The comment explaining why we modify VRSAVE is misleading, glibc
does rely on the behaviour. Update the comment.
Signed-off-by: Anton Blanchard <anton@samba.org>
Reviewed-by: Cyril Bur <cyrilbur@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This patch adds the ability to be able to save the VSX registers to the
thread struct without giving up (disabling the facility) next time the
process returns to userspace.
This patch builds on a previous optimisation for the FPU and VEC registers
in the thread copy path to avoid a possibly pointless reload of VSX state.
Signed-off-by: Cyril Bur <cyrilbur@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This patch adds the ability to be able to save the VEC registers to the
thread struct without giving up (disabling the facility) next time the
process returns to userspace.
This patch builds on a previous optimisation for the FPU registers in the
thread copy path to avoid a possibly pointless reload of VEC state.
Signed-off-by: Cyril Bur <cyrilbur@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Currently the FPU, VEC and VSX facilities are lazily loaded. This is not
a problem unless a process is using these facilities.
Modern versions of GCC are very good at automatically vectorising code,
new and modernised workloads make use of floating point and vector
facilities, even the kernel makes use of vectorised memcpy.
All this combined greatly increases the cost of a syscall since the
kernel uses the facilities sometimes even in syscall fast-path making it
increasingly common for a thread to take an *_unavailable exception soon
after a syscall, not to mention potentially taking all three.
The obvious overcompensation to this problem is to simply always load
all the facilities on every exit to userspace. Loading up all FPU, VEC
and VSX registers every time can be expensive and if a workload does
avoid using them, it should not be forced to incur this penalty.
An 8bit counter is used to detect if the registers have been used in the
past and the registers are always loaded until the value wraps to back
to zero.
Several versions of the assembly in entry_64.S were tested:
1. Always calling C.
2. Performing a common case check and then calling C.
3. A complex check in asm.
After some benchmarking it was determined that avoiding C in the common
case is a performance benefit (option 2). The full check in asm (option
3) greatly complicated that codepath for a negligible performance gain
and the trade-off was deemed not worth it.
Signed-off-by: Cyril Bur <cyrilbur@gmail.com>
[mpe: Move load_vec in the struct to fill an existing hole, reword change log]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
fixup
More consolidation of our MSR available bit handling.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Move the MSR modification into c. Removing it from the assembly
function will allow us to avoid costly MSR writes by batching them
up.
Check the FP and VMX bits before calling the relevant giveup_*()
function. This makes giveup_vsx() and flush_vsx_to_thread() perform
more like their sister functions, and allows us to use
flush_vsx_to_thread() in the signal code.
Move the check_if_tm_restore_required() check in.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Move the MSR modification into new c functions. Removing it from
the low level functions will allow us to avoid costly MSR writes
by batching them up.
Move the check_if_tm_restore_required() check into these new functions.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
We used to allow giveup_*() to be called with a NULL task struct
pointer. Now those cases are handled in the caller we can remove
the checks. We can also remove giveup_altivec_notask() which is also
unused.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The UP only lazy floating point and vector optimisations were written
back when SMP was not common, and neither glibc nor gcc used vector
instructions. Now SMP is very common, glibc aggressively uses vector
instructions and gcc autovectorises.
We want to add new optimisations that apply to both UP and SMP, but
in preparation for that remove these UP only optimisations.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Similar to the non TM load_up_*() functions, don't disable the MSR
bits on the way out.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
As our various loops (copy, string, crypto etc) get more complicated,
we want to share implementations between userspace (eg glibc) and
the kernel. We also want to write userspace test harnesses to put
in tools/testing/selftest.
One gratuitous difference between userspace and the kernel is the
VMX register definitions - the kernel uses vrX whereas both gcc and
glibc use vX.
Change the kernel to match userspace.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Currently, when we have a process using the transactional memory
facilities on POWER8 (that is, the processor is in transactional
or suspended state), and the process enters the kernel and the
kernel then uses the floating-point or vector (VMX/Altivec) facility,
we end up corrupting the user-visible FP/VMX/VSX state. This
happens, for example, if a page fault causes a copy-on-write
operation, because the copy_page function will use VMX to do the
copy on POWER8. The test program below demonstrates the bug.
The bug happens because when FP/VMX state for a transactional process
is stored in the thread_struct, we store the checkpointed state in
.fp_state/.vr_state and the transactional (current) state in
.transact_fp/.transact_vr. However, when the kernel wants to use
FP/VMX, it calls enable_kernel_fp() or enable_kernel_altivec(),
which saves the current state in .fp_state/.vr_state. Furthermore,
when we return to the user process we return with FP/VMX/VSX
disabled. The next time the process uses FP/VMX/VSX, we don't know
which set of state (the current register values, .fp_state/.vr_state,
or .transact_fp/.transact_vr) we should be using, since we have no
way to tell if we are still in the same transaction, and if not,
whether the previous transaction succeeded or failed.
