout);

        rq = to_request(fence);
        if (i915_request_completed(rq))
                goto out;

        /*
         * This client is about to stall waiting for the GPU. In many cases
         * this is undesirable and limits the throughput of the system, as
         * many clients cannot continue processing user input/output whilst
         * blocked. RPS autotuning may take tens of milliseconds to respond
         * to the GPU load and thus incurs additional latency for the client.
         * We can circumvent that by promoting the GPU frequency to maximum
         * before we wait. This makes the GPU throttle up much more quickly
         * (good for benchmarks and user experience, e.g. window animations),
         * but at a cost of spending more power processing the workload
         * (bad for battery). Not all clients even want their results
         * immediately and for them we should just let the GPU select its own
         * frequency to maximise efficiency. To prevent a single client from
         * forcing the clocks too high for the whole system, we only allow
         * each client to waitboost once in a busy period.
         */
        if (rps_client && !i915_request_started(rq)) {
                if (INTEL_GEN(rq->i915) >= 6)
                        gen6_rps_boost(rq, rps_client);
        }

        timeout = i915_request_wait(rq, flags, timeout);

out:
        if (flags & I915_WAIT_LOCKED && i915_request_completed(rq))
                i915_request_retire_upto(rq);

        return timeout;
}

static long
i915_gem_object_wait_reservation(struct reservation_object *resv,
                                 unsigned int flags,
                                 long timeout,
                                 struct intel_rps_client *rps_client)
{
        unsigned int seq = __read_seqcount_begin(&resv->seq);
        struct dma_fence *excl;
        bool prune_fences = false;

        if (flags & I915_WAIT_ALL) {
                struct dma_fence **shared;
                unsigned int count, i;
                int ret;

                ret = reservation_object_get_fences_rcu(resv,
                                                        &excl, &count, &shared);
                if (ret)
                        return ret;

                for (i = 0; i < count; i++) {
                        timeout = i915_gem_object_wait_fence(shared[i],
                                                             flags, timeout,
                                                             rps_client);
                        if (timeout < 0)
                                break;

                        dma_fence_put(shared[i]);
                }

                for (; i < count; i++)
                        dma_fence_put(shared[i]);
                kfree(shared);

                /*
                 * If both shared fences and an exclusive fence exist,
                 * then by construction the shared fences must be later
                 * than the exclusive fence. If we successfully wait for
                 * all the shared fences, we know that the exclusive fence
                 * must all be signaled. If all the shared fences are
                 * signaled, we can prune the array and recover the
                 * floating references on the fences/requests.
                 */
                prune_fences = count && timeout >= 0;
        } else {
                excl = reservation_object_get_excl_rcu(resv);
        }

        if (excl && timeout >= 0)
                timeout = i915_gem_object_wait_fence(excl, flags, timeout,
                                                     rps_client);

        dma_fence_put(excl);

        /*
         * Opportunistically prune the fences iff we know they have *all* been
         * signaled and that the reservation object has not been changed (i.e.
         * no new fences have been added).
         */
        if (prune_fences && !__read_seqcount_retry(&resv->seq, seq)) {
                if (reservation_object_trylock(resv)) {
                        if (!__read_seqcount_retry(&resv->seq, seq))
                                reservation_object_add_excl_fence(resv, NULL);
                        reservation_object_unlock(resv);
                }
        }

        return timeout;
}

static void __fence_set_priority(struct dma_fence *fence, int prio)
{
        struct i915_request *rq;
        struct intel_engine_cs *engine;

        if (dma_fence_is_signaled(fence) || !dma_fence_is_i915(fence))
                return;

        rq = to_request(fence);
        engine = rq->engine;

        rcu_read_lock();
        if (engine->schedule)
                engine->schedule(rq, prio);
        rcu_read_unlock();
}

static void fence_set_priority(struct dma_fence *fence, int prio)
{
        /* Recurse once into a fence-array */
        if (dma_fence_is_array(fence)) {
                struct dma_fence_array *array = to_dma_fence_array(fence);
                int i;

                for (i = 0; i < array->num_fences; i++)
                        __fence_set_priority(array->fences[i], prio);
        } else {
                __fence_set_priority(fence, prio);
        }
}

int
i915_gem_object_wait_priority(struct drm_i915_gem_object *obj,
                              unsigned int flags,
                              int prio)
{
        struct dma_fence *excl;

        if (flags & I915_WAIT_ALL) {
                struct dma_fence **shared;
                unsigned int count, i;
                int ret;

                ret = reservation_object_get_fences_rcu(obj->resv,
                                                        &excl, &count, &shared);
                if (ret)
                        return ret;

                for (i = 0; i < count; i++) {
                        fence_set_priority(shared[i], prio);
                        dma_fence_put(shared[i]);
                }

                kfree(shared);
        } else {
                excl = reservation_object_get_excl_rcu(obj->resv);
        }

        if (excl) {
                fence_set_priority(excl, prio);
                dma_fence_put(excl);
        }
        return 0;
}

/**
 * Waits for rendering to the object to be completed
 * @obj: i915 gem object
 * @flags: how to wait (under a lock, for all rendering or just for writes etc)
 * @timeout: how long to wait
 * @rps_client: client (user process) to charge for any waitboosting
 */
int
i915_gem_object_wait(struct drm_i915_gem_object *obj,
                     unsigned int flags,
                     long timeout,
                     struct intel_rps_client *rps_client)
{
        might_sleep();
#if IS_ENABLED(CONFIG_LOCKDEP)
        GEM_BUG_ON(debug_locks &&
                   !!lockdep_is_held(&obj->base.dev->struct_mutex) !=
                   !!(flags & I915_WAIT_LOCKED));
#endif
        GEM_BUG_ON(timeout < 0);

        timeout = i915_gem_object_wait_reservation(obj->resv,
                                                   flags, timeout,
                                                   rps_client);
        return timeout < 0 ? timeout : 0;
}

static struct intel_rps_client *to_rps_client(struct drm_file *file)
{
        struct drm_i915_file_private *fpriv = file->driver_priv;

        return &fpriv->rps_client;
}

static int
i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
                     struct drm_i915_gem_pwrite *args,
                     struct drm_file *file)
{
        void *vaddr = obj->phys_handle->vaddr + args->offset;
        char __user *user_data = u64_to_user_ptr(args->data_ptr);

        /* We manually control the domain here and pretend that it
         * remains coherent i.e. in the GTT domain, like shmem_pwrite.
         */
        intel_fb_obj_invalidate(obj, ORIGIN_CPU);
        if (copy_from_user(vaddr, user_data, args->size))
                return -EFAULT;

        drm_clflush_virt_range(vaddr, args->size);
        i915_gem_chipset_flush(to_i915(obj->base.dev));

        intel_fb_obj_flush(obj, ORIGIN_CPU);
        return 0;
}

void *i915_gem_object_alloc(struct drm_i915_private *dev_priv)
{
        return kmem_cache_zalloc(dev_priv->objects, GFP_KERNEL);
}

void i915_gem_object_free(struct drm_i915_gem_object *obj)
{
        struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
        kmem_cache_free(dev_priv->objects, obj);
}

static int
i915_gem_create(struct drm_file *file,
                struct drm_i915_private *dev_priv,
                uint64_t size,
                uint32_t *handle_p)
{
        struct drm_i915_gem_object *obj;
        int ret;
        u32 handle;

        size = roundup(size, PAGE_SIZE);
        if (size == 0)
                return -EINVAL;

        /* Allocate the new object */
        obj = i915_gem_object_create(dev_priv, size);
        if (IS_ERR(obj))
                return PTR_ERR(obj);

        ret = drm_gem_handle_create(file, &obj->base, &handle);
        /* drop reference from allocate - handle holds it now */
        i915_gem_object_put(obj);
        if (ret)
                return ret;

        *handle_p = handle;
        return 0;
}

int
i915_gem_dumb_create(struct drm_file *file,
                     struct drm_device *dev,
                     struct drm_mode_create_dumb *args)
{
        /* have to work out size/pitch and return them */
        args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64);
        args->size = args->pitch * args->height;
        return i915_gem_create(file, to_i915(dev),
                               args->size, &args->handle);
}

static bool gpu_write_needs_clflush(struct drm_i915_gem_object *obj)
{
        return !(obj->cache_level == I915_CACHE_NONE ||
                 obj->cache_level == I915_CACHE_WT);
}

/**
 * Creates a new mm object and returns a handle to it.
 * @dev: drm device pointer
 * @data: ioctl data blob
 * @file: drm file pointer
 */
int
i915_gem_create_ioctl(struct drm_device *dev, void *data,
                      struct drm_file *file)
{
        struct drm_i915_private *dev_priv = to_i915(dev);
        struct drm_i915_gem_create *args = data;

        i915_gem_flush_free_objects(dev_priv);

        return i915_gem_create(file, dev_priv,
                               args->size, &args->handle);
}

static inline enum fb_op_origin
fb_write_origin(struct drm_i915_gem_object *obj, unsigned int domain)
{
        return (domain == I915_GEM_DOMAIN_GTT ?
                obj->frontbuffer_ggtt_origin : ORIGIN_CPU);
}

void i915_gem_flush_ggtt_writes(struct drm_i915_private *dev_priv)
{
        /*
         * No actual flushing is required for the GTT write domain for reads
         * from the GTT domain. Writes to it "immediately" go to main memory
         * as far as we know, so there's no chipset flush. It also doesn't
         * land in the GPU render cache.
         *
         * However, we do have to enforce the order so that all writes through
         * the GTT land before any writes to the device, such as updates to
         * the GATT itself.
         *
         * We also have to wait a bit for the writes to land from the GTT.
         * An uncached read (i.e. mmio) seems to be ideal for the round-trip
         * timing. This issue has only been observed when switching quickly
         * between GTT writes and CPU reads from inside the kernel on recent hw,
         * and it appears to only affect discrete GTT blocks (i.e. on LLC
         * system agents we cannot reproduce this behaviour, until Cannonlake
         * that was!).
         */

        wmb();

        intel_runtime_pm_get(dev_priv);
        spin_lock_irq(&dev_priv->uncore.lock);

        POSTING_READ_FW(RING_HEAD(RENDER_RING_BASE));

        spin_unlock_irq(&dev_priv->uncore.lock);
        intel_runtime_pm_put(dev_priv);
}

static void
flush_write_domain(struct drm_i915_gem_object *obj, unsigned int flush_domains)
{
        struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
        struct i915_vma *vma;

        if (!(obj->write_domain & flush_domains))
                return;

        switch (obj->write_domain) {
        case I915_GEM_DOMAIN_GTT:
                i915_gem_flush_ggtt_writes(dev_priv);

                intel_fb_obj_flush(obj,
                                   fb_write_origin(obj, I915_GEM_DOMAIN_GTT));

                for_each_ggtt_vma(vma, obj) {
                        if (vma->iomap)
                                continue;

                        i915_vma_unset_ggtt_write(vma);
                }
                break;

        case I915_GEM_DOMAIN_CPU:
                i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC);
                break;

        case I915_GEM_DOMAIN_RENDER:
                if (gpu_write_needs_clflush(obj))
                        obj->cache_dirty = true;
                break;
        }

        obj->write_domain = 0;
}

static inline int
__copy_to_user_swizzled(char __user *cpu_vaddr,
                        const char *gpu_vaddr, int gpu_offset,
                        int length)
{
        int ret, cpu_offset = 0;

        while (length > 0) {
                int cacheline_end = ALIGN(gpu_offset + 1, 64);
                int this_length = min(cacheline_end - gpu_offset, length);
                int swizzled_gpu_offset = gpu_offset ^ 64;

                ret = __copy_to_user(cpu_vaddr + cpu_offset,
                                     gpu_vaddr + swizzled_gpu_offset,
                                     this_length);
                if (ret)
                        return ret + length;

                cpu_offset += this_length;
                gpu_offset += this_length;
                length -= this_length;
        }

        return 0;
}

static inline int
__copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
                          const char __user *cpu_vaddr,
                          int length)
{
        int ret, cpu_offset = 0;

        while (length > 0) {
                int cacheline_end = ALIGN(gpu_offset + 1, 64);
                int this_length = min(cacheline_end - gpu_offset, length);
                int swizzled_gpu_offset = gpu_offset ^ 64;

                ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
                                       cpu_vaddr + cpu_offset,
                                       this_length);
                if (ret)
                        return ret + length;

                cpu_offset += this_length;
                gpu_offset += this_length;
                length -= this_length;
        }

        return 0;
}

/*
 * Pins the specified object's pages and synchronizes the object with
 * GPU accesses. Sets needs_clflush to non-zero if the caller should
 * flush the object from the CPU cache.
 */
int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
                                    unsigned int *needs_clflush)
{
        int ret;

        lockdep_assert_held(&obj->base.dev->struct_mutex);

        *needs_clflush = 0;
        if (!i915_gem_object_has_struct_page(obj))
                return -ENODEV;

        ret = i915_gem_object_wait(obj,
                                   I915_WAIT_INTERRUPTIBLE |
                                   I915_WAIT_LOCKED,
                                   MAX_SCHEDULE_TIMEOUT,
                                   NULL);
        if (ret)
                return ret;

        ret = i915_gem_object_pin_pages(obj);
        if (ret)
                return ret;

        if (obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ ||
            !static_cpu_has(X86_FEATURE_CLFLUSH)) {
                ret = i915_gem_object_set_to_cpu_domain(obj, false);
                if (ret)
                        goto err_unpin;
                else
                        goto out;
        }

        flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU);

        /* If we're not in the cpu read domain, set ourself into the gtt
         * read domain and manually flush cachelines (if required). This
         * optimizes for the case when the gpu will dirty the data
         * anyway again before the next pread happens.
         */
        if (!obj->cache_dirty &&
            !(obj->read_domains & I915_GEM_DOMAIN_CPU))
                *needs_clflush = CLFLUSH_BEFORE;

out:
        /* return with the pages pinned */
        return 0;

err_unpin:
        i915_gem_object_unpin_pages(obj);
        return ret;
}

int i915_gem_obj_prepare_shmem_write(struct drm_i915_gem_object *obj,
                                     unsigned int *needs_clflush)
{
        int ret;

        lockdep_assert_held(&obj->base.dev->struct_mutex);

