openkylin
/
linux
mirror of https://gitee.com/openkylin/linux.git
9
0
Fork 0
Code Issues Projects Releases Wiki Activity

vmwgfx: Implement fence objects 

Will be needed for queries and drm event-driven throttling.

As a benefit, they help avoid stale user-space fence handles.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Jakob Bornecrantz <jakob@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This commit is contained in:
Thomas Hellstrom 2011-09-01 20:18:44 +00:00 committed by Dave Airlie
parent 4f73a96bd7
commit ae2a104058
10 changed files with 1010 additions and 87 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
drivers/gpu/drm/vmwgfx/Makefile
View File

@ -4,6 +4,7 @@ ccflags-y := -Iinclude/drm
vmwgfx-y := vmwgfx_execbuf.o vmwgfx_gmr.o vmwgfx_kms.o vmwgfx_drv.o \
vmwgfx_fb.o vmwgfx_ioctl.o vmwgfx_resource.o vmwgfx_buffer.o \
vmwgfx_fifo.o vmwgfx_irq.o vmwgfx_ldu.o vmwgfx_ttm_glue.o \
vmwgfx_overlay.o vmwgfx_marker.o vmwgfx_gmrid_manager.o
vmwgfx_overlay.o vmwgfx_marker.o vmwgfx_gmrid_manager.o \
vmwgfx_fence.o
obj-$(CONFIG_DRM_VMWGFX) := vmwgfx.o

26
drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
View File

@ -274,39 +274,39 @@ static int vmw_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
static void *vmw_sync_obj_ref(void *sync_obj)
{
return sync_obj;
return (void *)
vmw_fence_obj_reference((struct vmw_fence_obj *) sync_obj);
}
static void vmw_sync_obj_unref(void **sync_obj)
{
*sync_obj = NULL;
vmw_fence_obj_unreference((struct vmw_fence_obj **) sync_obj);
}
static int vmw_sync_obj_flush(void *sync_obj, void *sync_arg)
{
struct vmw_private *dev_priv = (struct vmw_private *)sync_arg;
mutex_lock(&dev_priv->hw_mutex);
vmw_write(dev_priv, SVGA_REG_SYNC, SVGA_SYNC_GENERIC);
mutex_unlock(&dev_priv->hw_mutex);
vmw_fence_obj_flush((struct vmw_fence_obj *) sync_obj);
return 0;
}
static bool vmw_sync_obj_signaled(void *sync_obj, void *sync_arg)
{
struct vmw_private *dev_priv = (struct vmw_private *)sync_arg;
uint32_t seqno = (unsigned long) sync_obj;
unsigned long flags = (unsigned long) sync_arg;
return vmw_fence_obj_signaled((struct vmw_fence_obj *) sync_obj,
(uint32_t) flags);
return vmw_seqno_passed(dev_priv, seqno);
}
static int vmw_sync_obj_wait(void *sync_obj, void *sync_arg,
bool lazy, bool interruptible)
{
struct vmw_private *dev_priv = (struct vmw_private *)sync_arg;
uint32_t seqno = (unsigned long) sync_obj;
unsigned long flags = (unsigned long) sync_arg;
return vmw_wait_seqno(dev_priv, false, seqno, false, 3*HZ);
return vmw_fence_obj_wait((struct vmw_fence_obj *) sync_obj,
(uint32_t) flags,
lazy, interruptible,
VMW_FENCE_WAIT_TIMEOUT);
}
struct ttm_bo_driver vmw_bo_driver = {

28
drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
View File

@ -82,12 +82,18 @@
#define DRM_IOCTL_VMW_EXECBUF \
DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_EXECBUF, \
struct drm_vmw_execbuf_arg)
#define DRM_IOCTL_VMW_GET_3D_CAP \
DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_GET_3D_CAP, \
struct drm_vmw_get_3d_cap_arg)
#define DRM_IOCTL_VMW_FENCE_WAIT \
DRM_IOWR(DRM_COMMAND_BASE + DRM_VMW_FENCE_WAIT, \
struct drm_vmw_fence_wait_arg)
#define DRM_IOCTL_VMW_GET_3D_CAP \
DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_GET_3D_CAP, \
struct drm_vmw_get_3d_cap_arg)
#define DRM_IOCTL_VMW_FENCE_SIGNALED \
DRM_IOWR(DRM_COMMAND_BASE + DRM_VMW_FENCE_SIGNALED, \
struct drm_vmw_fence_signaled_arg)
#define DRM_IOCTL_VMW_FENCE_UNREF \
DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_FENCE_UNREF, \
struct drm_vmw_fence_arg)
/**
* The core DRM version of this macro doesn't account for
@ -131,7 +137,12 @@ static struct drm_ioctl_desc vmw_ioctls[] = {
DRM_AUTH | DRM_UNLOCKED),
VMW_IOCTL_DEF(VMW_EXECBUF, vmw_execbuf_ioctl,
DRM_AUTH | DRM_UNLOCKED),
VMW_IOCTL_DEF(VMW_FENCE_WAIT, vmw_fence_wait_ioctl,
VMW_IOCTL_DEF(VMW_FENCE_WAIT, vmw_fence_obj_wait_ioctl,
DRM_AUTH | DRM_UNLOCKED),
VMW_IOCTL_DEF(VMW_FENCE_SIGNALED,
vmw_fence_obj_signaled_ioctl,
DRM_AUTH | DRM_UNLOCKED),
VMW_IOCTL_DEF(VMW_FENCE_UNREF, vmw_fence_obj_unref_ioctl,
DRM_AUTH | DRM_UNLOCKED),
VMW_IOCTL_DEF(VMW_GET_3D_CAP, vmw_get_cap_3d_ioctl,
DRM_AUTH | DRM_UNLOCKED),
@ -198,12 +209,14 @@ static int vmw_request_device(struct vmw_private *dev_priv)
DRM_ERROR("Unable to initialize FIFO.\n");
return ret;
}
vmw_fence_fifo_up(dev_priv->fman);
return 0;
}
static void vmw_release_device(struct vmw_private *dev_priv)
{
vmw_fence_fifo_down(dev_priv->fman);
vmw_fifo_release(dev_priv, &dev_priv->fifo);
}
@ -434,6 +447,10 @@ static int vmw_driver_load(struct drm_device *dev, unsigned long chipset)
goto out_no_device;
}
}
dev_priv->fman = vmw_fence_manager_init(dev_priv);
if (unlikely(dev_priv->fman == NULL))
goto out_no_fman;
ret = vmw_kms_init(dev_priv);
if (unlikely(ret != 0))
goto out_no_kms;
@ -475,6 +492,8 @@ static int vmw_driver_load(struct drm_device *dev, unsigned long chipset)
vmw_overlay_close(dev_priv);
vmw_kms_close(dev_priv);
out_no_kms:
vmw_fence_manager_takedown(dev_priv->fman);
out_no_fman:
if (dev_priv->stealth)
pci_release_region(dev->pdev, 2);
else
@ -518,6 +537,7 @@ static int vmw_driver_unload(struct drm_device *dev)
}
vmw_kms_close(dev_priv);
vmw_overlay_close(dev_priv);
vmw_fence_manager_takedown(dev_priv->fman);
if (dev_priv->stealth)
pci_release_region(dev->pdev, 2);
else

