Skip to contentSkip to conversation

ufs: Do not requeue while ungating the clock

PATCH Superseded Bart Van Assche 3 participants 16 messages May 17, 2023 – May 29, 2023

Subsystems (3)

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Bart Van Assche

author reviewer
#0 lore raw

v2 cover ufs: Do not requeue while ungating the clock

Hi Martin,

In the traces we recorded while testing zoned storage we noticed that UFS commands are requeued while the clock is being ungated. Command requeueing makes it harder than necessary to preserve the command order. Hence this patch series that modifies the SCSI core and also the UFS driver such that clock ungating does not trigger command requeueing.

Please consider this patch series for the next merge window.

Thanks,

Bart.

Changes compared to v1:
- Dropped patch "scsi: ufs: core: Unexport ufshcd_hold() and ufshcd_release()".
- Removed a ufshcd_scsi_block_requests() / ufshcd_scsi_unblock_requests() pair from patch "scsi: ufs: Ungate the clock synchronously".

Bart Van Assche (4):
scsi: core: Rework scsi_host_block()
scsi: core: Support setting BLK_MQ_F_BLOCKING
scsi: ufs: Enable the BLK_MQ_F_BLOCKING flag
scsi: ufs: Ungate the clock synchronously

 drivers/scsi/scsi_lib.c          | 26 ++++++-----
 drivers/ufs/core/ufs-sysfs.c     |  2 +-
 drivers/ufs/core/ufshcd-crypto.c |  2 +-
 drivers/ufs/core/ufshcd-priv.h   |  2 +-
 drivers/ufs/core/ufshcd.c        | 76 ++++++++++----------------------
 include/scsi/scsi_host.h         |  3 ++
 6 files changed, 45 insertions(+), 66 deletions(-)
author reviewer
#1 lore raw

PATCH v2 1/4 scsi: core: Rework scsi_host_block()

Make scsi_host_block() easier to read by converting it to the widely used early-return style. See also commit f983622ae605 ("scsi: core: Avoid calling synchronize_rcu() for each device in scsi_host_block()").

---
 drivers/scsi/scsi_lib.c | 18 +++++++++++++-----
 1 file changed, 13 insertions(+), 5 deletions(-)
drivers/scsi/scsi_lib.c +13 -5
index b7c569a42aa4..758a57616dd3 100644
@@ -2939,11 +2939,20 @@ scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
 }
 EXPORT_SYMBOL_GPL(scsi_target_unblock);
 
+/**
+ * scsi_host_block - Try to transition all logical units to the SDEV_BLOCK state
+ * @shost: device to block
+ *
+ * Pause SCSI command processing for all logical units associated with the SCSI
+ * host and wait until pending scsi_queue_rq() calls have finished.
+ *
+ * Returns zero if successful or a negative error code upon failure.
+ */
 int
 scsi_host_block(struct Scsi_Host *shost)
 {
 	struct scsi_device *sdev;
-	int ret = 0;
+	int ret;
 
 	/*
 	 * Call scsi_internal_device_block_nowait so we can avoid
@@ -2955,7 +2964,7 @@ scsi_host_block(struct Scsi_Host *shost)
 		mutex_unlock(&sdev->state_mutex);
 		if (ret) {
 			scsi_device_put(sdev);
-			break;
+			return ret;
 		}
 	}
 
@@ -2965,10 +2974,9 @@ scsi_host_block(struct Scsi_Host *shost)
 	 */
 	WARN_ON_ONCE(shost->tag_set.flags & BLK_MQ_F_BLOCKING);
 
-	if (!ret)
-		synchronize_rcu();
+	synchronize_rcu();
 
-	return ret;
+	return 0;
 }
 EXPORT_SYMBOL_GPL(scsi_host_block);
 
Reviewed-by on v2 1/4
#2 lore raw

Re: [PATCH v2 1/4] scsi: core: Rework scsi_host_block()

Quoted 9 lines · Bart Van Assche

Make scsi_host_block() easier to read by converting it to the widely used early-return style. See also commit f983622ae605 ("scsi: core: Avoid calling synchronize_rcu() for each device in scsi_host_block()").

Thanks,
Ming

author reviewer
#3 lore raw

PATCH v2 2/4 scsi: core: Support setting BLK_MQ_F_BLOCKING

In reply to Bart Van Assche

Prepare for adding code in ufshcd_queuecommand() that may sleep. This patch is similar to a patch posted last year by Mike Christie. See also
https://lore.kernel.org/all/[email protected]/

---
 drivers/scsi/scsi_lib.c  | 10 +++-------
 include/scsi/scsi_host.h |  3 +++
 2 files changed, 6 insertions(+), 7 deletions(-)
drivers/scsi/scsi_lib.c +3 -7
index 758a57616dd3..894af68babc2 100644
@@ -1982,6 +1982,8 @@ int scsi_mq_setup_tags(struct Scsi_Host *shost)
 	tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
 	tag_set->flags |=
 		BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
+	if (shost->hostt->queuecommand_may_block)
+		tag_set->flags |= BLK_MQ_F_BLOCKING;
 	tag_set->driver_data = shost;
 	if (shost->host_tagset)
 		tag_set->flags |= BLK_MQ_F_TAG_HCTX_SHARED;
@@ -2968,13 +2970,7 @@ scsi_host_block(struct Scsi_Host *shost)
 		}
 	}
 
