This tests perf hardware breakpoints (ie PERF_TYPE_BREAKPOINT) on
powerpc.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
We now have barrier_nospec as mitigation so print it in
cpu_show_spectre_v1() when enabled.
Signed-off-by: Michal Suchanek <msuchanek@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Our syscall entry is done in assembly so patch in an explicit
barrier_nospec.
Based on a patch by Michal Suchanek.
Signed-off-by: Michal Suchanek <msuchanek@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Based on the x86 commit doing the same.
See commit 304ec1b050 ("x86/uaccess: Use __uaccess_begin_nospec()
and uaccess_try_nospec") and b3bbfb3fb5 ("x86: Introduce
__uaccess_begin_nospec() and uaccess_try_nospec") for more detail.
In all cases we are ordering the load from the potentially
user-controlled pointer vs a previous branch based on an access_ok()
check or similar.
Base on a patch from Michal Suchanek.
Signed-off-by: Michal Suchanek <msuchanek@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Check what firmware told us and enable/disable the barrier_nospec as
appropriate.
We err on the side of enabling the barrier, as it's no-op on older
systems, see the comment for more detail.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Note that unlike RFI which is patched only in kernel the nospec state
reflects settings at the time the module was loaded.
Iterating all modules and re-patching every time the settings change
is not implemented.
Based on lwsync patching.
Signed-off-by: Michal Suchanek <msuchanek@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Based on the RFI patching. This is required to be able to disable the
speculation barrier.
Only one barrier type is supported and it does nothing when the
firmware does not enable it. Also re-patching modules is not supported
So the only meaningful thing that can be done is patching out the
speculation barrier at boot when the user says it is not wanted.
Signed-off-by: Michal Suchanek <msuchanek@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
A no-op form of ori (or immediate of 0 into r31 and the result stored
in r31) has been re-tasked as a speculation barrier. The instruction
only acts as a barrier on newer machines with appropriate firmware
support. On older CPUs it remains a harmless no-op.
Implement barrier_nospec using this instruction.
mpe: The semantics of the instruction are believed to be that it
prevents execution of subsequent instructions until preceding branches
have been fully resolved and are no longer executing speculatively.
There is no further documentation available at this time.
Signed-off-by: Michal Suchanek <msuchanek@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This now has new code in it written by Nick and I, and switch to a
SPDX tag.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
This allows eg. the RCU stall detector, or the soft/hardlockup
detectors to trigger a backtrace on all CPUs.
We implement this by sending a "safe" NMI, which will actually only
send an IPI. Unfortunately the generic code prints "NMI", so that's a
little confusing but we can probably live with it.
If one of the CPUs doesn't respond to the IPI, we then print some info
from it's paca and do a backtrace based on its saved_r1.
Example output:
INFO: rcu_sched detected stalls on CPUs/tasks:
2-...0: (0 ticks this GP) idle=1be/1/4611686018427387904 softirq=1055/1055 fqs=25735
(detected by 4, t=58847 jiffies, g=58, c=57, q=1258)
Sending NMI from CPU 4 to CPUs 2:
CPU 2 didn't respond to backtrace IPI, inspecting paca.
irq_soft_mask: 0x01 in_mce: 0 in_nmi: 0 current: 3623 (bash)
Back trace of paca->saved_r1 (0xc0000000e1c83ba0) (possibly stale):
Call Trace:
[c0000000e1c83ba0] [0000000000000014] 0x14 (unreliable)
[c0000000e1c83bc0] [c000000000765798] lkdtm_do_action+0x48/0x80
[c0000000e1c83bf0] [c000000000765a40] direct_entry+0x110/0x1b0
[c0000000e1c83c90] [c00000000058e650] full_proxy_write+0x90/0xe0
[c0000000e1c83ce0] [c0000000003aae3c] __vfs_write+0x6c/0x1f0
[c0000000e1c83d80] [c0000000003ab214] vfs_write+0xd4/0x240
[c0000000e1c83dd0] [c0000000003ab5cc] ksys_write+0x6c/0x110
[c0000000e1c83e30] [c00000000000b860] system_call+0x58/0x6c
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Currently the options we have for sending NMIs are not necessarily
safe, that is they can potentially interrupt a CPU in a
non-recoverable region of code, meaning the kernel must then panic().
But we'd like to use smp_send_nmi_ipi() to do cross-CPU calls in
situations where we don't want to risk a panic(), because it doesn't
have the requirement that interrupts must be enabled like
smp_call_function().
