On Sat, 21 May 2022 09:32:37 +0300 Kalle Valo wrote:
> >> >>>>> There are sleep in atomic context bugs when uploading device dump
> >> >>>>> data on usb interface. The root cause is that the operations that
> >> >>>>> may sleep are called in fw_dump_timer_fn which is a timer handler.
> >> >>>>> The call tree shows the execution paths that could lead to bugs:
> >> >>>>>
> >> >>>>> (Interrupt context)
> >> >>>>> fw_dump_timer_fn
> >> >>>>> mwifiex_upload_device_dump
> >> >>>>> dev_coredumpv(..., GFP_KERNEL)
> >> >>
> >> >> just looking at this description, why isn't the simple fix just to
> >> >> change this call to use GFP_ATOMIC?
> >> >
> >> > Because change the parameter of dev_coredumpv() to GFP_ATOMIC could only solve
> >> > partial problem. The following GFP_KERNEL parameters are in /lib/kobject.c
> >> > which is not influenced by dev_coredumpv().
> >> >
> >> > kobject_set_name_vargs
> >> > kvasprintf_const(GFP_KERNEL, ...); //may sleep
> >> > kstrdup(s, GFP_KERNEL); //may sleep
> >> Then it seems there is a problem with dev_coredumpm().
> >> dev_coredumpm() takes a gfp param which means it expects to be called in
> >> any context, but it then calls dev_set_name() which, as you point out,
> >> cannot be called from an atomic context.
> >> So if we cannot change the fact that dev_set_name() cannot be called
> >> from an atomic context, then it would seem to follow that
> >> dev_coredumpv also cannot be called from an atomic
> >> context and hence their gfp param is pointless and should presumably be
> >> removed.
> > Thanks for your time and suggestions! I think the gfp_t parameter of dev_coredumpv and
> > dev_coredumpm may not be removed, because it could be used to pass value to gfp_t
> > parameter of kzalloc in dev_coredumpm. What's more, there are also many other places
> > use dev_coredumpv and dev_coredumpm, if we remove the gfp_t parameter, there are too many
> > places that need to modify and these places are not in interrupt
> > context.
> "Too many users" is not a valid reason to leave a bug in place, either
> dev_coredumpv() should support GFP_ATOMIC or the gfp_t parameter should
> be removed.
The following is one method that letting dev_coredump() support GFP_ATOMIC:
I think dev_set_name() is used to allocate memory to set a device name,
which need only several bytes and there is little chance to sleep in the
real world. However the dev_coredumpv() is used to allocate memory to store
device coredump, which need lots of memory space and have more chances
to sleep. So I think only change the gfp_t parameter of dev_coredumpv()
from GFP_KERNEL to GFP_ATOMIC is ok.
diff --git a/drivers/net/wireless/marvell/mwifiex/main.c b/drivers/net/wireless/marvell/mwifiex/main.c
index ace7371c477..258906920a2 100644
@@ -1116,7 +1116,7 @@ void mwifiex_upload_device_dump(struct mwifiex_adapter *adapter)
"== mwifiex dump information to /sys/class/devcoredump start\n");
dev_coredumpv(adapter->dev, adapter->devdump_data, adapter->devdump_len,
"== mwifiex dump information to /sys/class/devcoredump end\n");
> > There are two solutions now: one is to moves the operations that may
> > sleep into a work item.
> That does not fix the root cause that dev_coredumpv() claims it can be
> called in atomic contexts.
I agree with you. There is not GFP_ATOMIC in lib/kobject.c. Should we modify
the gfp_t parameter in kobject.c in order to support atomic contexts? Do you have
any other good methods?
> > Another is to change the gfp_t parameter of dev_coredumpv from GFP_KERNEL to GFP_ATOMIC, and
> > change the gfp_t parameter of kvasprintf_const and kstrdup from GFP_KERNEL to
> > "in_interrupt() ? GFP_ATOMIC : GFP_KERNEL".
> in_interrupt() is deprecated and should not be used. And I don't think
> it detects all atomic contexts like spinlocks.
I agree with you, the in_interrupt() is not proper.
Thanks for your time and suggestions!