2005-10-10 20:50:37 +08:00
|
|
|
/*
|
|
|
|
*
|
|
|
|
* 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.
|
|
|
|
*/
|
|
|
|
|
2013-07-25 10:12:32 +08:00
|
|
|
#define DEBUG
|
2005-10-10 20:50:37 +08:00
|
|
|
|
2011-07-23 06:24:23 +08:00
|
|
|
#include <linux/export.h>
|
2005-10-10 20:50:37 +08:00
|
|
|
#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 10:16:44 +08:00
|
|
|
#include <linux/bootmem.h>
|
2006-11-11 14:25:02 +08:00
|
|
|
#include <linux/pci.h>
|
2008-04-17 12:35:01 +08:00
|
|
|
#include <linux/lockdep.h>
|
2010-07-12 12:36:09 +08:00
|
|
|
#include <linux/memblock.h>
|
2014-06-04 15:50:47 +08:00
|
|
|
#include <linux/memory.h>
|
2015-04-09 10:52:56 +08:00
|
|
|
#include <linux/nmi.h>
|
2011-10-10 18:50:43 +08:00
|
|
|
|
2005-10-10 20:50:37 +08:00
|
|
|
#include <asm/io.h>
|
2005-12-04 15:39:37 +08:00
|
|
|
#include <asm/kdump.h>
|
2005-10-10 20:50:37 +08:00
|
|
|
#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>
|
2017-05-09 11:16:52 +08:00
|
|
|
#include <asm/dt_cpu_ftrs.h>
|
2005-10-10 20:50:37 +08:00
|
|
|
#include <asm/sections.h>
|
|
|
|
#include <asm/btext.h>
|
|
|
|
#include <asm/nvram.h>
|
|
|
|
#include <asm/setup.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>
|
2005-10-28 20:53:37 +08:00
|
|
|
#include <asm/xmon.h>
|
2005-11-07 06:49:43 +08:00
|
|
|
#include <asm/udbg.h>
|
2005-11-11 21:06:06 +08:00
|
|
|
#include <asm/kexec.h>
|
2011-04-06 13:18:48 +08:00
|
|
|
#include <asm/code-patching.h>
|
2016-03-24 19:04:04 +08:00
|
|
|
#include <asm/livepatch.h>
|
2016-07-05 13:03:49 +08:00
|
|
|
#include <asm/opal.h>
|
2016-07-05 13:07:51 +08:00
|
|
|
#include <asm/cputhreads.h>
|
2005-10-10 20:50:37 +08:00
|
|
|
|
2017-10-24 19:44:44 +08:00
|
|
|
#include "setup.h"
|
|
|
|
|
2005-10-10 20:50:37 +08:00
|
|
|
#ifdef DEBUG
|
|
|
|
#define DBG(fmt...) udbg_printf(fmt)
|
|
|
|
#else
|
|
|
|
#define DBG(fmt...)
|
|
|
|
#endif
|
|
|
|
|
2013-03-20 14:30:12 +08:00
|
|
|
int spinning_secondaries;
|
2005-10-10 20:50:37 +08:00
|
|
|
u64 ppc64_pft_size;
|
|
|
|
|
2005-12-09 09:40:17 +08:00
|
|
|
struct ppc64_caches ppc64_caches = {
|
2017-01-09 07:31:47 +08:00
|
|
|
.l1d = {
|
|
|
|
.block_size = 0x40,
|
|
|
|
.log_block_size = 6,
|
|
|
|
},
|
|
|
|
.l1i = {
|
|
|
|
.block_size = 0x40,
|
|
|
|
.log_block_size = 6
|
|
|
|
},
|
2005-12-09 09:40:17 +08:00
|
|
|
};
|
2005-10-10 20:50:37 +08:00
|
|
|
EXPORT_SYMBOL_GPL(ppc64_caches);
|
|
|
|
|
2013-10-12 08:22:38 +08:00
|
|
|
#if defined(CONFIG_PPC_BOOK3E) && defined(CONFIG_SMP)
|
2016-07-05 13:07:51 +08:00
|
|
|
void __init setup_tlb_core_data(void)
|
2013-10-12 08:22:38 +08:00
|
|
|
{
|
|
|
|
int cpu;
|
|
|
|
|
2014-03-08 04:48:35 +08:00
|
|
|
BUILD_BUG_ON(offsetof(struct tlb_core_data, lock) != 0);
|
|
|
|
|
2013-10-12 08:22:38 +08:00
|
|
|
for_each_possible_cpu(cpu) {
|
|
|
|
int first = cpu_first_thread_sibling(cpu);
|
|
|
|
|
2015-10-07 11:48:09 +08:00
|
|
|
/*
|
|
|
|
* If we boot via kdump on a non-primary thread,
|
|
|
|
* make sure we point at the thread that actually
|
|
|
|
* set up this TLB.
|
|
|
|
*/
|
|
|
|
if (cpu_first_thread_sibling(boot_cpuid) == first)
|
|
|
|
first = boot_cpuid;
|
|
|
|
|
2013-10-12 08:22:38 +08:00
|
|
|
paca[cpu].tcd_ptr = &paca[first].tcd;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If we have threads, we need either tlbsrx.
|
|
|
|
* or e6500 tablewalk mode, or else TLB handlers
|
|
|
|
* will be racy and could produce duplicate entries.
|
2017-02-15 17:24:25 +08:00
|
|
|
* Should we panic instead?
