For some 32-bit architectures calculation of DTO and STO timeout can be incorrect
due to multiply overflow. Lets prevent this by casting multiply and result to u64.
Suggested by Jisheng Zhang.
Switch DIV_ROUND_UP macro to DIV_ROUND_UP_ULL is not reasonable
because overflow happens on multiply and DIV_ROUND_UP_ULL helps
with sum overflow.
---
Changes since v2:
-add fix for cto_ms
Evgeniy Didin (2):
mmc: dw_mmc: Fix the DTO timeout overflow calculation for 32-bit
systems
mmc: dw_mmc: Fix the CTO overflow calculation for 32-bit systems
drivers/mmc/host/dw_mmc.c | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
--
2.11.0
In commit 4c2357f57dd5 ("mmc: dw_mmc: Fix the CTO timeout calculation")
have been made changes which can cause multiply overflow for 32-bit systems.
The value of cto_ms is lower the drto_ms, but nevertheless overflow can occur.
Lets cast this multiply to u64 type which prevents overflow.
Signed-off-by: Evgeniy Didin <[email protected]>
CC: Alexey Brodkin <[email protected]>
CC: Eugeniy Paltsev <[email protected]>
CC: Douglas Anderson <[email protected]>
CC: Ulf Hansson <[email protected]>
CC: [email protected]
CC: [email protected]
Cc: <[email protected]> # 4c2357f57dd5 mmc: dw_mmc: Fix the CTO timeout calculation
---
drivers/mmc/host/dw_mmc.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/drivers/mmc/host/dw_mmc.c b/drivers/mmc/host/dw_mmc.c
index 194159219b32..775fb3ae1443 100644
--- a/drivers/mmc/host/dw_mmc.c
+++ b/drivers/mmc/host/dw_mmc.c
@@ -409,7 +409,8 @@ static inline void dw_mci_set_cto(struct dw_mci *host)
cto_div = (mci_readl(host, CLKDIV) & 0xff) * 2;
if (cto_div == 0)
cto_div = 1;
- cto_ms = DIV_ROUND_UP(MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
+
+ cto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
/* add a bit spare time */
cto_ms += 10;
--
2.11.0
In commit 9d9491a7da2a ("mmc: dw_mmc: Fix the DTO timeout calculation") have been made
changes which cause multiply overflow for 32-bit systems.
The broken timeout calculations caused false interrupt latency warnings
and stacktrace splat (such as below) when accessing the SD Card.
| Running : 4M-check-reassembly-tcp-cmykw2-rotatew2.out -v0 -w1
| - Info: Finished target initialization.
| mmcblk0: error -110 transferring data, sector 320544, nr 2048, cmd response
| 0x900, card status 0x0
| mmc_host mmc0: Bus speed (slot 0) = 50000000Hz (slot req 400000Hz, actual
| 396825HZ div = 63)
| mmc_host mmc0: Bus speed (slot 0) = 50000000Hz (slot req 25000000Hz, actual
| 25000000HZ div = 1)
| ------------[ cut here ]------------
| softirq: huh, entered softirq 6 TASKLET 6f6a9412 with preempt_count 00000101,
| exited with 00000100?
| WARNING: CPU: 2 PID: 0 at ../lib/scatterlist.c:652 sg_miter_next+0x28/0x20c
| Modules linked in:
| CPU: 2 PID: 0 Comm: swapper/2 Not tainted 4.15.0 #57
|
| Stack Trace:
| arc_unwind_core.constprop.1+0xd0/0xf4
| dump_stack+0x68/0x80
| warn_slowpath_null+0x4e/0xec
| sg_miter_next+0x28/0x20c
| dw_mci_read_data_pio+0x44/0x190
| dw_mmc f000a000.mmc: Unexpected interrupt latency
| dw_mci_interrupt+0x3ee/0x530
| __handle_irq_event_percpu+0x56/0x150
| handle_irq_event+0x34/0x78
| handle_level_irq+0x8e/0x120
| generic_handle_irq+0x1c/0x2c
| idu_cascade_isr+0x30/0x6c
| __handle_domain_irq+0x72/0xc8
| ret_from_exception+0x0/0x8
|---[ end trace 2a58c9af6c25fe51 ]---
Lets cast this multiply to u64 type which prevents overflow.
