| 1 | // Licensed to the .NET Foundation under one or more agreements. |
| 2 | // The .NET Foundation licenses this file to you under the MIT license. |
| 3 | // See the LICENSE file in the project root for more information. |
| 4 | |
| 5 | #include <cstdint> |
| 6 | #include <cstddef> |
| 7 | #include <cassert> |
| 8 | #include <memory> |
| 9 | #include <pthread.h> |
| 10 | #include <signal.h> |
| 11 | |
| 12 | #include "config.h" |
| 13 | #include "common.h" |
| 14 | |
| 15 | #include "gcenv.structs.h" |
| 16 | #include "gcenv.base.h" |
| 17 | #include "gcenv.os.h" |
| 18 | #include "gcenv.unix.inl" |
| 19 | #include "volatile.h" |
| 20 | |
| 21 | #if HAVE_SYS_TIME_H |
| 22 | #include <sys/time.h> |
| 23 | #else |
| 24 | #error "sys/time.h required by GC PAL for the time being" |
| 25 | #endif // HAVE_SYS_TIME_ |
| 26 | |
| 27 | #if HAVE_SYS_MMAN_H |
| 28 | #include <sys/mman.h> |
| 29 | #else |
| 30 | #error "sys/mman.h required by GC PAL" |
| 31 | #endif // HAVE_SYS_MMAN_H |
| 32 | |
| 33 | #ifdef __linux__ |
| 34 | #include <sys/syscall.h> |
| 35 | #endif // __linux__ |
| 36 | |
| 37 | #include <time.h> // nanosleep |
| 38 | #include <sched.h> // sched_yield |
| 39 | #include <errno.h> |
| 40 | #include <unistd.h> // sysconf |
| 41 | #include "globals.h" |
| 42 | #include "cgroup.h" |
| 43 | |
| 44 | #if defined(_ARM_) || defined(_ARM64_) |
| 45 | #define SYSCONF_GET_NUMPROCS _SC_NPROCESSORS_CONF |
| 46 | #else |
| 47 | #define SYSCONF_GET_NUMPROCS _SC_NPROCESSORS_ONLN |
| 48 | #endif |
| 49 | |
| 50 | // The cachced number of logical CPUs observed. |
| 51 | static uint32_t g_logicalCpuCount = 0; |
| 52 | |
| 53 | // Helper memory page used by the FlushProcessWriteBuffers |
| 54 | static uint8_t* g_helperPage = 0; |
| 55 | |
| 56 | // Mutex to make the FlushProcessWriteBuffersMutex thread safe |
| 57 | static pthread_mutex_t g_flushProcessWriteBuffersMutex; |
| 58 | |
| 59 | size_t GetRestrictedPhysicalMemoryLimit(); |
| 60 | bool GetPhysicalMemoryUsed(size_t* val); |
| 61 | bool GetCpuLimit(uint32_t* val); |
| 62 | |
| 63 | static size_t g_RestrictedPhysicalMemoryLimit = 0; |
| 64 | |
| 65 | uint32_t g_pageSizeUnixInl = 0; |
| 66 | |
| 67 | // Initialize the interface implementation |
| 68 | // Return: |
| 69 | // true if it has succeeded, false if it has failed |
| 70 | bool GCToOSInterface::Initialize() |
| 71 | { |
| 72 | int pageSize = sysconf( _SC_PAGE_SIZE ); |
| 73 | |
| 74 | g_pageSizeUnixInl = uint32_t((pageSize > 0) ? pageSize : 0x1000); |
| 75 | |
| 76 | // Calculate and cache the number of processors on this machine |
| 77 | int cpuCount = sysconf(SYSCONF_GET_NUMPROCS); |
| 78 | if (cpuCount == -1) |
| 79 | { |
| 80 | return false; |
| 81 | } |
| 82 | |
| 83 | g_logicalCpuCount = cpuCount; |
| 84 | |
| 85 | assert(g_helperPage == 0); |
| 86 | |
| 87 | g_helperPage = static_cast<uint8_t*>(mmap(0, OS_PAGE_SIZE, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0)); |
| 88 | |
| 89 | if(g_helperPage == MAP_FAILED) |
| 90 | { |
| 91 | return false; |
| 92 | } |
| 93 | |
| 94 | // Verify that the s_helperPage is really aligned to the g_SystemInfo.dwPageSize |
| 95 | assert((((size_t)g_helperPage) & (OS_PAGE_SIZE - 1)) == 0); |
| 96 | |
| 97 | // Locking the page ensures that it stays in memory during the two mprotect |
| 98 | // calls in the FlushProcessWriteBuffers below. If the page was unmapped between |
| 99 | // those calls, they would not have the expected effect of generating IPI. |
| 100 | int status = mlock(g_helperPage, OS_PAGE_SIZE); |
| 101 | |
| 102 | if (status != 0) |
| 103 | { |
