diff options
author | Nicholas Bellinger | 2011-06-13 14:46:09 -0700 |
---|---|---|
committer | Nicholas Bellinger | 2011-07-22 09:37:45 +0000 |
commit | 07bde79a5c355dbca66ca4318645aa17b4c0d859 (patch) | |
tree | e602132193959df436a2f5d325cd3da9984ada4f | |
parent | 695434e1cbd57f404110bf4ab187a5127ffd79bb (diff) |
target: Add SCF_EMULATE_QUEUE_FULL -> transport_handle_queue_full
This patch adds SCF_EMULATE_QUEUE_FULL support using -EAGAIN failures
via transport_handle_queue_full() to signal queue full in completion
path TFO->queue_data_in() and TFO->queue_status() callbacks.
This is done using a new se_cmd->transport_qf_callback() to handle
the following queue full exception cases within target core:
*) TRANSPORT_COMPLETE_OK (for completion path queue full)
*) TRANSPORT_COMPLETE_QF_WP (for TRANSPORT_WRITE_PENDING queue full)
*) transport_send_check_condition_and_sense() failure paths in
transport_generic_request_failure() and transport_generic_complete_ok()
All logic is driven using se_device->qf_work_queue -> target_qf_do_work()
to to requeue outstanding se_cmd at the head of se_dev->queue_obj->qobj_list
for transport_processing_thread() execution.
Tested using tcm_qla2xxx with MAX_OUTSTANDING_COMMANDS=128 for FCP READ
to trigger the TRANSPORT_COMPLETE_OK queue full cases, and a simulated
TFO->write_pending() -EAGAIN failure to trigger TRANSPORT_COMPLETE_QF_WP.
Reported-by: Roland Dreier <roland@purestorage.com>
Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
-rw-r--r-- | drivers/target/target_core_transport.c | 194 | ||||
-rw-r--r-- | include/target/target_core_base.h | 9 |
2 files changed, 188 insertions, 15 deletions
diff --git a/drivers/target/target_core_transport.c b/drivers/target/target_core_transport.c index 1ae6eb7a621b..056c4cb4736d 100644 --- a/drivers/target/target_core_transport.c +++ b/drivers/target/target_core_transport.c @@ -204,6 +204,9 @@ static int transport_generic_write_pending(struct se_cmd *); static int transport_processing_thread(void *param); static int __transport_execute_tasks(struct se_device *dev); static void transport_complete_task_attr(struct se_cmd *cmd); +static int transport_complete_qf(struct se_cmd *cmd); +static void transport_handle_queue_full(struct se_cmd *cmd, + struct se_device *dev, int (*qf_callback)(struct se_cmd *)); static void transport_direct_request_timeout(struct se_cmd *cmd); static void transport_free_dev_tasks(struct se_cmd *cmd); static u32 transport_allocate_tasks(struct se_cmd *cmd, @@ -768,7 +771,11 @@ static void transport_add_cmd_to_queue( } spin_lock_irqsave(&qobj->cmd_queue_lock, flags); - list_add_tail(&cmd->se_queue_node, &qobj->qobj_list); + if (cmd->se_cmd_flags & SCF_EMULATE_QUEUE_FULL) { + cmd->se_cmd_flags &= ~SCF_EMULATE_QUEUE_FULL; + list_add(&cmd->se_queue_node, &qobj->qobj_list); + } else + list_add_tail(&cmd->se_queue_node, &qobj->qobj_list); atomic_inc(&cmd->t_transport_queue_active); spin_unlock_irqrestore(&qobj->cmd_queue_lock, flags); @@ -1102,6 +1109,40 @@ void transport_remove_task_from_execute_queue( spin_unlock_irqrestore(&dev->execute_task_lock, flags); } +/* + * Handle QUEUE_FULL / -EAGAIN status + */ + +static void target_qf_do_work(struct work_struct *work) +{ + struct se_device *dev = container_of(work, struct se_device, + qf_work_queue); + struct se_cmd *cmd, *cmd_tmp; + + spin_lock_irq(&dev->qf_cmd_lock); + list_for_each_entry_safe(cmd, cmd_tmp, &dev->qf_cmd_list, se_qf_node) { + + list_del(&cmd->se_qf_node); + atomic_dec(&dev->dev_qf_count); + smp_mb__after_atomic_dec(); + spin_unlock_irq(&dev->qf_cmd_lock); + + printk(KERN_INFO "Processing %s cmd: %p QUEUE_FULL in work queue" + " context: %s\n", cmd->se_tfo->get_fabric_name(), cmd, + (cmd->t_state == TRANSPORT_COMPLETE_OK) ? "COMPLETE_OK" : + (cmd->t_state == TRANSPORT_COMPLETE_QF_WP) ? "WRITE_PENDING" + : "UNKNOWN"); + /* + * The SCF_EMULATE_QUEUE_FULL flag will be cleared once se_cmd + * has been added to head of queue + */ + transport_add_cmd_to_queue(cmd, cmd->t_state); + + spin_lock_irq(&dev->qf_cmd_lock); + } + spin_unlock_irq(&dev->qf_cmd_lock); +} + unsigned char *transport_dump_cmd_direction(struct se_cmd *cmd) { switch (cmd->data_direction) { @@ -1531,6 +1572,7 @@ struct se_device *transport_add_device_to_core_hba( INIT_LIST_HEAD(&dev->delayed_cmd_list); INIT_LIST_HEAD(&dev->ordered_cmd_list); INIT_LIST_HEAD(&dev->state_task_list); + INIT_LIST_HEAD(&dev->qf_cmd_list); spin_lock_init(&dev->execute_task_lock); spin_lock_init(&dev->delayed_cmd_lock); spin_lock_init(&dev->ordered_cmd_lock); @@ -1541,6 +1583,7 @@ struct se_device *transport_add_device_to_core_hba( spin_lock_init(&dev->dev_status_thr_lock); spin_lock_init(&dev->se_port_lock); spin_lock_init(&dev->se_tmr_lock); + spin_lock_init(&dev->qf_cmd_lock); dev->queue_depth = dev_limits->queue_depth; atomic_set(&dev->depth_left, dev->queue_depth); @@ -1584,7 +1627,10 @@ struct se_device *transport_add_device_to_core_hba( dev->transport->name); goto out; } - + /* + * Setup work_queue for QUEUE_FULL + */ + INIT_WORK(&dev->qf_work_queue, target_qf_do_work); /* * Preload the initial INQUIRY const values if we are doing * anything virtual (IBLOCK, FILEIO, RAMDISK), but not for TCM/pSCSI @@ -1697,6 +1743,7 @@ void transport_init_se_cmd( INIT_LIST_HEAD(&cmd->se_lun_node); INIT_LIST_HEAD(&cmd->se_delayed_node); INIT_LIST_HEAD(&cmd->se_ordered_node); + INIT_LIST_HEAD(&cmd->se_qf_node); INIT_LIST_HEAD(&cmd->t_mem_list); INIT_LIST_HEAD(&cmd->t_mem_bidi_list); @@ -2019,6 +2066,8 @@ static void transport_generic_request_failure( int complete, int sc) { + int ret = 0; + DEBUG_GRF("-----[ Storage Engine Exception for cmd: %p ITT: 0x%08x" " CDB: 0x%02x\n", cmd, cmd->se_tfo->get_task_tag(cmd), cmd->t_task_cdb[0]); @@ -2109,7 +2158,9 @@ static void transport_generic_request_failure( cmd->orig_fe_lun, 0x2C, ASCQ_2CH_PREVIOUS_RESERVATION_CONFLICT_STATUS); - cmd->se_tfo->queue_status(cmd); + ret = cmd->se_tfo->queue_status(cmd); + if (ret == -EAGAIN) + goto queue_full; goto check_stop; case PYX_TRANSPORT_USE_SENSE_REASON: /* @@ -2126,13 +2177,22 @@ static void transport_generic_request_failure( if (!sc) transport_new_cmd_failure(cmd); - else - transport_send_check_condition_and_sense(cmd, - cmd->scsi_sense_reason, 0); + else { + ret = transport_send_check_condition_and_sense(cmd, + cmd->scsi_sense_reason, 0); + if (ret == -EAGAIN) + goto queue_full; + } + check_stop: transport_lun_remove_cmd(cmd); if (!(transport_cmd_check_stop_to_fabric(cmd))) ; + return; + +queue_full: + cmd->t_state = TRANSPORT_COMPLETE_OK; + transport_handle_queue_full(cmd, cmd->se_dev, transport_complete_qf); } static void transport_direct_request_timeout(struct se_cmd *cmd) @@ -3637,9 +3697,53 @@ static void transport_complete_task_attr(struct se_cmd *cmd) wake_up_interruptible(&dev->dev_queue_obj.thread_wq); } +static int transport_complete_qf(struct se_cmd *cmd) +{ + int ret = 0; + + if (cmd->se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) + return cmd->se_tfo->queue_status(cmd); + + switch (cmd->data_direction) { + case DMA_FROM_DEVICE: + ret = cmd->se_tfo->queue_data_in(cmd); + break; + case DMA_TO_DEVICE: + if (!list_empty(&cmd->t_mem_bidi_list)) { + ret = cmd->se_tfo->queue_data_in(cmd); + if (ret < 0) + return ret; + } + /* Fall through for DMA_TO_DEVICE */ + case DMA_NONE: + ret = cmd->se_tfo->queue_status(cmd); + break; + default: + break; + } + + return ret; +} + +static void transport_handle_queue_full( + struct