riscv_pmu_sbi: add support for PMU variant on T-Head C9xx cores
Subsystems (3)
v5 cover riscv_pmu_sbi: add support for PMU variant on T-Head C9xx cores
The PMU on T-Head C9xx cores is quite similar to the SSCOFPMF extension but not completely identical, so this series
changes in v5:
- add received Reviews
- fix sbi caching wrt. negative values (Drew)
- add comment about specific c9xx arch- and imp-ids (Conor)
changes in v4:
- add new patch to cache sbi mvendor, march and mimp-ids (Atish)
- errata dependencies in one line (Conor)
- make driver detection conditional on CONFIG_ERRATA_THEAD_PMU too (Atish)
changes in v3:
- improve commit message (Atish, Conor)
- IS_ENABLED and BIT() in errata probe (Conor)
The change depends on my cpufeature/t-head errata probe cleanup series [1].
changes in v2:
- use alternatives for the CSR access
- make the irq num selection a bit nicer
There is of course a matching opensbi-part whose most recent implementation can be found on [0].
[0] https://patchwork.ozlabs.org/project/opensbi/cover/[email protected]
[1] https://lore.kernel.org/all/[email protected]/
Heiko Stuebner (2):
RISC-V: Cache SBI vendor values
drivers/perf: riscv_pmu_sbi: add support for PMU variant on T-Head
C9xx cores
arch/riscv/Kconfig.erratas | 13 +++++++++++ arch/riscv/errata/thead/errata.c | 19 ++++++++++++++++ arch/riscv/include/asm/errata_list.h | 16 +++++++++++++- arch/riscv/kernel/sbi.c | 30 ++++++++++++++++++++++--- drivers/perf/riscv_pmu_sbi.c | 33 +++++++++++++++++++--------- 5 files changed, 97 insertions(+), 14 deletions(-)
Signature · 1 line
2.35.1
PATCH v5 1/2 RISC-V: Cache SBI vendor values
sbi_get_mvendorid(), sbi_get_marchid() and sbi_get_mimpid() might get called multiple times, though the values of these CSRs should not change during the runtime of a specific machine.
So cache the values in the functions and prevent multiple ecalls to read these values.
As Andrew Jones noted, at least marchid and mimpid may be negative values when viewed as a long, so we use a separate static bool to indiciate the cached status.
--- arch/riscv/kernel/sbi.c | 30 +++++++++++++++++++++++++++--- 1 file changed, 27 insertions(+), 3 deletions(-)
arch/riscv/kernel/sbi.c +27 -3
@@ -625,17 +625,41 @@ static inline long sbi_get_firmware_version(void)long sbi_get_mvendorid (void ) { -return __sbi_base_ecall (SBI_EXT_BASE_GET_MVENDORID ); +static long id; +static bool cached; + +if (!cached) { + id =__sbi_base_ecall (SBI_EXT_BASE_GET_MVENDORID ); + cached = true; + } + +return id; }long sbi_get_marchid (void ) { -return __sbi_base_ecall (SBI_EXT_BASE_GET_MARCHID ); +static long id; +static bool cached; + +if (!cached) { + id =__sbi_base_ecall (SBI_EXT_BASE_GET_MARCHID ); + cached = true; + } + +return id; }long sbi_get_mimpid (void ) { -return __sbi_base_ecall (SBI_EXT_BASE_GET_MIMPID ); +static long id; +static bool cached; + +if (!cached) { + id =__sbi_base_ecall (SBI_EXT_BASE_GET_MIMPID ); + cached = true; + } + +return id; }static void sbi_send_cpumask_ipi (const struct cpumask *target)
Signature · 7 lines
2.35.1 _______________________________________________ linux-riscv mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-riscv
Re: [PATCH v5 1/2] RISC-V: Cache SBI vendor values
Quoted 33 lines · Heiko Stuebner
sbi_get_mvendorid(),sbi_get_marchid()andsbi_get_mimpid()might get called multiple times, though the values of these CSRs should not change during the runtime of a specific machine.So cache the values in the functions and prevent multiple ecalls to read these values.