        *needs_clflush = 0;
        if (!i915_gem_object_has_struct_page(obj))
                return -ENODEV;

        ret = i915_gem_object_wait(obj,
                                   I915_WAIT_INTERRUPTIBLE |
                                   I915_WAIT_LOCKED |
                                   I915_WAIT_ALL,
                                   MAX_SCHEDULE_TIMEOUT,
                                   NULL);
        if (ret)
                return ret;

        ret = i915_gem_object_pin_pages(obj);
        if (ret)
                return ret;

        if (obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE ||
            !static_cpu_has(X86_FEATURE_CLFLUSH)) {
                ret = i915_gem_object_set_to_cpu_domain(obj, true);
                if (ret)
                        goto err_unpin;
                else
                        goto out;
        }

        flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU);

        /* If we're not in the cpu write domain, set ourself into the
         * gtt write domain and manually flush cachelines (as required).
         * This optimizes for the case when the gpu will use the data
         * right away and we therefore have to clflush anyway.
         */
        if (!obj->cache_dirty) {
                *needs_clflush |= CLFLUSH_AFTER;

                /*
                 * Same trick applies to invalidate partially written
                 * cachelines read before writing.
                 */
                if (!(obj->read_domains & I915_GEM_DOMAIN_CPU))
                        *needs_clflush |= CLFLUSH_BEFORE;
        }

out:
        intel_fb_obj_invalidate(obj, ORIGIN_CPU);
        obj->mm.dirty = true;
        /* return with the pages pinned */
        return 0;

err_unpin:
        i915_gem_object_unpin_pages(obj);
        return ret;
}

static void
shmem_clflush_swizzled_range(char *addr, unsigned long length,
                             bool swizzled)
{
        if (unlikely(swizzled)) {
                unsigned long start = (unsigned long) addr;
                unsigned long end = (unsigned long) addr + length;

                /* For swizzling simply ensure that we always flush both
                 * channels. Lame, but simple and it works. Swizzled
                 * pwrite/pread is far from a hotpath - current userspace
                 * doesn't use it at all. */
                start = round_down(start, 128);
                end = round_up(end, 128);

                drm_clflush_virt_range((void *)start, end - start);
        } else {
                drm_clflush_virt_range(addr, length);
        }

}

/* Only difference to the fast-path function is that this can handle bit17
 * and uses non-atomic copy and kmap functions. */
static int
shmem_pread_slow(struct page *page, int offset, int length,
                 char __user *user_data,
                 bool page_do_bit17_swizzling, bool needs_clflush)
{
        char *vaddr;
        int ret;

        vaddr = kmap(page);
        if (needs_clflush)
                shmem_clflush_swizzled_range(vaddr + offset, length,
                                             page_do_bit17_swizzling);

        if (page_do_bit17_swizzling)
                ret = __copy_to_user_swizzled(user_data, vaddr, offset, length);
        else
                ret = __copy_to_user(user_data, vaddr + offset, length);
        kunmap(page);

        return ret ? - EFAULT : 0;
}

static int
shmem_pread(struct page *page, int offset, int length, char __user *user_data,
            bool page_do_bit17_swizzling, bool needs_clflush)
{
        int ret;

        ret = -ENODEV;
        if (!page_do_bit17_swizzling) {
                char *vaddr = kmap_atomic(page);

                if (needs_clflush)
                        drm_clflush_virt_range(vaddr + offset, length);
                ret = __copy_to_user_inatomic(user_data, vaddr + offset, length);
                kunmap_atomic(vaddr);
        }
        if (ret == 0)
                return 0;

        return shmem_pread_slow(page, offset, length, user_data,
                                page_do_bit17_swizzling, needs_clflush);
}

static int
i915_gem_shmem_pread(struct drm_i915_gem_object *obj,
                     struct drm_i915_gem_pread *args)
{
        char __user *user_data;
        u64 remain;
        unsigned int obj_do_bit17_swizzling;
        unsigned int needs_clflush;
        unsigned int idx, offset;
        int ret;

        obj_do_bit17_swizzling = 0;
        if (i915_gem_object_needs_bit17_swizzle(obj))
                obj_do_bit17_swizzling = BIT(17);

        ret = mutex_lock_interruptible(&obj->base.dev->struct_mutex);

        if (ret)
                return ret;

        ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush);
        mutex_unlock(&obj->base.dev->struct_mutex);
        if (ret)
                return ret;

        remain = args->size;
        user_data = u64_to_user_ptr(args->data_ptr);
        offset = offset_in_page(args->offset);
        for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
                struct page *page = i915_gem_object_get_page(obj, idx);
                int length;

                length = remain;
                if (offset + length > PAGE_SIZE)
                        length = PAGE_SIZE - offset;

                ret = shmem_pread(page, offset, length, user_data,
                                  page_to_phys(page) & obj_do_bit17_swizzling,
                                  needs_clflush);
                if (ret)
                        break;

                remain -= length;
                user_data += length;
                offset = 0;
        }

        i915_gem_obj_finish_shmem_access(obj);
        return ret;
}

static inline bool
gtt_user_read(struct io_mapping *mapping,
              loff_t base, int offset,
              char __user *user_data, int length)
{
        void __iomem *vaddr;
        unsigned long unwritten;

        /* We can use the cpu mem copy function because this is X86. */
        vaddr = io_mapping_map_atomic_wc(mapping, base);
        unwritten = __copy_to_user_inatomic(user_data,
                                            (void __force *)vaddr + offset,
                                            length);
        io_mapping_unmap_atomic(vaddr);
        if (unwritten) {
                vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
                unwritten = copy_to_user(user_data,
                                         (void __force *)vaddr + offset,
                                         length);
                io_mapping_unmap(vaddr);
        }
        return unwritten;
}

static int
i915_gem_gtt_pread(struct drm_i915_gem_object *obj,
                   const struct drm_i915_gem_pread *args)
{
        struct drm_i915_private *i915 = to_i915(obj->base.dev);
        struct i915_ggtt *ggtt = &i915->ggtt;
        struct drm_mm_node node;
        struct i915_vma *vma;
        void __user *user_data;
        u64 remain, offset;
        int ret;

        ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
        if (ret)
                return ret;

        intel_runtime_pm_get(i915);
        vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
                                       PIN_MAPPABLE |
                                       PIN_NONFAULT |
                                       PIN_NONBLOCK);
        if (!IS_ERR(vma)) {
                node.start = i915_ggtt_offset(vma);
                node.allocated = false;
                ret = i915_vma_put_fence(vma);
                if (ret) {
                        i915_vma_unpin(vma);
                        vma = ERR_PTR(ret);
                }
        }
        if (IS_ERR(vma)) {
                ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
                if (ret)
                        goto out_unlock;
                GEM_BUG_ON(!node.allocated);
        }

        ret = i915_gem_object_set_to_gtt_domain(obj, false);
        if (ret)
                goto out_unpin;

        mutex_unlock(&i915->drm.struct_mutex);

        user_data = u64_to_user_ptr(args->data_ptr);
        remain = args->size;
        offset = args->offset;

        while (remain > 0) {
                /* Operation in this page
                 *
                 * page_base = page offset within aperture
                 * page_offset = offset within page
                 * page_length = bytes to copy for this page
                 */
                u32 page_base = node.start;
                unsigned page_offset = offset_in_page(offset);
                unsigned page_length = PAGE_SIZE - page_offset;
                page_length = remain < page_length ? remain : page_length;
                if (node.allocated) {
                        wmb();
                        ggtt->base.insert_page(&ggtt->base,
                                               i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
                                               node.start, I915_CACHE_NONE, 0);
                        wmb();
                } else {
                        page_base += offset & PAGE_MASK;
                }

                if (gtt_user_read(&ggtt->iomap, page_base, page_offset,
                                  user_data, page_length)) {
                        ret = -EFAULT;
                        break;
                }

                remain -= page_length;
                user_data += page_length;
                offset += page_length;
        }

        mutex_lock(&i915->drm.struct_mutex);
out_unpin:
        if (node.allocated) {
                wmb();
                ggtt->base.clear_range(&ggtt->base,
                                       node.start, node.size);
                remove_mappable_node(&node);
        } else {
                i915_vma_unpin(vma);
        }
out_unlock:
        intel_runtime_pm_put(i915);
        mutex_unlock(&i915->drm.struct_mutex);

        return ret;
}

/**
 * Reads data from the object referenced by handle.
 * @dev: drm device pointer
 * @data: ioctl data blob
 * @file: drm file pointer
 *
 * On error, the contents of *data are undefined.
 */
int
i915_gem_pread_ioctl(struct drm_device *dev, void *data,
                     struct drm_file *file)
{
        struct drm_i915_gem_pread *args = data;
        struct drm_i915_gem_object *obj;
        int ret;

        if (args->size == 0)
                return 0;

        if (!access_ok(VERIFY_WRITE,
                       u64_to_user_ptr(args->data_ptr),
                       args->size))
                return -EFAULT;

        obj = i915_gem_object_lookup(file, args->handle);
        if (!obj)
                return -ENOENT;

        /* Bounds check source.  */
        if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
                ret = -EINVAL;
                goto out;
        }

        trace_i915_gem_object_pread(obj, args->offset, args->size);

        ret = i915_gem_object_wait(obj,
                                   I915_WAIT_INTERRUPTIBLE,
                                   MAX_SCHEDULE_TIMEOUT,
                                   to_rps_client(file));
        if (ret)
                goto out;

        ret = i915_gem_object_pin_pages(obj);
        if (ret)
                goto out;

        ret = i915_gem_shmem_pread(obj, args);
        if (ret == -EFAULT || ret == -ENODEV)
                ret = i915_gem_gtt_pread(obj, args);

        i915_gem_object_unpin_pages(obj);
out:
        i915_gem_object_put(obj);
        return ret;
}

/* This is the fast write path which cannot handle
 * page faults in the source data
 */

static inline bool
ggtt_write(struct io_mapping *mapping,
           loff_t base, int offset,
           char __user *user_data, int length)
{
        void __iomem *vaddr;
        unsigned long unwritten;

        /* We can use the cpu mem copy function because this is X86. */
        vaddr = io_mapping_map_atomic_wc(mapping, base);
        unwritten = __copy_from_user_inatomic_nocache((void __force *)vaddr + offset,
                                                      user_data, length);
        io_mapping_unmap_atomic(vaddr);
        if (unwritten) {
                vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
                unwritten = copy_from_user((void __force *)vaddr + offset,
                                           user_data, length);
                io_mapping_unmap(vaddr);
        }

        return unwritten;
}

/**
 * This is the fast pwrite path, where we copy the data directly from the
 * user into the GTT, uncached.
 * @obj: i915 GEM object
 * @args: pwrite arguments structure
 */
static int
i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
                         const struct drm_i915_gem_pwrite *args)
{
        struct drm_i915_private *i915 = to_i915(obj->base.dev);
        struct i915_ggtt *ggtt = &i915->ggtt;
        struct drm_mm_node node;
        struct i915_vma *vma;
        u64 remain, offset;
        void __user *user_data;
        int ret;

        ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
        if (ret)
                return ret;

        if (i915_gem_object_has_struct_page(obj)) {
                /*
                 * Avoid waking the device up if we can fallback, as
                 * waking/resuming is very slow (worst-case 10-100 ms
                 * depending on PCI sleeps and our own resume time).
                 * This easily dwarfs any performance advantage from
                 * using the cache bypass of indirect GGTT access.
                 */
                if (!intel_runtime_pm_get_if_in_use(i915)) {
                        ret = -EFAULT;
                        goto out_unlock;
                }
        } else {
                /* No backing pages, no fallback, we must force GGTT access */
                intel_runtime_pm_get(i915);
        }

        vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
                                       PIN_MAPPABLE |
                                       PIN_NONFAULT |
                                       PIN_NONBLOCK);
        if (!IS_ERR(vma)) {
                node.start = i915_ggtt_offset(vma);
                node.allocated = false;
                ret = i915_vma_put_fence(vma);
                if (ret) {
                        i915_vma_unpin(vma);
                        vma = ERR_PTR(ret);
                }
        }
        if (IS_ERR(vma)) {
                ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
                if (ret)
                        goto out_rpm;
                GEM_BUG_ON(!node.allocated);
        }

        ret = i915_gem_object_set_to_gtt_domain(obj, true);
        if (ret)
                goto out_unpin;

        mutex_unlock(&i915->drm.struct_mutex);

        intel_fb_obj_invalidate(obj, ORIGIN_CPU);

        user_data = u64_to_user_ptr(args->data_ptr);
        offset = args->offset;
        remain = args->size;
        while (remain) {
                /* Operation in this page
                 *
                 * page_base = page offset within aperture
                 * page_offset = offset within page
                 * page_length = bytes to copy for this page
                 */
                u32 page_base = node.start;
                unsigned int page_offset = offset_in_page(offset);
                unsigned int page_length = PAGE_SIZE - page_offset;
                page_length = remain < page_length ? remain : page_length;
                if (node.allocated) {
                        wmb(); /* flush the write before we modify the GGTT */
                        ggtt->base.insert_page(&ggtt->base,
                                               i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
                                               node.start, I915_CACHE_NONE, 0);
                        wmb(); /* flush modifications to the GGTT (insert_page) */
                } else {
                        page_base += offset & PAGE_MASK;
                }
                /* If we get a fault while copying data, then (presumably) our
                 * source page isn't available.  Return the error and we'll
                 * retry in the slow path.
                 * If the object is non-shmem backed, we retry again with the
                 * path that handles page fault.
                 */
                if (ggtt_write(&ggtt->iomap, page_base, page_offset,
                               user_data, page_length)) {
                        ret = -EFAULT;
                        break;
                }

