Merge branch 'bugfix/esp32p4_remove_fence' into 'master'

fix(riscv): Remove the memory barrier when changing interrupt threshold

Closes IDF-7898

See merge request espressif/esp-idf!29119
This commit is contained in:
Omar Chebib 2024-02-21 10:49:18 +08:00
commit 345bf79b4e

View File

@ -60,15 +60,6 @@ FORCE_INLINE_ATTR void __attribute__((always_inline)) rv_utils_wait_for_intr(voi
asm volatile ("wfi\n");
}
/* ------------------------------------------------- Memory Barrier ----------------------------------------------------
*
* ------------------------------------------------------------------------------------------------------------------ */
//TODO: IDF-7898
FORCE_INLINE_ATTR void rv_utils_memory_barrier(void)
{
asm volatile("fence iorw, iorw" : : : "memory");
}
/* -------------------------------------------------- CPU Registers ----------------------------------------------------
*
* ------------------------------------------------------------------------------------------------------------------ */
@ -191,15 +182,11 @@ FORCE_INLINE_ATTR uint32_t __attribute__((always_inline)) rv_utils_set_intlevel(
REG_SET_FIELD(CLIC_INT_THRESH_REG, CLIC_CPU_INT_THRESH, ((intlevel << (8 - NLBITS))) | 0x1f);
/**
* TODO: IDF-7898
* Here is an issue that,
* 1. Set the CLIC_INT_THRESH_REG to mask off interrupts whose level is lower than `intlevel`.
* 2. Set MSTATUS_MIE (global interrupt), then program may jump to interrupt vector.
* 3. The register value change in Step 1 may happen during Step 2.
*
* To prevent this, here a fence is used
* After writing the threshold register, the new threshold is not directly taken into account by the CPU.
* By executing ~8 nop instructions, or by performing a memory load right now, the previous memory write
* operations is forced, making the new threshold active. It is then safe to re-enable MIE bit in mstatus.
*/
rv_utils_memory_barrier();
REG_READ(CLIC_INT_THRESH_REG);
RV_SET_CSR(mstatus, old_mstatus & MSTATUS_MIE);
return old_thresh;