From: Dan Williams <dan.j.williams@intel.com>
To: James Bottomley <James.Bottomley@HansenPartnership.com>
Cc: Alan Stern <stern@rowland.harvard.edu>, Tejun Heo <tj@kernel.org>,
Len Brown <len.brown@intel.com>,
linux-scsi <linux-scsi@vger.kernel.org>,
Phillip Susi <psusi@ubuntu.com>,
IDE/ATA development list <linux-ide@vger.kernel.org>,
Todd Brandt <todd.e.brandt@linux.intel.com>
Subject: Re: [PATCH v5 3/3] scsi: async sd resume
Date: Thu, 20 Mar 2014 17:09:43 -0700 [thread overview]
Message-ID: <1395360583.5823.1.camel@dwillia2-mobl2.amr.corp.intel.com> (raw)
In-Reply-To: <1395080440.10864.51.camel@dwillia2-mobl2.amr.corp.intel.com>
On Mon, 2014-03-17 at 11:20 -0700, Dan Williams wrote:
> On Sun, 2014-03-16 at 11:21 -0700, James Bottomley wrote:
> > On Sat, 2014-03-15 at 16:35 -0700, Dan Williams wrote:
> > > On Sat, Mar 15, 2014 at 2:47 PM, James Bottomley
> > > > Hm, OK, if this is tied at the hip to async scanning, why do you need
> > > > another async domain for it? Why not just use the current async
> > > > scanning domain ... it will actually probably resolve a few nasty (but
> > > > wholly manufactured) races where scanning races with suspend.
> > >
> > > I considered that initially, but it ends up destroying most/all of the
> > > benefit of doing it asynchronously. This is due to the fact that
> > > scsi_sd_probe_domain is flushed by the async_synchronize_full()
> > > performed in dpm_resume(). We want userspace to resume while the disk
> > > may still be starting.
> >
> > OK, finally got it, the new domain doesn't participate in
> > async_synchronize_full() but scsi_sd_probe_domain does (and has to
> > because of the device and module operations). That might actually be
> > worth a comment somewhere.
>
> Fair enough, see the comment added to the declaration of
> scsi_sd_pm_domain below.
>
> BTW, these patches are independent. Patch 1 and 2 can go through
> Tejun/libata and this one can go through you/scsi by itself.
>
...and Tejun has now applied the ata bits for 3.15.
James, when you get a chance, please apply this latest version of patch
3 or otherwise let me know your disposition.
Regards,
Dan
(no changes since the version I sent on the 17th)
8<------
scsi: async sd resume
From: Dan Williams <dan.j.williams@intel.com>
async_schedule() sd resume work to allow disks and other devices to
resume in parallel.
This moves the entirety of scsi_device resume to an async context to
ensure that scsi_device_resume() remains ordered with respect to the
completion of the start/stop command. For the duration of the resume,
new command submissions (that do not originate from the scsi-core) will
be deferred (BLKPREP_DEFER).
It adds a new ASYNC_DOMAIN_EXCLUSIVE(scsi_sd_pm_domain) as a container
of these operations. Like scsi_sd_probe_domain it is flushed at
sd_remove() time to ensure async ops do not continue past the
end-of-life of the sdev. The implementation explicitly refrains from
reusing scsi_sd_probe_domain directly for this purpose as it is flushed
at the end of dpm_resume(), potentially defeating some of the benefit.
Given sdevs are quiesced it is permissible for these resume operations
to bleed past the async_synchronize_full() calls made by the driver
core.
We defer the resolution of which pm callback to call until
scsi_dev_type_{suspend|resume} time and guarantee that the callback
parameter is never NULL. With this in place the type of resume
operation is encoded in the async function identifier.
There is a concern that async resume could trigger PSU overload. In the
enterprise, storage enclosures enforce staggered spin-up regardless of
what the kernel does making async scanning safe by default. Outside of
that context a user can disable asynchronous scanning via a kernel
command line or CONFIG_SCSI_SCAN_ASYNC. Honor that setting when
deciding whether to do resume asynchronously.
