#include "drmP.h"
#include "drm.h"
+#include "drm_compat.h"
#include "i915_drm.h"
#include "i915_drv.h"
-
-#define WATCH_COHERENCY 0
-#define WATCH_BUF 0
-#define WATCH_EXEC 0
-#define WATCH_LRU 0
-#define WATCH_RELOC 0
+#include <linux/swap.h>
static int
i915_gem_object_set_domain(struct drm_gem_object *obj,
uint32_t read_domains,
uint32_t write_domain);
+static int
+i915_gem_object_set_domain_range(struct drm_gem_object *obj,
+ uint64_t offset,
+ uint64_t size,
+ uint32_t read_domains,
+ uint32_t write_domain);
+int
+i915_gem_set_domain(struct drm_gem_object *obj,
+ struct drm_file *file_priv,
+ uint32_t read_domains,
+ uint32_t write_domain);
+static int i915_gem_object_get_page_list(struct drm_gem_object *obj);
+static void i915_gem_object_free_page_list(struct drm_gem_object *obj);
+static int i915_gem_object_wait_rendering(struct drm_gem_object *obj);
int
i915_gem_init_ioctl(struct drm_device *dev, void *data,
return -EINVAL;
}
- drm_memrange_init(&dev_priv->mm.gtt_space, args->gtt_start,
+ drm_mm_init(&dev_priv->mm.gtt_space, args->gtt_start,
args->gtt_end - args->gtt_start);
+ dev->gtt_total = (uint32_t) (args->gtt_end - args->gtt_start);
+
+ mutex_unlock(&dev->struct_mutex);
+
+ return 0;
+}
+
+
+/**
+ * Creates a new mm object and returns a handle to it.
+ */
+int
+i915_gem_create_ioctl(struct drm_device *dev, void *data,
+ struct drm_file *file_priv)
+{
+ struct drm_i915_gem_create *args = data;
+ struct drm_gem_object *obj;
+ int handle, ret;
+
+ args->size = roundup(args->size, PAGE_SIZE);
+
+ /* Allocate the new object */
+ obj = drm_gem_object_alloc(dev, args->size);
+ if (obj == NULL)
+ return -ENOMEM;
+
+ ret = drm_gem_handle_create(file_priv, obj, &handle);
+ mutex_lock(&dev->struct_mutex);
+ drm_gem_object_handle_unreference(obj);
+ mutex_unlock(&dev->struct_mutex);
+
+ if (ret)
+ return ret;
+
+ args->handle = handle;
+
+ return 0;
+}
+
+/**
+ * Reads data from the object referenced by handle.
+ *
+ * On error, the contents of *data are undefined.
+ */
+int
+i915_gem_pread_ioctl(struct drm_device *dev, void *data,
+ struct drm_file *file_priv)
+{
+ struct drm_i915_gem_pread *args = data;
+ struct drm_gem_object *obj;
+ struct drm_i915_gem_object *obj_priv;
+ ssize_t read;
+ loff_t offset;
+ int ret;
+
+ obj = drm_gem_object_lookup(dev, file_priv, args->handle);
+ if (obj == NULL)
+ return -EBADF;
+ obj_priv = obj->driver_private;
+
+ /* Bounds check source.
+ *
+ * XXX: This could use review for overflow issues...
+ */
+ if (args->offset > obj->size || args->size > obj->size ||
+ args->offset + args->size > obj->size) {
+ drm_gem_object_unreference(obj);
+ return -EINVAL;
+ }
+
+ mutex_lock(&dev->struct_mutex);
+
+ ret = i915_gem_object_set_domain_range(obj, args->offset, args->size,
+ I915_GEM_DOMAIN_CPU, 0);
+ if (ret != 0) {
+ drm_gem_object_unreference(obj);
+ mutex_unlock(&dev->struct_mutex);
+ }
+
+ offset = args->offset;
+
+ read = vfs_read(obj->filp, (char __user *)(uintptr_t)args->data_ptr,
+ args->size, &offset);
+ if (read != args->size) {
+ drm_gem_object_unreference(obj);
+ mutex_unlock(&dev->struct_mutex);
+ if (read < 0)
+ return read;
+ else
+ return -EINVAL;
+ }
+
+ drm_gem_object_unreference(obj);
+ mutex_unlock(&dev->struct_mutex);
+
+ return 0;
+}
+
+#include "drm_compat.h"
+
+static int
+i915_gem_gtt_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
+ struct drm_i915_gem_pwrite *args,
+ struct drm_file *file_priv)
+{
+ struct drm_i915_gem_object *obj_priv = obj->driver_private;
+ ssize_t remain;
+ loff_t offset;
+ char __user *user_data;
+ char *vaddr;
+ int i, o, l;
+ int ret = 0;
+ unsigned long pfn;
+ unsigned long unwritten;
+
+ user_data = (char __user *) (uintptr_t) args->data_ptr;
+ remain = args->size;
+ if (!access_ok(VERIFY_READ, user_data, remain))
+ return -EFAULT;
+
+
+ mutex_lock(&dev->struct_mutex);
+ ret = i915_gem_object_pin(obj, 0);
+ if (ret) {
+ mutex_unlock(&dev->struct_mutex);
+ return ret;
+ }
+ ret = i915_gem_set_domain(obj, file_priv,
+ I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
+ if (ret)
+ goto fail;
+
+ obj_priv = obj->driver_private;
+ offset = obj_priv->gtt_offset + args->offset;
+ obj_priv->dirty = 1;
+
+ while (remain > 0) {
+ /* Operation in this page
+ *
+ * i = page number
+ * o = offset within page
+ * l = bytes to copy
+ */
+ i = offset >> PAGE_SHIFT;
+ o = offset & (PAGE_SIZE-1);
+ l = remain;
+ if ((o + l) > PAGE_SIZE)
+ l = PAGE_SIZE - o;
+
+ pfn = (dev->agp->base >> PAGE_SHIFT) + i;
+
+#ifdef DRM_KMAP_ATOMIC_PROT_PFN
+ /* kmap_atomic can't map IO pages on non-HIGHMEM kernels
+ */
+ vaddr = kmap_atomic_prot_pfn(pfn, KM_USER0,
+ __pgprot(__PAGE_KERNEL));
+#if WATCH_PWRITE
+ DRM_INFO("pwrite i %d o %d l %d pfn %ld vaddr %p\n",
+ i, o, l, pfn, vaddr);
+#endif
+ unwritten = __copy_from_user_inatomic_nocache(vaddr + o,
+ user_data, l);
+ kunmap_atomic(vaddr, KM_USER0);
+
+ if (unwritten)
+#endif
+ {
+ vaddr = ioremap(pfn << PAGE_SHIFT, PAGE_SIZE);
+#if WATCH_PWRITE
+ DRM_INFO("pwrite slow i %d o %d l %d "
+ "pfn %ld vaddr %p\n",
+ i, o, l, pfn, vaddr);
+#endif
+ if (vaddr == NULL) {
+ ret = -EFAULT;
+ goto fail;
+ }
+ unwritten = __copy_from_user(vaddr + o, user_data, l);
+#if WATCH_PWRITE
+ DRM_INFO("unwritten %ld\n", unwritten);
+#endif
+ iounmap(vaddr);
+ if (unwritten) {
+ ret = -EFAULT;
+ goto fail;
+ }
+ }
+
+ remain -= l;
+ user_data += l;
+ offset += l;
+ }
+#if WATCH_PWRITE && 1
+ i915_gem_clflush_object(obj);
+ i915_gem_dump_object(obj, args->offset + args->size, __func__, ~0);
+ i915_gem_clflush_object(obj);
+#endif
+
+fail:
+ i915_gem_object_unpin(obj);
+ mutex_unlock(&dev->struct_mutex);
+
+ return ret;
+}
+
+int
+i915_gem_shmem_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
+ struct drm_i915_gem_pwrite *args,
+ struct drm_file *file_priv)
+{
+ int ret;
+ loff_t offset;
+ ssize_t written;
+
+ mutex_lock(&dev->struct_mutex);
+
+ ret = i915_gem_set_domain(obj, file_priv,
+ I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
+ if (ret) {
+ mutex_unlock(&dev->struct_mutex);
+ return ret;
+ }
+
+ offset = args->offset;
+
+ written = vfs_write(obj->filp,
+ (char __user *)(uintptr_t) args->data_ptr,
+ args->size, &offset);
+ if (written != args->size) {
+ mutex_unlock(&dev->struct_mutex);
+ if (written < 0)
+ return written;
+ else
+ return -EINVAL;
+ }
+
+ mutex_unlock(&dev->struct_mutex);
+
+ return 0;
+}
+
+/**
+ * Writes data to the object referenced by handle.
