1 ===================================================
2 Scalable Vector Extension support for AArch64 Linux
3 ===================================================
5 Author: Dave Martin <Dave.Martin@arm.com>
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
14 This is an outline of the most important features and issues only and not
15 intended to be exhaustive.
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.
25 * SVE registers Z0..Z31, P0..P15 and FFR and the current vector length VL, are
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.
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".
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]
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.
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:
56 This list may be extended over time as the SVE architecture evolves.
58 These extensions are also reported via the CPU ID register ID_AA64ZFR0_EL1,
59 which userspace can read using an MRS instruction. See elf_hwcaps.txt and
60 cpu-feature-registers.txt for details.
62 * On hardware that supports the SME extensions, HWCAP2_SME will also be
63 reported in the AT_HWCAP2 aux vector entry. Among other things SME adds
64 streaming mode which provides a subset of the SVE feature set using a
65 separate SME vector length and the same Z/V registers. See sme.rst
68 * Debuggers should restrict themselves to interacting with the target via the
69 NT_ARM_SVE regset. The recommended way of detecting support for this regset
70 is to connect to a target process first and then attempt a
71 ptrace(PTRACE_GETREGSET, pid, NT_ARM_SVE, &iov). Note that when SME is
72 present and streaming SVE mode is in use the FPSIMD subset of registers
73 will be read via NT_ARM_SVE and NT_ARM_SVE writes will exit streaming mode
76 * Whenever SVE scalable register values (Zn, Pn, FFR) are exchanged in memory
77 between userspace and the kernel, the register value is encoded in memory in
78 an endianness-invariant layout, with bits [(8 * i + 7) : (8 * i)] encoded at
79 byte offset i from the start of the memory representation. This affects for
80 example the signal frame (struct sve_context) and ptrace interface
81 (struct user_sve_header) and associated data.
83 Beware that on big-endian systems this results in a different byte order than
84 for the FPSIMD V-registers, which are stored as single host-endian 128-bit
85 values, with bits [(127 - 8 * i) : (120 - 8 * i)] of the register encoded at
86 byte offset i. (struct fpsimd_context, struct user_fpsimd_state).
89 2. Vector length terminology
90 -----------------------------
92 The size of an SVE vector (Z) register is referred to as the "vector length".
94 To avoid confusion about the units used to express vector length, the kernel
95 adopts the following conventions:
97 * Vector length (VL) = size of a Z-register in bytes
99 * Vector quadwords (VQ) = size of a Z-register in units of 128 bits
103 The VQ convention is used where the underlying granularity is important, such
104 as in data structure definitions. In most other situations, the VL convention
105 is used. This is consistent with the meaning of the "VL" pseudo-register in
106 the SVE instruction set architecture.
109 3. System call behaviour
110 -------------------------
112 * On syscall, V0..V31 are preserved (as without SVE). Thus, bits [127:0] of
113 Z0..Z31 are preserved. All other bits of Z0..Z31, and all of P0..P15 and FFR
114 become zero on return from a syscall.
116 * The SVE registers are not used to pass arguments to or receive results from
119 * In practice the affected registers/bits will be preserved or will be replaced
120 with zeros on return from a syscall, but userspace should not make
121 assumptions about this. The kernel behaviour may vary on a case-by-case
124 * All other SVE state of a thread, including the currently configured vector
125 length, the state of the PR_SVE_VL_INHERIT flag, and the deferred vector
126 length (if any), is preserved across all syscalls, subject to the specific
127 exceptions for execve() described in section 6.
129 In particular, on return from a fork() or clone(), the parent and new child
130 process or thread share identical SVE configuration, matching that of the
131 parent before the call.
137 * A new signal frame record sve_context encodes the SVE registers on signal
140 * This record is supplementary to fpsimd_context. The FPSR and FPCR registers
141 are only present in fpsimd_context. For convenience, the content of V0..V31
142 is duplicated between sve_context and fpsimd_context.
144 * The record contains a flag field which includes a flag SVE_SIG_FLAG_SM which
145 if set indicates that the thread is in streaming mode and the vector length
146 and register data (if present) describe the streaming SVE data and vector
149 * The signal frame record for SVE always contains basic metadata, in particular
150 the thread's vector length (in sve_context.vl).
152 * The SVE registers may or may not be included in the record, depending on
153 whether the registers are live for the thread. The registers are present if
155 sve_context.head.size >= SVE_SIG_CONTEXT_SIZE(sve_vq_from_vl(sve_context.vl)).
