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 * Debuggers should restrict themselves to interacting with the target via the
38 NT_ARM_SVE regset. The recommended way of detecting support for this regset
39 is to connect to a target process first and then attempt a
40 ptrace(PTRACE_GETREGSET, pid, NT_ARM_SVE, &iov).
43 2. Vector length terminology
44 -----------------------------
46 The size of an SVE vector (Z) register is referred to as the "vector length".
48 To avoid confusion about the units used to express vector length, the kernel
49 adopts the following conventions:
51 * Vector length (VL) = size of a Z-register in bytes
53 * Vector quadwords (VQ) = size of a Z-register in units of 128 bits
57 The VQ convention is used where the underlying granularity is important, such
58 as in data structure definitions. In most other situations, the VL convention
59 is used. This is consistent with the meaning of the "VL" pseudo-register in
60 the SVE instruction set architecture.
63 3. System call behaviour
64 -------------------------
66 * On syscall, V0..V31 are preserved (as without SVE). Thus, bits [127:0] of
67 Z0..Z31 are preserved. All other bits of Z0..Z31, and all of P0..P15 and FFR
68 become unspecified on return from a syscall.
70 * The SVE registers are not used to pass arguments to or receive results from
73 * In practice the affected registers/bits will be preserved or will be replaced
74 with zeros on return from a syscall, but userspace should not make
75 assumptions about this. The kernel behaviour may vary on a case-by-case
78 * All other SVE state of a thread, including the currently configured vector
79 length, the state of the PR_SVE_VL_INHERIT flag, and the deferred vector
80 length (if any), is preserved across all syscalls, subject to the specific
81 exceptions for execve() described in section 6.
83 In particular, on return from a fork() or clone(), the parent and new child
84 process or thread share identical SVE configuration, matching that of the
85 parent before the call.
91 * A new signal frame record sve_context encodes the SVE registers on signal
94 * This record is supplementary to fpsimd_context. The FPSR and FPCR registers
95 are only present in fpsimd_context. For convenience, the content of V0..V31
96 is duplicated between sve_context and fpsimd_context.
98 * The signal frame record for SVE always contains basic metadata, in particular
99 the thread's vector length (in sve_context.vl).
101 * The SVE registers may or may not be included in the record, depending on
102 whether the registers are live for the thread. The registers are present if
104 sve_context.head.size >= SVE_SIG_CONTEXT_SIZE(sve_vq_from_vl(sve_context.vl)).
106 * If the registers are present, the remainder of the record has a vl-dependent
107 size and layout. Macros SVE_SIG_* are defined [1] to facilitate access to
110 * If the SVE context is too big to fit in sigcontext.__reserved[], then extra
111 space is allocated on the stack, an extra_context record is written in
112 __reserved[] referencing this space. sve_context is then written in the
113 extra space. Refer to [1] for further details about this mechanism.
119 When returning from a signal handler:
121 * If there is no sve_context record in the signal frame, or if the record is
122 present but contains no register data as desribed in the previous section,
123 then the SVE registers/bits become non-live and take unspecified values.
125 * If sve_context is present in the signal frame and contains full register
126 data, the SVE registers become live and are populated with the specified
127 data. However, for backward compatibility reasons, bits [127:0] of Z0..Z31
128 are always restored from the corresponding members of fpsimd_context.vregs[]
129 and not from sve_context. The remaining bits are restored from sve_context.
131 * Inclusion of fpsimd_context in the signal frame remains mandatory,
132 irrespective of whether sve_context is present or not.
134 * The vector length cannot be changed via signal return. If sve_context.vl in
135 the signal frame does not match the current vector length, the signal return
136 attempt is treated as illegal, resulting in a forced SIGSEGV.
142 Some new prctl() calls are added to allow programs to manage the SVE vector
145 prctl(PR_SVE_SET_VL, unsigned long arg)
147 Sets the vector length of the calling thread and related flags, where
148 arg == vl | flags. Other threads of the calling process are unaffected.
150 vl is the desired vector length, where sve_vl_valid(vl) must be true.
154 PR_SVE_SET_VL_INHERIT
156 Inherit the current vector length across execve(). Otherwise, the
157 vector length is reset to the system default at execve(). (See
162 Defer the requested vector length change until the next execve()
163 performed by this thread.
165 The effect is equivalent to implicit exceution of the following
166 call immediately after the next execve() (if any) by the thread:
168 prctl(PR_SVE_SET_VL, arg & ~PR_SVE_SET_VL_ONEXEC)
170 This allows launching of a new program with a different vector
171 length, while avoiding runtime side effects in the caller.
