Re: [PATCH 2/3] RISC-V: Detect unaligned vector accesses supported.

[Charlie Jenkins <charlie@rivosinc.com>]

On Thu, Jun 06, 2024 at 02:29:23PM -0700, Charlie Jenkins wrote:
> On Thu, Jun 06, 2024 at 02:32:14PM -0400, Jesse Taube wrote:
> > Run a unaligned vector access to test if the system supports
> > vector unaligned access. Add the result to a new key in hwprobe.
> > This is useful for usermode to know if vector misaligned accesses are
> > supported and if they are faster or slower than equivalent byte accesses.
> > 
> > Signed-off-by: Jesse Taube <jesse@rivosinc.com>
> > ---
> >  arch/riscv/include/asm/cpufeature.h        |  2 +
> >  arch/riscv/include/asm/hwprobe.h           |  2 +-
> >  arch/riscv/include/asm/vector.h            |  1 +
> >  arch/riscv/include/uapi/asm/hwprobe.h      |  6 ++
> >  arch/riscv/kernel/sys_hwprobe.c            | 34 +++++++++
> >  arch/riscv/kernel/traps_misaligned.c       | 84 ++++++++++++++++++++--
> >  arch/riscv/kernel/unaligned_access_speed.c |  4 ++
> >  arch/riscv/kernel/vector.c                 |  2 +-
> >  8 files changed, 129 insertions(+), 6 deletions(-)
> > 
> > diff --git a/arch/riscv/include/asm/cpufeature.h b/arch/riscv/include/asm/cpufeature.h
> > index 347805446151..a012c8490a27 100644
> > --- a/arch/riscv/include/asm/cpufeature.h
> > +++ b/arch/riscv/include/asm/cpufeature.h
> > @@ -35,9 +35,11 @@ void riscv_user_isa_enable(void);
> >  
> >  #if defined(CONFIG_RISCV_MISALIGNED)
> >  bool check_unaligned_access_emulated_all_cpus(void);
> > +bool check_vector_unaligned_access_all_cpus(void);
> >  void unaligned_emulation_finish(void);
> >  bool unaligned_ctl_available(void);
> >  DECLARE_PER_CPU(long, misaligned_access_speed);
> > +DECLARE_PER_CPU(long, vector_misaligned_access);
> >  #else
> >  static inline bool unaligned_ctl_available(void)
> >  {
> > diff --git a/arch/riscv/include/asm/hwprobe.h b/arch/riscv/include/asm/hwprobe.h
> > index 630507dff5ea..150a9877b0af 100644
> > --- a/arch/riscv/include/asm/hwprobe.h
> > +++ b/arch/riscv/include/asm/hwprobe.h
> > @@ -8,7 +8,7 @@
> >  
> >  #include <uapi/asm/hwprobe.h>
> >  
> > -#define RISCV_HWPROBE_MAX_KEY 6
> > +#define RISCV_HWPROBE_MAX_KEY 7
> >  
> >  static inline bool riscv_hwprobe_key_is_valid(__s64 key)
> >  {
