1 Scalable Vector Extension support for AArch64 Linux
2 ===================================================
4 Author: Dave Martin <Dave.Martin@arm.com>
7 This document outlines briefly the interface provided to userspace by Linux in
8 order to support use of the ARM Scalable Vector Extension (SVE).
10 This is an outline of the most important features and issues only and not
11 intended to be exhaustive.
13 This document does not aim to describe the SVE architecture or programmer's
14 model. To aid understanding, a minimal description of relevant programmer's
15 model features for SVE is included in Appendix A.
21 * SVE registers Z0..Z31, P0..P15 and FFR and the current vector length VL, are
24 * The presence of SVE is reported to userspace via HWCAP_SVE in the aux vector
25 AT_HWCAP entry. Presence of this flag implies the presence of the SVE
26 instructions and registers, and the Linux-specific system interfaces
27 described in this document. SVE is reported in /proc/cpuinfo as "sve".
29 * Support for the execution of SVE instructions in userspace can also be
30 detected by reading the CPU ID register ID_AA64PFR0_EL1 using an MRS
31 instruction, and checking that the value of the SVE field is nonzero. [3]
33 It does not guarantee the presence of the system interfaces described in the
34 following sections: software that needs to verify that those interfaces are
35 present must check for HWCAP_SVE instead.
37 * On hardware that supports the SVE2 extensions, HWCAP2_SVE2 will also
38 be reported in the AT_HWCAP2 aux vector entry. In addition to this,
39 optional extensions to SVE2 may be reported by the presence of:
48 This list may be extended over time as the SVE architecture evolves.
50 These extensions are also reported via the CPU ID register ID_AA64ZFR0_EL1,
51 which userspace can read using an MRS instruction. See elf_hwcaps.txt and
52 cpu-feature-registers.txt for details.
54 * Debuggers should restrict themselves to interacting with the target via the
55 NT_ARM_SVE regset. The recommended way of detecting support for this regset
56 is to connect to a target process first and then attempt a
57 ptrace(PTRACE_GETREGSET, pid, NT_ARM_SVE, &iov).
59 * Whenever SVE scalable register values (Zn, Pn, FFR) are exchanged in memory
60 between userspace and the kernel, the register value is encoded in memory in
61 an endianness-invariant layout, with bits [(8 * i + 7) : (8 * i)] encoded at
62 byte offset i from the start of the memory representation. This affects for
63 example the signal frame (struct sve_context) and ptrace interface
64 (struct user_sve_header) and associated data.
66 Beware that on big-endian systems this results in a different byte order than
67 for the FPSIMD V-registers, which are stored as single host-endian 128-bit
68 values, with bits [(127 - 8 * i) : (120 - 8 * i)] of the register encoded at
69 byte offset i. (struct fpsimd_context, struct user_fpsimd_state).
72 2. Vector length terminology
73 -----------------------------
75 The size of an SVE vector (Z) register is referred to as the "vector length".
77 To avoid confusion about the units used to express vector length, the kernel
78 adopts the following conventions:
80 * Vector length (VL) = size of a Z-register in bytes
82 * Vector quadwords (VQ) = size of a Z-register in units of 128 bits
86 The VQ convention is used where the underlying granularity is important, such
87 as in data structure definitions. In most other situations, the VL convention
88 is used. This is consistent with the meaning of the "VL" pseudo-register in
89 the SVE instruction set architecture.
92 3. System call behaviour
93 -------------------------
95 * On syscall, V0..V31 are preserved (as without SVE). Thus, bits [127:0] of
96 Z0..Z31 are preserved. All other bits of Z0..Z31, and all of P0..P15 and FFR
97 become unspecified on return from a syscall.
99 * The SVE registers are not used to pass arguments to or receive results from
102 * In practice the affected registers/bits will be preserved or will be replaced
103 with zeros on return from a syscall, but userspace should not make
104 assumptions about this. The kernel behaviour may vary on a case-by-case
107 * All other SVE state of a thread, including the currently configured vector
108 length, the state of the PR_SVE_VL_INHERIT flag, and the deferred vector
109 length (if any), is preserved across all syscalls, subject to the specific
110 exceptions for execve() described in section 6.
