1 =================================================== 2 Scalable Vector Extension support for AArch64 Linux 3 =================================================== 4 5 Author: Dave Martin <Dave.Martin@arm.com> 6 7 Date: 4 August 2017 8 9 This document outlines briefly the interface provided to userspace by Linux in 10 order to support use of the ARM Scalable Vector Extension (SVE), including 11 interactions with Streaming SVE mode added by the Scalable Matrix Extension 12 (SME). 13 14 This is an outline of the most important features and issues only and not 15 intended to be exhaustive. 16 17 This document does not aim to describe the SVE architecture or programmer's 18 model. To aid understanding, a minimal description of relevant programmer's 19 model features for SVE is included in Appendix A. 20 21 22 1. General 23 ----------- 24 25 * SVE registers Z0..Z31, P0..P15 and FFR and the current vector length VL, are 26 tracked per-thread. 27 28 * In streaming mode FFR is not accessible unless HWCAP2_SME_FA64 is present 29 in the system, when it is not supported and these interfaces are used to 30 access streaming mode FFR is read and written as zero. 31 32 * The presence of SVE is reported to userspace via HWCAP_SVE in the aux vector 33 AT_HWCAP entry. Presence of this flag implies the presence of the SVE 34 instructions and registers, and the Linux-specific system interfaces 35 described in this document. SVE is reported in /proc/cpuinfo as "sve". 36 37 * Support for the execution of SVE instructions in userspace can also be 38 detected by reading the CPU ID register ID_AA64PFR0_EL1 using an MRS 39 instruction, and checking that the value of the SVE field is nonzero. [3] 40 41 It does not guarantee the presence of the system interfaces described in the 42 following sections: software that needs to verify that those interfaces are 43 present must check for HWCAP_SVE instead. 44 45 * On hardware that supports the SVE2 extensions, HWCAP2_SVE2 will also 46 be reported in the AT_HWCAP2 aux vector entry. In addition to this, 47 optional extensions to SVE2 may be reported by the presence of: 48 49 HWCAP2_SVE2 50 HWCAP2_SVEAES 51 HWCAP2_SVEPMULL 52 HWCAP2_SVEBITPERM 53 HWCAP2_SVESHA3 54 HWCAP2_SVESM4 55 HWCAP2_SVE2P1 56 57 This list may be extended over time as the SVE architecture evolves. 58 59 These extensions are also reported via the CPU ID register ID_AA64ZFR0_EL1, 60 which userspace can read using an MRS instruction. See elf_hwcaps.txt and 61 cpu-feature-registers.txt for details. 62 63 * On hardware that supports the SME extensions, HWCAP2_SME will also be 64 reported in the AT_HWCAP2 aux vector entry. Among other things SME adds 65 streaming mode which provides a subset of the SVE feature set using a 66 separate SME vector length and the same Z/V registers. See sme.rst 67 for more details. 68 69 * Debuggers should restrict themselves to interacting with the target via the 70 NT_ARM_SVE regset. The recommended way of detecting support for this regset 71 is to connect to a target process first and then attempt a 72 ptrace(PTRACE_GETREGSET, pid, NT_ARM_SVE, &iov). Note that when SME is 73 present and streaming SVE mode is in use the FPSIMD subset of registers 74 will be read via NT_ARM_SVE and NT_ARM_SVE writes will exit streaming mode 75 in the target. 76 77 * Whenever SVE scalable register values (Zn, Pn, FFR) are exchanged in memory 78 between userspace and the kernel, the register value is encoded in memory in 79 an endianness-invariant layout, with bits [(8 * i + 7) : (8 * i)] encoded at 80 byte offset i from the start of the memory representation. This affects for 81 example the signal frame (struct sve_context) and ptrace interface 82 (struct user_sve_header) and associated data. 83 84 Beware that on big-endian systems this results in a different byte order than 85 for the FPSIMD V-registers, which are stored as single host-endian 128-bit 86 values, with bits [(127 - 8 * i) : (120 - 8 * i)] of the register encoded at 87 byte offset i. (struct fpsimd_context, struct user_fpsimd_state). 