| 104 | return false; |
| 105 | } |
| 106 | |
| 107 | status = pthread_mutex_init(&g_flushProcessWriteBuffersMutex, NULL); |
| 108 | if (status != 0) |
| 109 | { |
| 110 | munlock(g_helperPage, OS_PAGE_SIZE); |
| 111 | return false; |
| 112 | } |
| 113 | |
| 114 | #if HAVE_MACH_ABSOLUTE_TIME |
| 115 | kern_return_t machRet; |
| 116 | if ((machRet = mach_timebase_info(&g_TimebaseInfo)) != KERN_SUCCESS) |
| 117 | { |
| 118 | return false; |
| 119 | } |
| 120 | #endif // HAVE_MACH_ABSOLUTE_TIME |
| 121 | |
| 122 | InitializeCGroup(); |
| 123 | |
| 124 | return true; |
| 125 | } |
| 126 | |
| 127 | // Shutdown the interface implementation |
| 128 | void GCToOSInterface::Shutdown() |
| 129 | { |
| 130 | int ret = munlock(g_helperPage, OS_PAGE_SIZE); |
| 131 | assert(ret == 0); |
| 132 | ret = pthread_mutex_destroy(&g_flushProcessWriteBuffersMutex); |
| 133 | assert(ret == 0); |
| 134 | |
| 135 | munmap(g_helperPage, OS_PAGE_SIZE); |
| 136 | |
| 137 | CleanupCGroup(); |
| 138 | } |
| 139 | |
| 140 | // Get numeric id of the current thread if possible on the |
| 141 | // current platform. It is indended for logging purposes only. |
| 142 | // Return: |
| 143 | // Numeric id of the current thread, as best we can retrieve it. |
| 144 | uint64_t GCToOSInterface::GetCurrentThreadIdForLogging() |
| 145 | { |
| 146 | #if defined(__linux__) |
| 147 | return (uint64_t)syscall(SYS_gettid); |
| 148 | #elif HAVE_PTHREAD_GETTHREADID_NP |
| 149 | return (uint64_t)pthread_getthreadid_np(); |
| 150 | #elif HAVE_PTHREAD_THREADID_NP |
| 151 | unsigned long long tid; |
| 152 | pthread_threadid_np(pthread_self(), &tid); |
| 153 | return (uint64_t)tid; |
| 154 | #else |
| 155 | // Fallback in case we don't know how to get integer thread id on the current platform |
| 156 | return (uint64_t)pthread_self(); |
| 157 | #endif |
| 158 | } |
| 159 | |
| 160 | // Get the process ID of the process. |
| 161 | uint32_t GCToOSInterface::GetCurrentProcessId() |
| 162 | { |
| 163 | return getpid(); |
| 164 | } |
| 165 | |
| 166 | // Set ideal affinity for the current thread |
| 167 | // Parameters: |
| 168 | // affinity - ideal processor affinity for the thread |
| 169 | // Return: |
| 170 | // true if it has succeeded, false if it has failed |
| 171 | bool GCToOSInterface::SetCurrentThreadIdealAffinity(GCThreadAffinity* affinity) |
| 172 | { |
| 173 | // TODO(segilles) |
| 174 | return false; |
| 175 | } |
| 176 | |
| 177 | // Get the number of the current processor |
| 178 | uint32_t GCToOSInterface::GetCurrentProcessorNumber() |
| 179 | { |
| 180 | #if HAVE_SCHED_GETCPU |
| 181 | int processorNumber = sched_getcpu(); |
| 182 | assert(processorNumber != -1); |
| 183 | return processorNumber; |
| 184 | #else |
| 185 | return 0; |
| 186 | #endif |
| 187 | } |
| 188 | |
| 189 | // Check if the OS supports getting current processor number |
| 190 | bool GCToOSInterface::CanGetCurrentProcessorNumber() |
| 191 | { |
| 192 | return HAVE_SCHED_GETCPU; |
| 193 | } |
| 194 | |
| 195 | // Flush write buffers of processors that are executing threads of the current process |
| 196 | void GCToOSInterface::FlushProcessWriteBuffers() |
| 197 | { |
| 198 | int status = pthread_mutex_lock(&g_flushProcessWriteBuffersMutex); |
| 199 | assert(status == 0 && "Failed to lock the flushProcessWriteBuffersMutex lock" ); |
| 200 | |
| 201 | // Changing a helper memory page protection from read / write to no access |
| 202 | // causes the OS to issue IPI to flush TLBs on all processors. This also |
| 203 | // results in flushing the processor buffers. |