16
drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
View File

@ -38,6 +38,7 @@
#include "ttm/ttm_lock.h"
#include "ttm/ttm_execbuf_util.h"
#include "ttm/ttm_module.h"
#include "vmwgfx_fence.h"
#define VMWGFX_DRIVER_DATE "20100927"
#define VMWGFX_DRIVER_MAJOR 1
@ -53,6 +54,11 @@
#define VMW_PL_GMR TTM_PL_PRIV0
#define VMW_PL_FLAG_GMR TTM_PL_FLAG_PRIV0
#define VMW_RES_CONTEXT ttm_driver_type0
#define VMW_RES_SURFACE ttm_driver_type1
#define VMW_RES_STREAM ttm_driver_type2
#define VMW_RES_FENCE ttm_driver_type3
struct vmw_fpriv {
struct drm_master *locked_master;
struct ttm_object_file *tfile;
@ -245,6 +251,7 @@ struct vmw_private {
atomic_t fifo_queue_waiters;
uint32_t last_read_seqno;
spinlock_t irq_lock;
struct vmw_fence_manager *fman;
/*
* Device state
@ -456,8 +463,6 @@ extern int vmw_irq_postinstall(struct drm_device *dev);
extern void vmw_irq_uninstall(struct drm_device *dev);
extern bool vmw_seqno_passed(struct vmw_private *dev_priv,
uint32_t seqno);
extern int vmw_fence_wait_ioctl(struct drm_device *dev, void *data,
struct drm_file *file_priv);
extern int vmw_fallback_wait(struct vmw_private *dev_priv,
bool lazy,
bool fifo_idle,
@ -466,7 +471,8 @@ extern int vmw_fallback_wait(struct vmw_private *dev_priv,
unsigned long timeout);
extern void vmw_update_seqno(struct vmw_private *dev_priv,
struct vmw_fifo_state *fifo_state);
extern void vmw_seqno_waiter_add(struct vmw_private *dev_priv);
extern void vmw_seqno_waiter_remove(struct vmw_private *dev_priv);
/**
* Rudimentary fence-like objects currently used only for throttling -
@ -572,4 +578,8 @@ static inline struct vmw_dma_buffer *vmw_dmabuf_reference(struct vmw_dma_buffer
return NULL;
}
static inline struct ttm_mem_global *vmw_mem_glob(struct vmw_private *dev_priv)
{
return (struct ttm_mem_global *) dev_priv->mem_global_ref.object;
}
#endif

