On Thu, Mar 18, 2021 at 11:10:29AM +0100, Fabrice Gasnier wrote:
> On 3/18/21 10:21 AM, Fabrice Gasnier wrote:
> > On 3/14/21 10:08 AM, William Breathitt Gray wrote:
> >> On Sun, Mar 14, 2021 at 04:56:44PM +0900, William Breathitt Gray wrote:
> >>> On Fri, Mar 12, 2021 at 04:02:42PM +0100, Fabrice Gasnier wrote:
> >>>> On 3/9/21 2:19 PM, William Breathitt Gray wrote:
> >>>>> +static ssize_t enums_available_show(const u32 *const enums,
> >>>>> +				    const size_t num_enums,
> >>>>> +				    const char *const strs[], char *buf)
> >>>>> +{
> >>>>> +	size_t len = 0;
> >>>>> +	size_t index;
> >>>>> +
> >>>>> +	for (index = 0; index < num_enums; index++)
> >>>>> +		len += sysfs_emit(buf + len, "%s\n", strs[enums[index]]);
> >>>>> +
> >>>>> +	return len;
> >>>>> +}
> >>>>> +
> >>>>> +static ssize_t strs_available_show(const struct counter_available *const avail,
> >>>>> +				   char *buf)
> >>>>> +{
> >>>>> +	size_t len = 0;
> >>>>> +	size_t index;
> >>>>> +
> >>>>> +	for (index = 0; index < avail->num_items; index++)
> >>>>> +		len += sysfs_emit(buf + len, "%s\n", avail->strs[index]);
> >>>>> +
> >>>>> +	return len;
> >>>>> +}
> >>>>
> >>>> Hi William,
> >>>>
> >>>> I was willing to do some testing on this series, on the stm32 counter
> >>>> drivers, since we released few fixes around them.
> >>>>
> >>>> I tried to apply this series against current testing branch, with few
> >>>> patches applied (so it applies cleanly):
> >>>> - dt-bindings: counter: add interrupt-counter binding
> >>>> - counter: add IRQ or GPIO based counter
> >>>> - counter: stm32-timer-cnt: fix ceiling miss-alignment with reload register
> >>>> - counter: stm32-timer-cnt: fix ceiling write max value
> >>>>  counter: stm32-timer-cnt: Report count function when SLAVE_MODE_DISABLED
> >>>>
> >>>>
> >>>> For both the "stm32-lptimer-cnt" and "stm32-timer-cnt" drivers, I get a
> >>>> warning message and stack dump in "sysfs_emit" when reading the
> >>>> available functions from sysfs.
> >>>> I started to do some testing on v8 of this series last week. I didn't
> >>>> noticed that.
> >>>>
> >>>> For both the "stm32-lptimer-cnt", there are 2 functions currently I get
> >>>> 1 stack dump. Only the "increase" function is printed correctly.
