x86: re-introduce non-generic memcpy_{to,from}io
This has been broken forever, and nobody ever really noticed because
it's purely a performance issue.
Long long ago, in commit 6175ddf06b61 ("x86: Clean up mem*io functions")
Brian Gerst simplified the memory copies to and from iomem, since on
x86, the instructions to access iomem are exactly the same as the
regular instructions.
That is technically true, and things worked, and nobody said anything.
Besides, back then the regular memcpy was pretty simple and worked fine.
Nobody noticed except for David Laight, that is. David has a testing a
TLP monitor he was writing for an FPGA, and has been occasionally
complaining about how memcpy_toio() writes things one byte at a time.
Which is completely unacceptable from a performance standpoint, even if
it happens to technically work.
The reason it's writing one byte at a time is because while it's
technically true that accesses to iomem are the same as accesses to
regular memory on x86, the _granularity_ (and ordering) of accesses
matter to iomem in ways that they don't matter to regular cached memory.
In particular, when ERMS is set, we default to using "rep movsb" for
larger memory copies. That is indeed perfectly fine for real memory,
since the whole point is that the CPU is going to do cacheline
optimizations and executes the memory copy efficiently for cached
memory.
With iomem? Not so much. With iomem, "rep movsb" will indeed work, but
it will copy things one byte at a time. Slowly and ponderously.
Now, originally, back in 2010 when commit 6175ddf06b61 was done, we
didn't use ERMS, and this was much less noticeable.
Our normal memcpy() was simpler in other ways too.
Because in fact, it's not just about using the string instructions. Our
memcpy() these days does things like "read and write overlapping values"
to handle the last bytes of the copy. Again, for normal memory,
overlapping accesses isn't an issue. For iomem? It can be.
So this re-introduces the specialized memcpy_toio(), memcpy_fromio() and
memset_io() functions. It doesn't particularly optimize them, but it
tries to at least not be horrid, or do overlapping accesses. In fact,
this uses the existing __inline_memcpy() function that we still had
lying around that uses our very traditional "rep movsl" loop followed by
movsw/movsb for the final bytes.
Somebody may decide to try to improve on it, but if we've gone almost a
decade with only one person really ever noticing and complaining, maybe
it's not worth worrying about further, once it's not _completely_ broken?
Reported-by: David Laight <David.Laight@aculab.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-05 04:52:49 +03:00
|
|
|
#include <linux/string.h>
|
|
|
|
|
#include <linux/module.h>
|
|
|
|
|
#include <linux/io.h>
|
|
|
|
|
|
x86: explicitly align IO accesses in memcpy_{to,from}io
In commit 170d13ca3a2f ("x86: re-introduce non-generic memcpy_{to,from}io")
I made our copy from IO space use a separate copy routine rather than
rely on the generic memcpy. I did that because our generic memory copy
isn't actually well-defined when it comes to internal access ordering or
alignment, and will in fact depend on various CPUID flags.
In particular, the default memcpy() for a modern Intel CPU will
generally be just a "rep movsb", which works reasonably well for
medium-sized memory copies of regular RAM, since the CPU will turn it
into fairly optimized microcode.
However, for non-cached memory and IO, "rep movs" ends up being
horrendously slow and will just do the architectural "one byte at a
time" accesses implied by the movsb.
At the other end of the spectrum, if you _don't_ end up using the "rep
movsb" code, you'd likely fall back to the software copy, which does
overlapping accesses for the tail, and may copy things backwards.
Again, for regular memory that's fine, for IO memory not so much.
The thinking was that clearly nobody really cared (because things
worked), but some people had seen horrible performance due to the byte
accesses, so let's just revert back to our long ago version that dod
"rep movsl" for the bulk of the copy, and then fixed up the potentially
last few bytes of the tail with "movsw/b".
Interestingly (and perhaps not entirely surprisingly), while that was
our original memory copy implementation, and had been used before for
IO, in the meantime many new users of memcpy_*io() had come about. And
while the access patterns for the memory copy weren't well-defined (so
arguably _any_ access pattern should work), in practice the "rep movsb"
case had been very common for the last several years.