112
drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
View File

@ -256,7 +256,7 @@ static int vmw_translate_guest_ptr(struct vmw_private *dev_priv,
val_buf = &sw_context->val_bufs[cur_validate_node];
val_buf->bo = ttm_bo_reference(bo);
val_buf->usage = TTM_USAGE_READWRITE;
val_buf->new_sync_obj_arg = (void *) dev_priv;
val_buf->new_sync_obj_arg = (void *) DRM_VMW_FENCE_FLAG_EXEC;
list_add_tail(&val_buf->head, &sw_context->validate_nodes);
++sw_context->cur_val_buf;
}
@ -321,7 +321,6 @@ static int vmw_cmd_wait_query(struct vmw_private *dev_priv,
return 0;
}
static int vmw_cmd_dma(struct vmw_private *dev_priv,
struct vmw_sw_context *sw_context,
SVGA3dCmdHeader *header)
@ -676,6 +675,50 @@ static int vmw_resize_cmd_bounce(struct vmw_sw_context *sw_context,
return 0;
}
/**
* vmw_execbuf_fence_commands - create and submit a command stream fence
*
* Creates a fence object and submits a command stream marker.
* If this fails for some reason, We sync the fifo and return NULL.
* It is then safe to fence buffers with a NULL pointer.
*/
int vmw_execbuf_fence_commands(struct drm_file *file_priv,
struct vmw_private *dev_priv,
struct vmw_fence_obj **p_fence,
uint32_t *p_handle)
{
uint32_t sequence;
int ret;
bool synced = false;
ret = vmw_fifo_send_fence(dev_priv, &sequence);
if (unlikely(ret != 0)) {
DRM_ERROR("Fence submission error. Syncing.\n");
synced = true;
}
if (p_handle != NULL)
ret = vmw_user_fence_create(file_priv, dev_priv->fman,
sequence,
DRM_VMW_FENCE_FLAG_EXEC,
p_fence, p_handle);
else
ret = vmw_fence_create(dev_priv->fman, sequence,
DRM_VMW_FENCE_FLAG_EXEC,
p_fence);
if (unlikely(ret != 0 && !synced)) {
(void) vmw_fallback_wait(dev_priv, false, false,
sequence, false,
VMW_FENCE_WAIT_TIMEOUT);
*p_fence = NULL;
}
return 0;
}
int vmw_execbuf_ioctl(struct drm_device *dev, void *data,
struct drm_file *file_priv)
{
@ -686,9 +729,10 @@ int vmw_execbuf_ioctl(struct drm_device *dev, void *data,
int ret;
void *user_cmd;
void *cmd;
uint32_t seqno;
struct vmw_sw_context *sw_context = &dev_priv->ctx;
struct vmw_master *vmaster = vmw_master(file_priv->master);
struct vmw_fence_obj *fence;
uint32_t handle;
ret = ttm_read_lock(&vmaster->lock, true);
if (unlikely(ret != 0))
@ -755,34 +799,60 @@ int vmw_execbuf_ioctl(struct drm_device *dev, void *data,
memcpy(cmd, sw_context->cmd_bounce, arg->command_size);
vmw_fifo_commit(dev_priv, arg->command_size);
ret = vmw_fifo_send_fence(dev_priv, &seqno);
ttm_eu_fence_buffer_objects(&sw_context->validate_nodes,
(void *)(unsigned long) seqno);
vmw_clear_validations(sw_context);
mutex_unlock(&dev_priv->cmdbuf_mutex);
user_fence_rep = (struct drm_vmw_fence_rep __user *)
(unsigned long)arg->fence_rep;
ret = vmw_execbuf_fence_commands(file_priv, dev_priv,
&fence,
(user_fence_rep) ? &handle : NULL);
/*
* This error is harmless, because if fence submission fails,
* vmw_fifo_send_fence will sync.
* vmw_fifo_send_fence will sync. The error will be propagated to
* user-space in @fence_rep
*/
if (ret != 0)
DRM_ERROR("Fence submission error. Syncing.\n");
fence_rep.error = ret;
fence_rep.fence_seq = (uint64_t) seqno;
fence_rep.pad64 = 0;
ttm_eu_fence_buffer_objects(&sw_context->validate_nodes,
(void *) fence);
user_fence_rep = (struct drm_vmw_fence_rep __user *)
(unsigned long)arg->fence_rep;
vmw_clear_validations(sw_context);
mutex_unlock(&dev_priv->cmdbuf_mutex);
/*
* copy_to_user errors will be detected by user space not
* seeing fence_rep::error filled in.
*/
if (user_fence_rep) {
fence_rep.error = ret;
fence_rep.handle = handle;
fence_rep.seqno = fence->seqno;
vmw_update_seqno(dev_priv, &dev_priv->fifo);
fence_rep.passed_seqno = dev_priv->last_read_seqno;
ret = copy_to_user(user_fence_rep, &fence_rep, sizeof(fence_rep));
/*
* copy_to_user errors will be detected by user space not
* seeing fence_rep::error filled in. Typically
* user-space would have pre-set that member to -EFAULT.
*/
ret = copy_to_user(user_fence_rep, &fence_rep,
sizeof(fence_rep));
/*
* User-space lost the fence object. We need to sync
* and unreference the handle.
*/
if (unlikely(ret != 0) && (fence_rep.error == 0)) {
BUG_ON(fence == NULL);
ttm_ref_object_base_unref(vmw_fpriv(file_priv)->tfile,
handle, TTM_REF_USAGE);
DRM_ERROR("Fence copy error. Syncing.\n");
(void) vmw_fence_obj_wait(fence,
fence->signal_mask,
false, false,
VMW_FENCE_WAIT_TIMEOUT);
}
}
if (likely(fence != NULL))
vmw_fence_obj_unreference(&fence);
vmw_kms_cursor_post_execbuf(dev_priv);
ttm_read_unlock(&vmaster->lock);

