From 5f655322b1ba4bd46e26e307d04098f9c84df764 Mon Sep 17 00:00:00 2001 From: Mikulas Patocka Date: Tue, 14 Mar 2017 11:47:29 -0400 Subject: parisc: support R_PARISC_SECREL32 relocation in modules The parisc kernel doesn't work with CONFIG_MODVERSIONS since the commit 71810db27c1c853b335675bee335d893bc3d324b. It can't load modules with the error: "module unix: Unknown relocation: 41". The commit changes __kcrctab from 64-bit valus to 32-bit values. The assembler generates R_PARISC_SECREL32 secrel relocation for them and the module loader doesn't support this relocation. This patch adds the R_PARISC_SECREL32 relocation to the module loader. Signed-off-by: Mikulas Patocka Cc: stable@vger.kernel.org # v4.10+ Signed-off-by: Helge Deller --- arch/parisc/kernel/module.c | 8 ++++++++ 1 file changed, 8 insertions(+) (limited to 'arch/parisc') diff --git a/arch/parisc/kernel/module.c b/arch/parisc/kernel/module.c index a0ecdb4abcc8..c66c943d9322 100644 --- a/arch/parisc/kernel/module.c +++ b/arch/parisc/kernel/module.c @@ -620,6 +620,10 @@ int apply_relocate_add(Elf_Shdr *sechdrs, */ *loc = fsel(val, addend); break; + case R_PARISC_SECREL32: + /* 32-bit section relative address. */ + *loc = fsel(val, addend); + break; case R_PARISC_DPREL21L: /* left 21 bit of relative address */ val = lrsel(val - dp, addend); @@ -807,6 +811,10 @@ int apply_relocate_add(Elf_Shdr *sechdrs, */ *loc = fsel(val, addend); break; + case R_PARISC_SECREL32: + /* 32-bit section relative address. */ + *loc = fsel(val, addend); + break; case R_PARISC_FPTR64: /* 64-bit function address */ if(in_local(me, (void *)(val + addend))) { -- cgit v1.2.3 From 316ec0624f951166daedbe446988ef92ae72b59f Mon Sep 17 00:00:00 2001 From: John David Anglin Date: Sat, 11 Mar 2017 18:03:34 -0500 Subject: parisc: Optimize flush_kernel_vmap_range and invalidate_kernel_vmap_range The previously submitted patch did not resolve the random segmentation faults observed on the phantom buildd system. There are still unresolved problems with the Debian 4.8 and 4.9 kernels on C8000. The attached patch removes the flush of the offset map pages and does a whole data cache flush for large ranges. No other arch flushes the offset map in these routines as far as I can tell. I have not observed any random segmentation faults on rp3440 in two weeks of testing with 4.10.0 and 4.10.1. Signed-off-by: John David Anglin Cc: stable@vger.kernel.org # v4.8+ Signed-off-by: Helge Deller --- arch/parisc/include/asm/cacheflush.h | 23 ++--------------------- arch/parisc/kernel/cache.c | 22 ++++++++++++++++++++++ 2 files changed, 24 insertions(+), 21 deletions(-) (limited to 'arch/parisc') diff --git a/arch/parisc/include/asm/cacheflush.h b/arch/parisc/include/asm/cacheflush.h index 19c9c3c5f267..c7e15cc5c668 100644 --- a/arch/parisc/include/asm/cacheflush.h +++ b/arch/parisc/include/asm/cacheflush.h @@ -43,28 +43,9 @@ static inline void flush_kernel_dcache_page(struct page *page) #define flush_kernel_dcache_range(start,size) \ flush_kernel_dcache_range_asm((start), (start)+(size)); -/* vmap range flushes and invalidates. Architecturally, we don't need - * the invalidate, because the CPU should refuse to speculate once an - * area has been flushed, so invalidate is left empty */ -static inline void flush_kernel_vmap_range(void *vaddr, int size) -{ - unsigned long start = (unsigned long)vaddr; - - flush_kernel_dcache_range_asm(start, start + size); -} -static inline void invalidate_kernel_vmap_range(void *vaddr, int size) -{ - unsigned long start = (unsigned long)vaddr; - void *cursor = vaddr; - for ( ; cursor < vaddr + size; cursor += PAGE_SIZE) { - struct page *page = vmalloc_to_page(cursor); - - if (test_and_clear_bit(PG_dcache_dirty, &page->flags)) - flush_kernel_dcache_page(page); - } - flush_kernel_dcache_range_asm(start, start + size); -} +void flush_kernel_vmap_range(void *vaddr, int size); +void invalidate_kernel_vmap_range(void *vaddr, int size); #define flush_cache_vmap(start, end) flush_cache_all() #define flush_cache_vunmap(start, end) flush_cache_all() diff --git a/arch/parisc/kernel/cache.c b/arch/parisc/kernel/cache.c index 0dc72d5de861..c32a09095216 100644 --- a/arch/parisc/kernel/cache.c +++ b/arch/parisc/kernel/cache.c @@ -616,3 +616,25 @@ flush_cache_page(struct vm_area_struct *vma, unsigned long vmaddr, unsigned long __flush_cache_page(vma, vmaddr, PFN_PHYS(pfn)); } } + +void flush_kernel_vmap_range(void *vaddr, int size) +{ + unsigned long start = (unsigned long)vaddr; + + if ((unsigned long)size > parisc_cache_flush_threshold) + flush_data_cache(); + else + flush_kernel_dcache_range_asm(start, start + size); +} +EXPORT_SYMBOL(flush_kernel_vmap_range); + +void invalidate_kernel_vmap_range(void *vaddr, int size) +{ + unsigned long start = (unsigned long)vaddr; + + if ((unsigned long)size > parisc_cache_flush_threshold) + flush_data_cache(); + else + flush_kernel_dcache_range_asm(start, start + size); +} +EXPORT_SYMBOL(invalidate_kernel_vmap_range); -- cgit v1.2.3 From 63d32d1e09cb2fc65b084b261976c06b40d19115 Mon Sep 17 00:00:00 2001 From: Helge Deller Date: Wed, 15 Mar 2017 21:10:17 +0100 Subject: parisc: Wire up statx system call Signed-off-by: Helge Deller --- arch/parisc/include/uapi/asm/unistd.h | 3 ++- arch/parisc/kernel/syscall_table.S | 1 + 2 files changed, 3 insertions(+), 1 deletion(-) (limited to 'arch/parisc') diff --git a/arch/parisc/include/uapi/asm/unistd.h b/arch/parisc/include/uapi/asm/unistd.h index 6b0741e7a7ed..667c99421003 100644 --- a/arch/parisc/include/uapi/asm/unistd.h +++ b/arch/parisc/include/uapi/asm/unistd.h @@ -362,8 +362,9 @@ #define __NR_copy_file_range (__NR_Linux + 346) #define __NR_preadv2 (__NR_Linux + 347) #define __NR_pwritev2 (__NR_Linux + 348) +#define __NR_statx (__NR_Linux + 349) -#define __NR_Linux_syscalls (__NR_pwritev2 + 1) +#define __NR_Linux_syscalls (__NR_statx + 1) #define __IGNORE_select /* newselect */ diff --git a/arch/parisc/kernel/syscall_table.S b/arch/parisc/kernel/syscall_table.S index 3cfef1de8061..44aeaa9c039f 100644 --- a/arch/parisc/kernel/syscall_table.S +++ b/arch/parisc/kernel/syscall_table.S @@ -444,6 +444,7 @@ ENTRY_SAME(copy_file_range) ENTRY_COMP(preadv2) ENTRY_COMP(pwritev2) + ENTRY_SAME(statx) .ifne (. - 90b) - (__NR_Linux_syscalls * (91b - 90b)) -- cgit v1.2.3 From 186ecf14e58befba434f0774eea89e35f64d3c6a Mon Sep 17 00:00:00 2001 From: Helge Deller Date: Wed, 15 Mar 2017 21:48:42 +0100 Subject: parisc: Avoid compiler warnings with access_ok() Commit 09b871ffd4d8 (parisc: Define access_ok() as macro) missed to mark uaddr as used, which then gives compiler warnings about unused variables. Fix it by comparing uaddr to uaddr which then gets optimized away by the compiler. Signed-off-by: Helge Deller Fixes: 09b871ffd4d8 ("parisc: Define access_ok() as macro") --- arch/parisc/include/asm/uaccess.