linux/arch/um/kernel/smp.c

267 lines
5.6 KiB
C
Raw Normal View History

/*
* Copyright (C) 2000 - 2003 Jeff Dike (jdike@addtoit.com)
* Licensed under the GPL
*/
#include "linux/percpu.h"
#include "asm/pgalloc.h"
#include "asm/tlb.h"
/* For some reason, mmu_gathers are referenced when CONFIG_SMP is off. */
DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
#ifdef CONFIG_SMP
#include "linux/sched.h"
#include "linux/module.h"
#include "linux/threads.h"
#include "linux/interrupt.h"
#include "linux/err.h"
#include "linux/hardirq.h"
#include "asm/smp.h"
#include "asm/processor.h"
#include "asm/spinlock.h"
#include "kern_util.h"
#include "kern.h"
#include "irq_user.h"
#include "os.h"
/* CPU online map, set by smp_boot_cpus */
cpumask_t cpu_online_map = CPU_MASK_NONE;
cpumask_t cpu_possible_map = CPU_MASK_NONE;
EXPORT_SYMBOL(cpu_online_map);
EXPORT_SYMBOL(cpu_possible_map);
/* Per CPU bogomips and other parameters
* The only piece used here is the ipi pipe, which is set before SMP is
* started and never changed.
*/
struct cpuinfo_um cpu_data[NR_CPUS];
/* A statistic, can be a little off */
int num_reschedules_sent = 0;
/* Not changed after boot */
struct task_struct *idle_threads[NR_CPUS];
void smp_send_reschedule(int cpu)
{
uml: start fixing os_read_file and os_write_file This patch starts the removal of a very old, very broken piece of code. This stems from the problem of passing a userspace buffer into read() or write() on the host. If that buffer had not yet been faulted in, read and write will return -EFAULT. To avoid this problem, the solution was to fault the buffer in before the system call by touching the pages that hold the buffer by doing a copy-user of a byte to each page. This is obviously bogus, but it does usually work, in tt mode, since the kernel and process are in the same address space and userspace addresses can be accessed directly in the kernel. In skas mode, where the kernel and process are in separate address spaces, it is completely bogus because the userspace address, which is invalid in the kernel, is passed into the system call instead of the corresponding physical address, which would be valid. Here, it appears that this code, on every host read() or write(), tries to fault in a random process page. This doesn't seem to cause any correctness problems, but there is a performance impact. This patch, and the ones following, result in a 10-15% performance gain on a kernel build. This code can't be immediately tossed out because when it is, you can't log in. Apparently, there is some code in the console driver which depends on this somehow. However, we can start removing it by switching the code which does I/O using kernel addresses to using plain read() and write(). This patch introduces os_read_file_k and os_write_file_k for use with kernel buffers and converts all call locations which use obvious kernel buffers to use them. These include I/O using buffers which are local variables which are on the stack or kmalloc-ed. Later patches will handle the less obvious cases, followed by a mass conversion back to the original interface. Signed-off-by: Jeff Dike <jdike@linux.intel.com> Cc: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-06 21:51:32 +00:00
os_write_file_k(cpu_data[cpu].ipi_pipe[1], "R", 1);
num_reschedules_sent++;
}
void smp_send_stop(void)
{
int i;
printk(KERN_INFO "Stopping all CPUs...");
for(i = 0; i < num_online_cpus(); i++){
if(i == current_thread->cpu)
continue;
uml: start fixing os_read_file and os_write_file This patch starts the removal of a very old, very broken piece of code. This stems from the problem of passing a userspace buffer into read() or write() on the host. If that buffer had not yet been faulted in, read and write will return -EFAULT. To avoid this problem, the solution was to fault the buffer in before the system call by touching the pages that hold the buffer by doing a copy-user of a byte to each page. This is obviously bogus, but it does usually work, in tt mode, since the kernel and process are in the same address space and userspace addresses can be accessed directly in the kernel. In skas mode, where the kernel and process are in separate address spaces, it is completely bogus because the userspace address, which is invalid in the kernel, is passed into the system call instead of the corresponding physical address, which would be valid. Here, it appears that this code, on every host read() or write(), tries to fault in a random process page. This doesn't seem to cause any correctness problems, but there is a performance impact. This patch, and the ones following, result in a 10-15% performance gain on a kernel build. This code can't be immediately tossed out because when it is, you can't log in. Apparently, there is some code in the console driver which depends on this somehow. However, we can start removing it by switching the code which does I/O using kernel addresses to using plain read() and write(). This patch introduces os_read_file_k and os_write_file_k for use with kernel buffers and converts all call locations which use obvious kernel buffers to use them. These include I/O using buffers which are local variables which are on the stack or kmalloc-ed. Later patches will handle the less obvious cases, followed by a mass conversion back to the original interface. Signed-off-by: Jeff Dike <jdike@linux.intel.com> Cc: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-06 21:51:32 +00:00