-	/*
-	 * SCSI never enables blk-mq's BLK_MQ_F_BLOCKING flag so
-	 * calling synchronize_rcu() once is enough.
-	 */
-	WARN_ON_ONCE(shost->tag_set.flags & BLK_MQ_F_BLOCKING);
-
-	synchronize_rcu();
+	blk_mq_wait_quiesce_done(&shost->tag_set);
 
 	return 0;
 }
include/scsi/scsi_host.h +3 -0
index 0f29799efa02..37a8a2608dc2 100644
@@ -458,6 +458,9 @@ struct scsi_host_template {
 	/* True if the host uses host-wide tagspace */
 	unsigned host_tagset:1;
 
+	/* The queuecommand callback may block. See also BLK_MQ_F_BLOCKING. */
+	unsigned queuecommand_may_block:1;
+
 	/*
 	 * Countdown for host blocking with no commands outstanding.
 	 */
author reviewer
#4 lore raw

PATCH v2 3/4 scsi: ufs: Enable the BLK_MQ_F_BLOCKING flag

In reply to Bart Van Assche

Prepare for adding code in ufshcd_queuecommand() that may sleep.

---
 drivers/ufs/core/ufshcd.c | 1 +
 1 file changed, 1 insertion(+)
drivers/ufs/core/ufshcd.c +1 -0
index 7ee150d67d49..993034ac1696 100644
@@ -8756,6 +8756,7 @@ static const struct scsi_host_template ufshcd_driver_template = {
 	.max_host_blocked	= 1,
 	.track_queue_depth	= 1,
 	.skip_settle_delay	= 1,
+	.queuecommand_may_block = true,
 	.sdev_groups		= ufshcd_driver_groups,
 	.rpm_autosuspend_delay	= RPM_AUTOSUSPEND_DELAY_MS,
 };
#5 lore raw

Re: [PATCH v2 3/4] scsi: ufs: Enable the BLK_MQ_F_BLOCKING flag

Quoted 12 lines · Bart Van Assche

Prepare for adding code in ufshcd_queuecommand() that may sleep.

---
 drivers/ufs/core/ufshcd.c | 1 +
 1 file changed, 1 insertion(+)
drivers/ufs/core/ufshcd.c +0 -0
index 7ee150d67d49..993034ac1696 100644

drivers/ufs/core/ufshcd.c static const struct scsi_host_template ufshcd_driver_template = { · line 8759

Quoted 4 lines
drivers/ufs/core/ufshcd.c +1 -0
@@ -8756,6 +8756,7 @@ static const struct scsi_host_template ufshcd_driver_template = {
 	.max_host_blocked	= 1,
 	.track_queue_depth	= 1,
 	.skip_settle_delay	= 1,
+	.queuecommand_may_block = true,

Shouldn't this only be for controllers that support clock gating?

Quoted 4 lines
drivers/ufs/core/ufshcd.c +0 -0
 	.sdev_groups		= ufshcd_driver_groups,
 	.rpm_autosuspend_delay	= RPM_AUTOSUSPEND_DELAY_MS,
 };
author reviewer
#6 lore raw

Re: [PATCH v2 3/4] scsi: ufs: Enable the BLK_MQ_F_BLOCKING flag

Quoted 8 lines · Adrian Hunter
Quoted 7 lines · Bart Van Assche

Prepare for adding code in ufshcd_queuecommand() that may sleep.

---
  drivers/ufs/core/ufshcd.c | 1 +
  1 file changed, 1 insertion(+)
Quoted 12 lines

drivers/ufs/core/ufshcd.c static const struct scsi_host_template ufshcd_driver_template = { · line 8759

Quoted 9 lines
drivers/ufs/core/ufshcd.c +1 -0
index 7ee150d67d49..993034ac1696 100644
@@ -8756,6 +8756,7 @@ static const struct scsi_host_template ufshcd_driver_template = {
  	.max_host_blocked	= 1,
  	.track_queue_depth	= 1,
  	.skip_settle_delay	= 1,
+	.queuecommand_may_block = true,

Shouldn't this only be for controllers that support clock gating?

Hi Adrian,

The overhead of BLK_MQ_F_BLOCKING is small relative to the time required to queue a UFS command so I think enabling BLK_MQ_F_BLOCKING for all UFS host controllers is fine. BLK_MQ_F_BLOCKING causes the block layer to use SRCU instead of RCU. The cost of the sleepable RCU primitives is dominated by the memory barrier (smp_mb()) in the srcu_read_lock() and srcu_read_unlock() calls. From kernel/rcu/srcutree.c:

int __srcu_read_lock(struct srcu_struct *ssp)

{
	int idx;

	idx = READ_ONCE(ssp->srcu_idx) & 0x1;
	this_cpu_inc(ssp->sda->srcu_lock_count[idx].counter);
	smp_mb(); /* B */  /* Avoid leaking the critical section. */
	return idx;
}

EXPORT_SYMBOL_GPL(__srcu_read_lock);

void __srcu_read_unlock(struct srcu_struct *ssp, int idx)

{
	smp_mb(); /* C */  /* Avoid leaking the critical section. */
	this_cpu_inc(ssp->sda->srcu_unlock_count[idx].counter);
}

EXPORT_SYMBOL_GPL(__srcu_read_unlock);

The rcu_read_lock() and rcu_read_lock() implementations do not call smp_mb() as one can see in include/linux/rcupdate.h:

static inline void __rcu_read_lock(void)

{
	preempt_disable();
}

static inline void __rcu_read_unlock(void)

{
	preempt_enable();
	if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
		rcu_read_unlock_strict();
}

Thanks,

Bart.

