2019-06-03 13:44:50 +08:00
|
|
|
// SPDX-License-Identifier: GPL-2.0-only
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
/*
|
|
|
|
* Adapted from arm64 version.
|
|
|
|
*
|
|
|
|
* Copyright (C) 2012 ARM Limited
|
|
|
|
* Copyright (C) 2015 Mentor Graphics Corporation.
|
|
|
|
*/
|
|
|
|
|
2016-08-15 16:20:21 +08:00
|
|
|
#include <linux/cache.h>
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
#include <linux/elf.h>
|
|
|
|
#include <linux/err.h>
|
|
|
|
#include <linux/kernel.h>
|
|
|
|
#include <linux/mm.h>
|
|
|
|
#include <linux/of.h>
|
|
|
|
#include <linux/printk.h>
|
|
|
|
#include <linux/slab.h>
|
|
|
|
#include <linux/timekeeper_internal.h>
|
|
|
|
#include <linux/vmalloc.h>
|
|
|
|
#include <asm/arch_timer.h>
|
|
|
|
#include <asm/barrier.h>
|
|
|
|
#include <asm/cacheflush.h>
|
|
|
|
#include <asm/page.h>
|
|
|
|
#include <asm/vdso.h>
|
|
|
|
#include <asm/vdso_datapage.h>
|
|
|
|
#include <clocksource/arm_arch_timer.h>
|
2019-11-04 18:59:59 +08:00
|
|
|
#include <vdso/helpers.h>
|
|
|
|
#include <vdso/vsyscall.h>
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
|
|
|
|
#define MAX_SYMNAME 64
|
|
|
|
|
|
|
|
static struct page **vdso_text_pagelist;
|
|
|
|
|
ARM: 8748/1: mm: Define vdso_start, vdso_end as array
Define vdso_start, vdso_end as array to avoid compile-time analysis error
for the case of built with CONFIG_FORTIFY_SOURCE.
and, since vdso_start, vdso_end are used in vdso.c only,
move extern-declaration from vdso.h to vdso.c.
If kernel is built with CONFIG_FORTIFY_SOURCE,
compile-time error happens at this code.
- if (memcmp(&vdso_start, "177ELF", 4))
The size of "&vdso_start" is recognized as 1 byte, but n is 4,
So that compile-time error is reported.
Acked-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Jinbum Park <jinb.park7@gmail.com>
Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
2018-03-06 08:37:21 +08:00
|
|
|
extern char vdso_start[], vdso_end[];
|
|
|
|
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
/* Total number of pages needed for the data and text portions of the VDSO. */
|
2016-08-15 16:20:21 +08:00
|
|
|
unsigned int vdso_total_pages __ro_after_init;
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
|
|
|
|
/*
|
|
|
|
* The VDSO data page.
|
|
|
|
*/
|
|
|
|
static union vdso_data_store vdso_data_store __page_aligned_data;
|
2019-11-04 18:59:59 +08:00
|
|
|
struct vdso_data *vdso_data = vdso_data_store.data;
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
|
2016-08-15 16:20:21 +08:00
|
|
|
static struct page *vdso_data_page __ro_after_init;
|
|
|
|
static const struct vm_special_mapping vdso_data_mapping = {
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
.name = "[vvar]",
|
|
|
|
.pages = &vdso_data_page,
|
|
|
|
};
|
|
|
|
|
2017-06-20 00:32:42 +08:00
|
|
|
static int vdso_mremap(const struct vm_special_mapping *sm,
|
|
|
|
struct vm_area_struct *new_vma)
|
|
|
|
{
|
|
|
|
unsigned long new_size = new_vma->vm_end - new_vma->vm_start;
|
|
|
|
unsigned long vdso_size;
|
|
|
|
|
|
|
|
/* without VVAR page */
|
|
|
|
vdso_size = (vdso_total_pages - 1) << PAGE_SHIFT;
|
|
|
|
|
|
|
|
if (vdso_size != new_size)
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
current->mm->context.vdso = new_vma->vm_start;
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2016-08-15 16:20:21 +08:00
|
|
|
static struct vm_special_mapping vdso_text_mapping __ro_after_init = {
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
.name = "[vdso]",
|
2017-06-20 00:32:42 +08:00
|
|
|
.mremap = vdso_mremap,
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
struct elfinfo {
|
|
|
|
Elf32_Ehdr *hdr; /* ptr to ELF */
|
|
|
|
Elf32_Sym *dynsym; /* ptr to .dynsym section */
|
|
|
|
unsigned long dynsymsize; /* size of .dynsym section */
|
|
|
|
char *dynstr; /* ptr to .dynstr section */
|
|
|
|
};
|
|
|
|
|
|
|
|
/* Cached result of boot-time check for whether the arch timer exists,
|
|
|
|
* and if so, whether the virtual counter is useable.