Inspired by Todd's analysis and initial proposal [2]:
https://01.org/suspendresume/blogs/tebrandt/2013/hard-disk-resume-optimization-simpler-approach
Cc: Len Brown <len.brown@intel.com>
Cc: Phillip Susi <psusi@ubuntu.com>
[alan: bug fix and clean up suggestion]
Acked-by: Alan Stern <stern@rowland.harvard.edu>
Suggested-by: Todd Brandt <todd.e.brandt@linux.intel.com>
[djbw: kick all resume work to the async queue]
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
---
drivers/scsi/Kconfig | 3 +
drivers/scsi/scsi.c | 9 +++
drivers/scsi/scsi_pm.c | 128 ++++++++++++++++++++++++++++++++++++----------
drivers/scsi/scsi_priv.h | 2 +
drivers/scsi/scsi_scan.c | 2 -
drivers/scsi/sd.c | 1
6 files changed, 115 insertions(+), 30 deletions(-)
diff --git a/drivers/scsi/Kconfig b/drivers/scsi/Kconfig
index c8bd092fc945..02832d64d918 100644
--- a/drivers/scsi/Kconfig
+++ b/drivers/scsi/Kconfig
@@ -263,6 +263,9 @@ config SCSI_SCAN_ASYNC
You can override this choice by specifying "scsi_mod.scan=sync"
or async on the kernel's command line.
+ Note that this setting also affects whether resuming from
+ system suspend will be performed asynchronously.
+
menu "SCSI Transports"
depends on SCSI
diff --git a/drivers/scsi/scsi.c b/drivers/scsi/scsi.c
index d8afec8317cf..1b345bf41a91 100644
--- a/drivers/scsi/scsi.c
+++ b/drivers/scsi/scsi.c
@@ -91,6 +91,15 @@ EXPORT_SYMBOL(scsi_logging_level);
ASYNC_DOMAIN(scsi_sd_probe_domain);
EXPORT_SYMBOL(scsi_sd_probe_domain);
+/*
+ * Separate domain (from scsi_sd_probe_domain) to maximize the benefit of
+ * asynchronous system resume operations. It is marked 'exclusive' to avoid
+ * being included in the async_synchronize_full() that is invoked by
+ * dpm_resume()
+ */
+ASYNC_DOMAIN_EXCLUSIVE(scsi_sd_pm_domain);
+EXPORT_SYMBOL(scsi_sd_pm_domain);
+
/* NB: These are exposed through /proc/scsi/scsi and form part of the ABI.
* You may not alter any existing entry (although adding new ones is
* encouraged once assigned by ANSI/INCITS T10
diff --git a/drivers/scsi/scsi_pm.c b/drivers/scsi/scsi_pm.c
index 001e9ceda4c3..7454498c4091 100644
--- a/drivers/scsi/scsi_pm.c
+++ b/drivers/scsi/scsi_pm.c
@@ -18,35 +18,77 @@
#ifdef CONFIG_PM_SLEEP
-static int scsi_dev_type_suspend(struct device *dev, int (*cb)(struct device *))
+static int do_scsi_suspend(struct device *dev, const struct dev_pm_ops *pm)
{
+ return pm && pm->suspend ? pm->suspend(dev) : 0;
+}
+
+static int do_scsi_freeze(struct device *dev, const struct dev_pm_ops *pm)
+{
+ return pm && pm->freeze ? pm->freeze(dev) : 0;
+}
+
+static int do_scsi_poweroff(struct device *dev, const struct dev_pm_ops *pm)
+{
+ return pm && pm->poweroff ? pm->poweroff(dev) : 0;
+}
+
+static int do_scsi_resume(struct device *dev, const struct dev_pm_ops *pm)
+{
+ return pm && pm->resume ? pm->resume(dev) : 0;
+}
+
+static int do_scsi_thaw(struct device *dev, const struct dev_pm_ops *pm)
+{
+ return pm && pm->thaw ? pm->thaw(dev) : 0;
+}
+
+static int do_scsi_restore(struct device *dev, const struct dev_pm_ops *pm)
+{
+ return pm && pm->restore ? pm->restore(dev) : 0;
+}
+
+static int scsi_dev_type_suspend(struct device *dev,
+ int (*cb)(struct device *, const struct dev_pm_ops *))
+{
+ const struct dev_pm_ops *pm = dev->driver ? dev->driver->pm : NULL;
int err;
+ /* flush pending in-flight resume operations, suspend is synchronous */
+ async_synchronize_full_domain(&scsi_sd_pm_domain);
+
err = scsi_device_quiesce(to_scsi_device(dev));
if (err == 0) {
- if (cb) {
- err = cb(dev);
- if (err)
- scsi_device_resume(to_scsi_device(dev));
- }
+ err = cb(dev, pm);
+ if (err)
+ scsi_device_resume(to_scsi_device(dev));
}