In particular Jarkko Sakkinen reported that the memcpy_*io() change
resuled in weird errors from his Geminilake NUC TPM module.
And it turns out that the TPM TCG accesses according to spec require
that the accesses be
(a) done strictly sequentially
(b) be naturally aligned
otherwise the TPM chip will abort the PCI transaction.
And, in fact, the tpm_crb.c driver did this:
memcpy_fromio(buf, priv->rsp, 6);
...
memcpy_fromio(&buf[6], &priv->rsp[6], expected - 6);
which really should never have worked in the first place, but back
before commit 170d13ca3a2f it *happened* to work, because the
memcpy_fromio() would be expanded to a regular memcpy, and
(a) gcc would expand the first memcpy in-line, and turn it into a
4-byte and a 2-byte read, and they happened to be in the right
order, and the alignment was right.
(b) gcc would call "memcpy()" for the second one, and the machines that
had this TPM chip also apparently ended up always having ERMS
("Enhanced REP MOVSB/STOSB instructions"), so we'd use the "rep
movbs" for that copy.
In other words, basically by pure luck, the code happened to use the
right access sizes in the (two different!) memcpy() implementations to
make it all work.
But after commit 170d13ca3a2f, both of the memcpy_fromio() calls
resulted in a call to the routine with the consistent memory accesses,
and in both cases it started out transferring with 4-byte accesses.
Which worked for the first copy, but resulted in the second copy doing a
32-bit read at an address that was only 2-byte aligned.
Jarkko is actually fixing the fragile code in the TPM driver, but since
this is an excellent example of why we absolutely must not use a generic
memcpy for IO accesses, _and_ an IO-specific one really should strive to
align the IO accesses, let's do exactly that.
Side note: Jarkko also noted that the driver had been used on ARM
platforms, and had worked. That was because on 32-bit ARM, memcpy_*io()
ends up always doing byte accesses, and on 64-bit ARM it first does byte
accesses to align to 8-byte boundaries, and then does 8-byte accesses
for the bulk.
So ARM actually worked by design, and the x86 case worked by pure luck.
We *might* want to make x86-64 do the 8-byte case too. That should be a
pretty straightforward extension, but let's do one thing at a time. And
generally MMIO accesses aren't really all that performance-critical, as
shown by the fact that for a long time we just did them a byte at a
time, and very few people ever noticed.
Reported-and-tested-by: Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>
Tested-by: Jerry Snitselaar <jsnitsel@redhat.com>
Cc: David Laight <David.Laight@aculab.com>
Fixes: 170d13ca3a2f ("x86: re-introduce non-generic memcpy_{to,from}io")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-31 22:10:20 +03:00
|
|
|
#define movs(type,to,from) \
|
|
|
|
|
asm volatile("movs" type:"=&D" (to), "=&S" (from):"0" (to), "1" (from):"memory")
|
|
|
|
|
|
x86: re-introduce non-generic memcpy_{to,from}io
This has been broken forever, and nobody ever really noticed because
it's purely a performance issue.
Long long ago, in commit 6175ddf06b61 ("x86: Clean up mem*io functions")
Brian Gerst simplified the memory copies to and from iomem, since on
x86, the instructions to access iomem are exactly the same as the
regular instructions.
That is technically true, and things worked, and nobody said anything.
Besides, back then the regular memcpy was pretty simple and worked fine.
Nobody noticed except for David Laight, that is. David has a testing a
TLP monitor he was writing for an FPGA, and has been occasionally
complaining about how memcpy_toio() writes things one byte at a time.
Which is completely unacceptable from a performance standpoint, even if
it happens to technically work.
The reason it's writing one byte at a time is because while it's
technically true that accesses to iomem are the same as accesses to
regular memory on x86, the _granularity_ (and ordering) of accesses
matter to iomem in ways that they don't matter to regular cached memory.