                remain -= page_length;
                user_data += page_length;
                offset += page_length;
        }
        intel_fb_obj_flush(obj, ORIGIN_CPU);

        mutex_lock(&i915->drm.struct_mutex);
out_unpin:
        if (node.allocated) {
                wmb();
                ggtt->base.clear_range(&ggtt->base,
                                       node.start, node.size);
                remove_mappable_node(&node);
        } else {
                i915_vma_unpin(vma);
        }
out_rpm:
        intel_runtime_pm_put(i915);
out_unlock:
        mutex_unlock(&i915->drm.struct_mutex);
        return ret;
}

static int
shmem_pwrite_slow(struct page *page, int offset, int length,
                  char __user *user_data,
                  bool page_do_bit17_swizzling,
                  bool needs_clflush_before,
                  bool needs_clflush_after)
{
        char *vaddr;
        int ret;

        vaddr = kmap(page);
        if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
                shmem_clflush_swizzled_range(vaddr + offset, length,
                                             page_do_bit17_swizzling);
        if (page_do_bit17_swizzling)
                ret = __copy_from_user_swizzled(vaddr, offset, user_data,
                                                length);
        else
                ret = __copy_from_user(vaddr + offset, user_data, length);
        if (needs_clflush_after)
                shmem_clflush_swizzled_range(vaddr + offset, length,
                                             page_do_bit17_swizzling);
        kunmap(page);

        return ret ? -EFAULT : 0;
}

/* Per-page copy function for the shmem pwrite fastpath.
 * Flushes invalid cachelines before writing to the target if
 * needs_clflush_before is set and flushes out any written cachelines after
 * writing if needs_clflush is set.
 */
static int
shmem_pwrite(struct page *page, int offset, int len, char __user *user_data,
             bool page_do_bit17_swizzling,
             bool needs_clflush_before,
             bool needs_clflush_after)
{
        int ret;

        ret = -ENODEV;
        if (!page_do_bit17_swizzling) {
                char *vaddr = kmap_atomic(page);

                if (needs_clflush_before)
                        drm_clflush_virt_range(vaddr + offset, len);
                ret = __copy_from_user_inatomic(vaddr + offset, user_data, len);
                if (needs_clflush_after)
                        drm_clflush_virt_range(vaddr + offset, len);

                kunmap_atomic(vaddr);
        }
        if (ret == 0)
                return ret;

        return shmem_pwrite_slow(page, offset, len, user_data,
                                 page_do_bit17_swizzling,
                                 needs_clflush_before,
                                 needs_clflush_after);
}

static int
i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj,
                      const struct drm_i915_gem_pwrite *args)
{
        struct drm_i915_private *i915 = to_i915(obj->base.dev);
        void __user *user_data;
        u64 remain;
        unsigned int obj_do_bit17_swizzling;
        unsigned int partial_cacheline_write;
        unsigned int needs_clflush;
        unsigned int offset, idx;
        int ret;

        ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
        if (ret)
                return ret;

        ret = i915_gem_obj_prepare_shmem_write(obj, &needs_clflush);
        mutex_unlock(&i915->drm.struct_mutex);
        if (ret)
                return ret;

        obj_do_bit17_swizzling = 0;
        if (i915_gem_object_needs_bit17_swizzle(obj))
                obj_do_bit17_swizzling = BIT(17);

        /* If we don't overwrite a cacheline completely we need to be
         * careful to have up-to-date data by first clflushing. Don't
         * overcomplicate things and flush the entire patch.
         */
        partial_cacheline_write = 0;
        if (needs_clflush & CLFLUSH_BEFORE)
                partial_cacheline_write = boot_cpu_data.x86_clflush_size - 1;

        user_data = u64_to_user_ptr(args->data_ptr);
        remain = args->size;
        offset = offset_in_page(args->offset);
        for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
                struct page *page = i915_gem_object_get_page(obj, idx);
                int length;

                length = remain;
                if (offset + length > PAGE_SIZE)
                        length = PAGE_SIZE - offset;

                ret = shmem_pwrite(page, offset, length, user_data,
                                   page_to_phys(page) & obj_do_bit17_swizzling,
                                   (offset | length) & partial_cacheline_write,
                                   needs_clflush & CLFLUSH_AFTER);
                if (ret)
                        break;

                remain -= length;
                user_data += length;
                offset = 0;
        }

        intel_fb_obj_flush(obj, ORIGIN_CPU);
        i915_gem_obj_finish_shmem_access(obj);
        return ret;
}

/**
 * Writes data to the object referenced by handle.
 * @dev: drm device
 * @data: ioctl data blob
 * @file: drm file
 *
 * On error, the contents of the buffer that were to be modified are undefined.
 */
int
i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
                      struct drm_file *file)
{
        struct drm_i915_gem_pwrite *args = data;
        struct drm_i915_gem_object *obj;
        int ret;

        if (args->size == 0)
                return 0;

        if (!access_ok(VERIFY_READ,
                       u64_to_user_ptr(args->data_ptr),
                       args->size))
                return -EFAULT;

        obj = i915_gem_object_lookup(file, args->handle);
        if (!obj)
                return -ENOENT;

        /* Bounds check destination. */
        if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
                ret = -EINVAL;
                goto err;
        }

        trace_i915_gem_object_pwrite(obj, args->offset, args->size);

        ret = -ENODEV;
        if (obj->ops->pwrite)
                ret = obj->ops->pwrite(obj, args);
        if (ret != -ENODEV)
                goto err;

        ret = i915_gem_object_wait(obj,
                                   I915_WAIT_INTERRUPTIBLE |
                                   I915_WAIT_ALL,
                                   MAX_SCHEDULE_TIMEOUT,
                                   to_rps_client(file));
        if (ret)
                goto err;

        ret = i915_gem_object_pin_pages(obj);
        if (ret)
                goto err;

        ret = -EFAULT;
        /* We can only do the GTT pwrite on untiled buffers, as otherwise
         * it would end up going through the fenced access, and we'll get
         * different detiling behavior between reading and writing.
         * pread/pwrite currently are reading and writing from the CPU
         * perspective, requiring manual detiling by the client.
         */
        if (!i915_gem_object_has_struct_page(obj) ||
            cpu_write_needs_clflush(obj))
                /* Note that the gtt paths might fail with non-page-backed user
                 * pointers (e.g. gtt mappings when moving data between
                 * textures). Fallback to the shmem path in that case.
                 */
                ret = i915_gem_gtt_pwrite_fast(obj, args);

        if (ret == -EFAULT || ret == -ENOSPC) {
                if (obj->phys_handle)
                        ret = i915_gem_phys_pwrite(obj, args, file);
                else
                        ret = i915_gem_shmem_pwrite(obj, args);
        }

        i915_gem_object_unpin_pages(obj);
err:
        i915_gem_object_put(obj);
        return ret;
}

static void i915_gem_object_bump_inactive_ggtt(struct drm_i915_gem_object *obj)
{
        struct drm_i915_private *i915;
        struct list_head *list;
        struct i915_vma *vma;

        GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));

        for_each_ggtt_vma(vma, obj) {
                if (i915_vma_is_active(vma))
                        continue;

                if (!drm_mm_node_allocated(&vma->node))
                        continue;

                list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
        }

        i915 = to_i915(obj->base.dev);
        spin_lock(&i915->mm.obj_lock);
        list = obj->bind_count ? &i915->mm.bound_list : &i915->mm.unbound_list;
        list_move_tail(&obj->mm.link, list);
        spin_unlock(&i915->mm.obj_lock);
}

/**
 * Called when user space prepares to use an object with the CPU, either
 * through the mmap ioctl's mapping or a GTT mapping.
 * @dev: drm device
 * @data: ioctl data blob
 * @file: drm file
 */
int
i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
                          struct drm_file *file)
{
        struct drm_i915_gem_set_domain *args = data;
        struct drm_i915_gem_object *obj;
        uint32_t read_domains = args->read_domains;
        uint32_t write_domain = args->write_domain;
        int err;

        /* Only handle setting domains to types used by the CPU. */
        if ((write_domain | read_domains) & I915_GEM_GPU_DOMAINS)
                return -EINVAL;

        /* Having something in the write domain implies it's in the read
         * domain, and only that read domain.  Enforce that in the request.
         */
        if (write_domain != 0 && read_domains != write_domain)
                return -EINVAL;

        obj = i915_gem_object_lookup(file, args->handle);
        if (!obj)
                return -ENOENT;

        /* Try to flush the object off the GPU without holding the lock.
         * We will repeat the flush holding the lock in the normal manner
         * to catch cases where we are gazumped.
         */
        err = i915_gem_object_wait(obj,
                                   I915_WAIT_INTERRUPTIBLE |
                                   (write_domain ? I915_WAIT_ALL : 0),
                                   MAX_SCHEDULE_TIMEOUT,
                                   to_rps_client(file));
        if (err)
                goto out;

        /*
         * Proxy objects do not control access to the backing storage, ergo
         * they cannot be used as a means to manipulate the cache domain
         * tracking for that backing storage. The proxy object is always
         * considered to be outside of any cache domain.
         */
        if (i915_gem_object_is_proxy(obj)) {
                err = -ENXIO;
                goto out;
        }

        /*
         * Flush and acquire obj->pages so that we are coherent through
         * direct access in memory with previous cached writes through
         * shmemfs and that our cache domain tracking remains valid.
         * For example, if the obj->filp was moved to swap without us
         * being notified and releasing the pages, we would mistakenly
         * continue to assume that the obj remained out of the CPU cached
         * domain.
         */
        err = i915_gem_object_pin_pages(obj);
        if (err)
                goto out;

        err = i915_mutex_lock_interruptible(dev);
        if (err)
                goto out_unpin;

        if (read_domains & I915_GEM_DOMAIN_WC)
                err = i915_gem_object_set_to_wc_domain(obj, write_domain);
        else if (read_domains & I915_GEM_DOMAIN_GTT)
                err = i915_gem_object_set_to_gtt_domain(obj, write_domain);
        else
                err = i915_gem_object_set_to_cpu_domain(obj, write_domain);

        /* And bump the LRU for this access */
        i915_gem_object_bump_inactive_ggtt(obj);

        mutex_unlock(&dev->struct_mutex);

        if (write_domain != 0)
                intel_fb_obj_invalidate(obj,
                                        fb_write_origin(obj, write_domain));

out_unpin:
        i915_gem_object_unpin_pages(obj);
out:
        i915_gem_object_put(obj);
        return err;
}

/**
 * Called when user space has done writes to this buffer
 * @dev: drm device
 * @data: ioctl data blob
 * @file: drm file
 */
int
i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
                         struct drm_file *file)
{
        struct drm_i915_gem_sw_finish *args = data;
        struct drm_i915_gem_object *obj;

        obj = i915_gem_object_lookup(file, args->handle);
        if (!obj)
                return -ENOENT;

        /*
         * Proxy objects are barred from CPU access, so there is no
         * need to ban sw_finish as it is a nop.
         */

        /* Pinned buffers may be scanout, so flush the cache */
        i915_gem_object_flush_if_display(obj);
        i915_gem_object_put(obj);

        return 0;
}

/**
 * i915_gem_mmap_ioctl - Maps the contents of an object, returning the address
 *                       it is mapped to.
 * @dev: drm device
 * @data: ioctl data blob
 * @file: drm file
 *
 * While the mapping holds a reference on the contents of the object, it doesn't
 * imply a ref on the object itself.
 *
 * IMPORTANT:
 *
 * DRM driver writers who look a this function as an example for how to do GEM
 * mmap support, please don't implement mmap support like here. The modern way
 * to implement DRM mmap support is with an mmap offset ioctl (like
 * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly.
 * That way debug tooling like valgrind will understand what's going on, hiding
 * the mmap call in a driver private ioctl will break that. The i915 driver only
 * does cpu mmaps this way because we didn't know better.
 */
int
i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
                    struct drm_file *file)
{
        struct drm_i915_gem_mmap *args = data;
        struct drm_i915_gem_object *obj;
        unsigned long addr;

        if (args->flags & ~(I915_MMAP_WC))
                return -EINVAL;

        if (args->flags & I915_MMAP_WC && !boot_cpu_has(X86_FEATURE_PAT))
                return -ENODEV;

        obj = i915_gem_object_lookup(file, args->handle);
        if (!obj)
                return -ENOENT;

        /* prime objects have no backing filp to GEM mmap
         * pages from.
         */
        if (!obj->base.filp) {
                i915_gem_object_put(obj);
                return -ENXIO;
        }

        addr = vm_mmap(obj->base.filp, 0, args->size,
                       PROT_READ | PROT_WRITE, MAP_SHARED,
                       args->offset);
        if (args->flags & I915_MMAP_WC) {
                struct mm_struct *mm = current->mm;
                struct vm_area_struct *vma;

                if (down_write_killable(&mm->mmap_sem)) {
                        i915_gem_object_put(obj);
                        return -EINTR;
                }
                vma = find_vma(mm, addr);
                if (vma)
                        vma->vm_page_prot =
                                pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
                else
                        addr = -ENOMEM;
                up_write(&mm->mmap_sem);

                /* This may race, but that's ok, it only gets set */
                WRITE_ONCE(obj->frontbuffer_ggtt_origin, ORIGIN_CPU);
        }
        i915_gem_object_put(obj);
        if (IS_ERR((void *)addr))
                return addr;

        args->addr_ptr = (uint64_t) addr;

        return 0;
}

static unsigned int tile_row_pages(struct drm_i915_gem_object *obj)
{
        return i915_gem_object_get_tile_row_size(obj) >> PAGE_SHIFT;
}

