1 /*
   2  * Copyright (c) 2016, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2016, 2018 SAP SE. All rights reserved.
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This code is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 only, as
   8  * published by the Free Software Foundation.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  *
  24  */
  25 
  26 #ifndef OS_CPU_LINUX_S390_VM_ATOMIC_LINUX_S390_INLINE_HPP
  27 #define OS_CPU_LINUX_S390_VM_ATOMIC_LINUX_S390_INLINE_HPP
  28 
  29 #include "runtime/atomic.hpp"
  30 #include "runtime/os.hpp"
  31 #include "vm_version_s390.hpp"
  32 
  33 // Note that the compare-and-swap instructions on System z perform
  34 // a serialization function before the storage operand is fetched
  35 // and again after the operation is completed.
  36 //
  37 // Used constraint modifiers:
  38 // = write-only access: Value on entry to inline-assembler code irrelevant.
  39 // + read/write access: Value on entry is used; on exit value is changed.
  40 //   read-only  access: Value on entry is used and never changed.
  41 // & early-clobber access: Might be modified before all read-only operands
  42 //                         have been used.
  43 // a address register operand (not GR0).
  44 // d general register operand (including GR0)
  45 // Q memory operand w/o index register.
  46 // 0..9 operand reference (by operand position).
  47 //      Used for operands that fill multiple roles. One example would be a
  48 //      write-only operand receiving its initial value from a read-only operand.
  49 //      Refer to cmpxchg(..) operand #0 and variable cmp_val for a real-life example.
  50 //
  51 
  52 // On System z, all store operations are atomic if the address where the data is stored into
  53 // is an integer multiple of the data length. Furthermore, all stores are ordered:
  54 // a store which occurs conceptually before another store becomes visible to other CPUs
  55 // before the other store becomes visible.
  56 
  57 //------------
  58 // Atomic::add
  59 //------------
  60 // These methods force the value in memory to be augmented by the passed increment.
  61 // Both, memory value and increment, are treated as 32bit signed binary integers.
  62 // No overflow exceptions are recognized, and the condition code does not hold
  63 // information about the value in memory.
  64 //
  65 // The value in memory is updated by using a compare-and-swap instruction. The
  66 // instruction is retried as often as required.
  67 //
  68 // The return value of the method is the value that was successfully stored. At the
  69 // time the caller receives back control, the value in memory may have changed already.
  70 
  71 // New atomic operations only include specific-operand-serialization, not full
  72 // memory barriers. We can use the Fast-BCR-Serialization Facility for them.
  73 inline void z196_fast_sync() {
  74   __asm__ __volatile__ ("bcr 14, 0" : : : "memory");
  75 }
  76 
  77 template<size_t byte_size>
  78 struct Atomic::PlatformAdd
  79   : Atomic::AddAndFetch<Atomic::PlatformAdd<byte_size> >
  80 {
  81   template<typename I, typename D>
  82   D add_and_fetch(I add_value, D volatile* dest, atomic_memory_order order) const;
  83 };
  84 
  85 template<>
  86 template<typename I, typename D>
  87 inline D Atomic::PlatformAdd<4>::add_and_fetch(I inc, D volatile* dest,
  88                                                atomic_memory_order order) const {
  89   STATIC_ASSERT(4 == sizeof(I));
  90   STATIC_ASSERT(4 == sizeof(D));
  91 
  92   D old, upd;
  93 
  94   if (VM_Version::has_LoadAndALUAtomicV1()) {
  95     if (order == memory_order_conservative) { z196_fast_sync(); }
  96     __asm__ __volatile__ (
  97       "   LGFR     0,%[inc]                \n\t" // save increment
  98       "   LA       3,%[mem]                \n\t" // force data address into ARG2
  99 //    "   LAA      %[upd],%[inc],%[mem]    \n\t" // increment and get old value
 100 //    "   LAA      2,0,0(3)                \n\t" // actually coded instruction
 101       "   .byte    0xeb                    \n\t" // LAA main opcode
 102       "   .byte    0x20                    \n\t" // R1,R3
 103       "   .byte    0x30                    \n\t" // R2,disp1
 104       "   .byte    0x00                    \n\t" // disp2,disp3
 105       "   .byte    0x00                    \n\t" // disp4,disp5
 106       "   .byte    0xf8                    \n\t" // LAA minor opcode
 107       "   AR       2,0                     \n\t" // calc new value in register
 108       "   LR       %[upd],2                \n\t" // move to result register
 109       //---<  outputs  >---
 110       : [upd]  "=&d" (upd)    // write-only, updated counter value
 111       , [mem]  "+Q"  (*dest)  // read/write, memory to be updated atomically
 112       //---<  inputs  >---
 113       : [inc]  "a"   (inc)    // read-only.
 114       //---<  clobbered  >---
 115       : "cc", "r0", "r2", "r3", "memory"
 116     );
 117     if (order == memory_order_conservative) { z196_fast_sync(); }
 118   } else {
 119     __asm__ __volatile__ (
 120       "   LLGF     %[old],%[mem]           \n\t" // get old value
 121       "0: LA       %[upd],0(%[inc],%[old]) \n\t" // calc result
 122       "   CS       %[old],%[upd],%[mem]    \n\t" // try to xchg res with mem
 123       "   JNE      0b                      \n\t" // no success? -> retry
 124       //---<  outputs  >---
 125       : [old] "=&a" (old)    // write-only, old counter value
 126       , [upd] "=&d" (upd)    // write-only, updated counter value
 127       , [mem] "+Q"  (*dest)  // read/write, memory to be updated atomically
 128       //---<  inputs  >---
 129       : [inc] "a"   (inc)    // read-only.
 130       //---<  clobbered  >---
 131       : "cc", "memory"
 132     );
 133   }
 134 
 135   return upd;
 136 }
 137 
 138 
 139 template<>
 140 template<typename I, typename D>
 141 inline D Atomic::PlatformAdd<8>::add_and_fetch(I inc, D volatile* dest,
 142                                                atomic_memory_order order) const {
 143   STATIC_ASSERT(8 == sizeof(I));
 144   STATIC_ASSERT(8 == sizeof(D));
 145 
 146   D old, upd;
 147 
 148   if (VM_Version::has_LoadAndALUAtomicV1()) {
 149     if (order == memory_order_conservative) { z196_fast_sync(); }
 150     __asm__ __volatile__ (
 151       "   LGR      0,%[inc]                \n\t" // save increment
 152       "   LA       3,%[mem]                \n\t" // force data address into ARG2
 153 //    "   LAAG     %[upd],%[inc],%[mem]    \n\t" // increment and get old value
 154 //    "   LAAG     2,0,0(3)                \n\t" // actually coded instruction
 155       "   .byte    0xeb                    \n\t" // LAA main opcode
 156       "   .byte    0x20                    \n\t" // R1,R3
 157       "   .byte    0x30                    \n\t" // R2,disp1
 158       "   .byte    0x00                    \n\t" // disp2,disp3
 159       "   .byte    0x00                    \n\t" // disp4,disp5
 160       "   .byte    0xe8                    \n\t" // LAA minor opcode
 161       "   AGR      2,0                     \n\t" // calc new value in register
 162       "   LGR      %[upd],2                \n\t" // move to result register
 163       //---<  outputs  >---
 164       : [upd]  "=&d" (upd)    // write-only, updated counter value
 165       , [mem]  "+Q"  (*dest)  // read/write, memory to be updated atomically
 166       //---<  inputs  >---
 167       : [inc]  "a"   (inc)    // read-only.
 168       //---<  clobbered  >---
 169       : "cc", "r0", "r2", "r3", "memory"
 170     );
 171     if (order == memory_order_conservative) { z196_fast_sync(); }
 172   } else {
 173     __asm__ __volatile__ (
 174       "   LG       %[old],%[mem]           \n\t" // get old value
 175       "0: LA       %[upd],0(%[inc],%[old]) \n\t" // calc result
 176       "   CSG      %[old],%[upd],%[mem]    \n\t" // try to xchg res with mem
 177       "   JNE      0b                      \n\t" // no success? -> retry
 178       //---<  outputs  >---
 179       : [old] "=&a" (old)    // write-only, old counter value
 180       , [upd] "=&d" (upd)    // write-only, updated counter value
 181       , [mem] "+Q"  (*dest)  // read/write, memory to be updated atomically
 182       //---<  inputs  >---
 183       : [inc] "a"   (inc)    // read-only.
 184       //---<  clobbered  >---
 185       : "cc", "memory"
 186     );
 187   }
 188 
 189   return upd;
 190 }
 191 
 192 
 193 //-------------
 194 // Atomic::xchg
 195 //-------------
 196 // These methods force the value in memory to be replaced by the new value passed
 197 // in as argument.
 198 //
 199 // The value in memory is replaced by using a compare-and-swap instruction. The
 200 // instruction is retried as often as required. This makes sure that the new
 201 // value can be seen, at least for a very short period of time, by other CPUs.
 202 //
 203 // If we would use a normal "load(old value) store(new value)" sequence,
 204 // the new value could be lost unnoticed, due to a store(new value) from
 205 // another thread.
 206 //
 207 // The return value is the (unchanged) value from memory as it was when the
 208 // replacement succeeded.
 209 template<>
 210 template<typename T>
 211 inline T Atomic::PlatformXchg<4>::operator()(T exchange_value,
 212                                              T volatile* dest,
 213                                              atomic_memory_order unused) const {
 214   STATIC_ASSERT(4 == sizeof(T));
 215   T old;
 216 
 217   __asm__ __volatile__ (
 218     "   LLGF     %[old],%[mem]           \n\t" // get old value
 219     "0: CS       %[old],%[upd],%[mem]    \n\t" // try to xchg upd with mem
 220     "   JNE      0b                      \n\t" // no success? -> retry
 221     //---<  outputs  >---
 222     : [old] "=&d" (old)      // write-only, prev value irrelevant
 223     , [mem] "+Q"  (*dest)    // read/write, memory to be updated atomically
 224     //---<  inputs  >---
 225     : [upd] "d"   (exchange_value) // read-only, value to be written to memory
 226     //---<  clobbered  >---
 227     : "cc", "memory"
 228   );
 229 
 230   return old;
 231 }
 232 
 233 template<>
 234 template<typename T>
 235 inline T Atomic::PlatformXchg<8>::operator()(T exchange_value,
 236                                              T volatile* dest,
 237                                              atomic_memory_order unused) const {
 238   STATIC_ASSERT(8 == sizeof(T));
 239   T old;
 240 
 241   __asm__ __volatile__ (
 242     "   LG       %[old],%[mem]           \n\t" // get old value
 243     "0: CSG      %[old],%[upd],%[mem]    \n\t" // try to xchg upd with mem
 244     "   JNE      0b                      \n\t" // no success? -> retry
 245     //---<  outputs  >---
 246     : [old] "=&d" (old)      // write-only, init from memory
 247     , [mem] "+Q"  (*dest)    // read/write, memory to be updated atomically
 248     //---<  inputs  >---
 249     : [upd] "d"   (exchange_value) // read-only, value to be written to memory
 250     //---<  clobbered  >---
 251     : "cc", "memory"
 252   );
 253 
 254   return old;
 255 }
 256 
 257 //----------------
 258 // Atomic::cmpxchg
 259 //----------------
 260 // These methods compare the value in memory with a given compare value.
 261 // If both values compare equal, the value in memory is replaced with
 262 // the exchange value.
 263 //
 264 // The value in memory is compared and replaced by using a compare-and-swap
 265 // instruction. The instruction is NOT retried (one shot only).
 266 //
 267 // The return value is the (unchanged) value from memory as it was when the
 268 // compare-and-swap instruction completed. A successful exchange operation
 269 // is indicated by (return value == compare_value). If unsuccessful, a new
 270 // exchange value can be calculated based on the return value which is the
 271 // latest contents of the memory location.
 272 //
 273 // Inspecting the return value is the only way for the caller to determine
 274 // if the compare-and-swap instruction was successful:
 275 // - If return value and compare value compare equal, the compare-and-swap
 276 //   instruction was successful and the value in memory was replaced by the
 277 //   exchange value.
 278 // - If return value and compare value compare unequal, the compare-and-swap
 279 //   instruction was not successful. The value in memory was left unchanged.
 280 //
 281 // The s390 processors always fence before and after the csg instructions.
 282 // Thus we ignore the memory ordering argument. The docu says: "A serialization
 283 // function is performed before the operand is fetched and again after the
 284 // operation is completed."
 285 
 286 // No direct support for cmpxchg of bytes; emulate using int.
 287 template<>
 288 struct Atomic::PlatformCmpxchg<1> : Atomic::CmpxchgByteUsingInt {};
 289 
 290 template<>
 291 template<typename T>
 292 inline T Atomic::PlatformCmpxchg<4>::operator()(T xchg_val,
 293                                                 T volatile* dest,
 294                                                 T cmp_val,
 295                                                 atomic_memory_order unused) const {
 296   STATIC_ASSERT(4 == sizeof(T));
 297   T old;
 298 
 299   __asm__ __volatile__ (
 300     "   CS       %[old],%[upd],%[mem]    \n\t" // Try to xchg upd with mem.
 301     // outputs
 302     : [old] "=&d" (old)      // Write-only, prev value irrelevant.
 303     , [mem] "+Q"  (*dest)    // Read/write, memory to be updated atomically.
 304     // inputs
 305     : [upd] "d"   (xchg_val)
 306     ,       "0"   (cmp_val)  // Read-only, initial value for [old] (operand #0).
 307     // clobbered
 308     : "cc", "memory"
 309   );
 310 
 311   return old;
 312 }
 313 
 314 template<>
 315 template<typename T>
 316 inline T Atomic::PlatformCmpxchg<8>::operator()(T xchg_val,
 317                                                 T volatile* dest,
 318                                                 T cmp_val,
 319                                                 atomic_memory_order unused) const {
 320   STATIC_ASSERT(8 == sizeof(T));
 321   T old;
 322 
 323   __asm__ __volatile__ (
 324     "   CSG      %[old],%[upd],%[mem]    \n\t" // Try to xchg upd with mem.
 325     // outputs
 326     : [old] "=&d" (old)      // Write-only, prev value irrelevant.
 327     , [mem] "+Q"  (*dest)    // Read/write, memory to be updated atomically.
 328     // inputs
 329     : [upd] "d"   (xchg_val)
 330     ,       "0"   (cmp_val)  // Read-only, initial value for [old] (operand #0).
 331     // clobbered
 332     : "cc", "memory"
 333   );
 334 
 335   return old;
 336 }
 337 
 338 #endif // OS_CPU_LINUX_S390_VM_ATOMIC_LINUX_S390_INLINE_HPP