diff options
Diffstat (limited to 'arch/tile/lib')
30 files changed, 1518 insertions, 897 deletions
diff --git a/arch/tile/lib/Makefile b/arch/tile/lib/Makefile index 93122d5b155..c4211cbb202 100644 --- a/arch/tile/lib/Makefile +++ b/arch/tile/lib/Makefile @@ -2,17 +2,17 @@ # Makefile for TILE-specific library files.. # -lib-y = cacheflush.o checksum.o cpumask.o delay.o \ - mb_incoherent.o uaccess.o memmove.o \ - memcpy_$(BITS).o memchr_$(BITS).o memset_$(BITS).o \ - strchr_$(BITS).o strlen_$(BITS).o - -ifeq ($(CONFIG_TILEGX),y) -lib-y += memcpy_user_64.o -else -lib-y += atomic_32.o atomic_asm_32.o memcpy_tile64.o -endif +lib-y = cacheflush.o checksum.o cpumask.o delay.o uaccess.o \ + memmove.o memcpy_$(BITS).o memchr_$(BITS).o memset_$(BITS).o \ + strchr_$(BITS).o strlen_$(BITS).o strnlen_$(BITS).o +lib-$(CONFIG_TILEGX) += memcpy_user_64.o +lib-$(CONFIG_TILEPRO) += atomic_32.o atomic_asm_32.o lib-$(CONFIG_SMP) += spinlock_$(BITS).o usercopy_$(BITS).o obj-$(CONFIG_MODULES) += exports.o + +# The finv_buffer_remote() and copy_{to,from}_user() routines can't +# have -pg added, since they both rely on being leaf functions. +CFLAGS_REMOVE_cacheflush.o = -pg +CFLAGS_REMOVE_memcpy_user_64.o = -pg diff --git a/arch/tile/lib/atomic_32.c b/arch/tile/lib/atomic_32.c index 7a5cc706ab6..c89b211fd9e 100644 --- a/arch/tile/lib/atomic_32.c +++ b/arch/tile/lib/atomic_32.c @@ -17,55 +17,15 @@ #include <linux/uaccess.h> #include <linux/module.h> #include <linux/mm.h> -#include <asm/atomic.h> -#include <asm/futex.h> +#include <linux/atomic.h> #include <arch/chip.h> -/* See <asm/atomic_32.h> */ -#if ATOMIC_LOCKS_FOUND_VIA_TABLE() - -/* - * A block of memory containing locks for atomic ops. Each instance of this - * struct will be homed on a different CPU. - */ -struct atomic_locks_on_cpu { - int lock[ATOMIC_HASH_L2_SIZE]; -} __attribute__((aligned(ATOMIC_HASH_L2_SIZE * 4))); - -static DEFINE_PER_CPU(struct atomic_locks_on_cpu, atomic_lock_pool); - -/* The locks we'll use until __init_atomic_per_cpu is called. */ -static struct atomic_locks_on_cpu __initdata initial_atomic_locks; - -/* Hash into this vector to get a pointer to lock for the given atomic. */ -struct atomic_locks_on_cpu *atomic_lock_ptr[ATOMIC_HASH_L1_SIZE] - __write_once = { - [0 ... ATOMIC_HASH_L1_SIZE-1] (&initial_atomic_locks) -}; - -#else /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */ - /* This page is remapped on startup to be hash-for-home. */ -int atomic_locks[PAGE_SIZE / sizeof(int) /* Only ATOMIC_HASH_SIZE is used */] - __attribute__((aligned(PAGE_SIZE), section(".bss.page_aligned"))); - -#endif /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */ +int atomic_locks[PAGE_SIZE / sizeof(int)] __page_aligned_bss; -static inline int *__atomic_hashed_lock(volatile void *v) +int *__atomic_hashed_lock(volatile void *v) { - /* NOTE: this code must match "sys_cmpxchg" in kernel/intvec.S */ -#if ATOMIC_LOCKS_FOUND_VIA_TABLE() - unsigned long i = - (unsigned long) v & ((PAGE_SIZE-1) & -sizeof(long long)); - unsigned long n = __insn_crc32_32(0, i); - - /* Grab high bits for L1 index. */ - unsigned long l1_index = n >> ((sizeof(n) * 8) - ATOMIC_HASH_L1_SHIFT); - /* Grab low bits for L2 index. */ - unsigned long l2_index = n & (ATOMIC_HASH_L2_SIZE - 1); - - return &atomic_lock_ptr[l1_index]->lock[l2_index]; -#else + /* NOTE: this code must match "sys_cmpxchg" in kernel/intvec_32.S */ /* * Use bits [3, 3 + ATOMIC_HASH_SHIFT) as the lock index. * Using mm works here because atomic_locks is page aligned. @@ -74,26 +34,13 @@ static inline int *__atomic_hashed_lock(volatile void *v) (unsigned long)atomic_locks, 2, (ATOMIC_HASH_SHIFT + 2) - 1); return (int *)ptr; -#endif } #ifdef CONFIG_SMP /* Return whether the passed pointer is a valid atomic lock pointer. */ static int is_atomic_lock(int *p) { -#if ATOMIC_LOCKS_FOUND_VIA_TABLE() - int i; - for (i = 0; i < ATOMIC_HASH_L1_SIZE; ++i) { - - if (p >= &atomic_lock_ptr[i]->lock[0] && - p < &atomic_lock_ptr[i]->lock[ATOMIC_HASH_L2_SIZE]) { - return 1; - } - } - return 0; -#else return p >= &atomic_locks[0] && p < &atomic_locks[ATOMIC_HASH_SIZE]; -#endif } void __atomic_fault_unlock(int *irqlock_word) @@ -112,33 +59,32 @@ static inline int *__atomic_setup(volatile void *v) return __atomic_hashed_lock(v); } -int _atomic_xchg(atomic_t *v, int n) +int _atomic_xchg(int *v, int n) { - return __atomic_xchg(&v->counter, __atomic_setup(v), n).val; + return __atomic_xchg(v, __atomic_setup(v), n).val; } EXPORT_SYMBOL(_atomic_xchg); -int _atomic_xchg_add(atomic_t *v, int i) +int _atomic_xchg_add(int *v, int i) { - return __atomic_xchg_add(&v->counter, __atomic_setup(v), i).val; + return __atomic_xchg_add(v, __atomic_setup(v), i).val; } EXPORT_SYMBOL(_atomic_xchg_add); -int _atomic_xchg_add_unless(atomic_t *v, int a, int u) +int _atomic_xchg_add_unless(int *v, int a, int u) { /* * Note: argument order is switched here since it is easier * to use the first argument consistently as the "old value" * in the assembly, as is done for _atomic_cmpxchg(). */ - return __atomic_xchg_add_unless(&v->counter, __atomic_setup(v), u, a) - .val; + return __atomic_xchg_add_unless(v, __atomic_setup(v), u, a).val; } EXPORT_SYMBOL(_atomic_xchg_add_unless); -int _atomic_cmpxchg(atomic_t *v, int o, int n) +int _atomic_cmpxchg(int *v, int o, int n) { - return __atomic_cmpxchg(&v->counter, __atomic_setup(v), o, n).val; + return __atomic_cmpxchg(v, __atomic_setup(v), o, n).val; } EXPORT_SYMBOL(_atomic_cmpxchg); @@ -161,78 +107,36 @@ unsigned long _atomic_xor(volatile unsigned long *p, unsigned long mask) EXPORT_SYMBOL(_atomic_xor); -u64 _atomic64_xchg(atomic64_t *v, u64 n) +long long _atomic64_xchg(long long *v, long long n) { - return __atomic64_xchg(&v->counter, __atomic_setup(v), n); + return __atomic64_xchg(v, __atomic_setup(v), n); } EXPORT_SYMBOL(_atomic64_xchg); -u64 _atomic64_xchg_add(atomic64_t *v, u64 i) +long long _atomic64_xchg_add(long long *v, long long i) { - return __atomic64_xchg_add(&v->counter, __atomic_setup(v), i); + return __atomic64_xchg_add(v, __atomic_setup(v), i); } EXPORT_SYMBOL(_atomic64_xchg_add); -u64 _atomic64_xchg_add_unless(atomic64_t *v, u64 a, u64 u) +long long _atomic64_xchg_add_unless(long long *v, long long a, long long u) { /* * Note: argument order is switched here since it is easier * to use the first argument consistently as the "old value" * in the assembly, as is done for _atomic_cmpxchg(). */ - return __atomic64_xchg_add_unless(&v->counter, __atomic_setup(v), - u, a); + return __atomic64_xchg_add_unless(v, __atomic_setup(v), u, a); } EXPORT_SYMBOL(_atomic64_xchg_add_unless); -u64 _atomic64_cmpxchg(atomic64_t *v, u64 o, u64 n) +long long _atomic64_cmpxchg(long long *v, long long o, long long n) { - return __atomic64_cmpxchg(&v->counter, __atomic_setup(v), o, n); + return __atomic64_cmpxchg(v, __atomic_setup(v), o, n); } EXPORT_SYMBOL(_atomic64_cmpxchg); -static inline int *__futex_setup(int __user *v) -{ - /* - * Issue a prefetch to the counter to bring it into cache. - * As for __atomic_setup, but we can't do a read into the L1 - * since it might fault; instead we do a prefetch into the L2. - */ - __insn_prefetch(v); - return __atomic_hashed_lock((int __force *)v); -} - -struct __get_user futex_set(int __user *v, int i) -{ - return __atomic_xchg((int __force *)v, __futex_setup(v), i); -} - -struct __get_user futex_add(int __user *v, int n) -{ - return __atomic_xchg_add((int __force *)v, __futex_setup(v), n); -} - -struct __get_user futex_or(int __user *v, int n) -{ - return __atomic_or((int __force *)v, __futex_setup(v), n); -} - -struct __get_user futex_andn(int __user *v, int n) -{ - return __atomic_andn((int __force *)v, __futex_setup(v), n); -} - -struct __get_user futex_xor(int __user *v, int n) -{ - return __atomic_xor((int __force *)v, __futex_setup(v), n); -} - -struct __get_user futex_cmpxchg(int __user *v, int o, int n) -{ - return __atomic_cmpxchg((int __force *)v, __futex_setup(v), o, n); -} - /* * If any of the atomic or futex routines hit a bad address (not in * the page tables at kernel PL) this routine is called. The futex @@ -251,54 +155,8 @@ struct __get_user __atomic_bad_address(int __user *addr) } -#if CHIP_HAS_CBOX_HOME_MAP() -static int __init noatomichash(char *str) -{ - pr_warning("noatomichash is deprecated.\n"); - return 1; -} -__setup("noatomichash", noatomichash); -#endif - void __init __init_atomic_per_cpu(void) { -#if ATOMIC_LOCKS_FOUND_VIA_TABLE() - - unsigned int i; - int actual_cpu; - - /* - * Before this is called from setup, we just have one lock for - * all atomic objects/operations. Here we replace the - * elements of atomic_lock_ptr so that they point at per_cpu - * integers. This seemingly over-complex approach stems from - * the fact that DEFINE_PER_CPU defines an entry for each cpu - * in the grid, not each cpu from 0..ATOMIC_HASH_SIZE-1. But - * for efficient hashing of atomics to their locks we want a - * compile time constant power of 2 for the size of this - * table, so we use ATOMIC_HASH_SIZE. - * - * Here we populate atomic_lock_ptr from the per cpu - * atomic_lock_pool, interspersing by actual cpu so that - * subsequent elements are homed on consecutive cpus. - */ - - actual_cpu = cpumask_first(cpu_possible_mask); - - for (i = 0; i < ATOMIC_HASH_L1_SIZE; ++i) { - /* - * Preincrement to slightly bias against using cpu 0, - * which has plenty of stuff homed on it already. - */ - actual_cpu = cpumask_next(actual_cpu, cpu_possible_mask); - if (actual_cpu >= nr_cpu_ids) - actual_cpu = cpumask_first(cpu_possible_mask); - - atomic_lock_ptr[i] = &per_cpu(atomic_lock_pool, actual_cpu); - } - -#else /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */ - /* Validate power-of-two and "bigger than cpus" assumption */ BUILD_BUG_ON(ATOMIC_HASH_SIZE & (ATOMIC_HASH_SIZE-1)); BUG_ON(ATOMIC_HASH_SIZE < nr_cpu_ids); @@ -322,9 +180,4 @@ void __init __init_atomic_per_cpu(void) * That should not produce more indices than ATOMIC_HASH_SIZE. */ BUILD_BUG_ON((PAGE_SIZE >> 3) > ATOMIC_HASH_SIZE); - -#endif /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */ - - /* The futex code makes this assumption, so we validate it here. */ - BUILD_BUG_ON(sizeof(atomic_t) != sizeof(int)); } diff --git a/arch/tile/lib/atomic_asm_32.S b/arch/tile/lib/atomic_asm_32.S index 5a5514b77e7..6bda3132cd6 100644 --- a/arch/tile/lib/atomic_asm_32.S +++ b/arch/tile/lib/atomic_asm_32.S @@ -14,7 +14,7 @@ * Support routines for atomic operations. Each function takes: * * r0: address to manipulate - * r1: pointer to atomic lock guarding this operation (for FUTEX_LOCK_REG) + * r1: pointer