From d6ea88865d3e5b0c62040531310c1f2c6a994f46 Mon Sep 17 00:00:00 2001 From: Dave Airlie Date: Mon, 22 Nov 2010 13:24:40 +1000 Subject: drm/ttm: Add a bo list reserve fastpath (v2) Makes it possible to reserve a list of buffer objects with a single spin lock / unlock if there is no contention. Should improve cpu usage on SMP kernels. v2: Initialize private list members on reserve and don't call ttm_bo_list_ref_sub() with zero put_count. Signed-off-by: Thomas Hellstrom Signed-off-by: Dave Airlie --- include/drm/ttm/ttm_bo_api.h | 38 ++++++++++++++++++++++++++++++++++++++ include/drm/ttm/ttm_bo_driver.h | 14 ++++++++++++++ include/drm/ttm/ttm_execbuf_util.h | 6 +++++- 3 files changed, 57 insertions(+), 1 deletion(-) (limited to 'include') diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h index beafc156a535..b0fc9c12554b 100644 --- a/include/drm/ttm/ttm_bo_api.h +++ b/include/drm/ttm/ttm_bo_api.h @@ -364,6 +364,44 @@ extern int ttm_bo_validate(struct ttm_buffer_object *bo, */ extern void ttm_bo_unref(struct ttm_buffer_object **bo); + +/** + * ttm_bo_list_ref_sub + * + * @bo: The buffer object. + * @count: The number of references with which to decrease @bo::list_kref; + * @never_free: The refcount should not reach zero with this operation. + * + * Release @count lru list references to this buffer object. + */ +extern void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count, + bool never_free); + +/** + * ttm_bo_add_to_lru + * + * @bo: The buffer object. + * + * Add this bo to the relevant mem type lru and, if it's backed by + * system pages (ttms) to the swap list. + * This function must be called with struct ttm_bo_global::lru_lock held, and + * is typically called immediately prior to unreserving a bo. + */ +extern void ttm_bo_add_to_lru(struct ttm_buffer_object *bo); + +/** + * ttm_bo_del_from_lru + * + * @bo: The buffer object. + * + * Remove this bo from all lru lists used to lookup and reserve an object. + * This function must be called with struct ttm_bo_global::lru_lock held, + * and is usually called just immediately after the bo has been reserved to + * avoid recursive reservation from lru lists. + */ +extern int ttm_bo_del_from_lru(struct ttm_buffer_object *bo); + + /** * ttm_bo_lock_delayed_workqueue * diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h index 8e0c848326b6..95068e6024db 100644 --- a/include/drm/ttm/ttm_bo_driver.h +++ b/include/drm/ttm/ttm_bo_driver.h @@ -864,6 +864,20 @@ extern int ttm_bo_reserve(struct ttm_buffer_object *bo, bool interruptible, bool no_wait, bool use_sequence, uint32_t sequence); + +/** + * ttm_bo_reserve_locked: + * + * Similar to ttm_bo_reserve, but must be called with the glob::lru_lock + * spinlock held, and will not remove reserved buffers from the lru lists. + * The function may release the LRU spinlock if it needs to sleep. + */ + +extern int ttm_bo_reserve_locked(struct ttm_buffer_object *bo, + bool interruptible, + bool no_wait, bool use_sequence, + uint32_t sequence); + /** * ttm_bo_unreserve * diff --git a/include/drm/ttm/ttm_execbuf_util.h b/include/drm/ttm/ttm_execbuf_util.h index cd2c475da9ea..fd09b8438977 100644 --- a/include/drm/ttm/ttm_execbuf_util.h +++ b/include/drm/ttm/ttm_execbuf_util.h @@ -41,7 +41,9 @@ * @bo: refcounted buffer object pointer. * @new_sync_obj_arg: New sync_obj_arg for @bo, to be used once * adding a new sync object. - * @reservied: Indicates whether @bo has been reserved for validation. + * @reserved: Indicates whether @bo has been reserved for validation. + * @removed: Indicates whether @bo has been removed from lru lists. + * @put_count: Number of outstanding references on bo::list_kref. */ struct