| Message ID | 20220921063026.89619-5-matt.ranostay@konsulko.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | HID: mcp2221: iio support and device resource management | expand |
On Fri, Sep 23, 2022 at 1:45 AM Matt Ranostay <matt.ranostay@konsulko.com> wrote: > > On Wed, Sep 21, 2022 at 10:57 AM Matt Ranostay > <matt.ranostay@konsulko.com> wrote: > > > > On Wed, Sep 21, 2022 at 1:05 AM Benjamin Tissoires > > <benjamin.tissoires@redhat.com> wrote: > > > > > > [foreword: please keep Jiri and myself (the HID maintainers) CC-ed to > > > the series, as you will need ack from us and we don't necessarily monitor > > > every single message on linux-input] > > > > > > On Sep 20 2022, Matt Ranostay wrote: > > > > Switch from i2c_add_adapter() to resource managed devm_i2c_add_adapter() > > > > for matching rest of driver initialization, and more concise code. > > > > > > > > Signed-off-by: Matt Ranostay <matt.ranostay@konsulko.com> > > > > --- > > > > drivers/hid/hid-mcp2221.c | 45 +++++++++++++++++---------------------- > > > > 1 file changed, 19 insertions(+), 26 deletions(-) > > > > > > > > diff --git a/drivers/hid/hid-mcp2221.c b/drivers/hid/hid-mcp2221.c > > > > index de52e9f7bb8c..7ba63bcd66de 100644 > > > > --- a/drivers/hid/hid-mcp2221.c > > > > +++ b/drivers/hid/hid-mcp2221.c > > > > @@ -824,6 +824,14 @@ static int mcp2221_raw_event(struct hid_device *hdev, > > > > return 1; > > > > } > > > > > > > > +static void mcp2221_hid_remove(void *ptr) > > > > +{ > > > > + struct hid_device *hdev = ptr; > > > > + > > > > + hid_hw_close(hdev); > > > > + hid_hw_stop(hdev); > > > > > > By default, if you remove the .remove() callback, hid_hw_stop() will get > > > automatically called by hid-core.c. So we are now calling it twice, > > > which, in a way is not a big deal but it might be an issue in the long > > > run. > > > > > > Generally speaking, in the HID subsystem, that situation doesn't happen > > > a lot because hid_hw_start() is usually the last command of probe, and > > > we don't need to open the device in the driver itself. > > > > > > Here, I guess as soon as you add the i2c adapter, you might want to have > > > the communication channels ready, and thus you need to have it open > > > *before* i2c_add_adapter. > > > > > > I would suggest the following if you want to keep the devm release of > > > stop and close: please put a big fat warning before mcp2221_hid_remove() > > > explaining that this is called in devm management, *and* add a function > > > that would just return 0 as the .remove() callback with another big fat > > > warning explaining that we don't want hid-core.c to call hid_hw_stop() > > > because we are doing it ourself through devres. > > > > > > > Yeah maybe best to keep the non-devres if it isn't going to affect how the last > > change in this series is trying to implement with iio. > > > > I'll wait for Jonathan to chime in on this thread. > > > > > Last, in the HID subsystem, we often interleave non devres with devres > > > for resource allocation, given that .remove() will be called before any > > > devres release. But that is assuming this ordering is OK, which doesn't > > > seem to be the case here. We first need to unregister the i2c adapter > > > and then close/stop the HID device. > > On second thought I2C will be unregistered before the HID calls, since > unless I'm totally > incorrect device resource management unwinds backwards in the order actions are > registered. Yeah, sorry if it was not clear: - .remove() is called *before* any devres action takes place - devres action are LIFO, so unwinded backwards as you say In the general case, a driver does: int probe() { void *pointer = devm_alloc(...) some_more_devm_action(...) hid_hw_start() return 0; } and so the HID start action is the last one, meaning that .remove will first call stop and then devres unwind will get called. But here, in your case, you need hid_hw_start to be called *before* devm_i2c_add_adapter(), meaning that the implicit .remove() will mess up with the device, so you are forced to do something about it. You can either keep a non devm variant, or you can override the .remove() of HID to not do anything and do the stop/close in a specific devm task, which you did here. You are just missing the "let's override .remove() to ensure we keep the device open and started while we need it". Cheers, Benjamin