|
2013-10-12 08:22:38 +08:00
|
|
|
*/
|
2017-02-15 17:24:25 +08:00
|
|
|
WARN_ONCE(smt_enabled_at_boot >= 2 &&
|
|
|
|
!mmu_has_feature(MMU_FTR_USE_TLBRSRV) &&
|
|
|
|
book3e_htw_mode != PPC_HTW_E6500,
|
|
|
|
"%s: unsupported MMU configuration\n", __func__);
|
2013-10-12 08:22:38 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2005-10-10 20:50:37 +08:00
|
|
|
#ifdef CONFIG_SMP
|
|
|
|
|
2010-08-05 15:42:11 +08:00
|
|
|
static char *smt_enabled_cmdline;
|
2005-10-10 20:50:37 +08:00
|
|
|
|
|
|
|
/* Look for ibm,smt-enabled OF option */
|
2016-07-05 13:07:51 +08:00
|
|
|
void __init check_smt_enabled(void)
|
2005-10-10 20:50:37 +08:00
|
|
|
{
|
|
|
|
struct device_node *dn;
|
2006-07-12 13:35:54 +08:00
|
|
|
const char *smt_option;
|
2005-10-10 20:50:37 +08:00
|
|
|
|
2010-08-05 15:42:11 +08:00
|
|
|
/* Default to enabling all threads */
|
|
|
|
smt_enabled_at_boot = threads_per_core;
|
2005-10-10 20:50:37 +08:00
|
|
|
|
2010-08-05 15:42:11 +08:00
|
|
|
/* 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 {
|
2014-08-09 05:24:01 +08:00
|
|
|
int smt;
|
2010-08-05 15:42:11 +08:00
|
|
|
int rc;
|
|
|
|
|
2014-08-09 05:24:01 +08:00
|
|
|
rc = kstrtoint(smt_enabled_cmdline, 10, &smt);
|
2010-08-05 15:42:11 +08:00
|
|
|
if (!rc)
|
|
|
|
smt_enabled_at_boot =
|
2014-08-09 05:24:01 +08:00
|
|
|
min(threads_per_core, smt);
|
2010-08-05 15:42:11 +08:00
|
|
|
}
|
|
|
|
} 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);
|
|
|
|
}
|
|
|
|
}
|
2005-10-10 20:50:37 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/* Look for smt-enabled= cmdline option */
|
|
|
|
static int __init early_smt_enabled(char *p)
|
|
|
|
{
|
2010-08-05 15:42:11 +08:00
|
|
|
smt_enabled_cmdline = p;
|
2005-10-10 20:50:37 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
early_param("smt-enabled", early_smt_enabled);
|
|
|
|
|
|
|
|
#endif /* CONFIG_SMP */
|
|
|
|
|
2013-02-12 22:44:50 +08:00
|
|
|
/** Fix up paca fields required for the boot cpu */
|
2016-07-05 13:07:50 +08:00
|
|
|
static void __init fixup_boot_paca(void)
|
2013-02-12 22:44:50 +08:00
|
|
|
{
|
|
|
|
/* The boot cpu is started */
|
|
|
|
get_paca()->cpu_start = 1;
|
|
|
|
/* Allow percpu accesses to work until we setup percpu data */
|
|
|
|
get_paca()->data_offset = 0;
|
|
|
|
}
|
|
|
|
|
2016-07-05 13:07:50 +08:00
|
|
|
static void __init configure_exceptions(void)
|
2014-03-28 10:36:30 +08:00
|
|
|
{
|
2014-07-17 13:29:45 +08:00
|
|
|
/*
|
2016-07-05 13:03:49 +08:00
|
|
|
* Setup the trampolines from the lowmem exception vectors
|
|
|
|
* to the kdump kernel when not using a relocatable kernel.
|
2014-07-17 13:29:45 +08:00
|
|
|
*/
|
2016-07-05 13:03:49 +08:00
|
|
|
setup_kdump_trampoline();
|
|
|
|
|
|
|
|
/* Under a PAPR hypervisor, we need hypercalls */
|
|
|
|
if (firmware_has_feature(FW_FEATURE_SET_MODE)) {
|
|
|
|
/* Enable AIL if possible */
|
|
|
|
pseries_enable_reloc_on_exc();
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Tell the hypervisor that we want our exceptions to
|
|
|
|
* be taken in little endian mode.
|
|
|
|
*
|
|
|
|
* We don't call this for big endian as our calling convention
|
|
|
|
* makes us always enter in BE, and the call may fail under
|
|
|
|
* some circumstances with kdump.
|
|
|
|
*/
|
|
|
|
#ifdef __LITTLE_ENDIAN__
|
|
|
|
pseries_little_endian_exceptions();
|
|
|
|
#endif
|
|
|
|
} else {
|
|
|
|
/* Set endian mode using OPAL */
|
|
|
|
if (firmware_has_feature(FW_FEATURE_OPAL))
|
|
|
|
opal_configure_cores();
|
|
|
|
|
2016-11-15 12:28:33 +08:00
|
|
|
/* AIL on native is done in cpu_ready_for_interrupts() */
|
2014-03-28 10:36:30 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-07-05 13:03:49 +08:00
|
|
|
static void cpu_ready_for_interrupts(void)
|
|
|
|
{
|
2016-11-15 12:28:33 +08:00
|
|
|
/*
|
|
|
|
* Enable AIL if supported, and we are in hypervisor mode. This
|
|
|
|
* is called once for every processor.
|
|
|
|
*
|
|
|
|
* If we are not in hypervisor mode the job is done once for
|
|
|
|
* the whole partition in configure_exceptions().
|
|
|
|
*/
|
2017-03-21 13:24:38 +08:00
|
|
|
if (cpu_has_feature(CPU_FTR_HVMODE) &&
|
|
|
|
cpu_has_feature(CPU_FTR_ARCH_207S)) {
|
2016-11-15 12:28:33 +08:00
|
|
|
unsigned long lpcr = mfspr(SPRN_LPCR);
|
|
|
|
mtspr(SPRN_LPCR, lpcr | LPCR_AIL_3);
|
|
|
|
}
|
|
|
|
|
powerpc: Disable HFSCR[TM] if TM is not supported
On Power8 & Power9 the early CPU inititialisation in __init_HFSCR()
turns on HFSCR[TM] (Hypervisor Facility Status and Control Register
[Transactional Memory]), but that doesn't take into account that TM
might be disabled by CPU features, or disabled by the kernel being built
with CONFIG_PPC_TRANSACTIONAL_MEM=n.