+ *
+ * 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_priv)
+{
+ struct drm_i915_gem_pwrite *args = data;
+ struct drm_gem_object *obj;
+ struct drm_i915_gem_object *obj_priv;
+ int ret = 0;
+
+ obj = drm_gem_object_lookup(dev, file_priv, args->handle);
+ if (obj == NULL)
+ return -EBADF;
+ obj_priv = obj->driver_private;
+
+ /* Bounds check destination.
+ *
+ * XXX: This could use review for overflow issues...
+ */
+ if (args->offset > obj->size || args->size > obj->size ||
+ args->offset + args->size > obj->size) {
+ drm_gem_object_unreference(obj);
+ return -EINVAL;
+ }
+
+ /* 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 (obj_priv->tiling_mode == I915_TILING_NONE &&
+ dev->gtt_total != 0)
+ ret = i915_gem_gtt_pwrite(dev, obj, args, file_priv);
+ else
+ ret = i915_gem_shmem_pwrite(dev, obj, args, file_priv);
+
+#if WATCH_PWRITE
+ if (ret)
+ DRM_INFO("pwrite failed %d\n", ret);
+#endif
+
+ drm_gem_object_unreference(obj);
+
+ return ret;
+}
+
+/**
+ * Called when user space prepares to use an object
+ */
+int
+i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
+ struct drm_file *file_priv)
+{
+ struct drm_i915_gem_set_domain *args = data;
+ struct drm_gem_object *obj;
+ int ret;
+
+ if (!(dev->driver->driver_features & DRIVER_GEM))
+ return -ENODEV;
+
+ obj = drm_gem_object_lookup(dev, file_priv, args->handle);
+ if (obj == NULL)
+ return -EBADF;
+
+ mutex_lock(&dev->struct_mutex);
+#if WATCH_BUF
+ DRM_INFO("set_domain_ioctl %p(%d), %08x %08x\n",
+ obj, obj->size, args->read_domains, args->write_domain);
+#endif
+ ret = i915_gem_set_domain(obj, file_priv,
+ args->read_domains, args->write_domain);
+ drm_gem_object_unreference(obj);
+ mutex_unlock(&dev->struct_mutex);
+ return ret;
+}
+
+/**
+ * Called when user space has done writes to this buffer
+ */
+int
+i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
+ struct drm_file *file_priv)
+{
+ struct drm_i915_gem_sw_finish *args = data;
+ struct drm_gem_object *obj;
+ struct drm_i915_gem_object *obj_priv;
+ int ret = 0;
+
+ if (!(dev->driver->driver_features & DRIVER_GEM))
+ return -ENODEV;
+
+ mutex_lock(&dev->struct_mutex);
+ obj = drm_gem_object_lookup(dev, file_priv, args->handle);
+ if (obj == NULL) {
+ mutex_unlock(&dev->struct_mutex);
+ return -EBADF;
+ }
+
+#if WATCH_BUF
+ DRM_INFO("%s: sw_finish %d (%p %d)\n",
+ __func__, args->handle, obj, obj->size);
+#endif
+ obj_priv = obj->driver_private;
+
+ /* Pinned buffers may be scanout, so flush the cache */
+ if ((obj->write_domain & I915_GEM_DOMAIN_CPU) && obj_priv->pin_count) {
+ i915_gem_clflush_object(obj);
+ drm_agp_chipset_flush(dev);
+ }
+ drm_gem_object_unreference(obj);
+ mutex_unlock(&dev->struct_mutex);
+ return ret;
+}
+
+/**
+ * Maps the contents of an object, returning the address it is mapped
+ * into.
+ *
+ * While the mapping holds a reference on the contents of the object, it doesn't
+ * imply a ref on the object itself.
+ */
+int
+i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
+ struct drm_file *file_priv)
+{
+ struct drm_i915_gem_mmap *args = data;
+ struct drm_gem_object *obj;
+ loff_t offset;
+ unsigned long addr;
+
+ if (!(dev->driver->driver_features & DRIVER_GEM))
+ return -ENODEV;
+
+ obj = drm_gem_object_lookup(dev, file_priv, args->handle);
+ if (obj == NULL)
+ return -EBADF;
+
+ offset = args->offset;
+
+ down_write(¤t->mm->mmap_sem);
+ addr = do_mmap(obj->filp, 0, args->size,
+ PROT_READ | PROT_WRITE, MAP_SHARED,
+ args->offset);
+ up_write(¤t->mm->mmap_sem);
+ mutex_lock(&dev->struct_mutex);
+ drm_gem_object_unreference(obj);
mutex_unlock(&dev->struct_mutex);
+ if (IS_ERR((void *)addr))
+ return addr;
+
+ args->addr_ptr = (uint64_t) addr;
return 0;
}
for (i = 0; i < page_count; i++)
- if (obj_priv->page_list[i] != NULL)
+ if (obj_priv->page_list[i] != NULL) {
+ if (obj_priv->dirty)
+ set_page_dirty(obj_priv->page_list[i]);
+ mark_page_accessed(obj_priv->page_list[i]);
page_cache_release(obj_priv->page_list[i]);
+ }
+ obj_priv->dirty = 0;
drm_free(obj_priv->page_list,
page_count * sizeof(struct page *),
&dev_priv->mm.active_list);
}
+
static void
i915_gem_object_move_to_inactive(struct drm_gem_object *obj)
{
drm_i915_private_t *dev_priv = dev->dev_private;
struct drm_i915_gem_object *obj_priv = obj->driver_private;
+ i915_verify_inactive(dev, __FILE__, __LINE__);
if (obj_priv->pin_count != 0)
list_del_init(&obj_priv->list);
else
obj_priv->active = 0;
drm_gem_object_unreference(obj);
}
+ i915_verify_inactive(dev, __FILE__, __LINE__);
}
/**
drm_i915_private_t *dev_priv = dev->dev_private;
struct drm_i915_gem_request *request;
uint32_t seqno;
+ int was_empty;
RING_LOCALS;
request = drm_calloc(1, sizeof(*request), DRM_MEM_DRIVER);
dev_priv->mm.next_gem_seqno++;
BEGIN_LP_RING(4);
- OUT_RING(CMD_STORE_DWORD_IDX);
- OUT_RING(I915_GEM_HWS_INDEX << STORE_DWORD_INDEX_SHIFT);
+ OUT_RING(MI_STORE_DWORD_INDEX);
+ OUT_RING(I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
OUT_RING(seqno);
- OUT_RING(GFX_OP_USER_INTERRUPT);
+ OUT_RING(MI_USER_INTERRUPT);
ADVANCE_LP_RING();
DRM_DEBUG("%d\n", seqno);
request->seqno = seqno;
request->emitted_jiffies = jiffies;
request->flush_domains = flush_domains;
+ was_empty = list_empty(&dev_priv->mm.request_list);
list_add_tail(&request->list, &dev_priv->mm.request_list);
+ if (was_empty)
+ schedule_delayed_work(&dev_priv->mm.retire_work, HZ);
return seqno;
}
* Ensures that all commands in the ring are finished
* before signalling the CPU
*/
-
uint32_t
i915_retire_commands(struct drm_device *dev)
{
drm_i915_private_t *dev_priv = dev->dev_private;
- uint32_t cmd = CMD_MI_FLUSH | MI_NO_WRITE_FLUSH;
+ uint32_t cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
uint32_t flush_domains = 0;
RING_LOCALS;
/* The sampler always gets flushed on i965 (sigh) */
if (IS_I965G(dev))
- flush_domains |= DRM_GEM_DOMAIN_I915_SAMPLER;
+ flush_domains |= I915_GEM_DOMAIN_SAMPLER;
BEGIN_LP_RING(2);
OUT_RING(cmd);
OUT_RING(0); /* noop */
__func__, request->seqno, obj);
#endif
+ /* If this request flushes the write domain,
+ * clear the write domain from the object now
+ */
+ if (request->flush_domains & obj->write_domain)
+ obj->write_domain = 0;
+
if (obj->write_domain != 0) {
list_move_tail(&obj_priv->list,
&dev_priv->mm.flushing_list);
return (int32_t)(seq1 - seq2) >= 0;
}
-static uint32_t
+uint32_t
i915_get_gem_seqno(struct drm_device *dev)
{
drm_i915_private_t *dev_priv = dev->dev_private;
list);
retiring_seqno = request->seqno;
- if (i915_seqno_passed(seqno, retiring_seqno)) {
+ if (i915_seqno_passed(seqno, retiring_seqno) ||
+ dev_priv->mm.wedged) {
i915_gem_retire_request(dev, request);
list_del(&request->list);
drm_free(request, sizeof(*request), DRM_MEM_DRIVER);
} else
- break;
+ break;
}
}
+void
+i915_gem_retire_work_handler(struct work_struct *work)
+{
+ drm_i915_private_t *dev_priv;
+ struct drm_device *dev;
+
+ dev_priv = container_of(work, drm_i915_private_t,
+ mm.retire_work.work);
+ dev = dev_priv->dev;
+
+ mutex_lock(&dev->struct_mutex);
+ i915_gem_retire_requests(dev);
+ if (!list_empty(&dev_priv->mm.request_list))
+ schedule_delayed_work(&dev_priv->mm.retire_work, HZ);
+ mutex_unlock(&dev->struct_mutex);
+}
+
/**
* Waits for a sequence number to be signaled, and cleans up the
* request and object lists appropriately for that event.