112 In particular, on return from a fork() or clone(), the parent and new child
113 process or thread share identical SVE configuration, matching that of the
114 parent before the call.
120 * A new signal frame record sve_context encodes the SVE registers on signal
123 * This record is supplementary to fpsimd_context. The FPSR and FPCR registers
124 are only present in fpsimd_context. For convenience, the content of V0..V31
125 is duplicated between sve_context and fpsimd_context.
127 * The signal frame record for SVE always contains basic metadata, in particular
128 the thread's vector length (in sve_context.vl).
130 * The SVE registers may or may not be included in the record, depending on
131 whether the registers are live for the thread. The registers are present if
133 sve_context.head.size >= SVE_SIG_CONTEXT_SIZE(sve_vq_from_vl(sve_context.vl)).
135 * If the registers are present, the remainder of the record has a vl-dependent
136 size and layout. Macros SVE_SIG_* are defined [1] to facilitate access to
139 * Each scalable register (Zn, Pn, FFR) is stored in an endianness-invariant
140 layout, with bits [(8 * i + 7) : (8 * i)] stored at byte offset i from the
141 start of the register's representation in memory.
143 * If the SVE context is too big to fit in sigcontext.__reserved[], then extra
144 space is allocated on the stack, an extra_context record is written in
145 __reserved[] referencing this space. sve_context is then written in the
146 extra space. Refer to [1] for further details about this mechanism.
152 When returning from a signal handler:
154 * If there is no sve_context record in the signal frame, or if the record is
155 present but contains no register data as desribed in the previous section,
156 then the SVE registers/bits become non-live and take unspecified values.
158 * If sve_context is present in the signal frame and contains full register
159 data, the SVE registers become live and are populated with the specified
160 data. However, for backward compatibility reasons, bits [127:0] of Z0..Z31
161 are always restored from the corresponding members of fpsimd_context.vregs[]
162 and not from sve_context. The remaining bits are restored from sve_context.
164 * Inclusion of fpsimd_context in the signal frame remains mandatory,
165 irrespective of whether sve_context is present or not.
167 * The vector length cannot be changed via signal return. If sve_context.vl in
168 the signal frame does not match the current vector length, the signal return
169 attempt is treated as illegal, resulting in a forced SIGSEGV.
175 Some new prctl() calls are added to allow programs to manage the SVE vector
178 prctl(PR_SVE_SET_VL, unsigned long arg)
180 Sets the vector length of the calling thread and related flags, where
181 arg == vl | flags. Other threads of the calling process are unaffected.
183 vl is the desired vector length, where sve_vl_valid(vl) must be true.
187 PR_SVE_SET_VL_INHERIT
189 Inherit the current vector length across execve(). Otherwise, the
190 vector length is reset to the system default at execve(). (See
195 Defer the requested vector length change until the next execve()
196 performed by this thread.
198 The effect is equivalent to implicit exceution of the following
199 call immediately after the next execve() (if any) by the thread:
201 prctl(PR_SVE_SET_VL, arg & ~PR_SVE_SET_VL_ONEXEC)
203 This allows launching of a new program with a different vector
204 length, while avoiding runtime side effects in the caller.
207 Without PR_SVE_SET_VL_ONEXEC, the requested change takes effect
211 Return value: a nonnegative on success, or a negative value on error:
212 EINVAL: SVE not supported, invalid vector length requested, or
218 * Either the calling thread's vector length or the deferred vector length
219 to be applied at the next execve() by the thread (dependent on whether
220 PR_SVE_SET_VL_ONEXEC is present in arg), is set to the largest value
221 supported by the system that is less than or equal to vl. If vl ==
222 SVE_VL_MAX, the value set will be the largest value supported by the
225 * Any previously outstanding deferred vector length change in the calling
228 * The returned value describes the resulting configuration, encoded as for
229 PR_SVE_GET_VL. The vector length reported in this value is the new
230 current vector length for this thread if PR_SVE_SET_VL_ONEXEC was not
231 present in arg; otherwise, the reported vector length is the deferred
232 vector length that will be applied at the next execve() by the calling
235 * Changing the vector length causes all of P0..P15, FFR and all bits of
236 Z0..Z31 except for Z0 bits [127:0] .. Z31 bits [127:0] to become
237 unspecified. Calling PR_SVE_SET_VL with vl equal to the thread's current
238 vector length, or calling PR_SVE_SET_VL with the PR_SVE_SET_VL_ONEXEC
239 flag, does not constitute a change to the vector length for this purpose.