88 89 90 2. Vector length terminology 91 ----------------------------- 92 93 The size of an SVE vector (Z) register is referred to as the "vector length". 94 95 To avoid confusion about the units used to express vector length, the kernel 96 adopts the following conventions: 97 98 * Vector length (VL) = size of a Z-register in bytes 99 100 * Vector quadwords (VQ) = size of a Z-register in units of 128 bits 101 102 (So, VL = 16 * VQ.) 103 104 The VQ convention is used where the underlying granularity is important, such 105 as in data structure definitions. In most other situations, the VL convention 106 is used. This is consistent with the meaning of the "VL" pseudo-register in 107 the SVE instruction set architecture. 108 109 110 3. System call behaviour 111 ------------------------- 112 113 * On syscall, V0..V31 are preserved (as without SVE). Thus, bits [127:0] of 114 Z0..Z31 are preserved. All other bits of Z0..Z31, and all of P0..P15 and FFR 115 become zero on return from a syscall. 116 117 * The SVE registers are not used to pass arguments to or receive results from 118 any syscall. 119 120 * All other SVE state of a thread, including the currently configured vector 121 length, the state of the PR_SVE_VL_INHERIT flag, and the deferred vector 122 length (if any), is preserved across all syscalls, subject to the specific 123 exceptions for execve() described in section 6. 124 125 In particular, on return from a fork() or clone(), the parent and new child 126 process or thread share identical SVE configuration, matching that of the 127 parent before the call. 128 129 130 4. Signal handling 131 ------------------- 132 133 * A new signal frame record sve_context encodes the SVE registers on signal 134 delivery. [1] 135 136 * This record is supplementary to fpsimd_context. The FPSR and FPCR registers 137 are only present in fpsimd_context. For convenience, the content of V0..V31 138 is duplicated between sve_context and fpsimd_context. 139 140 * The record contains a flag field which includes a flag SVE_SIG_FLAG_SM which 141 if set indicates that the thread is in streaming mode and the vector length 142 and register data (if present) describe the streaming SVE data and vector 143 length. 144 145 * The signal frame record for SVE always contains basic metadata, in particular 146 the thread's vector length (in sve_context.vl). 147 148 * The SVE registers may or may not be included in the record, depending on 149 whether the registers are live for the thread. The registers are present if 150 and only if: 151 sve_context.head.size >= SVE_SIG_CONTEXT_SIZE(sve_vq_from_vl(sve_context.vl)). 152 153 * If the registers are present, the remainder of the record has a vl-dependent 154 size and layout. Macros SVE_SIG_* are defined [1] to facilitate access to 155 the members. 156 157 * Each scalable register (Zn, Pn, FFR) is stored in an endianness-invariant 158 layout, with bits [(8 * i + 7) : (8 * i)] stored at byte offset i from the 159 start of the register's representation in memory. 160 161 * If the SVE context is too big to fit in sigcontext.__reserved[], then extra 162 space is allocated on the stack, an extra_context record is written in 163 __reserved[] referencing this space. sve_context is then written in the 164 extra space. Refer to [1] for further details about this mechanism. 165 166 167 5. Signal return 168 ----------------- 169 170 When returning from a signal handler: 171 172 * If there is no sve_context record in the signal frame, or if the record is 173 present but contains no register data as described in the previous section, 174 then the SVE registers/bits become non-live and take unspecified values. 