2009-06-12 16:26:42 +08:00
|
|
|
/*
|
|
|
|
* Access kernel memory without faulting -- s390 specific implementation.
|
|
|
|
*
|
2015-03-13 20:13:36 +08:00
|
|
|
* Copyright IBM Corp. 2009, 2015
|
2009-06-12 16:26:42 +08:00
|
|
|
*
|
|
|
|
* Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>,
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include <linux/uaccess.h>
|
|
|
|
#include <linux/kernel.h>
|
|
|
|
#include <linux/types.h>
|
|
|
|
#include <linux/errno.h>
|
2011-10-30 22:16:39 +08:00
|
|
|
#include <linux/gfp.h>
|
s390: allow absolute memory access for /dev/mem
Currently dev/mem for s390 provides only real memory access. This means
that the CPU prefix pages are swapped. The prefix swap for real memory
works as follows:
Each CPU owns a prefix register that points to a page aligned memory
location "P". If this CPU accesses the address range [0,0x1fff], it is
translated by the hardware to [P,P+0x1fff]. Accordingly if this CPU
accesses the address range [P,P+0x1fff], it is translated by the hardware
to [0,0x1fff]. Therefore, if [P,P+0x1fff] or [0,0x1fff] is read from
the current /dev/mem device, the incorrectly swapped memory content is
returned.
With this patch the /dev/mem architecture code is modified to provide
absolute memory access. This is done via the arch specific functions
xlate_dev_mem_ptr() and unxlate_dev_mem_ptr(). For swapped pages on
s390 the function xlate_dev_mem_ptr() now returns a new buffer with a
copy of the requested absolute memory. In case the buffer was allocated,
the unxlate_dev_mem_ptr() function frees it after /dev/mem code has
called copy_to_user().
Signed-off-by: Michael Holzheu <holzheu@linux.vnet.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
2012-05-09 22:27:36 +08:00
|
|
|
#include <linux/cpu.h>
|
2012-03-29 01:30:02 +08:00
|
|
|
#include <asm/ctl_reg.h>
|
2013-09-07 01:10:48 +08:00
|
|
|
#include <asm/io.h>
|
2009-06-12 16:26:42 +08:00
|
|
|
|
2015-03-13 19:55:56 +08:00
|
|
|
static notrace long s390_kernel_write_odd(void *dst, const void *src, size_t size)
|
2009-06-12 16:26:42 +08:00
|
|
|
{
|
2015-03-13 20:13:36 +08:00
|
|
|
unsigned long aligned, offset, count;
|
|
|
|
char tmp[8];
|
2009-06-12 16:26:42 +08:00
|
|
|
|
2015-03-13 20:13:36 +08:00
|
|
|
aligned = (unsigned long) dst & ~7UL;
|
|
|
|
offset = (unsigned long) dst & 7UL;
|
|
|
|
size = min(8UL - offset, size);
|
|
|
|
count = size - 1;
|
2009-06-12 16:26:42 +08:00
|
|
|
asm volatile(
|
|
|
|
" bras 1,0f\n"
|
2015-03-13 20:13:36 +08:00
|
|
|
" mvc 0(1,%4),0(%5)\n"
|
|
|
|
"0: mvc 0(8,%3),0(%0)\n"
|
|
|
|
" ex %1,0(1)\n"
|
|
|
|
" lg %1,0(%3)\n"
|
|
|
|
" lra %0,0(%0)\n"
|
|
|
|
" sturg %1,%0\n"
|
|
|
|
: "+&a" (aligned), "+&a" (count), "=m" (tmp)
|
|
|
|
: "a" (&tmp), "a" (&tmp[offset]), "a" (src)
|
|
|
|
: "cc", "memory", "1");
|
|
|
|
return size;
|
2009-06-12 16:26:42 +08:00
|
|
|
}
|
|
|
|
|
2015-03-13 19:55:56 +08:00
|
|
|
/*
|
|
|
|
* s390_kernel_write - write to kernel memory bypassing DAT
|
|
|
|
* @dst: destination address
|
|
|
|
* @src: source address
|
|
|
|
* @size: number of bytes to copy
|
|
|
|
*
|
|
|
|
* This function writes to kernel memory bypassing DAT and possible page table
|
|
|
|
* write protection. It writes to the destination using the sturg instruction.