> > diff --git a/arch/riscv/include/asm/vector.h b/arch/riscv/include/asm/vector.h
> > index 731dcd0ed4de..776af9b37e23 100644
> > --- a/arch/riscv/include/asm/vector.h
> > +++ b/arch/riscv/include/asm/vector.h
> > @@ -21,6 +21,7 @@
> >  
> >  extern unsigned long riscv_v_vsize;
> >  int riscv_v_setup_vsize(void);
> > +bool insn_is_vector(u32 insn_buf);
> >  bool riscv_v_first_use_handler(struct pt_regs *regs);
> >  void kernel_vector_begin(void);
> >  void kernel_vector_end(void);
> > diff --git a/arch/riscv/include/uapi/asm/hwprobe.h b/arch/riscv/include/uapi/asm/hwprobe.h
> > index 060212331a03..ebacff86f134 100644
> > --- a/arch/riscv/include/uapi/asm/hwprobe.h
> > +++ b/arch/riscv/include/uapi/asm/hwprobe.h
> > @@ -68,6 +68,12 @@ struct riscv_hwprobe {
> >  #define		RISCV_HWPROBE_MISALIGNED_UNSUPPORTED	(4 << 0)
> >  #define		RISCV_HWPROBE_MISALIGNED_MASK		(7 << 0)
> >  #define RISCV_HWPROBE_KEY_ZICBOZ_BLOCK_SIZE	6
> > +#define RISCV_HWPROBE_VEC_KEY_MISALIGNED_PERF	7
> > +#define		RISCV_HWPROBE_VEC_MISALIGNED_UNKNOWN		0
> > +#define		RISCV_HWPROBE_VEC_MISALIGNED_SUPPORTED		1
> > +#define		RISCV_HWPROBE_VEC_MISALIGNED_SLOW		2
> > +#define		RISCV_HWPROBE_VEC_MISALIGNED_FAST		3
> > +#define		RISCV_HWPROBE_VEC_MISALIGNED_UNSUPPORTED	4
> >  /* Increase RISCV_HWPROBE_MAX_KEY when adding items. */
> >  
> >  /* Flags */
> > diff --git a/arch/riscv/kernel/sys_hwprobe.c b/arch/riscv/kernel/sys_hwprobe.c
> > index b286b73e763e..ce641cc6e47a 100644
> > --- a/arch/riscv/kernel/sys_hwprobe.c
> > +++ b/arch/riscv/kernel/sys_hwprobe.c
> > @@ -184,6 +184,36 @@ static u64 hwprobe_misaligned(const struct cpumask *cpus)
> >  }
> >  #endif
> >  
> > +#if defined(CONFIG_RISCV_PROBE_UNALIGNED_ACCESS)
> > +static u64 hwprobe_vec_misaligned(const struct cpumask *cpus)
> > +{
> > +	int cpu;
> > +	u64 perf = -1ULL;
> > +
> > +	for_each_cpu(cpu, cpus) {
> > +		int this_perf = per_cpu(vector_misaligned_access, cpu);
> > +
> > +		if (perf == -1ULL)
> > +			perf = this_perf;
> > +
> > +		if (perf != this_perf) {
> > +			perf = RISCV_HWPROBE_VEC_MISALIGNED_UNKNOWN;
> > +			break;
> > +		}
> > +	}
> > +
> > +	if (perf == -1ULL)
> > +		return RISCV_HWPROBE_VEC_MISALIGNED_UNKNOWN;
> > +
> > +	return perf;
> > +}
> > +#else
> > +static u64 hwprobe_vec_misaligned(const struct cpumask *cpus)
> > +{