se_cmd *cmd, + struct se_device *dev, + int (*qf_callback)(struct se_cmd *)) +{ + spin_lock_irq(&dev->qf_cmd_lock); + cmd->se_cmd_flags |= SCF_EMULATE_QUEUE_FULL; + cmd->transport_qf_callback = qf_callback; + list_add_tail(&cmd->se_qf_node, &cmd->se_dev->qf_cmd_list); + atomic_inc(&dev->dev_qf_count); + smp_mb__after_atomic_inc(); + spin_unlock_irq(&cmd->se_dev->qf_cmd_lock); + + schedule_work(&cmd->se_dev->qf_work_queue); +} + static void transport_generic_complete_ok(struct se_cmd *cmd) { - int reason = 0; + int reason = 0, ret; /* * Check if we need to move delayed/dormant tasks from cmds on the * delayed execution list after a HEAD_OF_QUEUE or ORDERED Task @@ -3648,6 +3752,21 @@ static void transport_generic_complete_ok(struct se_cmd *cmd) if (cmd->se_dev->dev_task_attr_type == SAM_TASK_ATTR_EMULATED) transport_complete_task_attr(cmd); /* + * Check to schedule QUEUE_FULL work, or execute an existing + * cmd->transport_qf_callback() + */ + if (atomic_read(&cmd->se_dev->dev_qf_count) != 0) + schedule_work(&cmd->se_dev->qf_work_queue); + + if (cmd->transport_qf_callback) { + ret = cmd->transport_qf_callback(cmd); + if (ret < 0) + goto queue_full; + + cmd->transport_qf_callback = NULL; + goto done; + } + /* * Check if we need to retrieve a sense buffer from * the struct se_cmd in question. */ @@ -3660,8 +3779,11 @@ static void transport_generic_complete_ok(struct se_cmd *cmd) * a non GOOD status. */ if (cmd->scsi_status) { - transport_send_check_condition_and_sense( + ret = transport_send_check_condition_and_sense( cmd, reason, 1); + if (ret == -EAGAIN) + goto queue_full; + transport_lun_remove_cmd(cmd); transport_cmd_check_stop_to_fabric(cmd); return; @@ -3693,7 +3815,9 @@ static void transport_generic_complete_ok(struct se_cmd *cmd) cmd->t_task_buf, cmd->data_length); - cmd->se_tfo->queue_data_in(cmd); + ret = cmd->se_tfo->queue_data_in(cmd); + if (ret == -EAGAIN) + goto queue_full; break; case DMA_TO_DEVICE: spin_lock(&cmd->se_lun->lun_sep_lock); @@ -3712,19 +3836,30 @@ static void transport_generic_complete_ok(struct se_cmd *cmd) cmd->data_length; } spin_unlock(&cmd->se_lun->lun_sep_lock); - cmd->se_tfo->queue_data_in(cmd); + ret = cmd->se_tfo->queue_data_in(cmd); + if (ret == -EAGAIN) + goto queue_full; break; } /* Fall through for DMA_TO_DEVICE */ case DMA_NONE: - cmd->se_tfo->queue_status(cmd); + ret = cmd->se_tfo->queue_status(cmd); + if (ret == -EAGAIN) + goto queue_full; break; default: break; } +done: transport_lun_remove_cmd(cmd); transport_cmd_check_stop_to_fabric(cmd); + return; + +queue_full: + printk(KERN_INFO "Handling complete_ok QUEUE_FULL: se_cmd: %p," + " data_direction: %d\n", cmd, cmd->data_direction); + transport_handle_queue_full(cmd, cmd->se_dev, transport_complete_qf); } static void transport_free_dev_tasks(struct se_cmd *cmd) @@ -4866,6 +5001,11 @@ void transport_generic_process_write(struct se_cmd *cmd) } EXPORT_SYMBOL(transport_generic_process_write); +static int transport_write_pending_qf(struct se_cmd *cmd) +{ + return cmd->se_tfo->write_pending(cmd); +} + /* transport_generic_write_pending(): * * @@ -4878,6 +5018,17 @@ static int transport_generic_write_pending(struct se_cmd *cmd) spin_lock_irqsave(&cmd->t_state_lock, flags); cmd->t_state = TRANSPORT_WRITE_PENDING; spin_unlock_irqrestore(&cmd->t_state_lock, flags); + + if (cmd->transport_qf_callback) { + ret = cmd->transport_qf_callback(cmd); + if (ret == -EAGAIN) + goto queue_full; + else if (ret < 0) + return ret; + + cmd->transport_qf_callback = NULL; + return 0; + } /* * For the TCM control CDBs using a contiguous buffer, do the memcpy * from the passed Linux/SCSI struct scatterlist located at @@ -4903,10 +5054,19 @@ static int