Re: mmc: freeing host while host->detect work queue is still active
From: Stephen Warren
Date: Fri Sep 12 2014 - 12:08:55 EST
On 09/11/2014 04:03 PM, Stephen Warren wrote:
Running Fedora rawhides's 3.17.0-0.rc4.git2.1.fc22.armv7hl kernel on
Jetson TK1 (an ARM board containing Tegra SoC), I see the following
during boot most times the Tegra SDHCI driver defers probe for the SD slot:
(and indeed I can reproduce the same issue in plain old v3.17-rc4, so
this isn't at all Fedora specific)
[ 8.377719] sdhci-tegra 700b0400.sdhci: Got CD GPIO #170.
[ 8.377780] sdhci-tegra 700b0400.sdhci: Got WP GPIO #132.
[ 8.377898] mmc1: Unknown controller version (3). You may
experience problems.
[ 8.379796] sdhci-tegra 700b0400.sdhci: No vmmc regulator found
[ 8.380225] ------------[ cut here ]------------
[ 8.380243] WARNING: CPU: 2 PID: 6 at lib/debugobjects.c:263
debug_print_object+0x8c/0xb4()
[ 8.380261] ODEBUG: free active (active state 0) object type:
timer_list hint: delayed_work_timer_fn+0x0/0x18
[ 8.380308] Modules linked in: ehci_tegra(+) sdhci_tegra
sdhci_pltfm sdhci phy_tegra_usb mmc_core i2c_tegra tegra_drm
drm_kms_helper drm host1x
[ 8.380319] CPU: 2 PID: 6 Comm: kworker/u8:0 Not tainted
3.17.0-0.rc4.git2.1.fc22.armv7hl #1
[ 8.380336] Workqueue: deferwq deferred_probe_work_func
[ 8.380358] [<c0218ffc>] (unwind_backtrace) from [<c0212bf0>]
(show_stack+0x18/0x1c)
[ 8.380374] [<c0212bf0>] (show_stack) from [<c094dbe0>]
(dump_stack+0x84/0xb0)
[ 8.380395] [<c094dbe0>] (dump_stack) from [<c0251690>]
(warn_slowpath_common+0x70/0x94)
[ 8.380413] [<c0251690>] (warn_slowpath_common) from [<c02516e8>]
(warn_slowpath_fmt+0x34/0x44)
[ 8.380426] [<c02516e8>] (warn_slowpath_fmt) from [<c057ec6c>]
(debug_print_object+0x8c/0xb4)
[ 8.380439] [<c057ec6c>] (debug_print_object) from [<c057f8d8>]
(debug_check_no_obj_freed+0xf4/0x1dc)
[ 8.380453] [<c057f8d8>] (debug_check_no_obj_freed) from
[<c0398174>] (kfree+0x160/0x308)
[ 8.380465] [<c0398174>] (kfree) from [<c0658094>]
(device_release+0x64/0x98)
[ 8.380479] [<c0658094>] (device_release) from [<c056a8a4>]
(kobject_release+0x12c/0x19c)
[ 8.380498] [<c056a8a4>] (kobject_release) from [<bf2994b0>]
(sdhci_tegra_probe+0x190/0x1bc [sdhci_tegra])
[ 8.380810] [<bf2994b0>] (sdhci_tegra_probe [sdhci_tegra]) from
[<c065e024>] (platform_drv_probe+0x34/0x68)
[ 8.380825] [<c065e024>] (platform_drv_probe) from [<c065c0bc>]
(driver_probe_device+0x13c/0x33c)
[ 8.380838] [<c065c0bc>] (driver_probe_device) from [<c065a5f8>]
(bus_for_each_drv+0x8c/0x9c)
[ 8.380852] [<c065a5f8>] (bus_for_each_drv) from [<c065bf08>]
(device_attach+0x70/0x94)
[ 8.380865] [<c065bf08>] (device_attach) from [<c065b464>]
(bus_probe_device+0x30/0xa4)
[ 8.380878] [<c065b464>] (bus_probe_device) from [<c065b954>]
(deferred_probe_work_func+0x88/0xb8)
[ 8.380894] [<c065b954>] (deferred_probe_work_func) from
[<c026b0e8>] (process_one_work+0x2a0/0x5f8)
[ 8.380908] [<c026b0e8>] (process_one_work) from [<c026c7e4>]
(worker_thread+0x2d0/0x40c)
[ 8.380922] [<c026c7e4>] (worker_thread) from [<c027153c>]
(kthread+0xd4/0xe8)
[ 8.380936] [<c027153c>] (kthread) from [<c020ecc8>]
(ret_from_fork+0x14/0x20)
[ 8.380941] ---[ end trace b1f4b0fe632eb3a4 ]---
The problem is as follows:
The following sequence of calls requests the CD (Change Detect) IRQ for
the SD slot:
sdhci_tegra_probe()
sdhci_tegra_parse_dt()
mmc_of_parse()
mmc_gpio_request_cd()
devm_request_threaded_irq()
The IRQ is triggered by a pin on the SD slot, so could be in any state,
and might fire immediately. This causes the IRQ handler
mmc_gpio_cd_irqt() to run, which calls mmc_detect_change() which calls
mmc_schedule_delayed_work(&host->detect, ...); Thus, the work queue can
be immediately active.