In particular, when ERMS is set, we default to using "rep movsb" for
larger memory copies. That is indeed perfectly fine for real memory,
since the whole point is that the CPU is going to do cacheline
optimizations and executes the memory copy efficiently for cached
memory.
With iomem? Not so much. With iomem, "rep movsb" will indeed work, but
it will copy things one byte at a time. Slowly and ponderously.
Now, originally, back in 2010 when commit 6175ddf06b61 was done, we
didn't use ERMS, and this was much less noticeable.
Our normal memcpy() was simpler in other ways too.
Because in fact, it's not just about using the string instructions. Our
memcpy() these days does things like "read and write overlapping values"
to handle the last bytes of the copy. Again, for normal memory,
overlapping accesses isn't an issue. For iomem? It can be.
So this re-introduces the specialized memcpy_toio(), memcpy_fromio() and
memset_io() functions. It doesn't particularly optimize them, but it
tries to at least not be horrid, or do overlapping accesses. In fact,
this uses the existing __inline_memcpy() function that we still had
lying around that uses our very traditional "rep movsl" loop followed by
movsw/movsb for the final bytes.
Somebody may decide to try to improve on it, but if we've gone almost a
decade with only one person really ever noticing and complaining, maybe
it's not worth worrying about further, once it's not _completely_ broken?
Reported-by: David Laight <David.Laight@aculab.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-05 04:52:49 +03:00
|
|
|
/* Originally from i386/string.h */
|
x86: explicitly align IO accesses in memcpy_{to,from}io
In commit 170d13ca3a2f ("x86: re-introduce non-generic memcpy_{to,from}io")
I made our copy from IO space use a separate copy routine rather than
rely on the generic memcpy. I did that because our generic memory copy
isn't actually well-defined when it comes to internal access ordering or
alignment, and will in fact depend on various CPUID flags.
In particular, the default memcpy() for a modern Intel CPU will
generally be just a "rep movsb", which works reasonably well for
medium-sized memory copies of regular RAM, since the CPU will turn it
into fairly optimized microcode.
However, for non-cached memory and IO, "rep movs" ends up being
horrendously slow and will just do the architectural "one byte at a
time" accesses implied by the movsb.
At the other end of the spectrum, if you _don't_ end up using the "rep
movsb" code, you'd likely fall back to the software copy, which does
overlapping accesses for the tail, and may copy things backwards.
Again, for regular memory that's fine, for IO memory not so much.
The thinking was that clearly nobody really cared (because things
worked), but some people had seen horrible performance due to the byte
accesses, so let's just revert back to our long ago version that dod
"rep movsl" for the bulk of the copy, and then fixed up the potentially
last few bytes of the tail with "movsw/b".
Interestingly (and perhaps not entirely surprisingly), while that was
our original memory copy implementation, and had been used before for
IO, in the meantime many new users of memcpy_*io() had come about. And
while the access patterns for the memory copy weren't well-defined (so
arguably _any_ access pattern should work), in practice the "rep movsb"
case had been very common for the last several years.
In particular Jarkko Sakkinen reported that the memcpy_*io() change
resuled in weird errors from his Geminilake NUC TPM module.
And it turns out that the TPM TCG accesses according to spec require
that the accesses be
(a) done strictly sequentially
(b) be naturally aligned
otherwise the TPM chip will abort the PCI transaction.
And, in fact, the tpm_crb.c driver did this:
memcpy_fromio(buf, priv->rsp, 6);
...
memcpy_fromio(&buf[6], &priv->rsp[6], expected - 6);
which really should never have worked in the first place, but back
before commit 170d13ca3a2f it *happened* to work, because the
memcpy_fromio() would be expanded to a regular memcpy, and
(a) gcc would expand the first memcpy in-line, and turn it into a
4-byte and a 2-byte read, and they happened to be in the right
order, and the alignment was right.
(b) gcc would call "memcpy()" for the second one, and the machines that
had this TPM chip also apparently ended up always having ERMS
("Enhanced REP MOVSB/STOSB instructions"), so we'd use the "rep
movbs" for that copy.