ttm_validate_buffer { @@ -49,6 +51,8 @@ struct ttm_validate_buffer { struct ttm_buffer_object *bo; void *new_sync_obj_arg; bool reserved; + bool removed; + int put_count; }; /** -- cgit v1.2.3 From ecf7ace9a8450303a987aa8364e53860cd50e554 Mon Sep 17 00:00:00 2001 From: Thomas Hellstrom Date: Tue, 16 Nov 2010 15:21:07 +0100 Subject: kref: Add a kref_sub function Makes it possible to optimize batched multiple unrefs. Initial user will be drivers/gpu/ttm which accumulates unrefs to be processed outside of atomic code. Signed-off-by: Thomas Hellstrom Signed-off-by: Dave Airlie --- include/linux/kref.h | 2 ++ 1 file changed, 2 insertions(+) (limited to 'include') diff --git a/include/linux/kref.h b/include/linux/kref.h index 6cc38fc07ab7..d4a62ab2ee5e 100644 --- a/include/linux/kref.h +++ b/include/linux/kref.h @@ -24,5 +24,7 @@ struct kref { void kref_init(struct kref *kref); void kref_get(struct kref *kref); int kref_put(struct kref *kref, void (*release) (struct kref *kref)); +int kref_sub(struct kref *kref, unsigned int count, + void (*release) (struct kref *kref)); #endif /* _KREF_H_ */ -- cgit v1.2.3 From 96726fe50feae74812a2ccf5d5da23cb01c0a413 Mon Sep 17 00:00:00 2001 From: Thomas Hellstrom Date: Wed, 17 Nov 2010 12:28:28 +0000 Subject: drm/ttm: Don't deadlock on recursive multi-bo reservations Add an aid for the driver to detect deadlocks on multi-bo reservations Update documentation. Signed-off-by: Thomas Hellstrom Reviewed-by: Jerome Glisse Signed-off-by: Dave Airlie --- include/drm/ttm/ttm_bo_driver.h | 25 ++++++++++++++++++++++--- 1 file changed, 22 insertions(+), 3 deletions(-) (limited to 'include') diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h index 95068e6024db..1e25a40c688e 100644 --- a/include/drm/ttm/ttm_bo_driver.h +++ b/include/drm/ttm/ttm_bo_driver.h @@ -859,6 +859,9 @@ extern void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo); * try again. (only if use_sequence == 1). * -ERESTARTSYS: A wait for the buffer to become unreserved was interrupted by * a signal. Release all buffer reservations and return to user-space. + * -EBUSY: The function needed to sleep, but @no_wait was true + * -EDEADLK: Bo already reserved using @sequence. This error code will only + * be returned if @use_sequence is set to true. */ extern int ttm_bo_reserve(struct ttm_buffer_object *bo, bool interruptible, @@ -868,11 +871,27 @@ extern int ttm_bo_reserve(struct ttm_buffer_object *bo, /** * ttm_bo_reserve_locked: * - * Similar to ttm_bo_reserve, but must be called with the glob::lru_lock - * spinlock held, and will not remove reserved buffers from the lru lists. + * @bo: A pointer to a struct ttm_buffer_object. + * @interruptible: Sleep interruptible if waiting. + * @no_wait: Don't sleep while trying to reserve, rather return -EBUSY. + * @use_sequence: If @bo is already reserved, Only sleep waiting for + * it to become unreserved if @sequence < (@bo)->sequence. + * + * Must be called with struct ttm_bo_global::lru_lock held, + * and will not remove reserved buffers from the lru lists. * The function may release the LRU spinlock if it needs to sleep. + * Otherwise identical to ttm_bo_reserve. + * + * Returns: + * -EAGAIN: The reservation may cause a deadlock. + * Release all buffer reservations, wait for @bo to become unreserved and + * try again. (only if use_sequence == 1). + * -ERESTARTSYS: A wait for the buffer to become unreserved was interrupted by + * a signal. Release all buffer reservations and return to user-space. + * -EBUSY: The function needed to sleep, but @no_wait was true + * -EDEADLK: Bo already reserved using @sequence. This error code will only + * be returned if @use_sequence