BUG_ON(seqno == 0);
if (!i915_seqno_passed(i915_get_gem_seqno(dev), seqno)) {
+ dev_priv->mm.waiting_gem_seqno = seqno;
i915_user_irq_on(dev_priv);
ret = wait_event_interruptible(dev_priv->irq_queue,
i915_seqno_passed(i915_get_gem_seqno(dev),
- seqno));
+ seqno) ||
+ dev_priv->mm.wedged);
i915_user_irq_off(dev_priv);
+ dev_priv->mm.waiting_gem_seqno = 0;
}
+ if (dev_priv->mm.wedged)
+ ret = -EIO;
+
+ if (ret && ret != -ERESTARTSYS)
+ DRM_ERROR("%s returns %d (awaiting %d at %d)\n",
+ __func__, ret, seqno, i915_get_gem_seqno(dev));
/* Directly dispatch request retiring. While we have the work queue
* to handle this, the waiter on a request often wants an associated
invalidate_domains, flush_domains);
#endif
- if (flush_domains & DRM_GEM_DOMAIN_CPU)
+ if (flush_domains & I915_GEM_DOMAIN_CPU)
drm_agp_chipset_flush(dev);
- if ((invalidate_domains|flush_domains) & ~DRM_GEM_DOMAIN_CPU) {
+ if ((invalidate_domains | flush_domains) & ~(I915_GEM_DOMAIN_CPU |
+ I915_GEM_DOMAIN_GTT)) {
/*
* read/write caches:
*
- * DRM_GEM_DOMAIN_I915_RENDER is always invalidated, but is
+ * I915_GEM_DOMAIN_RENDER is always invalidated, but is
* only flushed if MI_NO_WRITE_FLUSH is unset. On 965, it is
* also flushed at 2d versus 3d pipeline switches.
*
* read-only caches:
*
- * DRM_GEM_DOMAIN_I915_SAMPLER is flushed on pre-965 if
+ * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
* MI_READ_FLUSH is set, and is always flushed on 965.
*
- * DRM_GEM_DOMAIN_I915_COMMAND may not exist?
+ * I915_GEM_DOMAIN_COMMAND may not exist?
*
- * DRM_GEM_DOMAIN_I915_INSTRUCTION, which exists on 965, is
+ * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
* invalidated when MI_EXE_FLUSH is set.
*
- * DRM_GEM_DOMAIN_I915_VERTEX, which exists on 965, is
+ * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
* invalidated with every MI_FLUSH.
*
* TLBs:
*
- * On 965, TLBs associated with DRM_GEM_DOMAIN_I915_COMMAND
- * and DRM_GEM_DOMAIN_CPU in are invalidated at PTE write and
- * DRM_GEM_DOMAIN_I915_RENDER and DRM_GEM_DOMAIN_I915_SAMPLER
+ * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
+ * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
+ * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
* are flushed at any MI_FLUSH.
*/
- cmd = CMD_MI_FLUSH | MI_NO_WRITE_FLUSH;
+ cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
if ((invalidate_domains|flush_domains) &
- DRM_GEM_DOMAIN_I915_RENDER)
+ I915_GEM_DOMAIN_RENDER)
cmd &= ~MI_NO_WRITE_FLUSH;
if (!IS_I965G(dev)) {
/*
* On the 965, the sampler cache always gets flushed
* and this bit is reserved.
*/
- if (invalidate_domains & DRM_GEM_DOMAIN_I915_SAMPLER)
+ if (invalidate_domains & I915_GEM_DOMAIN_SAMPLER)
cmd |= MI_READ_FLUSH;
}
- if (invalidate_domains & DRM_GEM_DOMAIN_I915_INSTRUCTION)
+ if (invalidate_domains & I915_GEM_DOMAIN_INSTRUCTION)
cmd |= MI_EXE_FLUSH;
#if WATCH_EXEC
struct drm_device *dev = obj->dev;
struct drm_i915_gem_object *obj_priv = obj->driver_private;
int ret;
+ uint32_t write_domain;
/* If there are writes queued to the buffer, flush and
* create a new seqno to wait for.
*/
- if (obj->write_domain & ~(DRM_GEM_DOMAIN_CPU)) {
- uint32_t write_domain = obj->write_domain;
+ write_domain = obj->write_domain & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT);
+ if (write_domain) {
#if WATCH_BUF
DRM_INFO("%s: flushing object %p from write domain %08x\n",
__func__, obj, write_domain);
#endif
i915_gem_flush(dev, 0, write_domain);
- obj->write_domain = 0;
i915_gem_object_move_to_active(obj);
obj_priv->last_rendering_seqno = i915_add_request(dev,
DRM_INFO("%s: flush moves to exec list %p\n", __func__, obj);
#endif
}
+
/* If there is rendering queued on the buffer being evicted, wait for
* it.
*/
static int
i915_gem_object_unbind(struct drm_gem_object *obj)
{
+ struct drm_device *dev = obj->dev;
struct drm_i915_gem_object *obj_priv = obj->driver_private;
int ret = 0;
if (obj_priv->gtt_space == NULL)
return 0;
+ if (obj_priv->pin_count != 0) {
+ DRM_ERROR("Attempting to unbind pinned buffer\n");
+ return -EINVAL;
+ }
+
+ /* Wait for any rendering to complete
+ */
+ ret = i915_gem_object_wait_rendering(obj);
+ if (ret) {
+ DRM_ERROR("wait_rendering failed: %d\n", ret);
+ return ret;
+ }
+
/* Move the object to the CPU domain to ensure that
* any possible CPU writes while it's not in the GTT
* are flushed when we go to remap it. This will
* also ensure that all pending GPU writes are finished
* before we unbind.
*/
- ret = i915_gem_object_set_domain (obj, DRM_GEM_DOMAIN_CPU,
- DRM_GEM_DOMAIN_CPU);
- if (ret)
+ ret = i915_gem_object_set_domain(obj, I915_GEM_DOMAIN_CPU,
+ I915_GEM_DOMAIN_CPU);
+ if (ret) {
+ DRM_ERROR("set_domain failed: %d\n", ret);
return ret;
+ }
if (obj_priv->agp_mem != NULL) {
drm_unbind_agp(obj_priv->agp_mem);
obj_priv->agp_mem = NULL;
}
+ BUG_ON(obj_priv->active);
+
i915_gem_object_free_page_list(obj);
- drm_memrange_put_block(obj_priv->gtt_space);
- obj_priv->gtt_space = NULL;
+ if (obj_priv->gtt_space) {
+ atomic_dec(&dev->gtt_count);
+ atomic_sub(obj->size, &dev->gtt_memory);
+
+ drm_mm_put_block(obj_priv->gtt_space);
+ obj_priv->gtt_space = NULL;
+ }
/* Remove ourselves from the LRU list if present. */
- if (!list_empty(&obj_priv->list)) {
+ if (!list_empty(&obj_priv->list))
list_del_init(&obj_priv->list);
- if (obj_priv->active) {
- DRM_ERROR("Failed to wait on buffer when unbinding, "
- "continued anyway.\n");
- obj_priv->active = 0;
- drm_gem_object_unreference(obj);
- }
- }
- return 0;
-}
-
-#if WATCH_BUF | WATCH_EXEC
-static void
-i915_gem_dump_page(struct page *page, uint32_t start, uint32_t end,
- uint32_t bias, uint32_t mark)
-{
- uint32_t *mem = kmap_atomic(page, KM_USER0);
- int i;
- for (i = start; i < end; i += 4)
- DRM_INFO("%08x: %08x%s\n",
- (int) (bias + i), mem[i / 4],
- (bias + i == mark) ? " ********" : "");
- kunmap_atomic(mem, KM_USER0);
- /* give syslog time to catch up */
- msleep(1);
-}
-
-static void
-i915_gem_dump_object(struct drm_gem_object *obj, int len,
- const char *where, uint32_t mark)
-{
- struct drm_i915_gem_object *obj_priv = obj->driver_private;
- int page;
-
- DRM_INFO("%s: object at offset %08x\n", where, obj_priv->gtt_offset);
- for (page = 0; page < (len + PAGE_SIZE-1) / PAGE_SIZE; page++) {
- int page_len, chunk, chunk_len;
-
- page_len = len - page * PAGE_SIZE;
- if (page_len > PAGE_SIZE)
- page_len = PAGE_SIZE;
-
- for (chunk = 0; chunk < page_len; chunk += 128) {
- chunk_len = page_len - chunk;
- if (chunk_len > 128)
- chunk_len = 128;
- i915_gem_dump_page(obj_priv->page_list[page],