In other words, basically by pure luck, the code happened to use the
right access sizes in the (two different!) memcpy() implementations to
make it all work.
But after commit 170d13ca3a2f, both of the memcpy_fromio() calls
resulted in a call to the routine with the consistent memory accesses,
and in both cases it started out transferring with 4-byte accesses.
Which worked for the first copy, but resulted in the second copy doing a
32-bit read at an address that was only 2-byte aligned.
Jarkko is actually fixing the fragile code in the TPM driver, but since
this is an excellent example of why we absolutely must not use a generic
memcpy for IO accesses, _and_ an IO-specific one really should strive to
align the IO accesses, let's do exactly that.
Side note: Jarkko also noted that the driver had been used on ARM
platforms, and had worked. That was because on 32-bit ARM, memcpy_*io()
ends up always doing byte accesses, and on 64-bit ARM it first does byte
accesses to align to 8-byte boundaries, and then does 8-byte accesses
for the bulk.
So ARM actually worked by design, and the x86 case worked by pure luck.
We *might* want to make x86-64 do the 8-byte case too. That should be a
pretty straightforward extension, but let's do one thing at a time. And
generally MMIO accesses aren't really all that performance-critical, as
shown by the fact that for a long time we just did them a byte at a
time, and very few people ever noticed.
Reported-and-tested-by: Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>
Tested-by: Jerry Snitselaar <jsnitsel@redhat.com>
Cc: David Laight <David.Laight@aculab.com>
Fixes: 170d13ca3a2f ("x86: re-introduce non-generic memcpy_{to,from}io")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-31 22:10:20 +03:00
|
|
|
static __always_inline void rep_movs(void *to, const void *from, size_t n)
|
x86: re-introduce non-generic memcpy_{to,from}io
This has been broken forever, and nobody ever really noticed because
it's purely a performance issue.
Long long ago, in commit 6175ddf06b61 ("x86: Clean up mem*io functions")
Brian Gerst simplified the memory copies to and from iomem, since on
x86, the instructions to access iomem are exactly the same as the
regular instructions.
That is technically true, and things worked, and nobody said anything.
Besides, back then the regular memcpy was pretty simple and worked fine.
Nobody noticed except for David Laight, that is. David has a testing a
TLP monitor he was writing for an FPGA, and has been occasionally
complaining about how memcpy_toio() writes things one byte at a time.
Which is completely unacceptable from a performance standpoint, even if
it happens to technically work.
The reason it's writing one byte at a time is because while it's
technically true that accesses to iomem are the same as accesses to
regular memory on x86, the _granularity_ (and ordering) of accesses
matter to iomem in ways that they don't matter to regular cached memory.
In particular, when ERMS is set, we default to using "rep movsb" for
larger memory copies. That is indeed perfectly fine for real memory,
since the whole point is that the CPU is going to do cacheline
optimizations and executes the memory copy efficiently for cached
memory.
With iomem? Not so much. With iomem, "rep movsb" will indeed work, but
it will copy things one byte at a time. Slowly and ponderously.
Now, originally, back in 2010 when commit 6175ddf06b61 was done, we
didn't use ERMS, and this was much less noticeable.
Our normal memcpy() was simpler in other ways too.
Because in fact, it's not just about using the string instructions. Our
memcpy() these days does things like "read and write overlapping values"
to handle the last bytes of the copy. Again, for normal memory,
overlapping accesses isn't an issue. For iomem? It can be.
So this re-introduces the specialized memcpy_toio(), memcpy_fromio() and
memset_io() functions. It doesn't particularly optimize them, but it
tries to at least not be horrid, or do overlapping accesses. In fact,
this uses the existing __inline_memcpy() function that we still had
lying around that uses our very traditional "rep movsl" loop followed by
movsw/movsb for the final bytes.
Somebody may decide to try to improve on it, but if we've gone almost a
decade with only one person really ever noticing and complaining, maybe
it's not worth worrying about further, once it's not _completely_ broken?
