20 Dec, 2017
4 commits
-
A 'perf record' on an app continuously writing in the 'value'
attribute show that most of the time is spent in kstrtol()--17.99%--value_store
|
|--10.17%--kstrtoint
| |
| |--8.82%--kstrtoll
|
|--2.50%--gpiod_set_value_cansleep
|
|--1.82%--u16_gpio_set
|
|--1.46%--value_storeThe normal case is to write 0 or 1 in the attribute, therefore
this patch avoids the call to kstrtol() in the most common casesThen 'perf record' shows
--7.21%--value_store
|
|--2.69%--u16_gpio_set
|
|--1.47%--value_store
|
|--1.08%--gpiod_set_value_cansleep
|
|--0.60%--mutex_lock
|
--0.58%--mutex_unlockSigned-off-by: Christophe Leroy
Signed-off-by: Linus Walleij -
A bench with 'perf record' shows that most of time spent in value_show()
is spent in sprintf()--42.41%--sysfs_kf_read
|
|--39.73%--dev_attr_show
| |
| |--38.23%--value_show
| | |
| | |--29.22%--sprintf
| | |
| | |--2.94%--gpiod_get_value_cansleep
| | |value_show() only returns "0\n" or "1\n", therefore the use of
sprintf() can be avoidedWith this patch we get the following result with 'perf record'
--13.89%--sysfs_kf_read
|
|--10.72%--dev_attr_show
| |
| |--9.44%--value_show
| | |
| | |--4.61%--gpiod_get_value_cansleepSigned-off-by: Christophe Leroy
Signed-off-by: Linus Walleij -
'value' attribute is supposed to only return 0 or 1 according to
the documentation.
With today's implementation, if gpiod_get_value_cansleep() fails
the printed 'value' is a negative value.This patch ensures that an error is returned on read instead.
Signed-off-by: Christophe Leroy
Signed-off-by: Linus Walleij -
The GPIO 'value' attribute is time critical. A small bench with
'perf record' on the app below shows that 80% of the time spent in
sysfs_kf_seq_show() is spent in memset() for zeroising the buffer.|--67.48%--sysfs_kf_seq_show
| |
| |--54.40%--memset
| |
| |--11.49%--dev_attr_show
| | |
| | |--10.06%--value_show
| | | |
| | | |--4.75%--sprintf
| | | | |This patch changes the attribute type to prealloc, eliminating the
need to zeroise the buffer at each read. 'perf record' gives the
following result.|--42.41%--sysfs_kf_read
| |
| |--39.73%--dev_attr_show
| | |
| | |--38.23%--value_show
| | | |
| | | |--29.22%--sprintf
| | | | |Test done with the following small app:
int main(int argc, char **argv)
{
int fd = open(argv[1], O_RDONLY);for (;;) {
int buf[512];read(fd, buf, 512);
lseek(fd, 0, SEEK_SET);
}
exit(0);
}Signed-off-by: Christophe Leroy
Signed-off-by: Linus Walleij
03 Dec, 2017
1 commit
-
General support for state persistence is added to gpiolib with the
introduction of a new pinconf parameter to propagate the request to
hardware. The existing persistence support for sleep is adapted to
include hardware support if the GPIO driver provides it. Persistence
continues to be enabled by default; in-kernel consumers can opt out, but
userspace (currently) does not have a choice.The *_SLEEP_MAY_LOSE_VALUE and *_SLEEP_MAINTAIN_VALUE symbols are
renamed, dropping the SLEEP prefix to reflect that the concept is no
longer sleep-specific. I feel that renaming to just *_MAY_LOSE_VALUE
could initially be misinterpreted, so I've further changed the symbols
to *_TRANSITORY and *_PERSISTENT to address this.The sysfs interface is modified only to keep consistency with the
chardev interface in enforcing persistence for userspace exports.Signed-off-by: Andrew Jeffery
Reviewed-by: Charles Keepax
Acked-by: Rob Herring
Signed-off-by: Linus Walleij
06 Sep, 2017
1 commit
-