- chunk, chunk + chunk_len,
- obj_priv->gtt_offset +
- page * PAGE_SIZE,
- mark);
- }
- }
-}
-#endif
-#if WATCH_LRU
-static void
-i915_dump_lru(struct drm_device *dev, const char *where)
-{
- drm_i915_private_t *dev_priv = dev->dev_private;
- struct drm_i915_gem_object *obj_priv;
-
- DRM_INFO("active list %s {\n", where);
- list_for_each_entry(obj_priv, &dev_priv->mm.active_list,
- list)
- {
- DRM_INFO(" %p: %08x\n", obj_priv,
- obj_priv->last_rendering_seqno);
- }
- DRM_INFO("}\n");
- DRM_INFO("flushing list %s {\n", where);
- list_for_each_entry(obj_priv, &dev_priv->mm.flushing_list,
- list)
- {
- DRM_INFO(" %p: %08x\n", obj_priv,
- obj_priv->last_rendering_seqno);
- }
- DRM_INFO("}\n");
- DRM_INFO("inactive %s {\n", where);
- list_for_each_entry(obj_priv, &dev_priv->mm.inactive_list, list) {
- DRM_INFO(" %p: %08x\n", obj_priv,
- obj_priv->last_rendering_seqno);
- }
- DRM_INFO("}\n");
+ return 0;
}
-#endif
static int
i915_gem_evict_something(struct drm_device *dev)
drm_i915_private_t *dev_priv = dev->dev_private;
struct drm_gem_object *obj;
struct drm_i915_gem_object *obj_priv;
- int ret;
+ int ret = 0;
for (;;) {
/* If there's an inactive buffer available now, grab it
list);
obj = obj_priv->obj;
BUG_ON(obj_priv->pin_count != 0);
+#if WATCH_LRU
+ DRM_INFO("%s: evicting %p\n", __func__, obj);
+#endif
+ BUG_ON(obj_priv->active);
+
+ /* Wait on the rendering and unbind the buffer. */
+ ret = i915_gem_object_unbind(obj);
break;
}
*/
if (!list_empty(&dev_priv->mm.request_list)) {
struct drm_i915_gem_request *request;
- int ret;
request = list_first_entry(&dev_priv->mm.request_list,
struct drm_i915_gem_request,
list);
ret = i915_wait_request(dev, request->seqno);
- if (ret != 0)
- return ret;
+ if (ret)
+ break;
- continue;
+ /* if waiting caused an object to become inactive,
+ * then loop around and wait for it. Otherwise, we
+ * assume that waiting freed and unbound something,
+ * so there should now be some space in the GTT
+ */
+ if (!list_empty(&dev_priv->mm.inactive_list))
+ continue;
+ break;
}
/* If we didn't have anything on the request list but there
continue;
}
+ DRM_ERROR("inactive empty %d request empty %d "
+ "flushing empty %d\n",
+ list_empty(&dev_priv->mm.inactive_list),
+ list_empty(&dev_priv->mm.request_list),
+ list_empty(&dev_priv->mm.flushing_list));
/* If we didn't do any of the above, there's nothing to be done
* and we just can't fit it in.
*/
return -ENOMEM;
}
-
-#if WATCH_LRU
- DRM_INFO("%s: evicting %p\n", __func__, obj);
-#endif
-
- BUG_ON(obj_priv->active);
-
- /* Wait on the rendering and unbind the buffer. */
- ret = i915_gem_object_unbind(obj);
-
return ret;
}
{
struct drm_i915_gem_object *obj_priv = obj->driver_private;
int page_count, i;
+ struct address_space *mapping;
+ struct inode *inode;
+ struct page *page;
+ int ret;
+
if (obj_priv->page_list)
return 0;
BUG_ON(obj_priv->page_list != NULL);
obj_priv->page_list = drm_calloc(page_count, sizeof(struct page *),
DRM_MEM_DRIVER);
- if (obj_priv->page_list == NULL)
+ if (obj_priv->page_list == NULL) {
+ DRM_ERROR("Faled to allocate page list\n");
return -ENOMEM;
+ }
+ inode = obj->filp->f_path.dentry->d_inode;
+ mapping = inode->i_mapping;
for (i = 0; i < page_count; i++) {
- obj_priv->page_list[i] =
- find_or_create_page(obj->filp->f_mapping, i, GFP_HIGHUSER);
-
- if (obj_priv->page_list[i] == NULL) {
+ page = read_mapping_page(mapping, i, NULL);
+ if (IS_ERR(page)) {
+ ret = PTR_ERR(page);
+ DRM_ERROR("read_mapping_page failed: %d\n", ret);
i915_gem_object_free_page_list(obj);
- return -ENOMEM;
+ return ret;
}
- unlock_page(obj_priv->page_list[i]);
+ obj_priv->page_list[i] = page;
}
return 0;
}
struct drm_device *dev = obj->dev;
drm_i915_private_t *dev_priv = dev->dev_private;
struct drm_i915_gem_object *obj_priv = obj->driver_private;
- struct drm_memrange_node *free_space;
+ struct drm_mm_node *free_space;
int page_count, ret;
if (alignment == 0)
}
search_free:
- free_space = drm_memrange_search_free(&dev_priv->mm.gtt_space,
- obj->size,
- alignment, 0);
+ free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
+ obj->size, alignment, 0);
if (free_space != NULL) {
- obj_priv->gtt_space =
- drm_memrange_get_block(free_space, obj->size,
- alignment);
+ obj_priv->gtt_space = drm_mm_get_block(free_space, obj->size,
+ alignment);
if (obj_priv->gtt_space != NULL) {
obj_priv->gtt_space->private = obj;
obj_priv->gtt_offset = obj_priv->gtt_space->start;
DRM_INFO("%s: GTT full, evicting something\n", __func__);
#endif
if (list_empty(&dev_priv->mm.inactive_list) &&
+ list_empty(&dev_priv->mm.flushing_list) &&
list_empty(&dev_priv->mm.active_list)) {
DRM_ERROR("GTT full, but LRU list empty\n");
return -ENOMEM;
}
ret = i915_gem_evict_something(dev);
- if (ret != 0)
+ if (ret != 0) {
+ DRM_ERROR("Failed to evict a buffer %d\n", ret);
return ret;
+ }
goto search_free;
}
#endif
ret = i915_gem_object_get_page_list(obj);
if (ret) {
- drm_memrange_put_block(obj_priv->gtt_space);
+ drm_mm_put_block(obj_priv->gtt_space);
obj_priv->gtt_space = NULL;
return ret;
}
obj_priv->gtt_offset);
if (obj_priv->agp_mem == NULL) {
i915_gem_object_free_page_list(obj);
- drm_memrange_put_block(obj_priv->gtt_space);
+ drm_mm_put_block(obj_priv->gtt_space);
obj_priv->gtt_space = NULL;
return -ENOMEM;
}
+ atomic_inc(&dev->gtt_count);
+ atomic_add(obj->size, &dev->gtt_memory);
/* 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
*/
- BUG_ON(obj->read_domains & ~DRM_GEM_DOMAIN_CPU);
- BUG_ON(obj->write_domain & ~DRM_GEM_DOMAIN_CPU);
+ BUG_ON(obj->read_domains & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT));
+ BUG_ON(obj->write_domain & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT));
return 0;
}
-static void
+void
i915_gem_clflush_object(struct drm_gem_object *obj)
{
struct drm_i915_gem_object *obj_priv = obj->driver_private;
uint32_t write_domain)
{
struct drm_device *dev = obj->dev;
+ struct drm_i915_gem_object *obj_priv = obj->driver_private;
uint32_t invalidate_domains = 0;
uint32_t flush_domains = 0;
int ret;
#if WATCH_BUF
- DRM_INFO("%s: object %p read %08x write %08x\n",
- __func__, obj, read_domains, write_domain);
+ DRM_INFO("%s: object %p read %08x -> %08x write %08x -> %08x\n",
+ __func__, obj,
+ obj->read_domains, read_domains,
+ obj->write_domain, write_domain);
#endif
/*
* If the object isn't moving to a new write domain,
*/
if (write_domain == 0)
read_domains |= obj->read_domains;
+ else
+ obj_priv->dirty = 1;
/*
* Flush the current write domain if
* stale data. That is, any new read domains.