is set to true. */ - extern int ttm_bo_reserve_locked(struct ttm_buffer_object *bo, bool interruptible, bool no_wait, bool use_sequence, -- cgit v1.2.3 From 702adba22433c175e8429a47760f35ca16caf1cd Mon Sep 17 00:00:00 2001 From: Thomas Hellstrom Date: Wed, 17 Nov 2010 12:28:29 +0000 Subject: drm/ttm/radeon/nouveau: Kill the bo lock in favour of a bo device fence_lock The bo lock used only to protect the bo sync object members, and since it is a per bo lock, fencing a buffer list will see a lot of locks and unlocks. Replace it with a per-device lock that protects the sync object members on *all* bos. Reading and setting these members will always be very quick, so the risc of heavy lock contention is microscopic. Note that waiting for sync objects will always take place outside of this lock. The bo device fence lock will eventually be replaced with a seqlock / rcu mechanism so we can determine that a bo is idle under a rcu / read seqlock. However this change will allow us to batch fencing and unreserving of buffers with a minimal amount of locking. Signed-off-by: Thomas Hellstrom Reviewed-by: Jerome Glisse Signed-off-by: Dave Airlie --- include/drm/ttm/ttm_bo_api.h | 6 ++---- include/drm/ttm/ttm_bo_driver.h | 3 +++ 2 files changed, 5 insertions(+), 4 deletions(-) (limited to 'include') diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h index b0fc9c12554b..edacd483c59c 100644 --- a/include/drm/ttm/ttm_bo_api.h +++ b/include/drm/ttm/ttm_bo_api.h @@ -154,7 +154,6 @@ struct ttm_tt; * keeps one refcount. When this refcount reaches zero, * the object is destroyed. * @event_queue: Queue for processes waiting on buffer object status change. - * @lock: spinlock protecting mostly synchronization members. * @mem: structure describing current placement. * @persistant_swap_storage: Usually the swap storage is deleted for buffers * pinned in physical memory. If this behaviour is not desired, this member @@ -213,7 +212,6 @@ struct ttm_buffer_object { struct kref kref; struct kref list_kref; wait_queue_head_t event_queue; - spinlock_t lock; /** * Members protected by the bo::reserved lock. @@ -248,10 +246,10 @@ struct ttm_buffer_object { atomic_t reserved; /** - * Members protected by the bo::lock + * Members protected by struct buffer_object_device::fence_lock * In addition, setting sync_obj to anything else * than NULL requires bo::reserved to be held. This allows for - * checking NULL while reserved but not holding bo::lock. + * checking NULL while reserved but not holding the mentioned lock. */ void *sync_obj_arg; diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h index 1e25a40c688e..ca8131e98300 100644 --- a/include/drm/ttm/ttm_bo_driver.h +++ b/include/drm/ttm/ttm_bo_driver.h @@ -510,6 +510,8 @@ struct ttm_bo_global { * * @driver: Pointer to a struct ttm_bo_driver struct setup by the driver. * @man: An array of mem_type_managers. + * @fence_lock: Protects the synchronizing members on *all* bos belonging + * to this device. * @addr_space_mm: Range manager for the device address space. * lru_lock: Spinlock that protects the buffer+device lru lists and * ddestroy lists. @@ -531,6 +533,7 @@ struct ttm_bo_device { struct ttm_bo_driver *driver; rwlock_t vm_lock; struct ttm_mem_type_manager man[TTM_NUM_MEM_TYPES]; + spinlock_t fence_lock; /* * Protected by the vm lock. */ -- cgit v1.2.3 From 95762c2b34069bf4adb7929969f1f5f5fc8a38df Mon Sep 17 00:00:00 2001 From: Thomas Hellstrom Date: Wed, 17 Nov 2010 12:28:30 +0000 Subject: drm/ttm: Improved fencing of buffer object lists Drastically reduce the number of spin lock / unlock operations by performing unreserving and fencing under global locks. Signed-off-by: Thomas Hellstrom Reviewed-by: Jerome Glisse Signed-off-by: Dave Airlie --- include/drm/ttm/ttm_bo_driver.h | 10 ++++++++++ include/drm/ttm/ttm_execbuf_util.h | 2 ++ 2 files changed, 12 insertions(+) (limited to 'include') diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h index ca8131e98300..cfb9ca4ec1c4 100644 --- a/include/drm/ttm/ttm_bo_driver.h +++ b/include/drm/ttm/ttm_bo_driver.h @@ -909,6 +909,16 @@ extern int ttm_bo_reserve_locked(struct ttm_buffer_object *bo, */ extern void ttm_bo_unreserve(struct ttm_buffer_object *bo); +/** + * ttm_bo_unreserve_locked + * + * @bo: A pointer to a struct ttm_buffer_object. + * + * Unreserve a previous reservation of @bo. + * Needs to be called with struct ttm_bo_global::lru_lock held. + */ +extern void ttm_bo_unreserve_locked(struct ttm_buffer_object *bo); + /** * ttm_bo_wait_unreserved * diff --git a/include/drm/ttm/ttm_execbuf_util.h b/include/drm/ttm/ttm_execbuf_util.h index fd09b8438977..535ab00407e0 100644 --- a/include/drm/ttm/ttm_execbuf_util.h +++ b/include/drm/ttm/ttm_execbuf_util.h @@ -44,6 +44,7 @@ * @reserved: Indicates whether @bo has been reserved for validation. * @removed: Indicates whether @bo has been removed from lru lists. * @put_count: Number of outstanding references on bo::list_kref. + * @old_sync_obj: Pointer to a sync object about to be unreferenced */ struct ttm_validate_buffer { @@ -53,6 +54,7 @@ struct ttm_validate_buffer { bool reserved; bool removed; int put_count; + void *old_sync_obj; }; /** -- cgit v1.2.3 From 65705962025df490d13df59ec57c5329d1bd0a16 Mon Sep 17 00:00:00 2001 From: Thomas Hellstrom Date: Wed, 17 Nov 2010 12:28:31 +0000 Subject: drm/ttm/vmwgfx: Have TTM manage the validation sequence. Rather than having the driver supply the validation sequence, leave that responsibility to TTM. This saves some confusion and a function argument. Signed-off-by: Thomas Hellstrom Signed-off-by: Dave Airlie --- include/drm/ttm/ttm_bo_driver.h | 2 ++ include/drm/ttm/ttm_execbuf_util.h | 3 +-- 2 files changed, 3 insertions(+), 2 deletions(-) (limited to 'include') diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h index cfb9ca4ec1c4..e3b2e245db1b 100644 --- a/include/drm/ttm/ttm_bo_driver.h +++ b/include/drm/ttm/ttm_bo_driver.h @@ -515,6 +515,7 @@ struct ttm_bo_global { * @addr_space_mm: Range manager for the device address space. * lru_lock: Spinlock that protects the buffer+device lru lists and * ddestroy lists. + * @val_seq: Current validation sequence. * @nice_mode: Try nicely to wait for buffer idle when cleaning a manager. * If a GPU lockup has been detected, this is forced to 0. * @dev_mapping: A pointer to the struct address_space representing the @@ -544,6 +545,7 @@ struct ttm_bo_device { * Protected by the global:lru lock. */ struct list_head ddestroy; + uint32_t val_seq; /* * Protected by load / firstopen / lastclose /unload sync. diff --git a/include/drm/ttm/ttm_execbuf_util.h b/include/drm/ttm/ttm_execbuf_util.h index 535ab00407e0..26cc7f9ffa41 100644 --- a/include/drm/ttm/ttm_execbuf_util.h +++ b/include/drm/ttm/ttm_execbuf_util.h @@ -72,7 +72,6 @@ extern void ttm_eu_backoff_reservation(struct list_head *list); * function ttm_eu_reserve_buffers * * @list: thread private list of ttm_validate_buffer structs. - * @val_seq: A unique sequence number. * * Tries to reserve bos pointed to by the list entries for validation. * If the function returns 0, all buffers are marked as "unfenced", @@ -94,7 +93,7 @@ extern void ttm_eu_backoff_reservation(struct list_head *list); * has failed. */ -extern int ttm_eu_reserve_buffers(struct list_head *list, uint32_t val_seq); +extern int ttm_eu_reserve_buffers(struct list_head *list); /** * function ttm_eu_fence_buffer_objects. -- cgit v1.2.3 From eba67093f535322cb4f1c4b737319c0907a0c81d Mon Sep 17 00:00:00 2001 From: Thomas Hellstrom Date: Thu, 11 Nov 2010 09:41:57 +0100 Subject: drm/ttm: Fix up io_mem_reserve / io_mem_free calling This patch attempts to fix up shortcomings with the current calling sequences. 