175 176 * If sve_context is present in the signal frame and contains full register 177 data, the SVE registers become live and are populated with the specified 178 data. However, for backward compatibility reasons, bits [127:0] of Z0..Z31 179 are always restored from the corresponding members of fpsimd_context.vregs[] 180 and not from sve_context. The remaining bits are restored from sve_context. 181 182 * Inclusion of fpsimd_context in the signal frame remains mandatory, 183 irrespective of whether sve_context is present or not. 184 185 * The vector length cannot be changed via signal return. If sve_context.vl in 186 the signal frame does not match the current vector length, the signal return 187 attempt is treated as illegal, resulting in a forced SIGSEGV. 188 189 * It is permitted to enter or leave streaming mode by setting or clearing 190 the SVE_SIG_FLAG_SM flag but applications should take care to ensure that 191 when doing so sve_context.vl and any register data are appropriate for the 192 vector length in the new mode. 193 194 195 6. prctl extensions 196 -------------------- 197 198 Some new prctl() calls are added to allow programs to manage the SVE vector 199 length: 200 201 prctl(PR_SVE_SET_VL, unsigned long arg) 202 203 Sets the vector length of the calling thread and related flags, where 204 arg == vl | flags. Other threads of the calling process are unaffected. 205 206 vl is the desired vector length, where sve_vl_valid(vl) must be true. 207 208 flags: 209 210 PR_SVE_VL_INHERIT 211 212 Inherit the current vector length across execve(). Otherwise, the 213 vector length is reset to the system default at execve(). (See 214 Section 9.) 215 216 PR_SVE_SET_VL_ONEXEC 217 218 Defer the requested vector length change until the next execve() 219 performed by this thread. 220 221 The effect is equivalent to implicit execution of the following 222 call immediately after the next execve() (if any) by the thread: 223 224 prctl(PR_SVE_SET_VL, arg & ~PR_SVE_SET_VL_ONEXEC) 225 226 This allows launching of a new program with a different vector 227 length, while avoiding runtime side effects in the caller. 228 229 230 Without PR_SVE_SET_VL_ONEXEC, the requested change takes effect 231 immediately. 232 233 234 Return value: a nonnegative on success, or a negative value on error: 235 EINVAL: SVE not supported, invalid vector length requested, or 236 invalid flags. 237 238 239 On success: 240 241 * Either the calling thread's vector length or the deferred vector length 242 to be applied at the next execve() by the thread (dependent on whether 243 PR_SVE_SET_VL_ONEXEC is present in arg), is set to the largest value 244 supported by the system that is less than or equal to vl. If vl == 245 SVE_VL_MAX, the value set will be the largest value supported by the 246 system. 247 248 * Any previously outstanding deferred vector length change in the calling 249 thread is cancelled. 250 251 * The returned value describes the resulting configuration, encoded as for 252 PR_SVE_GET_VL. The vector length reported in this value is the new 253 current vector length for this thread if PR_SVE_SET_VL_ONEXEC was not 254 present in arg; otherwise, the reported vector length is the deferred 255 vector length that will be applied at the next execve() by the calling 256 thread. 257 258 * Changing the vector length causes all of P0..P15, FFR and all bits of 259 Z0..Z31 except for Z0 bits [127:0] .. Z31 bits [127:0] to become 260 unspecified. Calling PR_SVE_SET_VL with vl equal to the thread's current 261 vector length, or calling PR_SVE_SET_VL with the PR_SVE_SET_VL_ONEXEC 262 flag, does not constitute a change to the vector length for this purpose. 263 264 265 prctl(PR_SVE_GET_VL) 266 267 Gets the vector length of the calling thread. 268 269 The following flag may be OR-ed into the result: 270 271 PR_SVE_VL_INHERIT 272 273 Vector length will be inherited across execve(). 274 275 There is no way to determine whether there is an outstanding deferred 276 vector length change (which would only normally be the case between a 277 fork() or vfork() and the corresponding execve() in typical use). 