*
* - HL_RESET_DEVICE_RELEASE
* Set if reset is due to device release
+ *
+ * - HL_RESET_FW
+ * F/W will perform the reset. No need to ask it to reset the device. This is relevant
+ * only when running with secured f/w
*/
#define HL_RESET_HARD (1 << 0)
#define HL_RESET_FROM_RESET_THREAD (1 << 1)
#define HL_RESET_HEARTBEAT (1 << 2)
#define HL_RESET_TDR (1 << 3)
#define HL_RESET_DEVICE_RELEASE (1 << 4)
+#define HL_RESET_FW (1 << 5)
#define HL_MAX_SOBS_PER_MONITOR 8
CS_TYPE_DEFAULT,
CS_TYPE_SIGNAL,
CS_TYPE_WAIT,
- CS_TYPE_COLLECTIVE_WAIT
+ CS_TYPE_COLLECTIVE_WAIT,
+ CS_RESERVE_SIGNALS,
+ CS_UNRESERVE_SIGNALS
};
/*
* @hdev: habanalabs device structure.
* @kref: refcount of this SOB. The SOB will reset once the refcount is zero.
* @sob_id: id of this SOB.
+ * @sob_addr: the sob offset from the base address.
* @q_idx: the H/W queue that uses this SOB.
+ * @need_reset: reset indication set when switching to the other sob.
*/
struct hl_hw_sob {
struct hl_device *hdev;
struct kref kref;
u32 sob_id;
+ u32 sob_addr;
u32 q_idx;
+ bool need_reset;
};
enum hl_collective_mode {
* reserved for the user
* @first_available_cq: first available CQ for the user.
* @user_interrupt_count: number of user interrupts.
+ * @server_type: Server type that the ASIC is currently installed in.
+ * The value is according to enum hl_server_type in uapi file.
* @tpc_enabled_mask: which TPCs are enabled.
* @completion_queues_count: number of completion queues.
* @fw_security_enabled: true if security measures are enabled in firmware,
u16 first_available_user_msix_interrupt;
u16 first_available_cq[HL_MAX_DCORES];
u16 user_interrupt_count;
+ u16 server_type;
u8 tpc_enabled_mask;
u8 completion_queues_count;
u8 fw_security_enabled;
* @completion: fence is implemented using completion
* @refcount: refcount for this fence
* @cs_sequence: sequence of the corresponding command submission
+ * @stream_master_qid_map: streams masters QID bitmap to represent all streams
+ * masters QIDs that multi cs is waiting on
* @error: mark this fence with error
* @timestamp: timestamp upon completion
- *
*/
struct hl_fence {
struct completion completion;
struct kref refcount;
u64 cs_sequence;
+ u32 stream_master_qid_map;
int error;
ktime_t timestamp;
};
/**
* struct hl_cs_compl - command submission completion object.
- * @sob_reset_work: workqueue object to run SOB reset flow.
* @base_fence: hl fence object.
* @lock: spinlock to protect fence.
* @hdev: habanalabs device structure.
* @hw_sob: the H/W SOB used in this signal/wait CS.
+ * @encaps_sig_hdl: encaps signals hanlder.
* @cs_seq: command submission sequence number.
* @type: type of the CS - signal/wait.
* @sob_val: the SOB value that is used in this signal/wait CS.
* @sob_group: the SOB group that is used in this collective wait CS.
+ * @encaps_signals: indication whether it's a completion object of cs with
+ * encaps signals or not.
*/
struct hl_cs_compl {
- struct work_struct sob_reset_work;
struct hl_fence base_fence;
spinlock_t lock;
struct hl_device *hdev;
struct hl_hw_sob *hw_sob;
+ struct hl_cs_encaps_sig_handle *encaps_sig_hdl;
u64 cs_seq;
enum hl_cs_type type;
u16 sob_val;
u16 sob_group;
+ bool encaps_signals;
};
/*
u8 curr_sob_offset;
};
+/**
+ * struct hl_encaps_signals_mgr - describes sync stream encapsulated signals
+ * handlers manager
+ * @lock: protects handles.
+ * @handles: an idr to hold all encapsulated signals handles.
+ */
+struct hl_encaps_signals_mgr {
+ spinlock_t lock;
+ struct idr handles;
+};
+
/**
* struct hl_hw_queue - describes a H/W transport queue.
* @shadow_queue: pointer to a shadow queue that holds pointers to jobs.
u64 region_base;
u64 region_size;
u64 bar_size;
- u32 offset_in_bar;
+ u64 offset_in_bar;
u8 bar_id;
u8 used;
};
* @init_firmware_loader: initialize data for FW loader.