157 * If the registers are present, the remainder of the record has a vl-dependent
158 size and layout. Macros SVE_SIG_* are defined [1] to facilitate access to
161 * Each scalable register (Zn, Pn, FFR) is stored in an endianness-invariant
162 layout, with bits [(8 * i + 7) : (8 * i)] stored at byte offset i from the
163 start of the register's representation in memory.
165 * If the SVE context is too big to fit in sigcontext.__reserved[], then extra
166 space is allocated on the stack, an extra_context record is written in
167 __reserved[] referencing this space. sve_context is then written in the
168 extra space. Refer to [1] for further details about this mechanism.
174 When returning from a signal handler:
176 * If there is no sve_context record in the signal frame, or if the record is
177 present but contains no register data as desribed in the previous section,
178 then the SVE registers/bits become non-live and take unspecified values.
180 * If sve_context is present in the signal frame and contains full register
181 data, the SVE registers become live and are populated with the specified
182 data. However, for backward compatibility reasons, bits [127:0] of Z0..Z31
183 are always restored from the corresponding members of fpsimd_context.vregs[]
184 and not from sve_context. The remaining bits are restored from sve_context.
186 * Inclusion of fpsimd_context in the signal frame remains mandatory,
187 irrespective of whether sve_context is present or not.
189 * The vector length cannot be changed via signal return. If sve_context.vl in
190 the signal frame does not match the current vector length, the signal return
191 attempt is treated as illegal, resulting in a forced SIGSEGV.
193 * It is permitted to enter or leave streaming mode by setting or clearing
194 the SVE_SIG_FLAG_SM flag but applications should take care to ensure that
195 when doing so sve_context.vl and any register data are appropriate for the
196 vector length in the new mode.
202 Some new prctl() calls are added to allow programs to manage the SVE vector
205 prctl(PR_SVE_SET_VL, unsigned long arg)
207 Sets the vector length of the calling thread and related flags, where
208 arg == vl | flags. Other threads of the calling process are unaffected.
210 vl is the desired vector length, where sve_vl_valid(vl) must be true.
216 Inherit the current vector length across execve(). Otherwise, the
217 vector length is reset to the system default at execve(). (See
222 Defer the requested vector length change until the next execve()
223 performed by this thread.
225 The effect is equivalent to implicit exceution of the following
226 call immediately after the next execve() (if any) by the thread:
228 prctl(PR_SVE_SET_VL, arg & ~PR_SVE_SET_VL_ONEXEC)
230 This allows launching of a new program with a different vector
231 length, while avoiding runtime side effects in the caller.
234 Without PR_SVE_SET_VL_ONEXEC, the requested change takes effect
238 Return value: a nonnegative on success, or a negative value on error:
239 EINVAL: SVE not supported, invalid vector length requested, or
245 * Either the calling thread's vector length or the deferred vector length
246 to be applied at the next execve() by the thread (dependent on whether
247 PR_SVE_SET_VL_ONEXEC is present in arg), is set to the largest value
248 supported by the system that is less than or equal to vl. If vl ==
249 SVE_VL_MAX, the value set will be the largest value supported by the
252 * Any previously outstanding deferred vector length change in the calling
255 * The returned value describes the resulting configuration, encoded as for
256 PR_SVE_GET_VL. The vector length reported in this value is the new
257 current vector length for this thread if PR_SVE_SET_VL_ONEXEC was not
258 present in arg; otherwise, the reported vector length is the deferred
259 vector length that will be applied at the next execve() by the calling
262 * Changing the vector length causes all of P0..P15, FFR and all bits of
263 Z0..Z31 except for Z0 bits [127:0] .. Z31 bits [127:0] to become
264 unspecified. Calling PR_SVE_SET_VL with vl equal to the thread's current
265 vector length, or calling PR_SVE_SET_VL with the PR_SVE_SET_VL_ONEXEC
266 flag, does not constitute a change to the vector length for this purpose.
271 Gets the vector length of the calling thread.
273 The following flag may be OR-ed into the result:
277 Vector length will be inherited across execve().
279 There is no way to determine whether there is an outstanding deferred
280 vector length change (which would only normally be the case between a
281 fork() or vfork() and the corresponding execve() in typical use).
283 To extract the vector length from the result, bitwise and it with
286 Return value: a nonnegative value on success, or a negative value on error:
287 EINVAL: SVE not supported.
291 ---------------------
293 * New regsets NT_ARM_SVE and NT_ARM_SSVE are defined for use with
294 PTRACE_GETREGSET and PTRACE_SETREGSET. NT_ARM_SSVE describes the
295 streaming mode SVE registers and NT_ARM_SVE describes the
296 non-streaming mode SVE registers.