So later in boot, when we have setup the CPU features, clear HSCR[TM] if
the TM CPU feature has been disabled. We use CPU_FTR_TM_COMP to account
for the CONFIG_PPC_TRANSACTIONAL_MEM=n case.
Without this a KVM guest might try use TM, even if told not to, and
cause an oops in the host kernel. Typically the oops is seen in
__kvmppc_vcore_entry() and may or may not be fatal to the host, but is
always bad news.
In practice all shipping CPU revisions do support TM, and all host
kernels we are aware of build with TM support enabled, so no one should
actually be able to hit this in the wild.
Fixes: 2a3563b023e5 ("powerpc: Setup in HFSCR for POWER8")
Cc: stable@vger.kernel.org # v3.10+
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Tested-by: Sam Bobroff <sam.bobroff@au1.ibm.com>
[mpe: Rewrite change log with input from Sam, add Fixes/stable]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-03-20 14:49:03 +08:00
|
|
|
/*
|
|
|
|
* Fixup HFSCR:TM based on CPU features. The bit is set by our
|
|
|
|
* early asm init because at that point we haven't updated our
|
|
|
|
* CPU features from firmware and device-tree. Here we have,
|
|
|
|
* so let's do it.
|
|
|
|
*/
|
|
|
|
if (cpu_has_feature(CPU_FTR_HVMODE) && !cpu_has_feature(CPU_FTR_TM_COMP))
|
|
|
|
mtspr(SPRN_HFSCR, mfspr(SPRN_HFSCR) & ~HFSCR_TM);
|
|
|
|
|
2016-07-05 13:03:49 +08:00
|
|
|
/* Set IR and DR in PACA MSR */
|
|
|
|
get_paca()->kernel_msr = MSR_KERNEL;
|
|
|
|
}
|
|
|
|
|
2005-10-10 20:50:37 +08:00
|
|
|
/*
|
|
|
|
* 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
|
2010-07-12 12:36:09 +08:00
|
|
|
* some early parsing of the device-tree to setup out MEMBLOCK
|
2005-10-10 20:50:37 +08:00
|
|
|
* 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)
|
|
|
|
{
|
2013-02-14 01:03:16 +08:00
|
|
|
static __initdata struct paca_struct boot_paca;
|
|
|
|
|
2008-05-07 08:00:56 +08:00
|
|
|
/* -------- printk is _NOT_ safe to use here ! ------- */
|
|
|
|
|
2017-05-09 11:16:52 +08:00
|
|
|
/* Try new device tree based feature discovery ... */
|
|
|
|
if (!dt_cpu_ftrs_init(__va(dt_ptr)))
|
|
|
|
/* Otherwise use the old style CPU table */
|
|
|
|
identify_cpu(0, mfspr(SPRN_PVR));
|
2006-10-24 14:42:40 +08:00
|
|
|
|
2006-06-28 11:18:53 +08:00
|
|
|
/* Assume we're on cpu 0 for now. Don't write to the paca yet! */
|
2010-01-28 21:23:22 +08:00
|
|
|
initialise_paca(&boot_paca, 0);
|
|
|
|
setup_paca(&boot_paca);
|
2013-02-12 22:44:50 +08:00
|
|
|
fixup_boot_paca();
|
2006-06-28 11:18:53 +08:00
|
|
|
|
2008-05-07 08:00:56 +08:00
|
|
|
/* -------- printk is now safe to use ------- */
|
|
|
|
|
2008-05-07 08:25:34 +08:00
|
|
|
/* Enable early debugging if any specified (see udbg.h) */
|
|
|
|
udbg_early_init();
|
|
|
|
|
2006-03-28 20:15:54 +08:00
|
|
|
DBG(" -> early_setup(), dt_ptr: 0x%lx\n", dt_ptr);
|
2005-10-10 20:50:37 +08:00
|
|
|
|
|
|
|
/*
|
2007-09-07 01:47:29 +08:00
|
|
|
* Do early initialization using the flattened device
|
|
|
|
* tree, such as retrieving the physical memory map or
|
|
|
|
* calculating/retrieving the hash table size.
|
2005-10-10 20:50:37 +08:00
|
|
|
*/
|
|
|
|
early_init_devtree(__va(dt_ptr));
|
|
|
|
|
2006-03-25 14:25:17 +08:00
|
|
|
/* Now we know the logical id of our boot cpu, setup the paca. */
|
2010-01-28 21:23:22 +08:00
|
|
|
setup_paca(&paca[boot_cpuid]);
|
2013-02-12 22:44:50 +08:00
|
|
|
fixup_boot_paca();
|
2006-03-25 14:25:17 +08:00
|
|
|
|
2016-07-05 13:03:46 +08:00
|
|
|
/*
|
2016-07-05 13:03:49 +08:00
|
|
|
* Configure exception handlers. This include setting up trampolines
|
|
|
|
* if needed, setting exception endian mode, etc...