Tested-by: Vineet Gupta <[email protected]>
Fixes: ARC STAR 9001306872 HSDK, sdio: board crashes when copying big files
Signed-off-by: Evgeniy Didin <[email protected]>
CC: Alexey Brodkin <[email protected]>
CC: Eugeniy Paltsev <[email protected]>
CC: Douglas Anderson <[email protected]>
CC: Ulf Hansson <[email protected]>
CC: [email protected]
CC: [email protected]
Cc: <[email protected]> # 9d9491a7da2a mmc: dw_mmc: Fix the DTO timeout calculation
---
Nothing changed since v2.
drivers/mmc/host/dw_mmc.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/drivers/mmc/host/dw_mmc.c b/drivers/mmc/host/dw_mmc.c
index 0aa39975f33b..194159219b32 100644
--- a/drivers/mmc/host/dw_mmc.c
+++ b/drivers/mmc/host/dw_mmc.c
@@ -1944,7 +1944,8 @@ static void dw_mci_set_drto(struct dw_mci *host)
drto_div = (mci_readl(host, CLKDIV) & 0xff) * 2;
if (drto_div == 0)
drto_div = 1;
- drto_ms = DIV_ROUND_UP(MSEC_PER_SEC * drto_clks * drto_div,
+
+ drto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * drto_clks * drto_div,
host->bus_hz);
/* add a bit spare time */
--
2.11.0
On Mon, Feb 26, 2018 at 4:34 PM, Evgeniy Didin
<[email protected]> wrote:
> In commit 4c2357f57dd5 ("mmc: dw_mmc: Fix the CTO timeout calculation")
> have been made changes which can cause multiply overflow for 32-bit systems.
> The value of cto_ms is lower the drto_ms, but nevertheless overflow can occur.
> Lets cast this multiply to u64 type which prevents overflow.
> - cto_ms = DIV_ROUND_UP(MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
> +
> + cto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
IIRC, someone commented on this or similar, i.e.
DIV_ROUND_UP_ULL() ?
--
With Best Regards,
Andy Shevchenko
On Mon, Feb 26, 2018 at 4:34 PM, Evgeniy Didin
<[email protected]> wrote:
> In commit 9d9491a7da2a ("mmc: dw_mmc: Fix the DTO timeout calculation") have been made
> changes which cause multiply overflow for 32-bit systems.
> The broken timeout calculations caused false interrupt latency warnings
> and stacktrace splat (such as below) when accessing the SD Card.
> + drto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * drto_clks * drto_div,
> host->bus_hz);
Same as per patch 2.
--
With Best Regards,
Andy Shevchenko
On Mon, 2018-02-26 at 16:39 +0200, Andy Shevchenko wrote:
> On Mon, Feb 26, 2018 at 4:34 PM, Evgeniy Didin
> <[email protected]> wrote:
> > In commit 4c2357f57dd5 ("mmc: dw_mmc: Fix the CTO timeout calculation")
> > have been made changes which can cause multiply overflow for 32-bit systems.
> > The value of cto_ms is lower the drto_ms, but nevertheless overflow can occur.
> > Lets cast this multiply to u64 type which prevents overflow.
> > - cto_ms = DIV_ROUND_UP(MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
> > +
> > + cto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
>
> IIRC, someone commented on this or similar, i.e.
>
> DIV_ROUND_UP_ULL() ?
Switch DIV_ROUND_UP macro to DIV_ROUND_UP_ULL is not reasonable
because overflow happens on multiply and DIV_ROUND_UP_ULL helps
with sum overflow.
--
Best regards,
Evgeniy Didin
On Mon, Feb 26, 2018 at 5:14 PM, Evgeniy Didin
<[email protected]> wrote:
> On Mon, 2018-02-26 at 16:39 +0200, Andy Shevchenko wrote:
>> On Mon, Feb 26, 2018 at 4:34 PM, Evgeniy Didin
>> <[email protected]> wrote:
>> > In commit 4c2357f57dd5 ("mmc: dw_mmc: Fix the CTO timeout calculation")
>> > have been made changes which can cause multiply overflow for 32-bit systems.
>> > The value of cto_ms is lower the drto_ms, but nevertheless overflow can occur.
>> > Lets cast this multiply to u64 type which prevents overflow.
>> > - cto_ms = DIV_ROUND_UP(MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
>> > +
>> > + cto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
>>
>> IIRC, someone commented on this or similar, i.e.
>>
>> DIV_ROUND_UP_ULL() ?
> Switch DIV_ROUND_UP macro to DIV_ROUND_UP_ULL is not reasonable
> because overflow happens on multiply and DIV_ROUND_UP_ULL helps
> with sum overflow.
Did you try to compile your code for 32-bit target?