I meant to mention this in my last message!

The scalar version has cutouts for configs here like:

	if (IS_ENABLED(CONFIG_RISCV_EFFICIENT_UNALIGNED_ACCESS))
		return RISCV_HWPROBE_MISALIGNED_FAST;

Having this functionality on vector as well would be much appreciated.
I don't think it's valid to assume that vector and scalar have the same
speed, so this would require a vector version of the RISCV_MISALIGNED
tree in arch/riscv/Kconfig.

- Charlie

> > +	return RISCV_HWPROBE_VEC_MISALIGNED_UNKNOWN;
> > +}
> > +#endif
> > +
> >  static void hwprobe_one_pair(struct riscv_hwprobe *pair,
> >  			     const struct cpumask *cpus)
> >  {
> > @@ -211,6 +241,10 @@ static void hwprobe_one_pair(struct riscv_hwprobe *pair,
> >  		pair->value = hwprobe_misaligned(cpus);
> >  		break;
> >  
> > +	case RISCV_HWPROBE_VEC_KEY_MISALIGNED_PERF:
> > +		pair->value = hwprobe_vec_misaligned(cpus);
> > +		break;
> > +
> >  	case RISCV_HWPROBE_KEY_ZICBOZ_BLOCK_SIZE:
> >  		pair->value = 0;
> >  		if (hwprobe_ext0_has(cpus, RISCV_HWPROBE_EXT_ZICBOZ))
> > diff --git a/arch/riscv/kernel/traps_misaligned.c b/arch/riscv/kernel/traps_misaligned.c
> > index 2adb7c3e4dd5..8f26c3d92230 100644
> > --- a/arch/riscv/kernel/traps_misaligned.c
> > +++ b/arch/riscv/kernel/traps_misaligned.c
> > @@ -16,6 +16,7 @@
> >  #include <asm/entry-common.h>
> >  #include <asm/hwprobe.h>
> >  #include <asm/cpufeature.h>
> > +#include <asm/vector.h>
> >  
> >  #define INSN_MATCH_LB			0x3
> >  #define INSN_MASK_LB			0x707f
> > @@ -413,10 +414,6 @@ int handle_misaligned_load(struct pt_regs *regs)
> >  
> >  	perf_sw_event(PERF_COUNT_SW_ALIGNMENT_FAULTS, 1, regs, addr);
> >  
> > -#ifdef CONFIG_RISCV_PROBE_UNALIGNED_ACCESS
> > -	*this_cpu_ptr(&misaligned_access_speed) = RISCV_HWPROBE_MISALIGNED_EMULATED;
> > -#endif
> > -
> >  	if (!unaligned_enabled)
> >  		return -1;
> >  
> > @@ -426,6 +423,17 @@ int handle_misaligned_load(struct pt_regs *regs)
> >  	if (get_insn(regs, epc, &insn))
> >  		return -1;
> >  
> > +#ifdef CONFIG_RISCV_PROBE_UNALIGNED_ACCESS
> > +	if (insn_is_vector(insn) &&
> > +	    *this_cpu_ptr(&vector_misaligned_access) == RISCV_HWPROBE_VEC_MISALIGNED_SUPPORTED) {
> > +		*this_cpu_ptr(&vector_misaligned_access) = RISCV_HWPROBE_VEC_MISALIGNED_UNSUPPORTED;
> > +		regs->epc = epc + INSN_LEN(insn);
> > +		return 0;
> > +	}
> > +
> > +	*this_cpu_ptr(&misaligned_access_speed) = RISCV_HWPROBE_MISALIGNED_EMULATED;
> 
> This unconditionally sets scalar unaligned accesses even if the
> unaligned access is caused by vector. Scalar unaligned accesses should
> only be set to emulated if this function is entered from a scalar
> unaligned load.
> 
> The rest of this function handles how scalar unaligned accesses are
> emulated, and the equivalent needs to happen for vector. You need to add
> routines that manually load the data from the memory address into the
> vector register. When Clément did this for scalar, he provided a test
> case to help reviewers [1]. Please add onto these test cases or make
> your own for vector.
> 
> Link: https://github.com/clementleger/unaligned_test [1]
> 
> > +#endif
> > +
> >  	regs->epc = 0;
> >  
> >  	if ((insn & INSN_MASK_LW) == INSN_MATCH_LW) {
> > @@ -625,6 +633,74 @@ static bool check_unaligned_access_emulated(int cpu)
> >  	return misaligned_emu_detected;
> >  }
> >  
> > +#ifdef CONFIG_RISCV_ISA_V
> > +static void check_vector_unaligned_access(struct work_struct *unused)
> 
> Can you standardize this name with the scalar version by writing
> emulated in it?
> 
> "check_vector_unaligned_access_emulated_all_cpus"
> 
> > +{
> > +	int cpu = smp_processor_id();
> > +	long *mas_ptr = this_cpu_ptr(&vector_misaligned_access);
> > +	unsigned long tmp_var;
> > +
> > +	if (!riscv_isa_extension_available(hart_isa[cpu].isa, v))
> > +		return;
> > +
> > +	*mas_ptr = RISCV_HWPROBE_VEC_MISALIGNED_SUPPORTED;
> > +
> > +	local_irq_enable();
> > +	kernel_vector_begin();
> > +	__asm__ __volatile__ (
> > +		".balign 4\n\t"