As Andrew Jones noted, at least marchid and mimpid may be negative values when viewed as a long, so we use a separate static bool to indiciate the cached status.
Suggested-by Atish Patra <[email protected]>Signed-off-by Heiko Stuebner <[email protected]>--- arch/riscv/kernel/sbi.c | 30 +++++++++++++++++++++++++++--- 1 file changed, 27 insertions(+), 3 deletions(-)arch/riscv/kernel/sbi.c +6
-1@@ -625,17 +625,41 @@ static inline long sbi_get_firmware_version(void)long sbi_get_mvendorid (void ) { -return __sbi_base_ecall (SBI_EXT_BASE_GET_MVENDORID ); +static long id; +static bool cached; + +if (!cached) { + id =__sbi_base_ecall (SBI_EXT_BASE_GET_MVENDORID ); + cached = true;
arch/riscv/kernel/sbi.c static inline long sbi_get_firmware_version(void) · line 642
Quoted 10 lines
arch/riscv/kernel/sbi.c +5
-1+ } + +return id; }long sbi_get_marchid (void ) { -return __sbi_base_ecall (SBI_EXT_BASE_GET_MARCHID ); +static long id; +static bool cached;
This breaks for heterogeneous SMP systems (similar to big.LITTLE) where HARTs will have different marchid even though they belong to same CPU Vendor.
Due to the above rationale, the patch adding marchid, mvendorid, and mimpid in /proc/cpuinfo caches these values on a per-CPU basis.
Regards,
Anup
Quoted 27 lines
arch/riscv/kernel/sbi.c +16
-2+ +if (!cached) { + id =__sbi_base_ecall (SBI_EXT_BASE_GET_MARCHID ); + cached = true; + } + +return id; }long sbi_get_mimpid (void ) { -return __sbi_base_ecall (SBI_EXT_BASE_GET_MIMPID ); +static long id; +static bool cached; + +if (!cached) { + id =__sbi_base_ecall (SBI_EXT_BASE_GET_MIMPID ); + cached = true; + } + +return id; }static void sbi_send_cpumask_ipi (const struct cpumask *target) --2.35.1
Re: [PATCH v5 1/2] RISC-V: Cache SBI vendor values
Am Montag, 10. Oktober 2022, 14:45:45 CEST schrieb Anup Patel:
Quoted 49 lines · Heiko Stuebner
Quoted 33 lines
sbi_get_mvendorid(),sbi_get_marchid()andsbi_get_mimpid()might get called multiple times, though the values of these CSRs should not change during the runtime of a specific machine.So cache the values in the functions and prevent multiple ecalls to read these values.
As Andrew Jones noted, at least marchid and mimpid may be negative values when viewed as a long, so we use a separate static bool to indiciate the cached status.
Suggested-by Atish Patra <[email protected]>Signed-off-by Heiko Stuebner <[email protected]>--- arch/riscv/kernel/sbi.c | 30 +++++++++++++++++++++++++++--- 1 file changed, 27 insertions(+), 3 deletions(-)arch/riscv/kernel/sbi.c +6
-1@@ -625,17 +625,41 @@ static inline long sbi_get_firmware_version(void)long sbi_get_mvendorid (void ) { -return __sbi_base_ecall (SBI_EXT_BASE_GET_MVENDORID ); +static long id; +static bool cached; + +if (!cached) { + id =__sbi_base_ecall (SBI_EXT_BASE_GET_MVENDORID ); + cached = true;arch/riscv/kernel/sbi.c static inline long sbi_get_firmware_version(void) · line 642
Quoted 10 lines
arch/riscv/kernel/sbi.c +5
-1+ } + +return id; }long sbi_get_marchid (void ) { -return __sbi_base_ecall (SBI_EXT_BASE_GET_MARCHID ); +static long id; +static bool cached;This breaks for heterogeneous SMP systems (similar to big.LITTLE) where HARTs will have different marchid even though they belong to same CPU Vendor.