> >>>>
> >>>> For the "stm32-timer-cnt", there are 4 functions currently, I get 3
> >>>> stack dumps. Only the "increase" function is printed correctly
> >>>>
> >>>> Sample log for "stm32-timer-cnt:
> >>>>
> >>>> root@stm32mp1:/sys/devices/platform/soc/44000000.timer/44000000.timer:counter/counter0#
> >>>> cat count0/function_available
> >>>> [ 4689.195506] ------------[ cut here ]------------
> >>>> [ 4689.198747] WARNING: CPU: 1 PID: 5841 at fs/sysfs/file.c:737
> >>>> sysfs_emit+0x88/0x94
> >>>> [ 4689.206233] invalid sysfs_emit: buf:f4a66208
> >>>> [ 4689.210553] Modules linked in: sha256_generic libsha256 sha256_arm
> >>>> cfg80211 panel_orisetech_otm8009a snd_soc_hdmi_codec
> >>>> snd_soc_stm32_sai_sub stm32_lptimers
> >>>> [ 4689.261444] CPU: 1 PID: 5841 Comm: cat Tainted: G        W
> >>>> 5.12.0-rc1 #534
> >>>> [ 4689.268999] Hardware name: STM32 (Device Tree Support)
> >>>> [ 4689.274166] [<c0310b38>] (unwind_backtrace) from [<c030b4ec>]
> >>>> (show_stack+0x10/0x14)
> >>>> [ 4689.281942] [<c030b4ec>] (show_stack) from [<c0fede70>]
> >>>> (dump_stack+0xc0/0xd4)
> >>>> [ 4689.289199] [<c0fede70>] (dump_stack) from [<c0345624>]
> >>>> (__warn+0xec/0x148)
> >>>> [ 4689.296194] [<c0345624>] (__warn) from [<c0fe9e90>]
> >>>> (warn_slowpath_fmt+0x98/0xbc)
> >>>> [ 4689.303714] [<c0fe9e90>] (warn_slowpath_fmt) from [<c0548ee0>]
> >>>> (sysfs_emit+0x88/0x94)
> >>>> [ 4689.311586] [<c0548ee0>] (sysfs_emit) from [<bf115de8>]
> >>>> (counter_comp_available_show+0x11c/0x1a4 [counter])
> >>>> [ 4689.321382] [<bf115de8>] (counter_comp_available_show [counter]) from
> >>>> [<c0a21b70>] (dev_attr_show+0x18/0x48)
> >>>> [ 4689.331263] [<c0a21b70>] (dev_attr_show) from [<c0549014>]
> >>>> (sysfs_kf_seq_show+0x88/0xf0)
> >>>> [ 4689.339394] [<c0549014>] (sysfs_kf_seq_show) from [<c04da6e8>]
> >>>> (seq_read_iter+0x1a4/0x554)
> >>>> [ 4689.347703] [<c04da6e8>] (seq_read_iter) from [<c04af6f0>]
> >>>> (vfs_read+0x1ac/0x2c4)
> >>>> [ 4689.355224] [<c04af6f0>] (vfs_read) from [<c04afc20>]
> >>>> (ksys_read+0x64/0xdc)
> >>>> [ 4689.362219] [<c04afc20>] (ksys_read) from [<c03000c0>]
> >>>> (ret_fast_syscall+0x0/0x58)
> >>>> [ 4689.369827] Exception stack(0xc7261fa8 to 0xc7261ff0)
> >>>> [ 4689.374906] 1fa0:                   00000000 00020000 00000003
> >>>> b6f35000 00020000 00000000
> >>>> [ 4689.383126] 1fc0: 00000000 00020000 b6f56ce0 00000003 00000003
> >>>> 00000000 00020000 00000000
> >>>> [ 4689.391344] 1fe0: 00000003 be8239a8 410bff27 4104c066
> >>>> ...
> >>>> 2 more stack dumps follow
> >>>> ...
> >>>> [ 4689.810479] ---[ end trace 59ed79949efe984c ]---
> >>>> increase
> >>>>
> >>>> I get similar backtrace with other _available attributes:
> >>>> $ cat signal0_action_available
> >>>> $ cat signal1_action_available
> >>>>
> >>>> Do you think I'm doing something wrong ?
> >>>>
> >>>> I tested then "quadrature x4" on the timer driver... It seems all fine.