619
drivers/gpu/drm/vmwgfx/vmwgfx_fence.c Normal file
View File

@ -0,0 +1,619 @@
/**************************************************************************
*
* Copyright © 2011 VMware, Inc., Palo Alto, CA., USA
* All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the
* "Software"), to deal in the Software without restriction, including
* without limitation the rights to use, copy, modify, merge, publish,
* distribute, sub license, and/or sell copies of the Software, and to
* permit persons to whom the Software is furnished to do so, subject to
* the following conditions:
*
* The above copyright notice and this permission notice (including the
* next paragraph) shall be included in all copies or substantial portions
* of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
* THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
* USE OR OTHER DEALINGS IN THE SOFTWARE.
*
**************************************************************************/
#include "drmP.h"
#include "vmwgfx_drv.h"
#define VMW_FENCE_WRAP (1 << 31)
struct vmw_fence_manager {
int num_fence_objects;
struct vmw_private *dev_priv;
spinlock_t lock;
u32 next_seqno;
struct list_head fence_list;
struct work_struct work;
u32 user_fence_size;
u32 fence_size;
bool fifo_down;
struct list_head cleanup_list;
};
struct vmw_user_fence {
struct ttm_base_object base;
struct vmw_fence_obj fence;
};
/**
* vmw_fence_destroy_locked
*
*/
static void vmw_fence_obj_destroy_locked(struct kref *kref)
{
struct vmw_fence_obj *fence =
container_of(kref, struct vmw_fence_obj, kref);
struct vmw_fence_manager *fman = fence->fman;
unsigned int num_fences;
list_del_init(&fence->head);
num_fences = --fman->num_fence_objects;
spin_unlock_irq(&fman->lock);
if (fence->destroy)
fence->destroy(fence);
else
kfree(fence);
spin_lock_irq(&fman->lock);
}
/**
* Execute signal actions on fences recently signaled.
* This is done from a workqueue so we don't have to execute
* signal actions from atomic context.
*/
static void vmw_fence_work_func(struct work_struct *work)
{
struct vmw_fence_manager *fman =
container_of(work, struct vmw_fence_manager, work);
struct list_head list;
struct vmw_fence_action *action, *next_action;
do {
INIT_LIST_HEAD(&list);
spin_lock_irq(&fman->lock);
list_splice_init(&fman->cleanup_list, &list);
spin_unlock_irq(&fman->lock);
if (list_empty(&list))
return;
/*
* At this point, only we should be able to manipulate the
* list heads of the actions we have on the private list.
*/
list_for_each_entry_safe(action, next_action, &list, head) {
list_del_init(&action->head);
action->cleanup(action);
}
} while (1);
}
struct vmw_fence_manager *vmw_fence_manager_init(struct vmw_private *dev_priv)
{
struct vmw_fence_manager *fman = kzalloc(sizeof(*fman), GFP_KERNEL);
if (unlikely(fman == NULL))
return NULL;
fman->dev_priv = dev_priv;
spin_lock_init(&fman->lock);
INIT_LIST_HEAD(&fman->fence_list);
INIT_LIST_HEAD(&fman->cleanup_list);
INIT_WORK(&fman->work, &vmw_fence_work_func);
fman->fifo_down = true;
fman->user_fence_size = ttm_round_pot(sizeof(struct vmw_user_fence));
fman->fence_size = ttm_round_pot(sizeof(struct vmw_fence_obj));
return fman;
}
void vmw_fence_manager_takedown(struct vmw_fence_manager *fman)
{
unsigned long irq_flags;
bool lists_empty;
(void) cancel_work_sync(&fman->work);
spin_lock_irqsave(&fman->lock, irq_flags);
lists_empty = list_empty(&fman->fence_list) &&
list_empty(&fman->cleanup_list);
spin_unlock_irqrestore(&fman->lock, irq_flags);
BUG_ON(!lists_empty);
kfree(fman);
}
static int vmw_fence_obj_init(struct vmw_fence_manager *fman,
struct vmw_fence_obj *fence,
u32 seqno,
uint32_t mask,
void (*destroy) (struct vmw_fence_obj *fence))
{
unsigned long irq_flags;
unsigned int num_fences;
int ret = 0;
fence->seqno = seqno;
INIT_LIST_HEAD(&fence->seq_passed_actions);
fence->fman = fman;
fence->signaled = 0;
fence->signal_mask = mask;
kref_init(&fence->kref);
fence->destroy = destroy;
init_waitqueue_head(&fence->queue);
spin_lock_irqsave(&fman->lock, irq_flags);
if (unlikely(fman->fifo_down)) {
ret = -EBUSY;
goto out_unlock;
}
list_add_tail(&fence->head, &fman->fence_list);
num_fences = ++fman->num_fence_objects;
out_unlock:
spin_unlock_irqrestore(&fman->lock, irq_flags);
return ret;
}
struct vmw_fence_obj *vmw_fence_obj_reference(struct vmw_fence_obj *fence)
{
kref_get(&fence->kref);
return fence;
}
/**
* vmw_fence_obj_unreference
*
* Note that this function may not be entered with disabled irqs since
* it may re-enable them in the destroy function.
*
*/
void vmw_fence_obj_unreference(struct vmw_fence_obj **fence_p)
{
struct vmw_fence_obj *fence = *fence_p;
struct vmw_fence_manager *fman = fence->fman;
*fence_p = NULL;
spin_lock_irq(&fman->lock);
BUG_ON(atomic_read(&fence->kref.refcount) == 0);
kref_put(&fence->kref, vmw_fence_obj_destroy_locked);