So add an API for the caller to indicate that it wants to use the NMI
infrastructure, but doesn't want to do anything "unsafe".
Currently that is implemented by not actually calling cause_nmi_ipi(),
instead falling back to an IPI. In future we can pass the safe
parameter down to cause_nmi_ipi() and the individual backends can
potentially take it into account before deciding what to do.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
A CPU that gets stuck with interrupts hard disable can be difficult to
debug, as on some platforms we have no way to interrupt the CPU to
find out what it's doing.
A stop-gap is to have the CPU save it's stack pointer (r1) in its paca
when it hard disables interrupts. That way if we can't interrupt it,
we can at least trace the stack based on where it last disabled
interrupts.
In some cases that will be total junk, but the stack trace code should
handle that. In the simple case of a CPU that disable interrupts and
then gets stuck in a loop, the stack trace should be informative.
We could clear the saved stack pointer when we enable interrupts, but
that loses information which could be useful if we have nothing else
to go on.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
set_fs() sets the addr_limit, which is used in access_ok() to
determine if an address is a user or kernel address.
Some code paths use set_fs() to temporarily elevate the addr_limit so
that kernel code can read/write kernel memory as if it were user
memory. That is fine as long as the code can't ever return to
userspace with the addr_limit still elevated.
If that did happen, then userspace can read/write kernel memory as if
it were user memory, eg. just with write(2). In case it's not clear,
that is very bad. It has also happened in the past due to bugs.
Commit 5ea0727b16 ("x86/syscalls: Check address limit on user-mode
return") added a mechanism to check the addr_limit value before
returning to userspace. Any call to set_fs() sets a thread flag,
TIF_FSCHECK, and if we see that on the return to userspace we go out
of line to check that the addr_limit value is not elevated.
For further info see the above commit, as well as:
https://lwn.net/Articles/722267/https://bugs.chromium.org/p/project-zero/issues/detail?id=990
Verified to work on 64-bit Book3S using a POC that objdumps the system
call handler, and a modified lkdtm_CORRUPT_USER_DS() that doesn't kill
the caller.
Before:
$ sudo ./test-tif-fscheck
...
0000000000000000 <.data>:
0: e1 f7 8a 79 rldicl. r10,r12,30,63
4: 80 03 82 40 bne 0x384
8: 00 40 8a 71 andi. r10,r12,16384
c: 78 0b 2a 7c mr r10,r1
10: 10 fd 21 38 addi r1,r1,-752
14: 08 00 c2 41 beq- 0x1c
18: 58 09 2d e8 ld r1,2392(r13)
1c: 00 00 41 f9 std r10,0(r1)
20: 70 01 61 f9 std r11,368(r1)
24: 78 01 81 f9 std r12,376(r1)
28: 70 00 01 f8 std r0,112(r1)
2c: 78 00 41 f9 std r10,120(r1)
30: 20 00 82 41 beq 0x50
34: a6 42 4c 7d mftb r10
After:
$ sudo ./test-tif-fscheck
Killed
And in dmesg:
Invalid address limit on user-mode return
WARNING: CPU: 1 PID: 3689 at ../include/linux/syscalls.h:260 do_notify_resume+0x140/0x170
...
NIP [c00000000001ee50] do_notify_resume+0x140/0x170
LR [c00000000001ee4c] do_notify_resume+0x13c/0x170
Call Trace:
do_notify_resume+0x13c/0x170 (unreliable)
ret_from_except_lite+0x70/0x74
Performance overhead is essentially zero in the usual case, because
the bit is checked as part of the existing _TIF_USER_WORK_MASK check.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
It's called 'fs' for historical reasons, it's named after the x86 'FS'
register. But we don't have to use that name for the member of
thread_struct, and in fact arch/x86 doesn't even call it 'fs' anymore.
So rename it to 'addr_limit', which better reflects what it's used
for, and is also the name used on other arches.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
In PPC_PTRACE_GETHWDBGINFO and PPC_PTRACE_SETHWDEBUG we do an
access_ok() check and then __copy_{from,to}_user().
Instead we should just use copy_{from,to}_user() which does all that
for us and is less error prone.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Reviewed-by: Samuel Mendoza-Jonas <sam@mendozajonas.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The EEH report functions now share a fair bit of code around the start
and end of each function.
So factor out as much as possible, and move the traversal into a
custom function. This also allows accurate debug to be generated more
easily.
Signed-off-by: Sam Bobroff <sbobroff@linux.ibm.com>
[mpe: Format with clang-format]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
If a device without a driver is recovered via EEH, the flag
EEH_DEV_NO_HANDLER is incorrectly left set on the device after
recovery, because the test in eeh_report_resume() for the existence of
a bound driver is done before the flag is cleared. If a driver is
later bound, and EEH experienced again, some of the drivers EEH
handers are not called.
To correct this, clear the flag unconditionally after EEH processing
is complete.
Signed-off-by: Sam Bobroff <sbobroff@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
To ease future refactoring, extract calls to eeh_enable_irq() and
eeh_disable_irq() from the various report functions. This makes
the report functions initial sequences more similar, as well as making
the IRQ changes visible when reading eeh_handle_normal_event().
Signed-off-by: Sam Bobroff <sbobroff@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
To ease future refactoring, extract setting of the channel state
from the report functions out into their own functions. This increases
the amount of code that is identical across all of the report
functions.
Signed-off-by: Sam Bobroff <sbobroff@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The same test is done in every EEH report function, so factor it out.
Since eeh_dev_removed() needs to be moved higher up in the file,
simplify it a little while we're at it.
Signed-off-by: Sam Bobroff <sbobroff@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Add a for_each-style macro for iterating through PEs without the
boilerplate required by a traversal function. eeh_pe_next() is now
exported, as it is now used directly in place.
Signed-off-by: Sam Bobroff <sbobroff@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
As EEH event handling progresses, a cumulative result of type
pci_ers_result is built up by (some of) the eeh_report_*() functions
using either:
if (rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
if (*res == PCI_ERS_RESULT_NONE) *res = rc;
or:
if ((*res == PCI_ERS_RESULT_NONE) ||
(*res == PCI_ERS_RESULT_RECOVERED)) *res = rc;
if (*res == PCI_ERS_RESULT_DISCONNECT &&
rc == PCI_ERS_RESULT_NEED_RESET) *res = rc;
(Where *res is the accumulator.)
However, the intent is not immediately clear and the result in some
situations is order dependent.
Address this by assigning a priority to each result value, and always
merging to the highest priority. This renders the intent clear, and
provides a stable value for all orderings.
Signed-off-by: Sam Bobroff <sbobroff@linux.ibm.com>
[mpe: Minor formatting (clang-format)]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
To aid debugging, add a message to show when EEH processing for a PE
will be done at the device's parent, rather than directly at the
device.
Signed-off-by: Sam Bobroff <sbobroff@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The traversal functions eeh_pe_traverse() and eeh_pe_dev_traverse()
both provide their first argument as void * but every single user casts
it to the expected type.
Change the type of the first parameter from void * to the appropriate
type, and clean up all uses.
Signed-off-by: Sam Bobroff <sbobroff@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Correct two cases where eeh_pcid_get() is used to reference the driver's
module but the reference is dropped before the driver pointer is used.
In eeh_rmv_device() also refactor a little so that only two calls to
eeh_pcid_put() are needed, rather than three and the reference isn't
taken at all if it wasn't needed.
Signed-off-by: Sam Bobroff <sbobroff@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Add a single log line at the end of successful EEH recovery, so that
it's clear that event processing has finished.
Signed-off-by: Sam Bobroff <sbobroff@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Since thread-imc internally use the core-imc hardware infrastructure
and is depended on it, having thread-imc in the kernel in the
absence of core-imc is trivial. Patch disables thread-imc, if
core-imc is not registered.
Signed-off-by: Anju T Sudhakar <anju@linux.vnet.ibm.com>
Reviewed-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Replace the direct return statement in imc_mem_init() with goto, to adhere
to the kernel coding style.
Signed-off-by: Anju T Sudhakar <anju@linux.vnet.ibm.com>
Reviewed-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When any of the IMC (In-Memory Collection counter) devices fail
to initialize, imc_common_mem_free() frees set of memory. In doing so,
pmu_ptr pointer is also freed. But pmu_ptr pointer is used in subsequent
function (imc_common_cpuhp_mem_free()) which is wrong. Patch here reorders
the code to avoid such access.
Also free the memory which is dynamically allocated during imc
initialization, wherever required.
Signed-off-by: Anju T Sudhakar <anju@linux.vnet.ibm.com>
Reviewed-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The device node obtained with of_find_compatible_node() should be
released by calling of_node_put(). But it was not released when
of_get_property() failed.
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
[mpe: Invert the sense of the if so we only need one return path]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Structure platform_driver does not need to set the owner field, as this
will be populated by the driver core.
Generated by scripts/coccinelle/api/platform_no_drv_owner.cocci.
Signed-off-by: Fabio Estevam <fabio.estevam@nxp.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The GETFIELD and SETFIELD macros in xive-regs.h aren't used except for
a single instance of GETFIELD, so replace that and remove them.
These macros are also defined in vas.h, so either those should be
eventually replaced or the macros moved into bitops.h.
Signed-off-by: Russell Currey <ruscur@russell.cc>
[mpe: Rewrite the assignment to 'he' to avoid ffs() etc.]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
time_init() will set up tb_ticks_per_usec based on reality.
time_init() is called *after* udbg_init_opal_common() during boot.
from arch/powerpc/kernel/time.c:
unsigned long tb_ticks_per_usec = 100; /* sane default */
Currently, all powernv systems have a timebase frequency of 512mhz
(512000000/1000000 == 0x200) - although there's nothing written
down anywhere that I can find saying that we couldn't make that
different based on the requirements in the ISA.
So, we've been (accidentally) thwacking the (currently) correct
(for powernv at least) value for tb_ticks_per_usec earlier than
we otherwise would have.
The "sane default" seems to be adequate for our purposes between
udbg_init_opal_common() and time_init() being called, and if it isn't,
then we should probably be setting it somewhere that isn't hvc_opal.c!
Signed-off-by: Stewart Smith <stewart@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
to_tm() is now completely unused, the only reference being in the
_dump_time() helper that is also unused. This removes both, leaving
the rest of the powerpc RTC code y2038 safe to as far as the hardware
supports.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
update_persistent_clock() is deprecated because it suffers from overflow
in 2038 on 32-bit architectures. This changes powerpc to use the
update_persistent_clock64() replacement, and to pass down 64-bit
timestamps consistently.
This is now simpler, as we no longer have to worry about the offset
numbers in tm_year and tm_mon that are different between the Linux
conventions and RTAS.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Looking through the remaining users of the deprecated mktime()
function, I found the powerpc rtc handlers, which use it in
place of rtc_tm_to_time64().
To clean this up, I'm changing over the read_persistent_clock()
function to the read_persistent_clock64() variant, and change
all the platform specific handlers along with it.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The to_tm() helper function operates on a signed integer for the time,
so it will suffer from overflow in 2038, even on 64-bit kernels.
Rather than fix that function, this replaces its use in the rtas
procfs implementation with the standard rtc_time64_to_tm() helper
that is very similar but is not affected by the overflow.
In order to actually support long times, the parser function gets
changed to 64-bit user input and output as well. Note that the tm_mon
and tm_year representation is slightly different, so we have to manually
add an offset here.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
In order to use the rtc_tm_to_time64() and rtc_time64_to_tm()
helper functions in later patches, we have to ensure that
CONFIG_RTC_LIB is always built-in.
Note that this symbol only controls a couple of helper functions,
not the actual RTC subsystem, which remains optional and is
enabled with CONFIG_RTC_CLASS.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Needed on Amiga X1000 with SB600.
Reported-by: Christian Zigotzky <chzigotzky@xenosoft.de>
Signed-off-by: Olof Johansson <olof@lixom.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When inserting SLB entries for EA above 512TB, we need to hard disable irq.
This will make sure we don't take a PMU interrupt that can possibly touch
user space address via a stack dump. To prevent this, we need to hard disable
the interrupt.
Also add a comment explaining why we don't need context synchronizing isync
with slbmte.
Fixes: f384796c4 ("powerpc/mm: Add support for handling > 512TB address in SLB miss")
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
With split pmd page table lock enabled, we don't use mm->page_table_lock when
updating pmd entries. This patch update hugetlb path to use the right lock
when inserting huge page directory entries into page table.
ex: if we are using hugepd and inserting hugepd entry at the pmd level, we
use pmd_lockptr, which based on config can be split pmd lock.
For update huge page directory entries itself we use mm->page_table_lock. We
do have a helper huge_pte_lockptr() for that.
Fixes: 675d99529 ("powerpc/book3s64: Enable split pmd ptlock")
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Currently we do not have an isync, or any other context synchronizing
instruction prior to the slbie/slbmte in _switch() that updates the
SLB entry for the kernel stack.
However that is not correct as outlined in the ISA.
From Power ISA Version 3.0B, Book III, Chapter 11, page 1133:
"Changing the contents of ... the contents of SLB entries ... can
have the side effect of altering the context in which data
addresses and instruction addresses are interpreted, and in which
instructions are executed and data accesses are performed.
...
These side effects need not occur in program order, and therefore
may require explicit synchronization by software.
...
The synchronizing instruction before the context-altering
instruction ensures that all instructions up to and including that
synchronizing instruction are fetched and executed in the context
that existed before the alteration."
And page 1136:
"For data accesses, the context synchronizing instruction before the
slbie, slbieg, slbia, slbmte, tlbie, or tlbiel instruction ensures
that all preceding instructions that access data storage have
completed to a point at which they have reported all exceptions
they will cause."
We're not aware of any bugs caused by this, but it should be fixed
regardless.
Add the missing isync when updating kernel stack SLB entry.
Cc: stable@vger.kernel.org
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
[mpe: Flesh out change log with more ISA text & explanation]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The stores to update the SLB shadow area must be made as they appear
in the C code, so that the hypervisor does not see an entry with
mismatched vsid and esid. Use WRITE_ONCE for this.
GCC has been observed to elide the first store to esid in the update,
which means that if the hypervisor interrupts the guest after storing
to vsid, it could see an entry with old esid and new vsid, which may
possibly result in memory corruption.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When a single-threaded process has a non-local mm_cpumask, try to use
that point to flush the TLBs out of other CPUs in the cpumask.
An IPI is used for clearing remote CPUs for a few reasons:
- An IPI can end lazy TLB use of the mm, which is required to prevent
TLB entries being created on the remote CPU. The alternative is to
drop lazy TLB switching completely, which costs 7.5% in a context
switch ping-pong test betwee a process and kernel idle thread.
- An IPI can have remote CPUs flush the entire PID, but the local CPU
can flush a specific VA. tlbie would require over-flushing of the
local CPU (where the process is running).
- A single threaded process that is migrated to a different CPU is
likely to have a relatively small mm_cpumask, so IPI is reasonable.
No other thread can concurrently switch to this mm, because it must
have been given a reference to mm_users by the current thread before it
can use_mm. mm_users can be asynchronously incremented (by
mm_activate or mmget_not_zero), but those users must use remote mm
access and can't use_mm or access user address space. Existing code
makes the this assumption already, for example sparc64 has reset
mm_cpumask using this condition since the start of history, see
arch/sparc/kernel/smp_64.c.
This reduces tlbies for a kernel compile workload from 0.90M to 0.12M,
tlbiels are increased significantly due to the PID flushing for the
cleaning up remote CPUs, and increased local flushes (PID flushes take
128 tlbiels vs 1 tlbie).
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Implementing pte_update with pte_xchg (which uses cmpxchg) is
inefficient. A single larx/stcx. works fine, no need for the less
efficient cmpxchg sequence.
Then remove the memory barriers from the operation. There is a
requirement for TLB flushing to load mm_cpumask after the store
that reduces pte permissions, which is moved into the TLB flush
code.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The ISA suggests ptesync after setting a pte, to prevent a table walk
initiated by a subsequent access from missing that store and causing a
spurious fault. This is an architectual allowance that allows an
implementation's page table walker to be incoherent with the store
queue.
However there is no correctness problem in taking a spurious fault in
userspace -- the kernel copes with these at any time, so the updated
pte will be found eventually. Spurious kernel faults on vmap memory
must be avoided, so a ptesync is put into flush_cache_vmap.
On POWER9 so far I have not found a measurable window where this can
result in more minor faults, so as an optimisation, remove the costly
ptesync from pte updates. If an implementation benefits from ptesync,
it would be better to add it back in update_mmu_cache, so it's not
done for things like fork(2).
fork --fork --exec benchmark improved 5.2% (12400->13100).
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Prefetch the faulting address in update_mmu_cache to give the page
table walker perhaps 100 cycles head start as locks are dropped and
the interrupt completed.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This matches other architectures, when we know there will be no
further accesses to the address (e.g., for teardown), page table
entries can be cleared non-atomically.
The comments about NMMU are bogus: all MMU notifiers (including NMMU)
are released at this point, with their TLBs flushed. An NMMU access at
this point would be a bug.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