*/
invalidate_domains |= read_domains & ~obj->read_domains;
- if ((flush_domains | invalidate_domains) & DRM_GEM_DOMAIN_CPU) {
+ if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU) {
#if WATCH_BUF
DRM_INFO("%s: CPU domain flush %08x invalidate %08x\n",
__func__, flush_domains, invalidate_domains);
* then pause for rendering so that the GPU caches will be
* flushed before the cpu cache is invalidated
*/
- if ((invalidate_domains & DRM_GEM_DOMAIN_CPU) &&
- (flush_domains & ~DRM_GEM_DOMAIN_CPU)) {
+ if ((invalidate_domains & I915_GEM_DOMAIN_CPU) &&
+ (flush_domains & ~(I915_GEM_DOMAIN_CPU |
+ I915_GEM_DOMAIN_GTT))) {
ret = i915_gem_object_wait_rendering(obj);
if (ret)
return ret;
if ((write_domain | flush_domains) != 0)
obj->write_domain = write_domain;
+
+ /* If we're invalidating the CPU domain, clear the per-page CPU
+ * domain list as well.
+ */
+ if (obj_priv->page_cpu_valid != NULL &&
+ (obj->read_domains & I915_GEM_DOMAIN_CPU) &&
+ ((read_domains & I915_GEM_DOMAIN_CPU) == 0)) {
+ memset(obj_priv->page_cpu_valid, 0, obj->size / PAGE_SIZE);
+ }
obj->read_domains = read_domains;
+
dev->invalidate_domains |= invalidate_domains;
dev->flush_domains |= flush_domains;
+#if WATCH_BUF
+ DRM_INFO("%s: read %08x write %08x invalidate %08x flush %08x\n",
+ __func__,
+ obj->read_domains, obj->write_domain,
+ dev->invalidate_domains, dev->flush_domains);
+#endif
+ return 0;
+}
+
+/**
+ * Set the read/write domain on a range of the object.
+ *
+ * Currently only implemented for CPU reads, otherwise drops to normal
+ * i915_gem_object_set_domain().
+ */
+static int
+i915_gem_object_set_domain_range(struct drm_gem_object *obj,
+ uint64_t offset,
+ uint64_t size,
+ uint32_t read_domains,
+ uint32_t write_domain)
+{
+ struct drm_i915_gem_object *obj_priv = obj->driver_private;
+ int ret, i;
+
+ if (obj->read_domains & I915_GEM_DOMAIN_CPU)
+ return 0;
+
+ if (read_domains != I915_GEM_DOMAIN_CPU ||
+ write_domain != 0)
+ return i915_gem_object_set_domain(obj,
+ read_domains, write_domain);
+
+ /* Wait on any GPU rendering to the object to be flushed. */
+ if (obj->write_domain & ~(I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT)) {
+ ret = i915_gem_object_wait_rendering(obj);
+ if (ret)
+ return ret;
+ }
+
+ if (obj_priv->page_cpu_valid == NULL) {
+ obj_priv->page_cpu_valid = drm_calloc(1, obj->size / PAGE_SIZE,
+ DRM_MEM_DRIVER);
+ }
+
+ /* Flush the cache on any pages that are still invalid from the CPU's
+ * perspective.
+ */
+ for (i = offset / PAGE_SIZE; i < (offset + size - 1) / PAGE_SIZE; i++) {
+ if (obj_priv->page_cpu_valid[i])
+ continue;
+
+ drm_ttm_cache_flush(obj_priv->page_list + i, 1);
+
+ obj_priv->page_cpu_valid[i] = 1;
+ }
+
return 0;
}
return flush_domains;
}
-#if WATCH_COHERENCY
-static void
-i915_gem_object_check_coherency(struct drm_gem_object *obj, int handle)
-{
- struct drm_device *dev = obj->dev;
- struct drm_i915_gem_object *obj_priv = obj->driver_private;
- int page;
- uint32_t *gtt_mapping;
- uint32_t *backing_map = NULL;
- int bad_count = 0;
-
- DRM_INFO("%s: checking coherency of object %p@0x%08x (%d, %dkb):\n",
- __FUNCTION__, obj, obj_priv->gtt_offset, handle,
- obj->size / 1024);
-
- gtt_mapping = ioremap(dev->agp->base + obj_priv->gtt_offset,
- obj->size);
- if (gtt_mapping == NULL) {
- DRM_ERROR("failed to map GTT space\n");
- return;
- }
-
- for (page = 0; page < obj->size / PAGE_SIZE; page++) {
- int i;
-
- backing_map = kmap_atomic(obj_priv->page_list[page], KM_USER0);
-
- if (backing_map == NULL) {
- DRM_ERROR("failed to map backing page\n");
- goto out;
- }
-
- for (i = 0; i < PAGE_SIZE / 4; i++) {
- uint32_t cpuval = backing_map[i];
- uint32_t gttval = readl(gtt_mapping +
- page * 1024 + i);
-
- if (cpuval != gttval) {
- DRM_INFO("incoherent CPU vs GPU at 0x%08x: "
- "0x%08x vs 0x%08x\n",
- (int)(obj_priv->gtt_offset +
- page * PAGE_SIZE + i * 4),
- cpuval, gttval);
- if (bad_count++ >= 8) {
- DRM_INFO("...\n");
- goto out;
- }
- }
- }
- kunmap_atomic(backing_map, KM_USER0);
- backing_map = NULL;
- }
-
- out:
- if (backing_map != NULL)
- kunmap_atomic(backing_map, KM_USER0);
- iounmap(gtt_mapping);
-
- /* give syslog time to catch up */
- msleep(1);
-
- /* Directly flush the object, since we just loaded values with the CPU
- * from thebacking pages and we don't want to disturb the cache
- * management that we're trying to observe.
- */
-
- i915_gem_clflush_object(obj);
-}
-#endif
-
/**
- * Bind an object to the GTT and evaluate the relocations landing in it
- *
- *
+ * Pin an object to the GTT and evaluate the relocations landing in it.
*/
static int
-i915_gem_object_bind_and_relocate(struct drm_gem_object *obj,
- struct drm_file *file_priv,
- struct drm_i915_gem_exec_object *entry)
+i915_gem_object_pin_and_relocate(struct drm_gem_object *obj,
+ struct drm_file *file_priv,
+ struct drm_i915_gem_exec_object *entry)
{
struct drm_device *dev = obj->dev;
struct drm_i915_gem_relocation_entry reloc;
struct drm_i915_gem_relocation_entry __user *relocs;
struct drm_i915_gem_object *obj_priv = obj->driver_private;
- int i;
+ int i, ret;
uint32_t last_reloc_offset = -1;
void *reloc_page = NULL;
/* Choose the GTT offset for our buffer and put it there. */
- if (obj_priv->gtt_space == NULL) {
- i915_gem_object_bind_to_gtt(obj, (unsigned) entry->alignment);
- if (obj_priv->gtt_space == NULL)
- return -ENOMEM;
- }
+ ret = i915_gem_object_pin(obj, (uint32_t) entry->alignment);
+ if (ret)
+ return ret;
entry->offset = obj_priv->gtt_offset;
int ret;
ret = copy_from_user(&reloc, relocs + i, sizeof(reloc));
- if (ret != 0)
+ if (ret != 0) {
+ i915_gem_object_unpin(obj);
return ret;
+ }
target_obj = drm_gem_object_lookup(obj->dev, file_priv,
reloc.target_handle);
- if (target_obj == NULL)
- return -EINVAL;
+ if (target_obj == NULL) {
+ i915_gem_object_unpin(obj);
+ return -EBADF;
+ }
target_obj_priv = target_obj->driver_private;
/* The target buffer should have appeared before us in the
DRM_ERROR("No GTT space found for object %d\n",
reloc.target_handle);
drm_gem_object_unreference(target_obj);
+ i915_gem_object_unpin(obj);
return -EINVAL;
}
obj, reloc.target_handle,
(int) reloc.offset, (int) obj->size);
drm_gem_object_unreference(target_obj);
+ i915_gem_object_unpin(obj);
return -EINVAL;
}
if (reloc.offset & 3) {
obj, reloc.target_handle,
(int) reloc.offset);
drm_gem_object_unreference(target_obj);
+ i915_gem_object_unpin(obj);
return -EINVAL;
}
reloc.write_domain,
target_obj->pending_write_domain);
drm_gem_object_unreference(target_obj);
+ i915_gem_object_unpin(obj);
return -EINVAL;
}
/* As we're writing through the gtt, flush
* any CPU writes before we write the relocations
*/
- if (obj->write_domain & DRM_GEM_DOMAIN_CPU) {
+ if (obj->write_domain & I915_GEM_DOMAIN_CPU) {
i915_gem_clflush_object(obj);
drm_agp_chipset_flush(dev);
obj->write_domain = 0;
last_reloc_offset = reloc_offset;
if (reloc_page == NULL) {
drm_gem_object_unreference(target_obj);
+ i915_gem_object_unpin(obj);
return -ENOMEM;
}
}
ret = copy_to_user(relocs + i, &reloc, sizeof(reloc));
if (ret != 0) {
drm_gem_object_unreference(target_obj);
+ i915_gem_object_unpin(obj);
return ret;
}
return ret;
}
- if (dev_priv->use_mi_batchbuffer_start) {
+ if (IS_I830(dev) || IS_845G(dev)) {
+ BEGIN_LP_RING(4);
+ OUT_RING(MI_BATCH_BUFFER);
+ OUT_RING(exec_start | MI_BATCH_NON_SECURE);
+ OUT_RING(exec_start + exec_len - 4);
+ OUT_RING(0);
+ ADVANCE_LP_RING();
+ } else {
BEGIN_LP_RING(2);
if (IS_I965G(dev)) {
OUT_RING(MI_BATCH_BUFFER_START |
OUT_RING(exec_start | MI_BATCH_NON_SECURE);
}
ADVANCE_LP_RING();
-
- } else {
- BEGIN_LP_RING(4);
- OUT_RING(MI_BATCH_BUFFER);
- OUT_RING(exec_start | MI_BATCH_NON_SECURE);
- OUT_RING(exec_start + exec_len - 4);
- OUT_RING(0);
- ADVANCE_LP_RING();
}
}
* relatively low latency when blocking on a particular request to finish.