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'arch/parisc') diff --git a/arch/parisc/include/asm/uaccess.h b/arch/parisc/include/asm/uaccess.h index fb4382c28259..edfbf9d6a6dd 100644 --- a/arch/parisc/include/asm/uaccess.h +++ b/arch/parisc/include/asm/uaccess.h @@ -32,7 +32,8 @@ * that put_user is the same as __put_user, etc. */ -#define access_ok(type, uaddr, size) (1) +#define access_ok(type, uaddr, size) \ + ( (uaddr) == (uaddr) ) #define put_user __put_user #define get_user __get_user -- cgit v1.2.3 From 74e3f6e63da6c8e8246fba1689e040bc926b4a1a Mon Sep 17 00:00:00 2001 From: Arvind Yadav Date: Tue, 14 Mar 2017 15:24:51 +0530 Subject: parisc: perf: Fix potential NULL pointer dereference Fix potential NULL pointer dereference and clean up coding style errors (code indent, trailing whitespaces). Signed-off-by: Arvind Yadav Signed-off-by: Helge Deller --- arch/parisc/kernel/perf.c | 94 ++++++++++++++++++++++++----------------------- 1 file changed, 49 insertions(+), 45 deletions(-) (limited to 'arch/parisc') diff --git a/arch/parisc/kernel/perf.c b/arch/parisc/kernel/perf.c index e282a5131d77..6017a5af2e6e 100644 --- a/arch/parisc/kernel/perf.c +++ b/arch/parisc/kernel/perf.c @@ -39,7 +39,7 @@ * the PDC INTRIGUE calls. This is done to eliminate bugs introduced * in various PDC revisions. The code is much more maintainable * and reliable this way vs having to debug on every version of PDC - * on every box. + * on every box. */ #include @@ -195,8 +195,8 @@ static int perf_config(uint32_t *image_ptr); static int perf_release(struct inode *inode, struct file *file); static int perf_open(struct inode *inode, struct file *file); static ssize_t perf_read(struct file *file, char __user *buf, size_t cnt, loff_t *ppos); -static ssize_t perf_write(struct file *file, const char __user *buf, size_t count, - loff_t *ppos); +static ssize_t perf_write(struct file *file, const char __user *buf, + size_t count, loff_t *ppos); static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg); static void perf_start_counters(void); static int perf_stop_counters(uint32_t *raddr); @@ -222,7 +222,7 @@ extern void perf_intrigue_disable_perf_counters (void); /* * configure: * - * Configure the cpu with a given data image. First turn off the counters, + * Configure the cpu with a given data image. First turn off the counters, * then download the image, then turn the counters back on. */ static int perf_config(uint32_t *image_ptr) @@ -234,7 +234,7 @@ static int perf_config(uint32_t *image_ptr) error = perf_stop_counters(raddr); if (error != 0) { printk("perf_config: perf_stop_counters = %ld\n", error); - return -EINVAL; + return -EINVAL; } printk("Preparing to write image\n"); @@ -242,7 +242,7 @@ printk("Preparing to write image\n"); error = perf_write_image((uint64_t *)image_ptr); if (error != 0) { printk("perf_config: DOWNLOAD = %ld\n", error); - return -EINVAL; + return -EINVAL; } printk("Preparing to start counters\n"); @@ -254,7 +254,7 @@ printk("Preparing to start counters\n"); } /* - * Open the device and initialize all of its memory. The device is only + * Open the device and initialize all of its memory. The device is only * opened once, but can be "queried" by multiple processes that know its * file descriptor. */ @@ -298,19 +298,19 @@ static ssize_t perf_read(struct file *file, char __user *buf, size_t cnt, loff_t * called on the processor that the download should happen * on. */ -static ssize_t perf_write(struct file *file, const char __user *buf, size_t count, - loff_t *ppos) +static ssize_t perf_write(struct file *file, const char __user *buf, + size_t count, loff_t *ppos) { size_t image_size; uint32_t image_type; uint32_t interface_type; uint32_t test; - if (perf_processor_interface == ONYX_INTF) + if (perf_processor_interface == ONYX_INTF) image_size = PCXU_IMAGE_SIZE; - else if (perf_processor_interface == CUDA_INTF) + else if (perf_processor_interface == CUDA_INTF) image_size = PCXW_IMAGE_SIZE; - else + else return -EFAULT; if (!capable(CAP_SYS_ADMIN)) @@ -330,22 +330,22 @@ static ssize_t perf_write(struct file *file, const char __user *buf, size_t coun /* First check the machine type is correct for the requested image */ - if (((perf_processor_interface == CUDA_INTF) && - (interface_type != CUDA_INTF)) || - ((perf_processor_interface == ONYX_INTF) && - (interface_type != ONYX_INTF))) + if (((perf_processor_interface == CUDA_INTF) && + (interface_type != CUDA_INTF)) || + ((perf_processor_interface == ONYX_INTF) && + (interface_type != ONYX_INTF))) return -EINVAL; /* Next check to make sure the requested image is valid */ - if (((interface_type == CUDA_INTF) && + if (((interface_type == CUDA_INTF) && (test >= MAX_CUDA_IMAGES)) || - ((interface_type == ONYX_INTF) && - (test >= MAX_ONYX_IMAGES))) + ((interface_type == ONYX_INTF) && + (test >= MAX_ONYX_IMAGES))) return -EINVAL; /* Copy the image into the processor */ - if (interface_type == CUDA_INTF) + if (interface_type == CUDA_INTF) return perf_config(cuda_images[test]); else return perf_config(onyx_images[test]); @@ -359,7 +359,7 @@ static ssize_t perf_write(struct file *file, const char __user *buf, size_t coun static void perf_patch_images(void) { #if 0 /* FIXME!! */ -/* +/* * NOTE: this routine is VERY specific to the current TLB image. * If the image is changed, this routine might also need to be changed. */ @@ -367,9 +367,9 @@ static void perf_patch_images(void) extern void $i_dtlb_miss_2_0(); extern void PA2_0_iva(); - /* + /* * We can only use the lower 32-bits, the upper 32-bits should be 0 - * anyway given this is in the kernel + * anyway given this is in the kernel */ uint32_t itlb_addr = (uint32_t)&($i_itlb_miss_2_0); uint32_t dtlb_addr = (uint32_t)&($i_dtlb_miss_2_0); @@ -377,21 +377,21 @@ static void perf_patch_images(void) if (perf_processor_interface == ONYX_INTF) { /* clear last 2 bytes */ - onyx_images[TLBMISS][15] &= 0xffffff00; + onyx_images[TLBMISS][15] &= 0xffffff00; /* set 2 bytes */ onyx_images[TLBMISS][15] |= (0x000000ff&((dtlb_addr) >> 24)); onyx_images[TLBMISS][16] = (dtlb_addr << 8)&0xffffff00; onyx_images[TLBMISS][17] = itlb_addr; /* clear last 2 bytes */ - onyx_images[TLBHANDMISS][15] &= 0xffffff00; + onyx_images[TLBHANDMISS][15] &= 0xffffff00; /* set 2 bytes */ onyx_images[TLBHANDMISS][15] |= (0x000000ff&((dtlb_addr) >> 24)); onyx_images[TLBHANDMISS][16] = (dtlb_addr << 8)&0xffffff00; onyx_images[TLBHANDMISS][17] = itlb_addr; /* clear last 2 bytes */ - onyx_images[BIG_CPI][15] &= 0xffffff00; + onyx_images[BIG_CPI][15] &= 0xffffff00; /* set 2 bytes */ onyx_images[BIG_CPI][15] |= (0x000000ff&((dtlb_addr) >> 24)); onyx_images[BIG_CPI][16] = (dtlb_addr << 8)&0xffffff00; @@ -404,24 +404,24 @@ static void perf_patch_images(void) } else if (perf_processor_interface == CUDA_INTF) { /* Cuda interface */ - cuda_images[TLBMISS][16] = + cuda_images[TLBMISS][16] = (cuda_images[TLBMISS][16]&0xffff0000) | ((dtlb_addr >> 8)&0x0000ffff); - cuda_images[TLBMISS][17] = + cuda_images[TLBMISS][17] = ((dtlb_addr << 24)&0xff000000) | ((itlb_addr >> 16)&0x000000ff); cuda_images[TLBMISS][18] = (itlb_addr << 16)&0xffff0000; - cuda_images[TLBHANDMISS][16] = + cuda_images[TLBHANDMISS][16] = (cuda_images[TLBHANDMISS][16]&0xffff0000) | ((dtlb_addr >> 8)&0x0000ffff); - cuda_images[TLBHANDMISS][17] = + cuda_images[TLBHANDMISS][17] = ((dtlb_addr << 24)&0xff000000) | ((itlb_addr >> 16)&0x000000ff); cuda_images[TLBHANDMISS][18] = (itlb_addr << 16)&0xffff0000; - cuda_images[BIG_CPI][16] = + cuda_images[BIG_CPI][16] = (cuda_images[BIG_CPI][16]&0xffff0000) | ((dtlb_addr >> 8)&0x0000ffff); - cuda_images[BIG_CPI][17] = + cuda_images[BIG_CPI][17] = ((dtlb_addr << 24)&0xff000000) | ((itlb_addr >> 16)&0x000000ff); cuda_images[BIG_CPI][18] = (itlb_addr << 16)&0xffff0000; } else { @@ -433,7 +433,7 @@ static void perf_patch_images(void) /* * ioctl routine - * All routines effect the processor that they are executed on. Thus you + * All routines effect the processor that they are executed on. Thus you * must be running on the processor that you wish to change. */ @@ -459,7 +459,7 @@ static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) } /* copy out the Counters */ - if (copy_to_user((void __user *)arg, raddr, + if (copy_to_user((void __user *)arg, raddr, sizeof (raddr)) != 0) { error = -EFAULT; break; @@ -487,7 +487,7 @@ static const struct file_operations perf_fops = { .open = perf_open, .release = perf_release }; - + static struct miscdevice perf_dev = { MISC_DYNAMIC_MINOR, PA_PERF_DEV, @@ -595,7 +595,7 @@ static int perf_stop_counters(uint32_t *raddr) /* OR sticky2 (bit 1496) to counter2 bit 32 */ tmp64 |= (userbuf[23] >> 8) & 0x0000000080000000; raddr[2] = (uint32_t)tmp64; - + /* Counter3 is bits 1497 to 1528 */ tmp64 = (userbuf[23] >> 7) & 0x00000000ffffffff; /* OR sticky3 (bit 1529) to counter3 bit 32 */ @@ -617,7 +617,7 @@ static int perf_stop_counters(uint32_t *raddr) userbuf[22] = 0; userbuf[23] = 0; - /* + /* * Write back the zeroed bytes + the image given * the read was destructive. */ @@ -625,13 +625,13 @@ static int perf_stop_counters(uint32_t *raddr) } else { /* - * Read RDR-15 which contains the counters and sticky bits + * Read RDR-15 which contains the counters and sticky bits */ if (!perf_rdr_read_ubuf(15, userbuf)) { return -13; } - /* + /* * Clear out the counters */ perf_rdr_clear(15); @@ -644,7 +644,7 @@ static int perf_stop_counters(uint32_t *raddr) raddr[2] = (uint32_t)((userbuf[1] >> 32) & 0x00000000ffffffffUL); raddr[3] = (uint32_t)(userbuf[1] & 0x00000000ffffffffUL); } - + return 0; } @@ -682,7 +682,7 @@ static int perf_rdr_read_ubuf(uint32_t rdr_num, uint64_t *buffer) i = tentry->num_words; while (i--) { buffer[i] = 0; - } + } /* Check for bits an even number of 64 */ if ((xbits = width & 0x03f) != 0) { @@ -808,18 +808,22 @@ static int perf_write_image(uint64_t *memaddr) } runway = ioremap_nocache(cpu_device->hpa.start, 4096); + if (!runway) { + pr_err("perf_write_image: ioremap failed!\n"); + return -ENOMEM; + } /* Merge intrigue bits into Runway STATUS 0 */ tmp64 = __raw_readq(runway + RUNWAY_STATUS) & 0xffecfffffffffffful; - __raw_writeq(tmp64 | (*memaddr++ & 0x0013000000000000ul), + __raw_writeq(tmp64 | (*memaddr++ & 0x0013000000000000ul), runway + RUNWAY_STATUS); - + /* Write RUNWAY DEBUG registers */ for (i = 0; i < 8; i++) { __raw_writeq(*memaddr++, runway + RUNWAY_DEBUG); } - return 0; + return 0; } /* @@ -843,7 +847,7 @@ printk("perf_rdr_write\n"); perf_rdr_shift_out_U(rdr_num, buffer[i]); } else { perf_rdr_shift_out_W(rdr_num, buffer[i]); - } + } } printk("perf_rdr_write done\n"); } -- cgit v1.2.3 From 73580dac7618e4bcd21679f553cf3c97323fec46 Mon Sep 17 00:00:00 2001 From: Helge Deller Date: Sat, 18 Mar 2017 17:13:27 +0100 Subject: parisc: Fix system shutdown halt On those parisc machines which don't provide a software power off function, the system currently kills the init process at the end of a shutdown and unexpectedly restarts insteads of halting. Fix it by adding a loop which will not return. Signed-off-by: Helge Deller Cc: stable@vger.kernel.org # 4.9+ --- arch/parisc/kernel/process.c | 2 ++ 1 file changed, 2 insertions(+) (limited to 'arch/parisc') diff --git a/arch/parisc/kernel/process.c b/arch/parisc/kernel/process.c index 06f7ca7fe70b..b76f503eee4a 100644 --- a/arch/parisc/kernel/process.c +++ b/arch/parisc/kernel/process.c @@ -142,6 +142,8 @@ void machine_power_off(void) printk(KERN_EMERG "System shut down completed.\n" "Please power this system off now."); + + for (;;); } void (*pm_power_off)(void) = machine_power_off; -- cgit v1.2.3 From 554bfeceb8a22d448cd986fc9efce25e833278a1 Mon Sep 17 00:00:00 2001 From: Helge Deller Date: Wed, 29 Mar 2017 21:41:05 +0200 Subject: parisc: Fix access fault handling in pa_memcpy() pa_memcpy() is the major memcpy implementation in the parisc kernel which is used to do any kind of userspace/kernel memory copies. Al Viro noticed various bugs in the implementation of pa_mempcy(), most notably that in case of faults it may report back to have copied more bytes than it actually did. Fixing those bugs is quite hard in the C-implementation, because the compiler is messing around with the registers and we are not guaranteed that specific variables are always in the same processor registers. This makes proper fault handling complicated. This patch implements pa_memcpy() in assembler. That way we have correct fault handling and adding a 64-bit copy routine was quite easy. Runtime tested with 32- and 64bit kernels. Reported-by: Al Viro Cc: # v4.9+ Signed-off-by: John David Anglin Signed-off-by: Helge Deller --- arch/parisc/lib/lusercopy.S | 318 ++++++++++++++++++++++++++++++ arch/parisc/lib/memcpy.c | 461 +------------------------------------------- 2 files changed, 321 insertions(+), 458 deletions(-) (limited to 'arch/parisc') diff --git a/arch/parisc/lib/lusercopy.S b/arch/parisc/lib/lusercopy.S index 56845de6b5df..f01188c044ee 100644 --- a/arch/parisc/lib/lusercopy.S +++ b/arch/parisc/lib/lusercopy.S @@ -5,6 +5,8 @@ * Copyright (C) 2000 Richard Hirst * Copyright (C) 2001 Matthieu Delahaye * Copyright (C) 2003 Randolph Chung + * Copyright (C) 2017 Helge Deller + * Copyright (C) 2017 John David Anglin * * * This program is free software; you can redistribute it and/or modify @@ -132,4 +134,320 @@ ENDPROC_CFI(lstrnlen_user) .procend + + +/* + * unsigned long pa_memcpy(void *dstp, const void *srcp, unsigned long len) + * + * Inputs: + * - sr1 already contains space of source region + * - sr2 already contains space of destination region + * + * Returns: + * - number of bytes that