|
2016-07-05 13:03:46 +08:00
|
|
|
*/
|
2016-07-05 13:03:49 +08:00
|
|
|
configure_exceptions();
|
2005-12-04 15:39:37 +08:00
|
|
|
|
2016-07-05 13:03:42 +08:00
|
|
|
/* Apply all the dynamic patching */
|
|
|
|
apply_feature_fixups();
|
2016-08-10 15:27:34 +08:00
|
|
|
setup_feature_keys();
|
2016-07-05 13:03:42 +08:00
|
|
|
|
2016-07-26 19:55:48 +08:00
|
|
|
/* Initialize the hash table or TLB handling */
|
|
|
|
early_init_mmu();
|
|
|
|
|
2017-10-24 19:44:44 +08:00
|
|
|
/*
|
|
|
|
* After firmware and early platform setup code has set things up,
|
|
|
|
* we note the SPR values for configurable control/performance
|
|
|
|
* registers, and use those as initial defaults.
|
|
|
|
*/
|
|
|
|
record_spr_defaults();
|
|
|
|
|
2014-03-28 10:36:29 +08:00
|
|
|
/*
|
|
|
|
* At this point, we can let interrupts switch to virtual mode
|
|
|
|
* (the MMU has been setup), so adjust the MSR in the PACA to
|
2014-03-28 10:36:30 +08:00
|
|
|
* have IR and DR set and enable AIL if it exists
|
2014-03-28 10:36:29 +08:00
|
|
|
*/
|
2014-03-28 10:36:30 +08:00
|
|
|
cpu_ready_for_interrupts();
|
2014-03-28 10:36:29 +08:00
|
|
|
|
2005-10-10 20:50:37 +08:00
|
|
|
DBG(" <- early_setup()\n");
|
2013-07-25 10:12:32 +08:00
|
|
|
|
|
|
|
#ifdef CONFIG_PPC_EARLY_DEBUG_BOOTX
|
|
|
|
/*
|
|
|
|
* This needs to be done *last* (after the above DBG() even)
|
|
|
|
*
|
|
|
|
* Right after we return from this function, we turn on the MMU
|
|
|
|
* which means the real-mode access trick that btext does will
|
|
|
|
* no longer work, it needs to switch to using a real MMU
|
|
|
|
* mapping. This call will ensure that it does
|
|
|
|
*/
|
|
|
|
btext_map();
|
|
|
|
#endif /* CONFIG_PPC_EARLY_DEBUG_BOOTX */
|
2005-10-10 20:50:37 +08:00
|
|
|
}
|
|
|
|
|
2005-11-10 10:37:51 +08:00
|
|
|
#ifdef CONFIG_SMP
|
|
|
|
void early_setup_secondary(void)
|
|
|
|
{
|
2016-03-04 13:01:48 +08:00
|
|
|
/* Mark interrupts disabled in PACA */
|
2009-03-20 03:34:16 +08:00
|
|
|
get_paca()->soft_enabled = 0;
|
2005-11-10 10:37:51 +08:00
|
|
|
|
2009-03-20 03:34:16 +08:00
|
|
|
/* Initialize the hash table or TLB handling */
|
|
|
|
early_init_mmu_secondary();
|
2014-03-28 10:36:29 +08:00
|
|
|
|
|
|
|
/*
|
|
|
|
* At this point, we can let interrupts switch to virtual mode
|
|
|
|
* (the MMU has been setup), so adjust the MSR in the PACA to
|
|
|
|
* have IR and DR set.
|
|
|
|
*/
|
2014-03-28 10:36:30 +08:00
|
|
|
cpu_ready_for_interrupts();
|
2005-11-10 10:37:51 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
#endif /* CONFIG_SMP */
|
2005-10-10 20:50:37 +08:00
|
|
|
|
2016-11-29 20:45:50 +08:00
|
|
|
#if defined(CONFIG_SMP) || defined(CONFIG_KEXEC_CORE)
|
2015-10-07 11:48:19 +08:00
|
|
|
static bool use_spinloop(void)
|
|
|
|
{
|
2017-10-23 16:05:07 +08:00
|
|
|
if (IS_ENABLED(CONFIG_PPC_BOOK3S)) {
|
|
|
|
/*
|
|
|
|
* See comments in head_64.S -- not all platforms insert
|
|
|
|
* secondaries at __secondary_hold and wait at the spin
|
|
|
|
* loop.
|
|
|
|
*/
|
|
|
|
if (firmware_has_feature(FW_FEATURE_OPAL))
|
|
|
|
return false;
|
2015-10-07 11:48:19 +08:00
|
|
|
return true;
|
2017-10-23 16:05:07 +08:00
|
|
|
}
|
2015-10-07 11:48:19 +08:00
|
|
|
|
|
|
|
/*
|
|
|
|
* When book3e boots from kexec, the ePAPR spin table does
|
|
|
|
* not get used.
|
|
|
|
*/
|
|
|
|
return of_property_read_bool(of_chosen, "linux,booted-from-kexec");
|
|
|
|
}
|
|
|
|
|
2005-11-04 09:09:42 +08:00
|
|
|
void smp_release_cpus(void)
|
|
|
|
{
|
2005-12-06 05:49:00 +08:00
|
|
|
unsigned long *ptr;
|
2011-03-16 11:54:35 +08:00
|
|
|
int i;
|
2005-11-04 09:09:42 +08:00
|
|
|
|
2015-10-07 11:48:19 +08:00
|
|
|
if (!use_spinloop())
|
|
|
|
return;
|
|
|
|
|
2005-11-04 09:09:42 +08:00
|
|
|
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.