to atomic lock guarding this operation (for ATOMIC_LOCK_REG) * r2: new value to write, or for cmpxchg/add_unless, value to compare against * r3: (cmpxchg/xchg_add_unless) new value to write or add; * (atomic64 ops) high word of value to write @@ -59,7 +59,7 @@ * bad kernel addresses). * * Note that if the value we would store is the same as what we - * loaded, we bypass the load. Other platforms with true atomics can + * loaded, we bypass the store. Other platforms with true atomics can * make the guarantee that a non-atomic __clear_bit(), for example, * can safely race with an atomic test_and_set_bit(); this example is * from bit_spinlock.h in slub_lock() / slub_unlock(). We can't do @@ -70,7 +70,7 @@ */ #include <linux/linkage.h> -#include <asm/atomic.h> +#include <asm/atomic_32.h> #include <asm/page.h> #include <asm/processor.h> @@ -164,6 +164,7 @@ STD_ENTRY_SECTION(__atomic\name, .text.atomic) STD_ENDPROC(__atomic\name) .ifc \bitwidth,32 .pushsection __ex_table,"a" + .align 4 .word 1b, __atomic\name .word 2b, __atomic\name .word __atomic\name, __atomic_bad_address diff --git a/arch/tile/lib/cacheflush.c b/arch/tile/lib/cacheflush.c index 11b6164c209..9c0ec22009a 100644 --- a/arch/tile/lib/cacheflush.c +++ b/arch/tile/lib/cacheflush.c @@ -12,12 +12,162 @@ * more details. */ +#include <linux/export.h> #include <asm/page.h> #include <asm/cacheflush.h> #include <arch/icache.h> +#include <arch/spr_def.h> void __flush_icache_range(unsigned long start, unsigned long end) { invalidate_icache((const void *)start, end - start, PAGE_SIZE); } + + +/* Force a load instruction to issue. */ +static inline void force_load(char *p) +{ + *(volatile char *)p; +} + +/* + * Flush and invalidate a VA range that is homed remotely on a single + * core (if "!hfh") or homed via hash-for-home (if "hfh"), waiting + * until the memory controller holds the flushed values. + */ +void __attribute__((optimize("omit-frame-pointer"))) +finv_buffer_remote(void *buffer, size_t size, int hfh) +{ + char *p, *base; + size_t step_size, load_count; + + /* + * On TILEPro the striping granularity is a fixed 8KB; on + * TILE-Gx it is configurable, and we rely on the fact that + * the hypervisor always configures maximum striping, so that + * bits 9 and 10 of the PA are part of the stripe function, so + * every 512 bytes we hit a striping boundary. + * + */ +#ifdef __tilegx__ + const unsigned long STRIPE_WIDTH = 512; +#else + const unsigned long STRIPE_WIDTH = 8192; +#endif + +#ifdef __tilegx__ + /* + * On TILE-Gx, we must disable the dstream prefetcher before doing + * a cache flush; otherwise, we could end up with data in the cache + * that we don't want there. Note that normally we'd do an mf + * after the SPR write to disabling the prefetcher, but we do one + * below, before any further loads, so there's no need to do it + * here. + */ + uint_reg_t old_dstream_pf = __insn_mfspr(SPR_DSTREAM_PF); + __insn_mtspr(SPR_DSTREAM_PF, 0); +#endif + + /* + * Flush and invalidate the buffer out of the local L1/L2 + * and request the home cache to flush and invalidate as well. + */ + __finv_buffer(buffer, size); + + /* + * Wait for the home cache to acknowledge that it has processed + * all the flush-and-invalidate requests. This does not mean + * that the flushed data has reached the memory controller yet, + * but it does mean the home cache is processing the flushes. + */ + __insn_mf(); + + /* + * Issue a load to the last cache line, which can't complete + * until all the previously-issued flushes to the same memory + * controller have also completed. If we weren't striping + * memory, that one load would be sufficient, but since we may + * be, we also need to back up to the last load issued to + * another memory controller, which would be the point where + * we crossed a "striping" boundary (the granularity of striping + * across memory controllers). Keep backing up and doing this + * until we are before the beginning of the buffer, or have + * hit all the controllers. + * + * If we are flushing a hash-for-home buffer, it's even worse. + * Each line may be homed on a different tile, and each tile + * may have up to four lines that are on different + * controllers. So as we walk backwards, we have to touch + * enough cache lines to satisfy these constraints. In + * practice this ends up being close enough to "load from + * every cache line on a full memory stripe on each + * controller" that we simply do that, to simplify the logic. + * + * On TILE-Gx the hash-for-home function is much more complex, + * with the upshot being we can't readily guarantee we have + * hit both entries in the 128-entry AMT that were hit by any + * load in the entire range, so we just re-load them all. + * With larger buffers, we may want to consider using a hypervisor + * trap to issue loads directly to each hash-for-home tile for + * each controller (doing it from Linux would trash the TLB). + */ + if (hfh) { + step_size = L2_CACHE_BYTES; +#ifdef __tilegx__ + load_count = (size + L2_CACHE_BYTES - 1) / L2_CACHE_BYTES; +#else + load_count = (STRIPE_WIDTH / L2_CACHE_BYTES) * + (1 << CHIP_LOG_NUM_MSHIMS()); +#endif + } else { + step_size = STRIPE_WIDTH; + load_count = (1 << CHIP_LOG_NUM_MSHIMS()); + } + + /* Load the last byte of the buffer. */ + p = (char *)buffer + size - 1; + force_load(p); + + /* Bump down to the end of the previous stripe or cache line. */ + p -= step_size; + p = (char *)((unsigned long)p | (step_size - 1)); + + /* Figure out how far back we need to go. */ + base = p - (step_size * (load_count - 2)); + if ((unsigned long)base < (unsigned long)buffer) + base = buffer; + + /* + * Fire all the loads we need. The MAF only has eight entries + * so we can have at most eight outstanding loads, so we + * unroll by that amount. + */ +#pragma unroll 8 + for (; p >= base; p -= step_size) + force_load(p); + + /* + * Repeat, but with finv's instead of loads, to get rid of the + * data we just loaded into our own cache and the old home L3. + * No need to unroll since finv's don't target a register. + * The finv's are guaranteed not to actually flush the data in + * the buffer back to their home, since we just read it, so the + * lines are clean in cache; we will only invalidate those lines. + */ + p = (char *)buffer + size - 1; + __insn_finv(p); + p -= step_size; + p = (char *)((unsigned long)p | (step_size - 1)); + for (; p >= base; p -= step_size) + __insn_finv(p); + + /* Wait for these finv's (and thus the first finvs) to be done. */ + __insn_mf(); + +#ifdef __tilegx__ + /* Reenable the prefetcher. */ + __insn_mtspr(SPR_DSTREAM_PF, old_dstream_pf); +#endif +} +EXPORT_SYMBOL_GPL(finv_buffer_remote); diff --git a/arch/tile/lib/checksum.c b/arch/tile/lib/checksum.c index e4bab5bd3f3..c3ca3e64d9d 100644 --- a/arch/tile/lib/checksum.c +++ b/arch/tile/lib/checksum.c @@ -16,19 +16,6 @@ #include <net/checksum.h> #include <linux/module.h> -static inline unsigned int longto16(unsigned long x) -{ - unsigned long ret; -#ifdef __tilegx__ - ret = __insn_v2sadu(x, 0); - ret = __insn_v2sadu(ret, 0); -#else - ret = __insn_sadh_u(x, 0); - ret = __insn_sadh_u(ret, 0); -#endif - return ret; -} - __wsum do_csum(const unsigned char *buff, int len) { int odd, count; @@ -94,7 +81,7 @@ __wsum do_csum(const unsigned char *buff, int len) } if (len & 1) result += *buff; - result = longto16(result); + result = csum_long(result); if (odd) result = swab16(result); out: diff --git a/arch/tile/lib/cpumask.c b/arch/tile/lib/cpumask.c index fdc403614d1..75947edccb2 100644 --- a/arch/tile/lib/cpumask.c +++ b/arch/tile/lib/cpumask.c @@ -16,6 +16,7 @@ #include <linux/ctype.h> #include <linux/errno.h> #include <linux/smp.h> +#include <linux/export.h> /* * Allow cropping out bits beyond the end of the array. @@ -50,3 +51,4 @@ int bitmap_parselist_crop(const char *bp, unsigned long *maskp, int nmaskbits) } while (*bp != '\0' && *bp != '\n'); return 0; } +EXPORT_SYMBOL(bitmap_parselist_crop); diff --git a/arch/tile/lib/delay.c b/arch/tile/lib/delay.c index 5801b03c13e..cdacdd11d36 100644 --- a/arch/tile/lib/delay.c +++ b/arch/tile/lib/delay.c @@ -15,20 +15,31 @@ #include <linux/module.h> #include <linux/delay.h> #include <linux/thread_info.h> -#include <asm/fixmap.h> -#include <hv/hypervisor.h> +#include <asm/timex.h> void __udelay(unsigned long usecs) { - hv_nanosleep(usecs * 1000); + if (usecs > ULONG_MAX / 1000) { + WARN_ON_ONCE(usecs > ULONG_MAX / 1000); + usecs = ULONG_MAX / 1000; + } + __ndelay(usecs * 1000); } EXPORT_SYMBOL(__udelay); void __ndelay(unsigned long nsecs) { - hv_nanosleep(nsecs); + cycles_t target = get_cycles(); + target += ns2cycles(nsecs); + while (get_cycles() < target) + cpu_relax(); } EXPORT_SYMBOL(__ndelay); -/* FIXME: should be declared in a header somewhere. */ +void __delay(unsigned long cycles) +{ + cycles_t target = get_cycles() + cycles; + while (get_cycles() < target) + cpu_relax(); +} EXPORT_SYMBOL(__delay); diff --git a/arch/tile/lib/exports.c b/arch/tile/lib/exports.c index 1509c559765..82733c87d67 100644 --- a/arch/tile/lib/exports.c +++ b/arch/tile/lib/exports.c @@ -18,17 +18,11 @@ /* arch/tile/lib/usercopy.S */ #include <linux/uaccess.h> -EXPORT_SYMBOL(__get_user_1); -EXPORT_SYMBOL(__get_user_2); -EXPORT_SYMBOL(__get_user_4); -EXPORT_SYMBOL(__get_user_8); -EXPORT_SYMBOL(__put_user_1); -EXPORT_SYMBOL(__put_user_2); -EXPORT_SYMBOL(__put_user_4); -EXPORT_SYMBOL(__put_user_8); EXPORT_SYMBOL(strnlen_user_asm); EXPORT_SYMBOL(strncpy_from_user_asm); EXPORT_SYMBOL(clear_user_asm); +EXPORT_SYMBOL(flush_user_asm); +EXPORT_SYMBOL(finv_user_asm); /* arch/tile/kernel/entry.S */ #include <linux/kernel.h> @@ -36,6 +30,15 @@ EXPORT_SYMBOL(clear_user_asm); EXPORT_SYMBOL(current_text_addr); EXPORT_SYMBOL(dump_stack); +/* arch/tile/kernel/head.S */ +EXPORT_SYMBOL(empty_zero_page); + +#ifdef CONFIG_FUNCTION_TRACER +/* arch/tile/kernel/mcount_64.S */ +#include <asm/ftrace.h> +EXPORT_SYMBOL(__mcount); +#endif /* CONFIG_FUNCTION_TRACER */ + /* arch/tile/lib/, various memcpy files */ EXPORT_SYMBOL(memcpy); EXPORT_SYMBOL(__copy_to_user_inatomic); @@ -45,9 +48,6 @@ EXPORT_SYMBOL(__copy_from_user_zeroing); EXPORT_SYMBOL(__copy_in_user_inatomic); #endif -/* arch/tile/lib/mb_incoherent.S */ -EXPORT_SYMBOL(__mb_incoherent); - /* hypervisor glue */ #include <hv/hypervisor.h> EXPORT_SYMBOL(hv_dev_open); @@ -60,6 +60,8 @@ EXPORT_SYMBOL(hv_dev_poll_cancel); EXPORT_SYMBOL(hv_dev_close); EXPORT_SYMBOL(hv_sysconf); EXPORT_SYMBOL(hv_confstr); +EXPORT_SYMBOL(hv_get_rtc); +EXPORT_SYMBOL(hv_set_rtc); /* libgcc.a */ uint32_t __udivsi3(uint32_t dividend, uint32_t divisor); @@ -79,10 +81,14 @@ EXPORT_SYMBOL(__umoddi3); int64_t __moddi3(int64_t dividend, int64_t divisor); EXPORT_SYMBOL(__moddi3); #ifndef __tilegx__ -uint64_t __ll_mul(uint64_t n0, uint64_t n1); -EXPORT_SYMBOL(__ll_mul); int64_t __muldi3(int64_t, int64_t); EXPORT_SYMBOL(__muldi3); uint64_t __lshrdi3(uint64_t, unsigned int); EXPORT_SYMBOL(__lshrdi3); +uint64_t __ashrdi3(uint64_t, unsigned int); +EXPORT_SYMBOL(__ashrdi3); +uint64_t __ashldi3(uint64_t, unsigned int); +EXPORT_SYMBOL(__ashldi3); +int __ffsdi2(uint64_t); +EXPORT_SYMBOL(__ffsdi2); #endif diff --git a/arch/tile/lib/mb_incoherent.S b/arch/tile/lib/mb_incoherent.S deleted file mode 100644 index 989ad7b68d5..00000000000 --- a/arch/tile/lib/mb_incoherent.S +++ /dev/null @@ -1,34 +0,0 @@ -/* - * Copyright 2010 Tilera Corporation. All Rights Reserved. - * - * This program is free software; you can redistribute it and/or - * modify it under the terms of the GNU General Public License - * as published by the Free Software Foundation, version 2. - * - * This program is distributed in the hope that it will be useful, but - * WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or - * NON INFRINGEMENT. See the GNU General Public License for - * more details. - * - * Assembly code for invoking the HV's fence_incoherent syscall. - */ - -#include <linux/linkage.h> -#include <hv/syscall_public.h> -#include <arch/abi.h> -#include <arch/chip.h> - -#if !CHIP_HAS_MF_WAITS_FOR_VICTIMS() - -/* - * Invoke the hypervisor's fence_incoherent syscall, which guarantees - * that all victims for cachelines homed on this tile have reached memory. - */ -STD_ENTRY(__mb_incoherent) - moveli TREG_SYSCALL_NR_NAME, HV_SYS_fence_incoherent - swint2 - jrp lr - STD_ENDPROC(__mb_incoherent) - -#endif diff --git a/arch/tile/lib/memchr_32.c b/arch/tile/lib/memchr_32.c index 6235283b485..cc3d9badf03 100644 --- a/arch/tile/lib/memchr_32.c +++ b/arch/tile/lib/memchr_32.c @@ -18,12 +18,24 @@ void *memchr(const void *s, int c, size_t n) { + const uint32_t *last_word_ptr; + const uint32_t *p; + const char *last_byte_ptr; + uintptr_t s_int; + uint32_t goal, before_mask, v, bits; + char *ret; + + if (__builtin_expect(n == 0, 0)) { + /* Don't dereference any memory if the array is empty. */ + return NULL; + } + /* Get an aligned pointer. */ - const uintptr_t s_int = (uintptr_t) s; - const uint32_t *p = (const uint32_t *)(s_int & -4); + s_int = (uintptr_t) s; + p = (const uint32_t *)(s_int & -4); /* Create four copies of the byte for which we are looking. */ - const uint32_t goal = 0x01010101 * (uint8_t) c; + goal = 0x01010101 * (uint8_t) c; /* Read the first word, but munge it so that bytes before the array * will not match goal. @@ -31,23 +43,14 @@ void *memchr(const void *s, int c, size_t n) * Note that this shift count expression works because we know * shift counts are taken mod 32. */ - const uint32_t before_mask = (1 << (s_int << 3)) - 1; - uint32_t v = (*p | before_mask) ^ (goal & before_mask); + before_mask = (1 << (s_int << 3)) - 1; + v = (*p | before_mask) ^ (goal & before_mask); /* Compute the address of the last byte. */ - const char *const last_byte_ptr = (const char *)s + n - 1; + last_byte_ptr = (const char *)s + n - 1; /* Compute the address of the word containing the last byte. */ - const uint32_t *const last_word_ptr = - (const uint32_t *)((uintptr_t) last_byte_ptr & -4); - - uint32_t bits; - char *ret; - - if (__builtin_expect(n == 0, 0)) { - /* Don't dereference any memory if the array is empty. */ - return NULL; - } + last_word_ptr = (const uint32_t *)((uintptr_t) last_byte_ptr & -4); while ((bits = __insn_seqb(v, goal)) == 0) { if (__builtin_expect(p == last_word_ptr, 0)) { diff --git a/arch/tile/lib/memchr_64.c b/arch/tile/lib/memchr_64.c new file mode 100644 index 00000000000..f8196b3a950 --- /dev/null +++ b/arch/tile/lib/memchr_64.c @@ -0,0 +1,69 @@ +/* + * Copyright 2011 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + */ + +#include <linux/types.h> +#include <linux/string.h> +#include <linux/module.h> +#include "string-endian.h" + +void *memchr(const void *s, int c, size_t n) +{ + const uint64_t *last_word_ptr; + const uint64_t *p; + const char *last_byte_ptr; + uintptr_t s_int; + uint64_t goal, before_mask, v, bits; + char *ret; + + if (__builtin_expect(n == 0, 0)) { + /* Don't dereference any memory if the array is empty. */ + return NULL; + } + + /* Get an aligned pointer. */ + s_int = (uintptr_t) s; + p = (const uint64_t *)(s_int & -8); + + /* Create eight copies of the byte for which we are looking. */ + goal = copy_byte(c); + + /* Read the first word, but munge it so that bytes before the array + * will not match goal. + */ + before_mask = MASK(s_int); + v = (*p | before_mask) ^ (goal & before_mask); + + /* Compute the address of the last byte. */ + last_byte_ptr = (const char *)s + n - 1; + + /* Compute the address of the word containing the last byte. */ + last_word_ptr = (const uint64_t *)((uintptr_t) last_byte_ptr & -8); + + while ((bits = __insn_v1cmpeq(v, goal)) == 0) { + if (__builtin_expect(p == last_word_ptr, 0)) { + /* We already read the last word in the array, + * so give up. + */ + return NULL; + } + v = *++p; + } + + /* We found a match, but it might be in a byte past the end + * of the array. + */ + ret = ((char *)p) + (CFZ(bits) >> 3); + return (ret <= last_byte_ptr) ? ret : NULL; +} +EXPORT_SYMBOL(memchr); diff --git a/arch/tile/lib/memcpy_32.S b/arch/tile/lib/memcpy_32.S index 2a419a6122d..a2771ae5da5 100644 --- a/arch/tile/lib/memcpy_32.S +++ b/arch/tile/lib/memcpy_32.S @@ -22,14 +22,6 @@ #include <linux/linkage.h> -/* On TILE64, we wrap these functions via arch/tile/lib/memcpy_tile64.c */ -#if !CHIP_HAS_COHERENT_LOCAL_CACHE() -#define memcpy __memcpy_asm -#define __copy_to_user_inatomic __copy_to_user_inatomic_asm -#define __copy_from_user_inatomic __copy_from_user_inatomic_asm -#define __copy_from_user_zeroing __copy_from_user_zeroing_asm -#endif - #define IS_MEMCPY 0 #define IS_COPY_FROM_USER 1 #define IS_COPY_FROM_USER_ZEROING 2 @@ -44,6 +36,7 @@ */ #define EX \ .pushsection __ex_table, "a"; \ + .align 4; \ .word 9f, memcpy_common_fixup; \ .popsection; \ 9 @@ -158,12 +151,9 @@ EX: { sw r0, r3; addi r0, r0, 4; addi r2, r2, -4 } { addi r3, r1, 60; andi r9, r9, -64 } -#if CHIP_HAS_WH64() /* No need to prefetch dst, we'll just do the wh64 * right before we copy a line. */ -#endif - EX: { lw r5, r3; addi r3, r3, 64; movei r4, 1 } /* Intentionally stall for a few cycles to leave L2 cache alone. */ { bnzt zero, .; move r27, lr } @@ -171,21 +161,6 @@ EX: { lw r6, r3; addi r3, r3, 64 } /* Intentionally stall for a few cycles to leave L2 cache alone. */ { bnzt zero, . } EX: { lw r7, r3; addi r3, r3, 64 } -#if !CHIP_HAS_WH64() - /* Prefetch the dest */ - /* Intentionally stall for a few cycles to leave L2 cache alone. */ - { bnzt zero, . } - /* Use a real load to cause a TLB miss if necessary. We aren't using - * r28, so this should be fine. - */ -EX: { lw r28, r9; addi r9, r9, 64 } - /* Intentionally stall for a few cycles to leave L2 cache alone. */ - { bnzt zero, . } - { prefetch r9; addi r9, r9, 64 } - /* Intentionally stall for a few cycles to leave L2 cache alone. */ - { bnzt zero, . } - { prefetch r9; addi r9, r9, 64 } -#endif /* Intentionally stall for a few cycles to leave L2 cache alone. */ { bz zero, .Lbig_loop2 } @@ -286,13 +261,8 @@ EX: { lw r7, r3; addi r3, r3, 64 } /* Fill second L1D line. */ EX: { lw r17, r17; addi r1, r1, 48; mvz r3, r13, r1 } /* r17 = WORD_4 */ -#if CHIP_HAS_WH64() /* Prepare destination line for writing. */ EX: { wh64 r9; addi r9, r9, 64 } -#else - /* Prefetch dest line */ - { prefetch r9; addi r9, r9, 64 } -#endif /* Load seven words that are L1D hits to cover wh64 L2 usage. */ /* Load the three remaining words from the last L1D line, which @@ -330,16 +300,7 @@ EX: { lw r18, r1; addi r1, r1, 4 } /* r18 = WORD_8 */ EX: { sw r0, r16; addi r0, r0, 4; add r16, r0, r2 } /* store(WORD_0) */ EX: { sw r0, r13; addi r0, r0, 4; andi r16, r16, -64 } /* store(WORD_1) */ EX: { sw r0, r14; addi r0, r0, 4; slt_u r16, r9, r16 } /* store(WORD_2) */ -#if CHIP_HAS_WH64() EX: { sw r0, r15; addi r0, r0, 4; addi r13, sp, -64 } /* store(WORD_3) */ -#else - /* Back up the r9 to a cache line we are already storing to - * if it gets past the end of the dest vector. Strictly speaking, - * we don't need to back up to the start of a cache line, but it's free - * and tidy, so why not? - */ -EX: { sw r0, r15; addi r0, r0, 4; andi r13, r0, -64 } /* store(WORD_3) */ -#endif /* Store second L1D line. */ EX: { sw r0, r17; addi r0, r0, 4; mvz r9, r16, r13 }/* store(WORD_4) */ EX: { sw r0, r19; addi r0, r0, 4 } /* store(WORD_5) */ @@ -403,7 +364,6 @@ EX: { sb r0, r3; addi r0, r0, 1; addi r2, r2, -1 } .Ldest_is_word_aligned: -#if CHIP_HAS_DWORD_ALIGN() EX: { andi r8, r0, 63; lwadd_na r6, r1, 4} { slti_u r9, r2, 64; bz r8, .Ldest_is_L2_line_aligned } @@ -511,26 +471,6 @@ EX: { swadd r0, r13, 4; addi r2, r2, -32 } /* Move r1 back to the point where it corresponds to r0. */ { addi r1, r1, -4 } -#else /* !CHIP_HAS_DWORD_ALIGN() */ - - /* Compute right/left shift counts and load initial source words. */ - { andi r5, r1, -4; andi r3, r1, 3 } -EX: { lw r6, r5; addi r5, r5, 4; shli r3, r3, 3 } -EX: { lw r7, r5; addi r5, r5, 4; sub r4, zero, r3 } - - /* Load and store one word at a time, using shifts and ORs - * to correct for the misaligned src. - */ -.Lcopy_unaligned_src_loop: - { shr r6, r6, r3; shl r8, r7, r4 } -EX: { lw r7, r5; or r8, r8, r6; move r6, r7 } -EX: { sw r0, r8; addi r0, r0, 4; addi r2, r2, -4 } - { addi r5, r5, 4; slti_u r8, r2, 8 } - { bzt r8, .Lcopy_unaligned_src_loop; addi r1, r1, 4 } - - { bz r2, .Lcopy_unaligned_done } -#endif /* !CHIP_HAS_DWORD_ALIGN() */ - /* Fall through */ /* @@ -614,5 +554,6 @@ memcpy_fixup_loop: .size memcpy_common_fixup, . - memcpy_common_fixup .section __ex_table,"a" + .align 4 .word .Lcfu, .Lcopy_from_user_fixup_zero_remainder .word .Lctu, .Lcopy_to_user_fixup_done diff --git a/arch/tile/lib/memcpy_64.c b/arch/tile/lib/memcpy_64.c new file mode 100644 index 00000000000..4815354b8cd --- /dev/null +++ b/arch/tile/lib/memcpy_64.c @@ -0,0 +1,367 @@ +/* + * Copyright 2011 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + */ + +#include <linux/types.h> +#include <linux/string.h> +#include <linux/module.h> +/* EXPORT_SYMBOL() is in arch/tile/lib/exports.c since this should be asm. */ + +/* Must be 8 bytes in size. */ +#define op_t uint64_t + +/* Threshold value for when to enter the unrolled loops. */ +#define OP_T_THRES 16 + +#if CHIP_L2_LINE_SIZE() != 64 +#error "Assumes 64 byte line size" +#endif + +/* How many cache lines ahead should we prefetch? */ +#define PREFETCH_LINES_AHEAD 4 + +/* + * Provide "base versions" of load and store for the normal code path. + * The kernel provides other versions for userspace copies. + */ +#define ST(p, v) (*(p) = (v)) +#define LD(p) (*(p)) + +#ifndef USERCOPY_FUNC +#define ST1 ST +#define ST2 ST +#define ST4 ST +#define ST8 ST +#define LD1 LD +#define LD2 LD +#define LD4 LD +#define LD8 LD +#define RETVAL dstv +void *memcpy(void *__restrict dstv, const void *__restrict srcv, size_t n) +#else +/* + * Special kernel version will provide implementation of the LDn/STn + * macros to return a count of uncopied bytes due to mm fault. + */ +#define RETVAL 0 +int __attribute__((optimize("omit-frame-pointer"))) +USERCOPY_FUNC(void *__restrict dstv, const void *__restrict srcv, size_t n) +#endif +{ + char *__restrict dst1 = (char *)dstv; + const char *__restrict src1 = (const char *)srcv; + const char *__restrict src1_end; + const char *__restrict prefetch; + op_t *__restrict dst8; /* 8-byte pointer to destination memory. */ + op_t final; /* Final bytes to write to trailing word, if any */ + long i; + + if (n < 16) { + for (; n; n--) + ST1(dst1++, LD1(src1++)); + return RETVAL; + } + + /* + * Locate the end of source memory we will copy. Don't + * prefetch past this. + */ + src1_end = src1 + n - 1; + + /* Prefetch ahead a few cache lines, but not past the end. */ + prefetch = src1; + for (i = 0; i < PREFETCH_LINES_AHEAD; i++) { + __insn_prefetch(prefetch); + prefetch += CHIP_L2_LINE_SIZE(); + prefetch = (prefetch < src1_end) ? prefetch : src1; + } + + /* Copy bytes until dst is word-aligned. */ + for (; (uintptr_t)dst1 & (sizeof(op_t) - 1); n--) + ST1(dst1++, LD1(src1++)); + + /* 8-byte pointer to destination memory. */ + dst8 = (op_t *)dst1; + + if (__builtin_expect((uintptr_t)src1 & (sizeof(op_t) - 1), 0)) { + /* Unaligned copy. */ + + op_t tmp0 = 0, tmp1 = 0, tmp2, tmp3; + const op_t *src8 = (const op_t *) ((uintptr_t)src1 & + -sizeof(op_t)); + const void *srci = (void *)src1; + int m; + + m = (CHIP_L2_LINE_SIZE() << 2) - + (((uintptr_t)dst8) & ((CHIP_L2_LINE_SIZE() << 2) - 1)); + m = (n < m) ? n : m; + m /= sizeof(op_t); + + /* Copy until 'dst' is cache-line-aligned. */ + n -= (sizeof(op_t) * m); + + switch (m % 4) { + case 0: + if (__builtin_expect(!m, 0)) + goto _M0; + tmp1 = LD8(src8++); + tmp2 = LD8(src8++); + goto _8B3; + case 2: + m += 2; + tmp3 = LD8(src8++); + tmp0 = LD8(src8++); + goto _8B1; + case 3: + m += 1; + tmp2 = LD8(src8++); + tmp3 = LD8(src8++); + goto _8B2; + case 1: + m--; + tmp0 = LD8(src8++); + tmp1 = LD8(src8++); + if (__builtin_expect(!m, 0)) + goto _8B0; + } + + do { + tmp2 = LD8(src8++); + tmp0 = __insn_dblalign(tmp0, tmp1, srci); + ST8(dst8++, tmp0); +_8B3: + tmp3 = LD8(src8++); + tmp1 = __insn_dblalign(tmp1, tmp2, srci); + ST8(dst8++, tmp1); +_8B2: + tmp0 = LD8(src8++); + tmp2 = __insn_dblalign(tmp2, tmp3, srci); + ST8(dst8++, tmp2); +_8B1: + tmp1 = LD8(src8++); + tmp3 = __insn_dblalign(tmp3, tmp0, srci); + ST8(dst8++, tmp3); + m -= 4; + } while (m); + +_8B0: + tmp0 = __insn_dblalign(tmp0, tmp1, srci); + ST8(dst8++, tmp0); + src8--; + +_M0: + if (__builtin_expect(n >= CHIP_L2_LINE_SIZE(), 0)) { + op_t tmp4, tmp5, tmp6, tmp7, tmp8; + + prefetch = ((const char *)src8) + + CHIP_L2_LINE_SIZE() * PREFETCH_LINES_AHEAD; + + for (tmp0 = LD8(src8++); n >= CHIP_L2_LINE_SIZE(); + n -= CHIP_L2_LINE_SIZE()) { + /* Prefetch and advance to next line to + prefetch, but don't go past the end. */ + __insn_prefetch(prefetch); + + /* Make sure prefetch got scheduled + earlier. */ + __asm__ ("" : : : "memory"); + + prefetch += CHIP_L2_LINE_SIZE(); + prefetch = (prefetch < src1_end) ? prefetch : + (const char *) src8; + + tmp1 = LD8(src8++); + tmp2 = LD8(src8++); + tmp3 = LD8(src8++); + tmp4 = LD8(src8++); + tmp5 = LD8(src8++); + tmp6 = LD8(src8++); + tmp7 = LD8(src8++); + tmp8 = LD8(src8++); + + tmp0 = __insn_dblalign(tmp0, tmp1, srci); + tmp1 = __insn_dblalign(tmp1, tmp2, srci); + tmp2 = __insn_dblalign(tmp2, tmp3, srci); + tmp3 = __insn_dblalign(tmp3, tmp4, srci); + tmp4 = __insn_dblalign(tmp4, tmp5, srci); + tmp5 = __insn_dblalign(tmp5, tmp6, srci); + tmp6 = __insn_dblalign(tmp6, tmp7, srci); + tmp7 = __insn_dblalign(tmp7, tmp8, srci); + + __insn_wh64(dst8); + + ST8(dst8++, tmp0); + ST8(dst8++, tmp1); + ST8(dst8++, tmp2); + ST8(dst8++, tmp3); + ST8(dst8++, tmp4); + ST8(dst8++, tmp5); + ST8(dst8++, tmp6); + ST8(dst8++, tmp7); + + tmp0 = tmp8; + } + src8--; + } + + /* Copy the rest 8-byte chunks. */ + if (n >= sizeof(op_t)) { + tmp0 = LD8(src8++); + for (; n >= sizeof(op_t); n -= sizeof(op_t)) { + tmp1 = LD8(src8++); + tmp0 = __insn_dblalign(tmp0, tmp1, srci); + ST8(dst8++, tmp0); + tmp0 = tmp1; + } + src8--; + } + + if (n == 0) + return RETVAL; + + tmp0 = LD8(src8++); + tmp1 = ((const char *)src8 <= src1_end) + ? LD8((op_t *)src8) : 0; + final = __insn_dblalign(tmp0, tmp1, srci); + + } else { + /* Aligned copy. */ + + const op_t *__restrict src8 = (const op_t *)src1; + + /* src8 and dst8 are both word-aligned. */ + if (n >= CHIP_L2_LINE_SIZE()) { + /* Copy until 'dst' is cache-line-aligned. */ + for (; (uintptr_t)dst8 & (CHIP_L2_LINE_SIZE() - 1); + n -= sizeof(op_t)) + ST8(dst8++, LD8(src8++)); + + for (; n >= CHIP_L2_LINE_SIZE(); ) { + op_t tmp0, tmp1, tmp2, tmp3; + op_t tmp4, tmp5, tmp6, tmp7; + + /* + * Prefetch and advance to next line + * to prefetch, but don't go past the + * end. + */ + __insn_prefetch(prefetch); + + /* Make sure prefetch got scheduled + earlier. */ + __asm__ ("" : : : "memory"); + + prefetch += CHIP_L2_LINE_SIZE(); + prefetch = (prefetch < src1_end) ? prefetch : + (const char *)src8; + + /* + * Do all the loads before wh64. This + * is necessary if [src8, src8+7] and + * [dst8, dst8+7] share the same cache + * line and dst8 <= src8, as can be + * the case when called from memmove, + * or with code tested on x86 whose + * memcpy always works with forward + * copies. + */ + tmp0 = LD8(src8++); + tmp1 = LD8(src8++); + tmp2 = LD8(src8++); + tmp3 = LD8(src8++); + tmp4 = LD8(src8++); + tmp5 = LD8(src8++); + tmp6 = LD8(src8++); + tmp7 = LD8(src8++); + + /* wh64 and wait for tmp7 load completion. */ + __asm__ ("move %0, %0; wh64 %1\n" + : : "r"(tmp7), "r"(dst8)); + + ST8(dst8++, tmp0); + ST8(dst8++, tmp1); + ST8(dst8++, tmp2); + ST8(dst8++, tmp3); + ST8(dst8++, tmp4); + ST8(dst8++, tmp5); + ST8(dst8++, tmp6); + ST8(dst8++, tmp7); + + n -= CHIP_L2_LINE_SIZE(); + } +#if CHIP_L2_LINE_SIZE() != 64 +# error "Fix code that assumes particular L2 cache line size." +#endif + } + + for (; n >= sizeof(op_t); n -= sizeof(op_t)) + ST8(dst8++, LD8(src8++)); + + if (__builtin_expect(n == 0, 1)) + return RETVAL; + + final = LD8(src8); + } + + /* n != 0 if we get here. Write out any trailing bytes. */ + dst1 = (char *)dst8; +#ifndef __BIG_ENDIAN__ + if (n & 4) { + ST4((uint32_t *)dst1, final); + dst1 += 4; + final >>= 32; + n &= 3; + } + if (n & 2) { + ST2((uint16_t *)dst1, final); + dst1 += 2; + final >>= 16; + n &= 1; + } + if (n) + ST1((uint8_t *)dst1, final); +#else + if (n & 4) { + ST4((uint32_t *)dst1, final >> 32); + dst1 += 4; + } + else + { + final >>= 32; + } + if (n & 2) { + ST2((uint16_t *)dst1, final >> 16); + dst1 += 2; + } + else + { + final >>= 16; + } + if (n & 1) + ST1((uint8_t *)dst1, final >> 8); +#endif + + return RETVAL; +} + +#ifdef USERCOPY_FUNC +#undef ST1 +#undef ST2 +#undef ST4 +#undef ST8 +#undef LD1 +#undef LD2 +#undef LD4 +#undef LD8 +#undef USERCOPY_FUNC +#endif diff --git a/arch/tile/lib/memcpy_tile64.c b/arch/tile/lib/memcpy_tile64.c deleted file mode 100644 index f7d4a6ad61e..00000000000 --- a/arch/tile/lib/memcpy_tile64.c +++ /dev/null @@ -1,276 +0,0 @@ -/* - * Copyright 2010 Tilera Corporation. All Rights Reserved. - * - * This program is free software; you can redistribute it and/or - * modify it under the terms of the GNU General Public License - * as published by the Free Software Foundation, version 2. - * - * This program is distributed in the hope that it will be useful, but - * WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or - * NON INFRINGEMENT. See the GNU General Public License for - * more details. - */ - -#include <linux/string.h> -#include <linux/smp.h> -#include <linux/module.h> -#include <linux/uaccess.h> -#include <asm/fixmap.h> -#include <asm/kmap_types.h> -#include <asm/tlbflush.h> -#include <hv/hypervisor.h> -#include <arch/chip.h> - - -#if !CHIP_HAS_COHERENT_LOCAL_CACHE() - -/* Defined in memcpy.S */ -extern unsigned long __memcpy_asm(void *to, const void *from, unsigned long n); -extern unsigned long __copy_to_user_inatomic_asm( - void __user *to, const void *from, unsigned long n); -extern unsigned long __copy_from_user_inatomic_asm( - void *to, const void __user *from, unsigned long n); -extern unsigned long __copy_from_user_zeroing_asm( - void *to, const void __user *from, unsigned long n); - -typedef unsigned long (*memcpy_t)(void *, const void *, unsigned long); - -/* Size above which to consider TLB games for performance */ -#define LARGE_COPY_CUTOFF 2048 - -/* Communicate to the simulator what we are trying to do. */ -#define sim_allow_multiple_caching(b) \ - __insn_mtspr(SPR_SIM_CONTROL, \ - SIM_CONTROL_ALLOW_MULTIPLE_CACHING | ((b) << _SIM_CONTROL_OPERATOR_BITS)) - -/* - * Copy memory by briefly enabling incoherent cacheline-at-a-time mode. - * - * We set up our own source and destination PTEs that we fully control. - * This is the only way to guarantee that we don't race with another - * thread that is modifying the PTE; we can't afford to try the - * copy_{to,from}_user() technique of catching the interrupt, since - * we must run with interrupts disabled to avoid the risk of some - * other code seeing the incoherent data in our cache. (Recall that - * our cache is indexed by PA, so even if the other code doesn't use - * our kmap_atomic virtual addresses, they'll still hit in cache using - * the normal VAs that aren't supposed to hit in cache.) - */ -static void memcpy_multicache(void *dest, const void *source, - pte_t dst_pte, pte_t src_pte, int len) -{ - int idx; - unsigned long flags, newsrc, newdst; - pmd_t *pmdp; - pte_t *ptep; - int type0, type1; - int cpu = get_cpu(); - - /* - * Disable interrupts so that we don't recurse into memcpy() - * in an interrupt handler, nor accidentally reference - * the PA of the source from an interrupt routine. Also - * notify the simulator that we're playing games so we don't - * generate spurious coherency warnings. - */ - local_irq_save(flags); - sim_allow_multiple_caching(1); - - /* Set up the new dest mapping */ - type0 = kmap_atomic_idx_push(); - idx = FIX_KMAP_BEGIN + (KM_TYPE_NR * cpu) + type0; - newdst = __fix_to_virt(idx) + ((unsigned long)dest & (PAGE_SIZE-1)); - pmdp = pmd_offset(pud_offset(pgd_offset_k(newdst), newdst), newdst); - ptep = pte_offset_kernel(pmdp, newdst); - if (pte_val(*ptep) != pte_val(dst_pte)) { - set_pte(ptep, dst_pte); - local_flush_tlb_page(NULL, newdst, PAGE_SIZE); - } - - /* Set up the new source mapping */ - type1 = kmap_atomic_idx_push(); - idx += (type0 - type1); - src_pte = hv_pte_set_nc(src_pte); - src_pte = hv_pte_clear_writable(src_pte); /* be paranoid */ - newsrc = __fix_to_virt(idx) + ((unsigned long)source & (PAGE_SIZE-1)); - pmdp = pmd_offset(pud_offset(pgd_offset_k(newsrc), newsrc), newsrc); - ptep = pte_offset_kernel(pmdp, newsrc); - *ptep = src_pte; /* set_pte() would be confused by this */ - local_flush_tlb_page(NULL, newsrc, PAGE_SIZE); - - /* Actually move the data. */ - __memcpy_asm((void *)newdst, (const void *)newsrc, len); - - /* - * Remap the source as locally-cached and not OLOC'ed so that - * we can inval without also invaling the remote cpu's cache. - * This also avoids known errata with inv'ing cacheable oloc data. - */ - src_pte = hv_pte_set_mode(src_pte, HV_PTE_MODE_CACHE_NO_L3); - src_pte = hv_pte_set_writable(src_pte); /* need write access for inv */ - *ptep = src_pte; /* set_pte() would be confused by this */ - local_flush_tlb_page(NULL, newsrc, PAGE_SIZE); - - /* - * Do the actual invalidation, covering the full L2 cache line - * at the end since __memcpy_asm() is somewhat aggressive. - */ - __inv_buffer((void *)newsrc, len); - - /* - * We're done: notify the simulator that all is back to normal, - * and re-enable interrupts and pre-emption. - */ - kmap_atomic_idx_pop(); - kmap_atomic_idx_pop(); - sim_allow_multiple_caching(0); - local_irq_restore(flags); - put_cpu(); -} - -/* - * Identify large copies from remotely-cached memory, and copy them - * via memcpy_multicache() if they look good, otherwise fall back - * to the particular kind of copying passed as the memcpy_t function. - */ -static unsigned long fast_copy(void *dest, const void *source, int len, - memcpy_t func) -{ - /* - * Check if it's big enough to bother with. We may end up doing a - * small copy via TLB manipulation if we're near a page boundary, - * but presumably we'll make it up when we hit the second page. - */ - while (len >= LARGE_COPY_CUTOFF) { - int copy_size, bytes_left_on_page; - pte_t *src_ptep, *dst_ptep; - pte_t src_pte, dst_pte; - struct page *src_page, *dst_page; - - /* Is the source page oloc'ed to a remote cpu? */ -retry_source: - src_ptep = virt_to_pte(current->mm, (unsigned long)source); - if (src_ptep == NULL) - break; - src_pte = *src_ptep; - if (!hv_pte_get_present(src_pte) || - !hv_pte_get_readable(src_pte) || - hv_pte_get_mode(src_pte) != HV_PTE_MODE_CACHE_TILE_L3) - break; - if (get_remote_cache_cpu(src_pte) == smp_processor_id()) - break; - src_page = pfn_to_page(hv_pte_get_pfn(src_pte)); - get_page(src_page); - if (pte_val(src_pte) != pte_val(*src_ptep)) { - put_page(src_page); - goto retry_source; - } - if (pte_huge(src_pte)) { - /* Adjust the PTE to correspond to a small page */ - int pfn = hv_pte_get_pfn(src_pte); - pfn += (((unsigned long)source & (HPAGE_SIZE-1)) - >> PAGE_SHIFT); - src_pte = pfn_pte(pfn, src_pte); - src_pte = pte_mksmall(src_pte); - } - - /* Is the destination page writable? */ -retry_dest: - dst_ptep = virt_to_pte(current->mm, (unsigned long)dest); - if (dst_ptep == NULL) { - put_page(src_page); - break; - } - dst_pte = *dst_ptep; - if (!hv_pte_get_present(dst_pte) || - !hv_pte_get_writable(dst_pte)) { - put_page(src_page); - break; - } - dst_page = pfn_to_page(hv_pte_get_pfn(dst_pte)); - if (dst_page == src_page) { - /* - * Source and dest are on the same page; this - * potentially exposes us to incoherence if any - * part of src and dest overlap on a cache line. - * Just give up rather than trying to be precise. - */ - put_page(src_page); - break; - } - get_page(dst_page); - if (pte_val(dst_pte) != pte_val(*dst_ptep)) { - put_page(dst_page); - goto retry_dest; - } - if (pte_huge(dst_pte)) { - /* Adjust the PTE to correspond to a small page */ - int pfn = hv_pte_get_pfn(dst_pte); - pfn += (((unsigned long)dest & (HPAGE_SIZE-1)) - >> PAGE_SHIFT); - dst_pte = pfn_pte(pfn, dst_pte); - dst_pte = pte_mksmall(dst_pte); - } - - /* All looks good: create a cachable PTE and copy from it */ - copy_size = len; - bytes_left_on_page = - PAGE_SIZE - (((int)source) & (PAGE_SIZE-1)); - if (copy_size > bytes_left_on_page) - copy_size = bytes_left_on_page; - bytes_left_on_page = - PAGE_SIZE - (((int)dest) & (PAGE_SIZE-1)); - if (copy_size > bytes_left_on_page) - copy_size = bytes_left_on_page; - memcpy_multicache(dest, source, dst_pte, src_pte, copy_size); - - /* Release the pages */ - put_page(dst_page); - put_page(src_page); - - /* Continue on the next page */ - dest += copy_size; - source += copy_size; - len -= copy_size; - } - - return func(dest, source, len); -} - -void *memcpy(void *to, const void *from, __kernel_size_t n) -{ - if (n < LARGE_COPY_CUTOFF) - return (void *)__memcpy_asm(to, from, n); - else - return (void *)fast_copy(to, from, n, __memcpy_asm); -} - -unsigned long __copy_to_user_inatomic(void __user *to, const void *from, - unsigned long n) -{ - if (n < LARGE_COPY_CUTOFF) - return __copy_to_user_inatomic_asm(to, from, n); - else - return fast_copy(to, from, n, __copy_to_user_inatomic_asm); -} - -unsigned long __copy_from_user_inatomic(void *to, const void __user *from, - unsigned long n) -{ - if (n < LARGE_COPY_CUTOFF) - return __copy_from_user_inatomic_asm(to, from, n); - else - return fast_copy(to, from, n, __copy_from_user_inatomic_asm); -} - -unsigned long __copy_from_user_zeroing(void *to, const void __user *from, - unsigned long n) -{ - if (n < LARGE_COPY_CUTOFF) - return __copy_from_user_zeroing_asm(to, from, n); - else - return fast_copy(to, from, n, __copy_from_user_zeroing_asm); -} - -#endif /* !CHIP_HAS_COHERENT_LOCAL_CACHE() */ diff --git a/arch/tile/lib/memcpy_user_64.c b/arch/tile/lib/memcpy_user_64.c new file mode 100644 index 00000000000..88c7016492c --- /dev/null +++ b/arch/tile/lib/memcpy_user_64.c @@ -0,0 +1,94 @@ +/* + * Copyright 2011 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + * + * Do memcpy(), but trap and return "n" when a load or store faults. + * + * Note: this idiom only works when memcpy() compiles to a leaf function. + * Here leaf function not only means it does not have calls, but also + * requires no stack operations (sp, stack frame pointer) and no + * use of callee-saved registers, else "jrp lr" will be incorrect since + * unwinding stack frame is bypassed. Since memcpy() is not complex so + * these conditions are satisfied here, but we need to be careful when + * modifying this file. This is not a clean solution but is the best + * one so far. + * + * Also note that we are capturing "n" from the containing scope here. + */ + +#define _ST(p, inst, v) \ + ({ \ + asm("1: " #inst " %0, %1;" \ + ".pushsection .coldtext.memcpy,\"ax\";" \ + "2: { move r0, %2; jrp lr };" \ + ".section __ex_table,\"a\";" \ + ".align 8;" \ + ".quad 1b, 2b;" \ + ".popsection" \ + : "=m" (*(p)) : "r" (v), "r" (n)); \ + }) + +#define _LD(p, inst) \ + ({ \ + unsigned long __v; \ + asm("1: " #inst " %0, %1;" \ + ".pushsection .coldtext.memcpy,\"ax\";" \ + "2: { move r0, %2; jrp lr };" \ + ".section __ex_table,\"a\";" \ + ".align 8;" \ + ".quad 1b, 2b;" \ + ".popsection" \ + : "=r" (__v) : "m" (*(p)), "r" (n)); \ + __v; \ + }) + +#define USERCOPY_FUNC __copy_to_user_inatomic +#define ST1(p, v) _ST((p), st1, (v)) +#define ST2(p, v) _ST((p), st2, (v)) +#define ST4(p, v) _ST((p), st4, (v)) +#define ST8(p, v) _ST((p), st, (v)) +#define LD1 LD +#define LD2 LD +#define LD4 LD +#define LD8 LD +#include "memcpy_64.c" + +#define USERCOPY_FUNC __copy_from_user_inatomic +#define ST1 ST +#define ST2 ST +#define ST4 ST +#define ST8 ST +#define LD1(p) _LD((p), ld1u) +#define LD2(p) _LD((p), ld2u) +#define LD4(p) _LD((p), ld4u) +#define LD8(p) _LD((p), ld) +#include "memcpy_64.c" + +#define USERCOPY_FUNC __copy_in_user_inatomic +#define ST1(p, v) _ST((p), st1, (v)) +#define ST2(p, v) _ST((p), st2, (v)) +#define ST4(p, v) _ST((p), st4, (v)) +#define ST8(p, v) _ST((p), st, (v)) +#define LD1(p) _LD((p), ld1u) +#define LD2(p) _LD((p), ld2u) +#define LD4(p) _LD((p), ld4u) +#define LD8(p) _LD((p), ld) +#include "memcpy_64.c" + +unsigned long __copy_from_user_zeroing(void *to, const void __user *from, + unsigned long n) +{ + unsigned long rc = __copy_from_user_inatomic(to, from, n); + if (unlikely(rc)) + memset(to + n - rc, 0, rc); + return rc; +} diff --git a/arch/tile/lib/memset_32.c b/arch/tile/lib/memset_32.c index 57dbb3a5bff..2042bfe6595 100644 --- a/arch/tile/lib/memset_32.c +++ b/arch/tile/lib/memset_32.c @@ -12,13 +12,10 @@ * more details. */ -#include <arch/chip.h> - #include <linux/types.h> #include <linux/string.h> #include <linux/module.h> - -#undef memset +#include <arch/chip.h> void *memset(void *s, int c, size_t n) { @@ -26,11 +23,7 @@ void *memset(void *s, int c, size_t n) int n32; uint32_t v16, v32; uint8_t *out8 = s; -#if !CHIP_HAS_WH64() - int ahead32; -#else int to_align32; -#endif /* Experimentation shows that a trivial tight loop is a win up until * around a size of 20, where writing a word at a time starts to win. @@ -61,21 +54,6 @@ void *memset(void *s, int c, size_t n) return s; } -#if !CHIP_HAS_WH64() - /* Use a spare issue slot to start prefetching the first cache - * line early. This instruction is free as the store can be buried - * in otherwise idle issue slots doing ALU ops. - */ - __insn_prefetch(out8); - - /* We prefetch the end so that a short memset that spans two cache - * lines gets some prefetching benefit. Again we believe this is free - * to issue. - */ - __insn_prefetch(&out8[n - 1]); -#endif /* !CHIP_HAS_WH64() */ - - /* Align 'out8'. We know n >= 3 so this won't write past the end. */ while (((uintptr_t) out8 & 3) != 0) { *out8++ = c; @@ -96,90 +74,6 @@ void *memset(void *s, int c, size_t n) /* This must be at least 8 or the following loop doesn't work. */ #define CACHE_LINE_SIZE_IN_WORDS (CHIP_L2_LINE_SIZE() / 4) -#if !CHIP_HAS_WH64() - - ahead32 = CACHE_LINE_SIZE_IN_WORDS; - - /* We already prefetched the first and last cache lines, so - * we only need to do more prefetching if we are storing - * to more than two cache lines. - */ - if (n32 > CACHE_LINE_SIZE_IN_WORDS * 2) { - int i; - - /* Prefetch the next several cache lines. - * This is the setup code for the software-pipelined - * loop below. - */ -#define MAX_PREFETCH 5 - ahead32 = n32 & -CACHE_LINE_SIZE_IN_WORDS; - if (ahead32 > MAX_PREFETCH * CACHE_LINE_SIZE_IN_WORDS) - ahead32 = MAX_PREFETCH * CACHE_LINE_SIZE_IN_WORDS; - - for (i = CACHE_LINE_SIZE_IN_WORDS; - i < ahead32; i += CACHE_LINE_SIZE_IN_WORDS) - __insn_prefetch(&out32[i]); - } - - if (n32 > ahead32) { - while (1) { - int j; - - /* Prefetch by reading one word several cache lines - * ahead. Since loads are non-blocking this will - * cause the full cache line to be read while we are - * finishing earlier cache lines. Using a store - * here causes microarchitectural performance - * problems where a victimizing store miss goes to - * the head of the retry FIFO and locks the pipe for - * a few cycles. So a few subsequent stores in this - * loop go into the retry FIFO, and then later - * stores see other stores to the same cache line - * are already in the retry FIFO and themselves go - * into the retry FIFO, filling it up and grinding - * to a halt waiting for the original miss to be - * satisfied. - */ - __insn_prefetch(&out32[ahead32]); - -#if CACHE_LINE_SIZE_IN_WORDS % 4 != 0 -#error "Unhandled CACHE_LINE_SIZE_IN_WORDS" -#endif - - n32 -= CACHE_LINE_SIZE_IN_WORDS; - - /* Save icache space by only partially unrolling - * this loop. - */ - for (j = CACHE_LINE_SIZE_IN_WORDS / 4; j > 0; j--) { - *out32++ = v32; - *out32++ = v32; - *out32++ = v32; - *out32++ = v32; - } - - /* To save compiled code size, reuse this loop even - * when we run out of prefetching to do by dropping - * ahead32 down. - */ - if (n32 <= ahead32) { - /* Not even a full cache line left, - * so stop now. - */ - if (n32 < CACHE_LINE_SIZE_IN_WORDS) - break; - - /* Choose a small enough value that we don't - * prefetch past the end. There's no sense - * in touching cache lines we don't have to. - */ - ahead32 = CACHE_LINE_SIZE_IN_WORDS - 1; - } - } - } - -#else /* CHIP_HAS_WH64() */ - /* Determine how many words we need to emit before the 'out32' * pointer becomes aligned modulo the cache line size. */ @@ -236,8 +130,6 @@ void *memset(void *s, int c, size_t n) n32 &= CACHE_LINE_SIZE_IN_WORDS - 1; } -#endif /* CHIP_HAS_WH64() */ - /* Now handle any leftover values. */ if (n32 != 0) { do { diff --git a/arch/tile/lib/memset_64.c b/arch/tile/lib/memset_64.c new file mode 100644 index 00000000000..03ef69cd73d --- /dev/null +++ b/arch/tile/lib/memset_64.c @@ -0,0 +1,142 @@ +/* + * Copyright 2011 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + */ + +#include <linux/types.h> +#include <linux/string.h> +#include <linux/module.h> +#include <arch/chip.h> +#include "string-endian.h" + +void *memset(void *s, int c, size_t n) +{ + uint64_t *out64; + int n64, to_align64; + uint64_t v64; + uint8_t *out8 = s; + + /* Experimentation shows that a trivial tight loop is a win up until + * around a size of 20, where writing a word at a time starts to win. + */ +#define BYTE_CUTOFF 20 + +#if BYTE_CUTOFF < 7 + /* This must be at least at least this big, or some code later + * on doesn't work. + */ +#error "BYTE_CUTOFF is too small" +#endif + + if (n < BYTE_CUTOFF) { + /* Strangely, this turns out to be the tightest way to + * write this loop. + */ + if (n != 0) { + do { + /* Strangely, combining these into one line + * performs worse. + */ + *out8 = c; + out8++; + } while (--n != 0); + } + + return s; + } + + /* Align 'out8'. We know n >= 7 so this won't write past the end. */ + while (((uintptr_t) out8 & 7) != 0) { + *out8++ = c; + --n; + } + + /* Align 'n'. */ + while (n & 7) + out8[--n] = c; + + out64 = (uint64_t *) out8; + n64 = n >> 3; + + /* Tile input byte out to 64 bits. */ + v64 = copy_byte(c); + + /* This must be at least 8 or the following loop doesn't work. */ +#define CACHE_LINE_SIZE_IN_DOUBLEWORDS (CHIP_L2_LINE_SIZE() / 8) + + /* Determine how many words we need to emit before the 'out32' + * pointer becomes aligned modulo the cache line size. + */ + to_align64 = (-((uintptr_t)out64 >> 3)) & + (CACHE_LINE_SIZE_IN_DOUBLEWORDS - 1); + + /* Only bother aligning and using wh64 if there is at least + * one full cache line to process. This check also prevents + * overrunning the end of the buffer with alignment words. + */ + if (to_align64 <= n64 - CACHE_LINE_SIZE_IN_DOUBLEWORDS) { + int lines_left; + + /* Align out64 mod the cache line size so we can use wh64. */ + n64 -= to_align64; + for (; to_align64 != 0; to_align64--) { + *out64 = v64; + out64++; + } + + /* Use unsigned divide to turn this into a right shift. */ + lines_left = (unsigned)n64 / CACHE_LINE_SIZE_IN_DOUBLEWORDS; + + do { + /* Only wh64 a few lines at a time, so we don't + * exceed the maximum number of victim lines. + */ + int x = ((lines_left < CHIP_MAX_OUTSTANDING_VICTIMS()) + ? lines_left + : CHIP_MAX_OUTSTANDING_VICTIMS()); + uint64_t *wh = out64; + int i = x; + int j; + + lines_left -= x; + + do { + __insn_wh64(wh); + wh += CACHE_LINE_SIZE_IN_DOUBLEWORDS; + } while (--i); + + for (j = x * (CACHE_LINE_SIZE_IN_DOUBLEWORDS / 4); + j != 0; j--) { + *out64++ = v64; + *out64++ = v64; + *out64++ = v64; + *out64++ = v64; + } + } while (lines_left != 0); + + /* We processed all full lines above, so only this many + * words remain to be processed. + */ + n64 &= CACHE_LINE_SIZE_IN_DOUBLEWORDS - 1; + } + + /* Now handle any leftover values. */ + if (n64 != 0) { + do { + *out64 = v64; + out64++; + } while (--n64 != 0); + } + + return s; +} +EXPORT_SYMBOL(memset); diff --git a/arch/tile/lib/spinlock_32.c b/arch/tile/lib/spinlock_32.c index 485e24d62c6..b34f79aada4 100644 --- a/arch/tile/lib/spinlock_32.c +++ b/arch/tile/lib/spinlock_32.c @@ -15,6 +15,7 @@ #include <linux/spinlock.h> #include <linux/module.h> #include <asm/processor.h> +#include <arch/spr_def.h> #include "spinlock_common.h" @@ -91,75 +92,75 @@ EXPORT_SYMBOL(arch_spin_unlock_wait); #define RD_COUNT_MASK ((1 << RD_COUNT_WIDTH) - 1) -/* Lock the word, spinning until there are no tns-ers. */ -static inline u32 get_rwlock(arch_rwlock_t *rwlock) -{ - u32 iterations = 0; - for (;;) { - u32 val = __insn_tns((int *)&rwlock->lock); - if (unlikely(val & 1)) { - delay_backoff(iterations++); - continue; - } - return val; - } -} - -int arch_read_trylock_slow(arch_rwlock_t *rwlock) -{ - u32 val = get_rwlock(rwlock); - int locked = (val << RD_COUNT_WIDTH) == 0; - rwlock->lock = val + (locked << RD_COUNT_SHIFT); - return locked; -} -EXPORT_SYMBOL(arch_read_trylock_slow); - -void arch_read_unlock_slow(arch_rwlock_t *rwlock) -{ - u32 val = get_rwlock(rwlock); - rwlock->lock = val - (1 << RD_COUNT_SHIFT); -} -EXPORT_SYMBOL(arch_read_unlock_slow); - -void arch_write_unlock_slow(arch_rwlock_t *rwlock, u32 val) +/* + * We can get the read lock if everything but the reader bits (which + * are in the high part of the word) is zero, i.e. no active or + * waiting writers, no tns. + * + * We guard the tns/store-back with an interrupt critical section to + * preserve the semantic that the same read lock can be acquired in an + * interrupt context. + */ +int arch_read_trylock(arch_rwlock_t *rwlock) { - u32 eq, mask = 1 << WR_CURR_SHIFT; - while (unlikely(val & 1)) { - /* Limited backoff since we are the highest-priority task. */ - relax(4); - val = __insn_tns((int *)&rwlock->lock); + u32 val; + __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1); + val = __insn_tns((int *)&rwlock->lock); + if (likely((val << _RD_COUNT_WIDTH) == 0)) { + val += 1 << RD_COUNT_SHIFT; + rwlock->lock = val; + __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0); + BUG_ON(val == 0); /* we don't expect wraparound */ + return 1; } - val = __insn_addb(val, mask); - eq = __insn_seqb(val, val << (WR_CURR_SHIFT - WR_NEXT_SHIFT)); - val = __insn_mz(eq & mask, val); - rwlock->lock = val; + if ((val & 1) == 0) + rwlock->lock = val; + __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0); + return 0; } -EXPORT_SYMBOL(arch_write_unlock_slow); +EXPORT_SYMBOL(arch_read_trylock); /* - * We spin until everything but the reader bits (which are in the high - * part of the word) are zero, i.e. no active or waiting writers, no tns. - * + * Spin doing arch_read_trylock() until we acquire the lock. * ISSUE: This approach can permanently starve readers. A reader who sees * a writer could instead take a ticket lock (just like a writer would), * and atomically enter read mode (with 1 reader) when it gets the ticket. - * This way both readers and writers will always make forward progress + * This way both readers and writers would always make forward progress * in a finite time. */ -void arch_read_lock_slow(arch_rwlock_t *rwlock, u32 val) +void arch_read_lock(arch_rwlock_t *rwlock) { u32 iterations = 0; - do { - if (!(val & 1)) - rwlock->lock = val; + while (unlikely(!arch_read_trylock(rwlock))) delay_backoff(iterations++); +} +EXPORT_SYMBOL(arch_read_lock); + +void arch_read_unlock(arch_rwlock_t *rwlock) +{ + u32 val, iterations = 0; + + mb(); /* guarantee anything modified under the lock is visible */ + for (;;) { + __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1); val = __insn_tns((int *)&rwlock->lock); - } while ((val << RD_COUNT_WIDTH) != 0); - rwlock->lock = val + (1 << RD_COUNT_SHIFT); + if (likely((val & 1) == 0)) { + rwlock->lock = val - (1 << _RD_COUNT_SHIFT); + __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0); + break; + } + __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0); + delay_backoff(iterations++); + } } -EXPORT_SYMBOL(arch_read_lock_slow); +EXPORT_SYMBOL(arch_read_unlock); -void arch_write_lock_slow(arch_rwlock_t *rwlock, u32 val) +/* + * We don't need an interrupt critical section here (unlike for + * arch_read_lock) since we should never use a bare write lock where + * it could be interrupted by code that could try to re-acquire it. + */ +void arch_write_lock(arch_rwlock_t *rwlock) { /* * The trailing underscore on this variable (and curr_ below) @@ -167,23 +168,36 @@ void arch_write_lock_slow(arch_rwlock_t *rwlock, u32 val) * when we compare them. */ u32 my_ticket_; + u32 iterations = 0; + u32 val = __insn_tns((int *)&rwlock->lock); - /* Take out the next ticket; this will also stop would-be readers. */ - if (val & 1) - val = get_rwlock(rwlock); - rwlock->lock = __insn_addb(val, 1 << WR_NEXT_SHIFT); - - /* Extract my ticket value from the original word. */ - my_ticket_ = val >> WR_NEXT_SHIFT; + if (likely(val == 0)) { + rwlock->lock = 1 << _WR_NEXT_SHIFT; + return; + } /* - * Wait until the "current" field matches our ticket, and - * there are no remaining readers. + * Wait until there are no readers, then bump up the next + * field and capture the ticket value. */ for (;;) { + if (!