spin_unlock_irq(&fman->lock);
}
void vmw_fences_perform_actions(struct vmw_fence_manager *fman,
struct list_head *list)
{
struct vmw_fence_action *action, *next_action;
list_for_each_entry_safe(action, next_action, list, head) {
list_del_init(&action->head);
if (action->seq_passed != NULL)
action->seq_passed(action);
/*
* Add the cleanup action to the cleanup list so that
* it will be performed by a worker task.
*/
if (action->cleanup != NULL)
list_add_tail(&action->head, &fman->cleanup_list);
}
}
void vmw_fences_update(struct vmw_fence_manager *fman, u32 seqno)
{
unsigned long flags;
struct vmw_fence_obj *fence, *next_fence;
struct list_head action_list;
spin_lock_irqsave(&fman->lock, flags);
list_for_each_entry_safe(fence, next_fence, &fman->fence_list, head) {
if (seqno - fence->seqno < VMW_FENCE_WRAP) {
list_del_init(&fence->head);
fence->signaled |= DRM_VMW_FENCE_FLAG_EXEC;
INIT_LIST_HEAD(&action_list);
list_splice_init(&fence->seq_passed_actions,
&action_list);
vmw_fences_perform_actions(fman, &action_list);
wake_up_all(&fence->queue);
}
}
if (!list_empty(&fman->cleanup_list))
(void) schedule_work(&fman->work);
spin_unlock_irqrestore(&fman->lock, flags);
}
bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence,
uint32_t flags)
{
struct vmw_fence_manager *fman = fence->fman;
unsigned long irq_flags;
uint32_t signaled;
spin_lock_irqsave(&fman->lock, irq_flags);
signaled = fence->signaled;
spin_unlock_irqrestore(&fman->lock, irq_flags);
flags &= fence->signal_mask;
if ((signaled & flags) == flags)
return 1;
if ((signaled & DRM_VMW_FENCE_FLAG_EXEC) == 0) {
struct vmw_private *dev_priv = fman->dev_priv;
__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
u32 seqno;
seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
vmw_fences_update(fman, seqno);
}
spin_lock_irqsave(&fman->lock, irq_flags);
signaled = fence->signaled;
spin_unlock_irqrestore(&fman->lock, irq_flags);
return ((signaled & flags) == flags);
}
int vmw_fence_obj_wait(struct vmw_fence_obj *fence,
uint32_t flags, bool lazy,
bool interruptible, unsigned long timeout)
{
struct vmw_private *dev_priv = fence->fman->dev_priv;
long ret;
if (likely(vmw_fence_obj_signaled(fence, flags)))
return 0;
vmw_fifo_ping_host(dev_priv, SVGA_SYNC_GENERIC);
vmw_seqno_waiter_add(dev_priv);
if (interruptible)
ret = wait_event_interruptible_timeout
(fence->queue,
vmw_fence_obj_signaled(fence, flags),
timeout);
else
ret = wait_event_timeout
(fence->queue,
vmw_fence_obj_signaled(fence, flags),
timeout);
vmw_seqno_waiter_remove(dev_priv);
if (unlikely(ret == 0))
ret = -EBUSY;
else if (likely(ret > 0))
ret = 0;
return ret;
}
void vmw_fence_obj_flush(struct vmw_fence_obj *fence)
{
struct vmw_private *dev_priv = fence->fman->dev_priv;
vmw_fifo_ping_host(dev_priv, SVGA_SYNC_GENERIC);
}
static void vmw_fence_destroy(struct vmw_fence_obj *fence)
{
struct vmw_fence_manager *fman = fence->fman;
kfree(fence);
/*
* Free kernel space accounting.
*/
ttm_mem_global_free(vmw_mem_glob(fman->dev_priv),
fman->fence_size);
}
int vmw_fence_create(struct vmw_fence_manager *fman,
uint32_t seqno,
uint32_t mask,
struct vmw_fence_obj **p_fence)
{
struct ttm_mem_global *mem_glob = vmw_mem_glob(fman->dev_priv);
struct vmw_fence_obj *fence;
int ret;
ret = ttm_mem_global_alloc(mem_glob, fman->fence_size,
false, false);
if (unlikely(ret != 0))
return ret;
fence = kzalloc(sizeof(*fence), GFP_KERNEL);
if (unlikely(fence == NULL)) {
ret = -ENOMEM;
goto out_no_object;
}
ret = vmw_fence_obj_init(fman, fence, seqno, mask,
vmw_fence_destroy);
if (unlikely(ret != 0))
goto out_err_init;
*p_fence = fence;
return 0;
out_err_init:
kfree(fence);
out_no_object:
ttm_mem_global_free(mem_glob, fman->fence_size);
return ret;
}
static void vmw_user_fence_destroy(struct vmw_fence_obj *fence)
{
struct vmw_user_fence *ufence =
container_of(fence, struct vmw_user_fence, fence);
struct vmw_fence_manager *fman = fence->fman;
kfree(ufence);
/*
* Free kernel space accounting.
*/
ttm_mem_global_free(vmw_mem_glob(fman->dev_priv),
fman->user_fence_size);
}
static void vmw_user_fence_base_release(struct ttm_base_object **p_base)
{
struct ttm_base_object *base = *p_base;
struct vmw_user_fence *ufence =
container_of(base, struct vmw_user_fence, base);
struct vmw_fence_obj *fence = &ufence->fence;
*p_base = NULL;
vmw_fence_obj_unreference(&fence);
}
int vmw_user_fence_create(struct drm_file *file_priv,
struct vmw_fence_manager *fman,
uint32_t seqno,
uint32_t mask,
struct vmw_fence_obj **p_fence,
uint32_t *p_handle)
{
struct ttm_object_file *tfile = vmw_fpriv(file_priv)->tfile;
struct vmw_user_fence *ufence;
struct vmw_fence_obj *tmp;
struct ttm_mem_global *mem_glob = vmw_mem_glob(fman->dev_priv);
int ret;
/*
* Kernel memory space accounting, since this object may
* be created by a user-space request.
*/
ret = ttm_mem_global_alloc(mem_glob, fman->user_fence_size,