Pull GPIO updates from Linus Walleij:
"This is the bulk of the GPIO changes for the v4.14 cycle.Not so much changes this time, phew. David Daney and Bartosz
Golaszewski did all the really interesting work in infrastructure
improvement across GPIO and IRQ core, hats off for them and to tglx
and Marc Z for general help with these patch sets.Core changes:
- Allow the GPIO irqchip to allocate IRQs dynamically. This is an
important change on systems where only a restricted number of IRQs,
lesser than the number of GPIO lines, can be utilized. Now we can
allocate these on a first-come-first-served basis instead of
hogging up valuable IRQ lines.- Serious fix-up of the kerneldoc documentation and inclusion into
the kerneldoc builds.- Pulled in the IRQ simulator from the IRQ core tree and use this in
the GPIO mockup driver for exhaustive testing of interrupt
abilities.New drivers:
- New driver for ThunderX and OCTEON-TX. This is especially
interesting as it picks up improvements from the IRQ core that
allow us to handle fasteoi ACKs upwards in a hierarchy when there
are IRQ flag latches on several levels in a hierarchy. Very
interesting work here.- New subdriver for Renesas R-Car r8a7745 (RZ/G1E).
Misc:
- Several fixes and improvements for Xilinx Zynq GPIO.
- Support an enablement GPIO for the 74x164 GPIO.
- Switch a bunch of chips to use devres to allocate irq descriptors.
- A bunch of constification fixes"
* tag 'gpio-v4.14-1' of git://git.kernel.org/pub/scm/linux/kernel/git/linusw/linux-gpio: (63 commits)
gpio: mockup: remove unused variable gc
gpio: pl061: constify amba_id
Revert "gpiolib: request the gpio before querying its direction"
gpio: twl6040: remove unneeded forward declaration
gpio: zevio: make gpio_chip const
gpio: add gpio_add_lookup_tables() to add several tables at once
gpio: rcar: Add r8a7745 (RZ/G1E) support
gpio: brcmstb: check return value of gpiochip_irqchip_add()
MAINTAINERS: Add entry for THUNDERX GPIO Driver.
gpio: Add gpio driver support for ThunderX and OCTEON-TX
gpio: mockup: use irq_sim
gpio: mxs: use devres for irq generic chip
gpio: mxc: use devres for irq generic chip
gpio: pch: use devres for irq generic chip
gpio: ml-ioh: use devres for irq generic chip
gpio: sta2x11: use devres for irq generic chip
gpio: sta2x11: disallow unbinding the driver
gpio: mxs: disallow unbinding the driver
gpio: mxc: disallow unbinding the driver
gpio: aspeed: Remove reference to clock name in debounce warning message
...
14 Aug, 2017
2 commits
-
Fix up some references to parameters to match the code.
Signed-off-by: Thierry Reding
Signed-off-by: Linus Walleij -
Check user-given gpio number and reject it before
calling gpio_to_desc() because gpio_to_desc() is
for kernel driver and it expects given gpio number
is valid (means 0 to 511).
If given number is invalid, gpio_to_desc() calls
WARN() and dump registers and stack for debug.
This means user can easily kick WARN() just by
writing invalid gpio number (e.g. 512) to
/sys/class/gpio/export.Fixes: 0e9a5edf5d01 ("gpio: fix deferred probe detection for legacy API")
Signed-off-by: Masami Hiramatsu
Reviewed-by: Andy Shevchenko
Signed-off-by: Linus Walleij
09 Jun, 2017
1 commit
-
The class_attrs pointer is long depreciated, and is about to be finally
removed, so move to use the class_groups pointer instead.Acked-by: Linus Walleij
Cc: Alexandre Courbot
Cc:
Signed-off-by: Greg Kroah-Hartman
12 Sep, 2016
1 commit
-
Both gpio_export and gpio_free APIs are obsolete now.
