DMARC-Filter: OpenDMARC Filter v1.3.2 mx1.redhat.com 5BB50285C1
Subject: Re: [PATCH] tcmu: Add fifo type waiter list support to avoid
 starvation
To: lixiubo@cmss.chinamobile.com, nab@linux-iscsi.org
References: <1493952686-29396-1-git-send-email-lixiubo@cmss.chinamobile.com>
Cc: agrover@redhat.com, namei.unix@gmail.com, sheng@yasker.org,
        linux-scsi@vger.kernel.org, target-devel@vger.kernel.org,
        linux-kernel@vger.kernel.org
From: Mike Christie <mchristi@redhat.com>
Message-ID: <5933887D.5080708@redhat.com>
Date: Sat, 3 Jun 2017 23:11:41 -0500
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101
 Thunderbird/38.6.0
MIME-Version: 1.0
In-Reply-To: <1493952686-29396-1-git-send-email-lixiubo@cmss.chinamobile.com>
Content-Type: text/plain; charset=windows-1252
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 7197
Lines: 231

On 05/04/2017 09:51 PM, lixiubo@cmss.chinamobile.com wrote:
> From: Xiubo Li <lixiubo@cmss.chinamobile.com>
> 
> The fifo type waiter list will hold the udevs who are waiting for the
> blocks from the data global pool. The unmap thread will try to feed the
> first udevs in waiter list, if the global free blocks available are
> not enough, it will stop traversing the list and abort waking up the
> others.
> 
> Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com>
> ---
>  drivers/target/target_core_user.c | 82 +++++++++++++++++++++++++++++++--------
>  1 file changed, 66 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/target/target_core_user.c b/drivers/target/target_core_user.c
> index 9045837..10c99e7 100644
> --- a/drivers/target/target_core_user.c
> +++ b/drivers/target/target_core_user.c
> @@ -97,6 +97,7 @@ struct tcmu_hba {
>  
>  struct tcmu_dev {
>  	struct list_head node;
> +	struct list_head waiter;
>  
>  	struct se_device se_dev;
>  
> @@ -123,7 +124,7 @@ struct tcmu_dev {
>  	wait_queue_head_t wait_cmdr;
>  	struct mutex cmdr_lock;
>  
> -	bool waiting_global;
> +	uint32_t waiting_blocks;
>  	uint32_t dbi_max;
>  	uint32_t dbi_thresh;
>  	DECLARE_BITMAP(data_bitmap, DATA_BLOCK_BITS);
> @@ -165,6 +166,10 @@ struct tcmu_cmd {
>  static DEFINE_MUTEX(root_udev_mutex);
>  static LIST_HEAD(root_udev);
>  
> +/* The data blocks global pool waiter list */
> +static DEFINE_MUTEX(root_udev_waiter_mutex);

Sorry for the delay.


Could you just add a comment about the lock ordering. It will be helpful
to new engineers/reviewers to check for errors.


> +static LIST_HEAD(root_udev_waiter);
> +
>  static atomic_t global_db_count = ATOMIC_INIT(0);
>  
>  static struct kmem_cache *tcmu_cmd_cache;
> @@ -195,6 +200,11 @@ enum tcmu_multicast_groups {
>  #define tcmu_cmd_set_dbi(cmd, index) ((cmd)->dbi[(cmd)->dbi_cur++] = (index))
>  #define tcmu_cmd_get_dbi(cmd) ((cmd)->dbi[(cmd)->dbi_cur++])
>  
> +static inline bool is_in_waiter_list(struct tcmu_dev *udev)
> +{
> +	return !!udev->waiting_blocks;
> +}
> +
>  static void tcmu_cmd_free_data(struct tcmu_cmd *tcmu_cmd, uint32_t len)
>  {
>  	struct tcmu_dev *udev = tcmu_cmd->tcmu_dev;
> @@ -250,8 +260,6 @@ static bool tcmu_get_empty_blocks(struct tcmu_dev *udev,
>  {
>  	int i;
>  
> -	udev->waiting_global = false;
> -
>  	for (i = tcmu_cmd->dbi_cur; i < tcmu_cmd->dbi_cnt; i++) {
>  		if (!tcmu_get_empty_block(udev, tcmu_cmd))
>  			goto err;
> @@ -259,9 +267,7 @@ static bool tcmu_get_empty_blocks(struct tcmu_dev *udev,
>  	return true;
>  
>  err:
> -	udev->waiting_global = true;
> -	/* Try to wake up the unmap thread */
> -	wake_up(&unmap_wait);
> +	udev->waiting_blocks += tcmu_cmd->dbi_cnt - i;
>  	return false;
>  }
>  
> @@ -671,6 +677,7 @@ static inline size_t tcmu_cmd_get_cmd_size(struct tcmu_cmd *tcmu_cmd,
>  
>  	while (!is_ring_space_avail(udev, tcmu_cmd, command_size, data_length)) {
>  		int ret;
> +		bool is_waiting_blocks = !!udev->waiting_blocks;

You can use your helper is_in_waiter_list().