174 Without PR_SVE_SET_VL_ONEXEC, the requested change takes effect
178 Return value: a nonnegative on success, or a negative value on error:
179 EINVAL: SVE not supported, invalid vector length requested, or
185 * Either the calling thread's vector length or the deferred vector length
186 to be applied at the next execve() by the thread (dependent on whether
187 PR_SVE_SET_VL_ONEXEC is present in arg), is set to the largest value
188 supported by the system that is less than or equal to vl. If vl ==
189 SVE_VL_MAX, the value set will be the largest value supported by the
192 * Any previously outstanding deferred vector length change in the calling
195 * The returned value describes the resulting configuration, encoded as for
196 PR_SVE_GET_VL. The vector length reported in this value is the new
197 current vector length for this thread if PR_SVE_SET_VL_ONEXEC was not
198 present in arg; otherwise, the reported vector length is the deferred
199 vector length that will be applied at the next execve() by the calling
202 * Changing the vector length causes all of P0..P15, FFR and all bits of
203 Z0..Z31 except for Z0 bits [127:0] .. Z31 bits [127:0] to become
204 unspecified. Calling PR_SVE_SET_VL with vl equal to the thread's current
205 vector length, or calling PR_SVE_SET_VL with the PR_SVE_SET_VL_ONEXEC
206 flag, does not constitute a change to the vector length for this purpose.
211 Gets the vector length of the calling thread.
213 The following flag may be OR-ed into the result:
215 PR_SVE_SET_VL_INHERIT
217 Vector length will be inherited across execve().
219 There is no way to determine whether there is an outstanding deferred
220 vector length change (which would only normally be the case between a
221 fork() or vfork() and the corresponding execve() in typical use).
223 To extract the vector length from the result, and it with
226 Return value: a nonnegative value on success, or a negative value on error:
227 EINVAL: SVE not supported.
231 ---------------------
233 * A new regset NT_ARM_SVE is defined for use with PTRACE_GETREGSET and
236 Refer to [2] for definitions.
238 The regset data starts with struct user_sve_header, containing:
242 Size of the complete regset, in bytes.
243 This depends on vl and possibly on other things in the future.
245 If a call to PTRACE_GETREGSET requests less data than the value of
246 size, the caller can allocate a larger buffer and retry in order to
247 read the complete regset.
251 Maximum size in bytes that the regset can grow to for the target
252 thread. The regset won't grow bigger than this even if the target
253 thread changes its vector length etc.
257 Target thread's current vector length, in bytes.
261 Maximum possible vector length for the target thread.
269 SVE registers are not live (GETREGSET) or are to be made
270 non-live (SETREGSET).
272 The payload is of type struct user_fpsimd_state, with the same
273 meaning as for NT_PRFPREG, starting at offset
274 SVE_PT_FPSIMD_OFFSET from the start of user_sve_header.
276 Extra data might be appended in the future: the size of the
277 payload should be obtained using SVE_PT_FPSIMD_SIZE(vq, flags).
279 vq should be obtained using sve_vq_from_vl(vl).
285 SVE registers are live (GETREGSET) or are to be made live
288 The payload contains the SVE register data, starting at offset
289 SVE_PT_SVE_OFFSET from the start of user_sve_header, and with
290 size SVE_PT_SVE_SIZE(vq, flags);
292 ... OR-ed with zero or more of the following flags, which have the same
293 meaning and behaviour as the corresponding PR_SET_VL_* flags:
297 SVE_PT_VL_ONEXEC (SETREGSET only).
299 * The effects of changing the vector length and/or flags are equivalent to
300 those documented for PR_SVE_SET_VL.
302 The caller must make a further GETREGSET call if it needs to know what VL is
303 actually set by SETREGSET, unless is it known in advance that the requested
306 * In the SVE_PT_REGS_SVE case, the size and layout of the payload depends on
307 the header fields. The SVE_PT_SVE_*() macros are provided to facilitate
308 access to the members.
310 * In either case, for SETREGSET it is permissible to omit the payload, in which
311 case only the vector length and flags are changed (along with any
312 consequences of those changes).
314 * For SETREGSET, if an SVE_PT_REGS_SVE payload is present and the
315 requested VL is not supported, the effect will be the same as if the
316 payload were omitted, except that an EIO error is reported. No
317 attempt is made to translate the payload data to the correct layout
318 for the vector length actually set. The thread's FPSIMD state is
319 preserved, but the remaining bits of the SVE registers become
320 unspecified. It is up to the caller to translate the payload layout
321 for the actual VL and retry.
323 * The effect of writing a partial, incomplete payload is unspecified.