|
2015-03-13 20:13:36 +08:00
|
|
|
* Therefore we have a read-modify-write sequence: the function reads eight
|
|
|
|
* bytes from destination at an eight byte boundary, modifies the bytes
|
2015-03-13 19:55:56 +08:00
|
|
|
* requested and writes the result back in a loop.
|
|
|
|
*
|
|
|
|
* Note: this means that this function may not be called concurrently on
|
|
|
|
* several cpus with overlapping words, since this may potentially
|
|
|
|
* cause data corruption.
|
|
|
|
*/
|
|
|
|
void notrace s390_kernel_write(void *dst, const void *src, size_t size)
|
2009-06-12 16:26:42 +08:00
|
|
|
{
|
2015-03-13 20:13:36 +08:00
|
|
|
long copied;
|
2009-06-12 16:26:42 +08:00
|
|
|
|
|
|
|
while (size) {
|
2015-03-13 19:55:56 +08:00
|
|
|
copied = s390_kernel_write_odd(dst, src, size);
|
2009-06-12 16:26:42 +08:00
|
|
|
dst += copied;
|
|
|
|
src += copied;
|
|
|
|
size -= copied;
|
|
|
|
}
|
|
|
|
}
|
2010-03-24 18:49:50 +08:00
|
|
|
|
2012-04-11 20:28:06 +08:00
|
|
|
static int __memcpy_real(void *dest, void *src, size_t count)
|
2010-03-24 18:49:50 +08:00
|
|
|
{
|
|
|
|
register unsigned long _dest asm("2") = (unsigned long) dest;
|
|
|
|
register unsigned long _len1 asm("3") = (unsigned long) count;
|
|
|
|
register unsigned long _src asm("4") = (unsigned long) src;
|
|
|
|
register unsigned long _len2 asm("5") = (unsigned long) count;
|
|
|
|
int rc = -EFAULT;
|
|
|
|
|
|
|
|
asm volatile (
|
|
|
|
"0: mvcle %1,%2,0x0\n"
|
|
|
|
"1: jo 0b\n"
|
|
|
|
" lhi %0,0x0\n"
|
|
|
|
"2:\n"
|
|
|
|
EX_TABLE(1b,2b)
|
|
|
|
: "+d" (rc), "+d" (_dest), "+d" (_src), "+d" (_len1),
|
|
|
|
"+d" (_len2), "=m" (*((long *) dest))
|
|
|
|
: "m" (*((long *) src))
|
|
|
|
: "cc", "memory");
|
2012-04-11 20:28:06 +08:00
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Copy memory in real mode (kernel to kernel)
|
|
|
|
*/
|
|
|
|
int memcpy_real(void *dest, void *src, size_t count)
|
|
|
|
{
|
2016-02-12 19:40:31 +08:00
|
|
|
int irqs_disabled, rc;
|
2012-04-11 20:28:06 +08:00
|
|
|
unsigned long flags;
|
|
|
|
|
|
|
|
if (!count)
|
|
|
|
return 0;
|
2016-02-12 19:40:31 +08:00
|
|
|
flags = __arch_local_irq_stnsm(0xf8UL);
|
|
|
|
irqs_disabled = arch_irqs_disabled_flags(flags);
|
|
|
|
if (!irqs_disabled)
|
|
|
|
trace_hardirqs_off();
|
2012-04-11 20:28:06 +08:00
|
|
|
rc = __memcpy_real(dest, src, count);
|
2016-02-12 19:40:31 +08:00
|
|
|
if (!irqs_disabled)
|
2016-02-09 23:23:39 +08:00
|
|
|
trace_hardirqs_on();
|
|
|
|
__arch_local_irq_ssm(flags);
|
2010-03-24 18:49:50 +08:00
|
|
|
return rc;
|
|
|
|
}
|
2011-08-03 22:44:19 +08:00
|
|
|
|
|
|
|
/*
|
2012-05-24 20:35:16 +08:00
|
|
|
* Copy memory in absolute mode (kernel to kernel)
|
2011-08-03 22:44:19 +08:00
|
|
|
*/
|
2012-05-24 20:35:16 +08:00
|
|
|