|
|
|
|
*/
|
2019-11-04 18:59:59 +08:00
|
|
|
bool cntvct_ok __ro_after_init;
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
|
|
|
|
static bool __init cntvct_functional(void)
|
|
|
|
{
|
|
|
|
struct device_node *np;
|
|
|
|
bool ret = false;
|
|
|
|
|
|
|
|
if (!IS_ENABLED(CONFIG_ARM_ARCH_TIMER))
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
/* The arm_arch_timer core should export
|
|
|
|
* arch_timer_use_virtual or similar so we don't have to do
|
|
|
|
* this.
|
|
|
|
*/
|
|
|
|
np = of_find_compatible_node(NULL, NULL, "arm,armv7-timer");
|
2020-01-29 03:22:13 +08:00
|
|
|
if (!np)
|
|
|
|
np = of_find_compatible_node(NULL, NULL, "arm,armv8-timer");
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
if (!np)
|
|
|
|
goto out_put;
|
|
|
|
|
|
|
|
if (of_property_read_bool(np, "arm,cpu-registers-not-fw-configured"))
|
|
|
|
goto out_put;
|
|
|
|
|
|
|
|
ret = true;
|
|
|
|
|
|
|
|
out_put:
|
|
|
|
of_node_put(np);
|
|
|
|
out:
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void * __init find_section(Elf32_Ehdr *ehdr, const char *name,
|
|
|
|
unsigned long *size)
|
|
|
|
{
|
|
|
|
Elf32_Shdr *sechdrs;
|
|
|
|
unsigned int i;
|
|
|
|
char *secnames;
|
|
|
|
|
|
|
|
/* Grab section headers and strings so we can tell who is who */
|
|
|
|
sechdrs = (void *)ehdr + ehdr->e_shoff;
|
|
|
|
secnames = (void *)ehdr + sechdrs[ehdr->e_shstrndx].sh_offset;
|
|
|
|
|
|
|
|
/* Find the section they want */
|
|
|
|
for (i = 1; i < ehdr->e_shnum; i++) {
|
|
|
|
if (strcmp(secnames + sechdrs[i].sh_name, name) == 0) {
|
|
|
|
if (size)
|
|
|
|
*size = sechdrs[i].sh_size;
|
|
|
|
return (void *)ehdr + sechdrs[i].sh_offset;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (size)
|
|
|
|
*size = 0;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
static Elf32_Sym * __init find_symbol(struct elfinfo *lib, const char *symname)
|
|
|
|
{
|
|
|
|
unsigned int i;
|
|
|
|
|
|
|
|
for (i = 0; i < (lib->dynsymsize / sizeof(Elf32_Sym)); i++) {
|
|
|
|
char name[MAX_SYMNAME], *c;
|
|
|
|
|
|
|
|
if (lib->dynsym[i].st_name == 0)
|
|
|
|
continue;
|
|
|
|
strlcpy(name, lib->dynstr + lib->dynsym[i].st_name,
|
|
|
|
MAX_SYMNAME);
|
|
|
|
c = strchr(name, '@');
|
|
|
|
if (c)
|
|
|
|
*c = 0;
|
|
|
|
if (strcmp(symname, name) == 0)
|
|
|
|
return &lib->dynsym[i];
|
|
|
|
}
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void __init vdso_nullpatch_one(struct elfinfo *lib, const char *symname)
|
|
|
|
{
|
|
|
|
Elf32_Sym *sym;
|
|
|
|
|
|
|
|
sym = find_symbol(lib, symname);
|
|
|
|
if (!sym)
|
|
|
|
return;
|
|
|
|
|
|
|
|
sym->st_name = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void __init patch_vdso(void *ehdr)
|
|
|
|
{
|
|
|
|
struct elfinfo einfo;
|
|
|
|
|
|
|
|
einfo = (struct elfinfo) {
|
|
|
|
.hdr = ehdr,
|
|
|
|
};
|
|
|
|
|
|
|
|
einfo.dynsym = find_section(einfo.hdr, ".dynsym", &einfo.dynsymsize);
|
|
|
|
einfo.dynstr = find_section(einfo.hdr, ".dynstr", NULL);
|
|
|
|
|
|
|
|
/* If the virtual counter is absent or non-functional we don't
|
|
|
|
* want programs to incur the slight additional overhead of
|
|
|
|
* dispatching through the VDSO only to fall back to syscalls.
