Hi, > Am 07.06.2017 um 22:44 schrieb Sebastian Reichel <sre@xxxxxxxxxx>: > > Hi, > > On Sun, May 21, 2017 at 12:38:18PM +0200, H. Nikolaus Schaller wrote: >> This fixes an issue if both this twl4030_charger driver and >> phy-twl4030-usb are compiled as modules and loaded in random order. >> It has been observed on GTA04 and OpenPandora devices that in worst >> case the boot process hangs and in best case the AC detection fails >> with a warning. >> >> Therefore we add deferred probing checks for the usb_phy and the >> iio channel for AC detection. >> >> For implementing this we must reorder code because we can't safely >> return -EPROBE_DEFER after allocating any devm managed interrupt >> (it might already/still be enabled without working interrupt handler). >> >> So the check for required resources that may abort probing by >> returning -EPROBE_DEFER, must come first. > > This sounds fishy. EPROBE_DEFER should not be different from > other error codes in this regard and devm_ requested resources > should be free'd on any error. Why is irq handler not working? 1. there is no other error code involved, before we try to convert the driver to handle EPROBE_DEFER. So it is not that EPROBE_DEFER is special but that we add an error return path for it. 2. I don't know why it is not working - I just know that the handler seems to be triggered before all resources are available (if probe() is aborted with EPROBE_DEFER error paths) which ends up in kernel panics. Therefore just to be safe, I have reordered things a little (without changing the function): 1. check for resources (with some EPROBE_DEFER) 2. allocate non-devm (with optional EPROBE_DEFER) 3. allocate devm So this should be safe in any case. Please also compare a discussion https://lkml.org/lkml/2013/2/22/65 > >> Signed-off-by: H. Nikolaus Schaller <hns@xxxxxxxxxxxxx> >> --- >> drivers/power/supply/twl4030_charger.c | 37 ++++++++++++++++++++-------------- >> 1 file changed, 22 insertions(+), 15 deletions(-) >> >> diff --git a/drivers/power/supply/twl4030_charger.c b/drivers/power/supply/twl4030_charger.c >> index 785a07bc4f39..945eabdbbc89 100644 >> --- a/drivers/power/supply/twl4030_charger.c >> +++ b/drivers/power/supply/twl4030_charger.c >> @@ -984,6 +984,28 @@ static int twl4030_bci_probe(struct platform_device *pdev) >> >> platform_set_drvdata(pdev, bci); >> >> + if (bci->dev->of_node) { >> + struct device_node *phynode; >> + >> + phynode = of_find_compatible_node(bci->dev->of_node->parent, >> + NULL, "ti,twl4030-usb"); >> + if (phynode) { >> + bci->transceiver = devm_usb_get_phy_by_node( >> + bci->dev, phynode, &bci->usb_nb); >> + if (IS_ERR(bci->transceiver) && >> + PTR_ERR(bci->transceiver) == -EPROBE_DEFER) >> + return -EPROBE_DEFER; /* PHY not ready */ >> + } >> + } >> + >> + bci->channel_vac = iio_channel_get(&pdev->dev, "vac"); >> + if (IS_ERR(bci->channel_vac)) { >> + if (PTR_ERR(bci->channel_vac) == -EPROBE_DEFER) >> + return -EPROBE_DEFER; /* iio not ready */ >> + dev_warn(&pdev->dev, "could not request vac iio channel"); >> + bci->channel_vac = NULL; >> + } >> + > > You should not request non-devm managed iio_channel before > devm-managed power-supply. Well, it is not *me* doing that. It is the unpatched driver which already does. See: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/power/supply/twl4030_charger.c?h=v4.12-rc4#n1069 I have just moved this line to a different location to be able to add a proper EPROBE_DEFER return path. > This could be fixed by switching to > devm_iio_channel_get(), which also cleans up some code. Yes, this could be an alternative solution. We still need EPROBE_DEFER handling. > > I suspect, that this is also your IRQ problem, since iio_channel > is currently free'd before irqs are free'd, but its used in irq > code. Hm. No. In upstream code it is never freed on probe failure because there is only an error return (goto fail) after allocating irqs in case the twl_i2c_write_u8 fails. Which usually doesn't. The EPROBE_DEFER returned by iio_channel_get not being ready simply prints a warning and turns off AC detection (bci->channel_vac = NULL) forever. > >> bci->ac = devm_power_supply_register(&pdev->dev, &twl4030_bci_ac_desc, >> NULL); >> if (IS_ERR(bci->ac)) { >> @@ -1017,25 +1039,10 @@ static int twl4030_bci_probe(struct platform_device *pdev) >> return ret; >> } >> >> - bci->channel_vac = iio_channel_get(&pdev->dev, "vac"); >> - if (IS_ERR(bci->channel_vac)) { My first attempt to fix EPROBE_DEFER was to add this + if (PTR_ERR(bci->channel_vac) == -EPROBE_DEFER) + return -EPROBE_DEFER; /* iio not ready */ >> - bci->channel_vac = NULL; >> - dev_warn(&pdev->dev, "could not request vac iio channel"); >> - } >> - >> INIT_WORK(&bci->work, twl4030_bci_usb_work); >> INIT_DELAYED_WORK(&bci->current_worker, twl4030_current_worker); >> >> bci->usb_nb.notifier_call = twl4030_bci_usb_ncb; >> - if (bci->dev->of_node) { >> - struct device_node *phynode; >> - >> - phynode = of_find_compatible_node(bci->dev->of_node->parent, >> - NULL, "ti,twl4030-usb"); >> - if (phynode) >> - bci->transceiver = devm_usb_get_phy_by_node( >> - bci->dev, phynode, &bci->usb_nb); and this + if (IS_ERR(bci->transceiver) && + PTR_ERR(bci->transceiver) == -EPROBE_DEFER) { + ret = -EPROBE_DEFER; /* PHY not ready */ + goto fail; This did run (and potentially return) after installing devm_request_threaded_irq leading to observation of severe kernel panics by interrupts. Moving the iio channel allocation before devm_request_threaded_irq made them go away. >> - } >> >> /* Enable interrupts now. */ >> reg = ~(u32)(TWL4030_ICHGLOW | TWL4030_ICHGEOC | TWL4030_TBATOR2 | > > -- Sebastian How important is it to keep the current sequence of devm_request_threaded_irq() + (devm_)iio_channel_get() untouched? The reason I ask is that it is more likely for (devm_)iio_channel_get() to fail than devm_request_threaded_irq(). Hence the iio channel should imho be always allocated first, so that we skip allocating+freeing unused irq handlers in case of iio not being ready. Therefore, I suggest to keep the proposed reordering even if we think devm-conversion of iio_channel_get() alone would solve it equally well. Next: should there be two separate patches? One for converting the non-devm iio_channel_get() and cleaning up error path. And one for adding EPROBE_DEFER error path. Or keep it in a single patch because they only make sense if done together (you can't add/remove deferred probing without converting and/or moving iio_channel_get())? So please advise how to proceed. BR and thanks, Nikolaus
Attachment:
signature.asc
Description: Message signed with OpenPGP using GPGMail