244 Gets the vector length of the calling thread.
246 The following flag may be OR-ed into the result:
248 PR_SVE_SET_VL_INHERIT
250 Vector length will be inherited across execve().
252 There is no way to determine whether there is an outstanding deferred
253 vector length change (which would only normally be the case between a
254 fork() or vfork() and the corresponding execve() in typical use).
256 To extract the vector length from the result, and it with
259 Return value: a nonnegative value on success, or a negative value on error:
260 EINVAL: SVE not supported.
264 ---------------------
266 * A new regset NT_ARM_SVE is defined for use with PTRACE_GETREGSET and
269 Refer to [2] for definitions.
271 The regset data starts with struct user_sve_header, containing:
275 Size of the complete regset, in bytes.
276 This depends on vl and possibly on other things in the future.
278 If a call to PTRACE_GETREGSET requests less data than the value of
279 size, the caller can allocate a larger buffer and retry in order to
280 read the complete regset.
284 Maximum size in bytes that the regset can grow to for the target
285 thread. The regset won't grow bigger than this even if the target
286 thread changes its vector length etc.
290 Target thread's current vector length, in bytes.
294 Maximum possible vector length for the target thread.
302 SVE registers are not live (GETREGSET) or are to be made
303 non-live (SETREGSET).
305 The payload is of type struct user_fpsimd_state, with the same
306 meaning as for NT_PRFPREG, starting at offset
307 SVE_PT_FPSIMD_OFFSET from the start of user_sve_header.
309 Extra data might be appended in the future: the size of the
310 payload should be obtained using SVE_PT_FPSIMD_SIZE(vq, flags).
312 vq should be obtained using sve_vq_from_vl(vl).
318 SVE registers are live (GETREGSET) or are to be made live
321 The payload contains the SVE register data, starting at offset
322 SVE_PT_SVE_OFFSET from the start of user_sve_header, and with
323 size SVE_PT_SVE_SIZE(vq, flags);
325 ... OR-ed with zero or more of the following flags, which have the same
326 meaning and behaviour as the corresponding PR_SET_VL_* flags:
330 SVE_PT_VL_ONEXEC (SETREGSET only).
332 * The effects of changing the vector length and/or flags are equivalent to
333 those documented for PR_SVE_SET_VL.
335 The caller must make a further GETREGSET call if it needs to know what VL is
336 actually set by SETREGSET, unless is it known in advance that the requested
339 * In the SVE_PT_REGS_SVE case, the size and layout of the payload depends on
340 the header fields. The SVE_PT_SVE_*() macros are provided to facilitate
341 access to the members.
343 * In either case, for SETREGSET it is permissible to omit the payload, in which
344 case only the vector length and flags are changed (along with any
345 consequences of those changes).
347 * For SETREGSET, if an SVE_PT_REGS_SVE payload is present and the
348 requested VL is not supported, the effect will be the same as if the
349 payload were omitted, except that an EIO error is reported. No
350 attempt is made to translate the payload data to the correct layout
351 for the vector length actually set. The thread's FPSIMD state is
352 preserved, but the remaining bits of the SVE registers become
353 unspecified. It is up to the caller to translate the payload layout
354 for the actual VL and retry.
356 * The effect of writing a partial, incomplete payload is unspecified.
359 8. ELF coredump extensions
360 ---------------------------
362 * A NT_ARM_SVE note will be added to each coredump for each thread of the
363 dumped process. The contents will be equivalent to the data that would have
364 been read if a PTRACE_GETREGSET of NT_ARM_SVE were executed for each thread
365 when the coredump was generated.
368 9. System runtime configuration
369 --------------------------------
371 * To mitigate the ABI impact of expansion of the signal frame, a policy
372 mechanism is provided for administrators, distro maintainers and developers
373 to set the default vector length for userspace processes:
375 /proc/sys/abi/sve_default_vector_length
377 Writing the text representation of an integer to this file sets the system
378 default vector length to the specified value, unless the value is greater
379 than the maximum vector length supported by the system in which case the
380 default vector length is set to that maximum.