| 204 | status = mprotect(g_helperPage, OS_PAGE_SIZE, PROT_READ | PROT_WRITE); |
| 205 | assert(status == 0 && "Failed to change helper page protection to read / write" ); |
| 206 | |
| 207 | // Ensure that the page is dirty before we change the protection so that |
| 208 | // we prevent the OS from skipping the global TLB flush. |
| 209 | __sync_add_and_fetch((size_t*)g_helperPage, 1); |
| 210 | |
| 211 | status = mprotect(g_helperPage, OS_PAGE_SIZE, PROT_NONE); |
| 212 | assert(status == 0 && "Failed to change helper page protection to no access" ); |
| 213 | |
| 214 | status = pthread_mutex_unlock(&g_flushProcessWriteBuffersMutex); |
| 215 | assert(status == 0 && "Failed to unlock the flushProcessWriteBuffersMutex lock" ); |
| 216 | } |
| 217 | |
| 218 | // Break into a debugger. Uses a compiler intrinsic if one is available, |
| 219 | // otherwise raises a SIGTRAP. |
| 220 | void GCToOSInterface::DebugBreak() |
| 221 | { |
| 222 | // __has_builtin is only defined by clang. GCC doesn't have a debug |
| 223 | // trap intrinsic anyway. |
| 224 | #ifndef __has_builtin |
| 225 | #define __has_builtin(x) 0 |
| 226 | #endif // __has_builtin |
| 227 | |
| 228 | #if __has_builtin(__builtin_debugtrap) |
| 229 | __builtin_debugtrap(); |
| 230 | #else |
| 231 | raise(SIGTRAP); |
| 232 | #endif |
| 233 | } |
| 234 | |
| 235 | // Causes the calling thread to sleep for the specified number of milliseconds |
| 236 | // Parameters: |
| 237 | // sleepMSec - time to sleep before switching to another thread |
| 238 | void GCToOSInterface::Sleep(uint32_t sleepMSec) |
| 239 | { |
| 240 | if (sleepMSec == 0) |
| 241 | { |
| 242 | return; |
| 243 | } |
| 244 | |
| 245 | timespec requested; |
| 246 | requested.tv_sec = sleepMSec / tccSecondsToMilliSeconds; |
| 247 | requested.tv_nsec = (sleepMSec - requested.tv_sec * tccSecondsToMilliSeconds) * tccMilliSecondsToNanoSeconds; |
| 248 | |
| 249 | timespec remaining; |
| 250 | while (nanosleep(&requested, &remaining) == EINTR) |
| 251 | { |
| 252 | requested = remaining; |
| 253 | } |
| 254 | } |
| 255 | |
| 256 | // Causes the calling thread to yield execution to another thread that is ready to run on the current processor. |
| 257 | // Parameters: |
| 258 | // switchCount - number of times the YieldThread was called in a loop |
| 259 | void GCToOSInterface::YieldThread(uint32_t switchCount) |
| 260 | { |
| 261 | int ret = sched_yield(); |
| 262 | |
| 263 | // sched_yield never fails on Linux, unclear about other OSes |
| 264 | assert(ret == 0); |
| 265 | } |
| 266 | |
| 267 | // Reserve virtual memory range. |
| 268 | // Parameters: |
| 269 | // size - size of the virtual memory range |
| 270 | // alignment - requested memory alignment, 0 means no specific alignment requested |
| 271 | // flags - flags to control special settings like write watching |
| 272 | // Return: |
| 273 | // Starting virtual address of the reserved range |
| 274 | void* GCToOSInterface::VirtualReserve(size_t size, size_t alignment, uint32_t flags) |
| 275 | { |
| 276 | assert(!(flags & VirtualReserveFlags::WriteWatch) && "WriteWatch not supported on Unix" ); |
| 277 | if (alignment == 0) |
| 278 | { |
| 279 | alignment = OS_PAGE_SIZE; |
| 280 | } |
| 281 | |
| 282 | size_t alignedSize = size + (alignment - OS_PAGE_SIZE); |
| 283 | void * pRetVal = mmap(nullptr, alignedSize, PROT_NONE, MAP_ANON | MAP_PRIVATE, -1, 0); |
| 284 | |
| 285 | if (pRetVal != NULL) |
| 286 | { |
| 287 | void * pAlignedRetVal = (void *)(((size_t)pRetVal + (alignment - 1)) & ~(alignment - 1)); |
| 288 | size_t startPadding = (size_t)pAlignedRetVal - (size_t)pRetVal; |
| 289 | if (startPadding != 0) |
| 290 | { |
| 291 | int ret = munmap(pRetVal, startPadding); |
| 292 | assert(ret == 0); |
| 293 | } |
| 294 | |
| 295 | size_t endPadding = alignedSize - (startPadding + size); |
| 296 | if (endPadding != 0) |
| 297 | { |
| 298 | int ret = munmap((void *)((size_t)pAlignedRetVal + size), endPadding); |
| 299 | assert(ret == 0); |
| 300 | } |
| 301 | |
| 302 | pRetVal = pAlignedRetVal; |
| 303 | } |
| 304 | |
| 305 | return pRetVal; |
| 306 | } |
| 307 | |
| 308 | // Release virtual memory range previously reserved using VirtualReserve |
| 309 | // Parameters: |
| 310 | // address - starting virtual address |
| 311 | // size - size of the virtual memory range |
| 312 | // Return: |
| 313 | // true if it has succeeded, false if it has failed |
| 314 | bool GCToOSInterface::VirtualRelease(void* address, size_t size) |
| 315 | { |
| 316 | int ret = munmap(address, size); |
| 317 | |
| 318 | return (ret == 0); |
| 319 | } |
| 320 | |
| 321 | // Commit virtual memory range. It must be part of a range reserved using VirtualReserve. |
| 322 | // Parameters: |
| 323 | // address - starting virtual address |
| 324 | // size - size of the virtual memory range |
| 325 | // Return: |
| 326 | // true if it has succeeded, false if it has failed |
| 327 | bool GCToOSInterface::VirtualCommit(void* address, size_t size, uint32_t node) |
| 328 | { |
| 329 | assert(node == NUMA_NODE_UNDEFINED && "Numa allocation is not ported to local GC on unix yet" ); |
| 330 | return mprotect(address, size, PROT_WRITE | PROT_READ) == 0; |
| 331 | } |
| 332 | |
| 333 | // Decomit virtual memory range. |
| 334 | // Parameters: |
| 335 | // address - starting virtual address |
| 336 | // size - size of the virtual memory range |
| 337 | // Return: |
| 338 | // true if it has succeeded, false if it has failed |
| 339 | bool GCToOSInterface::VirtualDecommit(void* address, size_t size) |
| 340 | { |
| 341 | // TODO: This can fail, however the GC does not handle the failure gracefully |
| 342 | // Explicitly calling mmap instead of mprotect here makes it |
| 343 | // that much more clear to the operating system that we no |
| 344 | // longer need these pages. Also, GC depends on re-commited pages to |
| 345 | // be zeroed-out. |
| 346 | return mmap(address, size, PROT_NONE, MAP_FIXED | MAP_ANON | MAP_PRIVATE, -1, 0) != NULL; |
| 347 | } |
| 348 | |
| 349 | // Reset virtual memory range. Indicates that data in the memory range specified by address and size is no |
| 350 | // longer of interest, but it should not be decommitted. |
| 351 | // Parameters: |
| 352 | // address - starting virtual address |
| 353 | // size - size of the virtual memory range |
| 354 | // unlock - true if the memory range should also be unlocked |
| 355 | // Return: |
| 356 | // true if it has succeeded, false if it has failed |
| 357 | bool GCToOSInterface::VirtualReset(void * address, size_t size, bool unlock) |
| 358 | { |
| 359 | int st; |
| 360 | #if HAVE_MADV_FREE |
| 361 | // Try to use MADV_FREE if supported. It tells the kernel that the application doesn't |
| 362 | // need the pages in the range. Freeing the pages can be delayed until a memory pressure |
| 363 | // occurs. |
| 364 | st = madvise(address, size, MADV_FREE); |
| 365 | if (st != 0) |
| 366 | #endif |
| 367 | { |
| 368 | // In case the MADV_FREE is not supported, use MADV_DONTNEED |
| 369 | st = madvise(address, size, MADV_DONTNEED); |
| 370 | } |
| 371 | |
| 372 | return (st == 0); |
| 373 | } |
| 374 | |
| 375 | // Check if the OS supports write watching |
| 376 | bool GCToOSInterface::SupportsWriteWatch() |
| 377 | { |
| 378 | return false; |
| 379 | } |
| 380 | |
| 381 | // Reset the write tracking state for the specified virtual memory range. |
| 382 | // Parameters: |
| 383 | // address - starting virtual address |
| 384 | // size - size of the virtual memory range |
| 385 | void GCToOSInterface::ResetWriteWatch(void* address, size_t size) |
| 386 | { |
| 387 | assert(!"should never call ResetWriteWatch on Unix" ); |
| 388 | } |
| 389 | |
| 390 | // Retrieve addresses of the pages that are written to in a region of virtual memory |
| 391 | // Parameters: |
| 392 | // resetState - true indicates to reset the write tracking state |
| 393 | // address - starting virtual address |
| 394 | // size - size of the virtual memory range |
| 395 | // pageAddresses - buffer that receives an array of page addresses in the memory region |
| 396 | // pageAddressesCount - on input, size of the lpAddresses array, in array elements |
| 397 | // on output, the number of page addresses that are returned in the array. |
| 398 | // Return: |
| 399 | // true if it has succeeded, false if it has failed |
| 400 | bool GCToOSInterface::GetWriteWatch(bool resetState, void* address, size_t size, void** pageAddresses, uintptr_t* pageAddressesCount) |
| 401 | { |
| 402 | assert(!"should never call GetWriteWatch on Unix" ); |
| 403 | return false; |
| 404 | } |
| 405 | |
| 406 | // Get size of the largest cache on the processor die |
| 407 | // Parameters: |
| 408 | // trueSize - true to return true cache size, false to return scaled up size based on |
| 409 | // the processor architecture |
| 410 | // Return: |
| 411 | // Size of the cache |
| 412 | size_t GCToOSInterface::GetCacheSizePerLogicalCpu(bool trueSize) |
| 413 | { |
| 414 | // TODO(segilles) processor detection |
| 415 | return 0; |
| 416 | } |
| 417 | |
| 418 | // Sets the calling thread's affinity to only run on the processor specified |
| 419 | // in the GCThreadAffinity structure. |
| 420 | // Parameters: |
| 421 | // affinity - The requested affinity for the calling thread. At most one processor |
| 422 | // can be provided. |
| 423 | // Return: |
| 424 | // true if setting the affinity was successful, false otherwise. |
| 425 | bool GCToOSInterface::SetThreadAffinity(GCThreadAffinity* affinity) |
| 426 | { |
| 427 | // [LOCALGC TODO] Thread affinity for unix |
| 428 | return false; |
| 429 | } |
| 430 | |
| 431 | // Boosts the calling thread's thread priority to a level higher than the default |
| 432 | // for new threads. |
| 433 | // Parameters: |
| 434 | // None. |
| 435 | // Return: |
| 436 | // true if the priority boost was successful, false otherwise. |
| 437 | bool GCToOSInterface::BoostThreadPriority() |
| 438 | { |
| 439 | // [LOCALGC TODO] Thread priority for unix |
| 440 | return false; |
| 441 | } |
| 442 | |
| 443 | /*++ |
| 444 | Function: |
| 445 | GetFullAffinityMask |
| 446 | |
| 447 | Get affinity mask for the specified number of processors with all |
| 448 | the processors enabled. |
| 449 | --*/ |
| 450 | static uintptr_t GetFullAffinityMask(int cpuCount) |
| 451 | { |
| 452 | return ((uintptr_t)1 << (cpuCount)) - 1; |
| 453 | } |
| 454 | |
| 455 | // Get affinity mask of the current process |
| 456 | // Parameters: |
| 457 | // processMask - affinity mask for the specified process |
| 458 | // systemMask - affinity mask for the system |
| 459 | // Return: |
| 460 | // true if it has succeeded, false if it has failed |
| 461 | // Remarks: |
| 462 | // A process affinity mask is a bit vector in which each bit represents the processors that |
| 463 | // a process is allowed to run on. A system affinity mask is a bit vector in which each bit |
| 464 | // represents the processors that are configured into a system. |
| 465 | // A process affinity mask is a subset of the system affinity mask. A process is only allowed |
| 466 | // to run on the processors configured into a system. Therefore, the process affinity mask cannot |
| 467 | // specify a 1 bit for a processor when the system affinity mask specifies a 0 bit for that processor. |
| 468 | bool GCToOSInterface::GetCurrentProcessAffinityMask(uintptr_t* processAffinityMask, uintptr_t* systemAffinityMask) |
| 469 | { |
| 470 | if (g_logicalCpuCount > 64) |
| 471 | { |
| 472 | *processAffinityMask = 0; |
| 473 | *systemAffinityMask = 0; |
| 474 | return true; |
| 475 | } |
| 476 | |
| 477 | uintptr_t systemMask = GetFullAffinityMask(g_logicalCpuCount); |
| 478 | |
| 479 | #if HAVE_SCHED_GETAFFINITY |
| 480 | |
| 481 | int pid = getpid(); |
| 482 | cpu_set_t cpuSet; |
| 483 | int st = sched_getaffinity(pid, sizeof(cpu_set_t), &cpuSet); |
| 484 | if (st == 0) |
| 485 | { |
| 486 | uintptr_t processMask = 0; |
| 487 | |
| 488 | for (int i = 0; i < g_logicalCpuCount; i++) |
| 489 | { |
| 490 | if (CPU_ISSET(i, &cpuSet)) |
| 491 | { |
| 492 | processMask |= ((uintptr_t)1) << i; |
| 493 | } |
| 494 | } |
| 495 | |
| 496 | *processAffinityMask = processMask; |
| 497 | *systemAffinityMask = systemMask; |
| 498 | return true; |
| 499 | } |
| 500 | else if (errno == EINVAL) |
| 501 | { |
| 502 | // There are more processors than can fit in a cpu_set_t |
| 503 | // return zero in both masks. |
| 504 | *processAffinityMask = 0; |
| 505 | *systemAffinityMask = 0; |
| 506 | return true; |
| 507 | } |
| 508 | else |
| 509 | { |
| 510 | // We should not get any of the errors that the sched_getaffinity can return since none |
| 511 | // of them applies for the current thread, so this is an unexpected kind of failure. |
| 512 | return false; |
| 513 | } |
| 514 | |
| 515 | #else // HAVE_SCHED_GETAFFINITY |
| 516 | |
| 517 | // There is no API to manage thread affinity, so let's return both affinity masks |
| 518 | // with all the CPUs on the system set. |
| 519 | *systemAffinityMask = systemMask; |
| 520 | *processAffinityMask = systemMask; |
| 521 | return true; |
| 522 | |
| 523 | #endif // HAVE_SCHED_GETAFFINITY |
| 524 | } |
| 525 | |
| 526 | // Get number of processors assigned to the current process |
| 527 | // Return: |
| 528 | // The number of processors |
| 529 | uint32_t GCToOSInterface::GetCurrentProcessCpuCount() |
| 530 | { |
| 531 | uintptr_t pmask, smask; |
| 532 | uint32_t cpuLimit; |
| 533 | |
| 534 | if (!GetCurrentProcessAffinityMask(&pmask, &smask)) |
| 535 | return 1; |
| 536 | |
| 537 | pmask &= smask; |
| 538 | |
| 539 | int count = 0; |
| 540 | while (pmask) |
| 541 | { |
| 542 | pmask &= (pmask - 1); |
| 543 | count++; |
| 544 | } |
| 545 | |
| 546 | // GetProcessAffinityMask can return pmask=0 and smask=0 on systems with more |
| 547 | // than 64 processors, which would leave us with a count of 0. Since the GC |
| 548 | // expects there to be at least one processor to run on (and thus at least one |
| 549 | // heap), we'll return 64 here if count is 0, since there are likely a ton of |
| 550 | // processors available in that case. The GC also cannot (currently) handle |
| 551 | // the case where there are more than 64 processors, so we will return a |
| 552 | // maximum of 64 here. |
| 553 | if (count == 0 || count > 64) |
| 554 | count = 64; |
| 555 | |
| 556 | if (GetCpuLimit(&cpuLimit) && cpuLimit < count) |
| 557 | count = cpuLimit; |
| 558 | |
| 559 | return count; |
| 560 | } |
| 561 | |
| 562 | // Return the size of the user-mode portion of the virtual address space of this process. |
| 563 | // Return: |
| 564 | // non zero if it has succeeded, 0 if it has failed |
| 565 | size_t GCToOSInterface::GetVirtualMemoryLimit() |
| 566 | { |
| 567 | #ifdef BIT64 |
| 568 | // There is no API to get the total virtual address space size on |
| 569 | // Unix, so we use a constant value representing 128TB, which is |
| 570 | // the approximate size of total user virtual address space on |
| 571 | // the currently supported Unix systems. |
| 572 | static const uint64_t _128TB = (1ull << 47); |
| 573 | return _128TB; |
| 574 | #else |
| 575 | return (size_t)-1; |
| 576 | #endif |
| 577 | } |
| 578 | |
| 579 | // Get the physical memory that this process can use. |
| 580 | // Return: |
| 581 | // non zero if it has succeeded, 0 if it has failed |
| 582 | // Remarks: |
| 583 | // If a process runs with a restricted memory limit, it returns the limit. If there's no limit |
| 584 | // specified, it returns amount of actual physical memory. |
| 585 | uint64_t GCToOSInterface::GetPhysicalMemoryLimit() |
| 586 | { |
| 587 | size_t restricted_limit; |
| 588 | // The limit was not cached |
| 589 | if (g_RestrictedPhysicalMemoryLimit == 0) |
| 590 | { |
| 591 | restricted_limit = GetRestrictedPhysicalMemoryLimit(); |
| 592 | VolatileStore(&g_RestrictedPhysicalMemoryLimit, restricted_limit); |
| 593 | } |
| 594 | restricted_limit = g_RestrictedPhysicalMemoryLimit; |
| 595 | |
| 596 | if (restricted_limit != 0 && restricted_limit != SIZE_T_MAX) |
| 597 | return restricted_limit; |
| 598 | |
| 599 | long pages = sysconf(_SC_PHYS_PAGES); |
| 600 | if (pages == -1) |
| 601 | { |
| 602 | return 0; |
| 603 | } |
| 604 | |
| 605 | long pageSize = sysconf(_SC_PAGE_SIZE); |
| 606 | if (pageSize == -1) |
| 607 | { |
| 608 | return 0; |
| 609 | } |
| 610 | |
| 611 | return pages * pageSize; |
| 612 | } |
| 613 | |
| 614 | // Get memory status |
| 615 | // Parameters: |
| 616 | // memory_load - A number between 0 and 100 that specifies the approximate percentage of physical memory |
| 617 | // that is in use (0 indicates no memory use and 100 indicates full memory use). |
| 618 | // available_physical - The amount of physical memory currently available, in bytes. |
| 619 | // available_page_file - The maximum amount of memory the current process can commit, in bytes. |
| 620 | void GCToOSInterface::GetMemoryStatus(uint32_t* memory_load, uint64_t* available_physical, uint64_t* available_page_file) |
| 621 | { |
| 622 | if (memory_load != nullptr || available_physical != nullptr) |
| 623 | { |
| 624 | uint64_t total = GetPhysicalMemoryLimit(); |
| 625 | |
| 626 | uint64_t available = 0; |
| 627 | uint32_t load = 0; |
| 628 | size_t used; |
| 629 | |
| 630 | // Get the physical memory in use - from it, we can get the physical memory available. |
| 631 | // We do this only when we have the total physical memory available. |
| 632 | if (total > 0 && GetPhysicalMemoryUsed(&used)) |
| 633 | { |
| 634 | available = total > used ? total-used : 0; |
| 635 | load = (uint32_t)(((float)used * 100) / (float)total); |
| 636 | } |
| 637 | |
| 638 | if (memory_load != nullptr) |
| 639 | *memory_load = load; |
| 640 | if (available_physical != nullptr) |
| 641 | *available_physical = available; |
| 642 | } |
| 643 | |
| 644 | if (available_page_file != nullptr) |
| 645 | *available_page_file = 0; |
| 646 | } |
| 647 | |
| 648 | // Get a high precision performance counter |
| 649 | // Return: |
| 650 | // The counter value |
| 651 | int64_t GCToOSInterface::QueryPerformanceCounter() |
| 652 | { |
| 653 | // TODO: This is not a particularly efficient implementation - we certainly could |
| 654 | // do much more specific platform-dependent versions if we find that this method |
| 655 | // runs hot. However, most likely it does not. |
| 656 | struct timeval tv; |
| 657 | if (gettimeofday(&tv, NULL) == -1) |
| 658 | { |
| 659 | assert(!"gettimeofday() failed" ); |
| 660 | // TODO (segilles) unconditional asserts |
| 661 | return 0; |
| 662 | } |
| 663 | return (int64_t) tv.tv_sec * (int64_t) tccSecondsToMicroSeconds + (int64_t) tv.tv_usec; |
| 664 | } |
| 665 | |
| 666 | // Get a frequency of the high precision performance counter |
| 667 | // Return: |
| 668 | // The counter frequency |
| 669 | int64_t GCToOSInterface::QueryPerformanceFrequency() |
| 670 | { |
| 671 | // The counter frequency of gettimeofday is in microseconds. |
| 672 | return tccSecondsToMicroSeconds; |
| 673 | } |
| 674 | |
| 675 | // Get a time stamp with a low precision |
| 676 | // Return: |
| 677 | // Time stamp in milliseconds |
| 678 | uint32_t GCToOSInterface::GetLowPrecisionTimeStamp() |
| 679 | { |
| 680 | // TODO(segilles) this is pretty naive, we can do better |
| 681 | uint64_t retval = 0; |
| 682 | struct timeval tv; |
| 683 | if (gettimeofday(&tv, NULL) == 0) |
| 684 | { |
| 685 | retval = (tv.tv_sec * tccSecondsToMilliSeconds) + (tv.tv_usec / tccMilliSecondsToMicroSeconds); |
| 686 | } |
| 687 | else |
| 688 | { |
| 689 | assert(!"gettimeofday() failed\n" ); |
| 690 | } |
| 691 | |
| 692 | return retval; |
| 693 | } |
| 694 | |
| 695 | // Gets the total number of processors on the machine, not taking |
| 696 | // into account current process affinity. |
| 697 | // Return: |
| 698 | // Number of processors on the machine |
| 699 | uint32_t GCToOSInterface::GetTotalProcessorCount() |
| 700 | { |
| 701 | // Calculated in GCToOSInterface::Initialize using |
| 702 | // sysconf(_SC_NPROCESSORS_ONLN) |
| 703 | return g_logicalCpuCount; |
| 704 | } |
| 705 | |
| 706 | bool GCToOSInterface::CanEnableGCNumaAware() |
| 707 | { |
| 708 | return false; |
| 709 | } |
| 710 | |
| 711 | bool GCToOSInterface::GetNumaProcessorNode(PPROCESSOR_NUMBER proc_no, uint16_t *node_no) |
| 712 | { |
| 713 | assert(!"Numa has not been ported to local GC for unix" ); |
| 714 | return false; |
| 715 | } |
| 716 | |
| 717 | bool GCToOSInterface::CanEnableGCCPUGroups() |
| 718 | { |
| 719 | return false; |
| 720 | } |
| 721 | |
| 722 | void GCToOSInterface::GetGroupForProcessor(uint16_t processor_number, uint16_t* group_number, uint16_t* group_processor_number) |
| 723 | { |
| 724 | assert(!"CpuGroup has not been ported to local GC for unix" ); |
| 725 | } |
| 726 | |
| 727 | // Initialize the critical section |
| 728 | void CLRCriticalSection::Initialize() |
| 729 | { |
| 730 | int st = pthread_mutex_init(&m_cs.mutex, NULL); |
| 731 | assert(st == 0); |
| 732 | } |
| 733 | |
| 734 | // Destroy the critical section |
| 735 | void CLRCriticalSection::Destroy() |
| 736 | { |
| 737 | int st = pthread_mutex_destroy(&m_cs.mutex); |
| 738 | assert(st == 0); |
| 739 | } |
| 740 | |
| 741 | // Enter the critical section. Blocks until the section can be entered. |
| 742 | void CLRCriticalSection::Enter() |
| 743 | { |
| 744 | pthread_mutex_lock(&m_cs.mutex); |
| 745 | } |
| 746 | |
| 747 | // Leave the critical section |
| 748 | void CLRCriticalSection::Leave() |
| 749 | { |
| 750 | pthread_mutex_unlock(&m_cs.mutex); |
| 751 | } |
| 752 | |