On Fri, Sep 23, 2022 at 12:03 AM Benjamin Tissoires <benjamin.tissoires@redhat.com> wrote: > > On Fri, Sep 23, 2022 at 1:45 AM Matt Ranostay > <matt.ranostay@konsulko.com> wrote: > > > > On Wed, Sep 21, 2022 at 10:57 AM Matt Ranostay > > <matt.ranostay@konsulko.com> wrote: > > > > > > On Wed, Sep 21, 2022 at 1:05 AM Benjamin Tissoires > > > <benjamin.tissoires@redhat.com> wrote: > > > > > > > > [foreword: please keep Jiri and myself (the HID maintainers) CC-ed to > > > > the series, as you will need ack from us and we don't necessarily monitor > > > > every single message on linux-input] > > > > > > > > On Sep 20 2022, Matt Ranostay wrote: > > > > > Switch from i2c_add_adapter() to resource managed devm_i2c_add_adapter() > > > > > for matching rest of driver initialization, and more concise code. > > > > > > > > > > Signed-off-by: Matt Ranostay <matt.ranostay@konsulko.com> > > > > > --- > > > > > drivers/hid/hid-mcp2221.c | 45 +++++++++++++++++---------------------- > > > > > 1 file changed, 19 insertions(+), 26 deletions(-) > > > > > > > > > > diff --git a/drivers/hid/hid-mcp2221.c b/drivers/hid/hid-mcp2221.c > > > > > index de52e9f7bb8c..7ba63bcd66de 100644 > > > > > --- a/drivers/hid/hid-mcp2221.c > > > > > +++ b/drivers/hid/hid-mcp2221.c > > > > > @@ -824,6 +824,14 @@ static int mcp2221_raw_event(struct hid_device *hdev, > > > > > return 1; > > > > > } > > > > > > > > > > +static void mcp2221_hid_remove(void *ptr) > > > > > +{ > > > > > + struct hid_device *hdev = ptr; > > > > > + > > > > > + hid_hw_close(hdev); > > > > > + hid_hw_stop(hdev); > > > > > > > > By default, if you remove the .remove() callback, hid_hw_stop() will get > > > > automatically called by hid-core.c. So we are now calling it twice, > > > > which, in a way is not a big deal but it might be an issue in the long > > > > run. > > > > > > > > Generally speaking, in the HID subsystem, that situation doesn't happen > > > > a lot because hid_hw_start() is usually the last command of probe, and > > > > we don't need to open the device in the driver itself. > > > > > > > > Here, I guess as soon as you add the i2c adapter, you might want to have > > > > the communication channels ready, and thus you need to have it open > > > > *before* i2c_add_adapter. > > > > > > > > I would suggest the following if you want to keep the devm release of > > > > stop and close: please put a big fat warning before mcp2221_hid_remove() > > > > explaining that this is called in devm management, *and* add a function > > > > that would just return 0 as the .remove() callback with another big fat > > > > warning explaining that we don't want hid-core.c to call hid_hw_stop() > > > > because we are doing it ourself through devres. > > > > > > > > > > Yeah maybe best to keep the non-devres if it isn't going to affect how the last > > > change in this series is trying to implement with iio. > > > > > > I'll wait for Jonathan to chime in on this thread. > > > > > > > Last, in the HID subsystem, we often interleave non devres with devres > > > > for resource allocation, given that .remove() will be called before any > > > > devres release. But that is assuming this ordering is OK, which doesn't > > > > seem to be the case here. We first need to unregister the i2c adapter > > > > and then close/stop the HID device. > > > > On second thought I2C will be unregistered before the HID calls, since > > unless