/**
 * i915_gem_mmap_gtt_version - report the current feature set for GTT mmaps
 *
 * A history of the GTT mmap interface:
 *
 * 0 - Everything had to fit into the GTT. Both parties of a memcpy had to
 *     aligned and suitable for fencing, and still fit into the available
 *     mappable space left by the pinned display objects. A classic problem
 *     we called the page-fault-of-doom where we would ping-pong between
 *     two objects that could not fit inside the GTT and so the memcpy
 *     would page one object in at the expense of the other between every
 *     single byte.
 *
 * 1 - Objects can be any size, and have any compatible fencing (X Y, or none
 *     as set via i915_gem_set_tiling() [DRM_I915_GEM_SET_TILING]). If the
 *     object is too large for the available space (or simply too large
 *     for the mappable aperture!), a view is created instead and faulted
 *     into userspace. (This view is aligned and sized appropriately for
 *     fenced access.)
 *
 * 2 - Recognise WC as a separate cache domain so that we can flush the
 *     delayed writes via GTT before performing direct access via WC.
 *
 * Restrictions:
 *
 *  * snoopable objects cannot be accessed via the GTT. It can cause machine
 *    hangs on some architectures, corruption on others. An attempt to service
 *    a GTT page fault from a snoopable object will generate a SIGBUS.
 *
 *  * the object must be able to fit into RAM (physical memory, though no
 *    limited to the mappable aperture).
 *
 *
 * Caveats:
 *
 *  * a new GTT page fault will synchronize rendering from the GPU and flush
 *    all data to system memory. Subsequent access will not be synchronized.
 *
 *  * all mappings are revoked on runtime device suspend.
 *
 *  * there are only 8, 16 or 32 fence registers to share between all users
 *    (older machines require fence register for display and blitter access
 *    as well). Contention of the fence registers will cause the previous users
 *    to be unmapped and any new access will generate new page faults.
 *
 *  * running out of memory while servicing a fault may generate a SIGBUS,
 *    rather than the expected SIGSEGV.
 */
int i915_gem_mmap_gtt_version(void)
{
        return 2;
}

static inline struct i915_ggtt_view
compute_partial_view(struct drm_i915_gem_object *obj,
                     pgoff_t page_offset,
                     unsigned int chunk)
{
        struct i915_ggtt_view view;

        if (i915_gem_object_is_tiled(obj))
                chunk = roundup(chunk, tile_row_pages(obj));

        view.type = I915_GGTT_VIEW_PARTIAL;
        view.partial.offset = rounddown(page_offset, chunk);
        view.partial.size =
                min_t(unsigned int, chunk,
                      (obj->base.size >> PAGE_SHIFT) - view.partial.offset);

        /* If the partial covers the entire object, just create a normal VMA. */
        if (chunk >= obj->base.size >> PAGE_SHIFT)
                view.type = I915_GGTT_VIEW_NORMAL;

        return view;
}

/**
 * i915_gem_fault - fault a page into the GTT
 * @vmf: fault info
 *
 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
 * from userspace.  The fault handler takes care of binding the object to
 * the GTT (if needed), allocating and programming a fence register (again,
 * only if needed based on whether the old reg is still valid or the object
 * is tiled) and inserting a new PTE into the faulting process.
 *
 * Note that the faulting process may involve evicting existing objects
 * from the GTT and/or fence registers to make room.  So performance may
 * suffer if the GTT working set is large or there are few fence registers
 * left.
 *
 * The current feature set supported by i915_gem_fault() and thus GTT mmaps
 * is exposed via I915_PARAM_MMAP_GTT_VERSION (see i915_gem_mmap_gtt_version).
 */
int i915_gem_fault(struct vm_fault *vmf)
{
#define MIN_CHUNK_PAGES ((1 << 20) >> PAGE_SHIFT) /* 1 MiB */
        struct vm_area_struct *area = vmf->vma;
        struct drm_i915_gem_object *obj = to_intel_bo(area->vm_private_data);
        struct drm_device *dev = obj->base.dev;
        struct drm_i915_private *dev_priv = to_i915(dev);
        struct i915_ggtt *ggtt = &dev_priv->ggtt;
        bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
        struct i915_vma *vma;
        pgoff_t page_offset;
        unsigned int flags;
        int ret;

        /* We don't use vmf->pgoff since that has the fake offset */
        page_offset = (vmf->address - area->vm_start) >> PAGE_SHIFT;

        trace_i915_gem_object_fault(obj, page_offset, true, write);

        /* Try to flush the object off the GPU first without holding the lock.
         * Upon acquiring the lock, we will perform our sanity checks and then
         * repeat the flush holding the lock in the normal manner to catch cases
         * where we are gazumped.
         */
        ret = i915_gem_object_wait(obj,
                                   I915_WAIT_INTERRUPTIBLE,
                                   MAX_SCHEDULE_TIMEOUT,
                                   NULL);
        if (ret)
                goto err;

        ret = i915_gem_object_pin_pages(obj);
        if (ret)
                goto err;

        intel_runtime_pm_get(dev_priv);

        ret = i915_mutex_lock_interruptible(dev);
        if (ret)
                goto err_rpm;

        /* Access to snoopable pages through the GTT is incoherent. */
        if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev_priv)) {
                ret = -EFAULT;
                goto err_unlock;
        }

        /* If the object is smaller than a couple of partial vma, it is
         * not worth only creating a single partial vma - we may as well
         * clear enough space for the full object.
         */
        flags = PIN_MAPPABLE;
        if (obj->base.size > 2 * MIN_CHUNK_PAGES << PAGE_SHIFT)
                flags |= PIN_NONBLOCK | PIN_NONFAULT;

        /* Now pin it into the GTT as needed */
        vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, flags);
        if (IS_ERR(vma)) {
                /* Use a partial view if it is bigger than available space */
                struct i915_ggtt_view view =
                        compute_partial_view(obj, page_offset, MIN_CHUNK_PAGES);

                /* Userspace is now writing through an untracked VMA, abandon
                 * all hope that the hardware is able to track future writes.
                 */
                obj->frontbuffer_ggtt_origin = ORIGIN_CPU;

                vma = i915_gem_object_ggtt_pin(obj, &view, 0, 0, PIN_MAPPABLE);
        }
        if (IS_ERR(vma)) {
                ret = PTR_ERR(vma);
                goto err_unlock;
        }

        ret = i915_gem_object_set_to_gtt_domain(obj, write);
        if (ret)
                goto err_unpin;

        ret = i915_vma_pin_fence(vma);
        if (ret)
                goto err_unpin;

        /* Finally, remap it using the new GTT offset */
        ret = remap_io_mapping(area,
                               area->vm_start + (vma->ggtt_view.partial.offset << PAGE_SHIFT),
                               (ggtt->gmadr.start + vma->node.start) >> PAGE_SHIFT,
                               min_t(u64, vma->size, area->vm_end - area->vm_start),
                               &ggtt->iomap);
        if (ret)
                goto err_fence;

        /* Mark as being mmapped into userspace for later revocation */
        assert_rpm_wakelock_held(dev_priv);
        if (!i915_vma_set_userfault(vma) && !obj->userfault_count++)
                list_add(&obj->userfault_link, &dev_priv->mm.userfault_list);
        GEM_BUG_ON(!obj->userfault_count);

        i915_vma_set_ggtt_write(vma);

err_fence:
        i915_vma_unpin_fence(vma);
err_unpin:
        __i915_vma_unpin(vma);
err_unlock:
        mutex_unlock(&dev->struct_mutex);
err_rpm:
        intel_runtime_pm_put(dev_priv);
        i915_gem_object_unpin_pages(obj);

err:
        switch (ret) {
        case -EIO:
                /*
                 * We eat errors when the gpu is terminally wedged to avoid
                 * userspace unduly crashing (gl has no provisions for mmaps to
                 * fail). But any other -EIO isn't ours (e.g. swap in failure)
                 * and so needs to be reported.
                 */
                if (!i915_terminally_wedged(&dev_priv->gpu_error)) {
                        ret = VM_FAULT_SIGBUS;
                        break;
                }
        case -EAGAIN:
                /*
                 * EAGAIN means the gpu is hung and we'll wait for the error
                 * handler to reset everything when re-faulting in
                 * i915_mutex_lock_interruptible.
                 */
        case 0:
        case -ERESTARTSYS:
        case -EINTR:
        case -EBUSY:
                /*
                 * EBUSY is ok: this just means that another thread
                 * already did the job.
                 */
                ret = VM_FAULT_NOPAGE;
                break;
        case -ENOMEM:
                ret = VM_FAULT_OOM;
                break;
        case -ENOSPC:
        case -EFAULT:
                ret = VM_FAULT_SIGBUS;
                break;
        default:
                WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
                ret = VM_FAULT_SIGBUS;
                break;
        }
        return ret;
}

static void __i915_gem_object_release_mmap(struct drm_i915_gem_object *obj)
{
        struct i915_vma *vma;

        GEM_BUG_ON(!obj->userfault_count);

        obj->userfault_count = 0;
        list_del(&obj->userfault_link);
        drm_vma_node_unmap(&obj->base.vma_node,
                           obj->base.dev->anon_inode->i_mapping);

        for_each_ggtt_vma(vma, obj)
                i915_vma_unset_userfault(vma);
}

/**
 * i915_gem_release_mmap - remove physical page mappings
 * @obj: obj in question
 *
 * Preserve the reservation of the mmapping with the DRM core code, but
 * relinquish ownership of the pages back to the system.
 *
 * It is vital that we remove the page mapping if we have mapped a tiled
 * object through the GTT and then lose the fence register due to
 * resource pressure. Similarly if the object has been moved out of the
 * aperture, than pages mapped into userspace must be revoked. Removing the
 * mapping will then trigger a page fault on the next user access, allowing
 * fixup by i915_gem_fault().
 */
void
i915_gem_release_mmap(struct drm_i915_gem_object *obj)
{
        struct drm_i915_private *i915 = to_i915(obj->base.dev);

        /* Serialisation between user GTT access and our code depends upon
         * revoking the CPU's PTE whilst the mutex is held. The next user
         * pagefault then has to wait until we release the mutex.
         *
         * Note that RPM complicates somewhat by adding an additional
         * requirement that operations to the GGTT be made holding the RPM
         * wakeref.
         */
        lockdep_assert_held(&i915->drm.struct_mutex);
        intel_runtime_pm_get(i915);

        if (!obj->userfault_count)
                goto out;

        __i915_gem_object_release_mmap(obj);

        /* Ensure that the CPU's PTE are revoked and there are not outstanding
         * memory transactions from userspace before we return. The TLB
         * flushing implied above by changing the PTE above *should* be
         * sufficient, an extra barrier here just provides us with a bit
         * of paranoid documentation about our requirement to serialise
         * memory writes before touching registers / GSM.
         */
        wmb();

out:
        intel_runtime_pm_put(i915);
}

void i915_gem_runtime_suspend(struct drm_i915_private *dev_priv)
{
        struct drm_i915_gem_object *obj, *on;
        int i;

        /*
         * Only called during RPM suspend. All users of the userfault_list
         * must be holding an RPM wakeref to ensure that this can not
         * run concurrently with themselves (and use the struct_mutex for
         * protection between themselves).
         */

        list_for_each_entry_safe(obj, on,
                                 &dev_priv->mm.userfault_list, userfault_link)
                __i915_gem_object_release_mmap(obj);

        /* The fence will be lost when the device powers down. If any were
         * in use by hardware (i.e. they are pinned), we should not be powering
         * down! All other fences will be reacquired by the user upon waking.
         */
        for (i = 0; i < dev_priv->num_fence_regs; i++) {
                struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];

                /* Ideally we want to assert that the fence register is not
                 * live at this point (i.e. that no piece of code will be
                 * trying to write through fence + GTT, as that both violates
                 * our tracking of activity and associated locking/barriers,
                 * but also is illegal given that the hw is powered down).
                 *
                 * Previously we used reg->pin_count as a "liveness" indicator.
                 * That is not sufficient, and we need a more fine-grained
                 * tool if we want to have a sanity check here.
                 */

                if (!reg->vma)
                        continue;

                GEM_BUG_ON(i915_vma_has_userfault(reg->vma));
                reg->dirty = true;
        }
}

static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
{
        struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
        int err;

        err = drm_gem_create_mmap_offset(&obj->base);
        if (likely(!err))
                return 0;

        /* Attempt to reap some mmap space from dead objects */
        do {
                err = i915_gem_wait_for_idle(dev_priv, I915_WAIT_INTERRUPTIBLE);
                if (err)
                        break;

                i915_gem_drain_freed_objects(dev_priv);
                err = drm_gem_create_mmap_offset(&obj->base);
                if (!err)
                        break;

        } while (flush_delayed_work(&dev_priv->gt.retire_work));

        return err;
}

static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
{
        drm_gem_free_mmap_offset(&obj->base);
}

int
i915_gem_mmap_gtt(struct drm_file *file,
                  struct drm_device *dev,
                  uint32_t handle,
                  uint64_t *offset)
{
        struct drm_i915_gem_object *obj;
        int ret;

        obj = i915_gem_object_lookup(file, handle);
        if (!obj)
                return -ENOENT;

        ret = i915_gem_object_create_mmap_offset(obj);
        if (ret == 0)
                *offset = drm_vma_node_offset_addr(&obj->base.vma_node);

        i915_gem_object_put(obj);
        return ret;
}

/**
 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
 * @dev: DRM device
 * @data: GTT mapping ioctl data
 * @file: GEM object info
 *
 * Simply returns the fake offset to userspace so it can mmap it.
 * The mmap call will end up in drm_gem_mmap(), which will set things
 * up so we can get faults in the handler above.
 *
 * The fault handler will take care of binding the object into the GTT
 * (since it may have been evicted to make room for something), allocating
 * a fence register, and mapping the appropriate aperture address into
 * userspace.
 */
int
i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
                        struct drm_file *file)
{
        struct drm_i915_gem_mmap_gtt *args = data;

        return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
}

/* Immediately discard the backing storage */
static void
i915_gem_object_truncate(struct drm_i915_gem_object *obj)
{
        i915_gem_object_free_mmap_offset(obj);

        if (obj->base.filp == NULL)
                return;