could not be copied. + * On success, this will be zero. + * + * This code is based on a C-implementation of a copy routine written by + * Randolph Chung, which in turn was derived from the glibc. + * + * Several strategies are tried to try to get the best performance for various + * conditions. In the optimal case, we copy by loops that copy 32- or 16-bytes + * at a time using general registers. Unaligned copies are handled either by + * aligning the destination and then using shift-and-write method, or in a few + * cases by falling back to a byte-at-a-time copy. + * + * Testing with various alignments and buffer sizes shows that this code is + * often >10x faster than a simple byte-at-a-time copy, even for strangely + * aligned operands. It is interesting to note that the glibc version of memcpy + * (written in C) is actually quite fast already. This routine is able to beat + * it by 30-40% for aligned copies because of the loop unrolling, but in some + * cases the glibc version is still slightly faster. This lends more + * credibility that gcc can generate very good code as long as we are careful. + * + * Possible optimizations: + * - add cache prefetching + * - try not to use the post-increment address modifiers; they may create + * additional interlocks. Assumption is that those were only efficient on old + * machines (pre PA8000 processors) + */ + + dst = arg0 + src = arg1 + len = arg2 + end = arg3 + t1 = r19 + t2 = r20 + t3 = r21 + t4 = r22 + srcspc = sr1 + dstspc = sr2 + + t0 = r1 + a1 = t1 + a2 = t2 + a3 = t3 + a0 = t4 + + save_src = ret0 + save_dst = ret1 + save_len = r31 + +ENTRY_CFI(pa_memcpy) + .proc + .callinfo NO_CALLS + .entry + + /* Last destination address */ + add dst,len,end + + /* short copy with less than 16 bytes? */ + cmpib,>>=,n 15,len,.Lbyte_loop + + /* same alignment? */ + xor src,dst,t0 + extru t0,31,2,t1 + cmpib,<>,n 0,t1,.Lunaligned_copy + +#ifdef CONFIG_64BIT + /* only do 64-bit copies if we can get aligned. */ + extru t0,31,3,t1 + cmpib,<>,n 0,t1,.Lalign_loop32 + + /* loop until we are 64-bit aligned */ +.Lalign_loop64: + extru dst,31,3,t1 + cmpib,=,n 0,t1,.Lcopy_loop_16 +20: ldb,ma 1(srcspc,src),t1 +21: stb,ma t1,1(dstspc,dst) + b .Lalign_loop64 + ldo -1(len),len + + ASM_EXCEPTIONTABLE_ENTRY(20b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(21b,.Lcopy_done) + + ldi 31,t0 +.Lcopy_loop_16: + cmpb,COND(>>=),n t0,len,.Lword_loop + +10: ldd 0(srcspc,src),t1 +11: ldd 8(srcspc,src),t2 + ldo 16(src),src +12: std,ma t1,8(dstspc,dst) +13: std,ma t2,8(dstspc,dst) +14: ldd 0(srcspc,src),t1 +15: ldd 8(srcspc,src),t2 + ldo 16(src),src +16: std,ma t1,8(dstspc,dst) +17: std,ma t2,8(dstspc,dst) + + ASM_EXCEPTIONTABLE_ENTRY(10b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(11b,.Lcopy16_fault) + ASM_EXCEPTIONTABLE_ENTRY(12b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(13b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(14b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(15b,.Lcopy16_fault) + ASM_EXCEPTIONTABLE_ENTRY(16b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(17b,.Lcopy_done) + + b .Lcopy_loop_16 + ldo -32(len),len + +.Lword_loop: + cmpib,COND(>>=),n 3,len,.Lbyte_loop +20: ldw,ma 4(srcspc,src),t1 +21: stw,ma t1,4(dstspc,dst) + b .Lword_loop + ldo -4(len),len + + ASM_EXCEPTIONTABLE_ENTRY(20b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(21b,.Lcopy_done) + +#endif /* CONFIG_64BIT */ + + /* loop until we are 32-bit aligned */ +.Lalign_loop32: + extru dst,31,2,t1 + cmpib,=,n 0,t1,.Lcopy_loop_4 +20: ldb,ma 1(srcspc,src),t1 +21: stb,ma t1,1(dstspc,dst) + b .Lalign_loop32 + ldo -1(len),len + + ASM_EXCEPTIONTABLE_ENTRY(20b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(21b,.Lcopy_done) + + +.Lcopy_loop_4: + cmpib,COND(>>=),n 15,len,.Lbyte_loop + +10: ldw 0(srcspc,src),t1 +11: ldw 4(srcspc,src),t2 +12: stw,ma t1,4(dstspc,dst) +13: stw,ma t2,4(dstspc,dst) +14: ldw 8(srcspc,src),t1 +15: ldw 12(srcspc,src),t2 + ldo 16(src),src +16: stw,ma t1,4(dstspc,dst) +17: stw,ma t2,4(dstspc,dst) + + ASM_EXCEPTIONTABLE_ENTRY(10b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(11b,.Lcopy8_fault) + ASM_EXCEPTIONTABLE_ENTRY(12b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(13b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(14b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(15b,.Lcopy8_fault) + ASM_EXCEPTIONTABLE_ENTRY(16b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(17b,.Lcopy_done) + + b .Lcopy_loop_4 + ldo -16(len),len + +.Lbyte_loop: + cmpclr,COND(<>) len,%r0,%r0 + b,n .Lcopy_done +20: ldb 0(srcspc,src),t1 + ldo 1(src),src +21: stb,ma t1,1(dstspc,dst) + b .Lbyte_loop + ldo -1(len),len + + ASM_EXCEPTIONTABLE_ENTRY(20b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(21b,.Lcopy_done) + +.Lcopy_done: + bv %r0(%r2) + sub end,dst,ret0 + + + /* src and dst are not aligned the same way. */ + /* need to go the hard way */ +.Lunaligned_copy: + /* align until dst is 32bit-word-aligned */ + extru dst,31,2,t1 + cmpib,COND(=),n 0,t1,.Lcopy_dstaligned +20: ldb 0(srcspc,src),t1 + ldo 1(src),src +21: stb,ma t1,1(dstspc,dst) + b .Lunaligned_copy + ldo -1(len),len + + ASM_EXCEPTIONTABLE_ENTRY(20b,.Lcopy_done) + ASM_EXCEPTIONTABLE_ENTRY(21b,.Lcopy_done) + +.Lcopy_dstaligned: + + /* store src, dst and len in safe place */ + copy src,save_src + copy dst,save_dst + copy len,save_len + + /* len now needs give number of words to copy */ + SHRREG len,2,len + + /* + * Copy from a not-aligned src to an aligned dst using shifts. + * Handles 4 words per loop. + */ + + depw,z src,28,2,t0 + subi 32,t0,t0 + mtsar t0 + extru len,31,2,t0 + cmpib,= 2,t0,.Lcase2 + /* Make src aligned by rounding it down. */ + depi 0,31,2,src + + cmpiclr,<> 3,t0,%r0 + b,n .Lcase3 + cmpiclr,<> 1,t0,%r0 + b,n .Lcase1 +.Lcase0: + cmpb,= %r0,len,.Lcda_finish + nop + +1: ldw,ma 4(srcspc,src), a3 + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault) +1: ldw,ma 4(srcspc,src), a0 + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault) + b,n .Ldo3 +.Lcase1: +1: ldw,ma 4(srcspc,src), a2 + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault) +1: ldw,ma 4(srcspc,src), a3 + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault) + ldo -1(len),len + cmpb,=,n %r0,len,.Ldo0 +.Ldo4: +1: ldw,ma 4(srcspc,src), a0 + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault) + shrpw a2, a3, %sar, t0 +1: stw,ma t0, 4(dstspc,dst) + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcopy_done) +.Ldo3: +1: ldw,ma 4(srcspc,src), a1 + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault) + shrpw a3, a0, %sar, t0 +1: stw,ma t0, 4(dstspc,dst) + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcopy_done) +.Ldo2: +1: ldw,ma 4(srcspc,src), a2 + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault) + shrpw a0, a1, %sar, t0 +1: stw,ma t0, 4(dstspc,dst) + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcopy_done) +.Ldo1: +1: ldw,ma 4(srcspc,src), a3 + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault) + shrpw a1, a2, %sar, t0 +1: stw,ma t0, 4(dstspc,dst) + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcopy_done) + ldo -4(len),len + cmpb,<> %r0,len,.Ldo4 + nop +.Ldo0: + shrpw a2, a3, %sar, t0 +1: stw,ma t0, 4(dstspc,dst) + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcopy_done) + +.Lcda_rdfault: +.Lcda_finish: + /* calculate new src, dst and len and jump to byte-copy loop */ + sub dst,save_dst,t0 + add save_src,t0,src + b .Lbyte_loop + sub save_len,t0,len + +.Lcase3: +1: ldw,ma 4(srcspc,src), a0 + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault) +1: ldw,ma 4(srcspc,src), a1 + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault) + b .Ldo2 + ldo 1(len),len +.Lcase2: +1: ldw,ma 4(srcspc,src), a1 + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault) +1: ldw,ma 4(srcspc,src), a2 + ASM_EXCEPTIONTABLE_ENTRY(1b,.Lcda_rdfault) + b .Ldo1 + ldo 2(len),len + + + /* fault exception fixup handlers: */ +#ifdef CONFIG_64BIT +.Lcopy16_fault: +10: b .Lcopy_done + std,ma t1,8(dstspc,dst) + ASM_EXCEPTIONTABLE_ENTRY(10b,.Lcopy_done) +#endif + +.Lcopy8_fault: +10: b .Lcopy_done + stw,ma t1,4(dstspc,dst) + ASM_EXCEPTIONTABLE_ENTRY(10b,.Lcopy_done) + + .exit +ENDPROC_CFI(pa_memcpy) + .procend + .end diff --git a/arch/parisc/lib/memcpy.c b/arch/parisc/lib/memcpy.c index f82ff10ed974..b3d47ec1d80a 100644 --- a/arch/parisc/lib/memcpy.c +++ b/arch/parisc/lib/memcpy.c @@ -2,7 +2,7 @@ * Optimized memory copy routines. * * Copyright (C) 2004 Randolph Chung - * Copyright (C) 2013 Helge Deller + * Copyright (C) 2013-2017 Helge Deller * * 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 @@ -21,474 +21,21 @@ * Portions derived from the GNU C Library * Copyright (C) 1991, 1997, 2003 Free Software Foundation, Inc. * - * Several strategies are tried to try to get the best performance for various - * conditions. In the optimal case, we copy 64-bytes in an unrolled loop using - * fp regs. This is followed by loops that copy 32- or 16-bytes at a time using - * general registers. Unaligned copies are handled either by aligning the - * destination and then using shift-and-write method, or in a few cases by - * falling back to a byte-at-a-time copy. - * - * I chose to implement this in C because it is easier to maintain and debug, - * and in my experiments it appears that the C code generated by gcc (3.3/3.4 - * at the time of writing) is fairly optimal. Unfortunately some of the - * semantics of the copy routine (exception handling) is difficult to express - * in C, so we have to play some tricks to get it to work. - * - * All the loads and stores are done via explicit asm() code in order to use - * the right space registers. - * - * Testing with various alignments and buffer sizes shows that this code is - * often >10x faster than a simple byte-at-a-time copy, even for strangely - * aligned operands. It is interesting to note that the glibc version - * of memcpy (written in C) is actually quite fast already. This routine is - * able to beat it by 30-40% for aligned copies because of the loop unrolling, - * but in some cases the glibc version is still slightly faster. This lends - * more credibility that gcc can generate very good code as long as we are - * careful. - * - * TODO: - * - cache prefetching needs more experimentation to get optimal settings - * - try not to use the post-increment address modifiers; they create additional - * interlocks - * - replace byte-copy loops with stybs sequences */ -#ifdef __KERNEL__ #include #include #include -#define s_space "%%sr1" -#define d_space "%%sr2" -#else -#include "memcpy.h" -#define s_space "%%sr0" -#define d_space "%%sr0" -#define pa_memcpy new2_copy -#endif DECLARE_PER_CPU(struct exception_data, exception_data); -#define preserve_branch(label) do { \ - volatile int dummy = 0; \ - /* The following branch is never taken, it's just here to */ \ - /* prevent gcc from optimizing away our exception code. */ \ - if (unlikely(dummy != dummy)) \ - goto label; \ -} while (0) - #define get_user_space() (segment_eq(get_fs(), KERNEL_DS) ? 0 : mfsp(3)) #define get_kernel_space() (0) -#define MERGE(w0, sh_1, w1, sh_2) ({ \ - unsigned int _r; \ - asm volatile ( \ - "mtsar %3\n" \ - "shrpw %1, %2, %%sar, %0\n" \ - : "=r"(_r) \ - : "r"(w0), "r"(w1), "r"(sh_2) \ - ); \ - _r; \ -}) -#define THRESHOLD 16 - -#ifdef DEBUG_MEMCPY -#define DPRINTF(fmt, args...) do { printk(KERN_DEBUG "%s:%d:%s ", __FILE__, __LINE__, __func__ ); printk(KERN_DEBUG fmt, ##args ); } while (0) -#else -#define DPRINTF(fmt, args...) -#endif - -#define def_load_ai_insn(_insn,_sz,_tt,_s,_a,_t,_e) \ - __asm__ __volatile__ ( \ - "1:\t" #_insn ",ma " #_sz "(" _s ",%1), %0\n\t" \ - ASM_EXCEPTIONTABLE_ENTRY(1b,_e) \ - : _tt(_t), "+r"(_a) \ - : \ - : "r8") - -#define def_store_ai_insn(_insn,_sz,_tt,_s,_a,_t,_e) \ - __asm__ __volatile__ ( \ - "1:\t" #_insn ",ma %1, " #_sz "(" _s ",%0)\n\t" \ - ASM_EXCEPTIONTABLE_ENTRY(1b,_e) \ - : "+r"(_a) \ - : _tt(_t) \ - : "r8") - -#define ldbma(_s, _a, _t, _e) def_load_ai_insn(ldbs,1,"=r",_s,_a,_t,_e) -#define stbma(_s, _t, _a, _e) def_store_ai_insn(stbs,1,"r",_s,_a,_t,_e) -#define ldwma(_s, _a, _t, _e) def_load_ai_insn(ldw,4,"=r",_s,_a,_t,_e) -#define stwma(_s, _t, _a, _e) def_store_ai_insn(stw,4,"r",_s,_a,_t,_e) -#define flddma(_s, _a, _t, _e) def_load_ai_insn(fldd,8,"=f",_s,_a,_t,_e) -#define fstdma(_s, _t, _a, _e) def_store_ai_insn(fstd,8,"f",_s,_a,_t,_e) - -#define def_load_insn(_insn,_tt,_s,_o,_a,_t,_e) \ - __asm__ __volatile__ ( \ - "1:\t" #_insn " " #_o "(" _s ",%1), %0\n\t" \ - ASM_EXCEPTIONTABLE_ENTRY(1b,_e) \ - : _tt(_t) \ - : "r"(_a) \ - : "r8") - -#define def_store_insn(_insn,_tt,_s,_t,_o,_a,_e) \ - __asm__ __volatile__ ( \ - "1:\t" #_insn " %0, " #_o "(" _s ",%1)\n\t" \ - ASM_EXCEPTIONTABLE_ENTRY(1b,_e) \ - : \ - : _tt(_t), "r"(_a) \ - : "r8") - -#define ldw(_s,_o,_a,_t,_e) def_load_insn(ldw,"=r",_s,_o,_a,_t,_e) -#define stw(_s,_t,_o,_a,_e) def_store_insn(stw,"r",_s,_t,_o,_a,_e) - -#ifdef CONFIG_PREFETCH -static inline void prefetch_src(const void *addr) -{ - __asm__("ldw 0(" s_space ",%0), %%r0" : : "r" (addr)); -} - -static inline void prefetch_dst(const void *addr) -{ - __asm__("ldd 0(" d_space ",%0), %%r0" : : "r" (addr)); -} -#else -#define prefetch_src(addr) do { } while(0) -#define prefetch_dst(addr) do { } while(0) -#endif - -#define PA_MEMCPY_OK 0 -#define PA_MEMCPY_LOAD_ERROR 1 -#define PA_MEMCPY_STORE_ERROR 2 - -/* Copy from a not-aligned src to an aligned dst, using shifts. Handles 4 words - * per loop. This code is derived from glibc. - */ -static noinline unsigned long copy_dstaligned(unsigned long dst, - unsigned long src, unsigned long len) -{ - /* gcc complains that a2 and a3 may be uninitialized, but actually - * they cannot be. Initialize a2/a3 to shut gcc up. - */ - register unsigned int a0, a1, a2 = 0, a3 = 0; - int sh_1, sh_2; - - /* prefetch_src((const void *)src); */ - - /* Calculate how to shift a word read at the memory operation - aligned srcp to make it aligned for copy. */ - sh_1 = 8 * (src % sizeof(unsigned int)); - sh_2 = 8 * sizeof(unsigned int) - sh_1; - - /* Make src aligned by rounding it down. */ - src &= -sizeof(unsigned int); - - switch (len % 4) - { - case 2: - /* a1 = ((unsigned int *) src)[0]; - a2 = ((unsigned int *) src)[1]; */ - ldw(s_space, 0, src, a1, cda_ldw_exc); - ldw(s_space, 4, src, a2, cda_ldw_exc); - src -= 1 * sizeof(unsigned int); - dst -= 3 * sizeof(unsigned int); - len += 2; - goto do1; - case 3: - /* a0 = ((unsigned int *) src)[0]; - a1 = ((unsigned int *) src)[1]; */ - ldw(s_space, 0, src, a0, cda_ldw_exc); - ldw(s_space, 4, src, a1, cda_ldw_exc); - src -= 0 * sizeof(unsigned int); - dst -= 2 * sizeof(unsigned int); - len += 1; - goto do2; - case 0: - if (len == 0) - return PA_MEMCPY_OK; - /* a3 = ((unsigned int *) src)[0]; - a0 = ((unsigned int *) src)[1]; */ - ldw(s_space, 0, src, a3, cda_ldw_exc); - ldw(s_space, 4, src, a0, cda_ldw_exc); - src -=-1 * sizeof(unsigned int); - dst -= 1 * sizeof(unsigned int); - len += 0; - goto do3; - case 1: - /* a2 = ((unsigned int *) src)[0]; - a3 = ((unsigned int *) src)[1]; */ - ldw(s_space, 0, src, a2, cda_ldw_exc); - ldw(s_space, 4, src, a3, cda_ldw_exc); - src -=-2 * sizeof(unsigned int); - dst -= 0 * sizeof(unsigned int); - len -= 1; - if (len == 0) - goto do0; - goto do4; /* No-op. */ - } - - do - { - /* prefetch_src((const void *)(src + 4 * sizeof(unsigned int))); */ -do4: - /* a0 = ((unsigned int *) src)[0]; */ - ldw(s_space, 0, src, a0, cda_ldw_exc); - /* ((unsigned int *) dst)[0] = MERGE (a2, sh_1, a3, sh_2); */ - stw(d_space, MERGE (a2, sh_1, a3, sh_2), 0, dst, cda_stw_exc); -do3: - /* a1 = ((unsigned int *) src)[1]; */ - ldw(s_space, 4, src, a1, cda_ldw_exc); - /* ((unsigned int *) dst)[1] = MERGE (a3, sh_1, a0, sh_2); */ - stw(d_space, MERGE (a3, sh_1, a0, sh_2), 4, dst, cda_stw_exc); -do2: - /* a2 = ((unsigned int *) src)[2]; */ - ldw(s_space, 8, src, a2, cda_ldw_exc); - /* ((unsigned int *) dst)[2] = MERGE (a0, sh_1, a1, sh_2); */ - stw(d_space, MERGE (a0, sh_1, a1, sh_2), 8, dst, cda_stw_exc); -do1: - /* a3 = ((unsigned int *) src)[3]; */ - ldw(s_space, 12, src, a3, cda_ldw_exc); - /* ((unsigned int *) dst)[3] = MERGE (a1, sh_1, a2, sh_2); */ - stw(d_space, MERGE (a1, sh_1, a2, sh_2), 12, dst, cda_stw_exc); - - src += 4 * sizeof(unsigned int); - dst += 4 * sizeof(unsigned int); - len -= 4; - } - while (len != 0); - -do0: - /* ((unsigned int *) dst)[0] = MERGE (a2, sh_1, a3, sh_2); */ - stw(d_space, MERGE (a2, sh_1, a3, sh_2), 0, dst, cda_stw_exc); - - preserve_branch(handle_load_error); - preserve_branch(handle_store_error); - - return PA_MEMCPY_OK; - -handle_load_error: - __asm__ __volatile__ ("cda_ldw_exc:\n"); - return PA_MEMCPY_LOAD_ERROR; - -handle_store_error: - __asm__ __volatile__ ("cda_stw_exc:\n"); - return PA_MEMCPY_STORE_ERROR; -} - - -/* Returns PA_MEMCPY_OK, PA_MEMCPY_LOAD_ERROR or PA_MEMCPY_STORE_ERROR. - * In case of an access fault the faulty address can be read from the per_cpu - * exception data struct. */ -static noinline unsigned long pa_memcpy_internal(void *dstp, const void *srcp, - unsigned long len) -{ - register unsigned long src, dst, t1, t2, t3; - register unsigned char *pcs, *pcd; - register unsigned int *pws, *pwd; - register double *pds, *pdd; - unsigned long ret; - - src = (unsigned long)srcp; - dst = (unsigned long)dstp; - pcs = (unsigned char *)srcp; - pcd = (unsigned char *)dstp; - - /* prefetch_src((const void *)srcp); */ - - if (len < THRESHOLD) - goto byte_copy; - - /* Check alignment */ - t1 = (src ^ dst); - if (unlikely(t1 & (sizeof(double)-1))) - goto unaligned_copy; - - /* src and dst have same alignment. */ - - /* Copy bytes till we are double-aligned. */ - t2 = src & (sizeof(double) - 1); - if (unlikely(t2 != 0)) { - t2 = sizeof(double) - t2; - while (t2 && len) { - /* *pcd++ = *pcs++; */ - ldbma(s_space, pcs, t3, pmc_load_exc); - len--; - stbma(d_space, t3, pcd, pmc_store_exc); - t2--; - } - } - - pds = (double *)pcs; - pdd = (double *)pcd; - -#if 0 - /* Copy 8 doubles at a time */ - while (len >= 8*sizeof(double)) { - register double r1, r2, r3, r4, r5, r6, r7, r8; - /* prefetch_src((char *)pds + L1_CACHE_BYTES); */ - flddma(s_space, pds, r1, pmc_load_exc); - flddma(s_space, pds, r2, pmc_load_exc); - flddma(s_space, pds, r3, pmc_load_exc); - flddma(s_space, pds, r4, pmc_load_exc); - fstdma(d_space, r1, pdd, pmc_store_exc); - fstdma(d_space, r2, pdd, pmc_store_exc); - fstdma(d_space, r3, pdd, pmc_store_exc); - fstdma(d_space, r4, pdd, pmc_store_exc); - -#if 0 - if (L1_CACHE_BYTES <= 32) - prefetch_src((char *)pds + L1_CACHE_BYTES); -#endif - flddma(s_space, pds, r5, pmc_load_exc); - flddma(s_space, pds, r6, pmc_load_exc); - flddma(s_space, pds, r7, pmc_load_exc); - flddma(s_space, pds, r8, pmc_load_exc); - fstdma(d_space, r5, pdd, pmc_store_exc); - fstdma(d_space, r6, pdd, pmc_store_exc); - fstdma(d_space, r7, pdd, pmc_store_exc); - fstdma(d_space, r8, pdd, pmc_store_exc); - len -= 8*sizeof(double); - } -#endif - - pws = (unsigned int *)pds; - pwd = (unsigned int *)pdd; - -word_copy: - while (len >= 8*sizeof(unsigned int)) { - register unsigned int r1,r2,r3,r4,r5,r6,r7,r8; - /* prefetch_src((char *)pws + L1_CACHE_BYTES); */ - ldwma(s_space, pws, r1, pmc_load_exc); - ldwma(s_space, pws, r2, pmc_load_exc); - ldwma(s_space, pws, r3, pmc_load_exc); - ldwma(s_space, pws, r4, pmc_load_exc); - stwma(d_space, r1, pwd, pmc_store_exc); - stwma(d_space, r2, pwd, pmc_store_exc); - stwma(d_space, r3, pwd, pmc_store_exc); - stwma(d_space, r4, pwd, pmc_store_exc); - - ldwma(s_space, pws, r5, pmc_load_exc); - ldwma(s_space, pws, r6, pmc_load_exc); - ldwma(s_space, pws, r7, pmc_load_exc); - ldwma(s_space, pws, r8, pmc_load_exc); - stwma(d_space, r5, pwd, pmc_store_exc); - stwma(d_space, r6, pwd, pmc_store_exc); - stwma(d_space, r7, pwd, pmc_store_exc); - stwma(d_space, r8, pwd, pmc_store_exc); - len -= 8*sizeof(unsigned int); - } - - while (len >= 4*sizeof(unsigned int)) { - register unsigned int r1,r2,r3,r4; - ldwma(s_space, pws, r1, pmc_load_exc); - ldwma(s_space, pws, r2, pmc_load_exc); - ldwma(s_space, pws, r3, pmc_load_exc); - ldwma(s_space, pws, r4, pmc_load_exc); - stwma(d_space, r1, pwd, pmc_store_exc); - stwma(d_space, r2, pwd, pmc_store_exc); - stwma(d_space, r3, pwd, pmc_store_exc); - stwma(d_space, r4, pwd, pmc_store_exc); - len -= 4*sizeof(unsigned int); - } - - pcs = (unsigned char *)pws; - pcd = (unsigned char *)pwd; - -byte_copy: - while (len) { - /* *pcd++ = *pcs++; */ - ldbma(s_space, pcs, t3, pmc_load_exc); - stbma(d_space, t3, pcd, pmc_store_exc); - len--; - } - - return PA_MEMCPY_OK; - -unaligned_copy: - /* possibly we are aligned on a word, but not on a double... */ - if (likely((t1 & (sizeof(unsigned int)-1)) == 0)) { - t2 = src & (sizeof(unsigned int) - 1); - - if (unlikely(t2 != 0)) { - t2 = sizeof(unsigned int) - t2; - while (t2) { - /* *pcd++ = *pcs++; */ - ldbma(s_space, pcs, t3, pmc_load_exc); - stbma(d_space, t3, pcd, pmc_store_exc); - len--; - t2--; - } - } - - pws = (unsigned int *)pcs; - pwd = (unsigned int *)pcd; - goto word_copy; - } - - /* Align the destination. */ - if (unlikely((dst & (sizeof(unsigned int) - 1)) != 0)) { - t2 = sizeof(unsigned int) - (dst & (sizeof(unsigned int) - 1)); - while (t2) { - /* *pcd++ = *pcs++; */ - ldbma(s_space, pcs, t3, pmc_load_exc); - stbma(d_space, t3, pcd, pmc_store_exc); - len--; - t2--; - } - dst = (unsigned long)pcd; - src = (unsigned long)pcs; - } - - ret = copy_dstaligned(dst, src, len / sizeof(unsigned int)); - if (ret) - return ret; - - pcs += (len & -sizeof(unsigned int)); - pcd += (len & -sizeof(unsigned int)); - len %= sizeof(unsigned int); - - preserve_branch(handle_load_error); - preserve_branch(handle_store_error); - - goto byte_copy; - -handle_load_error: - __asm__ __volatile__ ("pmc_load_exc:\n"); - return PA_MEMCPY_LOAD_ERROR; - -handle_store_error: - __asm__ __volatile__ ("pmc_store_exc:\n"); - return PA_MEMCPY_STORE_ERROR; -} - - /* Returns 0 for success, otherwise, returns number of bytes not transferred. */ -static unsigned long pa_memcpy(void *dstp, const void *srcp, unsigned long len) -{ - unsigned long ret, fault_addr, reference; - struct exception_data *d; - - ret = pa_memcpy_internal(dstp, srcp, len); - if (likely(ret == PA_MEMCPY_OK)) - return 0; - - /* if a load or store fault occured we can get the faulty addr */ - d = this_cpu_ptr(&exception_data); - fault_addr = d->fault_addr; - - /* error in load or store? */ - if (ret == PA_MEMCPY_LOAD_ERROR) - reference = (unsigned long) srcp; - else - reference = (unsigned long) dstp; +extern unsigned long pa_memcpy(void *dst, const void *src, + unsigned long len); - DPRINTF("pa_memcpy: fault type = %lu, len=%lu fault_addr=%lu ref=%lu\n", - ret, len, fault_addr, reference); - - if (fault_addr >= reference) - return len - (fault_addr - reference); - else - return len; -} - -#ifdef __KERNEL__ unsigned long __copy_to_user(void __user *dst, const void *src, unsigned long len) { @@ -537,5 +84,3 @@ long probe_kernel_read(void *dst, const void *src, size_t size) return __probe_kernel_read(dst, src, size); } - -#endif -- cgit v1.2.3 From d19f5e41b344a057bb2450024a807476f30978d2 Mon Sep 17 00:00:00 2001 From: Helge Deller Date: Sat, 25 Mar 2017 11:59:15 +0100 Subject: parisc: Clean up fixup routines for get_user()/put_user() Al Viro noticed that userspace accesses via get_user()/put_user() can be simplified a lot with regard to usage of the exception handling. This patch implements a fixup routine for get_user() and put_user() in such that the exception handler will automatically load -EFAULT into the register %r8 (the error value) in case on a fault on userspace. Additionally the fixup routine will zero the target register on fault in case of a get_user() call. The target register is extracted out of the faulting assembly instruction. This patch brings a few benefits over the old implementation: 1. Exception handling gets much cleaner, easier and smaller in size. 2. Helper functions like fixup_get_user_skip_1 (all of fixup.S) can be dropped. 3. No need to hardcode %r9 as target register for get_user() any longer. This helps the compiler register allocator and thus creates less assembler statements. 4. No dependency on the exception_data contents any longer. 5. Nested faults will be handled cleanly. Reported-by: Al Viro Cc: # v4.9+ Signed-off-by: Helge Deller --- arch/parisc/include/asm/uaccess.h | 59 +++++++++++++---------- arch/parisc/kernel/parisc_ksyms.c | 10 ---- arch/parisc/lib/Makefile | 2 +- arch/parisc/lib/fixup.S | 98 --------------------------------------- arch/parisc/mm/fault.c | 17 +++++++ 5 files changed, 52 insertions(+), 134 deletions(-) delete mode 100644 arch/parisc/lib/fixup.S (limited to 'arch/parisc') diff --git a/arch/parisc/include/asm/uaccess.h b/arch/parisc/include/asm/uaccess.h index edfbf9d6a6dd..8442727f28d2 100644 --- a/arch/parisc/include/asm/uaccess.h +++ b/arch/parisc/include/asm/uaccess.h @@ -64,6 +64,15 @@ struct exception_table_entry { ".word (" #fault_addr " - .), (" #except_addr " - .)\n\t" \ ".previous\n" +/* + * ASM_EXCEPTIONTABLE_ENTRY_EFAULT() creates a special exception table entry + * (with lowest bit set) for which the fault handler in fixup_exception() will + * load -EFAULT into %r8 for a read or write fault, and zeroes the target + * register in case of a read fault in get_user(). + */ +#define ASM_EXCEPTIONTABLE_ENTRY_EFAULT( fault_addr, except_addr )\ + ASM_EXCEPTIONTABLE_ENTRY( fault_addr, except_addr + 1) + /* * The page fault handler stores, in a per-cpu area, the following information * if a fixup routine is available. @@ -91,7 +100,7 @@ struct exception_data { #define __get_user(x, ptr) \ ({ \ register long __gu_err __asm__ ("r8") = 0; \ - register long __gu_val __asm__ ("r9") = 0; \ + register long __gu_val; \ \ load_sr2(); \ switch (sizeof(*(ptr))) { \ @@ -107,22 +116,23 @@ struct exception_data { }) #define __get_user_asm(ldx, ptr) \ - __asm__("\n1:\t" ldx "\t0(%%sr2,%2),%0\n\t" \ - ASM_EXCEPTIONTABLE_ENTRY(1b, fixup_get_user_skip_1)\ + __asm__("1: " ldx " 0(%%sr2,%2),%0\n" \ + "9:\n" \ + ASM_EXCEPTIONTABLE_ENTRY_EFAULT(1b, 9b) \ : "=r"(__gu_val), "=r"(__gu_err) \ - : "r"(ptr), "1"(__gu_err) \ - : "r1"); + : "r"(ptr), "1"(__gu_err)); #if !defined(CONFIG_64BIT) #define __get_user_asm64(ptr) \ - __asm__("\n1:\tldw 0(%%sr2,%2),%0" \ - "\n2:\tldw 4(%%sr2,%2),%R0\n\t" \ - ASM_EXCEPTIONTABLE_ENTRY(1b, fixup_get_user_skip_2)\ - ASM_EXCEPTIONTABLE_ENTRY(2b, fixup_get_user_skip_1)\ + __asm__(" copy %%r0,%R0\n" \ + "1: ldw 0(%%sr2,%2),%0\n" \ + "2: ldw 4(%%sr2,%2),%R0\n" \ + "9:\n" \ + ASM_EXCEPTIONTABLE_ENTRY_EFAULT(1b, 9b) \ + ASM_EXCEPTIONTABLE_ENTRY_EFAULT(2b, 9b) \ : "=r"(__gu_val), "=r"(__gu_err) \ - : "r"(ptr), "1"(__gu_err) \ - : "r1"); + : "r"(ptr), "1"(__gu_err)); #endif /* !defined(CONFIG_64BIT) */ @@ -148,32 +158,31 @@ struct exception_data { * The "__put_user/kernel_asm()" macros tell gcc they read from memory * instead of writing. This is because they do not write to any memory * gcc knows about, so there are no aliasing issues. These macros must - * also be aware that "fixup_put_user_skip_[12]" are executed in the - * context of the fault, and any registers used there must be listed - * as clobbers. In this case only "r1" is used by the current routines. - * r8/r9 are already listed as err/val. + * also be aware that fixups are executed in the context of the fault, + * and any registers used there must be listed as clobbers. + * r8 is already listed as err. */ #define __put_user_asm(stx, x, ptr) \ __asm__ __volatile__ ( \ - "\n1:\t" stx "\t%2,0(%%sr2,%1)\n\t" \ - ASM_EXCEPTIONTABLE_ENTRY(1b, fixup_put_user_skip_1)\ + "1: " stx " %2,0(%%sr2,%1)\n" \ + "9:\n" \ + ASM_EXCEPTIONTABLE_ENTRY_EFAULT(1b, 9b) \ : "=r"(__pu_err) \ - : "r"(ptr), "r"(x), "0"(__pu_err) \ - : "r1") + : "r"(ptr), "r"(x), "0"(__pu_err)) #if !defined(CONFIG_64BIT) #define __put_user_asm64(__val, ptr) do { \ __asm__ __volatile__ ( \ - "\n1:\tstw %2,0(%%sr2,%1)" \ - "\n2:\tstw %R2,4(%%sr2,%1)\n\t" \ - ASM_EXCEPTIONTABLE_ENTRY(1b, fixup_put_user_skip_2)\ - ASM_EXCEPTIONTABLE_ENTRY(2b, fixup_put_user_skip_1)\ + "1: stw %2,0(%%sr2,%1)\n" \ + "2: stw %R2,4(%%sr2,%1)\n" \ + "9:\n" \ + ASM_EXCEPTIONTABLE_ENTRY_EFAULT(1b, 9b) \ + ASM_EXCEPTIONTABLE_ENTRY_EFAULT(2b, 9b) \ : "=r"(__pu_err) \ - : "r"(ptr), "r"(__val), "0"(__pu_err) \ - : "r1"); \ + : "r"(ptr), "r"(__val), "0"(__pu_err)); \ } while (0) #endif /* !defined(CONFIG_64BIT) */ diff --git a/arch/parisc/kernel/parisc_ksyms.c b/arch/parisc/kernel/parisc_ksyms.c index 7484b3d11e0d..c6d6272a934f 100644 --- a/arch/parisc/kernel/parisc_ksyms.c +++ b/arch/parisc/kernel/parisc_ksyms.c @@ -47,16 +47,6 @@ EXPORT_SYMBOL(__cmpxchg_u64); EXPORT_SYMBOL(lclear_user); EXPORT_SYMBOL(lstrnlen_user); -/* Global fixups - defined as int to avoid creation of function pointers */ -extern int fixup_get_user_skip_1; -extern int fixup_get_user_skip_2; -extern int fixup_put_user_skip_1; -extern int fixup_put_user_skip_2; -EXPORT_SYMBOL(fixup_get_user_skip_1); -EXPORT_SYMBOL(fixup_get_user_skip_2); -EXPORT_SYMBOL(fixup_put_user_skip_1); -EXPORT_SYMBOL(fixup_put_user_skip_2); - #ifndef CONFIG_64BIT /* Needed so insmod can set dp value */ extern int $global$; diff --git a/arch/parisc/lib/Makefile b/arch/parisc/lib/Makefile index 8fa92b8d839a..f2dac4d73b1b 100644 --- a/arch/parisc/lib/Makefile +++ b/arch/parisc/lib/Makefile @@ -2,7 +2,7 @@ # Makefile for parisc-specific library files # -lib-y := lusercopy.o bitops.o checksum.o io.o memset.o fixup.o memcpy.o \ +lib-y := lusercopy.o bitops.o checksum.o io.o memset.o memcpy.o \ ucmpdi2.o delay.o obj-y := iomap.o diff --git a/arch/parisc/lib/fixup.S b/arch/parisc/lib/fixup.S deleted file mode 100644 index a5b72f22c7a6..000000000000 --- a/arch/parisc/lib/fixup.S +++ /dev/null @@ -1,98 +0,0 @@ -/* - * Linux/PA-RISC Project (http://www.parisc-linux.org/) - * - * Copyright (C) 2004 Randolph Chung - * - * 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; either version 2, or (at your option) - * any later version. - * - * 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. See the - * GNU General Public License for more details. - * - * You should have received a copy of the GNU General Public License - * along with this program; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. - * - * Fixup routines for kernel exception handling. - */ -#include -#include -#include -#include - -#ifdef CONFIG_SMP - .macro get_fault_ip t1 t2 - loadgp - addil LT%__per_cpu_offset,%r27 - LDREG RT%__per_cpu_offset(%r1),\t1 - /* t2 = smp_processor_id() */ - mfctl 30,\t2 - ldw TI_CPU(\t2),\t2 -#ifdef CONFIG_64BIT - extrd,u \t2,63,32,\t2 -#endif - /* t2 = &__per_cpu_offset[smp_processor_id()]; */ - LDREGX \t2(\t1),\t2 - addil LT%exception_data,%r27 - LDREG RT%exception_data(%r1),\t1 - /* t1 = this_cpu_ptr(&exception_data) */ - add,l \t1,\t2,\t1 - /* %r27 = t1->fault_gp - restore gp */ - LDREG EXCDATA_GP(\t1), %r27 - /* t1 = t1->fault_ip */ - LDREG EXCDATA_IP(\t1), \t1 - .endm -#else - .macro get_fault_ip t1 t2 - loadgp - /* t1 = this_cpu_ptr(&exception_data) */ - addil LT%exception_data,%r27 - LDREG RT%exception_data(%r1),\t2 - /* %r27 = t2->fault_gp - restore gp */ - LDREG EXCDATA_GP(\t2), %r27 - /* t1 = t2->fault_ip */ - LDREG EXCDATA_IP(\t2), \t1 - .endm -#endif - - .level LEVEL - - .text - .section .fixup, "ax" - - /* get_user() fixups, store -EFAULT in r8, and 0 in r9 */ -ENTRY_CFI(fixup_get_user_skip_1) - get_fault_ip %r1,%r8 - ldo 4(%r1), %r1 - ldi -EFAULT, %r8 - bv %r0(%r1) - copy %r0, %r9 -ENDPROC_CFI(fixup_get_user_skip_1) - -ENTRY_CFI(fixup_get_user_skip_2) - get_fault_ip %r1,%r8 - ldo 8(%r1), %r1 - ldi -EFAULT, %r8 - bv %r0(%r1) - copy %r0, %r9 -ENDPROC_CFI(fixup_get_user_skip_2) - - /* put_user() fixups, store -EFAULT in r8 */ -ENTRY_CFI(fixup_put_user_skip_1) - get_fault_ip %r1,%r8 - ldo 4(%r1), %r1 - bv %r0(%r1) - ldi -EFAULT, %r8 -ENDPROC_CFI(fixup_put_user_skip_1) - -ENTRY_CFI(fixup_put_user_skip_2) - get_fault_ip %r1,%r8 - ldo 8(%r1), %r1 - bv %r0(%r1) - ldi -EFAULT, %r8 -ENDPROC_CFI(fixup_put_user_skip_2) - diff --git a/arch/parisc/mm/fault.c b/arch/parisc/mm/fault.c index deab89a8915a..32ec22146141 100644 --- a/arch/parisc/mm/fault.c +++ b/arch/parisc/mm/fault.c @@ -150,6 +150,23 @@ int fixup_exception(struct pt_regs *regs) d->fault_space = regs->isr; d->fault_addr = regs->ior; + /* + * Fix up get_user() and put_user(). + * ASM_EXCEPTIONTABLE_ENTRY_EFAULT() sets the least-significant + * bit in the relative address of the fixup routine to indicate + * that %r8 should be loaded with -EFAULT to report a userspace + * access error. + */ + if (fix->fixup & 1) { + regs->gr[8] = -EFAULT; + + /* zero target register for get_user() */ + if (parisc_acctyp(0, regs->iir) == VM_READ) { + int treg = regs->iir & 0x1f; + regs->gr[treg] = 0; + } + } + regs->iaoq[0] = (unsigned long)&fix->fixup + fix->fixup; regs->iaoq[0] &= ~3; /* -- cgit v1.2.3 From 476e75a44b56038bee9207242d4bc718f6b4de06 Mon Sep 17 00:00:00 2001 From: Helge Deller Date: Wed, 29 Mar 2017 08:25:30 +0200 Subject: parisc: Avoid stalled CPU warnings after system shutdown Commit 73580dac7618 ("parisc: Fix system shutdown halt") introduced an endless loop for systems which don't provide a software power off function. But the soft lockup detector will detect this and report stalled CPUs after some time. Avoid those unwanted warnings by disabling the soft lockup detector. Fixes: 73580dac7618 ("parisc: Fix system shutdown halt") Signed-off-by: Helge Deller Cc: stable@vger.kernel.org # 4.9+ --- arch/parisc/kernel/process.c | 2 ++ 1 file changed, 2 insertions(+) (limited to 'arch/parisc') diff --git a/arch/parisc/kernel/process.c b/arch/parisc/kernel/process.c index b76f503eee4a..4516a5b53f38 100644 --- a/arch/parisc/kernel/process.c +++ b/arch/parisc/kernel/process.c @@ -143,6 +143,8 @@ void machine_power_off(void) printk(KERN_EMERG "System shut down completed.\n" "Please power this system off now."); + /* prevent soft lockup/stalled CPU messages for endless loop. */ + rcu_sysrq_start(); for (;;); } -- cgit v1.2.3