However, if later part of sdhci_tegra_probe() fails, e.g.
sdhci_add_host() -> mmc_regulator_get_supply() ->
devm_regulator_get_optional() -> -EPROBE_DEFER, then sdhci_tegra_probe()
needs to tear everything down, and so calls sdhci_pltfm_free() ->
sdhci_free_host() -> mmc_free_host() -> put_device(&host->class_dev) ->
mmc_host_classdev_release() -> kfree(host). However, host->detect is
part of host and may still be active at this point, which is what
triggers the ODETECT log spew mentioned above.
This doesn't happen in linux-next, because mmc_gpiod_request_cd()
doesn't set up the IRQ handler. Rather this happens inside
mmc_gpiod_request_cd_irq() which is called from mmc_start_host(), which
I believe is well after deferred probe could occur.
Some possible options for fixing this in 3.17 are:
a) Back-port whatever changes in mainline stopped mmc_of_parse()() from
requesting the IRQ. That is commit 740a221ef0e5 "mmc: slot-gpio: Add
GPIO descriptor based CD GPIO API". I haven't investigated how many
other commits would be required for that commit to apply, but I'm
nervous it'd be too many to apply to 3.17.
Sorry, I quoted the wrong commit ID there. That commit is already
present in 3.17 (and some earlier). That commit adds the separate
mmc_gpiod_request_cd_irq() function that's necessary to solve the
problem, and adds a call to it from mmc_start_host(). However, it still
leaves a call to mmc_gpiod_request_cd_irq() in mmc_gpio_request_cd(),
thus leaving the problem in place.
The following commit from linux-next actually solves the problem for me,
by having mmc_of_parse() call mmc_gpiod_request_cd() [which doesn't call
mmc_gpiod_request_cd_irq()] rather than mmc_gpio_request_cd():
98e90de99a0c mmc: host: switch OF parser to use gpio descriptors
That commit relies on at least one of the following two commits in order
to successfully build directly on top of 3.17-rc4:
9d2fa2428ae1 mmc: slot-gpio: add gpiod variant to get wp GPIO
9fbc695075e9 mmc: slot-gpio: switch to use flags when getting GPIO
Is that too much to apply for 3.17 too?
However I note that this only solves the problem for any user of
mmc_of_parse(); any code still calling mmc_gpio_request_cd() directly
could still be exposed to the issue. I wonder if we shouldn't also
remove the call to mmc_gpiod_request_cd_irq() from mmc_gpio_request_cd().
So...
b) Fix the -EPROBE_DEFER cleanup path to explicitly free the IRQ (so it
can't schedule the work queue any more), and then cancel the work queue
(so it isn't active and doesn't trigger the ODETECT message).
A naive patch might be like:
diff --git a/drivers/mmc/core/host.c b/drivers/mmc/core/host.c
index 31969436d77c..8125f916be4a 100644
--- a/drivers/mmc/core/host.c
+++ b/drivers/mmc/core/host.c
@@ -586,6 +586,9 @@ void mmc_free_host(struct mmc_host *host)
idr_remove(&mmc_host_idr, host->index);
spin_unlock(&mmc_host_lock);
+ mmc_gpiod_free_cd(host);
+ cancel_delayed_work_sync(&host->detect);
+
put_device(&host->class_dev);
}
... something like that (but calling a new mmc_gpiod_free_cd_irq())
might still be the best approach. I guess I should just send a patch to
do that?
However, that's problematic, since I think mmc_free_host() would have to
somehow determine whether it needs to call mmc_gpio_free_cd() or
mmc_gpiod_free_cd()? Perhaps we should *just* call devm_free_irq()
there, rather than tearing down all the GPIO stuff too, i.e. split out a
separate mmc_gpio_free_cd_irq() function rather like there's a separate
IRQ request function?
What approach do people prefer and/or does anyone have any other ideas
for a fix?
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/