        /* Our goal here is to return as much of the memory as
         * is possible back to the system as we are called from OOM.
         * To do this we must instruct the shmfs to drop all of its
         * backing pages, *now*.
         */
        shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1);
        obj->mm.madv = __I915_MADV_PURGED;
        obj->mm.pages = ERR_PTR(-EFAULT);
}

/* Try to discard unwanted pages */
void __i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
{
        struct address_space *mapping;

        lockdep_assert_held(&obj->mm.lock);
        GEM_BUG_ON(i915_gem_object_has_pages(obj));

        switch (obj->mm.madv) {
        case I915_MADV_DONTNEED:
                i915_gem_object_truncate(obj);
        case __I915_MADV_PURGED:
                return;
        }

        if (obj->base.filp == NULL)
                return;

        mapping = obj->base.filp->f_mapping,
        invalidate_mapping_pages(mapping, 0, (loff_t)-1);
}

static void
i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj,
                              struct sg_table *pages)
{
        struct sgt_iter sgt_iter;
        struct page *page;

        __i915_gem_object_release_shmem(obj, pages, true);

        i915_gem_gtt_finish_pages(obj, pages);

        if (i915_gem_object_needs_bit17_swizzle(obj))
                i915_gem_object_save_bit_17_swizzle(obj, pages);

        for_each_sgt_page(page, sgt_iter, pages) {
                if (obj->mm.dirty)
                        set_page_dirty(page);

                if (obj->mm.madv == I915_MADV_WILLNEED)
                        mark_page_accessed(page);

                put_page(page);
        }
        obj->mm.dirty = false;

        sg_free_table(pages);
        kfree(pages);
}

static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj)
{
        struct radix_tree_iter iter;
        void __rcu **slot;

        rcu_read_lock();
        radix_tree_for_each_slot(slot, &obj->mm.get_page.radix, &iter, 0)
                radix_tree_delete(&obj->mm.get_page.radix, iter.index);
        rcu_read_unlock();
}

void __i915_gem_object_put_pages(struct drm_i915_gem_object *obj,
                                 enum i915_mm_subclass subclass)
{
        struct drm_i915_private *i915 = to_i915(obj->base.dev);
        struct sg_table *pages;

        if (i915_gem_object_has_pinned_pages(obj))
                return;

        GEM_BUG_ON(obj->bind_count);
        if (!i915_gem_object_has_pages(obj))
                return;

        /* May be called by shrinker from within get_pages() (on another bo) */
        mutex_lock_nested(&obj->mm.lock, subclass);
        if (unlikely(atomic_read(&obj->mm.pages_pin_count)))
                goto unlock;

        /* ->put_pages might need to allocate memory for the bit17 swizzle
         * array, hence protect them from being reaped by removing them from gtt
         * lists early. */
        pages = fetch_and_zero(&obj->mm.pages);
        GEM_BUG_ON(!pages);

        spin_lock(&i915->mm.obj_lock);
        list_del(&obj->mm.link);
        spin_unlock(&i915->mm.obj_lock);

        if (obj->mm.mapping) {
                void *ptr;

                ptr = page_mask_bits(obj->mm.mapping);
                if (is_vmalloc_addr(ptr))
                        vunmap(ptr);
                else
                        kunmap(kmap_to_page(ptr));

                obj->mm.mapping = NULL;
        }

        __i915_gem_object_reset_page_iter(obj);

        if (!IS_ERR(pages))
                obj->ops->put_pages(obj, pages);

        obj->mm.page_sizes.phys = obj->mm.page_sizes.sg = 0;

unlock:
        mutex_unlock(&obj->mm.lock);
}

static bool i915_sg_trim(struct sg_table *orig_st)
{
        struct sg_table new_st;
        struct scatterlist *sg, *new_sg;
        unsigned int i;

        if (orig_st->nents == orig_st->orig_nents)
                return false;

        if (sg_alloc_table(&new_st, orig_st->nents, GFP_KERNEL | __GFP_NOWARN))
                return false;

        new_sg = new_st.sgl;
        for_each_sg(orig_st->sgl, sg, orig_st->nents, i) {
                sg_set_page(new_sg, sg_page(sg), sg->length, 0);
                /* called before being DMA mapped, no need to copy sg->dma_* */
                new_sg = sg_next(new_sg);
        }
        GEM_BUG_ON(new_sg); /* Should walk exactly nents and hit the end */

        sg_free_table(orig_st);

        *orig_st = new_st;
        return true;
}

static int i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
{
        struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
        const unsigned long page_count = obj->base.size / PAGE_SIZE;
        unsigned long i;
        struct address_space *mapping;
        struct sg_table *st;
        struct scatterlist *sg;
        struct sgt_iter sgt_iter;
        struct page *page;
        unsigned long last_pfn = 0;     /* suppress gcc warning */
        unsigned int max_segment = i915_sg_segment_size();
        unsigned int sg_page_sizes;
        gfp_t noreclaim;
        int ret;

        /* Assert that the object is not currently in any GPU domain. As it
         * wasn't in the GTT, there shouldn't be any way it could have been in
         * a GPU cache
         */
        GEM_BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS);
        GEM_BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS);

        st = kmalloc(sizeof(*st), GFP_KERNEL);
        if (st == NULL)
                return -ENOMEM;

rebuild_st:
        if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
                kfree(st);
                return -ENOMEM;
        }

        /* Get the list of pages out of our struct file.  They'll be pinned
         * at this point until we release them.
         *
         * Fail silently without starting the shrinker
         */
        mapping = obj->base.filp->f_mapping;
        noreclaim = mapping_gfp_constraint(mapping, ~__GFP_RECLAIM);
        noreclaim |= __GFP_NORETRY | __GFP_NOWARN;

        sg = st->sgl;
        st->nents = 0;
        sg_page_sizes = 0;
        for (i = 0; i < page_count; i++) {
                const unsigned int shrink[] = {
                        I915_SHRINK_BOUND | I915_SHRINK_UNBOUND | I915_SHRINK_PURGEABLE,
                        0,
                }, *s = shrink;
                gfp_t gfp = noreclaim;

                do {
                        page = shmem_read_mapping_page_gfp(mapping, i, gfp);
                        if (likely(!IS_ERR(page)))
                                break;

                        if (!*s) {
                                ret = PTR_ERR(page);
                                goto err_sg;
                        }

                        i915_gem_shrink(dev_priv, 2 * page_count, NULL, *s++);
                        cond_resched();

                        /* We've tried hard to allocate the memory by reaping
                         * our own buffer, now let the real VM do its job and
                         * go down in flames if truly OOM.
                         *
                         * However, since graphics tend to be disposable,
                         * defer the oom here by reporting the ENOMEM back
                         * to userspace.
                         */
                        if (!*s) {
                                /* reclaim and warn, but no oom */
                                gfp = mapping_gfp_mask(mapping);

                                /* Our bo are always dirty and so we require
                                 * kswapd to reclaim our pages (direct reclaim
                                 * does not effectively begin pageout of our
                                 * buffers on its own). However, direct reclaim
                                 * only waits for kswapd when under allocation
                                 * congestion. So as a result __GFP_RECLAIM is
                                 * unreliable and fails to actually reclaim our
                                 * dirty pages -- unless you try over and over
                                 * again with !__GFP_NORETRY. However, we still
                                 * want to fail this allocation rather than
                                 * trigger the out-of-memory killer and for
                                 * this we want __GFP_RETRY_MAYFAIL.
                                 */
                                gfp |= __GFP_RETRY_MAYFAIL;
                        }
                } while (1);

                if (!i ||
                    sg->length >= max_segment ||
                    page_to_pfn(page) != last_pfn + 1) {
                        if (i) {
                                sg_page_sizes |= sg->length;
                                sg = sg_next(sg);
                        }
                        st->nents++;
                        sg_set_page(sg, page, PAGE_SIZE, 0);
                } else {
                        sg->length += PAGE_SIZE;
                }
                last_pfn = page_to_pfn(page);

                /* Check that the i965g/gm workaround works. */
                WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
        }
        if (sg) { /* loop terminated early; short sg table */
                sg_page_sizes |= sg->length;
                sg_mark_end(sg);
        }

        /* Trim unused sg entries to avoid wasting memory. */
        i915_sg_trim(st);

        ret = i915_gem_gtt_prepare_pages(obj, st);
        if (ret) {
                /* DMA remapping failed? One possible cause is that
                 * it could not reserve enough large entries, asking
                 * for PAGE_SIZE chunks instead may be helpful.
                 */
                if (max_segment > PAGE_SIZE) {
                        for_each_sgt_page(page, sgt_iter, st)
                                put_page(page);
                        sg_free_table(st);

                        max_segment = PAGE_SIZE;
                        goto rebuild_st;
                } else {
                        dev_warn(&dev_priv->drm.pdev->dev,
                                 "Failed to DMA remap %lu pages\n",
                                 page_count);
                        goto err_pages;
                }
        }

        if (i915_gem_object_needs_bit17_swizzle(obj))
                i915_gem_object_do_bit_17_swizzle(obj, st);

        __i915_gem_object_set_pages(obj, st, sg_page_sizes);

        return 0;

err_sg:
        sg_mark_end(sg);
err_pages:
        for_each_sgt_page(page, sgt_iter, st)
                put_page(page);
        sg_free_table(st);
        kfree(st);

        /* shmemfs first checks if there is enough memory to allocate the page
         * and reports ENOSPC should there be insufficient, along with the usual
         * ENOMEM for a genuine allocation failure.
         *
         * We use ENOSPC in our driver to mean that we have run out of aperture
         * space and so want to translate the error from shmemfs back to our
         * usual understanding of ENOMEM.
         */
        if (ret == -ENOSPC)
                ret = -ENOMEM;

        return ret;
}

void __i915_gem_object_set_pages(struct drm_i915_gem_object *obj,
                                 struct sg_table *pages,
                                 unsigned int sg_page_sizes)
{
        struct drm_i915_private *i915 = to_i915(obj->base.dev);
        unsigned long supported = INTEL_INFO(i915)->page_sizes;
        int i;

        lockdep_assert_held(&obj->mm.lock);

        obj->mm.get_page.sg_pos = pages->sgl;
        obj->mm.get_page.sg_idx = 0;

        obj->mm.pages = pages;

        if (i915_gem_object_is_tiled(obj) &&
            i915->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
                GEM_BUG_ON(obj->mm.quirked);
                __i915_gem_object_pin_pages(obj);
                obj->mm.quirked = true;
        }

        GEM_BUG_ON(!sg_page_sizes);
        obj->mm.page_sizes.phys = sg_page_sizes;

        /*
         * Calculate the supported page-sizes which fit into the given
         * sg_page_sizes. This will give us the page-sizes which we may be able
         * to use opportunistically when later inserting into the GTT. For
         * example if phys=2G, then in theory we should be able to use 1G, 2M,
         * 64K or 4K pages, although in practice this will depend on a number of
         * other factors.
         */
        obj->mm.page_sizes.sg = 0;
        for_each_set_bit(i, &supported, ilog2(I915_GTT_MAX_PAGE_SIZE) + 1) {
                if (obj->mm.page_sizes.phys & ~0u << i)
                        obj->mm.page_sizes.sg |= BIT(i);
        }
        GEM_BUG_ON(!HAS_PAGE_SIZES(i915, obj->mm.page_sizes.sg));

        spin_lock(&i915->mm.obj_lock);
        list_add(&obj->mm.link, &i915->mm.unbound_list);
        spin_unlock(&i915->mm.obj_lock);
}

static int ____i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
{
        int err;

        if (unlikely(obj->mm.madv != I915_MADV_WILLNEED)) {
                DRM_DEBUG("Attempting to obtain a purgeable object\n");
                return -EFAULT;
        }

        err = obj->ops->get_pages(obj);
        GEM_BUG_ON(!err && !i915_gem_object_has_pages(obj));

        return err;
}

/* Ensure that the associated pages are gathered from the backing storage
 * and pinned into our object. i915_gem_object_pin_pages() may be called
 * multiple times before they are released by a single call to
 * i915_gem_object_unpin_pages() - once the pages are no longer referenced
 * either as a result of memory pressure (reaping pages under the shrinker)
 * or as the object is itself released.
 */
int __i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
{
        int err;

        err = mutex_lock_interruptible(&obj->mm.lock);
        if (err)
                return err;

        if (unlikely(!i915_gem_object_has_pages(obj))) {
                GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));

                err = ____i915_gem_object_get_pages(obj);
                if (err)
                        goto unlock;

                smp_mb__before_atomic();
        }
        atomic_inc(&obj->mm.pages_pin_count);

unlock:
        mutex_unlock(&obj->mm.lock);
        return err;
}

/* The 'mapping' part of i915_gem_object_pin_map() below */
static void *i915_gem_object_map(const struct drm_i915_gem_object *obj,
                                 enum i915_map_type type)
{
        unsigned long n_pages = obj->base.size >> PAGE_SHIFT;
        struct sg_table *sgt = obj->mm.pages;
        struct sgt_iter sgt_iter;
        struct page *page;
        struct page *stack_pages[32];
        struct page **pages = stack_pages;
        unsigned long i = 0;
        pgprot_t pgprot;
        void *addr;

        /* A single page can always be kmapped */
        if (n_pages == 1 && type == I915_MAP_WB)
                return kmap(sg_page(sgt->sgl));

        if (n_pages > ARRAY_SIZE(stack_pages)) {
                /* Too big for stack -- allocate temporary array instead */
                pages = kvmalloc_array(n_pages, sizeof(*pages), GFP_KERNEL);
                if (!pages)
                        return NULL;
        }

        for_each_sgt_page(page, sgt_iter, sgt)
                pages[i++] = page;

