diff mbox series

[04/14] scsi: qla2xxx: qla82xx: Remove in_interrupt().

Message ID 20201126132952.2287996-5-bigeasy@linutronix.de (mailing list archive)
State Accepted
Headers show
Series scsi: Remove in_interrupt() usage. | expand

Commit Message

Sebastian Andrzej Siewior Nov. 26, 2020, 1:29 p.m. UTC
From: "Ahmed S. Darwish" <a.darwish@linutronix.de>

qla82xx_idc_lock() spins on a certain hardware state until it's updated. At
the end of each spin, if in_interrupt() is true, it does 20 loops of
cpu_relax(). Otherwise, it yields the CPU.

While in_interrupt() is ill-defined and does not provide what the name
suggests, it is not needed here: qla82xx_idc_lock() is always called
from process context. Below is an analysis of its callers, in order of
appearance:

  - qla_nx.c: qla82xx_device_bootstrap(), only called by
    qla82xx_device_state_handler(), has multiple msleep()s.

  - qla_nx.c: qla82xx_need_qsnt_handler(), has one second msleep()

  - qla_nx.c: qla82xx_wait_for_state_change(), one second msleep()

  - qla_nx.c: qla82xx_need_reset_handler(), can sleep up to 10 seconds

  - qla_nx.c: qla82xx_device_state_handler(), has multiple msleep()s

  - qla_nx.c: qla82xx_abort_isp(), if it's a qla82xx controller, calls
    qla82xx_device_state_handler(), which sleeps. It's also bound to
    isp_operations ->abort_isp() hook, where all the callers are in
    process context.

  - qla_nx.c: qla82xx_beacon_on(), bound to isp_operations ->beacon_on()
    hook.  That hook is only called once, in a mutex locked context,
    from qla2x00_beacon_store().

  - qla_nx.c: qla82xx_beacon_off(), bound to isp_operations
    ->beacon_off() hook.  Like ->beacon_on(), it's only called once, in
    a mutex locked context, from qla2x00_beacon_store().

  - qla_nx.c: qla82xx_fw_dump(), calls qla2x00_wait_for_chip_reset(),
    which has msleep() in a loop. It is bound to isp_operations
    ->fw_dump() hook. That hook *is* called from atomic context at
    qla_isr.c by multiple interrupt handlers. Nonetheless, it's other
    controllers interrupt handlers, and not the qla82xx.

  - qla_attr.c: qla2x00_sysfs_write_fw_dump(), and
    qla2x00_sysfs_write_reset(), process-context sysfs ->write() hooks.

  - qla_os.c: qla2x00_probe_one(). PCI ->probe(), process context.

  - qla_os.c: qla2x00_clear_drv_active(), called solely from
    qla2x00_remove_one(), which is PCI ->remove() hook, process context.

  - qla_os.c: qla2x00_do_dpc(), kthread function, process context.

Remove the in_interrupt() check. Change qla82xx_idc_lock() specification
to a purely process-context function. Mark it with "Context: task, might
sleep".

Change qla82xx_idc_lock() implementation to sleep 100ms, instead of a
schedule(), for each spin. This is more deterministic, and it matches
the other qla models idc_lock() functions.

Signed-off-by: Ahmed S. Darwish <a.darwish@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Nilesh Javali <njavali@marvell.com>
Cc: <GR-QLogic-Storage-Upstream@marvell.com>
---
 drivers/scsi/qla2xxx/qla_def.h |  5 +++++
 drivers/scsi/qla2xxx/qla_nx.c  | 25 +++++++++++--------------
 2 files changed, 16 insertions(+), 14 deletions(-)