Signed-off-by: Amitesh Singh
Signed-off-by: Linus Walleij
25 Feb, 2016
1 commit
-
We started to assign the gpio_device as parent for the sysfs
but this changes the expected layout of sysfs. Restore the
previous behaviour.Signed-off-by: Bamvor Jian Zhang
Signed-off-by: Linus Walleij
12 Feb, 2016
3 commits
-
Some information about the GPIO chip need to stay around also
after the gpio_chip has been removed and only the gpio_device
persist. The base and ngpio are such things, for example we
don't want a new chip arriving to overlap the number space
of a dangling gpio_device, and the chardev may still query
the device for the number of lines etc.Note that the code that assigns base and insert gpio_device
into the global list no longer check for a missing gpio_chip:
we respect the number space allocated by any other gpio_device.As a consequence of the gdev being referenced directly from
the gpio_desc, we need to verify it differently from all
in-kernel API calls that fall through to direct queries to
the gpio_chip vtable: we first check that desc is !NULL, then
that desc->gdev is !NULL, then, if desc->gdev->chip is NULL,
we *BAIL OUT* without any error, so as to manage the case
where operations are requested on a device that is gone.These checks were non-uniform and partly missing in the past:
so to simplify: create the macros VALIDATE_DESC() that will
return -EINVAL if the desc or desc->gdev is missing and just
0 if the chip is gone, and conversely VALIDATE_DESC_VOID()
for the case where the function does not return an error.
By using these macros, we get warning messages about missing
gdev with reference to the right function in the kernel log.Despite the macro business this simplifies the code and make
it more readable than if we copy/paste the same descriptor
checking code into all code ABI call sites (IMHO).Signed-off-by: Linus Walleij
-
We need gpio_device to hold the descriptors so that they can
be lifecycled with the struct gpio_device held from userspace.
Move the descriptor array into gpio_device. Also rename it from
"desc" (singularis) to "descs" (pluralis) to reflect the fact
that it is an array.Signed-off-by: Linus Walleij
-
Since gpio_device is the struct that survives if the backing
gpio_chip is removed, move the sysfs mock device to this state
container so it becomes part of the dangling state of the
GPIO device on removal.Signed-off-by: Linus Walleij
09 Feb, 2016
1 commit
-
GPIO chips have been around for years, but were never real devices,
instead they were piggy-backing on a parent device (such as a
platform_device or amba_device) but this was always optional.
GPIO chips could also exist without any device at all, with its
struct device *parent (ex *dev) pointer being set to null.When sysfs was in use, a mock device would be created, with the
optional parent assigned, or just floating orphaned with NULL
as parent.If sysfs is active, it will use this device as parent.
We now create a gpio_device struct containing a real
struct device and move the subsystem over to using that. The
list of struct gpio_chip:s is augmented to hold struct
gpio_device:s and we find gpio_chips:s by first looking up
the struct gpio_device.The struct gpio_device is designed to stay around even if the
gpio_chip is removed, so as to satisfy users in userspace
that need a backing data structure to hold the state of the
session initiated with e.g. a character device even if there is
no physical chip anymore.From this point on, gpiochips are devices.
Cc: Johan Hovold
Cc: Michael Welling
Cc: Markus Pargmann
Signed-off-by: Linus Walleij
19 Nov, 2015
1 commit
-
The name .dev in a struct is normally reserved for a struct device
that is let us say a superclass to the thing described by the struct.