I'm totally > > incorrect device resource management unwinds backwards in the order actions are > > registered. > > Yeah, sorry if it was not clear: > - .remove() is called *before* any devres action takes place > - devres action are LIFO, so unwinded backwards as you say > > In the general case, a driver does: > int probe() { > void *pointer = devm_alloc(...) > some_more_devm_action(...) > hid_hw_start() > return 0; > } > > and so the HID start action is the last one, meaning that .remove will > first call stop and then devres unwind will get called. > > But here, in your case, you need hid_hw_start to be called *before* > devm_i2c_add_adapter(), meaning that the implicit .remove() will mess > up with the device, so you are forced to do something about it. > > You can either keep a non devm variant, or you can override the > .remove() of HID to not do anything and do the stop/close in a > specific devm task, which you did here. You are just missing the > "let's override .remove() to ensure we keep the device open and > started while we need it". Ok, now I understand! Thanks, Matt > > Cheers, > Benjamin >
On Fri, 23 Sep 2022 14:22:18 -0700 Matt Ranostay <matt.ranostay@konsulko.com> wrote: > On Fri, Sep 23, 2022 at 12:03 AM Benjamin Tissoires > <benjamin.tissoires@redhat.com> wrote: > > > > On Fri, Sep 23, 2022 at 1:45 AM Matt Ranostay > > <matt.ranostay@konsulko.com> wrote: > > > > > > On Wed, Sep 21, 2022 at 10:57 AM Matt Ranostay > > > <matt.ranostay@konsulko.com> wrote: > > > > > > > > On Wed, Sep 21, 2022 at 1:05 AM Benjamin Tissoires > > > > <benjamin.tissoires@redhat.com> wrote: > > > > > > > > > > [foreword: please keep Jiri and myself (the HID maintainers) CC-ed to > > > > > the series, as you will need ack from us and we don't necessarily monitor > > > > > every single message on linux-input] > > > > > > > > > > On Sep 20 2022, Matt Ranostay wrote: > > > > > > Switch from i2c_add_adapter() to resource managed devm_i2c_add_adapter() > > > > > > for matching rest of driver initialization, and more concise code. > > > > > > > > > > > > Signed-off-by: Matt Ranostay <matt.ranostay@konsulko.com> > > > > > > --- > > > > > > drivers/hid/hid-mcp2221.c | 45 +++++++++++++++++---------------------- > > > > > > 1 file changed, 19 insertions(+), 26 deletions(-) > > > > > > > > > > > > diff --git a/drivers/hid/hid-mcp2221.c b/drivers/hid/hid-mcp2221.c > > > > > > index de52e9f7bb8c..7ba63bcd66de 100644 > > > > > > --- a/drivers/hid/hid-mcp2221.c > > > > > > +++ b/drivers/hid/hid-mcp2221.c > > > > > > @@ -824,6 +824,14 @@ static int mcp2221_raw_event(struct hid_device *hdev, > > > > > > return 1; > > > > > > } > > > > > > > > > > > > +static void mcp2221_hid_remove(void *ptr) > > > > > > +{ > > > > > > + struct hid_device *hdev = ptr; > > > > > > + > > > > > > + hid_hw_close(hdev); > > > > > > + hid_hw_stop(hdev); > > > > > > > > > > By default, if you remove the .remove() callback, hid_hw_stop() will get > > > > > automatically called by hid-core.c. So we are now calling it twice, > > > > > which, in a way is not a big deal but it might be an issue in the long > > > > > run. > > > > > > > > > > Generally speaking, in the HID subsystem, that situation doesn't happen > > > > > a lot because hid_hw_start() is usually the last command of probe, and > > > > > we don't need to open the device in the driver itself. > > > > > > > > > > Here, I guess as soon as you add the i2c adapter, you might want to have > > > > > the communication channels ready, and thus you need to have it open > > > > > *before* i2c_add_adapter. > > > > > > > > > > I would suggest the following if you want to keep the devm release of > > > > > stop and close: please put a big fat warning before mcp2221_hid_remove() > > > > > explaining that this is called in devm management, *and* add a function > > > > > that would just return 0 as the .remove() callback