--
With Best Regards,
Andy Shevchenko
On Mon, 2018-02-26 at 18:53 +0200, Andy Shevchenko wrote:
> On Mon, Feb 26, 2018 at 5:14 PM, Evgeniy Didin
> <[email protected]> wrote:
> > On Mon, 2018-02-26 at 16:39 +0200, Andy Shevchenko wrote:
> > > On Mon, Feb 26, 2018 at 4:34 PM, Evgeniy Didin
> > > <[email protected]> wrote:
> > > > In commit 4c2357f57dd5 ("mmc: dw_mmc: Fix the CTO timeout calculation")
> > > > have been made changes which can cause multiply overflow for 32-bit systems.
> > > > The value of cto_ms is lower the drto_ms, but nevertheless overflow can occur.
> > > > Lets cast this multiply to u64 type which prevents overflow.
> > > > - cto_ms = DIV_ROUND_UP(MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
> > > > +
> > > > + cto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
> > >
> > > IIRC, someone commented on this or similar, i.e.
> > >
> > > DIV_ROUND_UP_ULL() ?
> >
> > Switch DIV_ROUND_UP macro to DIV_ROUND_UP_ULL is not reasonable
> > because overflow happens on multiply and DIV_ROUND_UP_ULL helps
> > with sum overflow.
>
> Did you try to compile your code for 32-bit target?
Yes, we have compiled code for 32-bit system.
I am wondering why are you asking that?
--
Best regards,
Evgeniy Didin
On Mon, Feb 26, 2018 at 7:14 PM, Evgeniy Didin
<[email protected]> wrote:
> On Mon, 2018-02-26 at 18:53 +0200, Andy Shevchenko wrote:
>> On Mon, Feb 26, 2018 at 5:14 PM, Evgeniy Didin
>> <[email protected]> wrote:
>> > On Mon, 2018-02-26 at 16:39 +0200, Andy Shevchenko wrote:
>> > > On Mon, Feb 26, 2018 at 4:34 PM, Evgeniy Didin
>> > > <[email protected]> wrote:
>> > > > In commit 4c2357f57dd5 ("mmc: dw_mmc: Fix the CTO timeout calculation")
>> > > > have been made changes which can cause multiply overflow for 32-bit systems.
>> > > > The value of cto_ms is lower the drto_ms, but nevertheless overflow can occur.
>> > > > Lets cast this multiply to u64 type which prevents overflow.
>> > > > - cto_ms = DIV_ROUND_UP(MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
>> > > > +
>> > > > + cto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
>> > >
>> > > IIRC, someone commented on this or similar, i.e.
>> > >
>> > > DIV_ROUND_UP_ULL() ?
>> >
>> > Switch DIV_ROUND_UP macro to DIV_ROUND_UP_ULL is not reasonable
>> > because overflow happens on multiply and DIV_ROUND_UP_ULL helps
>> > with sum overflow.
>>
>> Did you try to compile your code for 32-bit target?
> Yes, we have compiled code for 32-bit system.
> I am wondering why are you asking that?
Because I simple didn't believe you.
ERROR: "__udivdi3" [drivers/mmc/host/dw_mmc.ko] undefined!
...scripts/Makefile.modpost:92: recipe for target '__modpost' failed
make[2]: *** [__modpost] Error 1
+++ b/drivers/mmc/host/dw_mmc.c
@@ -409,7 +409,7 @@ static inline void dw_mci_set_cto(struct dw_mci *host)
- cto_ms = DIV_ROUND_UP(MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
+ cto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * cto_clks * cto_div,
host->bus_hz);
--
With Best Regards,
Andy Shevchenko
Hi Andy,
On Mon, 2018-02-26 at 20:30 +0200, Andy Shevchenko wrote:
> On Mon, Feb 26, 2018 at 7:14 PM, Evgeniy Didin
> <[email protected]> wrote:
> > On Mon, 2018-02-26 at 18:53 +0200, Andy Shevchenko wrote:
> > > On Mon, Feb 26, 2018 at 5:14 PM, Evgeniy Didin
> > > <[email protected]> wrote:
> > > > On Mon, 2018-02-26 at 16:39 +0200, Andy Shevchenko wrote:
> > > > > On Mon, Feb 26, 2018 at 4:34 PM, Evgeniy Didin
> > > > > <[email protected]> wrote:
> > > > > > In commit 4c2357f57dd5 ("mmc: dw_mmc: Fix the CTO timeout calculation")
> > > > > > have been made changes which can cause multiply overflow for 32-bit systems.
> > > > > > The value of cto_ms is lower the drto_ms, but nevertheless overflow can occur.
> > > > > > Lets cast this multiply to u64 type which prevents overflow.
> > > > > > - cto_ms = DIV_ROUND_UP(MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
> > > > > > +
> > > > > > + cto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
> > > > >
> > > > > IIRC, someone commented on this or similar, i.e.