|
2008-08-30 09:40:24 +08:00
|
|
|
*/
|
2005-11-04 09:09:42 +08:00
|
|
|
|
2005-12-06 05:49:00 +08:00
|
|
|
ptr = (unsigned long *)((unsigned long)&__secondary_hold_spinloop
|
|
|
|
- PHYSICAL_START);
|
2014-03-11 08:54:06 +08:00
|
|
|
*ptr = ppc_function_entry(generic_secondary_smp_init);
|
2011-03-16 11:54:35 +08:00
|
|
|
|
|
|
|
/* And wait a bit for them to catch up */
|
|
|
|
for (i = 0; i < 100000; i++) {
|
|
|
|
mb();
|
|
|
|
HMT_low();
|
2011-05-26 02:09:12 +08:00
|
|
|
if (spinning_secondaries == 0)
|
2011-03-16 11:54:35 +08:00
|
|
|
break;
|
|
|
|
udelay(1);
|
|
|
|
}
|
2011-05-26 02:09:12 +08:00
|
|
|
DBG("spinning_secondaries = %d\n", spinning_secondaries);
|
2005-11-04 09:09:42 +08:00
|
|
|
|
|
|
|
DBG(" <- smp_release_cpus()\n");
|
|
|
|
}
|
2016-11-29 20:45:50 +08:00
|
|
|
#endif /* CONFIG_SMP || CONFIG_KEXEC_CORE */
|
2005-11-04 09:09:42 +08:00
|
|
|
|
2005-10-10 20:50:37 +08:00
|
|
|
/*
|
2005-11-10 10:37:51 +08:00
|
|
|
* Initialize some remaining members of the ppc64_caches and systemcfg
|
|
|
|
* structures
|
2005-10-10 20:50:37 +08:00
|
|
|
* (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
|
|
|
|
*/
|
2017-01-09 07:31:47 +08:00
|
|
|
|
|
|
|
static void init_cache_info(struct ppc_cache_info *info, u32 size, u32 lsize,
|
|
|
|
u32 bsize, u32 sets)
|
|
|
|
{
|
|
|
|
info->size = size;
|
|
|
|
info->sets = sets;
|
|
|
|
info->line_size = lsize;
|
|
|
|
info->block_size = bsize;
|
|
|
|
info->log_block_size = __ilog2(bsize);
|
2017-03-05 07:54:34 +08:00
|
|
|
if (bsize)
|
|
|
|
info->blocks_per_page = PAGE_SIZE / bsize;
|
|
|
|
else
|
|
|
|
info->blocks_per_page = 0;
|
2017-02-03 14:20:07 +08:00
|
|
|
|
|
|
|
if (sets == 0)
|
|
|
|
info->assoc = 0xffff;
|
|
|
|
else
|
|
|
|
info->assoc = size / (sets * lsize);
|
2017-01-09 07:31:47 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static bool __init parse_cache_info(struct device_node *np,
|
|
|
|
bool icache,
|
|
|
|
struct ppc_cache_info *info)
|
|
|
|
{
|
|
|
|
static const char *ipropnames[] __initdata = {
|
|
|
|
"i-cache-size",
|
|
|
|
"i-cache-sets",
|
|
|
|
"i-cache-block-size",
|
|
|
|
"i-cache-line-size",
|
|
|
|
};
|
|
|
|
static const char *dpropnames[] __initdata = {
|
|
|
|
"d-cache-size",
|
|
|
|
"d-cache-sets",
|
|
|
|
"d-cache-block-size",
|
|
|
|
"d-cache-line-size",
|
|
|
|
};
|
|
|
|
const char **propnames = icache ? ipropnames : dpropnames;
|
|
|
|
const __be32 *sizep, *lsizep, *bsizep, *setsp;
|
|
|
|
u32 size, lsize, bsize, sets;
|
|
|
|
bool success = true;
|
|
|
|
|
|
|
|
size = 0;
|
|
|
|
sets = -1u;
|
|
|
|
lsize = bsize = cur_cpu_spec->dcache_bsize;
|
|
|
|
sizep = of_get_property(np, propnames[0], NULL);
|
|
|
|
if (sizep != NULL)
|
|
|
|
size = be32_to_cpu(*sizep);
|
|
|
|
setsp = of_get_property(np, propnames[1], NULL);
|
|
|
|
if (setsp != NULL)
|
|
|
|
sets = be32_to_cpu(*setsp);
|
|
|
|
bsizep = of_get_property(np, propnames[2], NULL);
|
|
|
|
lsizep = of_get_property(np, propnames[3], NULL);
|
|
|
|
if (bsizep == NULL)
|
|
|
|
bsizep = lsizep;
|
|
|
|
if (lsizep != NULL)
|
|
|
|
lsize = be32_to_cpu(*lsizep);
|
|
|
|
if (bsizep != NULL)
|
|
|
|
bsize = be32_to_cpu(*bsizep);
|
|
|
|
if (sizep == NULL || bsizep == NULL || lsizep == NULL)
|
|
|
|
success = false;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* OF is weird .. it represents fully associative caches
|
|
|
|
* as "1 way" which doesn't make much sense and doesn't
|
|
|
|
* leave room for direct mapped. We'll assume that 0
|
|
|
|
* in OF means direct mapped for that reason.