#7 lore raw

Re: [PATCH v2 3/4] scsi: ufs: Enable the BLK_MQ_F_BLOCKING flag

Quoted 22 lines · Bart Van Assche
Quoted 8 lines · Adrian Hunter
Quoted 7 lines · Bart Van Assche

Prepare for adding code in ufshcd_queuecommand() that may sleep.

---
  drivers/ufs/core/ufshcd.c | 1 +

  1 file changed, 1 insertion(+)
Quoted 12 lines

drivers/ufs/core/ufshcd.c static const struct scsi_host_template ufshcd_driver_template = { · line 8759

Quoted 9 lines
drivers/ufs/core/ufshcd.c +1 -0
index 7ee150d67d49..993034ac1696 100644
@@ -8756,6 +8756,7 @@ static const struct scsi_host_template ufshcd_driver_template = {
       .max_host_blocked    = 1,
       .track_queue_depth    = 1,
       .skip_settle_delay    = 1,
+    .queuecommand_may_block = true,

Shouldn't this only be for controllers that support clock gating?

Quoted 5 lines

Hi Adrian,

The overhead of BLK_MQ_F_BLOCKING is small relative to the time required to queue a UFS command so I think enabling BLK_MQ_F_BLOCKING for all UFS host controllers is fine.

Doesn't it also force the queue to be run asynchronously always?

But in any case, it doesn't seem like something to force on drivers just because it would take a bit more coding to make it optional.

Quoted 42 lines

BLK_MQ_F_BLOCKING causes the block layer to use SRCU
instead of RCU. The cost of the sleepable RCU primitives is dominated by the memory barrier (smp_mb()) in the srcu_read_lock() and srcu_read_unlock() calls. From kernel/rcu/srcutree.c:

int __srcu_read_lock(struct srcu_struct *ssp)
{
    int idx;

    idx = READ_ONCE(ssp->srcu_idx) & 0x1; this_cpu_inc(ssp->sda->srcu_lock_count[idx].counter); smp_mb(); /* B */  /* Avoid leaking the critical section. */ return idx;
}
EXPORT_SYMBOL_GPL(__srcu_read_lock);

void __srcu_read_unlock(struct srcu_struct *ssp, int idx) {
    smp_mb(); /* C */  /* Avoid leaking the critical section. */ this_cpu_inc(ssp->sda->srcu_unlock_count[idx].counter); }
EXPORT_SYMBOL_GPL(__srcu_read_unlock);

The rcu_read_lock() and rcu_read_lock() implementations do not call smp_mb() as one can see in include/linux/rcupdate.h:

static inline void __rcu_read_lock(void)
{
    preempt_disable();
}

static inline void __rcu_read_unlock(void) {
    preempt_enable();
    if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) rcu_read_unlock_strict();
}

Thanks,

Bart.

author reviewer
#8 lore raw

Re: [PATCH v2 3/4] scsi: ufs: Enable the BLK_MQ_F_BLOCKING flag

Quoted 9 lines · Adrian Hunter
Quoted 3 lines · Bart Van Assche

The overhead of BLK_MQ_F_BLOCKING is small relative to the time required to queue a UFS command so I think enabling BLK_MQ_F_BLOCKING for all UFS host controllers is fine.

Doesn't it also force the queue to be run asynchronously always?

But in any case, it doesn't seem like something to force on drivers just because it would take a bit more coding to make it optional.

Making BLK_MQ_F_BLOCKING optional would complicate testing of the UFS driver. Although it is possible to make BLK_MQ_F_BLOCKING optional, I'm wondering whether it is worth it? I haven't noticed any performance difference in my tests with BLK_MQ_F_BLOCKING enabled compared to BLK_MQ_F_BLOCKING disabled.

Thanks,

Bart.

#9 lore raw

Re: [PATCH v2 3/4] scsi: ufs: Enable the BLK_MQ_F_BLOCKING flag

Quoted 11 lines · Bart Van Assche
Quoted 9 lines · Adrian Hunter
Quoted 3 lines · Bart Van Assche

The overhead of BLK_MQ_F_BLOCKING is small relative to the time required to queue a UFS command so I think enabling BLK_MQ_F_BLOCKING for all UFS host controllers is fine.

Doesn't it also force the queue to be run asynchronously always?

But in any case, it doesn't seem like something to force on drivers just because it would take a bit more coding to make it optional.

Making BLK_MQ_F_BLOCKING optional would complicate testing of the UFS driver. Although it is possible to make BLK_MQ_F_BLOCKING optional, I'm wondering whether it is worth it? I haven't noticed any performance difference in my tests with BLK_MQ_F_BLOCKING enabled compared to BLK_MQ_F_BLOCKING disabled.

It is hard to know the effects, especially wrt to future hardware.

