* [PATCH 0/2] MIPS Booting fixes @ 2023-02-25 22:10 Jiaxun Yang 2023-02-25 22:10 ` [PATCH 1/2] MIPS: smp-cps: Don't rely on CP0_CMGCRBASE Jiaxun Yang 2023-02-25 22:10 ` [PATCH 2/2] MIPS: cevt-r4k: Offset counter value for clearing compare interrupt Jiaxun Yang 0 siblings, 2 replies; 10+ messages in thread From: Jiaxun Yang @ 2023-02-25 22:10 UTC (permalink / raw) To: linux-mips; +Cc: tsbogend, Jiaxun Yang Hi all, This patchset fixes two issues that was found when doing reboot stress test on Malta/Boston board with various MIPS cores. Perhaps they should go through the mips-fixes tree. Thanks Jiaxun Yang (2): MIPS: smp-cps: Don't rely on CP0_CMGCRBASE MIPS: cevt-r4k: Offset counter value for clearing compare interrupt arch/mips/include/asm/smp-cps.h | 4 ++++ arch/mips/kernel/cevt-r4k.c | 3 +++ arch/mips/kernel/cps-vec.S | 35 ++++++++++++++------------------- arch/mips/kernel/smp-cps.c | 2 ++ 4 files changed, 24 insertions(+), 20 deletions(-) -- 2.37.1 (Apple Git-137.1) ^ permalink raw reply [flat|nested] 10+ messages in thread
* [PATCH 1/2] MIPS: smp-cps: Don't rely on CP0_CMGCRBASE 2023-02-25 22:10 [PATCH 0/2] MIPS Booting fixes Jiaxun Yang @ 2023-02-25 22:10 ` Jiaxun Yang 2023-02-27 11:59 ` Thomas Bogendoerfer 2023-02-25 22:10 ` [PATCH 2/2] MIPS: cevt-r4k: Offset counter value for clearing compare interrupt Jiaxun Yang 1 sibling, 1 reply; 10+ messages in thread From: Jiaxun Yang @ 2023-02-25 22:10 UTC (permalink / raw) To: linux-mips; +Cc: tsbogend, Jiaxun Yang CP0_CMGCRBASE is not always available on CPS enabled system such as early proAptiv. We patch the entry of CPS NMI vector to inject CMGCR address directly into register during early core bringup. For VPE bringup as the core is already coherenct at that point we just read the variable to obtain the address. Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com> --- arch/mips/include/asm/smp-cps.h | 4 ++++ arch/mips/kernel/cps-vec.S | 35 ++++++++++++++------------------- arch/mips/kernel/smp-cps.c | 2 ++ 3 files changed, 21 insertions(+), 20 deletions(-) diff --git a/arch/mips/include/asm/smp-cps.h b/arch/mips/include/asm/smp-cps.h index 7e5b9411faee..22a572b70fe3 100644 --- a/arch/mips/include/asm/smp-cps.h +++ b/arch/mips/include/asm/smp-cps.h @@ -7,6 +7,8 @@ #ifndef __MIPS_ASM_SMP_CPS_H__ #define __MIPS_ASM_SMP_CPS_H__ +#define CPS_ENTRY_PATCH_INSNS 6 + #ifndef __ASSEMBLY__ struct vpe_boot_config { @@ -30,6 +32,8 @@ extern void mips_cps_boot_vpes(struct core_boot_config *cfg, unsigned vpe); extern void mips_cps_pm_save(void); extern void mips_cps_pm_restore(void); +extern void *mips_cps_core_entry_patch_end; + #ifdef CONFIG_MIPS_CPS extern bool mips_cps_smp_in_use(void); diff --git a/arch/mips/kernel/cps-vec.S b/arch/mips/kernel/cps-vec.S index 975343240148..fece0cca5548 100644 --- a/arch/mips/kernel/cps-vec.S +++ b/arch/mips/kernel/cps-vec.S @@ -13,6 +13,7 @@ #include <asm/mipsregs.h> #include <asm/mipsmtregs.h> #include <asm/pm.h> +#include <asm/smp-cps.h> #define GCR_CPC_BASE_OFS 0x0088 #define GCR_CL_COHERENCE_OFS 0x2008 @@ -80,25 +81,20 @@ nop .endm - /* Calculate