with another big fat > > > > > warning explaining that we don't want hid-core.c to call hid_hw_stop() > > > > > because we are doing it ourself through devres. > > > > > > > > > > > > > Yeah maybe best to keep the non-devres if it isn't going to affect how the last > > > > change in this series is trying to implement with iio. > > > > > > > > I'll wait for Jonathan to chime in on this thread. Not my subsystem, so I'm happy if others have to take the headaches that mixing and matching causes :) Personally I'd rather not! Whilst devm_ brings it's own issues (the plumbers session on this was as ever fun) the particular fun set of bugs that turn up because of mixing it in probe() with manual removal in remove() was one where I've never heard a good argument against using devm_ until the first thing in probe() where you decide not to then not using devm_ calls after that. I have seen a handful of cases where a different order was needed, but far more bugs and / or difficult to reason out flows as a result of mixing them up. Obviously straight forward allocations are fine as freeing them late doesn't matter. Registration / consumer requests not so much. Jonathan
diff --git a/drivers/hid/hid-mcp2221.c b/drivers/hid/hid-mcp2221.c index de52e9f7bb8c..7ba63bcd66de 100644 --- a/drivers/hid/hid-mcp2221.c +++ b/drivers/hid/hid-mcp2221.c @@ -824,6 +824,14 @@ static int mcp2221_raw_event(struct hid_device *hdev, return 1; } +static void mcp2221_hid_remove(void *ptr) +{ + struct hid_device *hdev = ptr; + + hid_hw_close(hdev); + hid_hw_stop(hdev); +} + static int mcp2221_probe(struct hid_device *hdev, const struct hid_device_id *id) { @@ -849,7 +857,8 @@ static int mcp2221_probe(struct hid_device *hdev, ret = hid_hw_open(hdev); if (ret) { hid_err(hdev, "can't open device\n"); - goto err_hstop; + hid_hw_stop(hdev); + return ret; } mutex_init(&mcp->lock); @@ -857,6 +866,10 @@ static int mcp2221_probe(struct hid_device *hdev, hid_set_drvdata(hdev, mcp); mcp->hdev = hdev; + ret = devm_add_action_or_reset(&hdev->dev, mcp2221_hid_remove, hdev); + if (ret) + return ret; + /* Set I2C bus clock diviser */ if (i2c_clk_freq > 400) i2c_clk_freq = 400; @@ -873,19 +886,17 @@ static int mcp2221_probe(struct hid_device *hdev, "MCP2221 usb-i2c bridge on hidraw%d", ((struct hidraw *)hdev->hidraw)->minor); - ret = i2c_add_adapter(&mcp->adapter); + ret = devm_i2c_add_adapter(&hdev->dev, &mcp->adapter); if (ret) { hid_err(hdev, "can't add usb-i2c adapter: %d\n", ret); - goto err_i2c; + return ret; } i2c_set_adapdata(&mcp->adapter, mcp); /* Setup GPIO chip */ mcp->gc = devm_kzalloc(&hdev->dev, sizeof(*mcp->gc), GFP_KERNEL); - if (!mcp->gc) { - ret = -ENOMEM; - goto err_gc; - } + if (!mcp->gc) + return -ENOMEM; mcp->gc->label = "mcp2221_gpio"; mcp->gc->direction_input = mcp_gpio_direction_input; @@ -900,26 +911,9 @@ static int mcp2221_probe(struct hid_device *hdev, ret = devm_gpiochip_add_data(&hdev->dev, mcp->gc, mcp); if (ret) - goto err_gc; + return ret; return 0; - -err_gc: - i2c_del_adapter(&mcp->adapter); -err_i2c: - hid_hw_close(mcp->hdev); -err_hstop: - hid_hw_stop(mcp->hdev); - return ret; -} - -static void mcp2221_remove(struct hid_device *hdev) -{ - struct mcp2221 *mcp = hid_get_drvdata(hdev); - - i2c_del_adapter(&mcp->adapter); - hid_hw_close(mcp->hdev); - hid_hw_stop(mcp->hdev); } static const struct hid_device_id mcp2221_devices[] = { @@ -932,7 +926,6 @@ static struct hid_driver mcp2221_driver = { .name = "mcp2221", .id_table = mcp2221_devices, .probe = mcp2221_probe, - .remove = mcp2221_remove, .raw_event = mcp2221_raw_event, };
Switch from i2c_add_adapter() to resource managed devm_i2c_add_adapter() for matching rest of driver initialization, and more concise code. Signed-off-by: Matt Ranostay <matt.ranostay@konsulko.com> --- drivers/hid/hid-mcp2221.c | 45 +++++++++++++++++---------------------- 1 file changed, 19 insertions(+), 26 deletions(-)