Thus it is necessary to strictly adhere to the rule that if FP has
been enabled at any point in a transaction, we must keep FP enabled
for the user process with the current transactional state in the
FP registers, until we detect that it is no longer in a transaction.
Similarly for VMX; once enabled it must stay enabled until the
process is no longer transactional.
In order to keep this rule, we add a new thread_info flag which we
test when returning from the kernel to userspace, called TIF_RESTORE_TM.
This flag indicates that there is FP/VMX/VSX state to be restored
before entering userspace, and when it is set the .tm_orig_msr field
in the thread_struct indicates what state needs to be restored.
The restoration is done by restore_tm_state(). The TIF_RESTORE_TM
bit is set by new giveup_fpu/altivec_maybe_transactional helpers,
which are called from enable_kernel_fp/altivec, giveup_vsx, and
flush_fp/altivec_to_thread instead of giveup_fpu/altivec.
The other thing to be done is to get the transactional FP/VMX/VSX
state from .fp_state/.vr_state when doing reclaim, if that state
has been saved there by giveup_fpu/altivec_maybe_transactional.
Having done this, we set the FP/VMX bit in the thread's MSR after
reclaim to indicate that that part of the state is now valid
(having been reclaimed from the processor's checkpointed state).
Finally, in the signal handling code, we move the clearing of the
transactional state bits in the thread's MSR a bit earlier, before
calling flush_fp_to_thread(), so that we don't unnecessarily set
the TIF_RESTORE_TM bit.
This is the test program:
/* Michael Neuling 4/12/2013
*
* See if the altivec state is leaked out of an aborted transaction due to
* kernel vmx copy loops.
*
* gcc -m64 htm_vmxcopy.c -o htm_vmxcopy
*
*/
/* We don't use all of these, but for reference: */
int main(int argc, char *argv[])
{
long double vecin = 1.3;
long double vecout;
unsigned long pgsize = getpagesize();
int i;
int fd;
int size = pgsize*16;
char tmpfile[] = "/tmp/page_faultXXXXXX";
char buf[pgsize];
char *a;
uint64_t aborted = 0;
fd = mkstemp(tmpfile);
assert(fd >= 0);
memset(buf, 0, pgsize);
for (i = 0; i < size; i += pgsize)
assert(write(fd, buf, pgsize) == pgsize);
unlink(tmpfile);
a = mmap(NULL, size, PROT_READ|PROT_WRITE, MAP_PRIVATE, fd, 0);
assert(a != MAP_FAILED);
asm __volatile__(
"lxvd2x 40,0,%[vecinptr] ; " // set 40 to initial value
TBEGIN
"beq 3f ;"
TSUSPEND
"xxlxor 40,40,40 ; " // set 40 to 0
"std 5, 0(%[map]) ;" // cause kernel vmx copy page
TABORT
TRESUME
TEND
"li %[res], 0 ;"
"b 5f ;"
"3: ;" // Abort handler
"li %[res], 1 ;"
"5: ;"
"stxvd2x 40,0,%[vecoutptr] ; "
: [res]"=r"(aborted)
: [vecinptr]"r"(&vecin),
[vecoutptr]"r"(&vecout),
[map]"r"(a)
: "memory", "r0", "r3", "r4", "r5", "r6", "r7");
if (aborted && (vecin != vecout)){
printf("FAILED: vector state leaked on abort %f != %f\n",
(double)vecin, (double)vecout);
exit(1);
}
munmap(a, size);
close(fd);
printf("PASSED!\n");
return 0;
}
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Commit de79f7b9f6 ("powerpc: Put FP/VSX and VR state into structures")
modified load_up_fpu() and load_up_altivec() in such a way that they
now use r7 and r8. Unfortunately, the callers of these functions on
32-bit machines then return to userspace via fast_exception_return,
which doesn't restore all of the volatile GPRs, but only r1, r3 -- r6
and r9 -- r12. This was causing userspace segfaults and other
userspace misbehaviour on 32-bit machines.
This fixes the problem by changing the register usage of load_up_fpu()
and load_up_altivec() to avoid using r7 and r8 and instead use r6 and
r10. This also adds comments to those functions saying which registers
may be used.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Tested-by: Scott Wood <scottwood@freescale.com> (on e500mc, so no altivec)
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This provides a facility which is intended for use by KVM, where the
contents of the FP/VSX and VMX (Altivec) registers can be saved away
to somewhere other than the thread_struct when kernel code wants to
use floating point or VMX instructions. This is done by providing a
pointer in the thread_struct to indicate where the state should be
saved to. The giveup_fpu() and giveup_altivec() functions test these
pointers and save state to the indicated location if they are non-NULL.
Note that the MSR_FP/VEC bits in task->thread.regs->msr are still used
to indicate whether the CPU register state is live, even when an
alternate save location is being used.