What about something like this?

drivers/ufs/core/ufshcd.c +11 -5
index 82321b8b32bc..301c0b2a406d 100644
@@ -10049,13 +10049,16 @@ static int ufshcd_set_dma_mask(struct ufs_hba *hba)
 }
 
 /**
- * ufshcd_alloc_host - allocate Host Bus Adapter (HBA)
+ * __ufshcd_alloc_host - allocate Host Bus Adapter (HBA)
  * @dev: pointer to device handle
  * @hba_handle: driver private handle
+ * @vops: pointer to variant specific operations
  * Returns 0 on success, non-zero value on failure
  */
-int ufshcd_alloc_host(struct device *dev, struct ufs_hba **hba_handle)
+int __ufshcd_alloc_host(struct device *dev, struct ufs_hba **hba_handle,
+			const struct ufs_hba_variant_ops *vops)
 {
+	struct scsi_host_template sht = ufshcd_driver_template;
 	struct Scsi_Host *host;
 	struct ufs_hba *hba;
 	int err = 0;
@@ -10067,8 +10070,10 @@ int ufshcd_alloc_host(struct device *dev, struct ufs_hba **hba_handle)
 		goto out_error;
 	}
 
-	host = scsi_host_alloc(&ufshcd_driver_template,
-				sizeof(struct ufs_hba));
+	if (vops && vops->nonblocking)
+		sht.queuecommand_may_block = false;
+
+	host = scsi_host_alloc(&sht, sizeof(struct ufs_hba));
 	if (!host) {
 		dev_err(dev, "scsi_host_alloc failed\n");
 		err = -ENOMEM;
@@ -10078,6 +10083,7 @@ int ufshcd_alloc_host(struct device *dev, struct ufs_hba **hba_handle)
 	hba = shost_priv(host);
 	hba->host = host;
 	hba->dev = dev;
+	hba->vops = vops;
 	hba->dev_ref_clk_freq = REF_CLK_FREQ_INVAL;
 	hba->nop_out_timeout = NOP_OUT_TIMEOUT;
 	ufshcd_set_sg_entry_size(hba, sizeof(struct ufshcd_sg_entry));
@@ -10089,7 +10095,7 @@ int ufshcd_alloc_host(struct device *dev, struct ufs_hba **hba_handle)
 out_error:
 	return err;
 }
-EXPORT_SYMBOL(ufshcd_alloc_host);
+EXPORT_SYMBOL(__ufshcd_alloc_host);
 
 /* This function exists because blk_mq_alloc_tag_set() requires this. */
 static blk_status_t ufshcd_queue_tmf(struct blk_mq_hw_ctx *hctx,
drivers/ufs/host/ufshcd-pci.c +9 -3
index 9c911787f84c..ce42b822fcb6 100644
@@ -447,6 +447,7 @@ static int ufs_intel_mtl_init(struct ufs_hba *hba)
 
 static struct ufs_hba_variant_ops ufs_intel_cnl_hba_vops = {
 	.name                   = "intel-pci",
+	.nonblocking		= true,
 	.init			= ufs_intel_common_init,
 	.exit			= ufs_intel_common_exit,
 	.link_startup_notify	= ufs_intel_link_startup_notify,
@@ -455,6 +456,7 @@ static struct ufs_hba_variant_ops ufs_intel_cnl_hba_vops = {
 
 static struct ufs_hba_variant_ops ufs_intel_ehl_hba_vops = {
 	.name                   = "intel-pci",
+	.nonblocking		= true,
 	.init			= ufs_intel_ehl_init,
 	.exit			= ufs_intel_common_exit,
 	.link_startup_notify	= ufs_intel_link_startup_notify,
@@ -463,6 +465,7 @@ static struct ufs_hba_variant_ops ufs_intel_ehl_hba_vops = {
 
 static struct ufs_hba_variant_ops ufs_intel_lkf_hba_vops = {
 	.name                   = "intel-pci",
+	.nonblocking		= true,
 	.init			= ufs_intel_lkf_init,
 	.exit			= ufs_intel_common_exit,
 	.hce_enable_notify	= ufs_intel_hce_enable_notify,
@@ -475,6 +478,7 @@ static struct ufs_hba_variant_ops ufs_intel_lkf_hba_vops = {
 
 static struct ufs_hba_variant_ops ufs_intel_adl_hba_vops = {
 	.name			= "intel-pci",
+	.nonblocking		= true,
 	.init			= ufs_intel_adl_init,
 	.exit			= ufs_intel_common_exit,
 	.link_startup_notify	= ufs_intel_link_startup_notify,
@@ -484,6 +488,7 @@ static struct ufs_hba_variant_ops ufs_intel_adl_hba_vops = {
 
 static struct ufs_hba_variant_ops ufs_intel_mtl_hba_vops = {
 	.name                   = "intel-pci",
+	.nonblocking		= true,
 	.init			= ufs_intel_mtl_init,
 	.exit			= ufs_intel_common_exit,
 	.hce_enable_notify	= ufs_intel_hce_enable_notify,
@@ -538,6 +543,7 @@ static void ufshcd_pci_remove(struct pci_dev *pdev)
 static int
 ufshcd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 {
+	const struct ufs_hba_variant_ops *vops;
 	struct ufs_host *ufs_host;
 	struct ufs_hba *hba;
 	void __iomem *mmio_base;
@@ -559,14 +565,14 @@ ufshcd_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 
 	mmio_base = pcim_iomap_table(pdev)[0];
 
-	err = ufshcd_alloc_host(&pdev->dev, &hba);
+	vops = (const struct ufs_hba_variant_ops *)id->driver_data;
+
+	err = __ufshcd_alloc_host(&pdev->dev, &hba, vops);
 	if (err) {
 		dev_err(&pdev->dev, "Allocation failed\n");
 		return err;
 	}
 