dev_dbg(dev, "scsi suspend: %d\n", err);
return err;
}
-static int scsi_dev_type_resume(struct device *dev, int (*cb)(struct device *))
+static int scsi_dev_type_resume(struct device *dev,
+ int (*cb)(struct device *, const struct dev_pm_ops *))
{
+ const struct dev_pm_ops *pm = dev->driver ? dev->driver->pm : NULL;
int err = 0;
- if (cb)
- err = cb(dev);
+ err = cb(dev, pm);
scsi_device_resume(to_scsi_device(dev));
dev_dbg(dev, "scsi resume: %d\n", err);
+
+ if (err == 0) {
+ pm_runtime_disable(dev);
+ pm_runtime_set_active(dev);
+ pm_runtime_enable(dev);
+ }
+
return err;
}
static int
-scsi_bus_suspend_common(struct device *dev, int (*cb)(struct device *))
+scsi_bus_suspend_common(struct device *dev,
+ int (*cb)(struct device *, const struct dev_pm_ops *))
{
int err = 0;
@@ -66,20 +108,54 @@ scsi_bus_suspend_common(struct device *dev, int (*cb)(struct device *))
return err;
}
-static int
-scsi_bus_resume_common(struct device *dev, int (*cb)(struct device *))
+static void async_sdev_resume(void *dev, async_cookie_t cookie)
{
- int err = 0;
+ scsi_dev_type_resume(dev, do_scsi_resume);
+}
- if (scsi_is_sdev_device(dev))
- err = scsi_dev_type_resume(dev, cb);
+static void async_sdev_thaw(void *dev, async_cookie_t cookie)
+{
+ scsi_dev_type_resume(dev, do_scsi_thaw);
+}
- if (err == 0) {
+static void async_sdev_restore(void *dev, async_cookie_t cookie)
+{
+ scsi_dev_type_resume(dev, do_scsi_restore);
+}
+
+static int scsi_bus_resume_common(struct device *dev,
+ int (*cb)(struct device *, const struct dev_pm_ops *))
+{
+ async_func_t fn;
+
+ if (!scsi_is_sdev_device(dev))
+ fn = NULL;
+ else if (cb == do_scsi_resume)
+ fn = async_sdev_resume;
+ else if (cb == do_scsi_thaw)
+ fn = async_sdev_thaw;
+ else if (cb == do_scsi_restore)
+ fn = async_sdev_restore;
+ else
+ fn = NULL;
+
+ if (fn) {
+ async_schedule_domain(fn, dev, &scsi_sd_pm_domain);
+
+ /*
+ * If a user has disabled async probing a likely reason
+ * is due to a storage enclosure that does not inject
+ * staggered spin-ups. For safety, make resume
+ * synchronous as well in that case.
+ */
+ if (strncmp(scsi_scan_type, "async", 5) != 0)
+ async_synchronize_full_domain(&scsi_sd_pm_domain);
+ } else {
pm_runtime_disable(dev);
pm_runtime_set_active(dev);
pm_runtime_enable(dev);
}
- return err;
+ return 0;
}
static int scsi_bus_prepare(struct device *dev)
@@ -97,38 +173,32 @@ static int scsi_bus_prepare(struct device *dev)
static int scsi_bus_suspend(struct device *dev)
{
- const struct dev_pm_ops *pm = dev->driver ? dev->driver->pm : NULL;
- return scsi_bus_suspend_common(dev, pm ? pm->suspend : NULL);
+ return scsi_bus_suspend_common(dev, do_scsi_suspend);
}
static int scsi_bus_resume(struct device *dev)
{
- const struct dev_pm_ops *pm = dev->driver ? dev->driver->pm : NULL;
- return scsi_bus_resume_common(dev, pm ? pm->resume : NULL);
+ return scsi_bus_resume_common(dev, do_scsi_resume);
}
static int scsi_bus_freeze(struct device *dev)
{
- const struct dev_pm_ops *pm = dev->driver ? dev->driver->pm : NULL;
- return scsi_bus_suspend_common(dev, pm ? pm->freeze : NULL);
+ return scsi_bus_suspend_common(dev, do_scsi_freeze);
}
static int scsi_bus_thaw(struct device *dev)
{
- const struct dev_pm_ops *pm = dev->driver ? dev->driver->pm : NULL;
- return scsi_bus_resume_common(dev, pm ? pm->thaw : NULL);
+ return scsi_bus_resume_common(dev, do_scsi_thaw);
}
static int scsi_bus_poweroff(struct device *dev)
{
- const struct dev_pm_ops *pm = dev->driver ? dev->driver->pm : NULL;
- return scsi_bus_suspend_common(dev, pm ? pm->poweroff : NULL);
+ return scsi_bus_suspend_common(dev, do_scsi_poweroff);
}
static int scsi_bus_restore(struct device *dev)