1) There's a fastpath where no locking occurs and only io_mem_reserved is called to obtain needed info for mapping. The fastpath is set per memory type manager. 2) If the fastpath is disabled, io_mem_reserve and io_mem_free will be exactly balanced and not called recursively for the same struct ttm_mem_reg. 3) Optionally the driver can choose to enable a per memory type manager LRU eviction mechanism that, when io_mem_reserve returns -EAGAIN will attempt to kill user-space mappings of memory in that manager to free up needed resources Signed-off-by: Thomas Hellstrom Reviewed-by: Ben Skeggs Signed-off-by: Dave Airlie --- include/drm/ttm/ttm_bo_api.h | 6 ++- include/drm/ttm/ttm_bo_driver.h | 104 +++++++++++++++++++++------------------- 2 files changed, 60 insertions(+), 50 deletions(-) (limited to 'include') diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h index edacd483c59c..50852aad260a 100644 --- a/include/drm/ttm/ttm_bo_api.h +++ b/include/drm/ttm/ttm_bo_api.h @@ -74,6 +74,8 @@ struct ttm_placement { * @is_iomem: is this io memory ? * @size: size in byte * @offset: offset from the base address + * @io_reserved_vm: The VM system has a refcount in @io_reserved_count + * @io_reserved_count: Refcounting the numbers of callers to ttm_mem_io_reserve * * Structure indicating the bus placement of an object. */ @@ -83,7 +85,8 @@ struct ttm_bus_placement { unsigned long size; unsigned long offset; bool is_iomem; - bool io_reserved; + bool io_reserved_vm; + uint64_t io_reserved_count; }; @@ -235,6 +238,7 @@ struct ttm_buffer_object { struct list_head lru; struct list_head ddestroy; struct list_head swap; + struct list_head io_reserve_lru; uint32_t val_seq; bool seq_valid; diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h index e3b2e245db1b..1da8af6ac884 100644 --- a/include/drm/ttm/ttm_bo_driver.h +++ b/include/drm/ttm/ttm_bo_driver.h @@ -179,30 +179,6 @@ struct ttm_tt { #define TTM_MEMTYPE_FLAG_MAPPABLE (1 << 1) /* Memory mappable */ #define TTM_MEMTYPE_FLAG_CMA (1 << 3) /* Can't map aperture */ -/** - * struct ttm_mem_type_manager - * - * @has_type: The memory type has been initialized. - * @use_type: The memory type is enabled. - * @flags: TTM_MEMTYPE_XX flags identifying the traits of the memory - * managed by this memory type. - * @gpu_offset: If used, the GPU offset of the first managed page of - * fixed memory or the first managed location in an aperture. - * @size: Size of the managed region. - * @available_caching: A mask of available caching types, TTM_PL_FLAG_XX, - * as defined in ttm_placement_common.h - * @default_caching: The default caching policy used for a buffer object - * placed in this memory type if the user doesn't provide one. - * @manager: The range manager used for this memory type. FIXME: If the aperture - * has a page size different from the underlying system, the granularity - * of this manager should take care of this. But the range allocating code - * in ttm_bo.c needs to be modified for this. - * @lru: The lru list for this memory type. - * - * This structure is used to identify and manage memory types for a device. - * It's set up by the ttm_bo_driver::init_mem_type method. - */ - struct ttm_mem_type_manager; struct ttm_mem_type_manager_func { @@ -287,6 +263,36 @@ struct ttm_mem_type_manager_func { void (*debug)(struct