Comments

Daniel Wagner Nov. 30, 2020, 10:59 a.m. UTC | #1
On Thu, Nov 26, 2020 at 02:29:42PM +0100, Sebastian Andrzej Siewior wrote:
> From: "Ahmed S. Darwish" <a.darwish@linutronix.de>
> 
> qla82xx_idc_lock() spins on a certain hardware state until it's updated. At
> the end of each spin, if in_interrupt() is true, it does 20 loops of
> cpu_relax(). Otherwise, it yields the CPU.
> 
> While in_interrupt() is ill-defined and does not provide what the name
> suggests, it is not needed here: qla82xx_idc_lock() is always called
> from process context. Below is an analysis of its callers, in order of
> appearance:
> 
>   - qla_nx.c: qla82xx_device_bootstrap(), only called by
>     qla82xx_device_state_handler(), has multiple msleep()s.
> 
>   - qla_nx.c: qla82xx_need_qsnt_handler(), has one second msleep()
> 
>   - qla_nx.c: qla82xx_wait_for_state_change(), one second msleep()
> 
>   - qla_nx.c: qla82xx_need_reset_handler(), can sleep up to 10 seconds
> 
>   - qla_nx.c: qla82xx_device_state_handler(), has multiple msleep()s
> 
>   - qla_nx.c: qla82xx_abort_isp(), if it's a qla82xx controller, calls
>     qla82xx_device_state_handler(), which sleeps. It's also bound to
>     isp_operations ->abort_isp() hook, where all the callers are in
>     process context.
> 
>   - qla_nx.c: qla82xx_beacon_on(), bound to isp_operations ->beacon_on()
>     hook.  That hook is only called once, in a mutex locked context,
>     from qla2x00_beacon_store().
> 
>   - qla_nx.c: qla82xx_beacon_off(), bound to isp_operations
>     ->beacon_off() hook.  Like ->beacon_on(), it's only called once, in
>     a mutex locked context, from qla2x00_beacon_store().
> 
>   - qla_nx.c: qla82xx_fw_dump(), calls qla2x00_wait_for_chip_reset(),
>     which has msleep() in a loop. It is bound to isp_operations
>     ->fw_dump() hook. That hook *is* called from atomic context at
>     qla_isr.c by multiple interrupt handlers. Nonetheless, it's other
>     controllers interrupt handlers, and not the qla82xx.

qla82xx_msix_default() and qla82xx_msix_rsp_q() call
qla24xx_process_response_queue() which doesn't implement the firmware
dumping.

>   - qla_attr.c: qla2x00_sysfs_write_fw_dump(), and
>     qla2x00_sysfs_write_reset(), process-context sysfs ->write() hooks.
> 
>   - qla_os.c: qla2x00_probe_one(). PCI ->probe(), process context.
> 
>   - qla_os.c: qla2x00_clear_drv_active(), called solely from
>     qla2x00_remove_one(), which is PCI ->remove() hook, process context.
> 
>   - qla_os.c: qla2x00_do_dpc(), kthread function, process context.
> 
> Remove the in_interrupt() check. Change qla82xx_idc_lock() specification
> to a purely process-context function. Mark it with "Context: task, might
> sleep".
> 
> Change qla82xx_idc_lock() implementation to sleep 100ms, instead of a
> schedule(), for each spin. This is more deterministic, and it matches
> the other qla models idc_lock() functions.
> 
> Signed-off-by: Ahmed S. Darwish <a.darwish@linutronix.de>
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> Cc: Nilesh Javali <njavali@marvell.com>
> Cc: <GR-QLogic-Storage-Upstream@marvell.com>