278 279 To extract the vector length from the result, bitwise and it with 280 PR_SVE_VL_LEN_MASK. 281 282 Return value: a nonnegative value on success, or a negative value on error: 283 EINVAL: SVE not supported. 284 285 286 7. ptrace extensions 287 --------------------- 288 289 * New regsets NT_ARM_SVE and NT_ARM_SSVE are defined for use with 290 PTRACE_GETREGSET and PTRACE_SETREGSET. NT_ARM_SSVE describes the 291 streaming mode SVE registers and NT_ARM_SVE describes the 292 non-streaming mode SVE registers. 293 294 In this description a register set is referred to as being "live" when 295 the target is in the appropriate streaming or non-streaming mode and is 296 using data beyond the subset shared with the FPSIMD Vn registers. 297 298 Refer to [2] for definitions. 299 300 The regset data starts with struct user_sve_header, containing: 301 302 size 303 304 Size of the complete regset, in bytes. 305 This depends on vl and possibly on other things in the future. 306 307 If a call to PTRACE_GETREGSET requests less data than the value of 308 size, the caller can allocate a larger buffer and retry in order to 309 read the complete regset. 310 311 max_size 312 313 Maximum size in bytes that the regset can grow to for the target 314 thread. The regset won't grow bigger than this even if the target 315 thread changes its vector length etc. 316 317 vl 318 319 Target thread's current vector length, in bytes. 320 321 max_vl 322 323 Maximum possible vector length for the target thread. 324 325 flags 326 327 at most one of 328 329 SVE_PT_REGS_FPSIMD 330 331 SVE registers are not live (GETREGSET) or are to be made 332 non-live (SETREGSET). 333 334 The payload is of type struct user_fpsimd_state, with the same 335 meaning as for NT_PRFPREG, starting at offset 336 SVE_PT_FPSIMD_OFFSET from the start of user_sve_header. 337 338 Extra data might be appended in the future: the size of the 339 payload should be obtained using SVE_PT_FPSIMD_SIZE(vq, flags). 340 341 vq should be obtained using sve_vq_from_vl(vl). 342 343 or 344 345 SVE_PT_REGS_SVE 346 347 SVE registers are live (GETREGSET) or are to be made live 348 (SETREGSET). 349 350 The payload contains the SVE register data, starting at offset 351 SVE_PT_SVE_OFFSET from the start of user_sve_header, and with 352 size SVE_PT_SVE_SIZE(vq, flags); 353 354 ... OR-ed with zero or more of the following flags, which have the same 355 meaning and behaviour as the corresponding PR_SET_VL_* flags: 356 357 SVE_PT_VL_INHERIT 358 359 SVE_PT_VL_ONEXEC (SETREGSET only). 360 361 If neither FPSIMD nor SVE flags are provided then no register 362 payload is available, this is only possible when SME is implemented. 363 364 365 * The effects of changing the vector length and/or flags are equivalent to 366 those documented for PR_SVE_SET_VL. 367 368 The caller must make a further GETREGSET call if it needs to know what VL is 369 actually set by SETREGSET, unless is it known in advance that the requested 370 VL is supported. 371 372 * In the SVE_PT_REGS_SVE case, the size and layout of the payload depends on 373 the header fields. The SVE_PT_SVE_*() macros are provided to facilitate 374 access to the members. 375 376 * In either case, for SETREGSET it is permissible to omit the payload, in which 377 case only the vector length and flags are changed (along with any 378 consequences of those changes). 379 380 * In systems supporting SME when in streaming mode a GETREGSET for 381 NT_REG_SVE will return only the user_sve_header with no register data, 382 similarly a GETREGSET for NT_REG_SSVE will not return any register data 383 when not in streaming mode. 384 385 * A GETREGSET for NT_ARM_SSVE will never return SVE_PT_REGS_FPSIMD. 