        /* Check that we have the expected number of pages */
        GEM_BUG_ON(i != n_pages);

        switch (type) {
        default:
                MISSING_CASE(type);
                /* fallthrough to use PAGE_KERNEL anyway */
        case I915_MAP_WB:
                pgprot = PAGE_KERNEL;
                break;
        case I915_MAP_WC:
                pgprot = pgprot_writecombine(PAGE_KERNEL_IO);
                break;
        }
        addr = vmap(pages, n_pages, 0, pgprot);

        if (pages != stack_pages)
                kvfree(pages);

        return addr;
}

/* get, pin, and map the pages of the object into kernel space */
void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,
                              enum i915_map_type type)
{
        enum i915_map_type has_type;
        bool pinned;
        void *ptr;
        int ret;

        if (unlikely(!i915_gem_object_has_struct_page(obj)))
                return ERR_PTR(-ENXIO);

        ret = mutex_lock_interruptible(&obj->mm.lock);
        if (ret)
                return ERR_PTR(ret);

        pinned = !(type & I915_MAP_OVERRIDE);
        type &= ~I915_MAP_OVERRIDE;

        if (!atomic_inc_not_zero(&obj->mm.pages_pin_count)) {
                if (unlikely(!i915_gem_object_has_pages(obj))) {
                        GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));

                        ret = ____i915_gem_object_get_pages(obj);
                        if (ret)
                                goto err_unlock;

                        smp_mb__before_atomic();
                }
                atomic_inc(&obj->mm.pages_pin_count);
                pinned = false;
        }
        GEM_BUG_ON(!i915_gem_object_has_pages(obj));

        ptr = page_unpack_bits(obj->mm.mapping, &has_type);
        if (ptr && has_type != type) {
                if (pinned) {
                        ret = -EBUSY;
                        goto err_unpin;
                }

                if (is_vmalloc_addr(ptr))
                        vunmap(ptr);
                else
                        kunmap(kmap_to_page(ptr));

                ptr = obj->mm.mapping = NULL;
        }

        if (!ptr) {
                ptr = i915_gem_object_map(obj, type);
                if (!ptr) {
                        ret = -ENOMEM;
                        goto err_unpin;
                }

                obj->mm.mapping = page_pack_bits(ptr, type);
        }

out_unlock:
        mutex_unlock(&obj->mm.lock);
        return ptr;

err_unpin:
        atomic_dec(&obj->mm.pages_pin_count);
err_unlock:
        ptr = ERR_PTR(ret);
        goto out_unlock;
}

static int
i915_gem_object_pwrite_gtt(struct drm_i915_gem_object *obj,
                           const struct drm_i915_gem_pwrite *arg)
{
        struct address_space *mapping = obj->base.filp->f_mapping;
        char __user *user_data = u64_to_user_ptr(arg->data_ptr);
        u64 remain, offset;
        unsigned int pg;

        /* Before we instantiate/pin the backing store for our use, we
         * can prepopulate the shmemfs filp efficiently using a write into
         * the pagecache. We avoid the penalty of instantiating all the
         * pages, important if the user is just writing to a few and never
         * uses the object on the GPU, and using a direct write into shmemfs
         * allows it to avoid the cost of retrieving a page (either swapin
         * or clearing-before-use) before it is overwritten.
         */
        if (i915_gem_object_has_pages(obj))
                return -ENODEV;

        if (obj->mm.madv != I915_MADV_WILLNEED)
                return -EFAULT;

        /* Before the pages are instantiated the object is treated as being
         * in the CPU domain. The pages will be clflushed as required before
         * use, and we can freely write into the pages directly. If userspace
         * races pwrite with any other operation; corruption will ensue -
         * that is userspace's prerogative!
         */

        remain = arg->size;
        offset = arg->offset;
        pg = offset_in_page(offset);

        do {
                unsigned int len, unwritten;
                struct page *page;
                void *data, *vaddr;
                int err;

                len = PAGE_SIZE - pg;
                if (len > remain)
                        len = remain;

                err = pagecache_write_begin(obj->base.filp, mapping,
                                            offset, len, 0,
                                            &page, &data);
                if (err < 0)
                        return err;

                vaddr = kmap(page);
                unwritten = copy_from_user(vaddr + pg, user_data, len);
                kunmap(page);

                err = pagecache_write_end(obj->base.filp, mapping,
                                          offset, len, len - unwritten,
                                          page, data);
                if (err < 0)
                        return err;

                if (unwritten)
                        return -EFAULT;

                remain -= len;
                user_data += len;
                offset += len;
                pg = 0;
        } while (remain);

        return 0;
}

static void i915_gem_context_mark_guilty(struct i915_gem_context *ctx)
{
        bool banned;

        atomic_inc(&ctx->guilty_count);

        banned = false;
        if (i915_gem_context_is_bannable(ctx)) {
                unsigned int score;

                score = atomic_add_return(CONTEXT_SCORE_GUILTY,
                                          &ctx->ban_score);
                banned = score >= CONTEXT_SCORE_BAN_THRESHOLD;

                DRM_DEBUG_DRIVER("context %s marked guilty (score %d) banned? %s\n",
                                 ctx->name, score, yesno(banned));
        }
        if (!banned)
                return;

        i915_gem_context_set_banned(ctx);
        if (!IS_ERR_OR_NULL(ctx->file_priv)) {
                atomic_inc(&ctx->file_priv->context_bans);
                DRM_DEBUG_DRIVER("client %s has had %d context banned\n",
                                 ctx->name, atomic_read(&ctx->file_priv->context_bans));
        }
}

static void i915_gem_context_mark_innocent(struct i915_gem_context *ctx)
{
        atomic_inc(&ctx->active_count);
}

struct i915_request *
i915_gem_find_active_request(struct intel_engine_cs *engine)
{
        struct i915_request *request, *active = NULL;
        unsigned long flags;

        /* We are called by the error capture and reset at a random
         * point in time. In particular, note that neither is crucially
         * ordered with an interrupt. After a hang, the GPU is dead and we
         * assume that no more writes can happen (we waited long enough for
         * all writes that were in transaction to be flushed) - adding an
         * extra delay for a recent interrupt is pointless. Hence, we do
         * not need an engine->irq_seqno_barrier() before the seqno reads.
         */
        spin_lock_irqsave(&engine->timeline->lock, flags);
        list_for_each_entry(request, &engine->timeline->requests, link) {
                if (__i915_request_completed(request, request->global_seqno))
                        continue;

                GEM_BUG_ON(request->engine != engine);
                GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
                                    &request->fence.flags));

                active = request;
                break;
        }
        spin_unlock_irqrestore(&engine->timeline->lock, flags);

        return active;
}

static bool engine_stalled(struct intel_engine_cs *engine)
{
        if (!engine->hangcheck.stalled)
                return false;

        /* Check for possible seqno movement after hang declaration */
        if (engine->hangcheck.seqno != intel_engine_get_seqno(engine)) {
                DRM_DEBUG_DRIVER("%s pardoned\n", engine->name);
                return false;
        }

        return true;
}

/*
 * Ensure irq handler finishes, and not run again.
 * Also return the active request so that we only search for it once.
 */
struct i915_request *
i915_gem_reset_prepare_engine(struct intel_engine_cs *engine)
{
        struct i915_request *request = NULL;

        /*
         * During the reset sequence, we must prevent the engine from
         * entering RC6. As the context state is undefined until we restart
         * the engine, if it does enter RC6 during the reset, the state
         * written to the powercontext is undefined and so we may lose
         * GPU state upon resume, i.e. fail to restart after a reset.
         */
        intel_uncore_forcewake_get(engine->i915, FORCEWAKE_ALL);

        /*
         * Prevent the signaler thread from updating the request
         * state (by calling dma_fence_signal) as we are processing
         * the reset. The write from the GPU of the seqno is
         * asynchronous and the signaler thread may see a different
         * value to us and declare the request complete, even though
         * the reset routine have picked that request as the active
         * (incomplete) request. This conflict is not handled
         * gracefully!
         */
        kthread_park(engine->breadcrumbs.signaler);

        /*
         * Prevent request submission to the hardware until we have
         * completed the reset in i915_gem_reset_finish(). If a request
         * is completed by one engine, it may then queue a request
         * to a second via its execlists->tasklet *just* as we are
         * calling engine->init_hw() and also writing the ELSP.
         * Turning off the execlists->tasklet until the reset is over
         * prevents the race.
         *
         * Note that this needs to be a single atomic operation on the
         * tasklet (flush existing tasks, prevent new tasks) to prevent
         * a race between reset and set-wedged. It is not, so we do the best
         * we can atm and make sure we don't lock the machine up in the more
         * common case of recursively being called from set-wedged from inside
         * i915_reset.
         */
        if (!atomic_read(&engine->execlists.tasklet.count))
                tasklet_kill(&engine->execlists.tasklet);
        tasklet_disable(&engine->execlists.tasklet);

        /*
         * We're using worker to queue preemption requests from the tasklet in
         * GuC submission mode.
         * Even though tasklet was disabled, we may still have a worker queued.
         * Let's make sure that all workers scheduled before disabling the
         * tasklet are completed before continuing with the reset.
         */
        if (engine->i915->guc.preempt_wq)
                flush_workqueue(engine->i915->guc.preempt_wq);

        if (engine->irq_seqno_barrier)
                engine->irq_seqno_barrier(engine);

        request = i915_gem_find_active_request(engine);
        if (request && request->fence.error == -EIO)
                request = ERR_PTR(-EIO); /* Previous reset failed! */

        return request;
}

int i915_gem_reset_prepare(struct drm_i915_private *dev_priv)
{
        struct intel_engine_cs *engine;
        struct i915_request *request;
        enum intel_engine_id id;
        int err = 0;

        for_each_engine(engine, dev_priv, id) {
                request = i915_gem_reset_prepare_engine(engine);
                if (IS_ERR(request)) {
                        err = PTR_ERR(request);
                        continue;
                }

                engine->hangcheck.active_request = request;
        }

        i915_gem_revoke_fences(dev_priv);

        return err;
}

static void skip_request(struct i915_request *request)
{
        void *vaddr = request->ring->vaddr;
        u32 head;

        /* As this request likely depends on state from the lost
         * context, clear out all the user operations leaving the
         * breadcrumb at the end (so we get the fence notifications).
         */
        head = request->head;
        if (request->postfix < head) {
                memset(vaddr + head, 0, request->ring->size - head);
                head = 0;
        }
        memset(vaddr + head, 0, request->postfix - head);

        dma_fence_set_error(&request->fence, -EIO);
}

static void engine_skip_context(struct i915_request *request)
{
        struct intel_engine_cs *engine = request->engine;
        struct i915_gem_context *hung_ctx = request->ctx;
        struct intel_timeline *timeline;
        unsigned long flags;

        timeline = i915_gem_context_lookup_timeline(hung_ctx, engine);

        spin_lock_irqsave(&engine->timeline->lock, flags);
        spin_lock(&timeline->lock);

        list_for_each_entry_continue(request, &engine->timeline->requests, link)
                if (request->ctx == hung_ctx)
                        skip_request(request);

        list_for_each_entry(request, &timeline->requests, link)
                skip_request(request);

        spin_unlock(&timeline->lock);
        spin_unlock_irqrestore(&engine->timeline->lock, flags);
}

/* Returns the request if it was guilty of the hang */
static struct i915_request *
i915_gem_reset_request(struct intel_engine_cs *engine,
                       struct i915_request *request)
{
        /* The guilty request will get skipped on a hung engine.
         *
         * Users of client default contexts do not rely on logical
         * state preserved between batches so it is safe to execute
         * queued requests following the hang. Non default contexts
         * rely on preserved state, so skipping a batch loses the
         * evolution of the state and it needs to be considered corrupted.
         * Executing more queued batches on top of corrupted state is
         * risky. But we take the risk by trying to advance through
         * the queued requests in order to make the client behaviour
         * more predictable around resets, by not throwing away random
         * amount of batches it has prepared for execution. Sophisticated
         * clients can use gem_reset_stats_ioctl and dma fence status
         * (exported via sync_file info ioctl on explicit fences) to observe
         * when it loses the context state and should rebuild accordingly.
         *
         * The context ban, and ultimately the client ban, mechanism are safety
         * valves if client submission ends up resulting in nothing more than
         * subsequent hangs.
         */

        if (engine_stalled(engine)) {
                i915_gem_context_mark_guilty(request->ctx);
                skip_request(request);

                /* If this context is now banned, skip all pending requests. */
                if (i915_gem_context_is_banned(request->ctx))
                        engine_skip_context(request);
        } else {
                /*
                 * Since this is not the hung engine, it may have advanced
                 * since the hang declaration. Double check by refinding
                 * the active request at the time of the reset.
                 */
                request = i915_gem_find_active_request(engine);
                if (request) {
                        i915_gem_context_mark_innocent(request->ctx);
                        dma_fence_set_error(&request->fence, -EAGAIN);

                        /* Rewind the engine to replay the incomplete rq */
                        spin_lock_irq(&engine->timeline->lock);
                        request = list_prev_entry(request, link);
                        if (&request->link == &engine->timeline->requests)
                                request = NULL;
                        spin_unlock_irq(&engine->timeline->lock);
                }
        }

        return request;
}

void i915_gem_reset_engine(struct intel_engine_cs *engine,
                           struct i915_request *request)
{
        /*
         * Make sure this write is visible before we re-enable the interrupt
         * handlers on another CPU, as tasklet_enable() resolves to just
         * a compiler barrier which is insufficient for our purpose here.
         */
        smp_store_mb(engine->irq_posted, 0);

        if (request)
                request = i915_gem_reset_request(engine, request);

        if (request) {
                DRM_DEBUG_DRIVER("resetting %s to restart from tail of request 0x%x\n",
                                 engine->name, request->global_seqno);
        }

        /* Setup the CS to resume from the breadcrumb of the hung request */
        engine->reset_hw(engine, request);
}

void i915_gem_reset(struct drm_i915_private *dev_priv)
{
        struct intel_engine_cs *engine;
        enum intel_engine_id id;

        lockdep_assert_held(&dev_priv->drm.struct_mutex);

        i915_retire_requests(dev_priv);

        for_each_engine(engine, dev_priv, id) {
                struct i915_gem_context *ctx;

                i915_gem_reset_engine(engine, engine->hangcheck.active_request);
                ctx = fetch_and_zero(&engine->last_retired_context);
                if (ctx)
                        engine->context_unpin(engine, ctx);

