At the start of __tm_recheckpoint() we save the kernel stack pointer
(r1) in SPRG SCRATCH0 (SPRG2) so that we can restore it after the
trecheckpoint.
Unfortunately, the same SPRG is used in the SLB miss handler. If an
SLB miss is taken between the save and restore of r1 to the SPRG, the
SPRG is changed and hence r1 is also corrupted. We can end up with
the following crash when we start using r1 again after the restore
from the SPRG:
Oops: Bad kernel stack pointer, sig: 6 [#1]
SMP NR_CPUS=2048 NUMA pSeries
CPU: 658 PID: 143777 Comm: htm_demo Tainted: G EL X 4.4.13-0-default #1
task: c0000b56993a7810 ti: c00000000cfec000 task.ti: c0000b56993bc000
NIP: c00000000004f188 LR: 00000000100040b8 CTR: 0000000010002570
REGS: c00000000cfefd40 TRAP: 0300 Tainted: G EL X (4.4.13-0-default)
MSR: 8000000300001033 <SF,ME,IR,DR,RI,LE> CR: 02000424 XER: 20000000
CFAR: c000000000008468 DAR: 00003ffd84e66880 DSISR: 40000000 SOFTE: 0
PACATMSCRATCH: 00003ffbc865e680
GPR00: fffffffcfabc4268 00003ffd84e667a0 00000000100d8c38 000000030544bb80
GPR04: 0000000000000002 00000000100cf200 0000000000000449 00000000100cf100
GPR08: 000000000000c350 0000000000002569 0000000000002569 00000000100d6c30
GPR12: 00000000100d6c28 c00000000e6a6b00 00003ffd84660000 0000000000000000
GPR16: 0000000000000003 0000000000000449 0000000010002570 0000010009684f20
GPR20: 0000000000800000 00003ffd84e5f110 00003ffd84e5f7a0 00000000100d0f40
GPR24: 0000000000000000 0000000000000000 0000000000000000 00003ffff0673f50
GPR28: 00003ffd84e5e960 00000000003d0f00 00003ffd84e667a0 00003ffd84e5e680
NIP [c00000000004f188] restore_gprs+0x110/0x17c
LR [00000000100040b8] 0x100040b8
Call Trace:
Instruction dump:
f8a1fff0 e8e700a8 38a00000 7ca10164 e8a1fff8 e821fff0 7c0007dd 7c421378
7db142a6 7c3242a6 38800002 7c810164 <e9c100e0> e9e100e8 ea0100f0 ea2100f8
We hit this on large memory machines (> 2TB) but it can also be hit on
smaller machines when 1TB segments are disabled.
To hit this, you also need to be virtualised to ensure SLBs are
periodically removed by the hypervisor.
This patches moves the saving of r1 to the SPRG to the region where we
are guaranteed not to take any further SLB misses.
Fixes: 98ae22e15b ("powerpc: Add helper functions for transactional memory context switching")
Cc: stable@vger.kernel.org # v3.9+
Signed-off-by: Michael Neuling <mikey@neuling.org>
Acked-by: Cyril Bur <cyrilbur@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Currently we have 2 segments that are bolted for the kernel linear
mapping (ie 0xc000... addresses). This is 0 to 1TB and also the kernel
stacks. Anything accessed outside of these regions may need to be
faulted in. (In practice machines with TM always have 1T segments)
If a machine has < 2TB of memory we never fault on the kernel linear
mapping as these two segments cover all physical memory. If a machine
has > 2TB of memory, there may be structures outside of these two
segments that need to be faulted in. This faulting can occur when
running as a guest as the hypervisor may remove any SLB that's not
bolted.
When we treclaim and trecheckpoint we have a window where we need to
run with the userspace GPRs. This means that we no longer have a valid
stack pointer in r1. For this window we therefore clear MSR RI to
indicate that any exceptions taken at this point won't be able to be
handled. This means that we can't take segment misses in this RI=0
window.