false, false);
if (unlikely(ret != 0))
return ret;
ufence = kzalloc(sizeof(*ufence), GFP_KERNEL);
if (unlikely(ufence == NULL)) {
ret = -ENOMEM;
goto out_no_object;
}
ret = vmw_fence_obj_init(fman, &ufence->fence, seqno,
mask, vmw_user_fence_destroy);
if (unlikely(ret != 0)) {
kfree(ufence);
goto out_no_object;
}
/*
* The base object holds a reference which is freed in
* vmw_user_fence_base_release.
*/
tmp = vmw_fence_obj_reference(&ufence->fence);
ret = ttm_base_object_init(tfile, &ufence->base, false,
VMW_RES_FENCE,
&vmw_user_fence_base_release, NULL);
if (unlikely(ret != 0)) {
/*
* Free the base object's reference
*/
vmw_fence_obj_unreference(&tmp);
goto out_err;
}
*p_fence = &ufence->fence;
*p_handle = ufence->base.hash.key;
return 0;
out_err:
tmp = &ufence->fence;
vmw_fence_obj_unreference(&tmp);
out_no_object:
ttm_mem_global_free(mem_glob, fman->user_fence_size);
return ret;
}
/**
* vmw_fence_fifo_down - signal all unsignaled fence objects.
*/
void vmw_fence_fifo_down(struct vmw_fence_manager *fman)
{
unsigned long irq_flags;
struct list_head action_list;
int ret;
/*
* The list may be altered while we traverse it, so always
* restart when we've released the fman->lock.
*/
spin_lock_irqsave(&fman->lock, irq_flags);
fman->fifo_down = true;
while (!list_empty(&fman->fence_list)) {
struct vmw_fence_obj *fence =
list_entry(fman->fence_list.prev, struct vmw_fence_obj,
head);
kref_get(&fence->kref);
spin_unlock_irq(&fman->lock);
ret = vmw_fence_obj_wait(fence, fence->signal_mask,
false, false,
VMW_FENCE_WAIT_TIMEOUT);
if (unlikely(ret != 0)) {
list_del_init(&fence->head);
fence->signaled |= DRM_VMW_FENCE_FLAG_EXEC;
INIT_LIST_HEAD(&action_list);
list_splice_init(&fence->seq_passed_actions,
&action_list);
vmw_fences_perform_actions(fman, &action_list);
wake_up_all(&fence->queue);
}
spin_lock_irq(&fman->lock);
BUG_ON(!list_empty(&fence->head));
kref_put(&fence->kref, vmw_fence_obj_destroy_locked);
}
spin_unlock_irqrestore(&fman->lock, irq_flags);
}
void vmw_fence_fifo_up(struct vmw_fence_manager *fman)
{
unsigned long irq_flags;
spin_lock_irqsave(&fman->lock, irq_flags);
fman->fifo_down = false;
spin_unlock_irqrestore(&fman->lock, irq_flags);
}
int vmw_fence_obj_wait_ioctl(struct drm_device *dev, void *data,
struct drm_file *file_priv)
{
struct drm_vmw_fence_wait_arg *arg =
(struct drm_vmw_fence_wait_arg *)data;
unsigned long timeout;
struct ttm_base_object *base;
struct vmw_fence_obj *fence;
struct ttm_object_file *tfile = vmw_fpriv(file_priv)->tfile;
int ret;
uint64_t wait_timeout = ((uint64_t)arg->timeout_us * HZ);
/*
* 64-bit division not present on 32-bit systems, so do an
* approximation. (Divide by 1000000).
*/
wait_timeout = (wait_timeout >> 20) + (wait_timeout >> 24) -
(wait_timeout >> 26);
if (!arg->cookie_valid) {
arg->cookie_valid = 1;
arg->kernel_cookie = jiffies + wait_timeout;
}
base = ttm_base_object_lookup(tfile, arg->handle);
if (unlikely(base == NULL)) {
printk(KERN_ERR "Wait invalid fence object handle "
"0x%08lx.\n",
(unsigned long)arg->handle);
return -EINVAL;
}
fence = &(container_of(base, struct vmw_user_fence, base)->fence);
timeout = jiffies;
if (time_after_eq(timeout, (unsigned long)arg->kernel_cookie)) {
ret = ((vmw_fence_obj_signaled(fence, arg->flags)) ?
0 : -EBUSY);
goto out;
}
timeout = (unsigned long)arg->kernel_cookie - timeout;
ret = vmw_fence_obj_wait(fence, arg->flags, arg->lazy, true, timeout);
out:
ttm_base_object_unref(&base);
/*
* Optionally unref the fence object.
*/
if (ret == 0 && (arg->wait_options & DRM_VMW_WAIT_OPTION_UNREF))
return ttm_ref_object_base_unref(tfile, arg->handle,
TTM_REF_USAGE);
return ret;
}
int vmw_fence_obj_signaled_ioctl(struct drm_device *dev, void *data,
struct drm_file *file_priv)
{
struct drm_vmw_fence_signaled_arg *arg =
(struct drm_vmw_fence_signaled_arg *) data;
struct ttm_base_object *base;
struct vmw_fence_obj *fence;
struct vmw_fence_manager *fman;
struct ttm_object_file *tfile = vmw_fpriv(file_priv)->tfile;
struct vmw_private *dev_priv = vmw_priv(dev);
base = ttm_base_object_lookup(tfile, arg->handle);
if (unlikely(base == NULL)) {
printk(KERN_ERR "Fence signaled invalid fence object handle "
"0x%08lx.\n",
(unsigned long)arg->handle);
return -EINVAL;
}
fence = &(container_of(base, struct vmw_user_fence, base)->fence);
fman = fence->fman;
arg->signaled = vmw_fence_obj_signaled(fence, arg->flags);
spin_lock_irq(&fman->lock);
arg->signaled_flags = fence->signaled;
arg->passed_seqno = dev_priv->last_read_seqno;
spin_unlock_irq(&fman->lock);
ttm_base_object_unref(&base);
return 0;
}
int vmw_fence_obj_unref_ioctl(struct drm_device *dev, void *data,
struct drm_file *file_priv)
{
struct drm_vmw_fence_arg *arg =
(struct drm_vmw_fence_arg *) data;
return ttm_ref_object_base_unref(vmw_fpriv(file_priv)->tfile,
arg->handle,
TTM_REF_USAGE);
}

