Merge patch series "Optimize the hot path in the UFS driver"

Bart Van Assche <bvanassche@acm.org> says:

Hi Martin,

This patch series optimizes the hot path of the UFS driver by making
struct scsi_cmnd and struct ufshcd_lrb adjacent. Making these two data
structures adjacent is realized as follows:

@@ -9040,6 +9046,7 @@ static const struct scsi_host_template ufshcd_driver_template = {
     .name           = UFSHCD,
     .proc_name      = UFSHCD,
     .map_queues     = ufshcd_map_queues,
+    .cmd_size       = sizeof(struct ufshcd_lrb),
     .init_cmd_priv  = ufshcd_init_cmd_priv,
     .queuecommand   = ufshcd_queuecommand,
     .mq_poll        = ufshcd_poll,

The following changes had to be made prior to making these two data
structures adjacent:
* Add support for driver-internal and reserved commands in the SCSI core.
* Instead of making the reserved command slot (hba->reserved_slot)
  invisible to the SCSI core, let the SCSI core allocate a reserved command.
* Remove all UFS data structure members that are no longer needed
  because struct scsi_cmnd and struct ufshcd_lrb are now adjacent
* Call ufshcd_init_lrb() from inside the code for queueing a command instead of
  calling this function before I/O starts. This is necessary because
  ufshcd_memory_alloc() allocates fewer instances than the block layer
  allocates requests. See also the following code in the block layer
  core:

    if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
                hctx->numa_node))

  Although the UFS driver could be modified such that ufshcd_init_lrb()
  is called from ufshcd_init_cmd_priv(), realizing this would require
  moving the memory allocations that happen from inside
  ufshcd_memory_alloc() into ufshcd_init_cmd_priv(). That would make
  this patch series even larger. Although ufshcd_init_lrb() is called for each
  command, the benefits of reduced indirection and better cache efficiency
  outweigh the small overhead of per-command lrb initialization.
* ufshcd_add_scsi_host() happens now before any device management
  commands are submitted. This change is necessary because this patch
  makes device management command allocation happen when the SCSI host
  is allocated.
* Allocate as many command slots as the host controller supports. Decrease
  host->cmds_per_lun if necessary once it is clear whether or not the UFS
  device supports less command slots than the host controller.

Please consider this patch series for the next merge window.

Thanks,

Bart.

Link: https://patch.msgid.link/20251031204029.2883185-1-bvanassche@acm.org
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

2 files changed, 52 insertions, 1 deletions

diff --git a/include/scsi/scsi_device.h b/include/scsi/scsi_device.h
index 9e24a7dde0d3..d62265d12cfe 100644
--- a/include/scsi/scsi_device.h
+++ b/include/scsi/scsi_device.h
@@ -564,6 +564,10 @@ int scsi_execute_cmd(struct scsi_device *sdev, const unsigned char *cmd,
 		     const struct scsi_exec_args *args);
 void scsi_failures_reset_retries(struct scsi_failures *failures);
 
+struct scsi_cmnd *scsi_get_internal_cmd(struct scsi_device *sdev,
+					enum dma_data_direction data_direction,
+					blk_mq_req_flags_t flags);
+void scsi_put_internal_cmd(struct scsi_cmnd *scmd);
 extern void sdev_disable_disk_events(struct scsi_device *sdev);
 extern void sdev_enable_disk_events(struct scsi_device *sdev);
 extern int scsi_vpd_lun_id(struct scsi_device *, char *, size_t);
@@ -595,6 +599,22 @@ static inline unsigned int sdev_id(struct scsi_device *sdev)
 #define scmd_id(scmd) sdev_id((scmd)->device)
 #define scmd_channel(scmd) sdev_channel((scmd)->device)
 
+/**
+ * scsi_device_is_pseudo_dev() - Whether a device is a pseudo SCSI device.
+ * @sdev: SCSI device to examine
+ *
+ * A pseudo SCSI device can be used to allocate SCSI commands but does not show
+ * up in sysfs. Additionally, the logical unit information in *@sdev is made up.
+ *
+ * This function tests the LUN number instead of comparing @sdev with
+ * @sdev->host->pseudo_sdev because this function may be called before
+ * @sdev->host->pseudo_sdev has been initialized.
+ */
+static inline bool scsi_device_is_pseudo_dev(struct scsi_device *sdev)
+{
+	return sdev->lun == U64_MAX;
+}
+
 /*
  * checks for positions of the SCSI state machine
  */
diff --git a/include/scsi/scsi_host.h b/include/scsi/scsi_host.h
index f5a243261236..e87cf7eadd26 100644
--- a/include/scsi/scsi_host.h
+++ b/include/scsi/scsi_host.h
@@ -87,6 +87,12 @@ struct scsi_host_template {
 	int (* queuecommand)(struct Scsi_Host *, struct scsi_cmnd *);
 
 	/*
+	 * Queue a reserved command (BLK_MQ_REQ_RESERVED). The .queuecommand()
+	 * documentation also applies to the .queue_reserved_command() callback.
+	 */
+	int (*queue_reserved_command)(struct Scsi_Host *, struct scsi_cmnd *);
+
+	/*
 	 * The commit_rqs function is used to trigger a hardware
 	 * doorbell after some requests have been queued with
 	 * queuecommand, when an error is encountered before sending
@@ -375,11 +381,20 @@ struct scsi_host_template {
 	/*
 	 * This determines if we will use a non-interrupt driven
 	 * or an interrupt driven scheme.  It is set to the maximum number
-	 * of simultaneous commands a single hw queue in HBA will accept.
+	 * of simultaneous commands a single hw queue in HBA will accept
+	 * excluding internal commands.
 	 */
 	int can_queue;
 
 	/*
+	 * This determines how many commands the HBA will set aside
+	 * for internal commands. This number will be added to
+	 * @can_queue to calculate the maximum number of simultaneous
+	 * commands sent to the host.
+	 */
+	int nr_reserved_cmds;
+
+	/*
 	 * In many instances, especially where disconnect / reconnect are
 	 * supported, our host also has an ID on the SCSI bus.  If this is
 	 * the case, then it must be reserved.  Please set this_id to -1 if
@@ -611,7 +626,17 @@ struct Scsi_Host {
 	unsigned short max_cmd_len;
 
 	int this_id;
+
+	/*
+	 * Number of commands this host can handle at the same time.
+	 * This excludes reserved commands as specified by nr_reserved_cmds.
+	 */
 	int can_queue;
+	/*
+	 * Number of reserved commands to allocate, if any.
+	 */
+	unsigned int nr_reserved_cmds;
+
 	short cmd_per_lun;
 	short unsigned int sg_tablesize;
 	short unsigned int sg_prot_tablesize;
@@ -703,6 +728,12 @@ struct Scsi_Host {
 	struct device		shost_gendev, shost_dev;
 
 	/*
+	 * A SCSI device structure used for sending internal commands to the
+	 * HBA. There is no corresponding logical unit inside the SCSI device.
+	 */
+	struct scsi_device *pseudo_sdev;
+
+	/*
 	 * Points to the transport data (if any) which is allocated
 	 * separately
 	 */