*/
static int
-i915_gem_ring_throttle(struct drm_device *dev)
+i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file_priv)
{
- drm_i915_private_t *dev_priv = dev->dev_private;
+ struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
int ret = 0;
+ uint32_t seqno;
mutex_lock(&dev->struct_mutex);
- while (!list_empty(&dev_priv->mm.request_list)) {
- struct drm_i915_gem_request *request;
-
- request = list_first_entry(&dev_priv->mm.request_list,
- struct drm_i915_gem_request,
- list);
-
- /* Break out if we're close enough. */
- if ((long) (jiffies - request->emitted_jiffies) <= (20 * HZ) / 1000) {
- mutex_unlock(&dev->struct_mutex);
- return 0;
- }
-
- /* Wait on the last request if not. */
- ret = i915_wait_request(dev, request->seqno);
- if (ret != 0) {
- mutex_unlock(&dev->struct_mutex);
- return ret;
- }
- }
+ seqno = i915_file_priv->mm.last_gem_throttle_seqno;
+ i915_file_priv->mm.last_gem_throttle_seqno =
+ i915_file_priv->mm.last_gem_seqno;
+ if (seqno)
+ ret = i915_wait_request(dev, seqno);
mutex_unlock(&dev->struct_mutex);
return ret;
}
i915_gem_execbuffer(struct drm_device *dev, void *data,
struct drm_file *file_priv)
{
+ drm_i915_private_t *dev_priv = dev->dev_private;
+ struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
struct drm_i915_gem_execbuffer *args = data;
struct drm_i915_gem_exec_object *exec_list = NULL;
struct drm_gem_object **object_list = NULL;
struct drm_gem_object *batch_obj;
- int ret, i;
+ int ret, i, pinned = 0;
uint64_t exec_offset;
uint32_t seqno, flush_domains;
- LOCK_TEST_WITH_RETURN(dev, file_priv);
-
#if WATCH_EXEC
DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
(int) args->buffers_ptr, args->buffer_count, args->batch_len);
#endif
- i915_kernel_lost_context(dev);
-
- ret = i915_gem_ring_throttle(dev);
- if (ret)
- return ret;
/* Copy in the exec list from userland */
exec_list = drm_calloc(sizeof(*exec_list), args->buffer_count,
mutex_lock(&dev->struct_mutex);
+ i915_verify_inactive(dev, __FILE__, __LINE__);
+
+ if (dev_priv->mm.wedged) {
+ DRM_ERROR("Execbuf while wedged\n");
+ mutex_unlock(&dev->struct_mutex);
+ return -EIO;
+ }
+
+ if (dev_priv->mm.suspended) {
+ DRM_ERROR("Execbuf while VT-switched.\n");
+ mutex_unlock(&dev->struct_mutex);
+ return -EBUSY;
+ }
+
/* Zero the gloabl flush/invalidate flags. These
* will be modified as each object is bound to the
* gtt
if (object_list[i] == NULL) {
DRM_ERROR("Invalid object handle %d at index %d\n",
exec_list[i].handle, i);
- ret = -EINVAL;
+ ret = -EBADF;
goto err;
}
object_list[i]->pending_read_domains = 0;
object_list[i]->pending_write_domain = 0;
- ret = i915_gem_object_bind_and_relocate(object_list[i],
- file_priv,
- &exec_list[i]);
+ ret = i915_gem_object_pin_and_relocate(object_list[i],
+ file_priv,
+ &exec_list[i]);
if (ret) {
DRM_ERROR("object bind and relocate failed %d\n", ret);
goto err;
}
+ pinned = i + 1;
}
/* Set the pending read domains for the batch buffer to COMMAND */
batch_obj = object_list[args->buffer_count-1];
- batch_obj->pending_read_domains = DRM_GEM_DOMAIN_I915_COMMAND;
+ batch_obj->pending_read_domains = I915_GEM_DOMAIN_COMMAND;
batch_obj->pending_write_domain = 0;
+ i915_verify_inactive(dev, __FILE__, __LINE__);
+
for (i = 0; i < args->buffer_count; i++) {
struct drm_gem_object *obj = object_list[i];
struct drm_i915_gem_object *obj_priv = obj->driver_private;
goto err;
}
+ i915_verify_inactive(dev, __FILE__, __LINE__);
+
/* Flush/invalidate caches and chipset buffer */
flush_domains = i915_gem_dev_set_domain(dev);
+ i915_verify_inactive(dev, __FILE__, __LINE__);
+
#if WATCH_COHERENCY
for (i = 0; i < args->buffer_count; i++) {
i915_gem_object_check_coherency(object_list[i],
*/
flush_domains |= i915_retire_commands(dev);
+ i915_verify_inactive(dev, __FILE__, __LINE__);
+
/*
* Get a seqno representing the execution of the current buffer,
* which we can wait on. We would like to mitigate these interrupts,
*/
seqno = i915_add_request(dev, flush_domains);
BUG_ON(seqno == 0);
+ i915_file_priv->mm.last_gem_seqno = seqno;
for (i = 0; i < args->buffer_count; i++) {
struct drm_gem_object *obj = object_list[i];
struct drm_i915_gem_object *obj_priv = obj->driver_private;
i915_dump_lru(dev, __func__);
#endif
+ i915_verify_inactive(dev, __FILE__, __LINE__);
+
/* Copy the new buffer offsets back to the user's exec list. */
ret = copy_to_user((struct drm_i915_relocation_entry __user *)
(uintptr_t) args->buffers_ptr,
args->buffer_count, ret);
err:
if (object_list != NULL) {
+ for (i = 0; i < pinned; i++)
+ i915_gem_object_unpin(object_list[i]);
+
for (i = 0; i < args->buffer_count; i++)
drm_gem_object_unreference(object_list[i]);
}
}
int
+i915_gem_object_pin(struct drm_gem_object *obj, uint32_t alignment)
+{
+ struct drm_device *dev = obj->dev;
+ struct drm_i915_gem_object *obj_priv = obj->driver_private;
+ int ret;
+
+ i915_verify_inactive(dev, __FILE__, __LINE__);
+ if (obj_priv->gtt_space == NULL) {
+ ret = i915_gem_object_bind_to_gtt(obj, alignment);
+ if (ret != 0) {
+ DRM_ERROR("Failure to bind: %d", ret);
+ return ret;
+ }
+ }
+ obj_priv->pin_count++;
+
+ /* If the object is not active and not pending a flush,
+ * remove it from the inactive list
+ */
+ if (obj_priv->pin_count == 1) {
+ atomic_inc(&dev->pin_count);
+ atomic_add(obj->size, &dev->pin_memory);
+ if (!obj_priv->active &&
+ (obj->write_domain & ~(I915_GEM_DOMAIN_CPU |
+ I915_GEM_DOMAIN_GTT)) == 0 &&
+ !list_empty(&obj_priv->list))
+ list_del_init(&obj_priv->list);
+ }
+ i915_verify_inactive(dev, __FILE__, __LINE__);
+
+ return 0;
+}
+
+void
+i915_gem_object_unpin(struct drm_gem_object *obj)
+{
+ struct drm_device *dev = obj->dev;
+ drm_i915_private_t *dev_priv = dev->dev_private;
+ struct drm_i915_gem_object *obj_priv = obj->driver_private;
+
+ i915_verify_inactive(dev, __FILE__, __LINE__);
+ obj_priv->pin_count--;
+ BUG_ON(obj_priv->pin_count < 0);
+ BUG_ON(obj_priv->gtt_space == NULL);
+
+ /* If the object is no longer pinned, and is
+ * neither active nor being flushed, then stick it on
+ * the inactive list
+ */
+ if (obj_priv->pin_count == 0) {
+ if (!obj_priv->active &&
+ (obj->write_domain & ~(I915_GEM_DOMAIN_CPU |
+ I915_GEM_DOMAIN_GTT)) == 0)
+ list_move_tail(&obj_priv->list,
+ &dev_priv->mm.inactive_list);
+ atomic_dec(&dev->pin_count);
+ atomic_sub(obj->size, &dev->pin_memory);
+ }
+ i915_verify_inactive(dev, __FILE__, __LINE__);
+}
+
+int
i915_gem_pin_ioctl(struct drm_device *dev, void *data,
struct drm_file *file_priv)
{
mutex_lock(&dev->struct_mutex);
- i915_kernel_lost_context(dev);
obj = drm_gem_object_lookup(dev, file_priv, args->handle);
if (obj == NULL) {
DRM_ERROR("Bad handle in i915_gem_pin_ioctl(): %d\n",
args->handle);
mutex_unlock(&dev->struct_mutex);
- return -EINVAL;
+ return -EBADF;
}
-
obj_priv = obj->driver_private;
- if (obj_priv->gtt_space == NULL) {
- ret = i915_gem_object_bind_to_gtt(obj,