386 387 * For SETREGSET, if an SVE_PT_REGS_SVE payload is present and the 388 requested VL is not supported, the effect will be the same as if the 389 payload were omitted, except that an EIO error is reported. No 390 attempt is made to translate the payload data to the correct layout 391 for the vector length actually set. The thread's FPSIMD state is 392 preserved, but the remaining bits of the SVE registers become 393 unspecified. It is up to the caller to translate the payload layout 394 for the actual VL and retry. 395 396 * Where SME is implemented it is not possible to GETREGSET the register 397 state for normal SVE when in streaming mode, nor the streaming mode 398 register state when in normal mode, regardless of the implementation defined 399 behaviour of the hardware for sharing data between the two modes. 400 401 * Any SETREGSET of NT_ARM_SVE will exit streaming mode if the target was in 402 streaming mode and any SETREGSET of NT_ARM_SSVE will enter streaming mode 403 if the target was not in streaming mode. 404 405 * The effect of writing a partial, incomplete payload is unspecified. 406 407 408 8. ELF coredump extensions 409 --------------------------- 410 411 * NT_ARM_SVE and NT_ARM_SSVE notes will be added to each coredump for 412 each thread of the dumped process. The contents will be equivalent to the 413 data that would have been read if a PTRACE_GETREGSET of the corresponding 414 type were executed for each thread when the coredump was generated. 415 416 9. System runtime configuration 417 -------------------------------- 418 419 * To mitigate the ABI impact of expansion of the signal frame, a policy 420 mechanism is provided for administrators, distro maintainers and developers 421 to set the default vector length for userspace processes: 422 423 /proc/sys/abi/sve_default_vector_length 424 425 Writing the text representation of an integer to this file sets the system 426 default vector length to the specified value rounded to a supported value 427 using the same rules as for setting vector length via PR_SVE_SET_VL. 428 429 The result can be determined by reopening the file and reading its 430 contents. 431 432 At boot, the default vector length is initially set to 64 or the maximum 433 supported vector length, whichever is smaller. This determines the initial 434 vector length of the init process (PID 1). 435 436 Reading this file returns the current system default vector length. 437 438 * At every execve() call, the new vector length of the new process is set to 439 the system default vector length, unless 440 441 * PR_SVE_VL_INHERIT (or equivalently SVE_PT_VL_INHERIT) is set for the 442 calling thread, or 443 444 * a deferred vector length change is pending, established via the 445 PR_SVE_SET_VL_ONEXEC flag (or SVE_PT_VL_ONEXEC). 446 447 * Modifying the system default vector length does not affect the vector length 448 of any existing process or thread that does not make an execve() call. 449 450 10. Perf extensions 451 -------------------------------- 452 453 * The arm64 specific DWARF standard [5] added the VG (Vector Granule) register 454 at index 46. This register is used for DWARF unwinding when variable length 455 SVE registers are pushed onto the stack. 456 457 * Its value is equivalent to the current SVE vector length (VL) in bits divided 458 by 64. 459 460 * The value is included in Perf samples in the regs[46] field if 461 PERF_SAMPLE_REGS_USER is set and the sample_regs_user mask has bit 46 set. 462 463 * The value is the current value at the time the sample was taken, and it can 464 change over time. 465 466 * If the system doesn't support SVE when perf_event_open is called with these 467 settings, the event will fail to open. 468 469 Appendix A. SVE programmer's model (informative) 470 ================================================= 471 472 This section provides a minimal description of the additions made by SVE to the 473 ARMv8-A programmer's model that are relevant to this document. 474 475 Note: This section is for information only and not intended to be complete or 476 to replace any architectural specification. 