-	hba->vops = (struct ufs_hba_variant_ops *)id->driver_data;
-
 	err = ufshcd_init(hba, mmio_base, pdev->irq);
 	if (err) {
 		dev_err(&pdev->dev, "Initialization failed\n");
include/ufs/ufshcd.h +17 -1
index 8039c2b72502..dbf2312ad756 100644
@@ -274,6 +274,8 @@ struct ufs_pwr_mode_info {
 /**
  * struct ufs_hba_variant_ops - variant specific callbacks
  * @name: variant name
+ * @nonblocking: queuecommand will not block (incompatible with
+ *               UFSHCD_CAP_CLK_GATING)
  * @init: called when the driver is initialized
  * @exit: called to cleanup everything done in init
  * @get_ufs_hci_version: called to get UFS HCI version
@@ -310,6 +312,7 @@ struct ufs_pwr_mode_info {
  */
 struct ufs_hba_variant_ops {
 	const char *name;
+	bool nonblocking;
 	int	(*init)(struct ufs_hba *);
 	void    (*exit)(struct ufs_hba *);
 	u32	(*get_ufs_hci_version)(struct ufs_hba *);
@@ -1225,7 +1228,20 @@ static inline void ufshcd_rmwl(struct ufs_hba *hba, u32 mask, u32 val, u32 reg)
 	ufshcd_writel(hba, tmp, reg);
 }
 
-int ufshcd_alloc_host(struct device *, struct ufs_hba **);
+int __ufshcd_alloc_host(struct device *, struct ufs_hba **,
+			const struct ufs_hba_variant_ops *);
+
+/**
+ * ufshcd_alloc_host - allocate Host Bus Adapter (HBA)
+ * @dev: pointer to device handle
+ * @hba_handle: driver private handle
+ * Returns 0 on success, non-zero value on failure
+ */
+static inline int ufshcd_alloc_host(struct device *dev, struct ufs_hba **hba_handle)
+{
+	return __ufshcd_alloc_host(dev, hba_handle, NULL);
+}
+
 void ufshcd_dealloc_host(struct ufs_hba *);
 int ufshcd_hba_enable(struct ufs_hba *hba);
 int ufshcd_init(struct ufs_hba *, void __iomem *, unsigned int);
author reviewer
#10 lore raw

Re: [PATCH v2 3/4] scsi: ufs: Enable the BLK_MQ_F_BLOCKING flag

Quoted 6 lines · Adrian Hunter
Quoted 5 lines · Bart Van Assche
Quoted 4 lines · Adrian Hunter
Quoted 2 lines · Bart Van Assche

The overhead of BLK_MQ_F_BLOCKING is small relative to the
time required to queue a UFS command so I think enabling BLK_MQ_F_BLOCKING for all UFS host controllers is fine.

Quoted 7 lines
Quoted 5 lines
Quoted 3 lines

Doesn't it also force the queue to be run asynchronously always?

But in any case, it doesn't seem like something to force on drivers just because it would take a bit more coding to make it optional.

Making BLK_MQ_F_BLOCKING optional would complicate testing of the UFS driver. Although it is possible to make BLK_MQ_F_BLOCKING optional, I'm wondering whether it is worth it? I haven't noticed any performance difference in my tests with BLK_MQ_F_BLOCKING enabled compared to BLK_MQ_F_BLOCKING disabled.

It is hard to know the effects, especially wrt to future hardware.

Are you perhaps referring to performance effects? I think the block layer can be modified to run the queue synchronously in most cases even if BLK_MQ_F_BLOCKING is set. The patch below works fine but is probably not acceptable for upstream because it uses in_atomic():

block/blk-mq.c +15 -1
index 551e7760f45e..00fbe0aa56b5 100644
@@ -1471,9 +1471,23 @@ static void blk_mq_requeue_work(struct work_struct *work)
 	blk_mq_run_hw_queues(q, false);
 }
 
+static bool may_sleep(void)
+{
+#ifdef CONFIG_PREEMPT_COUNT
+	return !in_atomic() && !irqs_disabled();
+#else
+	return false;
+#endif
+}
+
 void blk_mq_kick_requeue_list(struct request_queue *q)
 {
+	if (may_sleep()) {
+		blk_mq_requeue_work(&q->requeue_work.work);
+		return;
+	}
 	kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
+
 }
 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 
@@ -2228,7 +2242,7 @@ void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 	if (!need_run)
 		return;
 
-	if (async || (hctx->flags & BLK_MQ_F_BLOCKING) ||
+	if (async || (hctx->flags & BLK_MQ_F_BLOCKING && !may_sleep()) ||
 	    !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
 		blk_mq_delay_run_hw_queue(hctx, 0);
 		return;
Quoted 1 line

What about something like this? [ ... ]

This introduces a redundancy and a potential for a conflict between the "nonblocking" flag and the UFSHCD_CAP_CLK_GATING flag. It is unfortunate that hba->caps is set so late otherwise we could use the result of (hba->caps & UFSHCD_CAP_CLK_GATING) to check whether or not BLK_MQ_F_BLOCKING is needed.

Additionally, this patch introduces a use-after-free issue since it causes scsi_host_alloc() to store a pointer to a stack variable (sht) into a SCSI host structure.

Bart.

#11 lore raw

Re: [PATCH v2 3/4] scsi: ufs: Enable the BLK_MQ_F_BLOCKING flag

Quoted 11 lines · Bart Van Assche
Quoted 10 lines · Adrian Hunter
Quoted 9 lines · Bart Van Assche
Quoted 7 lines · Adrian Hunter
Quoted 2 lines · Bart Van Assche

The overhead of BLK_MQ_F_BLOCKING is small relative to the
time required to queue a UFS command so I think enabling BLK_MQ_F_BLOCKING for all UFS host controllers is fine.