                /*
                 * Ostensibily, we always want a context loaded for powersaving,
                 * so if the engine is idle after the reset, send a request
                 * to load our scratch kernel_context.
                 *
                 * More mysteriously, if we leave the engine idle after a reset,
                 * the next userspace batch may hang, with what appears to be
                 * an incoherent read by the CS (presumably stale TLB). An
                 * empty request appears sufficient to paper over the glitch.
                 */
                if (intel_engine_is_idle(engine)) {
                        struct i915_request *rq;

                        rq = i915_request_alloc(engine,
                                                dev_priv->kernel_context);
                        if (!IS_ERR(rq))
                                __i915_request_add(rq, false);
                }
        }

        i915_gem_restore_fences(dev_priv);

        if (dev_priv->gt.awake) {
                intel_sanitize_gt_powersave(dev_priv);
                intel_enable_gt_powersave(dev_priv);
                if (INTEL_GEN(dev_priv) >= 6)
                        gen6_rps_busy(dev_priv);
        }
}

void i915_gem_reset_finish_engine(struct intel_engine_cs *engine)
{
        tasklet_enable(&engine->execlists.tasklet);
        kthread_unpark(engine->breadcrumbs.signaler);

        intel_uncore_forcewake_put(engine->i915, FORCEWAKE_ALL);
}

void i915_gem_reset_finish(struct drm_i915_private *dev_priv)
{
        struct intel_engine_cs *engine;
        enum intel_engine_id id;

        lockdep_assert_held(&dev_priv->drm.struct_mutex);

        for_each_engine(engine, dev_priv, id) {
                engine->hangcheck.active_request = NULL;
                i915_gem_reset_finish_engine(engine);
        }
}

static void nop_submit_request(struct i915_request *request)
{
        dma_fence_set_error(&request->fence, -EIO);

        i915_request_submit(request);
}

static void nop_complete_submit_request(struct i915_request *request)
{
        unsigned long flags;

        dma_fence_set_error(&request->fence, -EIO);

        spin_lock_irqsave(&request->engine->timeline->lock, flags);
        __i915_request_submit(request);
        intel_engine_init_global_seqno(request->engine, request->global_seqno);
        spin_unlock_irqrestore(&request->engine->timeline->lock, flags);
}

void i915_gem_set_wedged(struct drm_i915_private *i915)
{
        struct intel_engine_cs *engine;
        enum intel_engine_id id;

        if (drm_debug & DRM_UT_DRIVER) {
                struct drm_printer p = drm_debug_printer(__func__);

                for_each_engine(engine, i915, id)
                        intel_engine_dump(engine, &p, "%s\n", engine->name);
        }

        set_bit(I915_WEDGED, &i915->gpu_error.flags);
        smp_mb__after_atomic();

        /*
         * First, stop submission to hw, but do not yet complete requests by
         * rolling the global seqno forward (since this would complete requests
         * for which we haven't set the fence error to EIO yet).
         */
        for_each_engine(engine, i915, id) {
                i915_gem_reset_prepare_engine(engine);

                engine->submit_request = nop_submit_request;
                engine->schedule = NULL;
        }
        i915->caps.scheduler = 0;

        /*
         * Make sure no one is running the old callback before we proceed with
         * cancelling requests and resetting the completion tracking. Otherwise
         * we might submit a request to the hardware which never completes.
         */
        synchronize_rcu();

        for_each_engine(engine, i915, id) {
                /* Mark all executing requests as skipped */
                engine->cancel_requests(engine);

                /*
                 * Only once we've force-cancelled all in-flight requests can we
                 * start to complete all requests.
                 */
                engine->submit_request = nop_complete_submit_request;
        }

        /*
         * Make sure no request can slip through without getting completed by
         * either this call here to intel_engine_init_global_seqno, or the one
         * in nop_complete_submit_request.
         */
        synchronize_rcu();

        for_each_engine(engine, i915, id) {
                unsigned long flags;

                /*
                 * Mark all pending requests as complete so that any concurrent
                 * (lockless) lookup doesn't try and wait upon the request as we
                 * reset it.
                 */
                spin_lock_irqsave(&engine->timeline->lock, flags);
                intel_engine_init_global_seqno(engine,
                                               intel_engine_last_submit(engine));
                spin_unlock_irqrestore(&engine->timeline->lock, flags);

                i915_gem_reset_finish_engine(engine);
        }

        wake_up_all(&i915->gpu_error.reset_queue);
}

bool i915_gem_unset_wedged(struct drm_i915_private *i915)
{
        struct i915_gem_timeline *tl;
        int i;

        lockdep_assert_held(&i915->drm.struct_mutex);
        if (!test_bit(I915_WEDGED, &i915->gpu_error.flags))
                return true;

        /* Before unwedging, make sure that all pending operations
         * are flushed and errored out - we may have requests waiting upon
         * third party fences. We marked all inflight requests as EIO, and
         * every execbuf since returned EIO, for consistency we want all
         * the currently pending requests to also be marked as EIO, which
         * is done inside our nop_submit_request - and so we must wait.
         *
         * No more can be submitted until we reset the wedged bit.
         */
        list_for_each_entry(tl, &i915->gt.timelines, link) {
                for (i = 0; i < ARRAY_SIZE(tl->engine); i++) {
                        struct i915_request *rq;

                        rq = i915_gem_active_peek(&tl->engine[i].last_request,
                                                  &i915->drm.struct_mutex);
                        if (!rq)
                                continue;

                        /* We can't use our normal waiter as we want to
                         * avoid recursively trying to handle the current
                         * reset. The basic dma_fence_default_wait() installs
                         * a callback for dma_fence_signal(), which is
                         * triggered by our nop handler (indirectly, the
                         * callback enables the signaler thread which is
                         * woken by the nop_submit_request() advancing the seqno
                         * and when the seqno passes the fence, the signaler
                         * then signals the fence waking us up).
                         */
                        if (dma_fence_default_wait(&rq->fence, true,
                                                   MAX_SCHEDULE_TIMEOUT) < 0)
                                return false;
                }
        }

        /* Undo nop_submit_request. We prevent all new i915 requests from
         * being queued (by disallowing execbuf whilst wedged) so having
         * waited for all active requests above, we know the system is idle
         * and do not have to worry about a thread being inside
         * engine->submit_request() as we swap over. So unlike installing
         * the nop_submit_request on reset, we can do this from normal
         * context and do not require stop_machine().
         */
        intel_engines_reset_default_submission(i915);
        i915_gem_contexts_lost(i915);

        smp_mb__before_atomic(); /* complete takeover before enabling execbuf */
        clear_bit(I915_WEDGED, &i915->gpu_error.flags);

        return true;
}

static void
i915_gem_retire_work_handler(struct work_struct *work)
{
        struct drm_i915_private *dev_priv =
                container_of(work, typeof(*dev_priv), gt.retire_work.work);
        struct drm_device *dev = &dev_priv->drm;

        /* Come back later if the device is busy... */
        if (mutex_trylock(&dev->struct_mutex)) {
                i915_retire_requests(dev_priv);
                mutex_unlock(&dev->struct_mutex);
        }

        /*
         * Keep the retire handler running until we are finally idle.
         * We do not need to do this test under locking as in the worst-case
         * we queue the retire worker once too often.
         */
        if (READ_ONCE(dev_priv->gt.awake))
                queue_delayed_work(dev_priv->wq,
                                   &dev_priv->gt.retire_work,
                                   round_jiffies_up_relative(HZ));
}

static void shrink_caches(struct drm_i915_private *i915)
{
        /*
         * kmem_cache_shrink() discards empty slabs and reorders partially
         * filled slabs to prioritise allocating from the mostly full slabs,
         * with the aim of reducing fragmentation.
         */
        kmem_cache_shrink(i915->priorities);
        kmem_cache_shrink(i915->dependencies);
        kmem_cache_shrink(i915->requests);
        kmem_cache_shrink(i915->luts);
        kmem_cache_shrink(i915->vmas);
        kmem_cache_shrink(i915->objects);
}

struct sleep_rcu_work {
        union {
                struct rcu_head rcu;
                struct work_struct work;
        };
        struct drm_i915_private *i915;
        unsigned int epoch;
};

static inline bool
same_epoch(struct drm_i915_private *i915, unsigned int epoch)
{
        /*
         * There is a small chance that the epoch wrapped since we started
         * sleeping. If we assume that epoch is at least a u32, then it will
         * take at least 2^32 * 100ms for it to wrap, or about 326 years.
         */
        return epoch == READ_ONCE(i915->gt.epoch);
}

static void __sleep_work(struct work_struct *work)
{
        struct sleep_rcu_work *s = container_of(work, typeof(*s), work);
        struct drm_i915_private *i915 = s->i915;
        unsigned int epoch = s->epoch;

        kfree(s);
        if (same_epoch(i915, epoch))
                shrink_caches(i915);
}

static void __sleep_rcu(struct rcu_head *rcu)
{
        struct sleep_rcu_work *s = container_of(rcu, typeof(*s), rcu);
        struct drm_i915_private *i915 = s->i915;

        if (same_epoch(i915, s->epoch)) {
                INIT_WORK(&s->work, __sleep_work);
                queue_work(i915->wq, &s->work);
        } else {
                kfree(s);
        }
}

static inline bool
new_requests_since_last_retire(const struct drm_i915_private *i915)
{
        return (READ_ONCE(i915->gt.active_requests) ||
                work_pending(&i915->gt.idle_work.work));
}

static void
i915_gem_idle_work_handler(struct work_struct *work)
{
        struct drm_i915_private *dev_priv =
                container_of(work, typeof(*dev_priv), gt.idle_work.work);
        unsigned int epoch = I915_EPOCH_INVALID;
        bool rearm_hangcheck;

        if (!READ_ONCE(dev_priv->gt.awake))
                return;

        /*
         * Wait for last execlists context complete, but bail out in case a
         * new request is submitted. As we don't trust the hardware, we
         * continue on if the wait times out. This is necessary to allow
         * the machine to suspend even if the hardware dies, and we will
         * try to recover in resume (after depriving the hardware of power,
         * it may be in a better mmod).
         */
        __wait_for(if (new_requests_since_last_retire(dev_priv)) return,
                   intel_engines_are_idle(dev_priv),
                   I915_IDLE_ENGINES_TIMEOUT * 1000,
                   10, 500);

        rearm_hangcheck =
                cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work);

        if (!mutex_trylock(&dev_priv->drm.struct_mutex)) {
                /* Currently busy, come back later */
                mod_delayed_work(dev_priv->wq,
                                 &dev_priv->gt.idle_work,
                                 msecs_to_jiffies(50));
                goto out_rearm;
        }

        /*
         * New request retired after this work handler started, extend active
         * period until next instance of the work.
         */
        if (new_requests_since_last_retire(dev_priv))
                goto out_unlock;

        /*
         * Be paranoid and flush a concurrent interrupt to make sure
         * we don't reactivate any irq tasklets after parking.
         *
         * FIXME: Note that even though we have waited for execlists to be idle,
         * there may still be an in-flight interrupt even though the CSB
         * is now empty. synchronize_irq() makes sure that a residual interrupt
         * is completed before we continue, but it doesn't prevent the HW from
         * raising a spurious interrupt later. To complete the shield we should
         * coordinate disabling the CS irq with flushing the interrupts.
         */
        synchronize_irq(dev_priv->drm.irq);

        intel_engines_park(dev_priv);
        i915_gem_timelines_park(dev_priv);

        i915_pmu_gt_parked(dev_priv);

        GEM_BUG_ON(!dev_priv->gt.awake);
        dev_priv->gt.awake = false;
        epoch = dev_priv->gt.epoch;
        GEM_BUG_ON(epoch == I915_EPOCH_INVALID);
        rearm_hangcheck = false;

        if (INTEL_GEN(dev_priv) >= 6)
                gen6_rps_idle(dev_priv);

        intel_display_power_put(dev_priv, POWER_DOMAIN_GT_IRQ);

        intel_runtime_pm_put(dev_priv);
out_unlock:
        mutex_unlock(&dev_priv->drm.struct_mutex);

out_rearm:
        if (rearm_hangcheck) {
                GEM_BUG_ON(!dev_priv->gt.awake);
                i915_queue_hangcheck(dev_priv);
        }

        /*
         * When we are idle, it is an opportune time to reap our caches.
         * However, we have many objects that utilise RCU and the ordered
         * i915->wq that this work is executing on. To try and flush any
         * pending frees now we are idle, we first wait for an RCU grace
         * period, and then queue a task (that will run last on the wq) to
         * shrink and re-optimize the caches.
         */
        if (same_epoch(dev_priv, epoch)) {
                struct sleep_rcu_work *s = kmalloc(sizeof(*s), GFP_KERNEL);
                if (s) {
                        s->i915 = dev_priv;
                        s->epoch = epoch;
                        call_rcu(&s->rcu, __sleep_rcu);
                }
        }
}

void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file)
{
        struct drm_i915_private *i915 = to_i915(gem->dev);
        struct drm_i915_gem_object *obj = to_intel_bo(gem);
        struct drm_i915_file_private *fpriv = file->driver_priv;
        struct i915_lut_handle *lut, *ln;

        mutex_lock(&i915->drm.struct_mutex);

        list_for_each_entry_safe(lut, ln, &obj->lut_list, obj_link) {
                struct i915_gem_context *ctx = lut->ctx;
                struct i915_vma *vma;

                GEM_BUG_ON(ctx->file_priv == ERR_PTR(-EBADF));
                if (ctx->file_priv != fpriv)
                        continue;

                vma = radix_tree_delete(&ctx->handles_vma, lut->handle);
                GEM_BUG_ON(vma->obj != obj);