477 478 A.1. Registers 479 --------------- 480 481 In A64 state, SVE adds the following: 482 483 * 32 8VL-bit vector registers Z0..Z31 484 For each Zn, Zn bits [127:0] alias the ARMv8-A vector register Vn. 485 486 A register write using a Vn register name zeros all bits of the corresponding 487 Zn except for bits [127:0]. 488 489 * 16 VL-bit predicate registers P0..P15 490 491 * 1 VL-bit special-purpose predicate register FFR (the "first-fault register") 492 493 * a VL "pseudo-register" that determines the size of each vector register 494 495 The SVE instruction set architecture provides no way to write VL directly. 496 Instead, it can be modified only by EL1 and above, by writing appropriate 497 system registers. 498 499 * The value of VL can be configured at runtime by EL1 and above: 500 16 <= VL <= VLmax, where VL must be a multiple of 16. 501 502 * The maximum vector length is determined by the hardware: 503 16 <= VLmax <= 256. 504 505 (The SVE architecture specifies 256, but permits future architecture 506 revisions to raise this limit.) 507 508 * FPSR and FPCR are retained from ARMv8-A, and interact with SVE floating-point 509 operations in a similar way to the way in which they interact with ARMv8 510 floating-point operations:: 511 512 8VL-1 128 0 bit index 513 +---- //// -----------------+ 514 Z0 | : V0 | 515 : : 516 Z7 | : V7 | 517 Z8 | : * V8 | 518 : : : 519 Z15 | : *V15 | 520 Z16 | : V16 | 521 : : 522 Z31 | : V31 | 523 +---- //// -----------------+ 524 31 0 525 VL-1 0 +-------+ 526 +---- //// --+ FPSR | | 527 P0 | | +-------+ 528 : | | *FPCR | | 529 P15 | | +-------+ 530 +---- //// --+ 531 FFR | | +-----+ 532 +---- //// --+ VL | | 533 +-----+ 534 535 (*) callee-save: 536 This only applies to bits [63:0] of Z-/V-registers. 537 FPCR contains callee-save and caller-save bits. See [4] for details. 538 539 540 A.2. Procedure call standard 541 ----------------------------- 542 543 The ARMv8-A base procedure call standard is extended as follows with respect to 544 the additional SVE register state: 545 546 * All SVE register bits that are not shared with FP/SIMD are caller-save. 547 548 * Z8 bits [63:0] .. Z15 bits [63:0] are callee-save. 549 550 This follows from the way these bits are mapped to V8..V15, which are caller- 551 save in the base procedure call standard. 552 553 554 Appendix B. ARMv8-A FP/SIMD programmer's model 555 =============================================== 556 557 Note: This section is for information only and not intended to be complete or 558 to replace any architectural specification. 559 560 Refer to [4] for more information. 561 562 ARMv8-A defines the following floating-point / SIMD register state: 563 564 * 32 128-bit vector registers V0..V31 565 * 2 32-bit status/control registers FPSR, FPCR 566 567 :: 568 569 127 0 bit index 570 +---------------+ 571 V0 | | 572 : : : 573 V7 | | 574 * V8 | | 575 : : : : 576 *V15 | | 577 V16 | | 578 : : : 579 V31 | | 580 +---------------+ 581 582 31 0 583 +-------+ 584 FPSR | | 585 +-------+ 586 *FPCR | | 587 +-------+ 588 589 (*) callee-save: 590 This only applies to bits [63:0] of V-registers. 591 FPCR contains a mixture of callee-save and caller-save bits. 592 593 594 References 595 ========== 596 597 [1] arch/arm64/include/uapi/asm/sigcontext.h 598 AArch64 Linux signal ABI definitions 599 600 [2] arch/arm64/include/uapi/asm/ptrace.h 601 AArch64 Linux ptrace ABI definitions 602 603 [3] Documentation/arch/arm64/cpu-feature-registers.rst 604 605 [4] ARM IHI0055C 606 http://infocenter.arm.com/help/topic/com.arm.doc.ihi0055c/IHI0055C_beta_aapcs64.pdf 607 http://infocenter.arm.com/help/topic/com.arm.doc.subset.swdev.abi/index.html 608 Procedure Call Standard for the ARM 64-bit Architecture (AArch64) 609 610 [5] https://github.com/ARM-software/abi-aa/blob/main/aadwarf64/aadwarf64.rst
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.