On Fri, 2021-02-05 at 13:25 -0400, Jason Gunthorpe wrote: > On Fri, Feb 05, 2021 at 08:48:11AM -0800, James Bottomley wrote: [...] > > The practical consequence of this model is that if you allocate a > > chip structure with tpm_chip_alloc() you have to release it again > > by doing a put of *both* devices. > > The final put of the devs should be directly after the > cdev_device_del(), not in a devm. This became all confused because > the devs was created during alloc, not register. Having a device that > is initialized but will never be added is weird. > > See sketch below. > > > Stefan noticed the latter, so we got the bogus patch 8979b02aaf1d > > ("tpm: Fix reference count to main device") applied which simply > > breaks the master/slave model by not taking a reference on the > > master for the slave. I'm not sure why I didn't notice the problem > > with this fix at the time, but attention must have been elsewhere. > > Well, this is sort of OK because we never use the devs in TPM1, so we > end up freeing the chip with a positive refcount on the devs, which > is weird but not a functional bug. > > Jason > > diff --git a/drivers/char/tpm/tpm-chip.c b/drivers/char/tpm/tpm- > chip.c > index ddaeceb7e10910..e07193a0dd4438 100644 > --- a/drivers/char/tpm/tpm-chip.c > +++ b/drivers/char/tpm/tpm-chip.c > @@ -344,7 +344,6 @@ struct tpm_chip *tpm_chip_alloc(struct device > *pdev, > chip->dev_num = rc; > > device_initialize(&chip->dev); > - device_initialize(&chip->devs); > > chip->dev.class = tpm_class; > chip->dev.class->shutdown_pre = tpm_class_shutdown; > @@ -352,29 +351,12 @@ struct tpm_chip *tpm_chip_alloc(struct device > *pdev, > chip->dev.parent = pdev; > chip->dev.groups = chip->groups; > > - chip->devs.parent = pdev; > - chip->devs.class = tpmrm_class; > - chip->devs.release = tpm_devs_release; > - /* get extra reference on main device to hold on > - * behalf of devs. This holds the chip structure > - * while cdevs is in use. The corresponding put > - * is in the tpm_devs_release (TPM2 only) > - */ > - if (chip->flags & TPM_CHIP_FLAG_TPM2) > - get_device(&chip->dev); > - > if (chip->dev_num == 0) > chip->dev.devt = MKDEV(MISC_MAJOR, TPM_MINOR); > else > chip->dev.devt = MKDEV(MAJOR(tpm_devt), chip->dev_num); > > - chip->devs.devt = > - MKDEV(MAJOR(tpm_devt), chip->dev_num + > TPM_NUM_DEVICES); > - > rc = dev_set_name(&chip->dev, "tpm%d", chip->dev_num); > - if (rc) > - goto out; > - rc = dev_set_name(&chip->devs, "tpmrm%d", chip->dev_num); > if (rc) > goto out; > > @@ -382,9 +364,7 @@ struct tpm_chip *tpm_chip_alloc(struct device > *pdev, > chip->flags |= TPM_CHIP_FLAG_VIRTUAL; > > cdev_init(&chip->cdev, &tpm_fops); > - cdev_init(&chip->cdevs, &tpmrm_fops); > chip->cdev.owner = THIS_MODULE; > - chip->cdevs.owner = THIS_MODULE; > > rc = tpm2_init_space(&chip->work_space, > TPM2_SPACE_BUFFER_SIZE); > if (rc) { > @@ -396,7 +376,6 @@ struct tpm_chip *tpm_chip_alloc(struct device > *pdev, > return chip; > > out: > - put_device(&chip->devs); > put_device(&chip->dev); > return ERR_PTR(rc); > } > @@ -445,13 +424,33 @@ static int tpm_add_char_device(struct tpm_chip > *chip) > } > > if (chip->flags & TPM_CHIP_FLAG_TPM2) { > + device_initialize(&chip->devs); > + chip->devs.parent = pdev; > + chip->devs.class = tpmrm_class; > + rc = dev_set_name(&chip->devs, "tpmrm%d", chip- > >dev_num); > + if (rc) > + goto out_put_devs; > + > + /* > + * get extra reference on main device to hold on > behalf of devs. > + * This holds the chip structure while cdevs is in > use. The > + * corresponding put is in the tpm_devs_release. > + */ > + get_device(&chip->dev); > + chip->devs.release = tpm_devs_release; > + > + chip->devs.devt = > + MKDEV(MAJOR(tpm_devt), chip->dev_num + > TPM_NUM_DEVICES); > + cdev_init(&chip->cdevs, &tpmrm_fops); > + chip->cdevs.owner = THIS_MODULE; > + Effectively all of this shuffles the tpmrm device allocation from chip_alloc to chip_add ... I'm not averse to this but it does mean we can suffer allocation failures now in the add routine and it makes error handling a bit more complex. On the other hand we can now check the TPM2 flag correctly, so it's swings and roundabouts. > rc = cdev_device_add(&chip->cdevs, &chip->devs); > if (rc) { > dev_err(&chip->devs, > "unable to cdev_device_add() %s, major > %d, minor %d, err=%d\n", > dev_name(&chip->devs), MAJOR(chip- > >devs.devt), > MINOR(chip->devs.devt), rc); > - return rc; > + goto out_put_devs; > } > } > > @@ -460,6 +459,10 @@ static int tpm_add_char_device(struct tpm_chip > *chip) > idr_replace(&dev_nums_idr, chip, chip->dev_num); > mutex_unlock(&idr_lock); > > +out_put_devs: > + put_device(&chip->devs); I think there should be a if (chip->flags & TPM_CHIP_FLAG_TPM2) here. I realise you got everything semantically correct and you only ever go to this label from somewhere that already has the check, but guess what will happen when the bot rewriters get hold of this ... > +out_del_dev: > + cdev_device_del(&chip->cdev); > return rc; > } > > @@ -640,8 +643,10 @@ void tpm_chip_unregister(struct tpm_chip *chip) > if (IS_ENABLED(CONFIG_HW_RANDOM_TPM)) > hwrng_unregister(&chip->hwrng); > tpm_bios_log_teardown(chip); > - if (chip->flags & TPM_CHIP_FLAG_TPM2) > + if (chip->flags & TPM_CHIP_FLAG_TPM2) { > cdev_device_del(&chip->cdevs, &chip->devs); > + put_device(&chip->devs); > + } > tpm_del_char_device(chip); Actually, I think you want to go further here. If there's a put_device(&chips->dev) as the last statement (or moved into tpm_del_char_device) we should now have no active reference on the devices from the kernel and we can eliminate the rc = devm_add_action_or_reset(pdev, (void (*)(void *)) put_device, &chip->dev); In tpmm_chip_alloc(). That way both /dev/tpm and /dev/tpmrm have identical lifetime properties. James