struct gpio_chip stands out by confusingly using a struct device *dev
to point to the parent device (such as a platform_device) that
represents the hardware. As we want to give gpio_chip:s real devices,
this is not working. We need to rename this member to parent.This was done by two coccinelle scripts, I guess it is possible to
combine them into one, but I don't know such stuff. They look like
this:@@
struct gpio_chip *var;
@@
-var->dev
+var->parentand:
@@
struct gpio_chip var;
@@
-var.dev
+var.parentand:
@@
struct bgpio_chip *var;
@@
-var->gc.dev
+var->gc.parentPlus a few instances of bgpio that I couldn't figure out how
to teach Coccinelle to rewrite.This patch hits all over the place, but I *strongly* prefer this
solution to any piecemal approaches that just exercise patch
mechanics all over the place. It mainly hits drivers/gpio and
drivers/pinctrl which is my own backyard anyway.Cc: Haavard Skinnemoen
Cc: Rafał Miłecki
Cc: Richard Purdie
Cc: Mauro Carvalho Chehab
Cc: Alek Du
Cc: Jaroslav Kysela
Cc: Takashi Iwai
Acked-by: Dmitry Torokhov
Acked-by: Greg Kroah-Hartman
Acked-by: Lee Jones
Acked-by: Jiri Kosina
Acked-by: Hans-Christian Egtvedt
Acked-by: Jacek Anaszewski
Signed-off-by: Linus Walleij
12 May, 2015
21 commits
-
Move irq trigger flags, which as sysfs-interface specific, to the class
device data.This avoids accessing the gpio-descriptor flags field using non-atomic
operations without any locking, and allows for a more clear separation
of the sysfs interface from gpiolib core.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Remove FLAG_SYSFS_DIR, which is sysfs-interface specific, and store it
in the class-device data instead.Note that the flag is only used during export.
Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Rename active-low helper using common prefix.
Also remove unnecessary manipulation of value argument.
Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Make sure to deregister the class device (and release the irq) while
holding the sysfs lock in gpio_unexport to prevent racing with
gpio_export.Note that this requires the recently introduced per-gpio locking to
avoid a deadlock with the kernfs active protection when waiting for the
attribute operations to drain during deregistration.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Add a per-gpio mutex to serialise attribute operations rather than use
one global mutex for all gpios and chips.Having a single global lock for all gpios in a system adds unnecessary
latency to the sysfs interface, and especially when having gpio
controllers connected over slow buses.Now that the global gpio-sysfs interrupt table is gone and with per-gpio
data in place, we can easily switch to using a more fine-grained locking
scheme.Keep the global mutex to serialise the global (class) operations of gpio
export and unexport and chip removal.Also document the locking assumptions made.
Note that this is also needed to fix a race between gpiod_export and
gpiod_unexport.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Drop unnecessary locking from gpiod_export_link. If the class device has
not already been unregistered, class_find_device returns the ref-counted
class device so there's no need for locking.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Remove goto from success path.
Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Add separate helper functions for irq request and free.
Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Only call irq helper if actually reconfiguring interrupt state.
This is a preparatory step in introducing separate gpio-irq request and
free functions.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Store the value sysfs entry in the gpiod data rather than in a global
table accessed through an index stored in the overloaded gpio-descriptor
flag field.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Remove redundant gpio-descriptor parameters from sysfs_set_active_low and
gpio_setup_irq.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Add gpiod class-device data.
This is a first step in getting rid of the insane gpio-descriptor flag
overloading, backward irq-interface implementation, and course grained
sysfs-interface locking (a single static mutex for every operation on
all exported gpios in a system).Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
The attribute operations will never be called for an unregistered device
so remove redundant checks for FLAG_EXPORT.Note that kernfs will also guarantee that any active sysfs operation has
finished before the attribute is removed during deregistration.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Make sure to release any irq only after the class device has been
deregistered.This avoids a race between gpiod_unexport and edge_store, where an irq
could be allocated just before the gpio class device is deregistered
without relying on FLAG_EXPORT and the global sysfs lock.Note that there is no need to hold the sysfs lock when releasing the irq
after the class device is gone as kernfs will prevent further attribute
operations.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Use DEVICE_ATTR_RO and DEVICE_ATTR_RW rather than specifying masks and
callbacks directly.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Remove gpiod_sysfs_set_active_low (and gpio_sysfs_set_active_low) which
allowed code to change the polarity of a gpio line even after it had
been exported through sysfs.Drivers should not care, and generally does not know, about gpio-line
polarity which is a hardware feature that needs to be described by
firmware.It is currently possible to define gpio-line polarity in device-tree and
acpi firmware or using platform data. Userspace can also change the
polarity through sysfs.Note that drivers using the legacy gpio interface could still use
GPIOF_ACTIVE_LOW to change the polarity before exporting the gpio.There are no in-kernel users of this interface.