|
|
|
|
*/
|
|
|
|
if (sets == 1)
|
|
|
|
sets = 0;
|
|
|
|
else if (sets == 0)
|
|
|
|
sets = 1;
|
|
|
|
|
|
|
|
init_cache_info(info, size, lsize, bsize, sets);
|
|
|
|
|
|
|
|
return success;
|
|
|
|
}
|
|
|
|
|
2016-07-05 13:07:51 +08:00
|
|
|
void __init initialize_cache_info(void)
|
2005-10-10 20:50:37 +08:00
|
|
|
{
|
2017-01-09 07:31:49 +08:00
|
|
|
struct device_node *cpu = NULL, *l2, *l3 = NULL;
|
|
|
|
u32 pvr;
|
2005-10-10 20:50:37 +08:00
|
|
|
|
|
|
|
DBG(" -> initialize_cache_info()\n");
|
|
|
|
|
2017-01-09 07:31:49 +08:00
|
|
|
/*
|
|
|
|
* All shipping POWER8 machines have a firmware bug that
|
|
|
|
* puts incorrect information in the device-tree. This will
|
|
|
|
* be (hopefully) fixed for future chips but for now hard
|
|
|
|
* code the values if we are running on one of these
|
|
|
|
*/
|
|
|
|
pvr = PVR_VER(mfspr(SPRN_PVR));
|
|
|
|
if (pvr == PVR_POWER8 || pvr == PVR_POWER8E ||
|
|
|
|
pvr == PVR_POWER8NVL) {
|
|
|
|
/* size lsize blk sets */
|
|
|
|
init_cache_info(&ppc64_caches.l1i, 0x8000, 128, 128, 32);
|
|
|
|
init_cache_info(&ppc64_caches.l1d, 0x10000, 128, 128, 64);
|
|
|
|
init_cache_info(&ppc64_caches.l2, 0x80000, 128, 0, 512);
|
|
|
|
init_cache_info(&ppc64_caches.l3, 0x800000, 128, 0, 8192);
|
|
|
|
} else
|
|
|
|
cpu = of_find_node_by_type(NULL, "cpu");
|
2005-10-10 20:50:37 +08:00
|
|
|
|
2017-01-09 07:31:47 +08:00
|
|
|
/*
|
|
|
|
* We're assuming *all* of the CPUs have the same
|
|
|
|
* d-cache and i-cache sizes... -Peter
|
|
|
|
*/
|
2017-01-09 07:31:48 +08:00
|
|
|
if (cpu) {
|
|
|
|
if (!parse_cache_info(cpu, false, &ppc64_caches.l1d))
|
2017-01-09 07:31:47 +08:00
|
|
|
DBG("Argh, can't find dcache properties !\n");
|
|
|
|
|
2017-01-09 07:31:48 +08:00
|
|
|
if (!parse_cache_info(cpu, true, &ppc64_caches.l1i))
|
2017-01-09 07:31:47 +08:00
|
|
|
DBG("Argh, can't find icache properties !\n");
|
2017-01-09 07:31:48 +08:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Try to find the L2 and L3 if any. Assume they are
|
|
|
|
* unified and use the D-side properties.
|
|
|
|
*/
|
|
|
|
l2 = of_find_next_cache_node(cpu);
|
|
|
|
of_node_put(cpu);
|
|
|
|
if (l2) {
|
|
|
|
parse_cache_info(l2, false, &ppc64_caches.l2);
|
|
|
|
l3 = of_find_next_cache_node(l2);
|
|
|
|
of_node_put(l2);
|
|
|
|
}
|
|
|
|
if (l3) {
|
|
|
|
parse_cache_info(l3, false, &ppc64_caches.l3);
|
|
|
|
of_node_put(l3);
|
|
|
|
}
|
2005-10-10 20:50:37 +08:00
|
|
|
}
|
|
|
|
|
2016-07-05 13:04:08 +08:00
|
|
|
/* For use by binfmt_elf */
|
2017-01-09 07:31:47 +08:00
|
|
|
dcache_bsize = ppc64_caches.l1d.block_size;
|
|
|
|
icache_bsize = ppc64_caches.l1i.block_size;
|
2016-07-05 13:04:08 +08:00
|
|
|
|
2017-05-09 11:16:52 +08:00
|
|
|
cur_cpu_spec->dcache_bsize = dcache_bsize;
|
|
|
|
cur_cpu_spec->icache_bsize = icache_bsize;
|
|
|
|
|
2005-10-10 20:50:37 +08:00
|
|
|
DBG(" <- initialize_cache_info()\n");
|
|
|
|
}
|
|
|
|
|
2011-05-03 22:07:01 +08:00
|
|
|
/* 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.
|
|
|
|
*/
|
2016-07-05 13:07:50 +08:00
|
|
|
static __init u64 safe_stack_limit(void)
|
2010-05-11 02:59:18 +08:00
|
|
|
{
|
2011-05-03 22:07:01 +08:00
|
|
|
#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
|
2017-08-13 09:33:41 +08:00
|
|
|
if (early_radix_enabled())
|
|
|
|
return ULONG_MAX;
|
|
|
|
|
2011-05-03 22:07:01 +08:00
|
|
|
/* BookS, the first segment is bolted */
|
|
|
|
if (mmu_has_feature(MMU_FTR_1T_SEGMENT))
|
2010-05-11 02:59:18 +08:00
|
|
|
return 1UL << SID_SHIFT_1T;
|
|
|
|
return 1UL << SID_SHIFT;
|
2011-05-03 22:07:01 +08:00
|
|
|
#endif
|
2010-05-11 02:59:18 +08:00
|
|
|
}
|
|
|
|
|
2016-07-05 13:07:51 +08:00
|
|
|
void __init irqstack_early_init(void)
|
2005-10-10 20:50:37 +08:00
|
|
|
{
|
2011-05-03 22:07:01 +08:00
|
|
|
u64 limit = safe_stack_limit();
|
2005-10-10 20:50:37 +08:00
|
|
|
unsigned int i;
|
|
|
|
|
|
|
|
/*
|
2010-12-08 08:55:03 +08:00
|
|
|
* Interrupt stacks must be in the first segment since we
|
2017-08-13 09:33:41 +08:00
|
|
|
* cannot afford to take SLB misses on them. They are not
|
|
|
|
* accessed in realmode.