105
drivers/gpu/drm/vmwgfx/vmwgfx_fence.h Normal file
View File

@ -0,0 +1,105 @@
/**************************************************************************
*
* Copyright © 2011 VMware, Inc., Palo Alto, CA., USA
* All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the
* "Software"), to deal in the Software without restriction, including
* without limitation the rights to use, copy, modify, merge, publish,
* distribute, sub license, and/or sell copies of the Software, and to
* permit persons to whom the Software is furnished to do so, subject to
* the following conditions:
*
* The above copyright notice and this permission notice (including the
* next paragraph) shall be included in all copies or substantial portions
* of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
* THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
* USE OR OTHER DEALINGS IN THE SOFTWARE.
*
**************************************************************************/
#ifndef _VMWGFX_FENCE_H_
#define VMW_FENCE_WAIT_TIMEOUT (5*HZ)
struct vmw_private;
struct vmw_fence_manager;
/**
*
*
*/
struct vmw_fence_action {
struct list_head head;
void (*seq_passed) (struct vmw_fence_action *action);
void (*cleanup) (struct vmw_fence_action *action);
};
struct vmw_fence_obj {
struct kref kref;
u32 seqno;
struct vmw_fence_manager *fman;
struct list_head head;
uint32_t signaled;
uint32_t signal_mask;
struct list_head seq_passed_actions;
void (*destroy)(struct vmw_fence_obj *fence);
wait_queue_head_t queue;
};
extern struct vmw_fence_manager *
vmw_fence_manager_init(struct vmw_private *dev_priv);
extern void vmw_fence_manager_takedown(struct vmw_fence_manager *fman);
extern void vmw_fence_obj_unreference(struct vmw_fence_obj **fence_p);
extern struct vmw_fence_obj *
vmw_fence_obj_reference(struct vmw_fence_obj *fence);
extern void vmw_fences_update(struct vmw_fence_manager *fman,
u32 sequence);
extern bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence,
uint32_t flags);
extern int vmw_fence_obj_wait(struct vmw_fence_obj *fence, uint32_t flags,
bool lazy,
bool interruptible, unsigned long timeout);
extern void vmw_fence_obj_flush(struct vmw_fence_obj *fence);
extern int vmw_fence_create(struct vmw_fence_manager *fman,
uint32_t seqno,
uint32_t mask,
struct vmw_fence_obj **p_fence);
extern int vmw_user_fence_create(struct drm_file *file_priv,
struct vmw_fence_manager *fman,
uint32_t sequence,
uint32_t mask,
struct vmw_fence_obj **p_fence,
uint32_t *p_handle);
extern void vmw_fence_fifo_up(struct vmw_fence_manager *fman);
extern void vmw_fence_fifo_down(struct vmw_fence_manager *fman);
extern int vmw_fence_obj_wait_ioctl(struct drm_device *dev, void *data,
struct drm_file *file_priv);
extern int vmw_fence_obj_signaled_ioctl(struct drm_device *dev, void *data,
struct drm_file *file_priv);
extern int vmw_fence_obj_unref_ioctl(struct drm_device *dev, void *data,
struct drm_file *file_priv);
#endif /* _VMWGFX_FENCE_H_ */

35
drivers/gpu/drm/vmwgfx/vmwgfx_irq.c
View File

@ -40,8 +40,13 @@ irqreturn_t vmw_irq_handler(DRM_IRQ_ARGS)
status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
spin_unlock(&dev_priv->irq_lock);
if (status & SVGA_IRQFLAG_ANY_FENCE)
if (status & SVGA_IRQFLAG_ANY_FENCE) {
__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
uint32_t seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
vmw_fences_update(dev_priv->fman, seqno);
wake_up_all(&dev_priv->fence_queue);
}
if (status & SVGA_IRQFLAG_FIFO_PROGRESS)
wake_up_all(&dev_priv->fifo_queue);
@ -68,12 +73,12 @@ void vmw_update_seqno(struct vmw_private *dev_priv,
struct vmw_fifo_state *fifo_state)
{
__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
uint32_t seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
if (dev_priv->last_read_seqno != seqno) {
dev_priv->last_read_seqno = seqno;
vmw_marker_pull(&fifo_state->marker_queue, seqno);
vmw_fences_update(dev_priv->fman, seqno);
}
}
@ -175,7 +180,7 @@ int vmw_fallback_wait(struct vmw_private *dev_priv,
return ret;
}
static void vmw_seqno_waiter_add(struct vmw_private *dev_priv)
void vmw_seqno_waiter_add(struct vmw_private *dev_priv)
{
mutex_lock(&dev_priv->hw_mutex);
if (dev_priv->fence_queue_waiters++ == 0) {
@ -192,7 +197,7 @@ static void vmw_seqno_waiter_add(struct vmw_private *dev_priv)
mutex_unlock(&dev_priv->hw_mutex);
}
static void vmw_seqno_waiter_remove(struct vmw_private *dev_priv)
void vmw_seqno_waiter_remove(struct vmw_private *dev_priv)
{
mutex_lock(&dev_priv->hw_mutex);
if (--dev_priv->fence_queue_waiters == 0) {
@ -286,25 +291,3 @@ void vmw_irq_uninstall(struct drm_device *dev)
status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
}
#define VMW_FENCE_WAIT_TIMEOUT 3*HZ;
int vmw_fence_wait_ioctl(struct drm_device *dev, void *data,
struct drm_file *file_priv)
{
struct drm_vmw_fence_wait_arg *arg =
(struct drm_vmw_fence_wait_arg *)data;
unsigned long timeout;
if (!arg->cookie_valid) {
arg->cookie_valid = 1;
arg->kernel_cookie = jiffies + VMW_FENCE_WAIT_TIMEOUT;
}
timeout = jiffies;
if (time_after_eq(timeout, (unsigned long)arg->kernel_cookie))
return -EBUSY;
timeout = (unsigned long)arg->kernel_cookie - timeout;
return vmw_wait_seqno(vmw_priv(dev), true, arg->seqno, true, timeout);
}

4
drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
View File

@ -31,10 +31,6 @@
#include "ttm/ttm_placement.h"
#include "drmP.h"
#define VMW_RES_CONTEXT ttm_driver_type0
#define VMW_RES_SURFACE ttm_driver_type1
#define VMW_RES_STREAM ttm_driver_type2
struct vmw_user_context {
struct ttm_base_object base;
struct vmw_resource res;