|
|
|
|
*/
|
|
|
|
if (!cntvct_ok) {
|
|
|
|
vdso_nullpatch_one(&einfo, "__vdso_gettimeofday");
|
|
|
|
vdso_nullpatch_one(&einfo, "__vdso_clock_gettime");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static int __init vdso_init(void)
|
|
|
|
{
|
|
|
|
unsigned int text_pages;
|
|
|
|
int i;
|
|
|
|
|
ARM: 8748/1: mm: Define vdso_start, vdso_end as array
Define vdso_start, vdso_end as array to avoid compile-time analysis error
for the case of built with CONFIG_FORTIFY_SOURCE.
and, since vdso_start, vdso_end are used in vdso.c only,
move extern-declaration from vdso.h to vdso.c.
If kernel is built with CONFIG_FORTIFY_SOURCE,
compile-time error happens at this code.
- if (memcmp(&vdso_start, "177ELF", 4))
The size of "&vdso_start" is recognized as 1 byte, but n is 4,
So that compile-time error is reported.
Acked-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Jinbum Park <jinb.park7@gmail.com>
Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
2018-03-06 08:37:21 +08:00
|
|
|
if (memcmp(vdso_start, "\177ELF", 4)) {
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
pr_err("VDSO is not a valid ELF object!\n");
|
|
|
|
return -ENOEXEC;
|
|
|
|
}
|
|
|
|
|
ARM: 8748/1: mm: Define vdso_start, vdso_end as array
Define vdso_start, vdso_end as array to avoid compile-time analysis error
for the case of built with CONFIG_FORTIFY_SOURCE.
and, since vdso_start, vdso_end are used in vdso.c only,
move extern-declaration from vdso.h to vdso.c.
If kernel is built with CONFIG_FORTIFY_SOURCE,
compile-time error happens at this code.
- if (memcmp(&vdso_start, "177ELF", 4))
The size of "&vdso_start" is recognized as 1 byte, but n is 4,
So that compile-time error is reported.
Acked-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Jinbum Park <jinb.park7@gmail.com>
Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
2018-03-06 08:37:21 +08:00
|
|
|
text_pages = (vdso_end - vdso_start) >> PAGE_SHIFT;
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
|
|
|
|
/* Allocate the VDSO text pagelist */
|
|
|
|
vdso_text_pagelist = kcalloc(text_pages, sizeof(struct page *),
|
|
|
|
GFP_KERNEL);
|
|
|
|
if (vdso_text_pagelist == NULL)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
/* Grab the VDSO data page. */
|
|
|
|
vdso_data_page = virt_to_page(vdso_data);
|
|
|
|
|
|
|
|
/* Grab the VDSO text pages. */
|
|
|
|
for (i = 0; i < text_pages; i++) {
|
|
|
|
struct page *page;
|
|
|
|
|
ARM: 8748/1: mm: Define vdso_start, vdso_end as array
Define vdso_start, vdso_end as array to avoid compile-time analysis error
for the case of built with CONFIG_FORTIFY_SOURCE.
and, since vdso_start, vdso_end are used in vdso.c only,
move extern-declaration from vdso.h to vdso.c.
If kernel is built with CONFIG_FORTIFY_SOURCE,
compile-time error happens at this code.
- if (memcmp(&vdso_start, "177ELF", 4))
The size of "&vdso_start" is recognized as 1 byte, but n is 4,
So that compile-time error is reported.
Acked-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Jinbum Park <jinb.park7@gmail.com>
Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
2018-03-06 08:37:21 +08:00
|
|
|
page = virt_to_page(vdso_start + i * PAGE_SIZE);
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
vdso_text_pagelist[i] = page;
|
|
|
|
}
|
|
|
|
|
|
|
|
vdso_text_mapping.pages = vdso_text_pagelist;
|
|
|
|
|
|
|
|
vdso_total_pages = 1; /* for the data/vvar page */
|
|
|
|
vdso_total_pages += text_pages;
|
|
|
|
|
|
|
|
cntvct_ok = cntvct_functional();
|
|
|
|
|
ARM: 8748/1: mm: Define vdso_start, vdso_end as array
Define vdso_start, vdso_end as array to avoid compile-time analysis error
for the case of built with CONFIG_FORTIFY_SOURCE.
and, since vdso_start, vdso_end are used in vdso.c only,
move extern-declaration from vdso.h to vdso.c.