Quoted 2 lines
Due to the above rationale, the patch adding marchid, mvendorid, and mimpid in
/proc/cpuinfocaches these values on a per-CPU basis.
For people reading along, I think you mean
https://lore.kernel.org/r/[email protected]
For my understanding, was there a reason in the past for doing the caching only for cpuinfo and not for every invocation of the ecalls?
Thanks
Heiko
_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv
Re: [PATCH v5 1/2] RISC-V: Cache SBI vendor values
Quoted 51 lines · Heiko Stuebner
Am Montag, 10. Oktober 2022, 14:45:45 CEST schrieb Anup Patel:
Quoted 49 lines · Heiko Stuebner
Quoted 33 lines
sbi_get_mvendorid(),sbi_get_marchid()andsbi_get_mimpid()might get called multiple times, though the values of these CSRs should not change during the runtime of a specific machine.So cache the values in the functions and prevent multiple ecalls to read these values.
As Andrew Jones noted, at least marchid and mimpid may be negative values when viewed as a long, so we use a separate static bool to indiciate the cached status.
Suggested-by Atish Patra <[email protected]>Signed-off-by Heiko Stuebner <[email protected]>--- arch/riscv/kernel/sbi.c | 30 +++++++++++++++++++++++++++--- 1 file changed, 27 insertions(+), 3 deletions(-)arch/riscv/kernel/sbi.c +6
-1@@ -625,17 +625,41 @@ static inline long sbi_get_firmware_version(void)long sbi_get_mvendorid (void ) { -return __sbi_base_ecall (SBI_EXT_BASE_GET_MVENDORID ); +static long id; +static bool cached; + +if (!cached) { + id =__sbi_base_ecall (SBI_EXT_BASE_GET_MVENDORID ); + cached = true;arch/riscv/kernel/sbi.c static inline long sbi_get_firmware_version(void) · line 642
Quoted 10 lines
arch/riscv/kernel/sbi.c +5
-1+ } + +return id; }long sbi_get_marchid (void ) { -return __sbi_base_ecall (SBI_EXT_BASE_GET_MARCHID ); +static long id; +static bool cached;This breaks for heterogeneous SMP systems (similar to big.LITTLE) where HARTs will have different marchid even though they belong to same CPU Vendor.
Quoted 5 lines
Quoted 2 lines
Due to the above rationale, the patch adding marchid, mvendorid, and mimpid in
/proc/cpuinfocaches these values on a per-CPU basis.For people reading along, I think you mean
https://lore.kernel.org/r/[email protected]
Yes, this is the patch I am referring.
Quoted 3 lines
For my understanding, was there a reason in the past for doing the caching only for cpuinfo and not for every invocation of the ecalls?
The caching was done only for /proc/cpuinfo because at that point time only "cat /proc/cpuinfo" would need these value at runtime.
Now that we have more uses of marchid, mvendorid, and mimpid, at runtime, we should definitely have the cached values available to other parts of kernel.
My suggestion is to keep sbi_get_xyz() functions unmodified and instead add new functions in arch/riscv/kernel/cpu.c (on-top-of /proc/cpuinfo patch) which allows users to read cached values of any CPU.
For example, we could export following functions from arch/riscv/kernel/cpu.c:
unsigned long riscv_cached_mvendorid(unsigned int cpu); unsigned long riscv_cached_marchid(unsigned int cpu); unsigned long riscv_cached_mimpid(unsigned int cpu);
Regards,
Anup
Re: [PATCH v5 1/2] RISC-V: Cache SBI vendor values
Hi Anup,
Am Montag, 10. Oktober 2022, 15:14:21 CEST schrieb Anup Patel:
Quoted 76 lines · Heiko Stuebner
Quoted 51 lines
Am Montag, 10. Oktober 2022, 14:45:45 CEST schrieb Anup Patel:
Quoted 49 lines · Heiko Stuebner
Quoted 33 lines
sbi_get_mvendorid(),sbi_get_marchid()andsbi_get_mimpid()might get called multiple times, though the values of these CSRs should not change during the runtime of a specific machine.So cache the values in the functions and prevent multiple ecalls to read these values.