- (unsigned) args->alignment);
- if (ret != 0) {
- DRM_ERROR("Failure to bind in "
- "i915_gem_pin_ioctl(): %d\n",
- ret);
- drm_gem_object_unreference(obj);
- mutex_unlock(&dev->struct_mutex);
- return ret;
- }
+
+ ret = i915_gem_object_pin(obj, args->alignment);
+ if (ret != 0) {
+ drm_gem_object_unreference(obj);
+ mutex_unlock(&dev->struct_mutex);
+ return ret;
}
- obj_priv->pin_count++;
+ /* XXX - flush the CPU caches for pinned objects
+ * as the X server doesn't manage domains yet
+ */
+ if (obj->write_domain & I915_GEM_DOMAIN_CPU) {
+ i915_gem_clflush_object(obj);
+ drm_agp_chipset_flush(dev);
+ obj->write_domain = 0;
+ }
args->offset = obj_priv->gtt_offset;
drm_gem_object_unreference(obj);
mutex_unlock(&dev->struct_mutex);
{
struct drm_i915_gem_pin *args = data;
struct drm_gem_object *obj;
- struct drm_i915_gem_object *obj_priv;
mutex_lock(&dev->struct_mutex);
- i915_kernel_lost_context(dev);
obj = drm_gem_object_lookup(dev, file_priv, args->handle);
if (obj == NULL) {
DRM_ERROR("Bad handle in i915_gem_unpin_ioctl(): %d\n",
args->handle);
mutex_unlock(&dev->struct_mutex);
- return -EINVAL;
+ return -EBADF;
}
- obj_priv = obj->driver_private;
- obj_priv->pin_count--;
+ i915_gem_object_unpin(obj);
+
drm_gem_object_unreference(obj);
mutex_unlock(&dev->struct_mutex);
return 0;
DRM_ERROR("Bad handle in i915_gem_busy_ioctl(): %d\n",
args->handle);
mutex_unlock(&dev->struct_mutex);
- return -EINVAL;
+ return -EBADF;
}
obj_priv = obj->driver_private;
args->busy = obj_priv->active;
-
+
drm_gem_object_unreference(obj);
mutex_unlock(&dev->struct_mutex);
return 0;
}
+int
+i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
+ struct drm_file *file_priv)
+{
+ return i915_gem_ring_throttle(dev, file_priv);
+}
+
int i915_gem_init_object(struct drm_gem_object *obj)
{
struct drm_i915_gem_object *obj_priv;
if (obj_priv == NULL)
return -ENOMEM;
+ /*
+ * We've just allocated pages from the kernel,
+ * so they've just been written by the CPU with
+ * zeros. They'll need to be clflushed before we
+ * use them with the GPU.
+ */
+ obj->write_domain = I915_GEM_DOMAIN_CPU;
+ obj->read_domains = I915_GEM_DOMAIN_CPU;
+
obj->driver_private = obj_priv;
obj_priv->obj = obj;
INIT_LIST_HEAD(&obj_priv->list);
void i915_gem_free_object(struct drm_gem_object *obj)
{
- i915_kernel_lost_context(obj->dev);
+ struct drm_i915_gem_object *obj_priv = obj->driver_private;
+
+ while (obj_priv->pin_count > 0)
+ i915_gem_object_unpin(obj);
+
i915_gem_object_unbind(obj);
+ drm_free(obj_priv->page_cpu_valid, 1, DRM_MEM_DRIVER);
drm_free(obj->driver_private, 1, DRM_MEM_DRIVER);
}
{
struct drm_device *dev = obj->dev;
int ret;
+ uint32_t flush_domains;
BUG_ON(!mutex_is_locked(&dev->struct_mutex));
- drm_client_lock_take(dev, file_priv);
- i915_kernel_lost_context(dev);
ret = i915_gem_object_set_domain(obj, read_domains, write_domain);
- if (ret) {
- drm_client_lock_release(dev);
+ if (ret)
return ret;
+ flush_domains = i915_gem_dev_set_domain(obj->dev);
+
+ if (flush_domains & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT))
+ (void) i915_add_request(dev, flush_domains);
+
+ return 0;
+}
+
+/** Unbinds all objects that are on the given buffer list. */
+static int
+i915_gem_evict_from_list(struct drm_device *dev, struct list_head *head)
+{
+ struct drm_gem_object *obj;
+ struct drm_i915_gem_object *obj_priv;
+ int ret;
+
+ while (!list_empty(head)) {
+ obj_priv = list_first_entry(head,
+ struct drm_i915_gem_object,
+ list);
+ obj = obj_priv->obj;
+
+ if (obj_priv->pin_count != 0) {
+ DRM_ERROR("Pinned object in unbind list\n");
+ mutex_unlock(&dev->struct_mutex);
+ return -EINVAL;
+ }
+
+ ret = i915_gem_object_unbind(obj);
+ if (ret != 0) {
+ DRM_ERROR("Error unbinding object in LeaveVT: %d\n",
+ ret);
+ mutex_unlock(&dev->struct_mutex);
+ return ret;
+ }
}
- i915_gem_dev_set_domain(obj->dev);
- drm_client_lock_release(dev);
+
+
return 0;
}
-int
-i915_gem_flush_pwrite(struct drm_gem_object *obj,
- uint64_t offset, uint64_t size)
+static int
+i915_gem_idle(struct drm_device *dev)
{
-#if 0
- struct drm_device *dev = obj->dev;
- struct drm_i915_gem_object *obj_priv = obj->driver_private;
+ drm_i915_private_t *dev_priv = dev->dev_private;
+ uint32_t seqno, cur_seqno, last_seqno;
+ int stuck;
- /*
- * For writes much less than the size of the object and
- * which are already pinned in memory, do the flush right now
+ if (dev_priv->mm.suspended)
+ return 0;
+
+ /* Hack! Don't let anybody do execbuf while we don't control the chip.
+ * We need to replace this with a semaphore, or something.
*/
+ dev_priv->mm.suspended = 1;
- if ((size < obj->size >> 1) && obj_priv->page_list != NULL) {
- unsigned long first_page = offset / PAGE_SIZE;
- unsigned long beyond_page = roundup(offset + size, PAGE_SIZE) / PAGE_SIZE;
+ i915_kernel_lost_context(dev);
- drm_ttm_cache_flush(obj_priv->page_list + first_page,
- beyond_page - first_page);
- drm_agp_chipset_flush(dev);
- obj->write_domain = 0;
+ /* Flush the GPU along with all non-CPU write domains
+ */
+ i915_gem_flush(dev, ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT),
+ ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT));
+ seqno = i915_add_request(dev, ~(I915_GEM_DOMAIN_CPU |
+ I915_GEM_DOMAIN_GTT));
+
+ if (seqno == 0) {
+ mutex_unlock(&dev->struct_mutex);
+ return -ENOMEM;
}
-#endif
+
+ dev_priv->mm.waiting_gem_seqno = seqno;
+ last_seqno = 0;
+ stuck = 0;
+ for (;;) {
+ cur_seqno = i915_get_gem_seqno(dev);
+ if (i915_seqno_passed(cur_seqno, seqno))
+ break;
+ if (last_seqno == cur_seqno) {
+ if (stuck++ > 100) {
+ DRM_ERROR("hardware wedged\n");
+ dev_priv->mm.wedged = 1;
+ DRM_WAKEUP(&dev_priv->irq_queue);
+ break;
+ }
+ }
+ msleep(10);
+ last_seqno = cur_seqno;
+ }
+ dev_priv->mm.waiting_gem_seqno = 0;
+
+ i915_gem_retire_requests(dev);
+
+ /* Active and flushing should now be empty as we've
+ * waited for a sequence higher than any pending execbuffer
+ */
+ BUG_ON(!list_empty(&dev_priv->mm.active_list));
+ BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
+
+ /* Request should now be empty as we've also waited
+ * for the last request in the list
+ */
+ BUG_ON(!list_empty(&dev_priv->mm.request_list));
+
+ /* Move all buffers out of the GTT. */
+ i915_gem_evict_from_list(dev, &dev_priv->mm.inactive_list);
+
+ BUG_ON(!list_empty(&dev_priv->mm.active_list));
+ BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
+ BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
+ BUG_ON(!list_empty(&dev_priv->mm.request_list));
return 0;
}
-void
-i915_gem_lastclose(struct drm_device *dev)
+static int
+i915_gem_init_hws(struct drm_device *dev)
+{
+ drm_i915_private_t *dev_priv = dev->dev_private;
+ struct drm_gem_object *obj;
+ struct drm_i915_gem_object *obj_priv;
+ int ret;
+
+ /* If we need a physical address for the status page, it's already
+ * initialized at driver load time.