void memcpy_absolute(void *dest, void *src, size_t count)
|
2011-08-03 22:44:19 +08:00
|
|
|
{
|
2012-05-24 20:35:16 +08:00
|
|
|
unsigned long cr0, flags, prefix;
|
2011-08-03 22:44:19 +08:00
|
|
|
|
2012-05-24 20:35:16 +08:00
|
|
|
flags = arch_local_irq_save();
|
2011-08-03 22:44:19 +08:00
|
|
|
__ctl_store(cr0, 0, 0);
|
|
|
|
__ctl_clear_bit(0, 28); /* disable lowcore protection */
|
2012-05-24 20:35:16 +08:00
|
|
|
prefix = store_prefix();
|
|
|
|
if (prefix) {
|
|
|
|
local_mcck_disable();
|
|
|
|
set_prefix(0);
|
|
|
|
memcpy(dest, src, count);
|
|
|
|
set_prefix(prefix);
|
|
|
|
local_mcck_enable();
|
|
|
|
} else {
|
|
|
|
memcpy(dest, src, count);
|
|
|
|
}
|
2011-08-03 22:44:19 +08:00
|
|
|
__ctl_load(cr0, 0, 0);
|
2012-05-24 20:35:16 +08:00
|
|
|
arch_local_irq_restore(flags);
|
2011-08-03 22:44:19 +08:00
|
|
|
}
|
2011-10-30 22:16:39 +08:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Copy memory from kernel (real) to user (virtual)
|
|
|
|
*/
|
2014-01-24 19:51:27 +08:00
|
|
|
int copy_to_user_real(void __user *dest, void *src, unsigned long count)
|
2011-10-30 22:16:39 +08:00
|
|
|
{
|
|
|
|
int offs = 0, size, rc;
|
|
|
|
char *buf;
|
|
|
|
|
|
|
|
buf = (char *) __get_free_page(GFP_KERNEL);
|
|
|
|
if (!buf)
|
|
|
|
return -ENOMEM;
|
|
|
|
rc = -EFAULT;
|
|
|
|
while (offs < count) {
|
|
|
|
size = min(PAGE_SIZE, count - offs);
|
|
|
|
if (memcpy_real(buf, src + offs, size))
|
|
|
|
goto out;
|
|
|
|
if (copy_to_user(dest + offs, buf, size))
|
|
|
|
goto out;
|
|
|
|
offs += size;
|
|
|
|
}
|
|
|
|
rc = 0;
|
|
|
|
out:
|
|
|
|
free_page((unsigned long) buf);
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
s390: allow absolute memory access for /dev/mem
Currently dev/mem for s390 provides only real memory access. This means
that the CPU prefix pages are swapped. The prefix swap for real memory
works as follows:
Each CPU owns a prefix register that points to a page aligned memory
location "P". If this CPU accesses the address range [0,0x1fff], it is
translated by the hardware to [P,P+0x1fff]. Accordingly if this CPU
accesses the address range [P,P+0x1fff], it is translated by the hardware
to [0,0x1fff]. Therefore, if [P,P+0x1fff] or [0,0x1fff] is read from
the current /dev/mem device, the incorrectly swapped memory content is
returned.
With this patch the /dev/mem architecture code is modified to provide
absolute memory access. This is done via the arch specific functions
xlate_dev_mem_ptr() and unxlate_dev_mem_ptr(). For swapped pages on
s390 the function xlate_dev_mem_ptr() now returns a new buffer with a
copy of the requested absolute memory. In case the buffer was allocated,
the unxlate_dev_mem_ptr() function frees it after /dev/mem code has
called copy_to_user().