326 8. ELF coredump extensions
327 ---------------------------
329 * A NT_ARM_SVE note will be added to each coredump for each thread of the
330 dumped process. The contents will be equivalent to the data that would have
331 been read if a PTRACE_GETREGSET of NT_ARM_SVE were executed for each thread
332 when the coredump was generated.
335 9. System runtime configuration
336 --------------------------------
338 * To mitigate the ABI impact of expansion of the signal frame, a policy
339 mechanism is provided for administrators, distro maintainers and developers
340 to set the default vector length for userspace processes:
342 /proc/sys/abi/sve_default_vector_length
344 Writing the text representation of an integer to this file sets the system
345 default vector length to the specified value, unless the value is greater
346 than the maximum vector length supported by the system in which case the
347 default vector length is set to that maximum.
349 The result can be determined by reopening the file and reading its
352 At boot, the default vector length is initially set to 64 or the maximum
353 supported vector length, whichever is smaller. This determines the initial
354 vector length of the init process (PID 1).
356 Reading this file returns the current system default vector length.
358 * At every execve() call, the new vector length of the new process is set to
359 the system default vector length, unless
361 * PR_SVE_SET_VL_INHERIT (or equivalently SVE_PT_VL_INHERIT) is set for the
364 * a deferred vector length change is pending, established via the
365 PR_SVE_SET_VL_ONEXEC flag (or SVE_PT_VL_ONEXEC).
367 * Modifying the system default vector length does not affect the vector length
368 of any existing process or thread that does not make an execve() call.
371 Appendix A. SVE programmer's model (informative)
372 =================================================
374 This section provides a minimal description of the additions made by SVE to the
375 ARMv8-A programmer's model that are relevant to this document.
377 Note: This section is for information only and not intended to be complete or
378 to replace any architectural specification.
383 In A64 state, SVE adds the following:
385 * 32 8VL-bit vector registers Z0..Z31
386 For each Zn, Zn bits [127:0] alias the ARMv8-A vector register Vn.
388 A register write using a Vn register name zeros all bits of the corresponding
389 Zn except for bits [127:0].
391 * 16 VL-bit predicate registers P0..P15
393 * 1 VL-bit special-purpose predicate register FFR (the "first-fault register")
395 * a VL "pseudo-register" that determines the size of each vector register
397 The SVE instruction set architecture provides no way to write VL directly.
398 Instead, it can be modified only by EL1 and above, by writing appropriate
401 * The value of VL can be configured at runtime by EL1 and above:
402 16 <= VL <= VLmax, where VL must be a multiple of 16.
404 * The maximum vector length is determined by the hardware:
407 (The SVE architecture specifies 256, but permits future architecture
408 revisions to raise this limit.)
410 * FPSR and FPCR are retained from ARMv8-A, and interact with SVE floating-point
411 operations in a similar way to the way in which they interact with ARMv8
412 floating-point operations.
414 8VL-1 128 0 bit index
415 +---- //// -----------------+
425 +---- //// -----------------+
428 +---- //// --+ FPSR | |
434 +---- //// --+ VL | |
438 This only applies to bits [63:0] of Z-/V-registers.
439 FPCR contains callee-save and caller-save bits. See [4] for details.
442 A.2. Procedure call standard
443 -----------------------------
445 The ARMv8-A base procedure call standard is extended as follows with respect to
446 the additional SVE register state:
448 * All SVE register bits that are not shared with FP/SIMD are caller-save.
450 * Z8 bits [63:0] .. Z15 bits [63:0] are callee-save.
452 This follows from the way these bits are mapped to V8..V15, which are caller-
453 save in the base procedure call standard.
456 Appendix B. ARMv8-A FP/SIMD programmer's model
457 ===============================================
459 Note: This section is for information only and not intended to be complete or
460 to replace any architectural specification.
462 Refer to [4] for for more information.
464 ARMv8-A defines the following floating-point / SIMD register state:
466 * 32 128-bit vector registers V0..V31
467 * 2 32-bit status/control registers FPSR, FPCR
490 This only applies to bits [63:0] of V-registers.
491 FPCR contains a mixture of callee-save and caller-save bits.
497 [1] arch/arm64/include/uapi/asm/sigcontext.h
498 AArch64 Linux signal ABI definitions
500 [2] arch/arm64/include/uapi/asm/ptrace.h
501 AArch64 Linux ptrace ABI definitions
503 [3] Documentation/arm64/cpu-feature-registers.txt
506 http://infocenter.arm.com/help/topic/com.arm.doc.ihi0055c/IHI0055C_beta_aapcs64.pdf
507 http://infocenter.arm.com/help/topic/com.arm.doc.subset.swdev.abi/index.html
508 Procedure Call Standard for the ARM 64-bit Architecture (AArch64)