Reported-by: David Laight <David.Laight@aculab.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-05 04:52:49 +03:00
|
|
|
{
|
|
|
|
|
unsigned long d0, d1, d2;
|
|
|
|
|
asm volatile("rep ; movsl\n\t"
|
|
|
|
|
"testb $2,%b4\n\t"
|
|
|
|
|
"je 1f\n\t"
|
|
|
|
|
"movsw\n"
|
|
|
|
|
"1:\ttestb $1,%b4\n\t"
|
|
|
|
|
"je 2f\n\t"
|
|
|
|
|
"movsb\n"
|
|
|
|
|
"2:"
|
|
|
|
|
: "=&c" (d0), "=&D" (d1), "=&S" (d2)
|
|
|
|
|
: "0" (n / 4), "q" (n), "1" ((long)to), "2" ((long)from)
|
|
|
|
|
: "memory");
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void memcpy_fromio(void *to, const volatile void __iomem *from, size_t n)
|
|
|
|
|
{
|
x86: explicitly align IO accesses in memcpy_{to,from}io
In commit 170d13ca3a2f ("x86: re-introduce non-generic memcpy_{to,from}io")
I made our copy from IO space use a separate copy routine rather than
rely on the generic memcpy. I did that because our generic memory copy
isn't actually well-defined when it comes to internal access ordering or
alignment, and will in fact depend on various CPUID flags.
In particular, the default memcpy() for a modern Intel CPU will
generally be just a "rep movsb", which works reasonably well for
medium-sized memory copies of regular RAM, since the CPU will turn it
into fairly optimized microcode.
However, for non-cached memory and IO, "rep movs" ends up being
horrendously slow and will just do the architectural "one byte at a
time" accesses implied by the movsb.
At the other end of the spectrum, if you _don't_ end up using the "rep
movsb" code, you'd likely fall back to the software copy, which does
overlapping accesses for the tail, and may copy things backwards.
Again, for regular memory that's fine, for IO memory not so much.
The thinking was that clearly nobody really cared (because things
worked), but some people had seen horrible performance due to the byte
accesses, so let's just revert back to our long ago version that dod
"rep movsl" for the bulk of the copy, and then fixed up the potentially
last few bytes of the tail with "movsw/b".
Interestingly (and perhaps not entirely surprisingly), while that was
our original memory copy implementation, and had been used before for
IO, in the meantime many new users of memcpy_*io() had come about. And
while the access patterns for the memory copy weren't well-defined (so
arguably _any_ access pattern should work), in practice the "rep movsb"
case had been very common for the last several years.
In particular Jarkko Sakkinen reported that the memcpy_*io() change
resuled in weird errors from his Geminilake NUC TPM module.
And it turns out that the TPM TCG accesses according to spec require
that the accesses be
(a) done strictly sequentially
(b) be naturally aligned
otherwise the TPM chip will abort the PCI transaction.
And, in fact, the tpm_crb.c driver did this:
memcpy_fromio(buf, priv->rsp, 6);
...
memcpy_fromio(&buf[6], &priv->rsp[6], expected - 6);
which really should never have worked in the first place, but back
before commit 170d13ca3a2f it *happened* to work, because the
memcpy_fromio() would be expanded to a regular memcpy, and
(a) gcc would expand the first memcpy in-line, and turn it into a
4-byte and a 2-byte read, and they happened to be in the right
order, and the alignment was right.
(b) gcc would call "memcpy()" for the second one, and the machines that
had this TPM chip also apparently ended up always having ERMS
("Enhanced REP MOVSB/STOSB instructions"), so we'd use the "rep
movbs" for that copy.
In other words, basically by pure luck, the code happened to use the
right access sizes in the (two different!) memcpy() implementations to
make it all work.
But after commit 170d13ca3a2f, both of the memcpy_fromio() calls
resulted in a call to the routine with the consistent memory accesses,
and in both cases it started out transferring with 4-byte accesses.
Which worked for the first copy, but resulted in the second copy doing a
32-bit read at an address that was only 2-byte aligned.