This also provides load_fp_state() and load_vr_state() functions, which
load up FP/VSX and VMX state from memory into the CPU registers, and
corresponding store_fp_state() and store_vr_state() functions, which
store FP/VSX and VMX state into memory from the CPU registers.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This creates new 'thread_fp_state' and 'thread_vr_state' structures
to store FP/VSX state (including FPSCR) and Altivec/VSX state
(including VSCR), and uses them in the thread_struct. In the
thread_fp_state, the FPRs and VSRs are represented as u64 rather
than double, since we rarely perform floating-point computations
on the values, and this will enable the structures to be used
in KVM code as well. Similarly FPSCR is now a u64 rather than
a structure of two 32-bit values.
This takes the offsets out of the macros such as SAVE_32FPRS,
REST_32FPRS, etc. This enables the same macros to be used for normal
and transactional state, enabling us to delete the transactional
versions of the macros. This also removes the unused do_load_up_fpu
and do_load_up_altivec, which were in fact buggy since they didn't
create large enough stack frames to account for the fact that
load_up_fpu and load_up_altivec are not designed to be called from C
and assume that their caller's stack frame is an interrupt frame.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This adds functions to restore the state of the FP/VSX registers from
what's stored in the thread_struct. Two version for FP/VSX are required
since one restores them from transactional/checkpoint side of the
thread_struct and the other from the speculated side.
Similar functions are added for VMX registers.
Signed-off-by: Matt Evans <matt@ozlabs.org>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Add two optimisations to enable_kernel_altivec:
- enable_kernel_altivec has already determined if we need to
save the previous task's state but we call giveup_altivec
in both cases, requiring an extra branch in giveup_altivec. Create
giveup_altivec_notask which only turns on the VMX bit in the
MSR.
- We write the VMX MSR bit each time we call enable_kernel_altivec
even it was already set. Check the bit and branch out if we have
already set it. The classic case for this is vectored IO
where we have to copy multiple buffers to or from userspace.
The following testcase was used to confirm this patch improves
performance:
http://ozlabs.org/~anton/junkcode/copy_to_user.c
Since the current breakpoint for using VMX in copy_tofrom_user is
4096 bytes, I'm using buffers of 4096 + 1 cacheline (4224) bytes.
A benchmark of 16 entry readvs (-s 16):
time copy_to_user -l 4224 -s 16 -i 1000000
completes 5.2% faster on a POWER7 PS700.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
A static branch hint will override dynamic branch prediction on
recent POWER CPUs. Since we are about to use more altivec in the
kernel remove the static hint in giveup_altivec that assumes
a userspace task is using altivec.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Since STACK_FRAME_OVERHEAD is defined in asm/ptrace.h and that
is ASSEMBER safe, we can just include that instead of going via
asm-offsets.h.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
e821ea70f3 introduced a bug by copying
some 64-bit originated code as-is to be used by both 32 and 64-bit
but this code contains a 64-bit ony "cmpdi" instruction.
This changes it to cmpwi, which is fine since VRSAVE can only contains
a 32-bit value anyway.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
CC: <stable@kernel.org>
The kernel uses SPRG registers for various purposes, typically in
low level assembly code as scratch registers or to hold per-cpu
global infos such as the PACA or the current thread_info pointer.
We want to be able to easily shuffle the usage of those registers
as some implementations have specific constraints realted to some
of them, for example, some have userspace readable aliases, etc..
and the current choice isn't always the best.
This patch should not change any code generation, and replaces the
usage of SPRN_SPRGn everywhere in the kernel with a named replacement
and adds documentation next to the definition of the names as to
what those are used for on each processor family.
The only parts that still use the original numbers are bits of KVM
or suspend/resume code that just blindly needs to save/restore all
the SPRGs.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
When moving load_up_altivec to vector.S a typo in a comment caused a
thinko setting the wrong variable.
Signed-off-by: Andreas Schwab <schwab@linux-m68k.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Currently, load_up_altivec and give_up_altivec are duplicated
in 32-bit and 64-bit. This creates a common implementation that
is moved away from head_32.S, head_64.S and misc_64.S and into
vector.S, using the same macros we already use for our common
implementation of load_up_fpu.
I also moved the VSX code over to vector.S though in that case
I didn't make it build on 32-bit (yet).
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Forthcoming machines will extend the FPSCR to 64 bits. We already
had a 64-bit save area for the FPSCR, but we need to use a new form
of the mtfsf instruction. Fortunately this new form is decoded as
an ordinary mtfsf by existing 64-bit processors.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Now that the register names and bit definitions are all in reg.h,
use that instead of processor.h in assembly code in a few places.
Signed-off-by: Paul Mackerras <paulus@samba.org>
This creates the directory structure under arch/powerpc and a bunch
of Kconfig files. It does a first-cut merge of arch/powerpc/mm,
arch/powerpc/lib and arch/powerpc/platforms/powermac. This is enough
to build a 32-bit powermac kernel with ARCH=powerpc.
For now we are getting some unmerged files from arch/ppc/kernel and
arch/ppc/syslib, or arch/ppc64/kernel. This makes some minor changes
to files in those directories and files outside arch/powerpc.
The boot directory is still not merged. That's going to be interesting.
Signed-off-by: Paul Mackerras <paulus@samba.org>