If kernel is built with CONFIG_FORTIFY_SOURCE,
compile-time error happens at this code.
- if (memcmp(&vdso_start, "177ELF", 4))
The size of "&vdso_start" is recognized as 1 byte, but n is 4,
So that compile-time error is reported.
Acked-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Jinbum Park <jinb.park7@gmail.com>
Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
2018-03-06 08:37:21 +08:00
|
|
|
patch_vdso(vdso_start);
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
arch_initcall(vdso_init);
|
|
|
|
|
|
|
|
static int install_vvar(struct mm_struct *mm, unsigned long addr)
|
|
|
|
{
|
|
|
|
struct vm_area_struct *vma;
|
|
|
|
|
|
|
|
vma = _install_special_mapping(mm, addr, PAGE_SIZE,
|
|
|
|
VM_READ | VM_MAYREAD,
|
|
|
|
&vdso_data_mapping);
|
|
|
|
|
2015-12-09 00:30:25 +08:00
|
|
|
return PTR_ERR_OR_ZERO(vma);
|
ARM: 8331/1: VDSO initialization, mapping, and synchronization
Initialize the VDSO page list at boot, install the VDSO mapping at
exec time, and update the data page during timer ticks. This code is
not built if CONFIG_VDSO is not enabled.
Account for the VDSO length when randomizing the offset from the
stack. The [vdso] and [vvar] pages are placed immediately following
the sigpage with separate _install_special_mapping calls.
We want to "penalize" systems lacking the arch timer as little
as possible. Previous versions of this code installed the VDSO
unconditionally and unmodified, making it a measurably slower way for
glibc to invoke the real syscalls on such systems. E.g. calling
gettimeofday via glibc goes from ~560ns to ~630ns on i.MX6Q.
If we can indicate to glibc that the time-related APIs in the VDSO are
not accelerated, glibc can continue to invoke the syscalls directly
instead of dispatching through the VDSO only to fall back to the slow
path.
Thus, if the architected timer is unusable for whatever reason, patch
the VDSO at boot time so that symbol lookups for gettimeofday and
clock_gettime return NULL. (This is similar to what powerpc does and
borrows code from there.) This allows glibc to perform the syscall
directly instead of passing control to the VDSO, which minimizes the
penalty. In my measurements the time taken for a gettimeofday call
via glibc goes from ~560ns to ~580ns (again on i.MX6Q), and this is
solely due to adding a test and branch to glibc's gettimeofday syscall
wrapper.
An alternative to patching the VDSO at boot would be to not install
the VDSO at all when the arch timer isn't usable. Another alternative
is to include a separate "dummy" vdso.so without gettimeofday and
clock_gettime, which would be selected at boot time. Either of these
would get cumbersome if the VDSO were to gain support for an API such
as getcpu which is unrelated to arch timer support.
Signed-off-by: Nathan Lynch <nathan_lynch@mentor.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2015-03-26 02:15:08 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/* assumes mmap_sem is write-locked */
|
|
|
|
void arm_install_vdso(struct mm_struct *mm, unsigned long addr)
|
|
|
|
{
|
|
|
|
struct vm_area_struct *vma;
|
|
|
|
unsigned long len;
|
|
|
|
|
|
|
|
mm->context.vdso = 0;
|
|
|
|
|
|
|
|
if (vdso_text_pagelist == NULL)
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (install_vvar(mm, addr))
|
|
|
|
return;
|
|
|
|
|
|
|
|
/* Account for vvar page. */
|
|
|
|
addr += PAGE_SIZE;
|
|
|
|
len = (vdso_total_pages - 1) << PAGE_SHIFT;
|
|
|
|
|
|
|
|
vma = _install_special_mapping(mm, addr, len,
|
|
|
|
VM_READ | VM_EXEC | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC,
|
|
|
|
&vdso_text_mapping);
|
|
|
|
|
|
|
|
if (!IS_ERR(vma))
|
|
|
|
mm->context.vdso = addr;
|
|
|
|
}
|
|
|
|
|