+ */
+ if (!I915_NEED_GFX_HWS(dev))
+ return 0;
+
+ obj = drm_gem_object_alloc(dev, 4096);
+ if (obj == NULL) {
+ DRM_ERROR("Failed to allocate status page\n");
+ return -ENOMEM;
+ }
+ obj_priv = obj->driver_private;
+
+ ret = i915_gem_object_pin(obj, 4096);
+ if (ret != 0) {
+ drm_gem_object_unreference(obj);
+ return ret;
+ }
+
+ dev_priv->status_gfx_addr = obj_priv->gtt_offset;
+ dev_priv->hws_map.offset = dev->agp->base + obj_priv->gtt_offset;
+ dev_priv->hws_map.size = 4096;
+ dev_priv->hws_map.type = 0;
+ dev_priv->hws_map.flags = 0;
+ dev_priv->hws_map.mtrr = 0;
+
+ drm_core_ioremap(&dev_priv->hws_map, dev);
+ if (dev_priv->hws_map.handle == NULL) {
+ DRM_ERROR("Failed to map status page.\n");
+ memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
+ drm_gem_object_unreference(obj);
+ return -EINVAL;
+ }
+ dev_priv->hws_obj = obj;
+ dev_priv->hw_status_page = dev_priv->hws_map.handle;
+ memset(dev_priv->hw_status_page, 0, PAGE_SIZE);
+ I915_WRITE(HWS_PGA, dev_priv->status_gfx_addr);
+ DRM_DEBUG("hws offset: 0x%08x\n", dev_priv->status_gfx_addr);
+
+ return 0;
+}
+
+static int
+i915_gem_init_ringbuffer(struct drm_device *dev)
+{
+ drm_i915_private_t *dev_priv = dev->dev_private;
+ struct drm_gem_object *obj;
+ struct drm_i915_gem_object *obj_priv;
+ int ret;
+
+ ret = i915_gem_init_hws(dev);
+ if (ret != 0)
+ return ret;
+
+ obj = drm_gem_object_alloc(dev, 128 * 1024);
+ if (obj == NULL) {
+ DRM_ERROR("Failed to allocate ringbuffer\n");
+ return -ENOMEM;
+ }
+ obj_priv = obj->driver_private;
+
+ ret = i915_gem_object_pin(obj, 4096);
+ if (ret != 0) {
+ drm_gem_object_unreference(obj);
+ return ret;
+ }
+
+ /* Set up the kernel mapping for the ring. */
+ dev_priv->ring.Size = obj->size;
+ dev_priv->ring.tail_mask = obj->size - 1;
+
+ dev_priv->ring.map.offset = dev->agp->base + obj_priv->gtt_offset;
+ dev_priv->ring.map.size = obj->size;
+ dev_priv->ring.map.type = 0;
+ dev_priv->ring.map.flags = 0;
+ dev_priv->ring.map.mtrr = 0;
+
+ drm_core_ioremap(&dev_priv->ring.map, dev);
+ if (dev_priv->ring.map.handle == NULL) {
+ DRM_ERROR("Failed to map ringbuffer.\n");
+ memset(&dev_priv->ring, 0, sizeof(dev_priv->ring));
+ drm_gem_object_unreference(obj);
+ return -EINVAL;
+ }
+ dev_priv->ring.ring_obj = obj;
+ dev_priv->ring.virtual_start = dev_priv->ring.map.handle;
+
+ /* Stop the ring if it's running. */
+ I915_WRITE(PRB0_CTL, 0);
+ I915_WRITE(PRB0_HEAD, 0);
+ I915_WRITE(PRB0_TAIL, 0);
+ I915_WRITE(PRB0_START, 0);
+
+ /* Initialize the ring. */
+ I915_WRITE(PRB0_START, obj_priv->gtt_offset);
+ I915_WRITE(PRB0_CTL,
+ ((obj->size - 4096) & RING_NR_PAGES) |
+ RING_NO_REPORT |
+ RING_VALID);
+
+ /* Update our cache of the ring state */
+ i915_kernel_lost_context(dev);
+
+ return 0;
+}
+
+static void
+i915_gem_cleanup_ringbuffer(struct drm_device *dev)
{
drm_i915_private_t *dev_priv = dev->dev_private;
+ if (dev_priv->ring.ring_obj == NULL)
+ return;
+
+ drm_core_ioremapfree(&dev_priv->ring.map, dev);
+
+ i915_gem_object_unpin(dev_priv->ring.ring_obj);
+ drm_gem_object_unreference(dev_priv->ring.ring_obj);
+ dev_priv->ring.ring_obj = NULL;
+ memset(&dev_priv->ring, 0, sizeof(dev_priv->ring));
+
+ if (dev_priv->hws_obj != NULL) {
+ i915_gem_object_unpin(dev_priv->hws_obj);
+ drm_gem_object_unreference(dev_priv->hws_obj);
+ dev_priv->hws_obj = NULL;
+ memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
+
+ /* Write high address into HWS_PGA when disabling. */
+ I915_WRITE(HWS_PGA, 0x1ffff000);
+ }
+}
+
+int
+i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
+ struct drm_file *file_priv)
+{
+ drm_i915_private_t *dev_priv = dev->dev_private;
+ int ret;
+
+ if (dev_priv->mm.wedged) {
+ DRM_ERROR("Reenabling wedged hardware, good luck\n");
+ dev_priv->mm.wedged = 0;
+ }
+
+ ret = i915_gem_init_ringbuffer(dev);
+ if (ret != 0)
+ return ret;
+
mutex_lock(&dev->struct_mutex);
+ BUG_ON(!list_empty(&dev_priv->mm.active_list));
+ BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
+ BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
+ BUG_ON(!list_empty(&dev_priv->mm.request_list));
+ dev_priv->mm.suspended = 0;
+ mutex_unlock(&dev->struct_mutex);
+ return 0;
+}
- /* Assume that the chip has been idled at this point. Just pull them
- * off the execution list and unref them. Since this is the last
- * close, this is also the last ref and they'll go away.
- */
+int
+i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
+ struct drm_file *file_priv)
+{
+ int ret;
- while (!list_empty(&dev_priv->mm.active_list)) {
- struct drm_i915_gem_object *obj_priv;
+ mutex_lock(&dev->struct_mutex);
+ ret = i915_gem_idle(dev);
+ if (ret == 0)
+ i915_gem_cleanup_ringbuffer(dev);
+ mutex_unlock(&dev->struct_mutex);
- obj_priv = list_first_entry(&dev_priv->mm.active_list,
- struct drm_i915_gem_object,
- list);
+ return 0;
+}
- list_del_init(&obj_priv->list);
- obj_priv->active = 0;
- obj_priv->obj->write_domain = 0;
- drm_gem_object_unreference(obj_priv->obj);
+void
+i915_gem_lastclose(struct drm_device *dev)
+{
+ int ret;
+ drm_i915_private_t *dev_priv = dev->dev_private;
+
+ mutex_lock(&dev->struct_mutex);
+
+ if (dev_priv->ring.ring_obj != NULL) {
+ ret = i915_gem_idle(dev);
+ if (ret)
+ DRM_ERROR("failed to idle hardware: %d\n", ret);
+
+ i915_gem_cleanup_ringbuffer(dev);
}
mutex_unlock(&dev->struct_mutex);
}
+
+void i915_gem_load(struct drm_device *dev)
+{
+ drm_i915_private_t *dev_priv = dev->dev_private;
+
+ INIT_LIST_HEAD(&dev_priv->mm.active_list);
+ INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
+ INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
+ INIT_LIST_HEAD(&dev_priv->mm.request_list);
+ INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
+ i915_gem_retire_work_handler);
+ dev_priv->mm.next_gem_seqno = 1;
+
+ i915_gem_detect_bit_6_swizzle(dev);
+}