ttm_mem_type_manager *man, const char *prefix); }; +/** + * struct ttm_mem_type_manager + * + * @has_type: The memory type has been initialized. + * @use_type: The memory type is enabled. + * @flags: TTM_MEMTYPE_XX flags identifying the traits of the memory + * managed by this memory type. + * @gpu_offset: If used, the GPU offset of the first managed page of + * fixed memory or the first managed location in an aperture. + * @size: Size of the managed region. + * @available_caching: A mask of available caching types, TTM_PL_FLAG_XX, + * as defined in ttm_placement_common.h + * @default_caching: The default caching policy used for a buffer object + * placed in this memory type if the user doesn't provide one. + * @func: structure pointer implementing the range manager. See above + * @priv: Driver private closure for @func. + * @io_reserve_mutex: Mutex optionally protecting shared io_reserve structures + * @use_io_reserve_lru: Use an lru list to try to unreserve io_mem_regions + * reserved by the TTM vm system. + * @io_reserve_lru: Optional lru list for unreserving io mem regions. + * @io_reserve_fastpath: Only use bdev::driver::io_mem_reserve to obtain + * static information. bdev::driver::io_mem_free is never used. + * @lru: The lru list for this memory type. + * + * This structure is used to identify and manage memory types for a device. + * It's set up by the ttm_bo_driver::init_mem_type method. + */ + + + struct ttm_mem_type_manager { struct ttm_bo_device *bdev; @@ -303,6 +309,15 @@ struct ttm_mem_type_manager { uint32_t default_caching; const struct ttm_mem_type_manager_func *func; void *priv; + struct mutex io_reserve_mutex; + bool use_io_reserve_lru; + bool io_reserve_fastpath; + + /* + * Protected by @io_reserve_mutex: + */ + + struct list_head io_reserve_lru; /* * Protected by the global->lru_lock. @@ -758,31 +773,6 @@ extern void ttm_bo_mem_put_locked(struct ttm_buffer_object *bo, extern int ttm_bo_wait_cpu(struct ttm_buffer_object *bo, bool no_wait); -/** - * ttm_bo_pci_offset - Get the PCI offset for the buffer object memory. - * - * @bo Pointer to a struct ttm_buffer_object. - * @bus_base On return the base of the PCI region - * @bus_offset On return the byte offset into the PCI region - * @bus_size On return the byte size of the buffer object or zero if - * the buffer object memory is not accessible through a PCI region. - * - * Returns: - * -EINVAL if the buffer object is currently not mappable. - * 0 otherwise. - */ - -extern int ttm_bo_pci_offset(struct ttm_bo_device *bdev, - struct ttm_mem_reg *mem, - unsigned long *bus_base, - unsigned long *bus_offset, - unsigned long *bus_size); - -extern int ttm_mem_io_reserve(struct ttm_bo_device *bdev, - struct ttm_mem_reg *mem); -extern void ttm_mem_io_free(struct ttm_bo_device *bdev, - struct ttm_mem_reg *mem); - extern void ttm_bo_global_release(struct drm_global_reference *ref); extern int ttm_bo_global_init(struct drm_global_reference *ref); @@ -814,6 +804,22 @@ extern int ttm_bo_device_init(struct ttm_bo_device *bdev, */ extern void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo); +/** + * ttm_bo_unmap_virtual + * + * @bo: tear down the virtual mappings for this BO + * + * The caller must take ttm_mem_io_lock before calling this function. + */ +extern void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo); + +extern int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo); +extern void ttm_mem_io_free_vm(struct ttm_buffer_object *bo); +extern int ttm_mem_io_lock(struct ttm_mem_type_manager *man, + bool interruptible); +extern void ttm_mem_io_unlock(struct ttm_mem_type_manager *man); + + /** * ttm_bo_reserve: * -- cgit v1.2.3