os_write_file_k(cpu_data[i].ipi_pipe[1], "S", 1);
}
printk("done\n");
}
static cpumask_t smp_commenced_mask = CPU_MASK_NONE;
static cpumask_t cpu_callin_map = CPU_MASK_NONE;
static int idle_proc(void *cpup)
{
int cpu = (int) cpup, err;
err = os_pipe(cpu_data[cpu].ipi_pipe, 1, 1);
if(err < 0)
panic("CPU#%d failed to create IPI pipe, err = %d", cpu, -err);
os_set_fd_async(cpu_data[cpu].ipi_pipe[0],
current->thread.mode.tt.extern_pid);
wmb();
if (cpu_test_and_set(cpu, cpu_callin_map)) {
printk("huh, CPU#%d already present??\n", cpu);
BUG();
}
while (!cpu_isset(cpu, smp_commenced_mask))
cpu_relax();
cpu_set(cpu, cpu_online_map);
default_idle();
return(0);
}
static struct task_struct *idle_thread(int cpu)
{
struct task_struct *new_task;
unsigned char c;
current->thread.request.u.thread.proc = idle_proc;
current->thread.request.u.thread.arg = (void *) cpu;
new_task = fork_idle(cpu);
if(IS_ERR(new_task))
panic("copy_process failed in idle_thread, error = %ld",
PTR_ERR(new_task));
cpu_tasks[cpu] = ((struct cpu_task)
{ .pid = new_task->thread.mode.tt.extern_pid,
.task = new_task } );
idle_threads[cpu] = new_task;
uml: start fixing os_read_file and os_write_file This patch starts the removal of a very old, very broken piece of code. This stems from the problem of passing a userspace buffer into read() or write() on the host. If that buffer had not yet been faulted in, read and write will return -EFAULT. To avoid this problem, the solution was to fault the buffer in before the system call by touching the pages that hold the buffer by doing a copy-user of a byte to each page. This is obviously bogus, but it does usually work, in tt mode, since the kernel and process are in the same address space and userspace addresses can be accessed directly in the kernel. In skas mode, where the kernel and process are in separate address spaces, it is completely bogus because the userspace address, which is invalid in the kernel, is passed into the system call instead of the corresponding physical address, which would be valid. Here, it appears that this code, on every host read() or write(), tries to fault in a random process page. This doesn't seem to cause any correctness problems, but there is a performance impact. This patch, and the ones following, result in a 10-15% performance gain on a kernel build. This code can't be immediately tossed out because when it is, you can't log in. Apparently, there is some code in the console driver which depends on this somehow. However, we can start removing it by switching the code which does I/O using kernel addresses to using plain read() and write(). This patch introduces os_read_file_k and os_write_file_k for use with kernel buffers and converts all call locations which use obvious kernel buffers to use them. These include I/O using buffers which are local variables which are on the stack or kmalloc-ed. Later patches will handle the less obvious cases, followed by a mass conversion back to the original interface. Signed-off-by: Jeff Dike <jdike@linux.intel.com> Cc: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-06 21:51:32 +00:00
CHOOSE_MODE(os_write_file_k(new_task->thread.mode.tt.switch_pipe[1], &c,
sizeof(c)),
({ panic("skas mode doesn't support SMP"); }));
return(new_task);
}
void smp_prepare_cpus(unsigned int maxcpus)
{
struct task_struct *idle;
unsigned long waittime;
int err, cpu, me = smp_processor_id();
int i;
for (i = 0; i < ncpus; ++i)
cpu_set(i, cpu_possible_map);
cpu_clear(me, cpu_online_map);
cpu_set(me, cpu_online_map);
cpu_set(me, cpu_callin_map);
err = os_pipe(cpu_data[me].ipi_pipe, 1, 1);
if(err < 0)
panic("CPU#0 failed to create IPI pipe, errno = %d", -err);
os_set_fd_async(cpu_data[me].ipi_pipe[0],
current->thread.mode.tt.extern_pid);
for(cpu = 1; cpu < ncpus; cpu++){
printk("Booting processor %d...\n", cpu);
idle = idle_thread(cpu);
init_idle(idle, cpu);
waittime = 200000000;
while (waittime-- && !cpu_isset(cpu, cpu_callin_map))
cpu_relax();
if (cpu_isset(cpu, cpu_callin_map))
printk("done\n");
else printk("failed\n");
}
}
void smp_prepare_boot_cpu(void)
{
cpu_set(smp_processor_id(), cpu_online_map);
}
int __cpu_up(unsigned int cpu)
{
cpu_set(cpu, smp_commenced_mask);
while (!cpu_isset(cpu, cpu_online_map))
mb();
return(0);
}
int setup_profiling_timer(unsigned int multiplier)
{
printk(KERN_INFO "setup_profiling_timer\n");
return(0);
}
void smp_call_function_slave(int cpu);
void IPI_handler(int cpu)
{
unsigned char c;
int fd;
fd = cpu_data[cpu].ipi_pipe[0];