an uncached address for the CM GCRs */ - .macro cmgcrb dest - .set push - .set noat - MFC0 $1, CP0_CMGCRBASE - PTR_SLL $1, $1, 4 - PTR_LI \dest, UNCAC_BASE - PTR_ADDU \dest, \dest, $1 - .set pop - .endm .balign 0x1000 LEAF(mips_cps_core_entry) /* - * These first 4 bytes will be patched by cps_smp_setup to load the - * CCA to use into register s0. + * These first several instructions will be patched by cps_smp_setup to load the + * CCA to use into register s0 and GCR base address to register s1. */ - .word 0 + .rept CPS_ENTRY_PATCH_INSNS + nop + .endr + + .global mips_cps_core_entry_patch_end +mips_cps_core_entry_patch_end: /* Check whether we're here due to an NMI */ mfc0 k0, CP0_STATUS @@ -121,8 +117,7 @@ not_nmi: mtc0 t0, CP0_STATUS /* Skip cache & coherence setup if we're already coherent */ - cmgcrb v1 - lw s7, GCR_CL_COHERENCE_OFS(v1) + lw s7, GCR_CL_COHERENCE_OFS(s1) bnez s7, 1f nop @@ -132,7 +127,7 @@ not_nmi: /* Enter the coherent domain */ li t0, 0xff - sw t0, GCR_CL_COHERENCE_OFS(v1) + sw t0, GCR_CL_COHERENCE_OFS(s1) ehb /* Set Kseg0 CCA to that in s0 */ @@ -305,8 +300,7 @@ LEAF(mips_cps_core_init) */ LEAF(mips_cps_get_bootcfg) /* Calculate a pointer to this cores struct core_boot_config */ - cmgcrb t0 - lw t0, GCR_CL_ID_OFS(t0) + lw t0, GCR_CL_ID_OFS(s1) li t1, COREBOOTCFG_SIZE mul t0, t0, t1 PTR_LA t1, mips_cps_core_bootcfg @@ -366,8 +360,9 @@ LEAF(mips_cps_boot_vpes) has_vp t0, 5f /* Find base address of CPC */ - cmgcrb t3 - PTR_L t1, GCR_CPC_BASE_OFS(t3) + PTR_LA t1, mips_gcr_base + PTR_L t1, 0(t1) + PTR_L t1, GCR_CPC_BASE_OFS(t1) PTR_LI t2, ~0x7fff and t1, t1, t2 PTR_LI t2, UNCAC_BASE diff --git a/arch/mips/kernel/smp-cps.c b/arch/mips/kernel/smp-cps.c index f2df0cae1b4d..4fc288bb85b9 100644 --- a/arch/mips/kernel/smp-cps.c +++ b/arch/mips/kernel/smp-cps.c @@ -162,6 +162,8 @@ static void __init cps_prepare_cpus(unsigned int max_cpus) */ entry_code = (u32 *)&mips_cps_core_entry; uasm_i_addiu(&entry_code, 16, 0, cca); + UASM_i_LA(&entry_code, 17, (long)mips_gcr_base); + BUG_ON((void *)entry_code > (void *)&mips_cps_core_entry_patch_end); blast_dcache_range((unsigned long)&mips_cps_core_entry, (unsigned long)entry_code); bc_wback_inv((unsigned long)&mips_cps_core_entry, -- 2.37.1 (Apple Git-137.1) ^ permalink raw reply related [flat|nested] 10+ messages in thread
* Re: [PATCH 1/2] MIPS: smp-cps: Don't rely on CP0_CMGCRBASE 2023-02-25 22:10 ` [PATCH 1/2] MIPS: smp-cps: Don't rely on CP0_CMGCRBASE Jiaxun Yang @ 2023-02-27 11:59 ` Thomas Bogendoerfer 0 siblings, 0 replies; 10+ messages in thread From: Thomas Bogendoerfer @ 2023-02-27 11:59 UTC (permalink / raw) To: Jiaxun Yang; +Cc: linux-mips On Sat, Feb 25, 2023 at 10:10:07PM +0000, Jiaxun Yang wrote: > CP0_CMGCRBASE is not always available on CPS enabled system > such as early proAptiv. > > We patch the entry of CPS NMI vector to inject CMGCR address > directly into register during early core bringup. > > For VPE bringup as the core is already coherenct at that point > we just read the variable to obtain the address. > > Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com> > --- > arch/mips/include/asm/smp-cps.h | 4 ++++ > arch/mips/kernel/cps-vec.S | 35 ++++++++++++++------------------- > arch/mips/kernel/smp-cps.c | 2 ++ > 3 files changed, 21 insertions(+), 20 deletions(-) > > diff --git a/arch/mips/include/asm/smp-cps.h b/arch/mips/include/asm/smp-cps.h > index 7e5b9411faee..22a572b70fe3 100644 > --- a/arch/mips/include/asm/smp-cps.h > +++ b/arch/mips/include/asm/smp-cps.h > @@ -7,6 +7,8 @@ > #ifndef __MIPS_ASM_SMP_CPS_H__ > #define __MIPS_ASM_SMP_CPS_H__ > > +#define CPS_ENTRY_PATCH_INSNS 6 > + > #ifndef __ASSEMBLY__ > > struct vpe_boot_config { > @@ -30,6 +32,8 @@ extern void mips_cps_boot_vpes(struct core_boot_config *cfg, unsigned vpe); > extern void mips_cps_pm_save(void); > extern void mips_cps_pm_restore(void); > > +extern void *mips_cps_core_entry_patch_end; > + > #ifdef CONFIG_MIPS_CPS > > extern bool mips_cps_smp_in_use(void); > diff --git a/arch/mips/kernel/cps-vec.S b/arch/mips/kernel/cps-vec.S > index 975343240148..fece0cca5548 100644 > --- a/arch/mips/kernel/cps-vec.S > +++ b/arch/mips/kernel/cps-vec.S > @@ -13,6 +13,7 @@ > #include <asm/mipsregs.h> > #include <asm/mipsmtregs.h> > #include <asm/pm.h> > +#include <asm/smp-cps.h> > > #define GCR_CPC_BASE_OFS 0x0088 > #define GCR_CL_COHERENCE_OFS 0x2008 > @@ -80,25 +81,20 @@ > nop > .endm > > - /* Calculate an uncached address for the CM GCRs */ > - .macro cmgcrb dest > - .set push > - .set noat > - MFC0 $1, CP0_CMGCRBASE > - PTR_SLL $1, $1, 4 > - PTR_LI \dest, UNCAC_BASE > - PTR_ADDU \dest, \dest, $1 > - .set pop > - .endm > > .balign 0x1000 > > LEAF(mips_cps_core_entry) > /* > - * These first 4 bytes will be patched by cps_smp_setup to load the > - * CCA to use into register s0. > + * These first several instructions will be patched by cps_smp_setup to load the > + * CCA to use into register s0 and GCR base address to register s1. > */ > - .word 0 > + .rept CPS_ENTRY_PATCH_INSNS > + nop > + .endr > + > + .global mips_cps_core_entry_patch_end > +mips_cps_core_entry_patch_end: > > /* Check whether we're here due to an NMI */ > mfc0 k0, CP0_STATUS > @@ -121,8 +117,7 @@ not_nmi: > mtc0 t0, CP0_STATUS > > /* Skip cache & coherence setup if we're already coherent */ > - cmgcrb v1 > - lw s7, GCR_CL_COHERENCE_OFS(v1) > + lw s7, GCR_CL_COHERENCE_OFS(s1) > bnez s7, 1f > nop > > @@ -132,7 +127,7 @@ not_nmi: > > /* Enter the coherent domain */ > li t0, 0xff > - sw t0, GCR_CL_COHERENCE_OFS(v1) > + sw t0, GCR_CL_COHERENCE_OFS(s1) > ehb > > /* Set Kseg0 CCA to that in s0 */ > @@ -305,8 +300,7 @@ LEAF(mips_cps_core_init) > */ > LEAF(mips_cps_get_bootcfg) > /* Calculate a pointer to this cores struct core_boot_config */ > - cmgcrb t0 > - lw t0, GCR_CL_ID_OFS(t0) > + lw t0, GCR_CL_ID_OFS(s1) > li t1, COREBOOTCFG_SIZE > mul t0, t0, t1 > PTR_LA t1, mips_cps_core_bootcfg > @@ -366,8 +360,9 @@ LEAF(mips_cps_boot_vpes) > has_vp t0, 5f > > /* Find base address of CPC */ > - cmgcrb t3 > - PTR_L t1, GCR_CPC_BASE_OFS(t3) > + PTR_LA t1, mips_gcr_base > + PTR_L t1, 0(t1) > + PTR_L t1, GCR_CPC_BASE_OFS(t1) > PTR_LI t2, ~0x7fff > and t1, t1, t2 > PTR_LI t2, UNCAC_BASE > diff --git a/arch/mips/kernel/smp-cps.c b/arch/mips/kernel/smp-cps.c > index f2df0cae1b4d..4fc288bb85b9 100644 > --- a/arch/mips/kernel/smp-cps.c > +++ b/arch/mips/kernel/smp-cps.c > @@ -162,6 +162,8 @@ static void __init cps_prepare_cpus(unsigned int max_cpus) > */ > entry_code = (u32 *)&mips_cps_core_entry; > uasm_i_addiu(&entry_code, 16, 0, cca); > + UASM_i_LA(&entry_code, 17, (long)mips_gcr_base); > + BUG_ON((void *)entry_code > (void *)&mips_cps_core_entry_patch_end); > blast_dcache_range((unsigned long)&mips_cps_core_entry, > (unsigned long)entry_code); > bc_wback_inv((unsigned long)&mips_cps_core_entry, why do you need to patch that in ? Would replacing cmgcrb by a PTR_LA t1, mips_gcr_base in arch/mips/kernel/cps-vec.S do the same thing ? Thomas. -- Crap can work. Given enough thrust pigs will fly, but it's not necessarily a good idea. [ RFC1925, 2.3 ] ^ permalink raw reply [flat|nested] 10+ messages in thread
* [PATCH 2/2] MIPS: cevt-r4k: Offset counter value for clearing compare interrupt 2023-02-25 22:10 [PATCH 0/2] MIPS Booting fixes Jiaxun Yang 2023-02-25 22:10 ` [PATCH 1/2] MIPS: smp-cps: Don't rely on CP0_CMGCRBASE Jiaxun Yang @ 2023-02-25 22:10 ` Jiaxun Yang 2023-02-26 22:04 ` Philippe Mathieu-Daudé 2023-02-26 23:23 ` Thomas Bogendoerfer 1 sibling, 2 replies; 10+ messages in thread From: Jiaxun Yang @ 2023-02-25 22:10 UTC (permalink / raw) To: linux-mips; +Cc: tsbogend, Jiaxun Yang In c0_compare_int_usable we clear compare interrupt by write value just read out from counter to compare register. However sometimes if those all instructions are graduated together then it's possible that at the time compare register is written, the counter haven't progressed, thus the interrupt is triggered again. It also applies to QEMU that instructions is execuated significantly faster then counter. Offset the counter value a litlle bit to prevent that happen. Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com> --- arch/mips/kernel/cevt-r4k.c | 3 +++ 1 file changed, 3 insertions(+) diff --git a/arch/mips/kernel/cevt-r4k.c b/arch/mips/kernel/cevt-r4k.c index 32ec67c9ab67..bbc422376e97 100644 --- a/arch/mips/kernel/cevt-r4k.c +++ b/arch/mips/kernel/cevt-r4k.c @@ -200,6 +200,8 @@ int c0_compare_int_usable(void) */ if (c0_compare_int_pending()) { cnt = read_c0_count(); + // Drawdown a little bit in case counter haven't progressed + cnt -= COMPARE_INT_SEEN_TICKS; write_c0_compare(cnt); back_to_back_c0_hazard(); while (read_c0_count() < (cnt + COMPARE_INT_SEEN_TICKS)) @@ -228,6 +230,7 @@ int c0_compare_int_usable(void) if (!c0_compare_int_pending()) return 0; cnt = read_c0_count(); + cnt -= COMPARE_INT_SEEN_TICKS; write_c0_compare(cnt); back_to_back_c0_hazard(); while (read_c0_count() < (cnt + COMPARE_INT_SEEN_TICKS)) -- 2.37.1 (Apple Git-137.1) ^ permalink raw reply related [flat|nested] 10+ messages in thread
* Re: [PATCH 2/2] MIPS: cevt-r4k: Offset counter value for clearing compare interrupt 2023-02-25 22:10 ` [PATCH 2/2] MIPS: cevt-r4k: Offset counter value for clearing compare interrupt Jiaxun Yang @ 2023-02-26 22:04 ` Philippe Mathieu-Daudé 2023-02-26 23:23 ` Thomas Bogendoerfer 1 sibling, 0 replies; 10+ messages in thread From: Philippe Mathieu-Daudé @ 2023-02-26 22:04 UTC (permalink / raw) To: Jiaxun Yang, linux-mips; +Cc: tsbogend On 25/2/23 23:10, Jiaxun Yang wrote: > In c0_compare_int_usable we clear compare interrupt by write value > just read out from counter to compare register. > > However sometimes if those all instructions are graduated together > then it's possible that at the time compare register is written, the > counter haven't progressed, thus the interrupt is triggered again. > > It also applies to QEMU that instructions is execuated significantly Typo "executed", > faster then counter. > > Offset the counter value a litlle bit to prevent that happen. "little". > > Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com> > --- > arch/mips/kernel/cevt-r4k.c | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/arch/mips/kernel/cevt-r4k.c b/arch/mips/kernel/cevt-r4k.c > index 32ec67c9ab67..bbc422376e97 100644 > --- a/arch/mips/kernel/cevt-r4k.c > +++ b/arch/mips/kernel/cevt-r4k.c > @@ -200,6 +200,8 @@ int c0_compare_int_usable(void) > */ > if (c0_compare_int_pending()) { > cnt = read_c0_count(); > + // Drawdown a little bit in case counter haven't progressed > + cnt -= COMPARE_INT_SEEN_TICKS; > write_c0_compare(cnt); > back_to_back_c0_hazard(); > while (read_c0_count() < (cnt + COMPARE_INT_SEEN_TICKS)) > @@ -228,6 +230,7 @@ int c0_compare_int_usable(void) > if (!c0_compare_int_pending()) > return 0; > cnt = read_c0_count(); > + cnt -= COMPARE_INT_SEEN_TICKS; > write_c0_compare(cnt); > back_to_back_c0_hazard(); > while (read_c0_count() < (cnt + COMPARE_INT_SEEN_TICKS)) ^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: [PATCH 2/2] MIPS: cevt-r4k: Offset counter value for clearing compare interrupt 2023-02-25 22:10 ` [PATCH 2/2] MIPS: cevt-r4k: Offset counter value for clearing compare interrupt Jiaxun Yang 2023-02-26 22:04 ` Philippe Mathieu-Daudé @ 2023-02-26 23:23 ` Thomas Bogendoerfer 2023-02-27 1:22 ` Jiaxun Yang 1 sibling, 1 reply; 10+ messages in thread From: Thomas Bogendoerfer @ 2023-02-26 23:23 UTC (permalink / raw) To: Jiaxun Yang; +Cc: linux-mips On Sat, Feb 25, 2023 at 10:10:08PM +0000, Jiaxun Yang wrote: > In c0_compare_int_usable we clear compare interrupt by write value > just read out from counter to compare register. > > However sometimes if those all instructions are graduated together > then it's possible that at the time compare register is written, the > counter haven't progressed, thus the interrupt is triggered again. > > It also applies to QEMU that instructions is execuated significantly > faster then counter. > > Offset the counter value a litlle bit to prevent that happen. > > Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com> > --- > arch/mips/kernel/cevt-r4k.c | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/arch/mips/kernel/cevt-r4k.c b/arch/mips/kernel/cevt-r4k.c > index 32ec67c9ab67..bbc422376e97 100644 > --- a/arch/mips/kernel/cevt-r4k.c > +++ b/arch/mips/kernel/cevt-r4k.c > @@ -200,6 +200,8 @@ int c0_compare_int_usable(void) > */ > if (c0_compare_int_pending()) { > cnt = read_c0_count(); > + // Drawdown a little bit in case counter haven't progressed no C++ comments > + cnt -= COMPARE_INT_SEEN_TICKS; > write_c0_compare(cnt); > back_to_back_c0_hazard(); > while (read_c0_count() < (cnt + COMPARE_INT_SEEN_TICKS)) this doesn't make sense. clearing of the interrupts happes because of this loop. So either COMPARE_INT_SEEN_TICKS is too small or you are hunting a different bug. Thomas. -- Crap can work. Given enough thrust pigs will fly, but it's not necessarily a good idea. [ RFC1925, 2.3 ] ^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: [PATCH 2/2] MIPS: cevt-r4k: Offset counter value for clearing compare interrupt 2023-02-26 23:23 ` Thomas Bogendoerfer @ 2023-02-27 1:22 ` Jiaxun Yang 2023-02-27 9:31 ` Thomas Bogendoerfer 0 siblings, 1 reply; 10+ messages in thread From: Jiaxun Yang @ 2023-02-27 1:22 UTC (permalink / raw) To: Thomas Bogendoerfer; +Cc: linux-mips@vger.kernel.org > 2023年2月26日 23:23，Thomas Bogendoerfer <tsbogend@alpha.franken.de> 写道： > > On Sat, Feb 25, 2023 at 10:10:08PM +0000, Jiaxun Yang wrote: >> In c0_compare_int_usable we clear compare interrupt by write value >> just read out from counter to compare register. >> >> However sometimes if those all instructions are graduated together >> then it's possible that at the time compare register is written, the >> counter haven't progressed, thus the interrupt is triggered again. >> >> It also applies to QEMU that instructions is execuated significantly >> faster then counter. >> >> Offset the counter value a litlle bit to prevent that happen. >> >> Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com> >> --- >> arch/mips/kernel/cevt-r4k.c | 3 +++ >> 1 file changed, 3 insertions(+) >> >> diff --git a/arch/mips/kernel/cevt-r4k.c b/arch/mips/kernel/cevt-r4k.c >> index 32ec67c9ab67..bbc422376e97 100644 >> --- a/arch/mips/kernel/cevt-r4k.c >> +++ b/arch/mips/kernel/cevt-r4k.c >> @@ -200,6 +200,8 @@ int c0_compare_int_usable(void) >> */ >> if (c0_compare_int_pending()) { >> cnt = read_c0_count(); >> + // Drawdown a little bit in case counter haven't progressed > > no C++ comments > >> + cnt -= COMPARE_INT_SEEN_TICKS; >> write_c0_compare(cnt); >> back_to_back_c0_hazard(); >> while (read_c0_count() < (cnt + COMPARE_INT_SEEN_TICKS)) > > this doesn't make sense. clearing of the interrupts happes because of > this loop. So either COMPARE_INT_SEEN_TICKS is too small or you are > hunting a different bug. Clearing interrupt happens in write_c0_compare but the problem is the interrupt does really get cleared because it triggered again straight away. Had confirmed issue on MIPS I6500 uarch simulation trace. Ah I see, it makes more sense if I move minus into write_c0_compare so the loop will still run. Though this loop is not required on any MTI cores, TI is always clear immediately since very early 4Kc. Will send v2 later. Thanks. > > Thomas. > > -- > Crap can work. Given enough thrust pigs will fly, but it's not necessarily a > good