> > +		".option push\n\t"
> > +		".option arch, +v\n\t"
> > +		"       vsetivli zero, 1, e16, m1, ta, ma\n\t"	// Vectors of 16b
> > +		"	vle16.v v0, (%[ptr])\n\t"		// Load bytes
> > +		".option pop\n\t"
> > +		: : [ptr] "r" ((u8 *)&tmp_var + 1) : "v0", "memory");
> 
> memory is being read from, but not written to, so there is no need to
> have a memory clobber.
> 
> > +	kernel_vector_end();
> > +
> > +	if (*mas_ptr == RISCV_HWPROBE_VEC_MISALIGNED_UNKNOWN)
> > +		*mas_ptr = RISCV_HWPROBE_VEC_MISALIGNED_SUPPORTED;
> > +}
> > +
> > +bool check_vector_unaligned_access_all_cpus(void)
> > +{
> > +	int cpu;
> > +	bool ret = true;
> > +
> > +	for_each_online_cpu(cpu)
> > +		if (riscv_isa_extension_available(hart_isa[cpu].isa, ZICCLSM))
> 
> zicclsm is not specific to vector so it can be extracted out of this
> vector specific function. Assuming that hardware properly reports the
> extension, if zicclsm is present then it is known that both vector and
> scalar unaligned accesses are supported.
> 
> > +			per_cpu(vector_misaligned_access, cpu) = RISCV_HWPROBE_VEC_MISALIGNED_SUPPORTED;
> 
> Please use the exising UNKNOWN terminology instead of renaming to
> SUPPORTED. Any option that is not UNSUPPORTED implies that unaligned
> accesses are supported.
> 
> > +		else
> > +			ret = false;
> > +
> > +
> > +	if (ret)
> > +		return true;
> > +	ret = true;
> > +
> > +	schedule_on_each_cpu(check_vector_unaligned_access);
> > +
> > +	for_each_online_cpu(cpu)
> > +		if (per_cpu(vector_misaligned_access, cpu)
> > +		    != RISCV_HWPROBE_VEC_MISALIGNED_SUPPORTED)
> > +			return false;
> > +
> > +	return ret;
> > +}
> > +#else
> 
> If CONFIG_RISCV_ISA_V is not set, there is no value in checking if
> vector unaligned accesses are supported because userspace will not be
> allowed to use vector instructions anyway.
> 
> - Charlie
> 
> > +bool check_vector_unaligned_access_all_cpus(void)
> > +{
> > +	int cpu;
> > +
> > +	for_each_online_cpu(cpu)
> > +		if (riscv_isa_extension_available(hart_isa[cpu].isa, ZICCLSM))
> > +			per_cpu(vector_misaligned_access, cpu) = RISCV_HWPROBE_VEC_MISALIGNED_SUPPORTED;
> > +		else
> > +			per_cpu(vector_misaligned_access, cpu) = RISCV_HWPROBE_VEC_MISALIGNED_UNSUPPORTED;
> > +
> > +	return false;
> > +}
> > +#endif
> > +
> >  bool check_unaligned_access_emulated_all_cpus(void)
> >  {
> >  	int cpu;
> > diff --git a/arch/riscv/kernel/unaligned_access_speed.c b/arch/riscv/kernel/unaligned_access_speed.c
> > index a9a6bcb02acf..92a84239beaa 100644
> > --- a/arch/riscv/kernel/unaligned_access_speed.c
> > +++ b/arch/riscv/kernel/unaligned_access_speed.c
> > @@ -20,6 +20,7 @@
> >  #define MISALIGNED_COPY_SIZE ((MISALIGNED_BUFFER_SIZE / 2) - 0x80)
> >  
> >  DEFINE_PER_CPU(long, misaligned_access_speed);
> > +DEFINE_PER_CPU(long, vector_misaligned_access) = RISCV_HWPROBE_VEC_MISALIGNED_UNKNOWN;
> >  
> >  #ifdef CONFIG_RISCV_PROBE_UNALIGNED_ACCESS
> >  static cpumask_t fast_misaligned_access;
> > @@ -264,6 +265,8 @@ static int check_unaligned_access_all_cpus(void)
> >  {
> >  	bool all_cpus_emulated = check_unaligned_access_emulated_all_cpus();
> >  
> > +	check_vector_unaligned_access_all_cpus();
> > +
> >  	if (!all_cpus_emulated)
> >  		return check_unaligned_access_speed_all_cpus();
> >  
> > @@ -273,6 +276,7 @@ static int check_unaligned_access_all_cpus(void)
> >  static int check_unaligned_access_all_cpus(void)
> >  {
> >  	check_unaligned_access_emulated_all_cpus();
> > +	check_vector_unaligned_access_all_cpus();
> >  
> >  	return 0;
> >  }
> > diff --git a/arch/riscv/kernel/vector.c b/arch/riscv/kernel/vector.c
> > index 6727d1d3b8f2..2cceab739b2c 100644
> > --- a/arch/riscv/kernel/vector.c
> > +++ b/arch/riscv/kernel/vector.c
> > @@ -66,7 +66,7 @@ void __init riscv_v_setup_ctx_cache(void)
> >  #endif
> >  }
> >  
> > -static bool insn_is_vector(u32 insn_buf)
> > +bool insn_is_vector(u32 insn_buf)
> >  {
> >  	u32 opcode = insn_buf & __INSN_OPCODE_MASK;
> >  	u32 width, csr;
> > -- 
> > 2.43.0
> >