                /* We allow the process to have multiple handles to the same
                 * vma, in the same fd namespace, by virtue of flink/open.
                 */
                GEM_BUG_ON(!vma->open_count);
                if (!--vma->open_count && !i915_vma_is_ggtt(vma))
                        i915_vma_close(vma);

                list_del(&lut->obj_link);
                list_del(&lut->ctx_link);

                kmem_cache_free(i915->luts, lut);
                __i915_gem_object_release_unless_active(obj);
        }

        mutex_unlock(&i915->drm.struct_mutex);
}

static unsigned long to_wait_timeout(s64 timeout_ns)
{
        if (timeout_ns < 0)
                return MAX_SCHEDULE_TIMEOUT;

        if (timeout_ns == 0)
                return 0;

        return nsecs_to_jiffies_timeout(timeout_ns);
}

/**
 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
 * @dev: drm device pointer
 * @data: ioctl data blob
 * @file: drm file pointer
 *
 * Returns 0 if successful, else an error is returned with the remaining time in
 * the timeout parameter.
 *  -ETIME: object is still busy after timeout
 *  -ERESTARTSYS: signal interrupted the wait
 *  -ENONENT: object doesn't exist
 * Also possible, but rare:
 *  -EAGAIN: incomplete, restart syscall
 *  -ENOMEM: damn
 *  -ENODEV: Internal IRQ fail
 *  -E?: The add request failed
 *
 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
 * non-zero timeout parameter the wait ioctl will wait for the given number of
 * nanoseconds on an object becoming unbusy. Since the wait itself does so
 * without holding struct_mutex the object may become re-busied before this
 * function completes. A similar but shorter * race condition exists in the busy
 * ioctl
 */
int
i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
{
        struct drm_i915_gem_wait *args = data;
        struct drm_i915_gem_object *obj;
        ktime_t start;
        long ret;

        if (args->flags != 0)
                return -EINVAL;

        obj = i915_gem_object_lookup(file, args->bo_handle);
        if (!obj)
                return -ENOENT;

        start = ktime_get();

        ret = i915_gem_object_wait(obj,
                                   I915_WAIT_INTERRUPTIBLE | I915_WAIT_ALL,
                                   to_wait_timeout(args->timeout_ns),
                                   to_rps_client(file));

        if (args->timeout_ns > 0) {
                args->timeout_ns -= ktime_to_ns(ktime_sub(ktime_get(), start));
                if (args->timeout_ns < 0)
                        args->timeout_ns = 0;

                /*
                 * Apparently ktime isn't accurate enough and occasionally has a
                 * bit of mismatch in the jiffies<->nsecs<->ktime loop. So patch
                 * things up to make the test happy. We allow up to 1 jiffy.
                 *
                 * This is a regression from the timespec->ktime conversion.
                 */
                if (ret == -ETIME && !nsecs_to_jiffies(args->timeout_ns))
                        args->timeout_ns = 0;

                /* Asked to wait beyond the jiffie/scheduler precision? */
                if (ret == -ETIME && args->timeout_ns)
                        ret = -EAGAIN;
        }

        i915_gem_object_put(obj);
        return ret;
}

static int wait_for_timeline(struct i915_gem_timeline *tl, unsigned int flags)
{
        int ret, i;

        for (i = 0; i < ARRAY_SIZE(tl->engine); i++) {
                ret = i915_gem_active_wait(&tl->engine[i].last_request, flags);
                if (ret)
                        return ret;
        }

        return 0;
}

static int wait_for_engines(struct drm_i915_private *i915)
{
        if (wait_for(intel_engines_are_idle(i915), I915_IDLE_ENGINES_TIMEOUT)) {
                dev_err(i915->drm.dev,
                        "Failed to idle engines, declaring wedged!\n");
                if (drm_debug & DRM_UT_DRIVER) {
                        struct drm_printer p = drm_debug_printer(__func__);
                        struct intel_engine_cs *engine;
                        enum intel_engine_id id;

                        for_each_engine(engine, i915, id)
                                intel_engine_dump(engine, &p,
                                                  "%s\n", engine->name);
                }

                i915_gem_set_wedged(i915);
                return -EIO;
        }

        return 0;
}

int i915_gem_wait_for_idle(struct drm_i915_private *i915, unsigned int flags)
{
        int ret;

        /* If the device is asleep, we have no requests outstanding */
        if (!READ_ONCE(i915->gt.awake))
                return 0;

        if (flags & I915_WAIT_LOCKED) {
                struct i915_gem_timeline *tl;

                lockdep_assert_held(&i915->drm.struct_mutex);

                list_for_each_entry(tl, &i915->gt.timelines, link) {
                        ret = wait_for_timeline(tl, flags);
                        if (ret)
                                return ret;
                }
                i915_retire_requests(i915);

                ret = wait_for_engines(i915);
        } else {
                ret = wait_for_timeline(&i915->gt.global_timeline, flags);
        }

        return ret;
}

static void __i915_gem_object_flush_for_display(struct drm_i915_gem_object *obj)
{
        /*
         * We manually flush the CPU domain so that we can override and
         * force the flush for the display, and perform it asyncrhonously.
         */
        flush_write_domain(obj, ~I915_GEM_DOMAIN_CPU);
        if (obj->cache_dirty)
                i915_gem_clflush_object(obj, I915_CLFLUSH_FORCE);
        obj->write_domain = 0;
}

void i915_gem_object_flush_if_display(struct drm_i915_gem_object *obj)
{
        if (!READ_ONCE(obj->pin_global))
                return;

        mutex_lock(&obj->base.dev->struct_mutex);
        __i915_gem_object_flush_for_display(obj);
        mutex_unlock(&obj->base.dev->struct_mutex);
}

/**
 * Moves a single object to the WC read, and possibly write domain.
 * @obj: object to act on
 * @write: ask for write access or read only
 *
 * This function returns when the move is complete, including waiting on
 * flushes to occur.
 */
int
i915_gem_object_set_to_wc_domain(struct drm_i915_gem_object *obj, bool write)
{
        int ret;

        lockdep_assert_held(&obj->base.dev->struct_mutex);

        ret = i915_gem_object_wait(obj,
                                   I915_WAIT_INTERRUPTIBLE |
                                   I915_WAIT_LOCKED |
                                   (write ? I915_WAIT_ALL : 0),
                                   MAX_SCHEDULE_TIMEOUT,
                                   NULL);
        if (ret)
                return ret;

        if (obj->write_domain == I915_GEM_DOMAIN_WC)
                return 0;

        /* Flush and acquire obj->pages so that we are coherent through
         * direct access in memory with previous cached writes through
         * shmemfs and that our cache domain tracking remains valid.
         * For example, if the obj->filp was moved to swap without us
         * being notified and releasing the pages, we would mistakenly
         * continue to assume that the obj remained out of the CPU cached
         * domain.
         */
        ret = i915_gem_object_pin_pages(obj);
        if (ret)
                return ret;

        flush_write_domain(obj, ~I915_GEM_DOMAIN_WC);

        /* Serialise direct access to this object with the barriers for
         * coherent writes from the GPU, by effectively invalidating the
         * WC domain upon first access.
         */
        if ((obj->read_domains & I915_GEM_DOMAIN_WC) == 0)
                mb();

        /* It should now be out of any other write domains, and we can update
         * the domain values for our changes.
         */
        GEM_BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_WC) != 0);
        obj->read_domains |= I915_GEM_DOMAIN_WC;
        if (write) {
                obj->read_domains = I915_GEM_DOMAIN_WC;
                obj->write_domain = I915_GEM_DOMAIN_WC;
                obj->mm.dirty = true;
        }

        i915_gem_object_unpin_pages(obj);
        return 0;
}

/**
 * Moves a single object to the GTT read, and possibly write domain.
 * @obj: object to act on
 * @write: ask for write access or read only
 *
 * This function returns when the move is complete, including waiting on
 * flushes to occur.
 */
int
i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
{
        int ret;

        lockdep_assert_held(&obj->base.dev->struct_mutex);

        ret = i915_gem_object_wait(obj,
                                   I915_WAIT_INTERRUPTIBLE |
                                   I915_WAIT_LOCKED |
                                   (write ? I915_WAIT_ALL : 0),
                                   MAX_SCHEDULE_TIMEOUT,
                                   NULL);
        if (ret)
                return ret;

        if (obj->write_domain == I915_GEM_DOMAIN_GTT)
                return 0;

        /* Flush and acquire obj->pages so that we are coherent through
         * direct access in memory with previous cached writes through
         * shmemfs and that our cache domain tracking remains valid.
         * For example, if the obj->filp was moved to swap without us
         * being notified and releasing the pages, we would mistakenly
         * continue to assume that the obj remained out of the CPU cached
         * domain.
         */
        ret = i915_gem_object_pin_pages(obj);
        if (ret)
                return ret;

        flush_write_domain(obj, ~I915_GEM_DOMAIN_GTT);

        /* Serialise direct access to this object with the barriers for
         * coherent writes from the GPU, by effectively invalidating the
         * GTT domain upon first access.
         */
        if ((obj->read_domains & I915_GEM_DOMAIN_GTT) == 0)
                mb();

        /* It should now be out of any other write domains, and we can update
         * the domain values for our changes.
         */
        GEM_BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
        obj->read_domains |= I915_GEM_DOMAIN_GTT;
        if (write) {
                obj->read_domains = I915_GEM_DOMAIN_GTT;
                obj->write_domain = I915_GEM_DOMAIN_GTT;
                obj->mm.dirty = true;
        }

        i915_gem_object_unpin_pages(obj);
        return 0;
}

/**
 * Changes the cache-level of an object across all VMA.
 * @obj: object to act on
 * @cache_level: new cache level to set for the object
 *
 * After this function returns, the object will be in the new cache-level
 * across all GTT and the contents of the backing storage will be coherent,
 * with respect to the new cache-level. In order to keep the backing storage
 * coherent for all users, we only allow a single cache level to be set
 * globally on the object and prevent it from being changed whilst the
 * hardware is reading from the object. That is if the object is currently
 * on the scanout it will be set to uncached (or equivalent display
 * cache coherency) and all non-MOCS GPU access will also be uncached so
 * that all direct access to the scanout remains coherent.
 */
int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
                                    enum i915_cache_level cache_level)
{
        struct i915_vma *vma;
        int ret;

        lockdep_assert_held(&obj->base.dev->struct_mutex);

        if (obj->cache_level == cache_level)
                return 0;

        /* Inspect the list of currently bound VMA and unbind any that would
         * be invalid given the new cache-level. This is principally to
         * catch the issue of the CS prefetch crossing page boundaries and
         * reading an invalid PTE on older architectures.
         */
restart:
        list_for_each_entry(vma, &obj->vma_list, obj_link) {
                if (!drm_mm_node_allocated(&vma->node))
                        continue;

                if (i915_vma_is_pinned(vma)) {
                        DRM_DEBUG("can not change the cache level of pinned objects\n");
                        return -EBUSY;
                }

                if (!i915_vma_is_closed(vma) &&
                    i915_gem_valid_gtt_space(vma, cache_level))
                        continue;

                ret = i915_vma_unbind(vma);
                if (ret)
                        return ret;

                /* As unbinding may affect other elements in the
                 * obj->vma_list (due to side-effects from retiring
                 * an active vma), play safe and restart the iterator.
                 */
                goto restart;
        }

        /* We can reuse the existing drm_mm nodes but need to change the
         * cache-level on the PTE. We could simply unbind them all and
         * rebind with the correct cache-level on next use. However since
         * we already have a valid slot, dma mapping, pages etc, we may as
         * rewrite the PTE in the belief that doing so tramples upon less
         * state and so involves less work.
         */
        if (obj->bind_count) {
                /* Before we change the PTE, the GPU must not be accessing it.
                 * If we wait upon the object, we know that all the bound
                 * VMA are no longer active.
                 */
                ret = i915_gem_object_wait(obj,
                                           I915_WAIT_INTERRUPTIBLE |
                                           I915_WAIT_LOCKED |
                                           I915_WAIT_ALL,
                                           MAX_SCHEDULE_TIMEOUT,
                                           NULL);
                if (ret)
                        return ret;

                if (!HAS_LLC(to_i915(obj->base.dev)) &&
                    cache_level != I915_CACHE_NONE) {
                        /* Access to snoopable pages through the GTT is
                         * incoherent and on some machines causes a hard
                         * lockup. Relinquish the CPU mmaping to force
                         * userspace to refault in the pages and we can
                         * then double check if the GTT mapping is still
                         * valid for that pointer access.
                         */
                        i915_gem_release_mmap(obj);

                        /* As we no longer need a fence for GTT access,
                         * we can relinquish it now (and so prevent having
                         * to steal a fence from someone else on the next
                         * fence request). Note GPU activity would have
                         * dropped the fence as all snoopable access is
                         * supposed to be linear.
                         */
                        for_each_ggtt_vma(vma, obj) {
                                ret = i915_vma_put_fence(vma);
                                if (ret)
                                        return ret;
                        }
                } else {
                        /* We either have incoherent backing store and
                         * so no GTT access or the architecture is fully
                         * coherent. In such cases, existing GTT mmaps
                         * ignore the cache bit in the PTE and we can
                         * rewrite it without confusing the GPU or having
                         * to force userspace to fault back in its mmaps.
                         */
                }

                list_for_each_entry(vma, &obj->vma_list, obj_link) {
                        if (!drm_mm_node_allocated(&vma->node))
                                continue;

                        ret = i915_vma_bind(vma, cache_level, PIN_UPDATE);
                        if (ret)
                                return ret;
                }
        }

        list_for_each_entry(vma, &obj->vma_list, obj_link)
                vma->node.color = cache_level;
        i915_gem_object_set_cache_coherency(obj, cache_level);
        obj->cache_dirty = true; /* Always invalidate stale cachelines */

        return 0;
}

int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
                               struct drm_file *file)
{
        struct drm_i915_gem_caching *args = data;
        struct drm_i915_gem_object *obj;
        int err = 0;

        rcu_read_lock();
        obj = i915_gem_object_lookup_rcu(file, args->handle);
        if (!obj) {
                err = -ENOENT;