298 In this description a register set is referred to as being "live" when
299 the target is in the appropriate streaming or non-streaming mode and is
300 using data beyond the subset shared with the FPSIMD Vn registers.
302 Refer to [2] for definitions.
304 The regset data starts with struct user_sve_header, containing:
308 Size of the complete regset, in bytes.
309 This depends on vl and possibly on other things in the future.
311 If a call to PTRACE_GETREGSET requests less data than the value of
312 size, the caller can allocate a larger buffer and retry in order to
313 read the complete regset.
317 Maximum size in bytes that the regset can grow to for the target
318 thread. The regset won't grow bigger than this even if the target
319 thread changes its vector length etc.
323 Target thread's current vector length, in bytes.
327 Maximum possible vector length for the target thread.
335 SVE registers are not live (GETREGSET) or are to be made
336 non-live (SETREGSET).
338 The payload is of type struct user_fpsimd_state, with the same
339 meaning as for NT_PRFPREG, starting at offset
340 SVE_PT_FPSIMD_OFFSET from the start of user_sve_header.
342 Extra data might be appended in the future: the size of the
343 payload should be obtained using SVE_PT_FPSIMD_SIZE(vq, flags).
345 vq should be obtained using sve_vq_from_vl(vl).
351 SVE registers are live (GETREGSET) or are to be made live
354 The payload contains the SVE register data, starting at offset
355 SVE_PT_SVE_OFFSET from the start of user_sve_header, and with
356 size SVE_PT_SVE_SIZE(vq, flags);
358 ... OR-ed with zero or more of the following flags, which have the same
359 meaning and behaviour as the corresponding PR_SET_VL_* flags:
363 SVE_PT_VL_ONEXEC (SETREGSET only).
365 If neither FPSIMD nor SVE flags are provided then no register
366 payload is available, this is only possible when SME is implemented.
369 * The effects of changing the vector length and/or flags are equivalent to
370 those documented for PR_SVE_SET_VL.
372 The caller must make a further GETREGSET call if it needs to know what VL is
373 actually set by SETREGSET, unless is it known in advance that the requested
376 * In the SVE_PT_REGS_SVE case, the size and layout of the payload depends on
377 the header fields. The SVE_PT_SVE_*() macros are provided to facilitate
378 access to the members.
380 * In either case, for SETREGSET it is permissible to omit the payload, in which
381 case only the vector length and flags are changed (along with any
382 consequences of those changes).
384 * In systems supporting SME when in streaming mode a GETREGSET for
385 NT_REG_SVE will return only the user_sve_header with no register data,
386 similarly a GETREGSET for NT_REG_SSVE will not return any register data
387 when not in streaming mode.
389 * A GETREGSET for NT_ARM_SSVE will never return SVE_PT_REGS_FPSIMD.
391 * For SETREGSET, if an SVE_PT_REGS_SVE payload is present and the
392 requested VL is not supported, the effect will be the same as if the
393 payload were omitted, except that an EIO error is reported. No
394 attempt is made to translate the payload data to the correct layout
395 for the vector length actually set. The thread's FPSIMD state is
396 preserved, but the remaining bits of the SVE registers become
397 unspecified. It is up to the caller to translate the payload layout
398 for the actual VL and retry.
400 * Where SME is implemented it is not possible to GETREGSET the register
401 state for normal SVE when in streaming mode, nor the streaming mode
402 register state when in normal mode, regardless of the implementation defined
403 behaviour of the hardware for sharing data between the two modes.
405 * Any SETREGSET of NT_ARM_SVE will exit streaming mode if the target was in
406 streaming mode and any SETREGSET of NT_ARM_SSVE will enter streaming mode
407 if the target was not in streaming mode.
409 * The effect of writing a partial, incomplete payload is unspecified.
412 8. ELF coredump extensions
413 ---------------------------
415 * NT_ARM_SVE and NT_ARM_SSVE notes will be added to each coredump for
416 each thread of the dumped process. The contents will be equivalent to the
417 data that would have been read if a PTRACE_GETREGSET of the corresponding
418 type were executed for each thread when the coredump was generated.
420 9. System runtime configuration
421 --------------------------------
423 * To mitigate the ABI impact of expansion of the signal frame, a policy
424 mechanism is provided for administrators, distro maintainers and developers
425 to set the default vector length for userspace processes:
427 /proc/sys/abi/sve_default_vector_length
429 Writing the text representation of an integer to this file sets the system
430 default vector length to the specified value, unless the value is greater
431 than the maximum vector length supported by the system in which case the
432 default vector length is set to that maximum.
