aboutsummaryrefslogtreecommitdiffstats
path: root/Documentation/powerpc/transactional_memory.txt
blob: 9791e98ab49cf28826efb1c4551916280ac31b00 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
Transactional Memory support
============================

POWER kernel support for this feature is currently limited to supporting
its use by user programs.  It is not currently used by the kernel itself.

This file aims to sum up how it is supported by Linux and what behaviour you
can expect from your user programs.


Basic overview
==============

Hardware Transactional Memory is supported on POWER8 processors, and is a
feature that enables a different form of atomic memory access.  Several new
instructions are presented to delimit transactions; transactions are
guaranteed to either complete atomically or roll back and undo any partial
changes.

A simple transaction looks like this:

begin_move_money:
  tbegin
  beq   abort_handler

  ld    r4, SAVINGS_ACCT(r3)
  ld    r5, CURRENT_ACCT(r3)
  subi  r5, r5, 1
  addi  r4, r4, 1
  std   r4, SAVINGS_ACCT(r3)
  std   r5, CURRENT_ACCT(r3)

  tend

  b     continue

abort_handler:
  ... test for odd failures ...

  /* Retry the transaction if it failed because it conflicted with
   * someone else: */
  b     begin_move_money


The 'tbegin' instruction denotes the start point, and 'tend' the end point.
Between these points the processor is in 'Transactional' state; any memory
references will complete in one go if there are no conflicts with other
transactional or non-transactional accesses within the system.  In this
example, the transaction completes as though it were normal straight-line code
IF no other processor has touched SAVINGS_ACCT(r3) or CURRENT_ACCT(r3); an
atomic move of money from the current account to the savings account has been
performed.  Even though the normal ld/std instructions are used (note no
lwarx/stwcx), either *both* SAVINGS_ACCT(r3) and CURRENT_ACCT(r3) will be
updated, or neither will be updated.

If, in the meantime, there is a conflict with the locations accessed by the
transaction, the transaction will be aborted by the CPU.  Register and memory
state will roll back to that at the 'tbegin', and control will continue from
'tbegin+4'.  The branch to abort_handler will be taken this second time; the
abort handler can check the cause of the failure, and retry.

Checkpointed registers include all GPRs, FPRs, VRs/VSRs, LR, CCR/CR, CTR, FPCSR
and a few other status/flag regs; see the ISA for details.

Causes of transaction aborts
============================

- Conflicts with cache lines used by other processors
- Signals
- Context switches
- See the ISA for full documentation of everything that will abort transactions.


Syscalls
========

Performing syscalls from within transaction is not recommended, and can lead
to unpredictable results.

Syscalls do not by design abort transactions, but beware: The kernel code will
not be running in transactional state.  The effect of syscalls will always
remain visible, but depending on the call they may abort your transaction as a
side-effect, read soon-to-be-aborted transactional data that should not remain
invisible, etc.  If you constantly retry a transaction that constantly aborts
itself by calling a syscall, you'll have a livelock & make no progress.

Simple syscalls (e.g. sigprocmask()) "could" be OK.  Even things like write()
from, say, printf() should be OK as long as the kernel does not access any
memory that was accessed transactionally.

Consider any syscalls that happen to work as debug-only -- not recommended for
production use.  Best to queue them up till after the transaction is over.


Signals
=======

Delivery of signals (both sync and async) during transactions provides a second
thread state (ucontext/mcontext) to represent the second transactional register
state.  Signal delivery 'treclaim's to capture both register states, so signals
abort transactions.  The usual ucontext_t passed to the signal handler
represents the checkpointed/original register state; the signal appears to have
arisen at 'tbegin+4'.

If the sighandler ucontext has uc_link set, a second ucontext has been
delivered.  For future compatibility the MSR.TS field should be checked to
determine the transactional state -- if so, the second ucontext in uc->uc_link
represents the active transactional registers at the point of the signal.

For 64-bit processes, uc->uc_mcontext.regs->msr is a full 64-bit MSR and its TS
field shows the transactional mode.

For 32-bit processes, the mcontext's MSR register is only 32 bits; the top 32
bits are stored in the MSR of the second ucontext, i.e. in
uc->uc_link->uc_mcontext.regs->msr.  The top word contains the transactional
state TS.

However, basic signal handlers don't need to be aware of transactions
and simply returning from the handler will deal with things correctly:

Transaction-aware signal handlers can read the transactional register state
from the second ucontext.  This will be necessary for crash handlers to
determine, for example, the address of the instruction causing the SIGSEGV.

Example signal handler:

    void crash_handler(int sig, siginfo_t *si, void *uc)
    {
      ucontext_t *ucp = uc;
      ucontext_t *transactional_ucp = ucp->uc_link;

      if (ucp_link) {
        u64 msr = ucp->uc_mcontext.regs->msr;
        /* May have transactional ucontext! */
#ifndef __powerpc64__
        msr |= ((u64)transactional_ucp->uc_mcontext.regs->msr) << 32;
#endif
        if (MSR_TM_ACTIVE(msr)) {
           /* Yes, we crashed during a transaction.  Oops. */
   fprintf(stderr, "Transaction to be restarted at 0x%llx, but "
                           "crashy instruction was at 0x%llx\n",
                           ucp->uc_mcontext.regs->nip,
                           transactional_ucp->uc_mcontext.regs->nip);
        }
      }

      fix_the_problem(ucp->dar);
    }

When in an active transaction that takes a signal, we need to be careful with
the stack.  It's possible that the stack has moved back up after the tbegin.
The obvious case here is when the tbegin is called inside a function that
returns before a tend.  In this case, the stack is part of the checkpointed
transactional memory state.  If we write over this non transactionally or in
suspend, we are in trouble because if we get a tm abort, the program counter and
stack pointer will be back at the tbegin but our in memory stack won't be valid
anymore.

To avoid this, when taking a signal in an active transaction, we need to use
the stack pointer from the checkpointed state, rather than the speculated
state.  This ensures that the signal context (written tm suspended) will be
written below the stack required for the rollback.  The transaction is aborted
because of the treclaim, so any memory written between the tbegin and the
signal will be rolled back anyway.

For signals taken in non-TM or suspended mode, we use the
normal/non-checkpointed stack pointer.


Failure cause codes used by kernel
==================================

These are defined in <asm/reg.h>, and distinguish different reasons why the
kernel aborted a transaction:

 TM_CAUSE_RESCHED       Thread was rescheduled.
 TM_CAUSE_TLBI          Software TLB invalide.
 TM_CAUSE_FAC_UNAV      FP/VEC/VSX unavailable trap.
 TM_CAUSE_SYSCALL       Currently unused; future syscalls that must abort
                        transactions for consistency will use this.
 TM_CAUSE_SIGNAL        Signal delivered.
 TM_CAUSE_MISC          Currently unused.
 TM_CAUSE_ALIGNMENT     Alignment fault.
 TM_CAUSE_EMULATE       Emulation that touched memory.

These can be checked by the user program's abort handler as TEXASR[0:7].  If
bit 7 is set, it indicates that the error is consider persistent.  For example
a TM_CAUSE_ALIGNMENT will be persistent while a TM_CAUSE_RESCHED will not.q

GDB
===

GDB and ptrace are not currently TM-aware.  If one stops during a transaction,
it looks like the transaction has just started (the checkpointed state is
presented).  The transaction cannot then be continued and will take the failure
handler route.  Furthermore, the transactional 2nd register state will be
inaccessible.  GDB can currently be used on programs using TM, but not sensibly
in parts within transactions.

Privacy Policy