idea. [ RFC1925, 2.3 ] ^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: [PATCH 2/2] MIPS: cevt-r4k: Offset counter value for clearing compare interrupt 2023-02-27 1:22 ` Jiaxun Yang @ 2023-02-27 9:31 ` Thomas Bogendoerfer 2023-02-27 11:44 ` Jiaxun Yang 0 siblings, 1 reply; 10+ messages in thread From: Thomas Bogendoerfer @ 2023-02-27 9:31 UTC (permalink / raw) To: Jiaxun Yang; +Cc: linux-mips@vger.kernel.org On Mon, Feb 27, 2023 at 01:22:47AM +0000, Jiaxun Yang wrote: > > > > 2023年2月26日 23:23，Thomas Bogendoerfer <tsbogend@alpha.franken.de> 写道： > > > > On Sat, Feb 25, 2023 at 10:10:08PM +0000, Jiaxun Yang wrote: > >> In c0_compare_int_usable we clear compare interrupt by write value > >> just read out from counter to compare register. > >> > >> However sometimes if those all instructions are graduated together > >> then it's possible that at the time compare register is written, the > >> counter haven't progressed, thus the interrupt is triggered again. > >> > >> It also applies to QEMU that instructions is execuated significantly > >> faster then counter. > >> > >> Offset the counter value a litlle bit to prevent that happen. > >> > >> Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com> > >> --- > >> arch/mips/kernel/cevt-r4k.c | 3 +++ > >> 1 file changed, 3 insertions(+) > >> > >> diff --git a/arch/mips/kernel/cevt-r4k.c b/arch/mips/kernel/cevt-r4k.c > >> index 32ec67c9ab67..bbc422376e97 100644 > >> --- a/arch/mips/kernel/cevt-r4k.c > >> +++ b/arch/mips/kernel/cevt-r4k.c > >> @@ -200,6 +200,8 @@ int c0_compare_int_usable(void) > >> */ > >> if (c0_compare_int_pending()) { > >> cnt = read_c0_count(); > >> + // Drawdown a little bit in case counter haven't progressed > > > > no C++ comments > > > >> + cnt -= COMPARE_INT_SEEN_TICKS; > >> write_c0_compare(cnt); > >> back_to_back_c0_hazard(); > >> while (read_c0_count() < (cnt + COMPARE_INT_SEEN_TICKS)) > > > > this doesn't make sense. clearing of the interrupts happes because of > > this loop. So either COMPARE_INT_SEEN_TICKS is too small or you are > > hunting a different bug. > > Clearing interrupt happens in write_c0_compare but the problem is the interrupt > does really get cleared because it triggered again straight away. the function you are patching is a test function whether counter/compare is usable, so it's not in the normal timer handling. See a problem there would more to broken hardware than to a bug in that function. > Had confirmed issue on MIPS I6500 uarch simulation trace. so not seen on real hardware ? Thomas. -- Crap can work. Given enough thrust pigs will fly, but it's not necessarily a good idea. [ RFC1925, 2.3 ] ^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: [PATCH 2/2] MIPS: cevt-r4k: Offset counter value for clearing compare interrupt 2023-02-27 9:31 ` Thomas Bogendoerfer @ 2023-02-27 11:44 ` Jiaxun Yang 2023-02-27 16:48 ` Thomas Bogendoerfer 0 siblings, 1 reply; 10+ messages in thread From: Jiaxun Yang @ 2023-02-27 11:44 UTC (permalink / raw) To: Thomas Bogendoerfer; +Cc: linux-mips@vger.kernel.org > 2023年2月27日 09:31，Thomas Bogendoerfer <tsbogend@alpha.franken.de> 写道： > > On Mon, Feb 27, 2023 at 01:22:47AM +0000, Jiaxun Yang wrote: >> >> >> [...] >> Clearing interrupt happens in write_c0_compare but the problem is the interrupt >> does really get cleared because it triggered again straight away. > > the function you are patching is a test function whether counter/compare > is usable, so it's not in the normal timer handling. See a problem there > would more to broken hardware than to a bug in that function. The problem is sometimes on a multi-core system this function managed to run on one of the core but then fail on another. In our real use of compare interrupt, we are clearing interrupt by writing back old value of compare, which is guarenteed to be smaller then current count because there are so many instructions being executed from exception vector to here. Also as we have no way to disable compare interrupt, if this function determined compare to be broken then it won’t register event_handler function, which will cause panic for null pointer when unexpected-ish interrupt comes. > >> Had confirmed issue on MIPS I6500 uarch simulation trace. > > so not seen on real hardware ? It was discovered on FPGA and then I tried to debug with simulation to reveal what’s going on hardware side. Thanks - Jiaxun > > Thomas. > > -- > Crap can work. Given enough thrust pigs will fly, but it's not necessarily a > good idea. [ RFC1925, 2.3 ] ^ permalink raw reply [flat|nested] 10+ messages in thread
* Re: [PATCH 2/2] MIPS: cevt-r4k: Offset counter value for clearing compare interrupt 2023-02-27 11:44 ` Jiaxun Yang @ 2023-02-27 16:48 ` Thomas Bogendoerfer 0 siblings, 0 replies; 10+ messages in thread From: Thomas Bogendoerfer @ 2023-02-27 16:48 UTC (permalink / raw) To: Jiaxun Yang; +Cc: linux-mips@vger.kernel.org On Mon, Feb 27, 2023 at 11:44:48AM +0000, Jiaxun Yang wrote: > > so not seen on real hardware ? > > It was discovered on FPGA and then I tried to debug with simulation to > reveal what’s going on hardware side. ic, I think the best fix would be to do a write_c0_compare(cnt - 1). Subtracting COMPARE_INT_SEEN_TICKS feels like overkill to me. Thomas. -- Crap can work. Given enough thrust pigs will fly, but it's not necessarily a good idea. [ RFC1925, 2.3 ] ^ permalink raw reply [flat|nested] 10+ messages in thread
end of thread, other threads:[~2023-02-27 16:49 UTC | newest] Thread overview: 10+ messages (download: mbox.gz follow: Atom feed -- links below jump to the message on this page -- 2023-02-25 22:10 [PATCH 0/2] MIPS Booting fixes Jiaxun Yang 2023-02-25 22:10 ` [PATCH 1/2] MIPS: smp-cps: Don't rely on CP0_CMGCRBASE Jiaxun Yang 2023-02-27 11:59 ` Thomas Bogendoerfer 2023-02-25 22:10 ` [PATCH 2/2] MIPS: cevt-r4k: Offset counter value for clearing compare interrupt Jiaxun Yang 2023-02-26 22:04 ` Philippe Mathieu-Daudé 2023-02-26 23:23 ` Thomas Bogendoerfer 2023-02-27 1:22 ` Jiaxun Yang 2023-02-27 9:31 ` Thomas Bogendoerfer 2023-02-27 11:44 ` Jiaxun Yang 2023-02-27 16:48 ` Thomas Bogendoerfer
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for NNTP newsgroup(s).