Hello!
On 09/06/2018 03:18 PM, Geert Uytterhoeven wrote:
>> When trying to use CMT for clockevents on R-Car gen3 SoCs, I noticed
>> that 'max_delta_ns' for the broadcast timer (CMT) was shown as 1000 in
>> /proc/timer_list. It turned out that when calculating it, the driver did
>> 1 << 32 (causing what I think was undefined behavior) resulting in a zero
>> delta, later clamped to 1000 by cev_delta2ns(). The root cause turned out
>> to be that the driver abused *unsigned long* for the CMT register values
>> (which are 16/32-bit), so that the calculation of 'ch->max_match_value'
>> in sh_cmt_setup_channel() used the worng branch. Using more proper 'u32'
>> instead fixed 'max_delta_ns' and even fixed the switching an active
>> clocksource to CMT (which caused the system to turn non-interactive
>> before).
>>
>> Signed-off-by: Sergei Shtylyov <sergei.shtylyov@xxxxxxxxxxxxxxxxxx>
>>
>> ---
>> This patch is against the 'tip.git' repo's 'timers/core' branch.
>> The CMT driver wasn't ever used on SH64; on R-Car gen3 (ARM64) I'm only
>> enabling building of this driver now, so not sure how urgent is this...
>>
>> Changes in version 2:
>> - completely redid the patch, fixing abuse of *unsigned long* for the CMT
>> register values.
>
> Thanks for the update!
>
>> --- tip.orig/drivers/clocksource/sh_cmt.c
>> +++ tip/drivers/clocksource/sh_cmt.c
>> @@ -82,14 +82,13 @@ struct sh_cmt_info {
>> unsigned long clear_bits;
>
> "clear_bits" should be u32, as it corresponds to a register value.
OK.
>> @@ -103,9 +102,9 @@ struct sh_cmt_channel {
>>
>> unsigned int timer_bit;
>> unsigned long flags;
>
> While no register value, "flags" can be unsigned int, too.
OK to leave that for another patch?
>> - unsigned long match_value;
>> - unsigned long next_match_value;
>> - unsigned long max_match_value;
>> + u32 match_value;
>> + u32 next_match_value;
>> + u32 max_match_value;
>> raw_spinlock_t lock;
>> struct clock_event_device ced;
>> struct clocksource cs;
>
>> @@ -290,7 +284,7 @@ static inline void sh_cmt_write_cmcor(st
>> static unsigned long sh_cmt_get_counter(struct sh_cmt_channel *ch,
>
> Return type should be "u32".
Thanks, missed it...
>> int *has_wrapped)
>
> While not "unsigned long", "has_wrapped" is used to store a register value,
> so I think it should be changed to "u32 *".
Hm, indeed...
>> {
>> - unsigned long v1, v2, v3;
>> + u32 v1, v2, v3;
>> int o1, o2;
>
> While not "unsigned long", "o1" and "o2" contain register values, so I think
> they should be "u32".
OK, lemme try that...
>> o1 = sh_cmt_read_cmcsr(ch) & ch->cmt->info->overflow_bit;
>
> "overflow_bit" should become "u32", too.
I've figured. :-)
>
>> @@ -619,9 +614,10 @@ static struct sh_cmt_channel *cs_to_sh_c
>> static u64 sh_cmt_clocksource_read(struct clocksource *cs)
>> {
>> struct sh_cmt_channel *ch = cs_to_sh_cmt(cs);
>> - unsigned long flags, raw;
>> + unsigned long flags;
>> unsigned long value;
>> int has_wrapped;
>> + u32 raw;
>>
>> raw_spin_lock_irqsave(&ch->lock, flags);
>> value = ch->total_cycles;
>
> "value" and "total_cycles" should be "u64". But that's a separate bug fix.
And for 32-bit CPUs, I guess? Will look into it...
> Not in the context, so I cannot comment to it, but "sh_cmt_info.width" should
> be "unsigned int", too.
Separate patch?
> Gr{oetje,eeting}s,
>
> Geert
MBR, Sergei