Signed-off-by: Michael Holzheu <holzheu@linux.vnet.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
2012-05-09 22:27:36 +08:00
|
|
|
/*
|
|
|
|
* Check if physical address is within prefix or zero page
|
|
|
|
*/
|
|
|
|
static int is_swapped(unsigned long addr)
|
|
|
|
{
|
|
|
|
unsigned long lc;
|
|
|
|
int cpu;
|
|
|
|
|
2015-12-31 17:29:00 +08:00
|
|
|
if (addr < sizeof(struct lowcore))
|
s390: allow absolute memory access for /dev/mem
Currently dev/mem for s390 provides only real memory access. This means
that the CPU prefix pages are swapped. The prefix swap for real memory
works as follows:
Each CPU owns a prefix register that points to a page aligned memory
location "P". If this CPU accesses the address range [0,0x1fff], it is
translated by the hardware to [P,P+0x1fff]. Accordingly if this CPU
accesses the address range [P,P+0x1fff], it is translated by the hardware
to [0,0x1fff]. Therefore, if [P,P+0x1fff] or [0,0x1fff] is read from
the current /dev/mem device, the incorrectly swapped memory content is
returned.
With this patch the /dev/mem architecture code is modified to provide
absolute memory access. This is done via the arch specific functions
xlate_dev_mem_ptr() and unxlate_dev_mem_ptr(). For swapped pages on
s390 the function xlate_dev_mem_ptr() now returns a new buffer with a
copy of the requested absolute memory. In case the buffer was allocated,
the unxlate_dev_mem_ptr() function frees it after /dev/mem code has
called copy_to_user().
Signed-off-by: Michael Holzheu <holzheu@linux.vnet.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
2012-05-09 22:27:36 +08:00
|
|
|
return 1;
|
|
|
|
for_each_online_cpu(cpu) {
|
|
|
|
lc = (unsigned long) lowcore_ptr[cpu];
|
2015-12-31 17:29:00 +08:00
|
|
|
if (addr > lc + sizeof(struct lowcore) - 1 || addr < lc)
|
s390: allow absolute memory access for /dev/mem
Currently dev/mem for s390 provides only real memory access. This means
that the CPU prefix pages are swapped. The prefix swap for real memory
works as follows:
Each CPU owns a prefix register that points to a page aligned memory
location "P". If this CPU accesses the address range [0,0x1fff], it is
translated by the hardware to [P,P+0x1fff]. Accordingly if this CPU
accesses the address range [P,P+0x1fff], it is translated by the hardware
to [0,0x1fff]. Therefore, if [P,P+0x1fff] or [0,0x1fff] is read from
the current /dev/mem device, the incorrectly swapped memory content is
returned.
With this patch the /dev/mem architecture code is modified to provide
absolute memory access. This is done via the arch specific functions
xlate_dev_mem_ptr() and unxlate_dev_mem_ptr(). For swapped pages on
s390 the function xlate_dev_mem_ptr() now returns a new buffer with a
copy of the requested absolute memory. In case the buffer was allocated,
the unxlate_dev_mem_ptr() function frees it after /dev/mem code has
called copy_to_user().
Signed-off-by: Michael Holzheu <holzheu@linux.vnet.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
2012-05-09 22:27:36 +08:00
|
|
|
continue;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Convert a physical pointer for /dev/mem access
|
|
|
|
*
|
|
|
|
* For swapped prefix pages a new buffer is returned that contains a copy of
|
|
|
|
* the absolute memory. The buffer size is maximum one page large.
|
|
|
|
*/
|
2014-07-28 23:20:33 +08:00
|
|
|
void *xlate_dev_mem_ptr(phys_addr_t addr)
|
s390: allow absolute memory access for /dev/mem
Currently dev/mem for s390 provides only real memory access. This means
that the CPU prefix pages are swapped. The prefix swap for real memory
works as follows:
Each CPU owns a prefix register that points to a page aligned memory
location "P". If this CPU accesses the address range [0,0x1fff], it is
translated by the hardware to [P,P+0x1fff]. Accordingly if this CPU
accesses the address range [P,P+0x1fff], it is translated by the hardware
to [0,0x1fff]. Therefore, if [P,P+0x1fff] or [0,0x1fff] is read from
the current /dev/mem device, the incorrectly swapped memory content is
returned.