149
include/drm/vmwgfx_drm.h
View File

@ -48,8 +48,12 @@
#define DRM_VMW_UNREF_SURFACE 10
#define DRM_VMW_REF_SURFACE 11
#define DRM_VMW_EXECBUF 12
#define DRM_VMW_FENCE_WAIT 13
#define DRM_VMW_GET_3D_CAP 14
#define DRM_VMW_GET_3D_CAP 13
#define DRM_VMW_FENCE_WAIT 14
#define DRM_VMW_FENCE_SIGNALED 15
#define DRM_VMW_FENCE_UNREF 16
#define DRM_VMW_FENCE_EVENT 17
/*************************************************************************/
/**
@ -318,14 +322,23 @@ struct drm_vmw_execbuf_arg {
uint32_t command_size;
uint32_t throttle_us;
uint64_t fence_rep;
uint32_t version;
uint32_t flags;
uint32_t version;
uint32_t flags;
};
/**
* struct drm_vmw_fence_rep
*
* @fence_seq: Fence seqno associated with a command submission.
* @handle: Fence object handle for fence associated with a command submission.
* @mask: Fence flags relevant for this fence object.
* @seqno: Fence sequence number in fifo. A fence object with a lower
* seqno will signal the EXEC flag before a fence object with a higher
* seqno. This can be used by user-space to avoid kernel calls to determine
* whether a fence has signaled the EXEC flag. Note that @seqno will
* wrap at 32-bit.
* @passed_seqno: The highest seqno number processed by the hardware
* so far. This can be used to mark user-space fence objects as signaled, and
* to determine whether a fence seqno might be stale.
* @error: This member should've been set to -EFAULT on submission.
* The following actions should be take on completion:
* error == -EFAULT: Fence communication failed. The host is synchronized.
@ -339,9 +352,12 @@ struct drm_vmw_execbuf_arg {
*/
struct drm_vmw_fence_rep {
uint64_t fence_seq;
int32_t error;
uint32_t handle;
uint32_t mask;
uint32_t seqno;
uint32_t passed_seqno;
uint32_t pad64;
int32_t error;
};
/*************************************************************************/
@ -430,14 +446,6 @@ struct drm_vmw_unref_dmabuf_arg {
uint32_t pad64;
};
struct drm_vmw_fence_wait_arg {
uint64_t seqno;
uint64_t kernel_cookie;
int32_t cookie_valid;
int32_t pad64;
};
/*************************************************************************/
/**
* DRM_VMW_CONTROL_STREAM - Control overlays, aka streams.
@ -559,6 +567,7 @@ struct drm_vmw_stream_arg {
* Return a single stream that was claimed by this process. Also makes
* sure that the stream has been stopped.
*/
/*************************************************************************/
/**
* DRM_VMW_GET_3D_CAP
@ -607,4 +616,114 @@ struct drm_vmw_update_layout_arg {
uint64_t rects;
};
/*************************************************************************/
/**
* DRM_VMW_FENCE_WAIT
*
* Waits for a fence object to signal. The wait is interruptible, so that
* signals may be delivered during the interrupt. The wait may timeout,
* in which case the calls returns -EBUSY. If the wait is restarted,
* that is restarting without resetting @cookie_valid to zero,
* the timeout is computed from the first call.
*
* The flags argument to the DRM_VMW_FENCE_WAIT ioctl indicates what to wait
* on:
* DRM_VMW_FENCE_FLAG_EXEC: All commands ahead of the fence in the command
* stream
* have executed.
* DRM_VMW_FENCE_FLAG_QUERY: All query results resulting from query finish
* commands
* in the buffer given to the EXECBUF ioctl returning the fence object handle
* are available to user-space.
*
* DRM_VMW_WAIT_OPTION_UNREF: If this wait option is given, and the
* fenc wait ioctl returns 0, the fence object has been unreferenced after
* the wait.
*/
#define DRM_VMW_FENCE_FLAG_EXEC (1 << 0)
#define DRM_VMW_FENCE_FLAG_QUERY (1 << 1)
#define DRM_VMW_WAIT_OPTION_UNREF (1 << 0)
/**
* struct drm_vmw_fence_wait_arg
*
* @handle: Fence object handle as returned by the DRM_VMW_EXECBUF ioctl.
* @cookie_valid: Must be reset to 0 on first call. Left alone on restart.
* @kernel_cookie: Set to 0 on first call. Left alone on restart.
* @timeout_us: Wait timeout in microseconds. 0 for indefinite timeout.
* @lazy: Set to 1 if timing is not critical. Allow more than a kernel tick
* before returning.
* @flags: Fence flags to wait on.
* @wait_options: Options that control the behaviour of the wait ioctl.
*
* Input argument to the DRM_VMW_FENCE_WAIT ioctl.
*/
struct drm_vmw_fence_wait_arg {
uint32_t handle;
int32_t cookie_valid;
uint64_t kernel_cookie;
uint64_t timeout_us;
int32_t lazy;
int32_t flags;
int32_t wait_options;
int32_t pad64;
};
/*************************************************************************/
/**
* DRM_VMW_FENCE_SIGNALED
*
* Checks if a fence object is signaled..
*/
/**
* struct drm_vmw_fence_signaled_arg
*
* @handle: Fence object handle as returned by the DRM_VMW_EXECBUF ioctl.
* @flags: Fence object flags input to DRM_VMW_FENCE_SIGNALED ioctl
* @signaled: Out: Flags signaled.
* @sequence: Out: Highest sequence passed so far. Can be used to signal the
* EXEC flag of user-space fence objects.
*
* Input/Output argument to the DRM_VMW_FENCE_SIGNALED and DRM_VMW_FENCE_UNREF
* ioctls.
*/
struct drm_vmw_fence_signaled_arg {
uint32_t handle;
uint32_t flags;
int32_t signaled;
uint32_t passed_seqno;
uint32_t signaled_flags;
uint32_t pad64;
};
/*************************************************************************/
/**
* DRM_VMW_FENCE_UNREF
*
* Unreferences a fence object, and causes it to be destroyed if there are no
* other references to it.
*
*/
/**
* struct drm_vmw_fence_arg
*
* @handle: Fence object handle as returned by the DRM_VMW_EXECBUF ioctl.
*
* Input/Output argument to the DRM_VMW_FENCE_UNREF ioctl..
*/
struct drm_vmw_fence_arg {
uint32_t handle;
uint32_t pad64;
};
#endif