Jarkko is actually fixing the fragile code in the TPM driver, but since
this is an excellent example of why we absolutely must not use a generic
memcpy for IO accesses, _and_ an IO-specific one really should strive to
align the IO accesses, let's do exactly that.
Side note: Jarkko also noted that the driver had been used on ARM
platforms, and had worked. That was because on 32-bit ARM, memcpy_*io()
ends up always doing byte accesses, and on 64-bit ARM it first does byte
accesses to align to 8-byte boundaries, and then does 8-byte accesses
for the bulk.
So ARM actually worked by design, and the x86 case worked by pure luck.
We *might* want to make x86-64 do the 8-byte case too. That should be a
pretty straightforward extension, but let's do one thing at a time. And
generally MMIO accesses aren't really all that performance-critical, as
shown by the fact that for a long time we just did them a byte at a
time, and very few people ever noticed.
Reported-and-tested-by: Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>
Tested-by: Jerry Snitselaar <jsnitsel@redhat.com>
Cc: David Laight <David.Laight@aculab.com>
Fixes: 170d13ca3a2f ("x86: re-introduce non-generic memcpy_{to,from}io")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-31 22:10:20 +03:00
|
|
|
if (unlikely(!n))
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
/* Align any unaligned source IO */
|
|
|
|
|
if (unlikely(1 & (unsigned long)from)) {
|
|
|
|
|
movs("b", to, from);
|
|
|
|
|
n--;
|
|
|
|
|
}
|
|
|
|
|
if (n > 1 && unlikely(2 & (unsigned long)from)) {
|
|
|
|
|
movs("w", to, from);
|
|
|
|
|
n-=2;
|
|
|
|
|
}
|
|
|
|
|
rep_movs(to, (const void *)from, n);
|
x86: re-introduce non-generic memcpy_{to,from}io
This has been broken forever, and nobody ever really noticed because
it's purely a performance issue.
Long long ago, in commit 6175ddf06b61 ("x86: Clean up mem*io functions")
Brian Gerst simplified the memory copies to and from iomem, since on
x86, the instructions to access iomem are exactly the same as the
regular instructions.
That is technically true, and things worked, and nobody said anything.
Besides, back then the regular memcpy was pretty simple and worked fine.
Nobody noticed except for David Laight, that is. David has a testing a
TLP monitor he was writing for an FPGA, and has been occasionally
complaining about how memcpy_toio() writes things one byte at a time.
Which is completely unacceptable from a performance standpoint, even if
it happens to technically work.
The reason it's writing one byte at a time is because while it's
technically true that accesses to iomem are the same as accesses to
regular memory on x86, the _granularity_ (and ordering) of accesses
matter to iomem in ways that they don't matter to regular cached memory.
In particular, when ERMS is set, we default to using "rep movsb" for
larger memory copies. That is indeed perfectly fine for real memory,
since the whole point is that the CPU is going to do cacheline
optimizations and executes the memory copy efficiently for cached
memory.
With iomem? Not so much. With iomem, "rep movsb" will indeed work, but
it will copy things one byte at a time. Slowly and ponderously.
Now, originally, back in 2010 when commit 6175ddf06b61 was done, we
didn't use ERMS, and this was much less noticeable.
Our normal memcpy() was simpler in other ways too.
Because in fact, it's not just about using the string instructions. Our
memcpy() these days does things like "read and write overlapping values"
to handle the last bytes of the copy. Again, for normal memory,
overlapping accesses isn't an issue. For iomem? It can be.
So this re-introduces the specialized memcpy_toio(), memcpy_fromio() and
memset_io() functions. It doesn't particularly optimize them, but it
tries to at least not be horrid, or do overlapping accesses. In fact,
this uses the existing __inline_memcpy() function that we still had
lying around that uses our very traditional "rep movsl" loop followed by
movsw/movsb for the final bytes.
Somebody may decide to try to improve on it, but if we've gone almost a
decade with only one person really ever noticing and complaining, maybe
it's not worth worrying about further, once it's not _completely_ broken?