Reviewed-by: Daniel Wagner <dwagner@suse.de>
Himanshu Madhani Nov. 30, 2020, 7:11 p.m. UTC | #2
> On Nov 26, 2020, at 7:29 AM, Sebastian Andrzej Siewior <bigeasy@linutronix.de> wrote:
> 
> From: "Ahmed S. Darwish" <a.darwish@linutronix.de>
> 
> qla82xx_idc_lock() spins on a certain hardware state until it's updated. At
> the end of each spin, if in_interrupt() is true, it does 20 loops of
> cpu_relax(). Otherwise, it yields the CPU.
> 
> While in_interrupt() is ill-defined and does not provide what the name
> suggests, it is not needed here: qla82xx_idc_lock() is always called
> from process context. Below is an analysis of its callers, in order of
> appearance:
> 
>  - qla_nx.c: qla82xx_device_bootstrap(), only called by
>    qla82xx_device_state_handler(), has multiple msleep()s.
> 
>  - qla_nx.c: qla82xx_need_qsnt_handler(), has one second msleep()
> 
>  - qla_nx.c: qla82xx_wait_for_state_change(), one second msleep()
> 
>  - qla_nx.c: qla82xx_need_reset_handler(), can sleep up to 10 seconds
> 
>  - qla_nx.c: qla82xx_device_state_handler(), has multiple msleep()s
> 
>  - qla_nx.c: qla82xx_abort_isp(), if it's a qla82xx controller, calls
>    qla82xx_device_state_handler(), which sleeps. It's also bound to
>    isp_operations ->abort_isp() hook, where all the callers are in
>    process context.
> 
>  - qla_nx.c: qla82xx_beacon_on(), bound to isp_operations ->beacon_on()
>    hook.  That hook is only called once, in a mutex locked context,
>    from qla2x00_beacon_store().
> 
>  - qla_nx.c: qla82xx_beacon_off(), bound to isp_operations
>    ->beacon_off() hook.  Like ->beacon_on(), it's only called once, in
>    a mutex locked context, from qla2x00_beacon_store().
> 
>  - qla_nx.c: qla82xx_fw_dump(), calls qla2x00_wait_for_chip_reset(),
>    which has msleep() in a loop. It is bound to isp_operations
>    ->fw_dump() hook. That hook *is* called from atomic context at
>    qla_isr.c by multiple interrupt handlers. Nonetheless, it's other
>    controllers interrupt handlers, and not the qla82xx.
> 
>  - qla_attr.c: qla2x00_sysfs_write_fw_dump(), and
>    qla2x00_sysfs_write_reset(), process-context sysfs ->write() hooks.
> 
>  - qla_os.c: qla2x00_probe_one(). PCI ->probe(), process context.
> 
>  - qla_os.c: qla2x00_clear_drv_active(), called solely from
>    qla2x00_remove_one(), which is PCI ->remove() hook, process context.
> 
>  - qla_os.c: qla2x00_do_dpc(), kthread function, process context.
> 
> Remove the in_interrupt() check. Change qla82xx_idc_lock() specification
> to a purely process-context function. Mark it with "Context: task, might
> sleep".
> 
> Change qla82xx_idc_lock() implementation to sleep 100ms, instead of a
> schedule(), for each spin. This is more deterministic, and it matches
> the other qla models idc_lock() functions.
> 
> Signed-off-by: Ahmed S. Darwish <a.darwish@linutronix.de>
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> Cc: Nilesh Javali <njavali@marvell.com>
> Cc: <GR-QLogic-Storage-Upstream@marvell.com>
> ---
> drivers/scsi/qla2xxx/qla_def.h |  5 +++++
> drivers/scsi/qla2xxx/qla_nx.c  | 25 +++++++++++--------------
> 2 files changed, 16 insertions(+), 14 deletions(-)
> 
> diff --git a/drivers/scsi/qla2xxx/qla_def.h b/drivers/scsi/qla2xxx/qla_def.h
> index ed9b10f8537d6..fe3c0e2f1ce88 100644
> --- a/drivers/scsi/qla2xxx/qla_def.h
> +++ b/drivers/scsi/qla2xxx/qla_def.h
> @@ -3296,8 +3296,10 @@ struct isp_operations {
> 	void (*fw_dump)(struct scsi_qla_host *vha);
> 	void (*mpi_fw_dump)(struct scsi_qla_host *, int);
> 
> +	/* Context: task, might sleep */
> 	int (*beacon_on) (struct scsi_qla_host *);
> 	int (*beacon_off) (struct scsi_qla_host *);
> +
> 	void (*beacon_blink) (struct scsi_qla_host *);
> 
> 	void *(*read_optrom)(struct scsi_qla_host *, void *,
> @@ -3308,7 +3310,10 @@ struct isp_operations {
> 	int (*get_flash_version) (struct scsi_qla_host *, void *);
> 	int (*start_scsi) (srb_t *);
> 	int (*start_scsi_mq) (srb_t *);
> +
> +	/* Context: task, might sleep */
> 	int (*abort_isp) (struct scsi_qla_host *);
> +
> 	int (*iospace_config)(struct qla_hw_data *);
> 	int (*initialize_adapter)(struct scsi_qla_host *);
> };
> diff --git a/drivers/scsi/qla2xxx/qla_nx.c b/drivers/scsi/qla2xxx/qla_nx.c
> index b3ba0de5d4fb8..b2017f1c35044 100644
> --- a/drivers/scsi/qla2xxx/qla_nx.c
> +++ b/drivers/scsi/qla2xxx/qla_nx.c
> @@ -489,29 +489,26 @@ qla82xx_rd_32(struct qla_hw_data *ha, ulong off_in)
> 	return data;
> }
> 
> -#define IDC_LOCK_TIMEOUT 100000000
> +/*
> + * Context: task, might sleep
> + */
> int qla82xx_idc_lock(struct qla_hw_data *ha)
> {
> -	int i;
> -	int done = 0, timeout = 0;
> +	const int delay_ms = 100, timeout_ms = 2000;
> +	int done, total = 0;
> 
> -	while (!done) {
> +	might_sleep();
> +
> +	while (true) {
> 		/* acquire semaphore5 from PCI HW block */
> 		done = qla82xx_rd_32(ha, QLA82XX_PCIE_REG(PCIE_SEM5_LOCK));
> 		if (done == 1)
> 			break;
> -		if (timeout >= IDC_LOCK_TIMEOUT)
> +		if (WARN_ON_ONCE(total >= timeout_ms))
> 			return -1;
> 
> -		timeout++;
> -
> -		/* Yield CPU */
> -		if (!in_interrupt())
> -			schedule();
> -		else {
> -			for (i = 0; i < 20; i++)
> -				cpu_relax();
> -		}
> +		total += delay_ms;
> +		msleep(delay_ms);
> 	}
> 
> 	return 0;
> -- 
> 2.29.2
> 