(val & 1)) { + if ((val >> RD_COUNT_SHIFT) == 0) + break; + rwlock->lock = val; + } + delay_backoff(iterations++); + val = __insn_tns((int *)&rwlock->lock); + } + + /* Take out the next ticket and extract my ticket value. */ + rwlock->lock = __insn_addb(val, 1 << WR_NEXT_SHIFT); + my_ticket_ = val >> WR_NEXT_SHIFT; + + /* Wait until the "current" field matches our ticket. */ + for (;;) { u32 curr_ = val >> WR_CURR_SHIFT; - u32 readers = val >> RD_COUNT_SHIFT; - u32 delta = ((my_ticket_ - curr_) & WR_MASK) + !!readers; + u32 delta = ((my_ticket_ - curr_) & WR_MASK); if (likely(delta == 0)) break; @@ -199,23 +213,47 @@ void arch_write_lock_slow(arch_rwlock_t *rwlock, u32 val) relax(4); } } -EXPORT_SYMBOL(arch_write_lock_slow); +EXPORT_SYMBOL(arch_write_lock); -int __tns_atomic_acquire(atomic_t *lock) +int arch_write_trylock(arch_rwlock_t *rwlock) { - int ret; - u32 iterations = 0; + u32 val = __insn_tns((int *)&rwlock->lock); - BUG_ON(__insn_mfspr(SPR_INTERRUPT_CRITICAL_SECTION)); - __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1); + /* + * If a tns is in progress, or there's a waiting or active locker, + * or active readers, we can't take the lock, so give up. + */ + if (unlikely(val != 0)) { + if (!(val & 1)) + rwlock->lock = val; + return 0; + } - while ((ret = __insn_tns((void *)&lock->counter)) == 1) - delay_backoff(iterations++); - return ret; + /* Set the "next" field to mark it locked. */ + rwlock->lock = 1 << _WR_NEXT_SHIFT; + return 1; } +EXPORT_SYMBOL(arch_write_trylock); -void __tns_atomic_release(atomic_t *p, int v) +void arch_write_unlock(arch_rwlock_t *rwlock) { - p->counter = v; - __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0); + u32 val, eq, mask; + + mb(); /* guarantee anything modified under the lock is visible */ + val = __insn_tns((int *)&rwlock->lock); + if (likely(val == (1 << _WR_NEXT_SHIFT))) { + rwlock->lock = 0; + return; + } + while (unlikely(val & 1)) { + /* Limited backoff since we are the highest-priority task. */ + relax(4); + val = __insn_tns((int *)&rwlock->lock); + } + mask = 1 << WR_CURR_SHIFT; + val = __insn_addb(val, mask); + eq = __insn_seqb(val, val << (WR_CURR_SHIFT - WR_NEXT_SHIFT)); + val = __insn_mz(eq & mask, val); + rwlock->lock = val; } +EXPORT_SYMBOL(arch_write_unlock); diff --git a/arch/tile/lib/spinlock_64.c b/arch/tile/lib/spinlock_64.c new file mode 100644 index 00000000000..d6fb9581e98 --- /dev/null +++ b/arch/tile/lib/spinlock_64.c @@ -0,0 +1,104 @@ +/* + * Copyright 2011 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + */ + +#include <linux/spinlock.h> +#include <linux/module.h> +#include <asm/processor.h> + +#include "spinlock_common.h" + +/* + * Read the spinlock value without allocating in our cache and without + * causing an invalidation to another cpu with a copy of the cacheline. + * This is important when we are spinning waiting for the lock. + */ +static inline u32 arch_spin_read_noalloc(void *lock) +{ + return atomic_cmpxchg((atomic_t *)lock, -1, -1); +} + +/* + * Wait until the high bits (current) match my ticket. + * If we notice the overflow bit set on entry, we clear it. + */ +void arch_spin_lock_slow(arch_spinlock_t *lock, u32 my_ticket) +{ + if (unlikely(my_ticket & __ARCH_SPIN_NEXT_OVERFLOW)) { + __insn_fetchand4(&lock->lock, ~__ARCH_SPIN_NEXT_OVERFLOW); + my_ticket &= ~__ARCH_SPIN_NEXT_OVERFLOW; + } + + for (;;) { + u32 val = arch_spin_read_noalloc(lock); + u32 delta = my_ticket - arch_spin_current(val); + if (delta == 0) + return; + relax((128 / CYCLES_PER_RELAX_LOOP) * delta); + } +} +EXPORT_SYMBOL(arch_spin_lock_slow); + +/* + * Check the lock to see if it is plausible, and try to get it with cmpxchg(). + */ +int arch_spin_trylock(arch_spinlock_t *lock) +{ + u32 val = arch_spin_read_noalloc(lock); + if (unlikely(arch_spin_current(val) != arch_spin_next(val))) + return 0; + return cmpxchg(&lock->lock, val, (val + 1) & ~__ARCH_SPIN_NEXT_OVERFLOW) + == val; +} +EXPORT_SYMBOL(arch_spin_trylock); + +void arch_spin_unlock_wait(arch_spinlock_t *lock) +{ + u32 iterations = 0; + while (arch_spin_is_locked(lock)) + delay_backoff(iterations++); +} +EXPORT_SYMBOL(arch_spin_unlock_wait); + +/* + * If the read lock fails due to a writer, we retry periodically + * until the value is positive and we write our incremented reader count. + */ +void __read_lock_failed(arch_rwlock_t *rw) +{ + u32 val; + int iterations = 0; + do { + delay_backoff(iterations++); + val = __insn_fetchaddgez4(&rw->lock, 1); + } while (unlikely(arch_write_val_locked(val))); +} +EXPORT_SYMBOL(__read_lock_failed); + +/* + * If we failed because there were readers, clear the "writer" bit + * so we don't block additional readers. Otherwise, there was another + * writer anyway, so our "fetchor" made no difference. Then wait, + * issuing periodic fetchor instructions, till we get the lock. + */ +void __write_lock_failed(arch_rwlock_t *rw, u32 val) +{ + int iterations = 0; + do { + if (!arch_write_val_locked(val)) + val = __insn_fetchand4(&rw->lock, ~__WRITE_LOCK_BIT); + delay_backoff(iterations++); + val = __insn_fetchor4(&rw->lock, __WRITE_LOCK_BIT); + } while (val != 0); +} +EXPORT_SYMBOL(__write_lock_failed); diff --git a/arch/tile/lib/spinlock_common.h b/arch/tile/lib/spinlock_common.h index c1010980913..6ac37509fac 100644 --- a/arch/tile/lib/spinlock_common.h +++ b/arch/tile/lib/spinlock_common.h @@ -60,5 +60,5 @@ static void delay_backoff(int iterations) loops += __insn_crc32_32(stack_pointer, get_cycles_low()) & (loops - 1); - relax(1 << exponent); + relax(loops); } diff --git a/arch/tile/lib/strchr_32.c b/arch/tile/lib/strchr_32.c index c94e6f7ae7b..841fe696301 100644 --- a/arch/tile/lib/strchr_32.c +++ b/arch/tile/lib/strchr_32.c @@ -16,8 +16,6 @@ #include <linux/string.h> #include <linux/module.h> -#undef strchr - char *strchr(const char *s, int c) { int z, g; diff --git a/arch/tile/lib/strchr_64.c b/arch/tile/lib/strchr_64.c new file mode 100644 index 00000000000..fe6e31c06f8 --- /dev/null +++ b/arch/tile/lib/strchr_64.c @@ -0,0 +1,62 @@ +/* + * Copyright 2011 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + */ + +#include <linux/types.h> +#include <linux/string.h> +#include <linux/module.h> +#include "string-endian.h" + +char *strchr(const char *s, int c) +{ + int z, g; + + /* Get an aligned pointer. */ + const uintptr_t s_int = (uintptr_t) s; + const uint64_t *p = (const uint64_t *)(s_int & -8); + + /* Create eight copies of the byte for which we are looking. */ + const uint64_t goal = copy_byte(c); + + /* Read the first aligned word, but force bytes before the string to + * match neither zero nor goal (we make sure the high bit of each + * byte is 1, and the low 7 bits are all the opposite of the goal + * byte). + */ + const uint64_t before_mask = MASK(s_int); + uint64_t v = (*p | before_mask) ^ (goal & __insn_v1shrui(before_mask, 1)); + + uint64_t zero_matches, goal_matches; + while (1) { + /* Look for a terminating '\0'. */ + zero_matches = __insn_v1cmpeqi(v, 0); + + /* Look for the goal byte. */ + goal_matches = __insn_v1cmpeq(v, goal); + + if (__builtin_expect((zero_matches | goal_matches) != 0, 0)) + break; + + v = *++p; + } + + z = CFZ(zero_matches); + g = CFZ(goal_matches); + + /* If we found c before '\0' we got a match. Note that if c == '\0' + * then g == z, and we correctly return the address of the '\0' + * rather than NULL. + */ + return (g <= z) ? ((char *)p) + (g >> 3) : NULL; +} +EXPORT_SYMBOL(strchr); diff --git a/arch/tile/lib/string-endian.h b/arch/tile/lib/string-endian.h new file mode 100644 index 00000000000..2e49cbfe937 --- /dev/null +++ b/arch/tile/lib/string-endian.h @@ -0,0 +1,44 @@ +/* + * Copyright 2013 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + * + * Provide a mask based on the pointer alignment that + * sets up non-zero bytes before the beginning of the string. + * The MASK expression works because shift counts are taken mod 64. + * Also, specify how to count "first" and "last" bits + * when the bits have been read as a word. + */ + +#include <asm/byteorder.h> + +#ifdef __LITTLE_ENDIAN +#define MASK(x) (__insn_shl(1ULL, (x << 3)) - 1) +#define NULMASK(x) ((2ULL << x) - 1) +#define CFZ(x) __insn_ctz(x) +#define REVCZ(x) __insn_clz(x) +#else +#define MASK(x) (__insn_shl(-2LL, ((-x << 3) - 1))) +#define NULMASK(x) (-2LL << (63 - x)) +#define CFZ(x) __insn_clz(x) +#define REVCZ(x) __insn_ctz(x) +#endif + +/* + * Create eight copies of the byte in a uint64_t. Byte Shuffle uses + * the bytes of srcB as the index into the dest vector to select a + * byte. With all indices of zero, the first byte is copied into all + * the other bytes. + */ +static inline uint64_t copy_byte(uint8_t byte) +{ + return __insn_shufflebytes(byte, 0, 0); +} diff --git a/arch/tile/lib/strlen_32.c b/arch/tile/lib/strlen_32.c index 4974292a553..f26f88e11e4 100644 --- a/arch/tile/lib/strlen_32.c +++ b/arch/tile/lib/strlen_32.c @@ -16,8 +16,6 @@ #include <linux/string.h> #include <linux/module.h> -#undef strlen - size_t strlen(const char *s) { /* Get an aligned pointer. */ diff --git a/arch/tile/lib/strlen_64.c b/arch/tile/lib/strlen_64.c new file mode 100644 index 00000000000..9583fc3361f --- /dev/null +++ b/arch/tile/lib/strlen_64.c @@ -0,0 +1,35 @@ +/* + * Copyright 2011 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + */ + +#include <linux/types.h> +#include <linux/string.h> +#include <linux/module.h> +#include "string-endian.h" + +size_t strlen(const char *s) +{ + /* Get an aligned pointer. */ + const uintptr_t s_int = (uintptr_t) s; + const uint64_t *p = (const uint64_t *)(s_int & -8); + + /* Read and MASK the first word. */ + uint64_t v = *p | MASK(s_int); + + uint64_t bits; + while ((bits = __insn_v1cmpeqi(v, 0)) == 0) + v = *++p; + + return ((const char *)p) + (CFZ(bits) >> 3) - s; +} +EXPORT_SYMBOL(strlen); diff --git a/arch/tile/lib/strnlen_32.c b/arch/tile/lib/strnlen_32.c new file mode 100644 index 00000000000..1434141d9e0 --- /dev/null +++ b/arch/tile/lib/strnlen_32.c @@ -0,0 +1,47 @@ +/* + * Copyright 2013 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + */ + +#include <linux/types.h> +#include <linux/string.h> +#include <linux/module.h> + +size_t strnlen(const char *s, size_t count) +{ + /* Get an aligned pointer. */ + const uintptr_t s_int = (uintptr_t) s; + const uint32_t *p = (const uint32_t *)(s_int & -4); + size_t bytes_read = sizeof(*p) - (s_int & (sizeof(*p) - 1)); + size_t len; + uint32_t v, bits; + + /* Avoid page fault risk by not reading any bytes when count is 0. */ + if (count == 0) + return 0; + + /* Read first word, but force bytes before the string to be nonzero. */ + v = *p | ((1 << ((s_int << 3) & 31)) - 1); + + while ((bits = __insn_seqb(v, 0)) == 0) { + if (bytes_read >= count) { + /* Read COUNT bytes and didn't find the terminator. */ + return count; + } + v = *++p; + bytes_read += sizeof(v); + } + + len = ((const char *) p) + (__insn_ctz(bits) >> 3) - s; + return (len < count ? len : count); +} +EXPORT_SYMBOL(strnlen); diff --git a/arch/tile/lib/strnlen_64.c b/arch/tile/lib/strnlen_64.c new file mode 100644 index 00000000000..2e8de6a5136 --- /dev/null +++ b/arch/tile/lib/strnlen_64.c @@ -0,0 +1,48 @@ +/* + * Copyright 2013 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + */ + +#include <linux/types.h> +#include <linux/string.h> +#include <linux/module.h> +#include "string-endian.h" + +size_t strnlen(const char *s, size_t count) +{ + /* Get an aligned pointer. */ + const uintptr_t s_int = (uintptr_t) s; + const uint64_t *p = (const uint64_t *)(s_int & -8); + size_t bytes_read = sizeof(*p) - (s_int & (sizeof(*p) - 1)); + size_t len; + uint64_t v, bits; + + /* Avoid page fault risk by not reading any bytes when count is 0. */ + if (count == 0) + return 0; + + /* Read and MASK the first word. */ + v = *p | MASK(s_int); + + while ((bits = __insn_v1cmpeqi(v, 0)) == 0) { + if (bytes_read >= count) { + /* Read COUNT bytes and didn't find the terminator. */ + return count; + } + v = *++p; + bytes_read += sizeof(v); + } + + len = ((const char *) p) + (CFZ(bits) >> 3) - s; + return (len < count ? len : count); +} +EXPORT_SYMBOL(strnlen); diff --git a/arch/tile/lib/uaccess.c b/arch/tile/lib/uaccess.c index f8d398c9ee7..030abe3ee4f 100644 --- a/arch/tile/lib/uaccess.c +++ b/arch/tile/lib/uaccess.c @@ -22,11 +22,3 @@ int __range_ok(unsigned long addr, unsigned long size) is_arch_mappable_range(addr, size)); } EXPORT_SYMBOL(__range_ok); - -#ifdef CONFIG_DEBUG_COPY_FROM_USER -void copy_from_user_overflow(void) -{ - WARN(1, "Buffer overflow detected!\n"); -} -EXPORT_SYMBOL(copy_from_user_overflow); -#endif diff --git a/arch/tile/lib/usercopy_32.S b/arch/tile/lib/usercopy_32.S index 979f76d8374..1bc16222463 100644 --- a/arch/tile/lib/usercopy_32.S +++ b/arch/tile/lib/usercopy_32.S @@ -19,82 +19,6 @@ /* Access user memory, but use MMU to avoid propagating kernel exceptions. */ - .pushsection .fixup,"ax" - -get_user_fault: - { move r0, zero; move r1, zero } - { movei r2, -EFAULT; jrp lr } - ENDPROC(get_user_fault) - -put_user_fault: - { movei r0, -EFAULT; jrp lr } - ENDPROC(put_user_fault) - - .popsection - -/* - * __get_user_N functions take a pointer in r0, and return 0 in r2 - * on success, with the value in r0; or else -EFAULT in r2. - */ -#define __get_user_N(bytes, LOAD) \ - STD_ENTRY(__get_user_##bytes); \ -1: { LOAD r0, r0; move r1, zero; move r2, zero }; \ - jrp lr; \ - STD_ENDPROC(__get_user_##bytes); \ - .pushsection __ex_table,"a"; \ - .word 1b, get_user_fault; \ - .popsection - -__get_user_N(1, lb_u) -__get_user_N(2, lh_u) -__get_user_N(4, lw) - -/* - * __get_user_8 takes a pointer in r0, and returns 0 in r2 - * on success, with the value in r0/r1; or else -EFAULT in r2. - */ - STD_ENTRY(__get_user_8); -1: { lw r0, r0; addi r1, r0, 4 }; -2: { lw r1, r1; move r2, zero }; - jrp lr; - STD_ENDPROC(__get_user_8); - .pushsection __ex_table,"a"; - .word 1b, get_user_fault; - .word 2b, get_user_fault; - .popsection - -/* - * __put_user_N functions take a value in r0 and a pointer in r1, - * and return 0 in r0 on success or -EFAULT on failure. - */ -#define __put_user_N(bytes, STORE) \ - STD_ENTRY(__put_user_##bytes); \ -1: { STORE r1, r0; move r0, zero }; \ - jrp lr; \ - STD_ENDPROC(__put_user_##bytes); \ - .pushsection __ex_table,"a"; \ - .word 1b, put_user_fault; \ - .popsection - -__put_user_N(1, sb) -__put_user_N(2, sh) -__put_user_N(4, sw) - -/* - * __put_user_8 takes a value in r0/r1 and a pointer in r2, - * and returns 0 in r0 on success or -EFAULT on failure. - */ -STD_ENTRY(__put_user_8) -1: { sw r2, r0; addi r2, r2, 4 } -2: { sw r2, r1; move r0, zero } - jrp lr - STD_ENDPROC(__put_user_8) - .pushsection __ex_table,"a" - .word 1b, put_user_fault - .word 2b, put_user_fault - .popsection - - /* * strnlen_user_asm takes the pointer in r0, and the length bound in r1. * It returns the length, including the terminating NUL, or zero on exception. @@ -112,6 +36,7 @@ strnlen_user_fault: { move r0, zero; jrp lr } ENDPROC(strnlen_user_fault) .section __ex_table,"a" + .align 4 .word 1b, strnlen_user_fault .popsection @@ -123,18 +48,20 @@ strnlen_user_fault: */ STD_ENTRY(strncpy_from_user_asm) { bz r2, 2f; move r3, r0 } -1: { lb_u r4, r1; addi r1, r1, 1; addi r2, r2, -1 } +1: { lb_u r4, r1; addi r1, r1, 1; addi r2, r2, -1 } { sb r0, r4; addi r0, r0, 1 } - bz r2, 2f - bnzt r4, 1b - addi r0, r0, -1 /* don't count the trailing NUL */ -2: { sub r0, r0, r3; jrp lr } + bz r4, 2f + bnzt r2, 1b + { sub r0, r0, r3; jrp lr } +2: addi r0, r0, -1 /* don't count the trailing NUL */ + { sub r0, r0, r3; jrp lr } STD_ENDPROC(strncpy_from_user_asm) .pushsection .fixup,"ax" strncpy_from_user_fault: { movei r0, -EFAULT; jrp lr } ENDPROC(strncpy_from_user_fault) .section __ex_table,"a" + .align 4 .word 1b, strncpy_from_user_fault .popsection @@ -153,6 +80,7 @@ STD_ENTRY(clear_user_asm) bnzt r1, 1b 2: { move r0, r1; jrp lr } .pushsection __ex_table,"a" + .align 4 .word 1b, 2b .popsection @@ -162,6 +90,7 @@ STD_ENTRY(clear_user_asm) 2: { move r0, r1; jrp lr } STD_ENDPROC(clear_user_asm) .pushsection __ex_table,"a" + .align 4 .word 1b, 2b .popsection @@ -181,25 +110,7 @@ STD_ENTRY(flush_user_asm) 2: { move r0, r1; jrp lr } STD_ENDPROC(flush_user_asm) .pushsection __ex_table,"a" - .word 1b, 2b - .popsection - -/* - * inv_user_asm takes the user target address in r0 and the - * number of bytes to invalidate in r1. - * It returns the number of not inv'able bytes (hopefully zero) in r0. - */ -STD_ENTRY(inv_user_asm) - bz r1, 2f - { movei r2, L2_CACHE_BYTES; add r1, r0, r1 } - { sub r2, zero, r2; addi r1, r1, L2_CACHE_BYTES-1 } - { and r0, r0, r2; and r1, r1, r2 } - { sub r1, r1, r0 } -1: { inv r0; addi r1, r1, -CHIP_INV_STRIDE() } - { addi r0, r0, CHIP_INV_STRIDE(); bnzt r1, 1b } -2: { move r0, r1; jrp lr } - STD_ENDPROC(inv_user_asm) - .pushsection __ex_table,"a" + .align 4 .word 1b, 2b .popsection @@ -219,5 +130,6 @@ STD_ENTRY(finv_user_asm) 2: { move r0, r1; jrp lr } STD_ENDPROC(finv_user_asm) .pushsection __ex_table,"a" + .align 4 .word 1b, 2b .popsection diff --git a/arch/tile/lib/usercopy_64.S b/arch/tile/lib/usercopy_64.S new file mode 100644 index 00000000000..b3b31a3306f --- /dev/null +++ b/arch/tile/lib/usercopy_64.S @@ -0,0 +1,135 @@ +/* + * Copyright 2011 Tilera Corporation. All Rights Reserved. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation, version 2. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or + * NON INFRINGEMENT. See the GNU General Public License for + * more details. + */ + +#include <linux/linkage.h> +#include <asm/errno.h> +#include <asm/cache.h> +#include <arch/chip.h> + +/* Access user memory, but use MMU to avoid propagating kernel exceptions. */ + +/* + * strnlen_user_asm takes the pointer in r0, and the length bound in r1. + * It returns the length, including the terminating NUL, or zero on exception. + * If length is greater than the bound, returns one plus the bound. + */ +STD_ENTRY(strnlen_user_asm) + { beqz r1, 2f; addi r3, r0, -1 } /* bias down to include NUL */ +1: { ld1u r4, r0; addi r1, r1, -1 } + beqz r4, 2f + { bnezt r1, 1b; addi r0, r0, 1 } +2: { sub r0, r0, r3; jrp lr } + STD_ENDPROC(strnlen_user_asm) + .pushsection .fixup,"ax" +strnlen_user_fault: + { move r0, zero; jrp lr } + ENDPROC(strnlen_user_fault) + .section __ex_table,"a" + .align 8 + .quad 1b, strnlen_user_fault + .popsection + +/* + * strncpy_from_user_asm takes the kernel target pointer in r0, + * the userspace source pointer in r1, and the length bound (including + * the trailing NUL) in r2. On success, it returns the string length + * (not including the trailing NUL), or -EFAULT on failure. + */ +STD_ENTRY(strncpy_from_user_asm) + { beqz r2, 2f; move r3, r0 } +1: { ld1u r4, r1; addi r1, r1, 1; addi r2, r2, -1 } + { st1 r0, r4; addi r0, r0, 1 } + beqz r4, 2f + bnezt r2, 1b + { sub r0, r0, r3; jrp lr } +2: addi r0, r0, -1 /* don't count the trailing NUL */ + { sub r0, r0, r3; jrp lr } + STD_ENDPROC(strncpy_from_user_asm) + .pushsection .fixup,"ax" +strncpy_from_user_fault: + { movei r0, -EFAULT; jrp lr } + ENDPROC(strncpy_from_user_fault) + .section __ex_table,"a" + .align 8 + .quad 1b, strncpy_from_user_fault + .popsection + +/* + * clear_user_asm takes the user target address in r0 and the + * number of bytes to zero in r1. + * It returns the number of uncopiable bytes (hopefully zero) in r0. + * Note that we don't use a separate .fixup section here since we fall + * through into the "fixup" code as the last straight-line bundle anyway. + */ +STD_ENTRY(clear_user_asm) + { beqz r1, 2f; or r2, r0, r1 } + andi r2, r2, 7 + beqzt r2, .Lclear_aligned_user_asm +1: { st1 r0, zero; addi r0, r0, 1; addi r1, r1, -1 } + bnezt r1, 1b +2: { move r0, r1; jrp lr } + .pushsection __ex_table,"a" + .align 8 + .quad 1b, 2b + .popsection + +.Lclear_aligned_user_asm: +1: { st r0, zero; addi r0, r0, 8; addi r1, r1, -8 } + bnezt r1, 1b +2: { move r0, r1; jrp lr } + STD_ENDPROC(clear_user_asm) + .pushsection __ex_table,"a" + .align 8 + .quad 1b, 2b + .popsection + +/* + * flush_user_asm takes the user target address in r0 and the + * number of bytes to flush in r1. + * It returns the number of unflushable bytes (hopefully zero) in r0. + */ +STD_ENTRY(flush_user_asm) + beqz r1, 2f + { movei r2, L2_CACHE_BYTES; add r1, r0, r1 } + { sub r2, zero, r2; addi r1, r1, L2_CACHE_BYTES-1 } + { and r0, r0, r2; and r1, r1, r2 } + { sub r1, r1, r0 } +1: { flush r0; addi r1, r1, -CHIP_FLUSH_STRIDE() } + { addi r0, r0, CHIP_FLUSH_STRIDE(); bnezt r1, 1b } +2: { move r0, r1; jrp lr } + STD_ENDPROC(flush_user_asm) + .pushsection __ex_table,"a" + .align 8 + .quad 1b, 2b + .popsection + +/* + * finv_user_asm takes the user target address in r0 and the + * number of bytes to flush-invalidate in r1. + * It returns the number of not finv'able bytes (hopefully zero) in r0. + */ +STD_ENTRY(finv_user_asm) + beqz r1, 2f + { movei r2, L2_CACHE_BYTES; add r1, r0, r1 } + { sub r2, zero, r2; addi r1, r1, L2_CACHE_BYTES-1 } + { and r0, r0, r2; and r1, r1, r2 } + { sub r1, r1, r0 } +1: { finv r0; addi r1, r1, -CHIP_FINV_STRIDE() } + { addi r0, r0, CHIP_FINV_STRIDE(); bnezt r1, 1b } +2: { move r0, r1; jrp lr } + STD_ENDPROC(finv_user_asm) + .pushsection __ex_table,"a" + .align 8 + .quad 1b, 2b + .popsection |