|
2005-10-10 20:50:37 +08:00
|
|
|
*/
|
2006-03-29 06:50:51 +08:00
|
|
|
for_each_possible_cpu(i) {
|
2005-11-07 08:06:55 +08:00
|
|
|
softirq_ctx[i] = (struct thread_info *)
|
2010-07-12 12:36:09 +08:00
|
|
|
__va(memblock_alloc_base(THREAD_SIZE,
|
2010-05-11 02:59:18 +08:00
|
|
|
THREAD_SIZE, limit));
|
2005-11-07 08:06:55 +08:00
|
|
|
hardirq_ctx[i] = (struct thread_info *)
|
2010-07-12 12:36:09 +08:00
|
|
|
__va(memblock_alloc_base(THREAD_SIZE,
|
2010-05-11 02:59:18 +08:00
|
|
|
THREAD_SIZE, limit));
|
2005-10-10 20:50:37 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2009-07-24 07:15:59 +08:00
|
|
|
#ifdef CONFIG_PPC_BOOK3E
|
2016-07-05 13:07:51 +08:00
|
|
|
void __init exc_lvl_early_init(void)
|
2009-07-24 07:15:59 +08:00
|
|
|
{
|
|
|
|
unsigned int i;
|
2013-10-23 17:31:21 +08:00
|
|
|
unsigned long sp;
|
2009-07-24 07:15:59 +08:00
|
|
|
|
|
|
|
for_each_possible_cpu(i) {
|
2013-10-23 17:31:21 +08:00
|
|
|
sp = memblock_alloc(THREAD_SIZE, THREAD_SIZE);
|
|
|
|
critirq_ctx[i] = (struct thread_info *)__va(sp);
|
|
|
|
paca[i].crit_kstack = __va(sp + THREAD_SIZE);
|
|
|
|
|
|
|
|
sp = memblock_alloc(THREAD_SIZE, THREAD_SIZE);
|
|
|
|
dbgirq_ctx[i] = (struct thread_info *)__va(sp);
|
|
|
|
paca[i].dbg_kstack = __va(sp + THREAD_SIZE);
|
|
|
|
|
|
|
|
sp = memblock_alloc(THREAD_SIZE, THREAD_SIZE);
|
|
|
|
mcheckirq_ctx[i] = (struct thread_info *)__va(sp);
|
|
|
|
paca[i].mc_kstack = __va(sp + THREAD_SIZE);
|
2009-07-24 07:15:59 +08:00
|
|
|
}
|
2011-04-06 13:18:48 +08:00
|
|
|
|
|
|
|
if (cpu_has_feature(CPU_FTR_DEBUG_LVL_EXC))
|
2013-05-12 07:26:23 +08:00
|
|
|
patch_exception(0x040, exc_debug_debug_book3e);
|
2009-07-24 07:15:59 +08:00
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2017-06-21 13:58:29 +08:00
|
|
|
/*
|
|
|
|
* Emergency stacks are used for a range of things, from asynchronous
|
|
|
|
* NMIs (system reset, machine check) to synchronous, process context.
|
|
|
|
* We set preempt_count to zero, even though that isn't necessarily correct. To
|
|
|
|
* get the right value we'd need to copy it from the previous thread_info, but
|
|
|
|
* doing that might fault causing more problems.
|
|
|
|
* TODO: what to do with accounting?
|
|
|
|
*/
|
|
|
|
static void emerg_stack_init_thread_info(struct thread_info *ti, int cpu)
|
|
|
|
{
|
|
|
|
ti->task = NULL;
|
|
|
|
ti->cpu = cpu;
|
|
|
|
ti->preempt_count = 0;
|
|
|
|
ti->local_flags = 0;
|
|
|
|
ti->flags = 0;
|
|
|
|
klp_init_thread_info(ti);
|
|
|
|
}
|
|
|
|
|
2005-10-10 20:50:37 +08:00
|
|
|
/*
|
|
|
|
* Stack space used when we detect a bad kernel stack pointer, and
|
2013-10-30 22:34:00 +08:00
|
|
|
* early in SMP boots before relocation is enabled. Exclusive emergency
|
|
|
|
* stack for machine checks.
|
2005-10-10 20:50:37 +08:00
|
|
|
*/
|
2016-07-05 13:07:51 +08:00
|
|
|
void __init emergency_stack_init(void)
|
2005-10-10 20:50:37 +08:00
|
|
|
{
|
2010-05-11 02:59:18 +08:00
|
|
|
u64 limit;
|
2005-10-10 20:50:37 +08:00
|
|
|
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
|
2017-08-13 09:33:41 +08:00
|
|
|
* bringup, machine check, system reset, and HMI, we need to get
|
|
|
|
* at them in real mode. This means they must also be within the RMO
|
|
|
|
* region.
|
2017-06-21 13:58:29 +08:00
|
|
|
*
|
|
|
|
* The IRQ stacks allocated elsewhere in this file are zeroed and
|
|
|
|
* initialized in kernel/irq.c. These are initialized here in order
|
|
|
|
* to have emergency stacks available as early as possible.