uml: start fixing os_read_file and os_write_file This patch starts the removal of a very old, very broken piece of code. This stems from the problem of passing a userspace buffer into read() or write() on the host. If that buffer had not yet been faulted in, read and write will return -EFAULT. To avoid this problem, the solution was to fault the buffer in before the system call by touching the pages that hold the buffer by doing a copy-user of a byte to each page. This is obviously bogus, but it does usually work, in tt mode, since the kernel and process are in the same address space and userspace addresses can be accessed directly in the kernel. In skas mode, where the kernel and process are in separate address spaces, it is completely bogus because the userspace address, which is invalid in the kernel, is passed into the system call instead of the corresponding physical address, which would be valid. Here, it appears that this code, on every host read() or write(), tries to fault in a random process page. This doesn't seem to cause any correctness problems, but there is a performance impact. This patch, and the ones following, result in a 10-15% performance gain on a kernel build. This code can't be immediately tossed out because when it is, you can't log in. Apparently, there is some code in the console driver which depends on this somehow. However, we can start removing it by switching the code which does I/O using kernel addresses to using plain read() and write(). This patch introduces os_read_file_k and os_write_file_k for use with kernel buffers and converts all call locations which use obvious kernel buffers to use them. These include I/O using buffers which are local variables which are on the stack or kmalloc-ed. Later patches will handle the less obvious cases, followed by a mass conversion back to the original interface. Signed-off-by: Jeff Dike <jdike@linux.intel.com> Cc: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-06 21:51:32 +00:00
while (os_read_file_k(fd, &c, 1) == 1) {
switch (c) {
case 'C':
smp_call_function_slave(cpu);
break;
case 'R':
set_tsk_need_resched(current);
break;
case 'S':
printk("CPU#%d stopping\n", cpu);
while(1)
pause();
break;
default:
printk("CPU#%d received unknown IPI [%c]!\n", cpu, c);
break;
}
}
}
int hard_smp_processor_id(void)
{
return(pid_to_processor_id(os_getpid()));
}
static DEFINE_SPINLOCK(call_lock);
static atomic_t scf_started;
static atomic_t scf_finished;
static void (*func)(void *info);
static void *info;
void smp_call_function_slave(int cpu)
{
atomic_inc(&scf_started);
(*func)(info);
atomic_inc(&scf_finished);
}
int smp_call_function(void (*_func)(void *info), void *_info, int nonatomic,
int wait)
{
int cpus = num_online_cpus() - 1;
int i;
if (!cpus)
return 0;
/* Can deadlock when called with interrupts disabled */
WARN_ON(irqs_disabled());
spin_lock_bh(&call_lock);
atomic_set(&scf_started, 0);
atomic_set(&scf_finished, 0);
func = _func;
info = _info;
for_each_online_cpu(i)
uml: start fixing os_read_file and os_write_file This patch starts the removal of a very old, very broken piece of code. This stems from the problem of passing a userspace buffer into read() or write() on the host. If that buffer had not yet been faulted in, read and write will return -EFAULT. To avoid this problem, the solution was to fault the buffer in before the system call by touching the pages that hold the buffer by doing a copy-user of a byte to each page. This is obviously bogus, but it does usually work, in tt mode, since the kernel and process are in the same address space and userspace addresses can be accessed directly in the kernel. In skas mode, where the kernel and process are in separate address spaces, it is completely bogus because the userspace address, which is invalid in the kernel, is passed into the system call instead of the corresponding physical address, which would be valid. Here, it appears that this code, on every host read() or write(), tries to fault in a random process page. This doesn't seem to cause any correctness problems, but there is a performance impact. This patch, and the ones following, result in a 10-15% performance gain on a kernel build. This code can't be immediately tossed out because when it is, you can't log in. Apparently, there is some code in the console driver which depends on this somehow. However, we can start removing it by switching the code which does I/O using kernel addresses to using plain read() and write(). This patch introduces os_read_file_k and os_write_file_k for use with kernel buffers and converts all call locations which use obvious kernel buffers to use them. These include I/O using buffers which are local variables which are on the stack or kmalloc-ed. Later patches will handle the less obvious cases, followed by a mass conversion back to the original interface. Signed-off-by: Jeff Dike <jdike@linux.intel.com> Cc: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-06 21:51:32 +00:00
os_write_file_k(cpu_data[i].ipi_pipe[1], "C", 1);
while (atomic_read(&scf_started) != cpus)
barrier();
if (wait)
while (atomic_read(&scf_finished) != cpus)
barrier();
spin_unlock_bh(&call_lock);
return 0;
}
#endif
/*
* Overrides for Emacs so that we follow Linus's tabbing style.
* Emacs will notice this stuff at the end of the file and automatically
* adjust the settings for this buffer only. This must remain at the end
* of the file.
* ---------------------------------------------------------------------------
* Local variables:
* c-file-style: "linux"
* End:
*/