Doesn't it also force the queue to be run asynchronously always?

But in any case, it doesn't seem like something to force on drivers just because it would take a bit more coding to make it optional.

Quoted 8 lines
Quoted 3 lines
Quoted 1 line

Making BLK_MQ_F_BLOCKING optional would complicate testing of the UFS driver. Although it is possible to make BLK_MQ_F_BLOCKING optional, I'm wondering whether it is worth it? I haven't noticed any performance difference in my tests with BLK_MQ_F_BLOCKING enabled compared to BLK_MQ_F_BLOCKING disabled.

It is hard to know the effects, especially wrt to future hardware.

Are you perhaps referring to performance effects? I think the block layer can be modified to run the queue synchronously in most cases even if BLK_MQ_F_BLOCKING is set. The patch below works fine but is probably not acceptable for upstream because it uses in_atomic(): [ ... ]

Why would we want to when we don't have to?

Quoted 11 lines
Quoted 1 line

What about something like this? [ ... ]

This introduces a redundancy and a potential for a conflict between the "nonblocking" flag and the UFSHCD_CAP_CLK_GATING flag. It is unfortunate that hba->caps is set so late otherwise we could use the result of (hba->caps & UFSHCD_CAP_CLK_GATING) to check whether or not BLK_MQ_F_BLOCKING is needed.

Additionally, this patch introduces a use-after-free issue since it causes scsi_host_alloc() to store a pointer to a stack variable (sht) into a SCSI host structure.

So with those fixed and the vops->may_block instead of vops->nonblocking:

drivers/ufs/core/ufshcd.c +52 -31
index 82321b8b32bc..6a4389f02b4e 100644
@@ -2027,6 +2027,11 @@ static void ufshcd_init_clk_gating(struct ufs_hba *hba)
 	if (!ufshcd_is_clkgating_allowed(hba))
 		return;
 
+	if (!hba->host->hostt->queuecommand_may_block) {
+		dev_warn(hba->dev, "Nonblocking queue, so clock gating is not enabled!\n");
+		return;
+	}
+
 	hba->clk_gating.state = CLKS_ON;
 
 	hba->clk_gating.delay_ms = 150;
@@ -8708,33 +8713,41 @@ static struct ufs_hba_variant_params ufs_hba_vps = {
 	.ondemand_data.downdifferential	= 5,
 };
 
-static const struct scsi_host_template ufshcd_driver_template = {
-	.module			= THIS_MODULE,
-	.name			= UFSHCD,
-	.proc_name		= UFSHCD,
-	.map_queues		= ufshcd_map_queues,
-	.queuecommand		= ufshcd_queuecommand,
-	.mq_poll		= ufshcd_poll,
-	.slave_alloc		= ufshcd_slave_alloc,
-	.slave_configure	= ufshcd_slave_configure,
-	.slave_destroy		= ufshcd_slave_destroy,
-	.change_queue_depth	= ufshcd_change_queue_depth,
-	.eh_abort_handler	= ufshcd_abort,
-	.eh_device_reset_handler = ufshcd_eh_device_reset_handler,
-	.eh_host_reset_handler   = ufshcd_eh_host_reset_handler,
-	.eh_timed_out		= ufshcd_eh_timed_out,
-	.this_id		= -1,
-	.sg_tablesize		= SG_ALL,
-	.cmd_per_lun		= UFSHCD_CMD_PER_LUN,
-	.can_queue		= UFSHCD_CAN_QUEUE,
-	.max_segment_size	= PRDT_DATA_BYTE_COUNT_MAX,
-	.max_sectors		= (1 << 20) / SECTOR_SIZE, /* 1 MiB */
-	.max_host_blocked	= 1,
-	.track_queue_depth	= 1,
-	.skip_settle_delay	= 1,
+#define UFSHCD_SHT_COMMON \
+	.module			= THIS_MODULE,				\
+	.name			= UFSHCD,				\
+	.proc_name		= UFSHCD,				\
+	.map_queues		= ufshcd_map_queues,			\
+	.queuecommand		= ufshcd_queuecommand,			\
+	.mq_poll		= ufshcd_poll,				\
+	.slave_alloc		= ufshcd_slave_alloc,			\
+	.slave_configure	= ufshcd_slave_configure,		\
+	.slave_destroy		= ufshcd_slave_destroy,			\
+	.change_queue_depth	= ufshcd_change_queue_depth,		\
+	.eh_abort_handler	= ufshcd_abort,				\
+	.eh_device_reset_handler = ufshcd_eh_device_reset_handler,	\
+	.eh_host_reset_handler   = ufshcd_eh_host_reset_handler,	\
+	.eh_timed_out		= ufshcd_eh_timed_out,			\
+	.this_id		= -1,					\
+	.sg_tablesize		= SG_ALL,				\
+	.cmd_per_lun		= UFSHCD_CMD_PER_LUN,			\
+	.can_queue		= UFSHCD_CAN_QUEUE,			\
+	.max_segment_size	= PRDT_DATA_BYTE_COUNT_MAX,		\
+	.max_sectors		= (1 << 20) / SECTOR_SIZE, /* 1 MiB */	\
+	.max_host_blocked	= 1,					\
+	.track_queue_depth	= 1,					\
+	.skip_settle_delay	= 1,					\
+	.sdev_groups		= ufshcd_driver_groups,			\
+	.rpm_autosuspend_delay	= RPM_AUTOSUSPEND_DELAY_MS
+
+static const struct scsi_host_template ufshcd_sht_blocking = {
+	UFSHCD_SHT_COMMON,
 	.queuecommand_may_block = true,
-	.sdev_groups		= ufshcd_driver_groups,
-	.rpm_autosuspend_delay	= RPM_AUTOSUSPEND_DELAY_MS,
+};
+
+static const struct scsi_host_template ufshcd_sht_nonblocking = {
+	UFSHCD_SHT_COMMON,
+	.queuecommand_may_block = false,
 };
 
 static int ufshcd_config_vreg_load(struct device *dev, struct ufs_vreg *vreg,
@@ -10049,13 +10062,16 @@ static int ufshcd_set_dma_mask(struct ufs_hba *hba)
 }
 