> >>>>
> >>>> Best regards
> >>>> Fabrice
> >>>>
> >>>>> +
> >>>>> +static ssize_t counter_comp_available_show(struct device *dev,
> >>>>> +					   struct device_attribute *attr,
> >>>>> +					   char *buf)
> >>>>> +{
> >>>>> +	const struct counter_attribute *const a = to_counter_attribute(attr);
> >>>>> +	const struct counter_count *const count = a->parent;
> >>>>> +	const struct counter_synapse *const synapse = a->comp.priv;
> >>>>> +	const struct counter_available *const avail = a->comp.priv;
> >>>>> +
> >>>>> +	switch (a->comp.type) {
> >>>>> +	case COUNTER_COMP_FUNCTION:
> >>>>> +		return enums_available_show(count->functions_list,
> >>>>> +					    count->num_functions,
> >>>>> +					    counter_function_str, buf);
> >>>>> +	case COUNTER_COMP_SYNAPSE_ACTION:
> >>>>> +		return enums_available_show(synapse->actions_list,
> >>>>> +					    synapse->num_actions,
> >>>>> +					    counter_synapse_action_str, buf);
> >>>>> +	case COUNTER_COMP_ENUM:
> >>>>> +		return strs_available_show(avail, buf);
> >>>>> +	case COUNTER_COMP_COUNT_MODE:
> >>>>> +		return enums_available_show(avail->enums, avail->num_items,
> >>>>> +					    counter_count_mode_str, buf);
> >>>>> +	default:
> >>>>> +		return -EINVAL;
> >>>>> +	}
> >>>>> +}
> >>>
> >>> Hi Fabrice,
> >>>
> >>> I can confirm that I'm hitting this regression as well with the
> >>> 104-quad-8 driver. The warning seems to be caused by the
> >>> offset_in_page(buf) check in sysfs_emit(). It looks like the first loop
> >>> in enums_available_show() calls sysfs_emit() correctly, but subsequent
> >>> loops have an invalid buf offset.
> >>>
> >>> The enums_available_show() callback is rather simple: call sysfs_emit()
> >>> for each enum string and increment buf by the length written each time.
> >>> I haven't modified this function since v8, so I am somewhat confused
> >>> about why the buf offset would be invalid here now. I wonder if there
> >>> has been a change somewhere else in the kernel that is causing
> >>> sysfs_emit() to now return an incorrect length.
> >>>
> >>> William Breathitt Gray
> >>
> >> Fabrice,
> >>
> >> Would you be able to check the values of buf and len before they enter
> >> sysfs_emit()? I think redefining the enums_available_show() function
> >> like this should suffice:
> >>
> >> static ssize_t enums_available_show(const u32 *const enums,
> >>                                     const size_t num_enums,
> >>                                     const char *const strs[], char *buf)
> >> {
> >>         size_t len = 0;
> >>         size_t index;
> >>
> >>         for (index = 0; index < num_enums; index++){
> >>                 pr_info("buf: %p\tbuf+len: %p\tlen: %zu\n", buf, buf + len, len);
> >>                 len += sysfs_emit(buf + len, "%s\n", strs[enums[index]]);
> >>         }
> >>
> >>         return len;
> >> }
> >>
> >> I want to see whether the issue is due to the sysfs_emit() return value
> >> or the value of buf.
> > 
> > Hi William,
> > 
> > Sorry for the delay,
> > 
> > I'm getting strange results on buf+len. Here's the result I'm getting
> > with same test as above:
> > 
> > [  170.190995] buf: 5daf3333    buf+len: 5daf3333       len: 0
> > [  170.194383] buf: 5daf3333    buf+len: 22c37039       len: 9
> > [  170.199268] ------------[ cut here ]------------
> > ...
> > [  170.404810] buf: 5daf3333    buf+len: 22c37039       len: 9
> > [  170.409663] ------------[ cut here ]------------
> > ...
> > [  170.615265] buf: 5daf3333    buf+len: 22c37039       len: 9
> > [  170.620117] ------------[ cut here ]------------
> > ...
> > increase
> 
> William,
> 
> I did the same, with %px instead of %p, and i'm getting:
> 
> [  124.001041] buf: c60fb000    buf+len: c60fb000       len: 0
> [  124.009442] buf: c60fb000    buf+len: c60fb009       len: 9
> [  124.019118] ------------[ cut here ]------------
> ...
> So, I believe this is caused by the offset_in_page(buf) check, in
> sysfs_emit().
> 
> I also double checked it on the v8 patchset, and I already had the same
> behavior. So I likely didn't checked the available attrs earlier. Sorry
> for this confusion.
> 
> Best Regards,
> Fabrice

Ah, I forgot %p doesn't show the true address. Okay so it looks like we
can't use sysfs_emit() with an offset. I'll change these sysfs_emit()
lines to use scnprintf() instead and that should prevent the warnings
from triggering.

Thanks,

William Breathitt Gray