Cc: Jonathan Corbet
Cc: Harry Wei
Cc: Arnd Bergmann
Cc: linux-doc@vger.kernel.org
Cc: linux-kernel@zh-kernel.org
Cc: linux-arch@vger.kernel.org
Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Rename the gpio-chip export/unexport functions to the more descriptive
names gpiochip_sysfs_register and gpiochip_sysfs_unregister.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Clean gpio-chip class device registration and deregistration.
The class device is registered when a gpio-chip is added (or from
gpiolib_sysfs_init post-core init call), and deregistered when the chip
is removed.Store the class device in struct gpio_chip directly rather than do a
class-device lookup on deregistration. This also removes the need for
the exported flag.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Reduce scope of sysfs_lock protection during chip export and unexport,
which is only needed to prevent gpiod (re-)exports during chip removal.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Put the recently introduced gpio-chip pointer to some more use in
gpiod_export.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij -
Drivers should call gpiochip_lock_as_irq (which prevents the pin
direction from being changed) in their irq_request_resources callbacks
but some drivers currently fail to do so.Instead a second, explicit and often redundant call to lock-as-irq is
made by the sysfs-interface implementation after an irq has been
requested.Move the explicit call before the irq-request to match the unlock done
after the irq is later released. Note that this also fixes an irq leak,
should the explicit call ever have failed.Also add a comment about removing the redundant call once the broken
drivers have been fixed.Signed-off-by: Johan Hovold
Reviewed-by: Alexandre Courbot
Signed-off-by: Linus Walleij
30 Apr, 2015
1 commit
-
Unregister GPIOs requested through sysfs at chip remove to avoid leaking
the associated memory and sysfs entries.The stale sysfs entries prevented the gpio numbers from being exported
when the gpio range was later reused (e.g. at device reconnect).This also fixes the related module-reference leak.
Note that kernfs makes sure that any on-going sysfs operations finish
before the class devices are unregistered and that further accesses
fail.The chip exported flag is used to prevent gpiod exports during removal.
This also makes it harder to trigger, but does not fix, the related race
between gpiochip_remove and export_store, which is really a race with
gpiod_request that needs to be addressed separately.Also note that this would prevent the crashes (e.g. NULL-dereferences)
at reconnect that affects pre-3.18 kernels, as well as use-after-free on
operations on open attribute files on pre-3.14 kernels (prior to
kernfs).Fixes: d8f388d8dc8d ("gpio: sysfs interface")
Cc: stable # v2.6.27: 01cca93a9491
Signed-off-by: Johan Hovold
Signed-off-by: Linus Walleij
30 Jan, 2015
2 commits
-
Fix memory leak in the gpio sysfs interface due to failure to drop
reference to device returned by class_find_device when setting the
gpio-line polarity.Fixes: 0769746183ca ("gpiolib: add support for changing value polarity
in sysfs")
Cc: stable # v2.6.33
Signed-off-by: Johan Hovold
Signed-off-by: Linus Walleij -
Fix memory leak in the gpio sysfs interface due to failure to drop
reference to device returned by class_find_device when creating a link.Fixes: a4177ee7f1a8 ("gpiolib: allow exported GPIO nodes to be named
using sysfs links")
Cc: stable # v2.6.32
Signed-off-by: Johan Hovold
Signed-off-by: Linus Walleij