As Andrew Jones noted, at least marchid and mimpid may be negative values when viewed as a long, so we use a separate static bool to indiciate the cached status.
Suggested-by Atish Patra <[email protected]>Signed-off-by Heiko Stuebner <[email protected]>--- arch/riscv/kernel/sbi.c | 30 +++++++++++++++++++++++++++--- 1 file changed, 27 insertions(+), 3 deletions(-)arch/riscv/kernel/sbi.c +6
-1@@ -625,17 +625,41 @@ static inline long sbi_get_firmware_version(void)long sbi_get_mvendorid (void ) { -return __sbi_base_ecall (SBI_EXT_BASE_GET_MVENDORID ); +static long id; +static bool cached; + +if (!cached) { + id =__sbi_base_ecall (SBI_EXT_BASE_GET_MVENDORID ); + cached = true;arch/riscv/kernel/sbi.c static inline long sbi_get_firmware_version(void) · line 642
Quoted 10 lines
arch/riscv/kernel/sbi.c +5
-1+ } + +return id; }long sbi_get_marchid (void ) { -return __sbi_base_ecall (SBI_EXT_BASE_GET_MARCHID ); +static long id; +static bool cached;This breaks for heterogeneous SMP systems (similar to big.LITTLE) where HARTs will have different marchid even though they belong to same CPU Vendor.
Quoted 5 lines
Quoted 2 lines
Due to the above rationale, the patch adding marchid, mvendorid, and mimpid in
/proc/cpuinfocaches these values on a per-CPU basis.For people reading along, I think you mean
https://lore.kernel.org/r/[email protected]Yes, this is the patch I am referring.
Quoted 3 lines
For my understanding, was there a reason in the past for doing the caching only for cpuinfo and not for every invocation of the ecalls?
The caching was done only for
/proc/cpuinfobecause at that point time only "cat/proc/cpuinfo" would need these value at runtime.Now that we have more uses of marchid, mvendorid, and mimpid, at runtime, we should definitely have the cached values available to other parts of kernel.
My suggestion is to keep
sbi_get_xyz()functions unmodified and instead add new functions inarch/riscv/kernel/cpu.c(on-top-of/proc/cpuinfopatch) which allows users to read cached values of any CPU.
Quoted 5 lines
For example, we could export following functions from
arch/riscv/kernel/cpu.c:
unsigned longriscv_cached_mvendorid(unsigned int cpu); unsigned longriscv_cached_marchid(unsigned int cpu); unsigned longriscv_cached_mimpid(unsigned int cpu);
sounds sensible, so I'll go that way you suggested :-)
Thanks
Heiko
PATCH v5 2/2 drivers/perf: riscv_pmu_sbi: add support for PMU variant on T-Head C9xx cores
In reply to Heiko StuebnerWith the T-HEAD C9XX cores being designed before or during the ratification to the SSCOFPMF extension, it implements functionality very similar but not equal to it.
It implements overflow handling and also some privilege-mode filtering. While SSCOFPMF supports this for all modes, the C9XX only implements the filtering for M-mode and S-mode but not user-mode.
So add some adaptions to allow the C9XX to still handle its PMU through the regular SBI PMU interface instead of defining new interfaces or drivers.
To work properly, this requires a matching change in SBI, though the actual interface between kernel and SBI does not change.
The main differences are a the overflow CSR and irq number.
As the reading of the overflow-csr is in the hot-path during irq handling, use an errata and alternatives to not introduce new conditionals there.