Looks Good. 

Reviewed-by: Himanshu Madhani <himanshu.madhani@oracle.com>

--
Himanshu Madhani	 Oracle Linux Engineering
diff mbox series

Patch

diff --git a/drivers/scsi/qla2xxx/qla_def.h b/drivers/scsi/qla2xxx/qla_def.h
index ed9b10f8537d6..fe3c0e2f1ce88 100644
--- a/drivers/scsi/qla2xxx/qla_def.h
+++ b/drivers/scsi/qla2xxx/qla_def.h
@@ -3296,8 +3296,10 @@  struct isp_operations {
 	void (*fw_dump)(struct scsi_qla_host *vha);
 	void (*mpi_fw_dump)(struct scsi_qla_host *, int);
 
+	/* Context: task, might sleep */
 	int (*beacon_on) (struct scsi_qla_host *);
 	int (*beacon_off) (struct scsi_qla_host *);
+
 	void (*beacon_blink) (struct scsi_qla_host *);
 
 	void *(*read_optrom)(struct scsi_qla_host *, void *,
@@ -3308,7 +3310,10 @@  struct isp_operations {
 	int (*get_flash_version) (struct scsi_qla_host *, void *);
 	int (*start_scsi) (srb_t *);
 	int (*start_scsi_mq) (srb_t *);
+
+	/* Context: task, might sleep */
 	int (*abort_isp) (struct scsi_qla_host *);
+
 	int (*iospace_config)(struct qla_hw_data *);
 	int (*initialize_adapter)(struct scsi_qla_host *);
 };
diff --git a/drivers/scsi/qla2xxx/qla_nx.c b/drivers/scsi/qla2xxx/qla_nx.c
index b3ba0de5d4fb8..b2017f1c35044 100644
--- a/drivers/scsi/qla2xxx/qla_nx.c
+++ b/drivers/scsi/qla2xxx/qla_nx.c
@@ -489,29 +489,26 @@  qla82xx_rd_32(struct qla_hw_data *ha, ulong off_in)
 	return data;
 }
 
-#define IDC_LOCK_TIMEOUT 100000000
+/*
+ * Context: task, might sleep
+ */
 int qla82xx_idc_lock(struct qla_hw_data *ha)
 {
-	int i;
-	int done = 0, timeout = 0;
+	const int delay_ms = 100, timeout_ms = 2000;
+	int done, total = 0;
 
-	while (!done) {
+	might_sleep();
+
+	while (true) {
 		/* acquire semaphore5 from PCI HW block */
 		done = qla82xx_rd_32(ha, QLA82XX_PCIE_REG(PCIE_SEM5_LOCK));
 		if (done == 1)
 			break;
-		if (timeout >= IDC_LOCK_TIMEOUT)
+		if (WARN_ON_ONCE(total >= timeout_ms))
 			return -1;
 
-		timeout++;
-
-		/* Yield CPU */
-		if (!in_interrupt())
-			schedule();
-		else {
-			for (i = 0; i < 20; i++)
-				cpu_relax();
-		}
+		total += delay_ms;
+		msleep(delay_ms);
 	}
 
 	return 0;