434 The result can be determined by reopening the file and reading its
437 At boot, the default vector length is initially set to 64 or the maximum
438 supported vector length, whichever is smaller. This determines the initial
439 vector length of the init process (PID 1).
441 Reading this file returns the current system default vector length.
443 * At every execve() call, the new vector length of the new process is set to
444 the system default vector length, unless
446 * PR_SVE_VL_INHERIT (or equivalently SVE_PT_VL_INHERIT) is set for the
449 * a deferred vector length change is pending, established via the
450 PR_SVE_SET_VL_ONEXEC flag (or SVE_PT_VL_ONEXEC).
452 * Modifying the system default vector length does not affect the vector length
453 of any existing process or thread that does not make an execve() call.
456 --------------------------------
458 * The arm64 specific DWARF standard [5] added the VG (Vector Granule) register
459 at index 46. This register is used for DWARF unwinding when variable length
460 SVE registers are pushed onto the stack.
462 * Its value is equivalent to the current SVE vector length (VL) in bits divided
465 * The value is included in Perf samples in the regs[46] field if
466 PERF_SAMPLE_REGS_USER is set and the sample_regs_user mask has bit 46 set.
468 * The value is the current value at the time the sample was taken, and it can
471 * If the system doesn't support SVE when perf_event_open is called with these
472 settings, the event will fail to open.
474 Appendix A. SVE programmer's model (informative)
475 =================================================
477 This section provides a minimal description of the additions made by SVE to the
478 ARMv8-A programmer's model that are relevant to this document.
480 Note: This section is for information only and not intended to be complete or
481 to replace any architectural specification.
486 In A64 state, SVE adds the following:
488 * 32 8VL-bit vector registers Z0..Z31
489 For each Zn, Zn bits [127:0] alias the ARMv8-A vector register Vn.
491 A register write using a Vn register name zeros all bits of the corresponding
492 Zn except for bits [127:0].
494 * 16 VL-bit predicate registers P0..P15
496 * 1 VL-bit special-purpose predicate register FFR (the "first-fault register")
498 * a VL "pseudo-register" that determines the size of each vector register
500 The SVE instruction set architecture provides no way to write VL directly.
501 Instead, it can be modified only by EL1 and above, by writing appropriate
504 * The value of VL can be configured at runtime by EL1 and above:
505 16 <= VL <= VLmax, where VL must be a multiple of 16.
507 * The maximum vector length is determined by the hardware:
510 (The SVE architecture specifies 256, but permits future architecture
511 revisions to raise this limit.)
513 * FPSR and FPCR are retained from ARMv8-A, and interact with SVE floating-point
514 operations in a similar way to the way in which they interact with ARMv8
515 floating-point operations::
517 8VL-1 128 0 bit index
518 +---- //// -----------------+
528 +---- //// -----------------+
531 +---- //// --+ FPSR | |
537 +---- //// --+ VL | |
541 This only applies to bits [63:0] of Z-/V-registers.
542 FPCR contains callee-save and caller-save bits. See [4] for details.
545 A.2. Procedure call standard
546 -----------------------------
548 The ARMv8-A base procedure call standard is extended as follows with respect to
549 the additional SVE register state:
551 * All SVE register bits that are not shared with FP/SIMD are caller-save.
553 * Z8 bits [63:0] .. Z15 bits [63:0] are callee-save.
555 This follows from the way these bits are mapped to V8..V15, which are caller-
556 save in the base procedure call standard.
559 Appendix B. ARMv8-A FP/SIMD programmer's model
560 ===============================================
562 Note: This section is for information only and not intended to be complete or
563 to replace any architectural specification.
565 Refer to [4] for more information.
567 ARMv8-A defines the following floating-point / SIMD register state:
569 * 32 128-bit vector registers V0..V31
570 * 2 32-bit status/control registers FPSR, FPCR
595 This only applies to bits [63:0] of V-registers.
596 FPCR contains a mixture of callee-save and caller-save bits.
602 [1] arch/arm64/include/uapi/asm/sigcontext.h
603 AArch64 Linux signal ABI definitions
605 [2] arch/arm64/include/uapi/asm/ptrace.h
606 AArch64 Linux ptrace ABI definitions
608 [3] Documentation/arm64/cpu-feature-registers.rst
611 http://infocenter.arm.com/help/topic/com.arm.doc.ihi0055c/IHI0055C_beta_aapcs64.pdf
612 http://infocenter.arm.com/help/topic/com.arm.doc.subset.swdev.abi/index.html
613 Procedure Call Standard for the ARM 64-bit Architecture (AArch64)
615 [5] https://github.com/ARM-software/abi-aa/blob/main/aadwarf64/aadwarf64.rst