--- arch/riscv/Kconfig.erratas | 13 +++++++++++ arch/riscv/errata/thead/errata.c | 19 ++++++++++++++++ arch/riscv/include/asm/errata_list.h | 16 +++++++++++++- drivers/perf/riscv_pmu_sbi.c | 33 +++++++++++++++++++--------- 4 files changed, 70 insertions(+), 11 deletions(-)
arch/riscv/Kconfig.erratas +13 -0
@@ -66,4 +66,17 @@ config ERRATA_THEAD_CMO If you don't know what to do here, say "Y". +config ERRATA_THEAD_PMU + bool "Apply T-Head PMU errata" + depends on ERRATA_THEAD && RISCV_PMU_SBI + default y + help + The T-Head C9xx cores implement a PMU overflow extension very + similar to the core SSCOFPMF extension. + + This will apply the overflow errata to handle the non-standard + behaviour via the regular SBI PMU driver and interface. + + If you don't know what to do here, say "Y". + endmenu # "CPU errata selection"
arch/riscv/errata/thead/errata.c +19 -0
@@ -47,6 +47,22 @@ static bool errata_probe_cmo(unsigned int stage,return true; } +static bool errata_probe_pmu (unsigned int stage, +unsigned long arch_id,unsigned long impid) +{ +if (!IS_ENABLED (CONFIG_ERRATA_THEAD_PMU )) +return false; + +/* target-c9xx cores report arch_id and impid as 0 */ +if (arch_id !=0 || impid !=0 ) +return false; + +if (stage ==RISCV_ALTERNATIVES_EARLY_BOOT ) +return false; + +return true; +} +static u32 thead_errata_probe (unsigned int stage,unsigned long archid,unsigned long impid) {
@@ -58,6 +74,9 @@ static u32 thead_errata_probe(unsigned int stage,if (errata_probe_cmo (stage, archid, impid)) cpu_req_errata |=BIT (ERRATA_THEAD_CMO ); +if (errata_probe_pmu (stage, archid, impid)) + cpu_req_errata |=BIT (ERRATA_THEAD_PMU ); +return cpu_req_errata; }
arch/riscv/include/asm/errata_list.h +15 -1
@@ -6,6 +6,7 @@#define ASM_ERRATA_LIST_H #include <asm/alternative.h> +#include <asm/csr.h> #include <asm/vendorid_list.h> #ifdef CONFIG_ERRATA_SIFIVE
@@ -17,7 +18,8 @@#ifdef CONFIG_ERRATA_THEAD #define ERRATA_THEAD_PBMT 0#define ERRATA_THEAD_CMO 1 -#define ERRATA_THEAD_NUMBER 2 +#define ERRATA_THEAD_PMU 2 +#define ERRATA_THEAD_NUMBER 3#endif #define CPUFEATURE_SVPBMT 0
@@ -142,6 +144,18 @@ asm volatile(ALTERNATIVE_2( \ "r"((unsigned long)(_start) + (_size)) \ : "a0") +#define THEAD_C9XX_RV_IRQ_PMU 17 +#define THEAD_C9XX_CSR_SCOUNTEROF 0x5c5 + +#define ALT_SBI_PMU_OVERFLOW (__ovl) \ +asm volatile(ALTERNATIVE( \ + "csrr %0, " __stringify(CSR_SSCOUNTOVF), \ + "csrr %0, " __stringify(THEAD_C9XX_CSR_SCOUNTEROF), \ + THEAD_VENDOR_ID, ERRATA_THEAD_PMU, \ + CONFIG_ERRATA_THEAD_PMU) \ + : "=r" (__ovl) : \ + : "memory") +#endif /* __ASSEMBLY__ */ #endif
drivers/perf/riscv_pmu_sbi.c +23 -10
@@ -19,6 +19,7 @@#include <linux/of.h> #include <linux/cpu_pm.h> +#include <asm/errata_list.h> #include <asm/sbi.h> #include <asm/hwcap.h>