 /**
- * ufshcd_alloc_host - allocate Host Bus Adapter (HBA)
+ * __ufshcd_alloc_host - allocate Host Bus Adapter (HBA)
  * @dev: pointer to device handle
  * @hba_handle: driver private handle
+ * @vops: pointer to variant specific operations
  * Returns 0 on success, non-zero value on failure
  */
-int ufshcd_alloc_host(struct device *dev, struct ufs_hba **hba_handle)
+int __ufshcd_alloc_host(struct device *dev, struct ufs_hba **hba_handle,
+			const struct ufs_hba_variant_ops *vops)
 {
+	const struct scsi_host_template *sht;
 	struct Scsi_Host *host;
 	struct ufs_hba *hba;
 	int err = 0;
@@ -10067,8 +10083,12 @@ int ufshcd_alloc_host(struct device *dev, struct ufs_hba **hba_handle)
 		goto out_error;
 	}
 
-	host = scsi_host_alloc(&ufshcd_driver_template,
-				sizeof(struct ufs_hba));
+	if (vops && vops->may_block)
+		sht = &ufshcd_sht_blocking;
+	else
+		sht = &ufshcd_sht_nonblocking;
+
+	host = scsi_host_alloc(sht, sizeof(struct ufs_hba));
 	if (!host) {
 		dev_err(dev, "scsi_host_alloc failed\n");
 		err = -ENOMEM;
@@ -10078,6 +10098,7 @@ int ufshcd_alloc_host(struct device *dev, struct ufs_hba **hba_handle)
 	hba = shost_priv(host);
 	hba->host = host;
 	hba->dev = dev;
+	hba->vops = vops;
 	hba->dev_ref_clk_freq = REF_CLK_FREQ_INVAL;
 	hba->nop_out_timeout = NOP_OUT_TIMEOUT;
 	ufshcd_set_sg_entry_size(hba, sizeof(struct ufshcd_sg_entry));
@@ -10089,7 +10110,7 @@ int ufshcd_alloc_host(struct device *dev, struct ufs_hba **hba_handle)
 out_error:
 	return err;
 }
-EXPORT_SYMBOL(ufshcd_alloc_host);
+EXPORT_SYMBOL(__ufshcd_alloc_host);
 
 /* This function exists because blk_mq_alloc_tag_set() requires this. */
 static blk_status_t ufshcd_queue_tmf(struct blk_mq_hw_ctx *hctx,
include/ufs/ufshcd.h +16 -1
index 8039c2b72502..dbeb879b47d1 100644
@@ -274,6 +274,7 @@ struct ufs_pwr_mode_info {
 /**
  * struct ufs_hba_variant_ops - variant specific callbacks
  * @name: variant name
+ * @may_block: queuecommand may block (required with UFSHCD_CAP_CLK_GATING)
  * @init: called when the driver is initialized
  * @exit: called to cleanup everything done in init
  * @get_ufs_hci_version: called to get UFS HCI version
@@ -310,6 +311,7 @@ struct ufs_pwr_mode_info {
  */
 struct ufs_hba_variant_ops {
 	const char *name;
+	bool	may_block;
 	int	(*init)(struct ufs_hba *);
 	void    (*exit)(struct ufs_hba *);
 	u32	(*get_ufs_hci_version)(struct ufs_hba *);
@@ -1225,7 +1227,20 @@ static inline void ufshcd_rmwl(struct ufs_hba *hba, u32 mask, u32 val, u32 reg)
 	ufshcd_writel(hba, tmp, reg);
 }
 
-int ufshcd_alloc_host(struct device *, struct ufs_hba **);
+int __ufshcd_alloc_host(struct device *, struct ufs_hba **,
+			const struct ufs_hba_variant_ops *);
+
+/**
+ * ufshcd_alloc_host - allocate Host Bus Adapter (HBA)
+ * @dev: pointer to device handle
+ * @hba_handle: driver private handle
+ * Returns 0 on success, non-zero value on failure
+ */
+static inline int ufshcd_alloc_host(struct device *dev, struct ufs_hba **hba_handle)
+{
+	return __ufshcd_alloc_host(dev, hba_handle, NULL);
+}
+
 void ufshcd_dealloc_host(struct ufs_hba *);
 int ufshcd_hba_enable(struct ufs_hba *hba);
 int ufshcd_init(struct ufs_hba *, void __iomem *, unsigned int);
author reviewer
#12 lore raw

Re: [PATCH v2 3/4] scsi: ufs: Enable the BLK_MQ_F_BLOCKING flag

Quoted 1 line · Adrian Hunter

Why would we want to when we don't have to?