382 The result can be determined by reopening the file and reading its
385 At boot, the default vector length is initially set to 64 or the maximum
386 supported vector length, whichever is smaller. This determines the initial
387 vector length of the init process (PID 1).
389 Reading this file returns the current system default vector length.
391 * At every execve() call, the new vector length of the new process is set to
392 the system default vector length, unless
394 * PR_SVE_SET_VL_INHERIT (or equivalently SVE_PT_VL_INHERIT) is set for the
397 * a deferred vector length change is pending, established via the
398 PR_SVE_SET_VL_ONEXEC flag (or SVE_PT_VL_ONEXEC).
400 * Modifying the system default vector length does not affect the vector length
401 of any existing process or thread that does not make an execve() call.
404 Appendix A. SVE programmer's model (informative)
405 =================================================
407 This section provides a minimal description of the additions made by SVE to the
408 ARMv8-A programmer's model that are relevant to this document.
410 Note: This section is for information only and not intended to be complete or
411 to replace any architectural specification.
416 In A64 state, SVE adds the following:
418 * 32 8VL-bit vector registers Z0..Z31
419 For each Zn, Zn bits [127:0] alias the ARMv8-A vector register Vn.
421 A register write using a Vn register name zeros all bits of the corresponding
422 Zn except for bits [127:0].
424 * 16 VL-bit predicate registers P0..P15
426 * 1 VL-bit special-purpose predicate register FFR (the "first-fault register")
428 * a VL "pseudo-register" that determines the size of each vector register
430 The SVE instruction set architecture provides no way to write VL directly.
431 Instead, it can be modified only by EL1 and above, by writing appropriate
434 * The value of VL can be configured at runtime by EL1 and above:
435 16 <= VL <= VLmax, where VL must be a multiple of 16.
437 * The maximum vector length is determined by the hardware:
440 (The SVE architecture specifies 256, but permits future architecture
441 revisions to raise this limit.)
443 * FPSR and FPCR are retained from ARMv8-A, and interact with SVE floating-point
444 operations in a similar way to the way in which they interact with ARMv8
445 floating-point operations.
447 8VL-1 128 0 bit index
448 +---- //// -----------------+
458 +---- //// -----------------+
461 +---- //// --+ FPSR | |
467 +---- //// --+ VL | |
471 This only applies to bits [63:0] of Z-/V-registers.
472 FPCR contains callee-save and caller-save bits. See [4] for details.
475 A.2. Procedure call standard
476 -----------------------------
478 The ARMv8-A base procedure call standard is extended as follows with respect to
479 the additional SVE register state:
481 * All SVE register bits that are not shared with FP/SIMD are caller-save.
483 * Z8 bits [63:0] .. Z15 bits [63:0] are callee-save.
485 This follows from the way these bits are mapped to V8..V15, which are caller-
486 save in the base procedure call standard.
489 Appendix B. ARMv8-A FP/SIMD programmer's model
490 ===============================================
492 Note: This section is for information only and not intended to be complete or
493 to replace any architectural specification.
495 Refer to [4] for for more information.
497 ARMv8-A defines the following floating-point / SIMD register state:
499 * 32 128-bit vector registers V0..V31
500 * 2 32-bit status/control registers FPSR, FPCR
523 This only applies to bits [63:0] of V-registers.
524 FPCR contains a mixture of callee-save and caller-save bits.
530 [1] arch/arm64/include/uapi/asm/sigcontext.h
531 AArch64 Linux signal ABI definitions
533 [2] arch/arm64/include/uapi/asm/ptrace.h
534 AArch64 Linux ptrace ABI definitions
536 [3] Documentation/arm64/cpu-feature-registers.txt
539 http://infocenter.arm.com/help/topic/com.arm.doc.ihi0055c/IHI0055C_beta_aapcs64.pdf
540 http://infocenter.arm.com/help/topic/com.arm.doc.subset.swdev.abi/index.html
541 Procedure Call Standard for the ARM 64-bit Architecture (AArch64)