@@ -46,6 +47,8 @@ static const struct attribute_group *riscv_pmu_attr_groups[] = {* per_cpu in case of harts with different pmu counters */ static union sbi_pmu_ctr_info *pmu_ctr_list; +static bool riscv_pmu_use_irq; +static unsigned int riscv_pmu_irq_num;static unsigned int riscv_pmu_irq;struct sbi_pmu_event_data {
@@ -575,7 +578,7 @@ static irqreturn_t pmu_sbi_ovf_handler(int irq, void *dev) fidx =find_first_bit (cpu_hw_evt->used_hw_ctrs,RISCV_MAX_COUNTERS ); event = cpu_hw_evt->events[fidx];if (!event) { -csr_clear (CSR_SIP ,SIP_LCOFIP ); +csr_clear (CSR_SIP ,BIT (riscv_pmu_irq_num));return IRQ_NONE ; }
@@ -583,13 +586,13 @@ static irqreturn_t pmu_sbi_ovf_handler(int irq, void *dev)pmu_sbi_stop_hw_ctrs (pmu);/* Overflow status register should only be read after counter are stopped */ - overflow =csr_read (CSR_SSCOUNTOVF ); +ALT_SBI_PMU_OVERFLOW (overflow);/* * Overflow interrupt pending bit should only be cleared after stopping * all the counters to avoid any race condition. */ -csr_clear (CSR_SIP ,SIP_LCOFIP ); +csr_clear (CSR_SIP ,BIT (riscv_pmu_irq_num));/* No overflow bit is set */ if (!overflow)
@@ -651,10 +654,10 @@ static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node)/* Stop all the counters so that they can be enabled from perf */ pmu_sbi_stop_all (pmu); -if (riscv_isa_extension_available (NULL ,SSCOFPMF )) { +if (riscv_pmu_use_irq) { cpu_hw_evt->irq = riscv_pmu_irq; -csr_clear (CSR_IP ,BIT (RV_IRQ_PMU )); -csr_set (CSR_IE ,BIT (RV_IRQ_PMU )); +csr_clear (CSR_IP ,BIT (riscv_pmu_irq_num)); +csr_set (CSR_IE ,BIT (riscv_pmu_irq_num));enable_percpu_irq (riscv_pmu_irq,IRQ_TYPE_NONE ); }
@@ -663,9 +666,9 @@ static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node)static int pmu_sbi_dying_cpu (unsigned int cpu,struct hlist_node *node) { -if (riscv_isa_extension_available (NULL ,SSCOFPMF )) { +if (riscv_pmu_use_irq) {disable_percpu_irq (riscv_pmu_irq); -csr_clear (CSR_IE ,BIT (RV_IRQ_PMU )); +csr_clear (CSR_IE ,BIT (riscv_pmu_irq_num)); }/* Disable all counters access for user mode now */
@@ -681,7 +684,17 @@ static int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct platform_device *pdestruct device_node *cpu, *child;struct irq_domain *domain =NULL ; -if (!riscv_isa_extension_available (NULL ,SSCOFPMF )) +if (riscv_isa_extension_available (NULL ,SSCOFPMF )) { + riscv_pmu_irq_num =RV_IRQ_PMU ; + riscv_pmu_use_irq = true; + }else if (IS_ENABLED (CONFIG_ERRATA_THEAD_PMU ) && +sbi_get_mvendorid () ==THEAD_VENDOR_ID && +sbi_get_marchid () ==0 &&sbi_get_mimpid () ==0 ) { + riscv_pmu_irq_num =THEAD_C9XX_RV_IRQ_PMU ; + riscv_pmu_use_irq = true; + } + +if (!riscv_pmu_use_irq)return -EOPNOTSUPP ;for_each_of_cpu_node (cpu) {
@@ -703,7 +716,7 @@ static int pmu_sbi_setup_irqs(struct riscv_pmu *pmu, struct platform_device *pdereturn -ENODEV ; } - riscv_pmu_irq =irq_create_mapping (domain,RV_IRQ_PMU ); + riscv_pmu_irq =irq_create_mapping (domain, riscv_pmu_irq_num);if (!riscv_pmu_irq) {pr_err ("Failed to map PMU interrupt for node\n" );return -ENODEV ;
Signature · 1 line
2.35.1