Reported-by: David Laight <David.Laight@aculab.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-05 04:52:49 +03:00
|
|
|
}
|
|
|
|
|
EXPORT_SYMBOL(memcpy_fromio);
|
|
|
|
|
|
|
|
|
|
void memcpy_toio(volatile void __iomem *to, const void *from, size_t n)
|
|
|
|
|
{
|
x86: explicitly align IO accesses in memcpy_{to,from}io
In commit 170d13ca3a2f ("x86: re-introduce non-generic memcpy_{to,from}io")
I made our copy from IO space use a separate copy routine rather than
rely on the generic memcpy. I did that because our generic memory copy
isn't actually well-defined when it comes to internal access ordering or
alignment, and will in fact depend on various CPUID flags.
In particular, the default memcpy() for a modern Intel CPU will
generally be just a "rep movsb", which works reasonably well for
medium-sized memory copies of regular RAM, since the CPU will turn it
into fairly optimized microcode.
However, for non-cached memory and IO, "rep movs" ends up being
horrendously slow and will just do the architectural "one byte at a
time" accesses implied by the movsb.
At the other end of the spectrum, if you _don't_ end up using the "rep
movsb" code, you'd likely fall back to the software copy, which does
overlapping accesses for the tail, and may copy things backwards.
Again, for regular memory that's fine, for IO memory not so much.
The thinking was that clearly nobody really cared (because things
worked), but some people had seen horrible performance due to the byte
accesses, so let's just revert back to our long ago version that dod
"rep movsl" for the bulk of the copy, and then fixed up the potentially
last few bytes of the tail with "movsw/b".
Interestingly (and perhaps not entirely surprisingly), while that was
our original memory copy implementation, and had been used before for
IO, in the meantime many new users of memcpy_*io() had come about. And
while the access patterns for the memory copy weren't well-defined (so
arguably _any_ access pattern should work), in practice the "rep movsb"
case had been very common for the last several years.
In particular Jarkko Sakkinen reported that the memcpy_*io() change
resuled in weird errors from his Geminilake NUC TPM module.
And it turns out that the TPM TCG accesses according to spec require
that the accesses be
(a) done strictly sequentially
(b) be naturally aligned
otherwise the TPM chip will abort the PCI transaction.
And, in fact, the tpm_crb.c driver did this:
memcpy_fromio(buf, priv->rsp, 6);
...
memcpy_fromio(&buf[6], &priv->rsp[6], expected - 6);
which really should never have worked in the first place, but back
before commit 170d13ca3a2f it *happened* to work, because the
memcpy_fromio() would be expanded to a regular memcpy, and
(a) gcc would expand the first memcpy in-line, and turn it into a
4-byte and a 2-byte read, and they happened to be in the right
order, and the alignment was right.
(b) gcc would call "memcpy()" for the second one, and the machines that
had this TPM chip also apparently ended up always having ERMS
("Enhanced REP MOVSB/STOSB instructions"), so we'd use the "rep
movbs" for that copy.
In other words, basically by pure luck, the code happened to use the
right access sizes in the (two different!) memcpy() implementations to
make it all work.
But after commit 170d13ca3a2f, both of the memcpy_fromio() calls
resulted in a call to the routine with the consistent memory accesses,
and in both cases it started out transferring with 4-byte accesses.
Which worked for the first copy, but resulted in the second copy doing a
32-bit read at an address that was only 2-byte aligned.
Jarkko is actually fixing the fragile code in the TPM driver, but since
this is an excellent example of why we absolutely must not use a generic
memcpy for IO accesses, _and_ an IO-specific one really should strive to
align the IO accesses, let's do exactly that.
Side note: Jarkko also noted that the driver had been used on ARM
platforms, and had worked. That was because on 32-bit ARM, memcpy_*io()
ends up always doing byte accesses, and on 64-bit ARM it first does byte
accesses to align to 8-byte boundaries, and then does 8-byte accesses
for the bulk.
So ARM actually worked by design, and the x86 case worked by pure luck.