Anyone is allowed to formulate objections to a patch but when objecting to a patch or an approach please explain *why*. I think I already explained that unconditionally enabling BLK_MQ_F_BLOCKING makes testing the UFS host controller driver easier.

Quoted 1 line

So with those fixed and the vops->may_block instead of vops->nonblocking:

[ ... ]

I think a much simpler solution exists. I plan to post a new version of this patch series soon.

Bart.

#13 lore raw

Re: [PATCH v2 3/4] scsi: ufs: Enable the BLK_MQ_F_BLOCKING flag

Quoted 4 lines · Bart Van Assche
Quoted 1 line · Adrian Hunter

Why would we want to when we don't have to?

Anyone is allowed to formulate objections to a patch but when objecting to a patch or an approach please explain *why*. I think I already explained that unconditionally enabling BLK_MQ_F_BLOCKING makes testing the UFS host controller driver easier.

I don't think you are claiming blocking queues are better for block drivers than non-blocking queues, so it is quite reasonable to stay with non-blocking queues.

I am not sure testing is really relevant to this discussion, but it seems like different combinations need to be tested anyway e.g. testing with / without clock gating is the same testing with / without blocking.

Quoted 6 lines
Quoted 2 lines

So with those fixed and the vops->may_block instead of vops->nonblocking:

[ ... ]

I think a much simpler solution exists. I plan to post a new version of this patch series soon.

Thank you!

author reviewer
#14 lore raw

PATCH v2 4/4 scsi: ufs: Ungate the clock synchronously

In reply to Bart Van Assche

Ungating the clock asynchronously causes ufshcd_queuecommand() to return SCSI_MLQUEUE_HOST_BUSY and hence causes commands to be requeued. This is suboptimal. Allow ufshcd_queuecommand() to sleep such that clock ungating does not trigger command requeuing. Remove the ufshcd_scsi_block_requests() and ufshcd_scsi_unblock_requests() because these are no longer needed. The flush_work(&hba->clk_gating.ungate_work) call is sufficient to make the SCSI core wait for clock ungating to complete.

---
 drivers/ufs/core/ufs-sysfs.c     |  2 +-
 drivers/ufs/core/ufshcd-crypto.c |  2 +-
 drivers/ufs/core/ufshcd-priv.h   |  2 +-
 drivers/ufs/core/ufshcd.c        | 84 ++++++++++----------------------
 include/ufs/ufshcd.h             |  2 +-
 5 files changed, 30 insertions(+), 62 deletions(-)
drivers/ufs/core/ufs-sysfs.c +1 -1
index 883f0e44b54e..cdf3d5f2b77b 100644
@@ -168,7 +168,7 @@ static ssize_t auto_hibern8_show(struct device *dev,
 	}
 
 	pm_runtime_get_sync(hba->dev);
-	ufshcd_hold(hba, false);
+	ufshcd_hold(hba);
 	ahit = ufshcd_readl(hba, REG_AUTO_HIBERNATE_IDLE_TIMER);
 	ufshcd_release(hba);
 	pm_runtime_put_sync(hba->dev);
drivers/ufs/core/ufshcd-crypto.c +1 -1
index 198360fe5e8e..f2c4422cab86 100644
@@ -24,7 +24,7 @@ static int ufshcd_program_key(struct ufs_hba *hba,
 	u32 slot_offset = hba->crypto_cfg_register + slot * sizeof(*cfg);
 	int err = 0;
 
-	ufshcd_hold(hba, false);
+	ufshcd_hold(hba);
 
 	if (hba->vops && hba->vops->program_key) {
 		err = hba->vops->program_key(hba, cfg, slot);
drivers/ufs/core/ufshcd-priv.h +1 -1
index d53b93c21a0c..22cac71090ae 100644
@@ -84,7 +84,7 @@ unsigned long ufshcd_mcq_poll_cqe_lock(struct ufs_hba *hba,
 int ufshcd_read_string_desc(struct ufs_hba *hba, u8 desc_index,
 			    u8 **buf, bool ascii);
 
-int ufshcd_hold(struct ufs_hba *hba, bool async);
+void ufshcd_hold(struct ufs_hba *hba);
 void ufshcd_release(struct ufs_hba *hba);
 
 int ufshcd_send_uic_cmd(struct ufs_hba *hba, struct uic_command *uic_cmd);
drivers/ufs/core/ufshcd.c +26 -58

268 lines not on this page · raw message

include/ufs/ufshcd.h +1 -1
index db2e669985d5..40c537a3880b 100644
@@ -1357,7 +1357,7 @@ void ufshcd_fixup_dev_quirks(struct ufs_hba *hba,
 int ufshcd_read_string_desc(struct ufs_hba *hba, u8 desc_index,
 			    u8 **buf, bool ascii);
 
-int ufshcd_hold(struct ufs_hba *hba, bool async);
+void ufshcd_hold(struct ufs_hba *hba);
 void ufshcd_release(struct ufs_hba *hba);
 
 void ufshcd_clkgate_delay_set(struct device *dev, unsigned long value);
author reviewer
#15 lore raw
In reply to Bart Van Assche
Quoted 5 lines · Bart Van Assche

In the traces we recorded while testing zoned storage we noticed that UFS commands are requeued while the clock is being ungated. Command requeueing makes it harder than necessary to preserve the command order. Hence this patch series that modifies the SCSI core and also the UFS driver such that clock ungating does not trigger command requeueing.

Can anyone please help with reviewing this patch series? I will be OoO in June so it would help if review comments would be posted soon.

Thanks,

Bart.