In this RI=0 region, we currently access the thread_struct for the
process being context switched to or from. This thread_struct access
may cause a segment fault since it's not guaranteed to be covered by
the two bolted segment entries described above.
We've seen this with a crash when running as a guest with > 2TB of
memory on PowerVM:
Unrecoverable exception 4100 at c00000000004f138
Oops: Unrecoverable exception, sig: 6 [#1]
SMP NR_CPUS=2048 NUMA pSeries
CPU: 1280 PID: 7755 Comm: kworker/1280:1 Tainted: G X 4.4.13-46-default #1
task: c000189001df4210 ti: c000189001d5c000 task.ti: c000189001d5c000
NIP: c00000000004f138 LR: 0000000010003a24 CTR: 0000000010001b20
REGS: c000189001d5f730 TRAP: 4100 Tainted: G X (4.4.13-46-default)
MSR: 8000000100001031 <SF,ME,IR,DR,LE> CR: 24000048 XER: 00000000
CFAR: c00000000004ed18 SOFTE: 0
GPR00: ffffffffc58d7b60 c000189001d5f9b0 00000000100d7d00 000000003a738288
GPR04: 0000000000002781 0000000000000006 0000000000000000 c0000d1f4d889620
GPR08: 000000000000c350 00000000000008ab 00000000000008ab 00000000100d7af0
GPR12: 00000000100d7ae8 00003ffe787e67a0 0000000000000000 0000000000000211
GPR16: 0000000010001b20 0000000000000000 0000000000800000 00003ffe787df110
GPR20: 0000000000000001 00000000100d1e10 0000000000000000 00003ffe787df050
GPR24: 0000000000000003 0000000000010000 0000000000000000 00003fffe79e2e30
GPR28: 00003fffe79e2e68 00000000003d0f00 00003ffe787e67a0 00003ffe787de680
NIP [c00000000004f138] restore_gprs+0xd0/0x16c
LR [0000000010003a24] 0x10003a24
Call Trace:
[c000189001d5f9b0] [c000189001d5f9f0] 0xc000189001d5f9f0 (unreliable)
[c000189001d5fb90] [c00000000001583c] tm_recheckpoint+0x6c/0xa0
[c000189001d5fbd0] [c000000000015c40] __switch_to+0x2c0/0x350
[c000189001d5fc30] [c0000000007e647c] __schedule+0x32c/0x9c0
[c000189001d5fcb0] [c0000000007e6b58] schedule+0x48/0xc0
[c000189001d5fce0] [c0000000000deabc] worker_thread+0x22c/0x5b0
[c000189001d5fd80] [c0000000000e7000] kthread+0x110/0x130
[c000189001d5fe30] [c000000000009538] ret_from_kernel_thread+0x5c/0xa4
Instruction dump:
7cb103a6 7cc0e3a6 7ca222a6 78a58402 38c00800 7cc62838 08860000 7cc000a6
38a00006 78c60022 7cc62838 0b060000 <e8c701a0> 7ccff120 e8270078 e8a70098
---[ end trace 602126d0a1dedd54 ]---
This fixes this by copying the required data from the thread_struct to
the stack before we clear MSR RI. Then once we clear RI, we only access
the stack, guaranteeing there's no segment miss.
We also tighten the region over which we set RI=0 on the treclaim()
path. This may have a slight performance impact since we're adding an
mtmsr instruction.
Fixes: 090b9284d7 ("powerpc/tm: Clear MSR RI in non-recoverable TM code")
Signed-off-by: Michael Neuling <mikey@neuling.org>
Reviewed-by: Cyril Bur <cyrilbur@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
PACA_DSCR offset macro tracks dscr_default element in the paca
structure. Better change the name of this macro to match that of the
data element it tracks. Makes the code more readable.
Signed-off-by: Anshuman Khandual <khandual@linux.vnet.ibm.com>
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>
Since commit "efcac65 powerpc: Per process DSCR + some fixes (try#4)"
it is no longer possible to set the DSCR on a per-CPU basis.
The old behaviour was to minipulate the DSCR SPR directly but this is no
longer sufficient: the value is quickly overwritten by context switching.
This patch stores the per-CPU DSCR value in a kernel variable rather than
directly in the SPR and it is used whenever a process has not set the DSCR
itself. The sysfs interface (/sys/devices/system/cpu/cpuN/dscr) is unchanged.
Writes to the old global default (/sys/devices/system/cpu/dscr_default)
now set all of the per-CPU values and reads return the last written value.
The new per-CPU default is added to the paca_struct and is used everywhere
outside of sysfs.c instead of the old global default.
Signed-off-by: Sam Bobroff <sam.bobroff@au1.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This series adds support for building the powerpc 64-bit
LE kernel using the new ABI v2. We already supported
running ABI v2 userspace programs but this adds support
for building the kernel itself using the new ABI.
If we do a treclaim and we are not in TM suspend mode, it results in a TM bad
thing (ie. a 0x700 program check). Similarly if we do a trechkpt and we have
an active transaction or TEXASR Failure Summary (FS) is not set, we also take a
TM bad thing.
This should never happen, but if it does (ie. a kernel bug), the cause is
almost impossible to debug as the GPR state is mostly userspace and hence we
don't get a call chain.
This adds some checks in these cases case a BUG_ON() (in asm) in case we ever
hit these cases. It moves the register saving around to preserve r1 till later
also.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
We save r1 to the scratch SPR and restore it from there after the trechkpt so
saving r1 to the paca is not needed.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
We can't take an IRQ when we're about to do a trechkpt as our GPR state is set
to user GPR values.
We've hit this when running some IBM Java stress tests in the lab resulting in
the following dump:
cpu 0x3f: Vector: 700 (Program Check) at [c000000007eb3d40]
pc: c000000000050074: restore_gprs+0xc0/0x148
lr: 00000000b52a8184
sp: ac57d360
msr: 8000000100201030
current = 0xc00000002c500000
paca = 0xc000000007dbfc00 softe: 0 irq_happened: 0x00
pid = 34535, comm = Pooled Thread #
R00 = 00000000b52a8184 R16 = 00000000b3e48fda
R01 = 00000000ac57d360 R17 = 00000000ade79bd8
R02 = 00000000ac586930 R18 = 000000000fac9bcc
R03 = 00000000ade60000 R19 = 00000000ac57f930
R04 = 00000000f6624918 R20 = 00000000ade79be8
R05 = 00000000f663f238 R21 = 00000000ac218a54
R06 = 0000000000000002 R22 = 000000000f956280
R07 = 0000000000000008 R23 = 000000000000007e
R08 = 000000000000000a R24 = 000000000000000c
R09 = 00000000b6e69160 R25 = 00000000b424cf00
R10 = 0000000000000181 R26 = 00000000f66256d4
R11 = 000000000f365ec0 R27 = 00000000b6fdcdd0
R12 = 00000000f66400f0 R28 = 0000000000000001
R13 = 00000000ada71900 R29 = 00000000ade5a300
R14 = 00000000ac2185a8 R30 = 00000000f663f238
R15 = 0000000000000004 R31 = 00000000f6624918
pc = c000000000050074 restore_gprs+0xc0/0x148
cfar= c00000000004fe28 dont_restore_vec+0x1c/0x1a4
lr = 00000000b52a8184
msr = 8000000100201030 cr = 24804888
ctr = 0000000000000000 xer = 0000000000000000 trap = 700
This moves tm_recheckpoint to a C function and moves the tm_restore_sprs into
that function. It then adds IRQ disabling over the trechkpt critical section.
It also sets the TEXASR FS in the signals code to ensure this is never set now
that we explictly write the TM sprs in tm_recheckpoint.
Signed-off-by: Michael Neuling <mikey@neuling.org>
cc: stable@vger.kernel.org
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The condition register (CR) is a 32 bit quantity so we should use
32 bit loads and stores.
Signed-off-by: Anton Blanchard <anton@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>
When we do a treclaim or trecheckpoint we end up running with userspace
PPR and DSCR values. Currently we don't do anything special to avoid
running with user values which could cause a severe performance
degradation.
This patch moves the PPR and DSCR save and restore around treclaim and
trecheckpoint so that we run with user values for a much shorter period.
More care is taken with the PPR as it's impact is greater than the DSCR.
This is similar to user exceptions, where we run HTM_MEDIUM early to
ensure that we don't run with a userspace PPR values in the kernel.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Cc: <stable@vger.kernel.org> # 3.9+
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
We can't take IRQs in tm_reclaim as we might have a bogus r13 and r1.
This turns IRQs hard off in this function.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Cc: <stable@vger.kernel.org> # 3.9+
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Since 2002, the kernel has not saved VRSAVE on exception entry and
restored it on exit; rather, VRSAVE gets context-switched in _switch.
This means that when executing in process context in the kernel, the
userspace VRSAVE value is live in the VRSAVE register.
However, the signal code assumes that current->thread.vrsave holds
the current VRSAVE value, which is incorrect. Therefore, this
commit changes it to use the actual VRSAVE register instead. (It
still uses current->thread.vrsave as a temporary location to store
it in, as __get_user and __put_user can only transfer to/from a
variable, not an SPR.)
This also modifies the transactional memory code to save and restore
VRSAVE regardless of whether VMX is enabled in the MSR. This is
because accesses to VRSAVE are not controlled by the MSR.VEC bit,
but can happen at any time.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
If a transaction is rolled back, the Target Address Register (TAR), Processor
Priority Register (PPR) and Data Stream Control Register (DSCR) should be
restored to the checkpointed values before the transaction began. Any changes
to these SPRs inside the transaction should not be visible in the abort
handler.
Currently Linux doesn't save or restore the checkpointed TAR, PPR or DSCR. If
we preempt a processes inside a transaction which has modified any of these, on
process restore, that same transaction may be aborted we but we won't see the
checkpointed versions of these SPRs.
This adds checkpointed versions of these SPRs to the thread_struct and adds the
save/restore of these three SPRs to the treclaim/trechkpt code.
Without this if any of these SPRs are modified during a transaction, users may
incorrectly see a speculated SPR value even if the transaction is aborted.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Cc: <stable@vger.kernel.org> [v3.10]
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
When we treclaim and trecheckpoint there's an unavoidable period when r1
will not be a valid kernel stack pointer.
This patch clears the MSR recoverable interrupt (RI) bit over these
regions to indicate we have an invalid kernel stack pointer.
For treclaim, the region over which we clear MSR RI is larger than
required to avoid the need for an extra costly mtmsrd.
Thanks to Paulus for suggesting this change.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
We can't compile a kernel with CONFIG_ALTIVEC=n when
CONFIG_PPC_TRANSACTIONAL_MEM=y. We currently get:
arch/powerpc/kernel/tm.S:320: Error: unsupported relocation against THREAD_VSCR
arch/powerpc/kernel/tm.S:323: Error: unsupported relocation against THREAD_VR0
arch/powerpc/kernel/tm.S:323: Error: unsupported relocation against THREAD_VR0
etc.
The below fixes this with a sprinkling of #ifdefs.
This was found by mpe with kisskb:
http://kisskb.ellerman.id.au/kisskb/buildresult/8539442/
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Here we add the helper functions to be used when context switching. These
allow us to fully reclaim and recheckpoint a transaction.
We introduce a new paca field called tm_scratch to help us store away register
values when doing the low level tm reclaim register save.
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>
Michael Neuling