> > > > >
> > > > > DIV_ROUND_UP_ULL() ?
> > > >
> > > > Switch DIV_ROUND_UP macro to DIV_ROUND_UP_ULL is not reasonable
> > > > because overflow happens on multiply and DIV_ROUND_UP_ULL helps
> > > > with sum overflow.
> > >
> > > Did you try to compile your code for 32-bit target?
> >
> > Yes, we have compiled code for 32-bit system.
> > I am wondering why are you asking that?
>
> Because I simple didn't believe you.
Well world around us is much more complicated than it sometimes looks like :)
> ERROR: "__udivdi3" [drivers/mmc/host/dw_mmc.ko] undefined!
> ...scripts/Makefile.modpost:92: recipe for target '__modpost' failed
> make[2]: *** [__modpost] Error 1
That's right __udivdi3() is not defined for some architectures but it is defined for
some others including ARC, Xtensa etc, see
https://elixir.bootlin.com/linux/latest/ident/__udivdi3
What happens __udivdi3() is implemented in libgcc in one form or another form
be it pure assembly or generic implementation written in C.
So maybe we need to add export of __udivdi3() for other arches, what do you think?
-Alexey
> Tested-by: Vineet Gupta <[email protected]>
> Fixes: ARC STAR 9001306872 HSDK, sdio: board crashes when copying big files
>
> Signed-off-by: Evgeniy Didin <[email protected]>
>
> CC: Alexey Brodkin <[email protected]>
> CC: Eugeniy Paltsev <[email protected]>
> CC: Douglas Anderson <[email protected]>
> CC: Ulf Hansson <[email protected]>
> CC: [email protected]
> CC: [email protected]
> Cc: <[email protected]> # 9d9491a7da2a mmc: dw_mmc: Fix the DTO timeout calculation
As I said, the correct tag may be:
Reported-by: Vineet Gupta <[email protected]> # ARC STAR
9001306872 HSDK, sdio: board crashes when copying big files
Tested-by: Vineet Gupta <[email protected]>
Fixes: 9d9491a7da2a ("mmc: dw_mmc: Fix the DTO timeout calculation")
Cc: <[email protected]>
Signed-off-by: Evgeniy Didin <[email protected]>
...
...
> ---
> Nothing changed since v2.
> drivers/mmc/host/dw_mmc.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> - drto_ms = DIV_ROUND_UP(MSEC_PER_SEC * drto_clks * drto_div,
> +
> + drto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * drto_clks * drto_div,
> host->bus_hz);
>
Hmm?
#define DIV_ROUND_DOWN_ULL(ll, d) \
({ unsigned long long _tmp = (ll); do_div(_tmp, d); _tmp; })
#define DIV_ROUND_UP_ULL(ll, d) DIV_ROUND_DOWN_ULL((ll) + (d) - 1, (d))
It uses intermediate unsigned long long _tmp for your "multiply", namely
MSEC_PER_SEC * drto_clks * drto_div, which could solves the problem.
So I don't see why DIV_ROUND_UP_ULL can't work for you?
> /* add a bit spare time */
>
On Mon, 26 Feb 2018 17:34:11 +0300 Evgeniy Didin wrote:
> For some 32-bit architectures calculation of DTO and STO timeout can be incorrect
> due to multiply overflow. Lets prevent this by casting multiply and result to u64.
>
> Suggested by Jisheng Zhang.
> Switch DIV_ROUND_UP macro to DIV_ROUND_UP_ULL is not reasonable
> because overflow happens on multiply and DIV_ROUND_UP_ULL helps
> with sum overflow.
hmmm, I mean something as below:
-cto_ms = DIV_ROUND_UP(MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
+cto_ms = DIV_ROUND_UP_ULL((u64)MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
This could avoid build error in arm case.
>
> ---
> Changes since v2:
> -add fix for cto_ms
>
> Evgeniy Didin (2):
> mmc: dw_mmc: Fix the DTO timeout overflow calculation for 32-bit
> systems
> mmc: dw_mmc: Fix the CTO overflow calculation for 32-bit systems
These two patches could be folded into one patch?