{
- const struct dev_pm_ops *pm = dev->driver ? dev->driver->pm : NULL;
- return scsi_bus_resume_common(dev, pm ? pm->restore : NULL);
+ return scsi_bus_resume_common(dev, do_scsi_restore);
}
#else /* CONFIG_PM_SLEEP */
diff --git a/drivers/scsi/scsi_priv.h b/drivers/scsi/scsi_priv.h
index f079a598bed4..48e5b657e79f 100644
--- a/drivers/scsi/scsi_priv.h
+++ b/drivers/scsi/scsi_priv.h
@@ -112,6 +112,7 @@ extern void scsi_exit_procfs(void);
#endif /* CONFIG_PROC_FS */
/* scsi_scan.c */
+extern char scsi_scan_type[];
extern int scsi_complete_async_scans(void);
extern int scsi_scan_host_selected(struct Scsi_Host *, unsigned int,
unsigned int, unsigned int, int);
@@ -166,6 +167,7 @@ static inline int scsi_autopm_get_host(struct Scsi_Host *h) { return 0; }
static inline void scsi_autopm_put_host(struct Scsi_Host *h) {}
#endif /* CONFIG_PM_RUNTIME */
+extern struct async_domain scsi_sd_pm_domain;
extern struct async_domain scsi_sd_probe_domain;
/*
diff --git a/drivers/scsi/scsi_scan.c b/drivers/scsi/scsi_scan.c
index 307a81137607..6b2f51f52af6 100644
--- a/drivers/scsi/scsi_scan.c
+++ b/drivers/scsi/scsi_scan.c
@@ -97,7 +97,7 @@ MODULE_PARM_DESC(max_luns,
#define SCSI_SCAN_TYPE_DEFAULT "sync"
#endif
-static char scsi_scan_type[6] = SCSI_SCAN_TYPE_DEFAULT;
+char scsi_scan_type[6] = SCSI_SCAN_TYPE_DEFAULT;
module_param_string(scan, scsi_scan_type, sizeof(scsi_scan_type), S_IRUGO);
MODULE_PARM_DESC(scan, "sync, async or none");
diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c
index 470954aba728..700c595c603e 100644
--- a/drivers/scsi/sd.c
+++ b/drivers/scsi/sd.c
@@ -3020,6 +3020,7 @@ static int sd_remove(struct device *dev)
devt = disk_devt(sdkp->disk);
scsi_autopm_get_device(sdkp->device);
+ async_synchronize_full_domain(&scsi_sd_pm_domain);
async_synchronize_full_domain(&scsi_sd_probe_domain);
blk_queue_prep_rq(sdkp->device->request_queue, scsi_prep_fn);
blk_queue_unprep_rq(sdkp->device->request_queue, NULL);
prev parent reply other threads:[~2014-03-21 0:09 UTC|newest]
Thread overview: 28+ messages / expand[flat|nested] mbox.gz Atom feed top
2014-03-05 20:17 [PATCH v5 0/3] Accelerate Storage Resume (2x or more) Dan Williams
2014-03-05 20:17 ` [PATCH v5 1/3] libata, libsas: kill pm_result and related cleanup Dan Williams
2014-03-10 20:29 ` Tejun Heo
2014-03-10 20:44 ` Tejun Heo
2014-03-05 20:17 ` [PATCH v5 2/3] libata: async resume Dan Williams
2014-12-19 23:46 ` Phillip Susi
2014-12-20 1:42 ` Phillip Susi
2014-03-05 20:17 ` [PATCH v5 3/3] scsi: async sd resume Dan Williams
2014-03-07 15:42 ` Alan Stern
2014-03-08 0:41 ` Dan Williams
2014-03-08 2:52 ` Dan Williams
2014-03-10 15:49 ` Alan Stern
2014-03-10 20:43 ` Tejun Heo
2014-03-10 20:56 ` Dan Williams
2014-03-10 20:59 ` Alan Stern
2014-03-10 21:06 ` Dan Williams
2014-03-10 22:41 ` Phillip Susi
2014-03-11 4:19 ` James Bottomley
2014-03-11 4:47 ` Dan Williams
2014-03-11 5:16 ` James Bottomley
2014-03-11 5:26 ` Dan Williams
2014-03-11 6:29 ` James Bottomley
2014-03-14 20:11 ` Dan Williams
2014-03-15 21:47 ` James Bottomley
2014-03-15 23:35 ` Dan Williams
2014-03-16 18:21 ` James Bottomley
2014-03-17 18:20 ` Dan Williams
2014-03-21 0:09 ` Dan Williams [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=1395360583.5823.1.camel@dwillia2-mobl2.amr.corp.intel.com \
--to=dan.j.williams@intel.com \
--cc=James.Bottomley@HansenPartnership.com \
--cc=len.brown@intel.com \
--cc=linux-ide@vger.kernel.org \
--cc=linux-scsi@vger.kernel.org \
--cc=psusi@ubuntu.com \
--cc=stern@rowland.harvard.edu \
--cc=tj@kernel.org \
--cc=todd.e.brandt@linux.intel.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox