linux/arch/powerpc/kernel/setup_64.c

673 lines
17 KiB
C
Raw Normal View History

/*
*
* Common boot and setup code.
*
* Copyright (C) 2001 PPC64 Team, IBM Corp
*
* 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 of the License, or (at your option) any later version.
*/
#undef DEBUG
#include <linux/module.h>
#include <linux/string.h>
#include <linux/sched.h>
#include <linux/init.h>
#include <linux/kernel.h>
#include <linux/reboot.h>
#include <linux/delay.h>
#include <linux/initrd.h>
#include <linux/seq_file.h>
#include <linux/ioport.h>
#include <linux/console.h>
#include <linux/utsname.h>
#include <linux/tty.h>
#include <linux/root_dev.h>
#include <linux/notifier.h>
#include <linux/cpu.h>
#include <linux/unistd.h>
#include <linux/serial.h>
#include <linux/serial_8250.h>
[PATCH] powerpc/64: per cpu data optimisations The current ppc64 per cpu data implementation is quite slow. eg: lhz 11,18(13) /* smp_processor_id() */ ld 9,.LC63-.LCTOC1(30) /* per_cpu__variable_name */ ld 8,.LC61-.LCTOC1(30) /* __per_cpu_offset */ sldi 11,11,3 /* form index into __per_cpu_offset */ mr 10,9 ldx 9,11,8 /* __per_cpu_offset[smp_processor_id()] */ ldx 0,10,9 /* load per cpu data */ 5 loads for something that is supposed to be fast, pretty awful. One reason for the large number of loads is that we have to synthesize 2 64bit constants (per_cpu__variable_name and __per_cpu_offset). By putting __per_cpu_offset into the paca we can avoid the 2 loads associated with it: ld 11,56(13) /* paca->data_offset */ ld 9,.LC59-.LCTOC1(30) /* per_cpu__variable_name */ ldx 0,9,11 /* load per cpu data Longer term we can should be able to do even better than 3 loads. If per_cpu__variable_name wasnt a 64bit constant and paca->data_offset was in a register we could cut it down to one load. A suggestion from Rusty is to use gcc's __thread extension here. In order to do this we would need to free up r13 (the __thread register and where the paca currently is). So far Ive had a few unsuccessful attempts at doing that :) The patch also allocates per cpu memory node local on NUMA machines. This patch from Rusty has been sitting in my queue _forever_ but stalled when I hit the compiler bug. Sorry about that. Finally I also only allocate per cpu data for possible cpus, which comes straight out of the x86-64 port. On a pseries kernel (with NR_CPUS == 128) and 4 possible cpus we see some nice gains: total used free shared buffers cached Mem: 4012228 212860 3799368 0 0 162424 total used free shared buffers cached Mem: 4016200 212984 3803216 0 0 162424 A saving of 3.75MB. Quite nice for smaller machines. Note: we now have to be careful of per cpu users that touch data for !possible cpus. At this stage it might be worth making the NUMA and possible cpu optimisations generic, but per cpu init is done so early we have to be careful that all architectures have their possible map setup correctly. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Paul Mackerras <paulus@samba.org>
2006-01-11 02:16:44 +00:00
#include <linux/bootmem.h>
#include <linux/pci.h>
#include <linux/lockdep.h>
#include <linux/memblock.h>
#include <asm/io.h>
#include <asm/kdump.h>
#include <asm/prom.h>
#include <asm/processor.h>
#include <asm/pgtable.h>
#include <asm/smp.h>
#include <asm/elf.h>
#include <asm/machdep.h>
#include <asm/paca.h>
#include <asm/time.h>
#include <asm/cputable.h>
#include <asm/sections.h>
#include <asm/btext.h>
#include <asm/nvram.h>
#include <asm/setup.h>
#include <asm/system.h>
#include <asm/rtas.h>
#include <asm/iommu.h>
#include <asm/serial.h>
#include <asm/cache.h>
#include <asm/page.h>
#include <asm/mmu.h>
#include <asm/firmware.h>
#include <asm/xmon.h>
#include <asm/udbg.h>
#include <asm/kexec.h>
#include <asm/mmu_context.h>
#include <asm/code-patching.h>
#include "setup.h"
#ifdef DEBUG
#define DBG(fmt...) udbg_printf(fmt)
#else
#define DBG(fmt...)
#endif
int boot_cpuid = 0;
int __initdata boot_cpu_count;
u64 ppc64_pft_size;
/* Pick defaults since we might want to patch instructions
* before we've read this from the device tree.
*/
struct ppc64_caches ppc64_caches = {
.dline_size = 0x40,
.log_dline_size = 6,
.iline_size = 0x40,
.log_iline_size = 6
};
EXPORT_SYMBOL_GPL(ppc64_caches);
/*
* These are used in binfmt_elf.c to put aux entries on the stack
* for each elf executable being started.
*/
int dcache_bsize;
int icache_bsize;
int ucache_bsize;
#ifdef CONFIG_SMP
static char *smt_enabled_cmdline;
/* Look for ibm,smt-enabled OF option */
static void check_smt_enabled(void)
{
struct device_node *dn;
const char *smt_option;
/* Default to enabling all threads */
smt_enabled_at_boot = threads_per_core;
/* Allow the command line to overrule the OF option */
if (smt_enabled_cmdline) {
if (!strcmp(smt_enabled_cmdline, "on"))
smt_enabled_at_boot = threads_per_core;
else if (!strcmp(smt_enabled_cmdline, "off"))
smt_enabled_at_boot = 0;
else {
long smt;
int rc;
rc = strict_strtol(smt_enabled_cmdline, 10, &smt);
if (!rc)
smt_enabled_at_boot =
min(threads_per_core, (int)smt);
}
} else {
dn = of_find_node_by_path("/options");
if (dn) {
smt_option = of_get_property(dn, "ibm,smt-enabled",
NULL);
if (smt_option) {
if (!strcmp(smt_option, "on"))
smt_enabled_at_boot = threads_per_core;
else if (!strcmp(smt_option, "off"))
smt_enabled_at_boot = 0;
}
of_node_put(dn);
}
}
}
/* Look for smt-enabled= cmdline option */
static int __init early_smt_enabled(char *p)
{
smt_enabled_cmdline = p;
return 0;
}
early_param("smt-enabled", early_smt_enabled);
#else
#define check_smt_enabled()
#endif /* CONFIG_SMP */
/*
* Early initialization entry point. This is called by head.S
* with MMU translation disabled. We rely on the "feature" of
* the CPU that ignores the top 2 bits of the address in real
* mode so we can access kernel globals normally provided we
* only toy with things in the RMO region. From here, we do
* some early parsing of the device-tree to setup out MEMBLOCK
* data structures, and allocate & initialize the hash table
* and segment tables so we can start running with translation
* enabled.
*
* It is this function which will call the probe() callback of
* the various platform types and copy the matching one to the
* global ppc_md structure. Your platform can eventually do
* some very early initializations from the probe() routine, but
* this is not recommended, be very careful as, for example, the
* device-tree is not accessible via normal means at this point.
*/
void __init early_setup(unsigned long dt_ptr)
{
/* -------- printk is _NOT_ safe to use here ! ------- */
/* Identify CPU type */
identify_cpu(0, mfspr(SPRN_PVR));
/* Assume we're on cpu 0 for now. Don't write to the paca yet! */
initialise_paca(&boot_paca, 0);
setup_paca(&boot_paca);
/* Initialize lockdep early or else spinlocks will blow */
lockdep_init();
/* -------- printk is now safe to use ------- */
/* Enable early debugging if any specified (see udbg.h) */
udbg_early_init();
DBG(" -> early_setup(), dt_ptr: 0x%lx\n", dt_ptr);
/*
* Do early initialization using the flattened device
* tree, such as retrieving the physical memory map or
* calculating/retrieving the hash table size.
*/
early_init_devtree(__va(dt_ptr));
/* Now we know the logical id of our boot cpu, setup the paca. */
setup_paca(&paca[boot_cpuid]);
/* Fix up paca fields required for the boot cpu */
get_paca()->cpu_start = 1;
/* Probe the machine type */
probe_machine();
setup_kdump_trampoline();
DBG("Found, Initializing memory management...\n");
/* Initialize the hash table or TLB handling */
early_init_mmu();
DBG(" <- early_setup()\n");
}
#ifdef CONFIG_SMP
void early_setup_secondary(void)
{
/* Mark interrupts enabled in PACA */
get_paca()->soft_enabled = 0;
/* Initialize the hash table or TLB handling */
early_init_mmu_secondary();
}
#endif /* CONFIG_SMP */
#if defined(CONFIG_SMP) || defined(CONFIG_KEXEC)
void smp_release_cpus(void)
{
unsigned long *ptr;
int i;
DBG(" -> smp_release_cpus()\n");
/* All secondary cpus are spinning on a common spinloop, release them
* all now so they can start to spin on their individual paca
* spinloops. For non SMP kernels, the secondary cpus never get out
* of the common spinloop.
powerpc: Make it possible to move the interrupt handlers away from the kernel This changes the way that the exception prologs transfer control to the handlers in 64-bit kernels with the aim of making it possible to have the prologs separate from the main body of the kernel. Now, instead of computing the address of the handler by taking the top 32 bits of the paca address (to get the 0xc0000000........ part) and ORing in something in the bottom 16 bits, we get the base address of the kernel by doing a load from the paca and add an offset. This also replaces an mfmsr and an ori to compute the MSR value for the handler with a load from the paca. That makes it unnecessary to have a separate version of EXCEPTION_PROLOG_PSERIES that forces 64-bit mode. We can no longer use a direct branches in the exception prolog code, which means that the SLB miss handlers can't branch directly to .slb_miss_realmode any more. Instead we have to compute the address and do an indirect branch. This is conditional on CONFIG_RELOCATABLE; for non-relocatable kernels we use a direct branch as before. (A later change will allow CONFIG_RELOCATABLE to be set on 64-bit powerpc.) Since the secondary CPUs on pSeries start execution in the first 0x100 bytes of real memory and then have to get to wherever the kernel is, we can't use a direct branch to get there. Instead this changes __secondary_hold_spinloop from a flag to a function pointer. When it is set to a non-NULL value, the secondary CPUs jump to the function pointed to by that value. Finally this eliminates one code difference between 32-bit and 64-bit by making __secondary_hold be the text address of the secondary CPU spinloop rather than a function descriptor for it. Signed-off-by: Paul Mackerras <paulus@samba.org>
2008-08-30 01:40:24 +00:00
*/
ptr = (unsigned long *)((unsigned long)&__secondary_hold_spinloop
- PHYSICAL_START);
powerpc: Make it possible to move the interrupt handlers away from the kernel This changes the way that the exception prologs transfer control to the handlers in 64-bit kernels with the aim of making it possible to have the prologs separate from the main body of the kernel. Now, instead of computing the address of the handler by taking the top 32 bits of the paca address (to get the 0xc0000000........ part) and ORing in something in the bottom 16 bits, we get the base address of the kernel by doing a load from the paca and add an offset. This also replaces an mfmsr and an ori to compute the MSR value for the handler with a load from the paca. That makes it unnecessary to have a separate version of EXCEPTION_PROLOG_PSERIES that forces 64-bit mode. We can no longer use a direct branches in the exception prolog code, which means that the SLB miss handlers can't branch directly to .slb_miss_realmode any more. Instead we have to compute the address and do an indirect branch. This is conditional on CONFIG_RELOCATABLE; for non-relocatable kernels we use a direct branch as before. (A later change will allow CONFIG_RELOCATABLE to be set on 64-bit powerpc.) Since the secondary CPUs on pSeries start execution in the first 0x100 bytes of real memory and then have to get to wherever the kernel is, we can't use a direct branch to get there. Instead this changes __secondary_hold_spinloop from a flag to a function pointer. When it is set to a non-NULL value, the secondary CPUs jump to the function pointed to by that value. Finally this eliminates one code difference between 32-bit and 64-bit by making __secondary_hold be the text address of the secondary CPU spinloop rather than a function descriptor for it. Signed-off-by: Paul Mackerras <paulus@samba.org>
2008-08-30 01:40:24 +00:00
*ptr = __pa(generic_secondary_smp_init);
/* And wait a bit for them to catch up */
for (i = 0; i < 100000; i++) {
mb();
HMT_low();
if (boot_cpu_count == 0)
break;
udelay(1);
}
DBG("boot_cpu_count = %d\n", boot_cpu_count);
DBG(" <- smp_release_cpus()\n");
}
#endif /* CONFIG_SMP || CONFIG_KEXEC */
/*
* Initialize some remaining members of the ppc64_caches and systemcfg
* structures
* (at least until we get rid of them completely). This is mostly some
* cache informations about the CPU that will be used by cache flush
* routines and/or provided to userland
*/
static void __init initialize_cache_info(void)
{
struct device_node *np;
unsigned long num_cpus = 0;
DBG(" -> initialize_cache_info()\n");
for (np = NULL; (np = of_find_node_by_type(np, "cpu"));) {
num_cpus += 1;
/* We're assuming *all* of the CPUs have the same
* d-cache and i-cache sizes... -Peter
*/
if ( num_cpus == 1 ) {
const u32 *sizep, *lsizep;
u32 size, lsize;
size = 0;
lsize = cur_cpu_spec->dcache_bsize;
sizep = of_get_property(np, "d-cache-size", NULL);
if (sizep != NULL)
size = *sizep;
lsizep = of_get_property(np, "d-cache-block-size", NULL);
/* fallback if block size missing */
if (lsizep == NULL)
lsizep = of_get_property(np, "d-cache-line-size", NULL);
if (lsizep != NULL)
lsize = *lsizep;
if (sizep == 0 || lsizep == 0)
DBG("Argh, can't find dcache properties ! "
"sizep: %p, lsizep: %p\n", sizep, lsizep);
ppc64_caches.dsize = size;
ppc64_caches.dline_size = lsize;
ppc64_caches.log_dline_size = __ilog2(lsize);
ppc64_caches.dlines_per_page = PAGE_SIZE / lsize;
size = 0;
lsize = cur_cpu_spec->icache_bsize;
sizep = of_get_property(np, "i-cache-size", NULL);
if (sizep != NULL)
size = *sizep;
lsizep = of_get_property(np, "i-cache-block-size", NULL);
if (lsizep == NULL)
lsizep = of_get_property(np, "i-cache-line-size", NULL);
if (lsizep != NULL)
lsize = *lsizep;
if (sizep == 0 || lsizep == 0)
DBG("Argh, can't find icache properties ! "
"sizep: %p, lsizep: %p\n", sizep, lsizep);
ppc64_caches.isize = size;
ppc64_caches.iline_size = lsize;
ppc64_caches.log_iline_size = __ilog2(lsize);
ppc64_caches.ilines_per_page = PAGE_SIZE / lsize;
}
}
DBG(" <- initialize_cache_info()\n");
}
/*
* Do some initial setup of the system. The parameters are those which
* were passed in from the bootloader.
*/
void __init setup_system(void)
{
DBG(" -> setup_system()\n");
/* Apply the CPUs-specific and firmware specific fixups to kernel
* text (nop out sections not relevant to this CPU or this firmware)
*/
[POWERPC] Support feature fixups in vdso's This patch reworks the feature fixup mecanism so vdso's can be fixed up. The main issue was that the construct: .long label (or .llong on 64 bits) will not work in the case of a shared library like the vdso. It will generate an empty placeholder in the fixup table along with a reloc, which is not something we can deal with in the vdso. The idea here (thanks Alan Modra !) is to instead use something like: 1: .long label - 1b That is, the feature fixup tables no longer contain addresses of bits of code to patch, but offsets of such code from the fixup table entry itself. That is properly resolved by ld when building the .so's. I've modified the fixup mecanism generically to use that method for the rest of the kernel as well. Another trick is that the 32 bits vDSO included in the 64 bits kernel need to have a table in the 64 bits format. However, gas does not support 32 bits code with a statement of the form: .llong label - 1b (Or even just .llong label) That is, it cannot emit the right fixup/relocation for the linker to use to assign a 32 bits address to an .llong field. Thus, in the specific case of the 32 bits vdso built as part of the 64 bits kernel, we are using a modified macro that generates: .long 0xffffffff .llong label - 1b Note that is assumes that the value is negative which is enforced by the .lds (those offsets are always negative as the .text is always before the fixup table and gas doesn't support emiting the reloc the other way around). Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@samba.org>
2006-10-20 01:47:18 +00:00
do_feature_fixups(cur_cpu_spec->cpu_features,
&__start___ftr_fixup, &__stop___ftr_fixup);
do_feature_fixups(cur_cpu_spec->mmu_features,
&__start___mmu_ftr_fixup, &__stop___mmu_ftr_fixup);
do_feature_fixups(powerpc_firmware_features,
&__start___fw_ftr_fixup, &__stop___fw_ftr_fixup);
do_lwsync_fixups(cur_cpu_spec->cpu_features,
&__start___lwsync_fixup, &__stop___lwsync_fixup);
/*
* Unflatten the device-tree passed by prom_init or kexec
*/
unflatten_device_tree();
/*
* Fill the ppc64_caches & systemcfg structures with informations
2006-07-03 11:36:01 +00:00
* retrieved from the device-tree.
*/
initialize_cache_info();
#ifdef CONFIG_PPC_RTAS
/*
* Initialize RTAS if available
*/
rtas_initialize();
#endif /* CONFIG_PPC_RTAS */
/*
* Check if we have an initrd provided via the device-tree
*/
check_for_initrd();
/*
* Do some platform specific early initializations, that includes
* setting up the hash table pointers. It also sets up some interrupt-mapping
* related options that will be used by finish_device_tree()
*/
if (ppc_md.init_early)
ppc_md.init_early();
/*
* We can discover serial ports now since the above did setup the
* hash table management for us, thus ioremap works. We do that early
* so that further code can be debugged
*/
find_legacy_serial_ports();
/*
* Register early console
*/
register_early_udbg_console();
/*
* Initialize xmon
*/
xmon_setup();
smp_setup_cpu_maps();
check_smt_enabled();
#ifdef CONFIG_SMP
/* Release secondary cpus out of their spinloops at 0x60 now that
* we can map physical -> logical CPU ids
*/
smp_release_cpus();
#endif
printk("Starting Linux PPC64 %s\n", init_utsname()->version);
printk("-----------------------------------------------------\n");
printk("ppc64_pft_size = 0x%llx\n", ppc64_pft_size);
printk("physicalMemorySize = 0x%llx\n", memblock_phys_mem_size());
if (ppc64_caches.dline_size != 0x80)
printk("ppc64_caches.dcache_line_size = 0x%x\n",
ppc64_caches.dline_size);
if (ppc64_caches.iline_size != 0x80)
printk("ppc64_caches.icache_line_size = 0x%x\n",
ppc64_caches.iline_size);
#ifdef CONFIG_PPC_STD_MMU_64
if (htab_address)
printk("htab_address = 0x%p\n", htab_address);
printk("htab_hash_mask = 0x%lx\n", htab_hash_mask);
#endif /* CONFIG_PPC_STD_MMU_64 */
if (PHYSICAL_START > 0)
printk("physical_start = 0x%llx\n",
(unsigned long long)PHYSICAL_START);
printk("-----------------------------------------------------\n");
DBG(" <- setup_system()\n");
}
/* This returns the limit below which memory accesses to the linear
* mapping are guarnateed not to cause a TLB or SLB miss. This is
* used to allocate interrupt or emergency stacks for which our
* exception entry path doesn't deal with being interrupted.
*/
static u64 safe_stack_limit(void)
{
#ifdef CONFIG_PPC_BOOK3E
/* Freescale BookE bolts the entire linear mapping */
if (mmu_has_feature(MMU_FTR_TYPE_FSL_E))
return linear_map_top;
/* Other BookE, we assume the first GB is bolted */
return 1ul << 30;
#else
/* BookS, the first segment is bolted */
if (mmu_has_feature(MMU_FTR_1T_SEGMENT))
return 1UL << SID_SHIFT_1T;
return 1UL << SID_SHIFT;
#endif
}
static void __init irqstack_early_init(void)
{
u64 limit = safe_stack_limit();
unsigned int i;
/*
* Interrupt stacks must be in the first segment since we
* cannot afford to take SLB misses on them.
*/
for_each_possible_cpu(i) {
softirq_ctx[i] = (struct thread_info *)
__va(memblock_alloc_base(THREAD_SIZE,
THREAD_SIZE, limit));
hardirq_ctx[i] = (struct thread_info *)
__va(memblock_alloc_base(THREAD_SIZE,
THREAD_SIZE, limit));
}
}
#ifdef CONFIG_PPC_BOOK3E
static void __init exc_lvl_early_init(void)
{
extern unsigned int interrupt_base_book3e;
extern unsigned int exc_debug_debug_book3e;
unsigned int i;
for_each_possible_cpu(i) {
critirq_ctx[i] = (struct thread_info *)
__va(memblock_alloc(THREAD_SIZE, THREAD_SIZE));
dbgirq_ctx[i] = (struct thread_info *)
__va(memblock_alloc(THREAD_SIZE, THREAD_SIZE));
mcheckirq_ctx[i] = (struct thread_info *)
__va(memblock_alloc(THREAD_SIZE, THREAD_SIZE));
}
if (cpu_has_feature(CPU_FTR_DEBUG_LVL_EXC))
patch_branch(&interrupt_base_book3e + (0x040 / 4) + 1,
(unsigned long)&exc_debug_debug_book3e, 0);
}
#else
#define exc_lvl_early_init()
#endif
/*
* Stack space used when we detect a bad kernel stack pointer, and
* early in SMP boots before relocation is enabled.
*/
static void __init emergency_stack_init(void)
{
u64 limit;
unsigned int i;
/*
* Emergency stacks must be under 256MB, we cannot afford to take
* SLB misses on them. The ABI also requires them to be 128-byte
* aligned.
*
* Since we use these as temporary stacks during secondary CPU
* bringup, we need to get at them in real mode. This means they
* must also be within the RMO region.
*/
limit = min(safe_stack_limit(), ppc64_rma_size);
for_each_possible_cpu(i) {
unsigned long sp;
sp = memblock_alloc_base(THREAD_SIZE, THREAD_SIZE, limit);
sp += THREAD_SIZE;
paca[i].emergency_sp = __va(sp);
}
}
/*
* Called into from start_kernel this initializes bootmem, which is used
* to manage page allocation until mem_init is called.
*/
void __init setup_arch(char **cmdline_p)
{
ppc64_boot_msg(0x12, "Setup Arch");
*cmdline_p = cmd_line;
/*
* Set cache line size based on type of cpu as a default.
* Systems with OF can look in the properties on the cpu node(s)
* for a possibly more accurate value.
*/
dcache_bsize = ppc64_caches.dline_size;
icache_bsize = ppc64_caches.iline_size;
/* reboot on panic */
panic_timeout = 180;
if (ppc_md.panic)
setup_panic();
init_mm.start_code = (unsigned long)_stext;
init_mm.end_code = (unsigned long) _etext;
init_mm.end_data = (unsigned long) _edata;
init_mm.brk = klimit;
irqstack_early_init();
exc_lvl_early_init();
emergency_stack_init();
#ifdef CONFIG_PPC_STD_MMU_64
stabs_alloc();
#endif
/* set up the bootmem stuff with available memory */
do_init_bootmem();
sparse_init();
#ifdef CONFIG_DUMMY_CONSOLE
conswitchp = &dummy_con;
#endif
if (ppc_md.setup_arch)
ppc_md.setup_arch();
paging_init();
/* Initialize the MMU context management stuff */
mmu_context_init();
ppc64_boot_msg(0x15, "Setup Done");
}
/* ToDo: do something useful if ppc_md is not yet setup. */
#define PPC64_LINUX_FUNCTION 0x0f000000
#define PPC64_IPL_MESSAGE 0xc0000000
#define PPC64_TERM_MESSAGE 0xb0000000
static void ppc64_do_msg(unsigned int src, const char *msg)
{
if (ppc_md.progress) {
char buf[128];
sprintf(buf, "%08X\n", src);
ppc_md.progress(buf, 0);
snprintf(buf, 128, "%s", msg);
ppc_md.progress(buf, 0);
}
}
/* Print a boot progress message. */
void ppc64_boot_msg(unsigned int src, const char *msg)
{
ppc64_do_msg(PPC64_LINUX_FUNCTION|PPC64_IPL_MESSAGE|src, msg);
printk("[boot]%04x %s\n", src, msg);
}
[PATCH] powerpc/64: per cpu data optimisations The current ppc64 per cpu data implementation is quite slow. eg: lhz 11,18(13) /* smp_processor_id() */ ld 9,.LC63-.LCTOC1(30) /* per_cpu__variable_name */ ld 8,.LC61-.LCTOC1(30) /* __per_cpu_offset */ sldi 11,11,3 /* form index into __per_cpu_offset */ mr 10,9 ldx 9,11,8 /* __per_cpu_offset[smp_processor_id()] */ ldx 0,10,9 /* load per cpu data */ 5 loads for something that is supposed to be fast, pretty awful. One reason for the large number of loads is that we have to synthesize 2 64bit constants (per_cpu__variable_name and __per_cpu_offset). By putting __per_cpu_offset into the paca we can avoid the 2 loads associated with it: ld 11,56(13) /* paca->data_offset */ ld 9,.LC59-.LCTOC1(30) /* per_cpu__variable_name */ ldx 0,9,11 /* load per cpu data Longer term we can should be able to do even better than 3 loads. If per_cpu__variable_name wasnt a 64bit constant and paca->data_offset was in a register we could cut it down to one load. A suggestion from Rusty is to use gcc's __thread extension here. In order to do this we would need to free up r13 (the __thread register and where the paca currently is). So far Ive had a few unsuccessful attempts at doing that :) The patch also allocates per cpu memory node local on NUMA machines. This patch from Rusty has been sitting in my queue _forever_ but stalled when I hit the compiler bug. Sorry about that. Finally I also only allocate per cpu data for possible cpus, which comes straight out of the x86-64 port. On a pseries kernel (with NR_CPUS == 128) and 4 possible cpus we see some nice gains: total used free shared buffers cached Mem: 4012228 212860 3799368 0 0 162424 total used free shared buffers cached Mem: 4016200 212984 3803216 0 0 162424 A saving of 3.75MB. Quite nice for smaller machines. Note: we now have to be careful of per cpu users that touch data for !possible cpus. At this stage it might be worth making the NUMA and possible cpu optimisations generic, but per cpu init is done so early we have to be careful that all architectures have their possible map setup correctly. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Paul Mackerras <paulus@samba.org>
2006-01-11 02:16:44 +00:00
#ifdef CONFIG_SMP
#define PCPU_DYN_SIZE ()
static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
[PATCH] powerpc/64: per cpu data optimisations The current ppc64 per cpu data implementation is quite slow. eg: lhz 11,18(13) /* smp_processor_id() */ ld 9,.LC63-.LCTOC1(30) /* per_cpu__variable_name */ ld 8,.LC61-.LCTOC1(30) /* __per_cpu_offset */ sldi 11,11,3 /* form index into __per_cpu_offset */ mr 10,9 ldx 9,11,8 /* __per_cpu_offset[smp_processor_id()] */ ldx 0,10,9 /* load per cpu data */ 5 loads for something that is supposed to be fast, pretty awful. One reason for the large number of loads is that we have to synthesize 2 64bit constants (per_cpu__variable_name and __per_cpu_offset). By putting __per_cpu_offset into the paca we can avoid the 2 loads associated with it: ld 11,56(13) /* paca->data_offset */ ld 9,.LC59-.LCTOC1(30) /* per_cpu__variable_name */ ldx 0,9,11 /* load per cpu data Longer term we can should be able to do even better than 3 loads. If per_cpu__variable_name wasnt a 64bit constant and paca->data_offset was in a register we could cut it down to one load. A suggestion from Rusty is to use gcc's __thread extension here. In order to do this we would need to free up r13 (the __thread register and where the paca currently is). So far Ive had a few unsuccessful attempts at doing that :) The patch also allocates per cpu memory node local on NUMA machines. This patch from Rusty has been sitting in my queue _forever_ but stalled when I hit the compiler bug. Sorry about that. Finally I also only allocate per cpu data for possible cpus, which comes straight out of the x86-64 port. On a pseries kernel (with NR_CPUS == 128) and 4 possible cpus we see some nice gains: total used free shared buffers cached Mem: 4012228 212860 3799368 0 0 162424 total used free shared buffers cached Mem: 4016200 212984 3803216 0 0 162424 A saving of 3.75MB. Quite nice for smaller machines. Note: we now have to be careful of per cpu users that touch data for !possible cpus. At this stage it might be worth making the NUMA and possible cpu optimisations generic, but per cpu init is done so early we have to be careful that all architectures have their possible map setup correctly. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Paul Mackerras <paulus@samba.org>
2006-01-11 02:16:44 +00:00
{
return __alloc_bootmem_node(NODE_DATA(cpu_to_node(cpu)), size, align,
__pa(MAX_DMA_ADDRESS));
}
[PATCH] powerpc/64: per cpu data optimisations The current ppc64 per cpu data implementation is quite slow. eg: lhz 11,18(13) /* smp_processor_id() */ ld 9,.LC63-.LCTOC1(30) /* per_cpu__variable_name */ ld 8,.LC61-.LCTOC1(30) /* __per_cpu_offset */ sldi 11,11,3 /* form index into __per_cpu_offset */ mr 10,9 ldx 9,11,8 /* __per_cpu_offset[smp_processor_id()] */ ldx 0,10,9 /* load per cpu data */ 5 loads for something that is supposed to be fast, pretty awful. One reason for the large number of loads is that we have to synthesize 2 64bit constants (per_cpu__variable_name and __per_cpu_offset). By putting __per_cpu_offset into the paca we can avoid the 2 loads associated with it: ld 11,56(13) /* paca->data_offset */ ld 9,.LC59-.LCTOC1(30) /* per_cpu__variable_name */ ldx 0,9,11 /* load per cpu data Longer term we can should be able to do even better than 3 loads. If per_cpu__variable_name wasnt a 64bit constant and paca->data_offset was in a register we could cut it down to one load. A suggestion from Rusty is to use gcc's __thread extension here. In order to do this we would need to free up r13 (the __thread register and where the paca currently is). So far Ive had a few unsuccessful attempts at doing that :) The patch also allocates per cpu memory node local on NUMA machines. This patch from Rusty has been sitting in my queue _forever_ but stalled when I hit the compiler bug. Sorry about that. Finally I also only allocate per cpu data for possible cpus, which comes straight out of the x86-64 port. On a pseries kernel (with NR_CPUS == 128) and 4 possible cpus we see some nice gains: total used free shared buffers cached Mem: 4012228 212860 3799368 0 0 162424 total used free shared buffers cached Mem: 4016200 212984 3803216 0 0 162424 A saving of 3.75MB. Quite nice for smaller machines. Note: we now have to be careful of per cpu users that touch data for !possible cpus. At this stage it might be worth making the NUMA and possible cpu optimisations generic, but per cpu init is done so early we have to be careful that all architectures have their possible map setup correctly. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Paul Mackerras <paulus@samba.org>
2006-01-11 02:16:44 +00:00
static void __init pcpu_fc_free(void *ptr, size_t size)
{
free_bootmem(__pa(ptr), size);
}
[PATCH] powerpc/64: per cpu data optimisations The current ppc64 per cpu data implementation is quite slow. eg: lhz 11,18(13) /* smp_processor_id() */ ld 9,.LC63-.LCTOC1(30) /* per_cpu__variable_name */ ld 8,.LC61-.LCTOC1(30) /* __per_cpu_offset */ sldi 11,11,3 /* form index into __per_cpu_offset */ mr 10,9 ldx 9,11,8 /* __per_cpu_offset[smp_processor_id()] */ ldx 0,10,9 /* load per cpu data */ 5 loads for something that is supposed to be fast, pretty awful. One reason for the large number of loads is that we have to synthesize 2 64bit constants (per_cpu__variable_name and __per_cpu_offset). By putting __per_cpu_offset into the paca we can avoid the 2 loads associated with it: ld 11,56(13) /* paca->data_offset */ ld 9,.LC59-.LCTOC1(30) /* per_cpu__variable_name */ ldx 0,9,11 /* load per cpu data Longer term we can should be able to do even better than 3 loads. If per_cpu__variable_name wasnt a 64bit constant and paca->data_offset was in a register we could cut it down to one load. A suggestion from Rusty is to use gcc's __thread extension here. In order to do this we would need to free up r13 (the __thread register and where the paca currently is). So far Ive had a few unsuccessful attempts at doing that :) The patch also allocates per cpu memory node local on NUMA machines. This patch from Rusty has been sitting in my queue _forever_ but stalled when I hit the compiler bug. Sorry about that. Finally I also only allocate per cpu data for possible cpus, which comes straight out of the x86-64 port. On a pseries kernel (with NR_CPUS == 128) and 4 possible cpus we see some nice gains: total used free shared buffers cached Mem: 4012228 212860 3799368 0 0 162424 total used free shared buffers cached Mem: 4016200 212984 3803216 0 0 162424 A saving of 3.75MB. Quite nice for smaller machines. Note: we now have to be careful of per cpu users that touch data for !possible cpus. At this stage it might be worth making the NUMA and possible cpu optimisations generic, but per cpu init is done so early we have to be careful that all architectures have their possible map setup correctly. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Paul Mackerras <paulus@samba.org>
2006-01-11 02:16:44 +00:00
static int pcpu_cpu_distance(unsigned int from, unsigned int to)
{
if (cpu_to_node(from) == cpu_to_node(to))
return LOCAL_DISTANCE;
else
return REMOTE_DISTANCE;
}
powerpc: Optimise per cpu accesses on 64bit Now we dynamically allocate the paca array, it takes an extra load whenever we want to access another cpu's paca. One place we do that a lot is per cpu variables. A simple example: DEFINE_PER_CPU(unsigned long, vara); unsigned long test4(int cpu) { return per_cpu(vara, cpu); } This takes 4 loads, 5 if you include the actual load of the per cpu variable: ld r11,-32760(r30) # load address of paca pointer ld r9,-32768(r30) # load link address of percpu variable sldi r3,r29,9 # get offset into paca (each entry is 512 bytes) ld r0,0(r11) # load paca pointer add r3,r0,r3 # paca + offset ld r11,64(r3) # load paca[cpu].data_offset ldx r3,r9,r11 # load per cpu variable If we remove the ppc64 specific per_cpu_offset(), we get the generic one which indexes into a statically allocated array. This removes one load and one add: ld r11,-32760(r30) # load address of __per_cpu_offset ld r9,-32768(r30) # load link address of percpu variable sldi r3,r29,3 # get offset into __per_cpu_offset (each entry 8 bytes) ldx r11,r11,r3 # load __per_cpu_offset[cpu] ldx r3,r9,r11 # load per cpu variable Having all the offsets in one array also helps when iterating over a per cpu variable across a number of cpus, such as in the scheduler. Before we would need to load one paca cacheline when calculating each per cpu offset. Now we have 16 (128 / sizeof(long)) per cpu offsets in each cacheline. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2010-05-31 18:45:11 +00:00
unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
EXPORT_SYMBOL(__per_cpu_offset);
void __init setup_per_cpu_areas(void)
{
const size_t dyn_size = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
size_t atom_size;
unsigned long delta;
unsigned int cpu;
int rc;
/*
* Linear mapping is one of 4K, 1M and 16M. For 4K, no need
* to group units. For larger mappings, use 1M atom which
* should be large enough to contain a number of units.
*/
if (mmu_linear_psize == MMU_PAGE_4K)
atom_size = PAGE_SIZE;
else
atom_size = 1 << 20;
rc = pcpu_embed_first_chunk(0, dyn_size, atom_size, pcpu_cpu_distance,
pcpu_fc_alloc, pcpu_fc_free);
if (rc < 0)
panic("cannot initialize percpu area (err=%d)", rc);
delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
powerpc: Optimise per cpu accesses on 64bit Now we dynamically allocate the paca array, it takes an extra load whenever we want to access another cpu's paca. One place we do that a lot is per cpu variables. A simple example: DEFINE_PER_CPU(unsigned long, vara); unsigned long test4(int cpu) { return per_cpu(vara, cpu); } This takes 4 loads, 5 if you include the actual load of the per cpu variable: ld r11,-32760(r30) # load address of paca pointer ld r9,-32768(r30) # load link address of percpu variable sldi r3,r29,9 # get offset into paca (each entry is 512 bytes) ld r0,0(r11) # load paca pointer add r3,r0,r3 # paca + offset ld r11,64(r3) # load paca[cpu].data_offset ldx r3,r9,r11 # load per cpu variable If we remove the ppc64 specific per_cpu_offset(), we get the generic one which indexes into a statically allocated array. This removes one load and one add: ld r11,-32760(r30) # load address of __per_cpu_offset ld r9,-32768(r30) # load link address of percpu variable sldi r3,r29,3 # get offset into __per_cpu_offset (each entry 8 bytes) ldx r11,r11,r3 # load __per_cpu_offset[cpu] ldx r3,r9,r11 # load per cpu variable Having all the offsets in one array also helps when iterating over a per cpu variable across a number of cpus, such as in the scheduler. Before we would need to load one paca cacheline when calculating each per cpu offset. Now we have 16 (128 / sizeof(long)) per cpu offsets in each cacheline. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2010-05-31 18:45:11 +00:00
for_each_possible_cpu(cpu) {
__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
paca[cpu].data_offset = __per_cpu_offset[cpu];
}
[PATCH] powerpc/64: per cpu data optimisations The current ppc64 per cpu data implementation is quite slow. eg: lhz 11,18(13) /* smp_processor_id() */ ld 9,.LC63-.LCTOC1(30) /* per_cpu__variable_name */ ld 8,.LC61-.LCTOC1(30) /* __per_cpu_offset */ sldi 11,11,3 /* form index into __per_cpu_offset */ mr 10,9 ldx 9,11,8 /* __per_cpu_offset[smp_processor_id()] */ ldx 0,10,9 /* load per cpu data */ 5 loads for something that is supposed to be fast, pretty awful. One reason for the large number of loads is that we have to synthesize 2 64bit constants (per_cpu__variable_name and __per_cpu_offset). By putting __per_cpu_offset into the paca we can avoid the 2 loads associated with it: ld 11,56(13) /* paca->data_offset */ ld 9,.LC59-.LCTOC1(30) /* per_cpu__variable_name */ ldx 0,9,11 /* load per cpu data Longer term we can should be able to do even better than 3 loads. If per_cpu__variable_name wasnt a 64bit constant and paca->data_offset was in a register we could cut it down to one load. A suggestion from Rusty is to use gcc's __thread extension here. In order to do this we would need to free up r13 (the __thread register and where the paca currently is). So far Ive had a few unsuccessful attempts at doing that :) The patch also allocates per cpu memory node local on NUMA machines. This patch from Rusty has been sitting in my queue _forever_ but stalled when I hit the compiler bug. Sorry about that. Finally I also only allocate per cpu data for possible cpus, which comes straight out of the x86-64 port. On a pseries kernel (with NR_CPUS == 128) and 4 possible cpus we see some nice gains: total used free shared buffers cached Mem: 4012228 212860 3799368 0 0 162424 total used free shared buffers cached Mem: 4016200 212984 3803216 0 0 162424 A saving of 3.75MB. Quite nice for smaller machines. Note: we now have to be careful of per cpu users that touch data for !possible cpus. At this stage it might be worth making the NUMA and possible cpu optimisations generic, but per cpu init is done so early we have to be careful that all architectures have their possible map setup correctly. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Paul Mackerras <paulus@samba.org>
2006-01-11 02:16:44 +00:00
}
#endif
[POWERPC] Allow hooking of PCI MMIO & PIO accessors on 64 bits This patch reworks the way iSeries hooks on PCI IO operations (both MMIO and PIO) and provides a generic way for other platforms to do so (we have need to do that for various other platforms). While reworking the IO ops, I ended up doing some spring cleaning in io.h and eeh.h which I might want to split into 2 or 3 patches (among others, eeh.h had a lot of useless stuff in it). A side effect is that EEH for PIO should work now (it used to pass IO ports down to the eeh address check functions which is bogus). Also, new are MMIO "repeat" ops, which other archs like ARM already had, and that we have too now: readsb, readsw, readsl, writesb, writesw, writesl. In the long run, I might also make EEH use the hooks instead of wrapping at the toplevel, which would make things even cleaner and relegate EEH completely in platforms/iseries, but we have to measure the performance impact there (though it's really only on MMIO reads) Since I also need to hook on ioremap, I shuffled the functions a bit there. I introduced ioremap_flags() to use by drivers who want to pass explicit flags to ioremap (and it can be hooked). The old __ioremap() is still there as a low level and cannot be hooked, thus drivers who use it should migrate unless they know they want the low level version. The patch "arch provides generic iomap missing accessors" (should be number 4 in this series) is a pre-requisite to provide full iomap API support with this patch. Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> Signed-off-by: Paul Mackerras <paulus@samba.org>
2006-11-11 06:25:10 +00:00
#ifdef CONFIG_PPC_INDIRECT_IO
struct ppc_pci_io ppc_pci_io;
EXPORT_SYMBOL(ppc_pci_io);
#endif /* CONFIG_PPC_INDIRECT_IO */