Thanks
On Mon, 26 Feb 2018 20:27:22 +0000 Alexey Brodkin wrote:
> Hi Andy,
>
> On Mon, 2018-02-26 at 20:30 +0200, Andy Shevchenko wrote:
> > On Mon, Feb 26, 2018 at 7:14 PM, Evgeniy Didin
> > <[email protected]> wrote:
> > > On Mon, 2018-02-26 at 18:53 +0200, Andy Shevchenko wrote:
> > > > On Mon, Feb 26, 2018 at 5:14 PM, Evgeniy Didin
> > > > <[email protected]> wrote:
> > > > > On Mon, 2018-02-26 at 16:39 +0200, Andy Shevchenko wrote:
> > > > > > On Mon, Feb 26, 2018 at 4:34 PM, Evgeniy Didin
> > > > > > <[email protected]> wrote:
> > > > > > > In commit 4c2357f57dd5 ("mmc: dw_mmc: Fix the CTO timeout calculation")
> > > > > > > have been made changes which can cause multiply overflow for 32-bit systems.
> > > > > > > The value of cto_ms is lower the drto_ms, but nevertheless overflow can occur.
> > > > > > > Lets cast this multiply to u64 type which prevents overflow.
> > > > > > > - cto_ms = DIV_ROUND_UP(MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
> > > > > > > +
> > > > > > > + cto_ms = DIV_ROUND_UP((u64)MSEC_PER_SEC * cto_clks * cto_div, host->bus_hz);
> > > > > >
> > > > > > IIRC, someone commented on this or similar, i.e.
> > > > > >
> > > > > > DIV_ROUND_UP_ULL() ?
> > > > >
> > > > > Switch DIV_ROUND_UP macro to DIV_ROUND_UP_ULL is not reasonable
> > > > > because overflow happens on multiply and DIV_ROUND_UP_ULL helps
> > > > > with sum overflow.
> > > >
> > > > Did you try to compile your code for 32-bit target?
> > >
> > > Yes, we have compiled code for 32-bit system.
> > > I am wondering why are you asking that?
> >
> > Because I simple didn't believe you.
>
> Well world around us is much more complicated than it sometimes looks like :)
>
> > ERROR: "__udivdi3" [drivers/mmc/host/dw_mmc.ko] undefined!
> > ...scripts/Makefile.modpost:92: recipe for target '__modpost' failed
> > make[2]: *** [__modpost] Error 1
>
> That's right __udivdi3() is not defined for some architectures but it is defined for
> some others including ARC, Xtensa etc, see
> https://elixir.bootlin.com/linux/latest/ident/__udivdi3
>
> What happens __udivdi3() is implemented in libgcc in one form or another form
> be it pure assembly or generic implementation written in C.
>
> So maybe we need to add export of __udivdi3() for other arches, what do you think?
Per my understanding, Linux kernel prefer to make use of do_div or implementations
in <linux/math64.h> for 64bit divide
Thanks
On Tue, Feb 27, 2018 at 5:52 AM, Jisheng Zhang
<[email protected]> wrote:
> On Mon, 26 Feb 2018 20:27:22 +0000 Alexey Brodkin wrote:
>> On Mon, 2018-02-26 at 20:30 +0200, Andy Shevchenko wrote:
>> > On Mon, Feb 26, 2018 at 7:14 PM, Evgeniy Didin
>> > <[email protected]> wrote:
>> > > On Mon, 2018-02-26 at 18:53 +0200, Andy Shevchenko wrote:
>> > > > On Mon, Feb 26, 2018 at 5:14 PM, Evgeniy Didin
>> > > > > > IIRC, someone commented on this or similar, i.e.
>> > > > > > DIV_ROUND_UP_ULL() ?
^^^^^^^^^^^^
>> So maybe we need to add export of __udivdi3() for other arches, what do you think?
> Per my understanding, Linux kernel prefer to make use of do_div or implementations
> in <linux/math64.h> for 64bit divide
To everyone in this thread. See just above. That's solution which will work.
Other than that, drop COMPILE_TEST from the Kconfig and put a comment
that driver will not compile on architectures w/o division library.
--
With Best Regards,
Andy Shevchenko
Hi Evgeniy,
Thank you for the patch! Yet something to improve:
[auto build test ERROR on linus/master]
[also build test ERROR on v4.16-rc3 next-20180228]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system]
url: https://github.com/0day-ci/linux/commits/Evgeniy-Didin/mmc-dw_mmc-Fix-the-DTO-timeout-overflow-calculation-for-32-bit-systems/20180228-141447
config: arm-exynos_defconfig (attached as .config)
compiler: arm-linux-gnueabi-gcc (Debian 7.2.0-11) 7.2.0
reproduce:
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross
chmod +x ~/bin/make.cross
# save the attached .config to linux build tree
make.cross ARCH=arm
All errors (new ones prefixed by >>):
drivers/mmc/host/dw_mmc.o: In function `dw_mci_set_drto':
>> dw_mmc.c:(.text+0x1644): undefined reference to `__aeabi_uldivmod'
---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation