1 // Copyright 2006 Google Inc. All Rights Reserved.
2 // Author: nsanders, menderico
4 // Licensed under the Apache License, Version 2.0 (the "License");
5 // you may not use this file except in compliance with the License.
6 // You may obtain a copy of the License at
8 // http://www.apache.org/licenses/LICENSE-2.0
10 // Unless required by applicable law or agreed to in writing, software
11 // distributed under the License is distributed on an "AS IS" BASIS,
12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 // See the License for the specific language governing permissions and
14 // limitations under the License.
16 #ifndef STRESSAPPTEST_OS_H_ // NOLINT
17 #define STRESSAPPTEST_OS_H_
21 #include <sys/syscall.h>
28 // This file must work with autoconf on its public version,
29 // so these includes are correct.
30 #include "adler32memcpy.h" // NOLINT
31 #include "sattypes.h" // NOLINT
32 #include "clock.h" // NOLINT
34 const char kPagemapPath[] = "/proc/self/pagemap";
47 typedef vector<PCIDevice*> PCIDevices;
53 // This class implements OS/Platform specific funtions.
59 // Set the minimum amount of hugepages that should be available for testing.
60 // Must be set before Initialize().
61 void SetMinimumHugepagesSize(int64 min_bytes) {
62 min_hugepages_bytes_ = min_bytes;
65 // Set the minium amount of memory that should not be allocated. This only
66 // has any affect if hugepages are not used.
67 // Must be set before Initialize().
68 void SetReserveSize(int64 reserve_mb) {
69 reserve_mb_ = reserve_mb;
72 // Set parameters needed to translate physical address to memory module.
73 void SetDramMappingParams(uintptr_t channel_hash, int channel_width,
74 vector< vector<string> > *channels) {
75 channel_hash_ = channel_hash;
76 channel_width_ = channel_width;
80 // Initializes data strctures and open files.
81 // Returns false on error.
82 virtual bool Initialize();
84 // Virtual to physical. This implementation is optional for
85 // subclasses to implement.
86 // Takes a pointer, and returns the corresponding bus address.
87 virtual uint64 VirtualToPhysical(void *vaddr);
89 // Prints failed dimm. This implementation is optional for
90 // subclasses to implement.
91 // Takes a bus address and string, and prints the DIMM name
92 // into the string. Returns the DIMM number that corresponds to the
93 // address given, or -1 if unable to identify the DIMM number.
94 // Note that subclass implementations of FindDimm() MUST fill
95 // buf with at LEAST one non-whitespace character (provided len > 0).
96 virtual int FindDimm(uint64 addr, char *buf, int len);
98 // Classifies addresses according to "regions"
99 // This may mean different things on different platforms.
100 virtual int32 FindRegion(uint64 paddr);
101 // Find cpu cores associated with a region. Either NUMA or arbitrary.
102 virtual cpu_set_t *FindCoreMask(int32 region);
103 // Return cpu cores associated with a region in a hex string.
104 virtual string FindCoreMaskFormat(int32 region);
106 // Returns the HD device that contains this file.
107 virtual string FindFileDevice(string filename);
109 // Returns a list of paths coresponding to HD devices found on this machine.
110 virtual list<string> FindFileDevices();
112 // Polls for errors. This implementation is optional.
113 // This will poll once for errors and return zero iff no errors were found.
114 virtual int ErrorPoll();
116 // Delay an appropriate amount of time between polling.
117 virtual void ErrorWait();
119 // Report errors. This implementation is mandatory.
120 // This will output a machine readable line regarding the error.
121 virtual bool ErrorReport(const char *part, const char *symptom, int count);
123 // Flushes page cache. Used to circumvent the page cache when doing disk
124 // I/O. This will be a NOP until ActivateFlushPageCache() is called, which
125 // is typically done when opening a file with O_DIRECT fails.
126 // Returns false on error, true on success or NOP.
127 // Subclasses may implement this in machine specific ways..
128 virtual bool FlushPageCache(void);
129 // Enable FlushPageCache() to actually do the flush instead of being a NOP.
130 virtual void ActivateFlushPageCache(void);
132 // Flushes cacheline. Used to distinguish read or write errors.
133 // Subclasses may implement this in machine specific ways..
134 // Takes a pointer, and flushed the cacheline containing that pointer.
135 virtual void Flush(void *vaddr);
137 // Fast flush, for use in performance critical code.
138 // This is bound at compile time, and will not pick up
139 // any runtime machine configuration info.
140 inline static void FastFlush(void *vaddr) {
141 #ifdef STRESSAPPTEST_CPU_PPC
142 asm volatile("dcbf 0,%0; sync" : : "r" (vaddr));
143 #elif defined(STRESSAPPTEST_CPU_X86_64) || defined(STRESSAPPTEST_CPU_I686)
144 // Put mfence before and after clflush to make sure:
145 // 1. The write before the clflush is committed to memory bus;
146 // 2. The read after the clflush is hitting the memory bus.
148 // From Intel manual:
149 // CLFLUSH is only ordered by the MFENCE instruction. It is not guaranteed
150 // to be ordered by any other fencing, serializing or other CLFLUSH
151 // instruction. For example, software can use an MFENCE instruction to
152 // insure that previous stores are included in the write-back.
153 asm volatile("mfence");
154 asm volatile("clflush (%0)" : : "r" (vaddr));
155 asm volatile("mfence");
156 #elif defined(STRESSAPPTEST_CPU_ARMV7A)
157 // ARMv7a cachelines are 8 words (32 bytes).
158 syscall(__ARM_NR_cacheflush, vaddr, reinterpret_cast<char*>(vaddr) + 32, 0);
160 #warning "Unsupported CPU type: Unable to force cache flushes."
164 // Fast flush, for use in performance critical code.
165 // This is bound at compile time, and will not pick up
166 // any runtime machine configuration info. Takes a NULL-terminated
167 // array of addresses to flush.
168 inline static void FastFlushList(void **vaddrs) {
169 #ifdef STRESSAPPTEST_CPU_PPC
171 asm volatile("dcbf 0,%0" : : "r" (*vaddrs++));
173 asm volatile("sync");
174 #elif defined(STRESSAPPTEST_CPU_X86_64) || defined(STRESSAPPTEST_CPU_I686)
175 // Put mfence before and after clflush to make sure:
176 // 1. The write before the clflush is committed to memory bus;
177 // 2. The read after the clflush is hitting the memory bus.
179 // From Intel manual:
180 // CLFLUSH is only ordered by the MFENCE instruction. It is not guaranteed
181 // to be ordered by any other fencing, serializing or other CLFLUSH
182 // instruction. For example, software can use an MFENCE instruction to
183 // insure that previous stores are included in the write-back.
184 asm volatile("mfence");
186 asm volatile("clflush (%0)" : : "r" (*vaddrs++));
188 asm volatile("mfence");
189 #elif defined(STRESSAPPTEST_CPU_ARMV7A)
191 FastFlush(*vaddrs++);
194 #warning "Unsupported CPU type: Unable to force cache flushes."
198 // Fast flush hint, for use in performance critical code.
199 // This is bound at compile time, and will not pick up
200 // any runtime machine configuration info. Note that this
201 // will not guarantee that a flush happens, but will at least
202 // hint that it should. This is useful for speeding up
203 // parallel march algorithms.
204 inline static void FastFlushHint(void *vaddr) {
205 #ifdef STRESSAPPTEST_CPU_PPC
206 asm volatile("dcbf 0,%0" : : "r" (vaddr));
207 #elif defined(STRESSAPPTEST_CPU_X86_64) || defined(STRESSAPPTEST_CPU_I686)
208 // From Intel manual:
209 // CLFLUSH is only ordered by the MFENCE instruction. It is not guaranteed
210 // to be ordered by any other fencing, serializing or other CLFLUSH
211 // instruction. For example, software can use an MFENCE instruction to
212 // insure that previous stores are included in the write-back.
213 asm volatile("clflush (%0)" : : "r" (vaddr));
214 #elif defined(STRESSAPPTEST_CPU_ARMV7A)
217 #warning "Unsupported CPU type: Unable to force cache flushes."
221 // Fast flush, for use in performance critical code.
222 // This is bound at compile time, and will not pick up
223 // any runtime machine configuration info. Sync's any
224 // transactions for ordering FastFlushHints.
225 inline static void FastFlushSync() {
226 #ifdef STRESSAPPTEST_CPU_PPC
227 asm volatile("sync");
228 #elif defined(STRESSAPPTEST_CPU_X86_64) || defined(STRESSAPPTEST_CPU_I686)
229 // Put mfence before and after clflush to make sure:
230 // 1. The write before the clflush is committed to memory bus;
231 // 2. The read after the clflush is hitting the memory bus.
233 // From Intel manual:
234 // CLFLUSH is only ordered by the MFENCE instruction. It is not guaranteed
235 // to be ordered by any other fencing, serializing or other CLFLUSH
236 // instruction. For example, software can use an MFENCE instruction to
237 // insure that previous stores are included in the write-back.
238 asm volatile("mfence");
239 #elif defined(STRESSAPPTEST_CPU_ARMV7A)
240 // This is a NOP, FastFlushHint() always does a full flush, so there's
241 // nothing to do for FastFlushSync().
243 #warning "Unsupported CPU type: Unable to force cache flushes."
247 // Get time in cpu timer ticks. Useful for matching MCEs with software
249 inline static uint64 GetTimestamp(void) {
251 #ifdef STRESSAPPTEST_CPU_PPC
252 uint32 tbl, tbu, temp;
260 : "=r"(tbl), "=r"(tbu), "=r"(temp)
264 tsc = (static_cast<uint64>(tbu) << 32) | static_cast<uint64>(tbl);
265 #elif defined(STRESSAPPTEST_CPU_X86_64) || defined(STRESSAPPTEST_CPU_I686)
267 __asm __volatile("rdtsc" : "=a" (data.l32.l), "=d"(data.l32.h));
269 #elif defined(STRESSAPPTEST_CPU_ARMV7A)
270 #warning "Unsupported CPU type ARMV7A: your timer may not function correctly"
273 #warning "Unsupported CPU type: your timer may not function correctly"
279 // Find the free memory on the machine.
280 virtual int64 FindFreeMemSize();
282 // Allocates test memory of length bytes.
283 // Subclasses must implement this.
284 // Call PepareTestMem to get a pointer.
285 virtual int64 AllocateAllMem(); // Returns length.
287 virtual bool AllocateTestMem(int64 length, uint64 paddr_base);
288 virtual void FreeTestMem();
290 // Prepares the memory for use. You must call this
291 // before using test memory, and after you are done.
292 virtual void *PrepareTestMem(uint64 offset, uint64 length);
293 virtual void ReleaseTestMem(void *addr, uint64 offset, uint64 length);
295 // Machine type detected. Can we implement all these functions correctly?
296 // Returns true if machine type is detected and implemented.
297 virtual bool IsSupported();
299 // Returns 32 for 32-bit, 64 for 64-bit.
300 virtual int AddressMode();
301 // Update OsLayer state regarding cpu support for various features.
302 virtual void GetFeatures();
304 // Open, read, write pci cfg through /proc/bus/pci. fd is /proc/pci file.
305 virtual int PciOpen(int bus, int device, int function);
306 virtual void PciWrite(int fd, uint32 offset, uint32 value, int width);
307 virtual uint32 PciRead(int fd, uint32 offset, int width);
310 virtual bool ReadMSR(uint32 core, uint32 address, uint64 *data);
311 virtual bool WriteMSR(uint32 core, uint32 address, uint64 *data);
313 // Extract bits [n+len-1, n] from a 32 bit word.
314 // so GetBitField(0x0f00, 8, 4) == 0xf.
315 virtual uint32 GetBitField(uint32 val, uint32 n, uint32 len);
317 // Platform and CPU specific CPU-stressing function.
318 // Returns true on success, false otherwise.
319 virtual bool CpuStressWorkload();
321 // Causes false errors for unittesting.
322 // Setting to "true" causes errors to be injected.
323 void set_error_injection(bool errors) { error_injection_ = errors; }
324 bool error_injection() const { return error_injection_; }
326 // Is SAT using normal malloc'd memory, or exotic mmap'd memory.
327 bool normal_mem() const { return normal_mem_; }
329 // Get numa config, if available..
330 int num_nodes() const { return num_nodes_; }
331 int num_cpus() const { return num_cpus_; }
333 // Handle to platform-specific error diagnoser.
334 ErrorDiag *error_diagnoser_;
336 // Disambiguate between different "warm" memcopies.
337 virtual bool AdlerMemcpyWarm(uint64 *dstmem, uint64 *srcmem,
338 unsigned int size_in_bytes,
339 AdlerChecksum *checksum);
341 // Store a callback to use to print
342 // app-specific info about the last error location.
343 // This call back is called with a physical address, and the app can fill in
344 // the most recent transaction that occurred at that address.
345 typedef bool (*ErrCallback)(uint64 paddr, string *buf);
346 void set_err_log_callback(
347 ErrCallback err_log_callback) {
348 err_log_callback_ = err_log_callback;
350 ErrCallback get_err_log_callback() { return err_log_callback_; }
352 // Set a clock object that can be overridden for use with unit tests.
353 void SetClock(Clock *clock) {
358 time_initialized_ = clock_->Now();
362 void *testmem_; // Location of test memory.
363 uint64 testmemsize_; // Size of test memory.
364 int64 totalmemsize_; // Size of available memory.
365 int64 min_hugepages_bytes_; // Minimum hugepages size.
366 int64 reserve_mb_; // Minimum amount of memory to reserve in MB.
367 bool error_injection_; // Do error injection?
368 bool normal_mem_; // Memory DMA capable?
369 bool use_hugepages_; // Use hugepage shmem?
370 bool use_posix_shm_; // Use 4k page shmem?
371 bool dynamic_mapped_shmem_; // Conserve virtual address space.
372 bool mmapped_allocation_; // Was memory allocated using mmap()?
373 int shmid_; // Handle to shmem
374 vector< vector<string> > *channels_; // Memory module names per channel.
375 uint64 channel_hash_; // Mask of address bits XORed for channel.
376 int channel_width_; // Channel width in bits.
378 int64 regionsize_; // Size of memory "regions"
379 int regioncount_; // Number of memory "regions"
380 int num_cpus_; // Number of cpus in the system.
381 int num_nodes_; // Number of nodes in the system.
382 int num_cpus_per_node_; // Number of cpus per node in the system.
383 int address_mode_; // Are we running 32 or 64 bit?
384 bool has_vector_; // Do we have sse2/neon instructions?
385 bool has_clflush_; // Do we have clflush instructions?
386 bool use_flush_page_cache_; // Do we need to flush the page cache?
389 time_t time_initialized_; // Start time of test.
391 vector<cpu_set_t> cpu_sets_; // Cache for cpu masks.
392 vector<bool> cpu_sets_valid_; // If the cpu mask cache is valid.
394 // Get file descriptor for dev msr.
395 virtual int OpenMSR(uint32 core, uint32 address);
397 // Look up how many hugepages there are.
398 virtual int64 FindHugePages();
400 // Link to find last transaction at an error location.
401 ErrCallback err_log_callback_;
403 // Object to wrap the time function.
407 DISALLOW_COPY_AND_ASSIGN(OsLayer);
410 // Selects and returns the proper OS and hardware interface. Does not call
411 // OsLayer::Initialize() on the new object.
412 OsLayer *OsLayerFactory(const std::map<std::string, std::string> &options);
414 #endif // STRESSAPPTEST_OS_H_ NOLINT