|
2005-10-10 20:50:37 +08:00
|
|
|
*/
|
2011-05-03 22:07:01 +08:00
|
|
|
limit = min(safe_stack_limit(), ppc64_rma_size);
|
2005-10-10 20:50:37 +08:00
|
|
|
|
2008-04-30 11:21:45 +08:00
|
|
|
for_each_possible_cpu(i) {
|
2016-03-24 19:04:04 +08:00
|
|
|
struct thread_info *ti;
|
|
|
|
ti = __va(memblock_alloc_base(THREAD_SIZE, THREAD_SIZE, limit));
|
2017-06-21 13:58:29 +08:00
|
|
|
memset(ti, 0, THREAD_SIZE);
|
|
|
|
emerg_stack_init_thread_info(ti, i);
|
2016-03-24 19:04:04 +08:00
|
|
|
paca[i].emergency_sp = (void *)ti + THREAD_SIZE;
|
2013-10-30 22:34:00 +08:00
|
|
|
|
|
|
|
#ifdef CONFIG_PPC_BOOK3S_64
|
2016-12-20 02:30:06 +08:00
|
|
|
/* emergency stack for NMI exception handling. */
|
|
|
|
ti = __va(memblock_alloc_base(THREAD_SIZE, THREAD_SIZE, limit));
|
2017-06-21 13:58:29 +08:00
|
|
|
memset(ti, 0, THREAD_SIZE);
|
|
|
|
emerg_stack_init_thread_info(ti, i);
|
2016-12-20 02:30:06 +08:00
|
|
|
paca[i].nmi_emergency_sp = (void *)ti + THREAD_SIZE;
|
|
|
|
|
2013-10-30 22:34:00 +08:00
|
|
|
/* emergency stack for machine check exception handling. */
|
2016-03-24 19:04:04 +08:00
|
|
|
ti = __va(memblock_alloc_base(THREAD_SIZE, THREAD_SIZE, limit));
|
2017-06-21 13:58:29 +08:00
|
|
|
memset(ti, 0, THREAD_SIZE);
|
|
|
|
emerg_stack_init_thread_info(ti, i);
|
2016-03-24 19:04:04 +08:00
|
|
|
paca[i].mc_emergency_sp = (void *)ti + THREAD_SIZE;
|
2013-10-30 22:34:00 +08:00
|
|
|
#endif
|
2008-04-30 11:21:45 +08:00
|
|
|
}
|
2005-10-10 20:50:37 +08:00
|
|
|
}
|
|
|
|
|
[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 10:16:44 +08:00
|
|
|
#ifdef CONFIG_SMP
|
2009-08-14 14:00:53 +08:00
|
|
|
#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 10:16:44 +08:00
|
|
|
{
|
2017-06-06 18:23:57 +08:00
|
|
|
return __alloc_bootmem_node(NODE_DATA(early_cpu_to_node(cpu)), size, align,
|
2009-08-14 14:00:53 +08:00
|
|
|
__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 10:16:44 +08:00
|
|
|
|
2009-08-14 14:00:53 +08:00
|
|
|
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 10:16:44 +08:00
|
|
|
|
2009-08-14 14:00:53 +08:00
|
|
|
static int pcpu_cpu_distance(unsigned int from, unsigned int to)
|
|
|
|
{
|
2017-06-06 18:23:57 +08:00
|
|
|
if (early_cpu_to_node(from) == early_cpu_to_node(to))
|
2009-08-14 14:00:53 +08:00
|
|
|
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>
2010-06-01 02:45:11 +08:00
|
|
|
unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
|
|
|
|
EXPORT_SYMBOL(__per_cpu_offset);
|
|
|
|
|
2009-08-14 14:00:53 +08:00
|
|
|
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>
2010-06-01 02:45:11 +08:00
|
|
|
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 10:16:44 +08:00
|
|
|
}
|
|
|
|
#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 14:25:10 +08:00
|
|
|
|
2014-06-04 15:50:47 +08:00
|
|
|
#ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
|
|
|
|
unsigned long memory_block_size_bytes(void)
|
|
|
|
{
|
|
|
|
if (ppc_md.memory_block_size)
|
|
|
|
return ppc_md.memory_block_size();
|
|
|
|
|
|
|
|
return MIN_MEMORY_BLOCK_SIZE;
|
|
|
|
}
|
|
|
|
#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 14:25:10 +08:00
|
|
|
|
2013-07-15 11:03:08 +08:00
|
|
|
#if defined(CONFIG_PPC_INDIRECT_PIO) || defined(CONFIG_PPC_INDIRECT_MMIO)
|
[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 14:25:10 +08:00
|
|
|
struct ppc_pci_io ppc_pci_io;
|
|
|
|
EXPORT_SYMBOL(ppc_pci_io);
|
2013-07-15 11:03:08 +08:00
|
|
|
#endif
|
2017-08-28 12:27:19 +08:00
|
|
|
|
|
|
|
#ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
|
|
|
|
u64 hw_nmi_get_sample_period(int watchdog_thresh)
|
|
|
|
{
|
|
|
|
return ppc_proc_freq * watchdog_thresh;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The perf based hardlockup detector breaks PMU event based branches, so
|
|
|
|
* disable it by default. Book3S has a soft-nmi hardlockup detector based
|
|
|
|
* on the decrementer interrupt, so it does not suffer from this problem.
|
|
|
|
*
|
|
|
|
* It is likely to get false positives in VM guests, so disable it there
|
|
|
|
* by default too.
|
|
|
|
*/
|
|
|
|
static int __init disable_hardlockup_detector(void)
|
|
|
|
{
|
|
|
|
#ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
|
|
|
|
hardlockup_detector_disable();
|
|
|
|
#else
|
|
|
|
if (firmware_has_feature(FW_FEATURE_LPAR))
|
|
|
|
hardlockup_detector_disable();
|
|
|
|
#endif
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
early_initcall(disable_hardlockup_detector);
|