>  		DEFINE_WAIT(__wait);
>  
>  		prepare_to_wait(&udev->wait_cmdr, &__wait, TASK_INTERRUPTIBLE);
> @@ -688,6 +695,18 @@ static inline size_t tcmu_cmd_get_cmd_size(struct tcmu_cmd *tcmu_cmd,
>  			return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
>  		}
>  
> +		/*
> +		 * Try to insert the current udev to waiter list and
> +		 * then wake up the unmap thread
> +		 */
> +		if (is_waiting_blocks) {
> +			mutex_lock(&root_udev_waiter_mutex);
> +			list_add_tail(&udev->waiter, &root_udev_waiter);
> +			mutex_unlock(&root_udev_waiter_mutex);
> +
> +			wake_up(&unmap_wait);
> +		}

Is this supposed to go before the schedule_timeout call?

If we are here and schedule_timeout returned non zero then our wait_cmdr
has been woken up from the unmap thread because it freed some space, so
we do not want to again add ourself and call unmap again.


> +
>  		mutex_lock(&udev->cmdr_lock);
>  
>  		/* We dropped cmdr_lock, cmd_head is stale */
> @@ -902,8 +921,6 @@ static unsigned int tcmu_handle_completions(struct tcmu_dev *udev)
>  	if (mb->cmd_tail == mb->cmd_head)
>  		del_timer(&udev->timeout); /* no more pending cmds */
>  
> -	wake_up(&udev->wait_cmdr);
> -
>  	return handled;
>  }
>  
> @@ -996,7 +1013,17 @@ static int tcmu_irqcontrol(struct uio_info *info, s32 irq_on)
>  	struct tcmu_dev *tcmu_dev = container_of(info, struct tcmu_dev, uio_info);
>  
>  	mutex_lock(&tcmu_dev->cmdr_lock);
> -	tcmu_handle_completions(tcmu_dev);
> +	/*
> +	 * If the current udev is also in waiter list, this will
> +	 * make sure that the other waiters in list be feeded ahead
> +	 * of it.
> +	 */
> +	if (is_in_waiter_list(tcmu_dev)) {
> +		wake_up(&unmap_wait);
> +	} else {
> +		tcmu_handle_completions(tcmu_dev);
> +		wake_up(&tcmu_dev->wait_cmdr);
> +	}
>  	mutex_unlock(&tcmu_dev->cmdr_lock);
>  
>  	return 0;
> @@ -1231,7 +1258,7 @@ static int tcmu_configure_device(struct se_device *dev)
>  	udev->data_off = CMDR_SIZE;
>  	udev->data_size = DATA_SIZE;
>  	udev->dbi_thresh = 0; /* Default in Idle state */
> -	udev->waiting_global = false;
> +	udev->waiting_blocks = 0;
>  
>  	/* Initialise the mailbox of the ring buffer */
>  	mb = udev->mb_addr;
> @@ -1345,6 +1372,13 @@ static void tcmu_free_device(struct se_device *dev)
>  	list_del(&udev->node);
>  	mutex_unlock(&root_udev_mutex);
>  
> +	mutex_lock(&root_udev_waiter_mutex);
> +	mutex_lock(&udev->cmdr_lock);
> +	if (is_in_waiter_list(udev))
> +		list_del(&udev->waiter);
> +	mutex_unlock(&udev->cmdr_lock);
> +	mutex_unlock(&root_udev_waiter_mutex);
> +
>  	vfree(udev->mb_addr);
>  
>  	/* Upper layer should drain all requests before calling this */
> @@ -1550,6 +1584,7 @@ static int unmap_thread_fn(void *data)
>  	struct tcmu_dev *udev;
>  	loff_t off;
>  	uint32_t start, end, block;
> +	static uint32_t free_blocks;
>  	struct page *page;
>  	int i;
>  
> @@ -1571,7 +1606,7 @@ static int unmap_thread_fn(void *data)
>  			tcmu_handle_completions(udev);
>  
>  			/* Skip the udevs waiting the global pool or in idle */
> -			if (udev->waiting_global || !udev->dbi_thresh) {
> +			if (is_in_waiter_list(udev) || !udev->dbi_thresh) {
>  				mutex_unlock(&udev->cmdr_lock);
>  				continue;
>  			}
> @@ -1607,17 +1642,32 @@ static int unmap_thread_fn(void *data)
>  				}
>  			}
>  			mutex_unlock(&udev->cmdr_lock);
> +
> +			free_blocks += end - start;
>  		}
> +		mutex_unlock(&root_udev_mutex);
>  
>  		/*
>  		 * Try to wake up the udevs who are waiting
> -		 * for the global data pool.
> +		 * for the global data pool blocks.
>  		 */
> -		list_for_each_entry(udev, &root_udev, node) {
> -			if (udev->waiting_global)
> -				wake_up(&udev->wait_cmdr);
> +		mutex_lock(&root_udev_waiter_mutex);
> +		list_for_each_entry(udev, &root_udev_waiter, waiter) {
> +			mutex_lock(&udev->cmdr_lock);
> +			if (udev->waiting_blocks < free_blocks) {
> +				mutex_unlock(&udev->cmdr_lock);
> +				break;
> +			}
> +
> +			free_blocks -= udev->waiting_blocks;
> +			udev->waiting_blocks = 0;
> +			mutex_unlock(&udev->cmdr_lock);
> +
> +			list_del(&udev->waiter);
> +
> +			wake_up(&udev->wait_cmdr);
>  		}
> -		mutex_unlock(&root_udev_mutex);
> +		mutex_unlock(&root_udev_waiter_mutex);
>  	}
>  
>  	return 0;
>