* @init_cpu_scrambler_dram: Enable CPU specific DRAM scrambling
* @state_dump_init: initialize constants required for state dump
+ * @get_sob_addr: get SOB base address offset.
+ * @set_pci_memory_regions: setting properties of PCI memory regions
+ * @get_stream_master_qid_arr: get pointer to stream masters QID array
*/
struct hl_asic_funcs {
int (*early_init)(struct hl_device *hdev);
int (*sw_init)(struct hl_device *hdev);
int (*sw_fini)(struct hl_device *hdev);
int (*hw_init)(struct hl_device *hdev);
- void (*hw_fini)(struct hl_device *hdev, bool hard_reset);
- void (*halt_engines)(struct hl_device *hdev, bool hard_reset);
+ void (*hw_fini)(struct hl_device *hdev, bool hard_reset, bool fw_reset);
+ void (*halt_engines)(struct hl_device *hdev, bool hard_reset, bool fw_reset);
int (*suspend)(struct hl_device *hdev);
int (*resume)(struct hl_device *hdev);
int (*mmap)(struct hl_device *hdev, struct vm_area_struct *vma,
void (*reset_sob_group)(struct hl_device *hdev, u16 sob_group);
void (*set_dma_mask_from_fw)(struct hl_device *hdev);
u64 (*get_device_time)(struct hl_device *hdev);
- void (*collective_wait_init_cs)(struct hl_cs *cs);
+ int (*collective_wait_init_cs)(struct hl_cs *cs);
int (*collective_wait_create_jobs)(struct hl_device *hdev,
- struct hl_ctx *ctx, struct hl_cs *cs, u32 wait_queue_id,
- u32 collective_engine_id);
+ struct hl_ctx *ctx, struct hl_cs *cs,
+ u32 wait_queue_id, u32 collective_engine_id,
+ u32 encaps_signal_offset);
u64 (*scramble_addr)(struct hl_device *hdev, u64 addr);
u64 (*descramble_addr)(struct hl_device *hdev, u64 addr);
void (*ack_protection_bits_errors)(struct hl_device *hdev);
void (*init_firmware_loader)(struct hl_device *hdev);
void (*init_cpu_scrambler_dram)(struct hl_device *hdev);
void (*state_dump_init)(struct hl_device *hdev);
+ u32 (*get_sob_addr)(struct hl_device *hdev, u32 sob_id);
+ void (*set_pci_memory_regions)(struct hl_device *hdev);
+ u32* (*get_stream_master_qid_arr)(void);
};
atomic64_t validation_drop_cnt;
};
-/**
- * struct hl_pending_cb - pending command buffer structure
- * @cb_node: cb node in pending cb list
- * @cb: command buffer to send in next submission
- * @cb_size: command buffer size
- * @hw_queue_id: destination queue id
- */
-struct hl_pending_cb {
- struct list_head cb_node;
- struct hl_cb *cb;
- u32 cb_size;
- u32 hw_queue_id;
-};
-
/**
* struct hl_ctx - user/kernel context.
* @mem_hash: holds mapping from virtual address to virtual memory area
* MMU hash or walking the PGT requires talking this lock.
* @hw_block_list_lock: protects the HW block memory list.
* @debugfs_list: node in debugfs list of contexts.
- * pending_cb_list: list of pending command buffers waiting to be sent upon
- * next user command submission context.
* @hw_block_mem_list: list of HW block virtual mapped addresses.
* @cs_counters: context command submission counters.
* @cb_va_pool: device VA pool for command buffers which are mapped to the
* device's MMU.
+ * @sig_mgr: encaps signals handle manager.
* @cs_sequence: sequence number for CS. Value is assigned to a CS and passed
* to user so user could inquire about CS. It is used as
* index to cs_pending array.
* @dram_default_hops: array that holds all hops addresses needed for default
* DRAM mapping.
- * @pending_cb_lock: spinlock to protect pending cb list
* @cs_lock: spinlock to protect cs_sequence.
* @dram_phys_mem: amount of used physical DRAM memory by this context.
* @thread_ctx_switch_token: token to prevent multiple threads of the same
* context from running the context switch phase.
* Only a single thread should run it.
- * @thread_pending_cb_token: token to prevent multiple threads from processing
- * the pending CB list. Only a single thread should
- * process the list since it is protected by a
- * spinlock and we don't want to halt the entire
- * command submission sequence.
* @thread_ctx_switch_wait_token: token to prevent the threads that didn't run
* the context switch phase from moving to their
* execution phase before the context switch phase
struct mutex mmu_lock;
struct mutex hw_block_list_lock;
struct list_head debugfs_list;
- struct list_head pending_cb_list;
struct list_head hw_block_mem_list;
struct hl_cs_counters_atomic cs_counters;
struct gen_pool *cb_va_pool;
+ struct hl_encaps_signals_mgr sig_mgr;
u64 cs_sequence;
u64 *dram_default_hops;
- spinlock_t pending_cb_lock;
spinlock_t cs_lock;
atomic64_t dram_phys_mem;
atomic_t thread_ctx_switch_token;
- atomic_t thread_pending_cb_token;
u32 thread_ctx_switch_wait_token;
u32 asid;
u32 handle;
* @sgt: pointer to the scatter-gather table that holds the pages.
* @dir: for DMA unmapping, the direction must be supplied, so save it.
* @debugfs_list: node in debugfs list of command submissions.
+ * @pid: the pid of the user process owning the memory
* @addr: user-space virtual address of the start of the memory area.
* @size: size of the memory area to pin & map.
* @dma_mapped: true if the SG was mapped to DMA addresses, false otherwise.
struct sg_table *sgt;
enum dma_data_direction dir;
struct list_head debugfs_list;
+ pid_t pid;
u64 addr;
u64 size;
u8 dma_mapped;
* @mirror_node : node in device mirror list of command submissions.
* @staged_cs_node: node in the staged cs list.
* @debugfs_list: node in debugfs list of command submissions.
+ * @encaps_sig_hdl: holds the encaps signals handle.
* @sequence: the sequence number of this CS.
* @staged_sequence: the sequence of the staged submission this CS is part of,
* relevant only if staged_cs is set.
* @timeout_jiffies: cs timeout in jiffies.
* @submission_time_jiffies: submission time of the cs
* @type: CS_TYPE_*.
+ * @encaps_sig_hdl_id: encaps signals handle id, set for the first staged cs.
* @submitted: true if CS was submitted to H/W.
* @completed: true if CS was completed by device.
* @timedout : true if CS was timedout.
* @staged_cs: true if this CS is part of a staged submission.
* @skip_reset_on_timeout: true if we shall not reset the device in case
* timeout occurs (debug scenario).
+ * @encaps_signals: true if this CS has encaps reserved signals.
*/
struct hl_cs {
u16 *jobs_in_queue_cnt;
struct list_head mirror_node;
struct list_head staged_cs_node;
struct list_head debugfs_list;
+ struct hl_cs_encaps_sig_handle *encaps_sig_hdl;
u64 sequence;
u64 staged_sequence;
u64 timeout_jiffies;
u64 submission_time_jiffies;
enum hl_cs_type type;
+ u32 encaps_sig_hdl_id;
u8 submitted;
u8 completed;
u8 timedout;
u8 staged_first;
u8 staged_cs;
u8 skip_reset_on_timeout;
+ u8 encaps_signals;
};
/**
* @hw_queue_id: the id of the H/W queue this job is submitted to.
* @user_cb_size: the actual size of the CB we got from the user.
* @job_cb_size: the actual size of the CB that we put on the queue.
+ * @encaps_sig_wait_offset: encapsulated signals offset, which allow user
+ * to wait on part of the reserved signals.
* @is_kernel_allocated_cb: true if the CB handle we got from the user holds a
* handle to a kernel-allocated CB object, false
* otherwise (SRAM/DRAM/host address).
u32 hw_queue_id;
u32 user_cb_size;
u32 job_cb_size;
+ u32 encaps_sig_wait_offset;
u8 is_kernel_allocated_cb;
u8 contains_dma_pkt;
};
* @state_dump_sem: protects state_dump.
* @addr: next address to read/write from/to in read/write32.
* @mmu_addr: next virtual address to translate to physical address in mmu_show.
+ * @userptr_lookup: the target user ptr to look up for on demand.
* @mmu_asid: ASID to use while translating in mmu_show.
* @state_dump_head: index of the latest state dump
* @i2c_bus: generic u8 debugfs file for bus value to use in i2c_data_read.
struct rw_semaphore state_dump_sem;
u64 addr;
u64 mmu_addr;
+ u64 userptr_lookup;
u32 mmu_asid;
u32 state_dump_head;
u8 i2c_bus;
#define HL_STR_MAX 32
-#define HL_DEV_STS_MAX (HL_DEVICE_STATUS_NEEDS_RESET + 1)
+#define HL_DEV_STS_MAX (HL_DEVICE_STATUS_LAST + 1)
/* Theoretical limit only. A single host can only contain up to 4 or 8 PCIe
* x16 cards. In extreme cases, there are hosts that can accommodate 16 cards.
* @wq: work queue for device reset procedure.
* @reset_work: reset work to be done.
* @hdev: habanalabs device structure.
+ * @fw_reset: whether f/w will do the reset without us sending them a message to do it.
*/
struct hl_device_reset_work {
struct workqueue_struct *wq;
struct delayed_work reset_work;
struct hl_device *hdev;
+ bool fw_reset;
};
/**
u64 virt_addr, struct hl_mmu_hop_info *hops);
};
+/**
+ * number of user contexts allowed to call wait_for_multi_cs ioctl in
+ * parallel
+ */
+#define MULTI_CS_MAX_USER_CTX 2
+
+/**
+ * struct multi_cs_completion - multi CS wait completion.
+ * @completion: completion of any of the CS in the list
+ * @lock: spinlock for the completion structure
+ * @timestamp: timestamp for the multi-CS completion
+ * @stream_master_qid_map: bitmap of all stream masters on which the multi-CS
+ * is waiting
+ * @used: 1 if in use, otherwise 0
+ */
+struct multi_cs_completion {
+ struct completion completion;
+ spinlock_t lock;
+ s64 timestamp;
+ u32 stream_master_qid_map;
+ u8 used;
+};
+
+/**
+ * struct multi_cs_data - internal data for multi CS call
+ * @ctx: pointer to the context structure
+ * @fence_arr: array of fences of all CSs
+ * @seq_arr: array of CS sequence numbers
+ * @timeout_us: timeout in usec for waiting for CS to complete
+ * @timestamp: timestamp of first completed CS
+ * @wait_status: wait for CS status
+ * @completion_bitmap: bitmap of completed CSs (1- completed, otherwise 0)
+ * @stream_master_qid_map: bitmap of all stream master QIDs on which the
+ * multi-CS is waiting
+ * @arr_len: fence_arr and seq_arr array length
+ * @gone_cs: indication of gone CS (1- there was gone CS, otherwise 0)
+ * @update_ts: update timestamp. 1- update the timestamp, otherwise 0.
+ */
+struct multi_cs_data {
+ struct hl_ctx *ctx;
+ struct hl_fence **fence_arr;
+ u64 *seq_arr;
+ s64 timeout_us;
+ s64 timestamp;
+ long wait_status;
+ u32 completion_bitmap;
+ u32 stream_master_qid_map;
+ u8 arr_len;
+ u8 gone_cs;
+ u8 update_ts;
+};
+
/**
* struct hl_device - habanalabs device structure.
* @pdev: pointer to PCI device, can be NULL in case of simulator device.
* @fw_loader: FW loader manager.
* @pci_mem_region: array of memory regions in the PCI
* @state_dump_specs: constants and dictionaries needed to dump system state.
+ * @multi_cs_completion: array of multi-CS completion.
* @dram_used_mem: current DRAM memory consumption.
* @timeout_jiffies: device CS timeout value.
* @max_power: the max power of the device, as configured by the sysadmin. This
* halted. We can't halt it again because the COMMS
* protocol will throw an error. Relevant only for
* cases where Linux was not loaded to device CPU
+ * @supports_wait_for_multi_cs: true if wait for multi CS is supported
*/
struct hl_device {
struct pci_dev *pdev;
struct hl_state_dump_specs state_dump_specs;
+ struct multi_cs_completion multi_cs_completion[
+ MULTI_CS_MAX_USER_CTX];
+ u32 *stream_master_qid_arr;
atomic64_t dram_used_mem;
u64 timeout_jiffies;
u64 max_power;
u8 curr_reset_cause;
u8 skip_reset_on_timeout;
u8 device_cpu_is_halted;
+ u8 supports_wait_for_multi_cs;
+ u8 stream_master_qid_arr_size;
/* Parameters for bring-up */
u64 nic_ports_mask;
};
+/**
+ * struct hl_cs_encaps_sig_handle - encapsulated signals handle structure
+ * @refcount: refcount used to protect removing this id when several
+ * wait cs are used to wait of the reserved encaps signals.
+ * @hdev: pointer to habanalabs device structure.
+ * @hw_sob: pointer to H/W SOB used in the reservation.
+ * @cs_seq: staged cs sequence which contains encapsulated signals
+ * @id: idr handler id to be used to fetch the handler info
+ * @q_idx: stream queue index
+ * @pre_sob_val: current SOB value before reservation
+ * @count: signals number
+ */
+struct hl_cs_encaps_sig_handle {
+ struct kref refcount;
+ struct hl_device *hdev;
+ struct hl_hw_sob *hw_sob;
+ u64 cs_seq;
+ u32 id;
+ u32 q_idx;
+ u32 pre_sob_val;
+ u32 count;
+};
+
/*
* IOCTLs
*/
* Kernel module functions that can be accessed by entire module
*/
+/**
+ * hl_get_sg_info() - get number of pages and the DMA address from SG list.
+ * @sg: the SG list.
+ * @dma_addr: pointer to DMA address to return.
+ *
+ * Calculate the number of consecutive pages described by the SG list. Take the
+ * offset of the address in the first page, add to it the length and round it up
+ * to the number of needed pages.
+ */
+static inline u32 hl_get_sg_info(struct scatterlist *sg, dma_addr_t *dma_addr)
+{
+ *dma_addr = sg_dma_address(sg);
+
+ return ((((*dma_addr) & (PAGE_SIZE - 1)) + sg_dma_len(sg)) +
+ (PAGE_SIZE - 1)) >> PAGE_SHIFT;
+}
+
/**
* hl_mem_area_inside_range() - Checks whether address+size are inside a range.
* @address: The start address of the area we want to validate.
int hl_hw_queues_create(struct hl_device *hdev);
void hl_hw_queues_destroy(struct hl_device *hdev);
int hl_hw_queue_send_cb_no_cmpl(struct hl_device *hdev, u32 hw_queue_id,
- u32 cb_size, u64 cb_ptr);
+ u32 cb_size, u64 cb_ptr);
+void hl_hw_queue_submit_bd(struct hl_device *hdev, struct hl_hw_queue *q,
+ u32 ctl, u32 len, u64 ptr);
int hl_hw_queue_schedule_cs(struct hl_cs *cs);
u32 hl_hw_queue_add_ptr(u32 ptr, u16 val);
void hl_hw_queue_inc_ci_kernel(struct hl_device *hdev, u32 hw_queue_id);
void hl_ctx_get(struct hl_device *hdev, struct hl_ctx *ctx);
int hl_ctx_put(struct hl_ctx *ctx);
struct hl_fence *hl_ctx_get_fence(struct hl_ctx *ctx, u64 seq);
+int hl_ctx_get_fences(struct hl_ctx *ctx, u64 *seq_arr,
+ struct hl_fence **fence, u32 arr_len);
void hl_ctx_mgr_init(struct hl_ctx_mgr *mgr);
void hl_ctx_mgr_fini(struct hl_device *hdev, struct hl_ctx_mgr *mgr);
void hl_cb_va_pool_fini(struct hl_ctx *ctx);
void hl_cs_rollback_all(struct hl_device *hdev);
-void hl_pending_cb_list_flush(struct hl_ctx *ctx);
struct hl_cs_job *hl_cs_allocate_job(struct hl_device *hdev,
enum hl_queue_type queue_type, bool is_kernel_allocated_cb);
void hl_sob_reset_error(struct kref *ref);
int hl_gen_sob_mask(u16 sob_base, u8 sob_mask, u8 *mask);
void hl_fence_put(struct hl_fence *fence);
+void hl_fences_put(struct hl_fence **fence, int len);
void hl_fence_get(struct hl_fence *fence);
void cs_get(struct hl_cs *cs);
bool cs_needs_completion(struct hl_cs *cs);
bool cs_needs_timeout(struct hl_cs *cs);
bool is_staged_cs_last_exists(struct hl_device *hdev, struct hl_cs *cs);
struct hl_cs *hl_staged_cs_find_first(struct hl_device *hdev, u64 cs_seq);
+void hl_multi_cs_completion_init(struct hl_device *hdev);
void goya_set_asic_funcs(struct hl_device *hdev);
void gaudi_set_asic_funcs(struct hl_device *hdev);
int sensor_index, u32 attr, long value);
int hl_set_current(struct hl_device *hdev,
int sensor_index, u32 attr, long value);
+void hw_sob_get(struct hl_hw_sob *hw_sob);
+void hw_sob_put(struct hl_hw_sob *hw_sob);
+void hl_encaps_handle_do_release(struct kref *ref);
+void hl_hw_queue_encaps_sig_set_sob_info(struct hl_device *hdev,
+ struct hl_cs *cs, struct hl_cs_job *job,
+ struct hl_cs_compl *cs_cmpl);
void hl_release_pending_user_interrupts(struct hl_device *hdev);
int hl_cs_signal_sob_wraparound_handler(struct hl_device *hdev, u32 q_idx,
- struct hl_hw_sob **hw_sob, u32 count);
+ struct hl_hw_sob **hw_sob, u32 count, bool encaps_sig);
int hl_state_dump(struct hl_device *hdev);
const char *hl_state_dump_get_sync_name(struct hl_device *hdev, u32 sync_id);