With this patch the /dev/mem architecture code is modified to provide
absolute memory access. This is done via the arch specific functions
xlate_dev_mem_ptr() and unxlate_dev_mem_ptr(). For swapped pages on
s390 the function xlate_dev_mem_ptr() now returns a new buffer with a
copy of the requested absolute memory. In case the buffer was allocated,
the unxlate_dev_mem_ptr() function frees it after /dev/mem code has
called copy_to_user().
Signed-off-by: Michael Holzheu <holzheu@linux.vnet.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
2012-05-09 22:27:36 +08:00
|
|
|
{
|
|
|
|
void *bounce = (void *) addr;
|
|
|
|
unsigned long size;
|
|
|
|
|
|
|
|
get_online_cpus();
|
|
|
|
preempt_disable();
|
|
|
|
if (is_swapped(addr)) {
|
|
|
|
size = PAGE_SIZE - (addr & ~PAGE_MASK);
|
|
|
|
bounce = (void *) __get_free_page(GFP_ATOMIC);
|
|
|
|
if (bounce)
|
2012-05-24 20:35:16 +08:00
|
|
|
memcpy_absolute(bounce, (void *) addr, size);
|
s390: allow absolute memory access for /dev/mem
Currently dev/mem for s390 provides only real memory access. This means
that the CPU prefix pages are swapped. The prefix swap for real memory
works as follows:
Each CPU owns a prefix register that points to a page aligned memory
location "P". If this CPU accesses the address range [0,0x1fff], it is
translated by the hardware to [P,P+0x1fff]. Accordingly if this CPU
accesses the address range [P,P+0x1fff], it is translated by the hardware
to [0,0x1fff]. Therefore, if [P,P+0x1fff] or [0,0x1fff] is read from
the current /dev/mem device, the incorrectly swapped memory content is
returned.
With this patch the /dev/mem architecture code is modified to provide
absolute memory access. This is done via the arch specific functions
xlate_dev_mem_ptr() and unxlate_dev_mem_ptr(). For swapped pages on
s390 the function xlate_dev_mem_ptr() now returns a new buffer with a
copy of the requested absolute memory. In case the buffer was allocated,
the unxlate_dev_mem_ptr() function frees it after /dev/mem code has
called copy_to_user().
Signed-off-by: Michael Holzheu <holzheu@linux.vnet.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
2012-05-09 22:27:36 +08:00
|
|
|
}
|
|
|
|
preempt_enable();
|
|
|
|
put_online_cpus();
|
|
|
|
return bounce;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Free converted buffer for /dev/mem access (if necessary)
|
|
|
|
*/
|
2014-07-28 23:20:33 +08:00
|
|
|
void unxlate_dev_mem_ptr(phys_addr_t addr, void *buf)
|
s390: allow absolute memory access for /dev/mem
Currently dev/mem for s390 provides only real memory access. This means
that the CPU prefix pages are swapped. The prefix swap for real memory
works as follows:
Each CPU owns a prefix register that points to a page aligned memory
location "P". If this CPU accesses the address range [0,0x1fff], it is
translated by the hardware to [P,P+0x1fff]. Accordingly if this CPU
accesses the address range [P,P+0x1fff], it is translated by the hardware
to [0,0x1fff]. Therefore, if [P,P+0x1fff] or [0,0x1fff] is read from
the current /dev/mem device, the incorrectly swapped memory content is
returned.
With this patch the /dev/mem architecture code is modified to provide
absolute memory access. This is done via the arch specific functions
xlate_dev_mem_ptr() and unxlate_dev_mem_ptr(). For swapped pages on
s390 the function xlate_dev_mem_ptr() now returns a new buffer with a
copy of the requested absolute memory. In case the buffer was allocated,
the unxlate_dev_mem_ptr() function frees it after /dev/mem code has
called copy_to_user().
Signed-off-by: Michael Holzheu <holzheu@linux.vnet.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
2012-05-09 22:27:36 +08:00
|
|
|
{
|
|
|
|
if ((void *) addr != buf)
|
|
|
|
free_page((unsigned long) buf);
|
|
|
|
}
|