[PATCH RFC 00/10] ARM: V7M: Support caches

[Alexandre Torgue]

Hi Vladimir,

On 04/21/2016 10:18 AM, Vladimir Murzin wrote:
> Hi,
>
> This patch set allows M-class cpus benefit of optional cache support.
> It originaly was written by Jonny, I've been keeping it localy mainly
> rebasing over Linux versions.
>
> The main idea behind patches is to reuse existing cache handling code
> from v7A/R. In case v7M cache operations are provided via memory
> mapped interface rather than co-processor instructions, so extra
> macros have been introduced to factor out cache handling logic and
> low-level operations.
>
> Along with the v7M cache support the first user (Cortex-M7) is
> introduced.
>
> Patches were tested on MPS2 platform with Cortex-M3/M4/M7. The later
> one showed significant boot speed-up.
>


Thanks for the series. Maxime and I have just tested it on STM32F7 hardware.
We found an issue on this M7 chip, which is not linked to your patches 
(see below patch that will be send shortly).
Once fixed, we are able to boot but we observe instabilities which 
disappear when D-cache is disabled from kernel config.
We observe random crashes.
Most of them look like this (hardfault):

(gdb) info reg
r0             0xc018c3fc    -1072118788
r1             0x3e0    992
r2             0x4100000b    1090519051
r3             0x0    0
r4             0x0    0
r5             0xc018c3fc    -1072118788
r6             0xc018c7dc    -1072117796
r7             0xc018c3fc    -1072118788
r8             0xc018bb30    -1072121040
r9             0x3e0    992
r10            0x4100000b    1090519051
r11            0x0    0
r12            0x8303c003    -2096906237
sp             0xc016bed0    0xc016bed0 <init_thread_union+7888>
lr             0xfffffff1    -15
pc             0xc000b6bc    0xc000b6bc <__invalid_entry>
xPSR           0x81000003    -2130706429
(gdb) bt
#0  __invalid_entry () at arch/arm/kernel/entry-v7m.S:24
#1  <signal handler called>
#2  vsnprintf (buf=0xc018c3fc <textbuf> "", size=992, fmt=0x4100000b "", 
args=...) at lib/vsprintf.c:1999
#3  0xc0095910 in vscnprintf (buf=<optimized out>, size=992, 
fmt=<optimized out>, args=...) at lib/vsprintf.c:2152
#4  0xc00269dc in vprintk_emit (facility=0, level=-1, dict=0x0, 
dictlen=<optimized out>, fmt=0x4100000b "", args=...) at 
kernel/printk/printk.c:1673
#5  0xc000a20a in arch_local_irq_enable () at 
./arch/arm/include/asm/irqflags.h:38
#6  arch_cpu_idle () at arch/arm/kernel/process.c:73
#7  0x00000000 in ?? ()

We continue to investigate this issue. If you have any idea, please let 
us know.

Thanks.

alex

Patch which will be sent shortly:


Author: Alexandre TORGUE <alexandre.torgue at gmail.com>
Date:   Wed May 25 14:02:35 2016 +0200

     ARM: V7M: Add dsb before jumping in handler mode

     According to ARM AN321 (section 4.12):

     "If the vector table is in writable memory such as SRAM, either 
relocated
     by VTOR or a device dependent memory remapping mechanism, then
     architecturally a memory barrier instruction is required after the 
vector
     table entry is updated, and if the exception is to be activated 
immediately"

     Signed-off-by: Maxime Coquelin <mcoquelin.stm32 at gmail.com>
     Signed-off-by: Alexandre TORGUE <alexandre.torgue at gmail.com>

diff --git a/arch/arm/mm/proc-v7m.S b/arch/arm/mm/proc-v7m.S
index 60387a0..807949f 100644
--- a/arch/arm/mm/proc-v7m.S
+++ b/arch/arm/mm/proc-v7m.S
@@ -124,6 +124,7 @@ __v7m_setup:
         badr    r1, 1f
         ldr     r5, [r12, #11 * 4]      @ read the SVC vector entry
         str     r1, [r12, #11 * 4]      @ write the temporary SVC 
vector entry
+       dsb
         mov     r6, lr                  @ save LR
         ldr     sp, =init_thread_union + THREAD_START_SP
         cpsie   i