transport_generic_write_pending(struct se_cmd *cmd) * frontend know that WRITE buffers are ready. */ ret = cmd->se_tfo->write_pending(cmd); - if (ret < 0) + if (ret == -EAGAIN) + goto queue_full; + else if (ret < 0) return ret; return PYX_TRANSPORT_WRITE_PENDING; + +queue_full: + printk(KERN_INFO "Handling write_pending QUEUE__FULL: se_cmd: %p\n", cmd); + cmd->t_state = TRANSPORT_COMPLETE_QF_WP; + transport_handle_queue_full(cmd, cmd->se_dev, + transport_write_pending_qf); + return ret; } void transport_release_cmd(struct se_cmd *cmd) @@ -5410,8 +5570,7 @@ int transport_send_check_condition_and_sense( cmd->scsi_sense_length = TRANSPORT_SENSE_BUFFER + offset; after_reason: - cmd->se_tfo->queue_status(cmd); - return 0; + return cmd->se_tfo->queue_status(cmd); } EXPORT_SYMBOL(transport_send_check_condition_and_sense); @@ -5733,7 +5892,9 @@ get_cmd: /* Fall through */ case TRANSPORT_NEW_CMD: ret = transport_generic_new_cmd(cmd); - if (ret < 0) { + if (ret == -EAGAIN) + break; + else if (ret < 0) { cmd->transport_error_status = ret; transport_generic_request_failure(cmd, NULL, 0, (cmd->data_direction != @@ -5763,6 +5924,9 @@ get_cmd: transport_stop_all_task_timers(cmd); transport_generic_request_timeout(cmd); break; + case TRANSPORT_COMPLETE_QF_WP: + transport_generic_write_pending(cmd); + break; default: printk(KERN_ERR "Unknown t_state: %d deferred_t_state:" " %d for ITT: 0x%08x i_state: %d on SE LUN:" diff --git a/include/target/target_core_base.h b/include/target/target_core_base.h index 71abc4c5e2b4..cd163dd94cd4 100644 --- a/include/target/target_core_base.h +++ b/include/target/target_core_base.h @@ -99,6 +99,7 @@ enum transport_state_table { TRANSPORT_FREE = 15, TRANSPORT_NEW_CMD_MAP = 16, TRANSPORT_FREE_CMD_INTR = 17, + TRANSPORT_COMPLETE_QF_WP = 18, }; /* Used for struct se_cmd->se_cmd_flags */ @@ -125,6 +126,7 @@ enum se_cmd_flags_table { SCF_PASSTHROUGH_CONTIG_TO_SG = 0x00200000, SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC = 0x00400000, SCF_EMULATE_CDB_ASYNC = 0x01000000, + SCF_EMULATE_QUEUE_FULL = 0x02000000, }; /* struct se_dev_entry->lun_flags and struct se_lun->lun_access */ @@ -466,6 +468,7 @@ struct se_cmd { struct list_head se_delayed_node; struct list_head se_ordered_node; struct list_head se_lun_node; + struct list_head se_qf_node; struct se_device *se_dev; struct se_dev_entry *se_deve; struct se_device *se_obj_ptr; @@ -480,6 +483,8 @@ struct se_cmd { void (*transport_split_cdb)(unsigned long long, u32 *, unsigned char *); void (*transport_wait_for_tasks)(struct se_cmd *, int, int); void (*transport_complete_callback)(struct se_cmd *); + int (*transport_qf_callback)(struct se_cmd *); + unsigned char *t_task_cdb; unsigned char __t_task_cdb[TCM_MAX_COMMAND_SIZE]; unsigned long long t_task_lba; @@ -743,6 +748,7 @@ struct se_device { atomic_t dev_status_thr_count; atomic_t dev_hoq_count; atomic_t dev_ordered_sync; + atomic_t dev_qf_count; struct se_obj dev_obj; struct se_obj dev_access_obj; struct se_obj dev_export_obj; @@ -758,6 +764,7 @@ struct se_device { spinlock_t dev_status_thr_lock; spinlock_t se_port_lock; spinlock_t se_tmr_lock; + spinlock_t qf_cmd_lock; /* Used for legacy SPC-2 reservationsa */ struct se_node_acl *dev_reserved_node_acl; /* Used for ALUA Logical Unit Group membership */ @@ -771,10 +778,12 @@ struct se_device { struct task_struct *process_thread; pid_t process_thread_pid; struct task_struct *dev_mgmt_thread; + struct work_struct qf_work_queue; struct list_head delayed_cmd_list; struct list_head ordered_cmd_list; struct list_head execute_task_list; struct list_head state_task_list; + struct list_head qf_cmd_list; /* Pointer to associated SE HBA */ struct se_hba *se_hba; struct se_subsystem_dev *se_sub_dev; |