We *might* want to make x86-64 do the 8-byte case too. That should be a
pretty straightforward extension, but let's do one thing at a time. And
generally MMIO accesses aren't really all that performance-critical, as
shown by the fact that for a long time we just did them a byte at a
time, and very few people ever noticed.
Reported-and-tested-by: Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>
Tested-by: Jerry Snitselaar <jsnitsel@redhat.com>
Cc: David Laight <David.Laight@aculab.com>
Fixes: 170d13ca3a2f ("x86: re-introduce non-generic memcpy_{to,from}io")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-31 22:10:20 +03:00
|
|
|
if (unlikely(!n))
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
/* Align any unaligned destination IO */
|
|
|
|
|
if (unlikely(1 & (unsigned long)to)) {
|
|
|
|
|
movs("b", to, from);
|
|
|
|
|
n--;
|
|
|
|
|
}
|
|
|
|
|
if (n > 1 && unlikely(2 & (unsigned long)to)) {
|
|
|
|
|
movs("w", to, from);
|
|
|
|
|
n-=2;
|
|
|
|
|
}
|
|
|
|
|
rep_movs((void *)to, (const void *) from, n);
|
x86: re-introduce non-generic memcpy_{to,from}io
This has been broken forever, and nobody ever really noticed because
it's purely a performance issue.
Long long ago, in commit 6175ddf06b61 ("x86: Clean up mem*io functions")
Brian Gerst simplified the memory copies to and from iomem, since on
x86, the instructions to access iomem are exactly the same as the
regular instructions.
That is technically true, and things worked, and nobody said anything.
Besides, back then the regular memcpy was pretty simple and worked fine.
Nobody noticed except for David Laight, that is. David has a testing a
TLP monitor he was writing for an FPGA, and has been occasionally
complaining about how memcpy_toio() writes things one byte at a time.
Which is completely unacceptable from a performance standpoint, even if
it happens to technically work.
The reason it's writing one byte at a time is because while it's
technically true that accesses to iomem are the same as accesses to
regular memory on x86, the _granularity_ (and ordering) of accesses
matter to iomem in ways that they don't matter to regular cached memory.
In particular, when ERMS is set, we default to using "rep movsb" for
larger memory copies. That is indeed perfectly fine for real memory,
since the whole point is that the CPU is going to do cacheline
optimizations and executes the memory copy efficiently for cached
memory.
With iomem? Not so much. With iomem, "rep movsb" will indeed work, but
it will copy things one byte at a time. Slowly and ponderously.
Now, originally, back in 2010 when commit 6175ddf06b61 was done, we
didn't use ERMS, and this was much less noticeable.
Our normal memcpy() was simpler in other ways too.
Because in fact, it's not just about using the string instructions. Our
memcpy() these days does things like "read and write overlapping values"
to handle the last bytes of the copy. Again, for normal memory,
overlapping accesses isn't an issue. For iomem? It can be.
So this re-introduces the specialized memcpy_toio(), memcpy_fromio() and
memset_io() functions. It doesn't particularly optimize them, but it
tries to at least not be horrid, or do overlapping accesses. In fact,
this uses the existing __inline_memcpy() function that we still had
lying around that uses our very traditional "rep movsl" loop followed by
movsw/movsb for the final bytes.
Somebody may decide to try to improve on it, but if we've gone almost a
decade with only one person really ever noticing and complaining, maybe
it's not worth worrying about further, once it's not _completely_ broken?
Reported-by: David Laight <David.Laight@aculab.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-05 04:52:49 +03:00
|
|
|
}
|
|
|
|
|
EXPORT_SYMBOL(memcpy_toio);
|
|
|
|
|
|
|
|
|
|
void memset_io(volatile void __iomem *a, int b, size_t c)
|
|
|
|
|
{
|
|
|
|
|
/*
|
|
|
|
|
* TODO: memset can mangle the IO patterns quite a bit.
|
|
|
|
|
* perhaps it would be better to use a dumb one:
|
|
|
|
|
*/
|
|
|
|
|
memset((void *)a, b, c);
|
|
|
|
|
}
|
|
|
|
|
EXPORT_SYMBOL(memset_io);
|