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).
12 This is an outline of the most important features and issues only and not
13 intended to be exhaustive.
15 This document does not aim to describe the SVE architecture or programmer's
16 model. To aid understanding, a minimal description of relevant programmer's
17 model features for SVE is included in Appendix A.
23 * SVE registers Z0..Z31, P0..P15 and FFR and the current vector length VL, are
26 * The presence of SVE is reported to userspace via HWCAP_SVE in the aux vector
27 AT_HWCAP entry. Presence of this flag implies the presence of the SVE
28 instructions and registers, and the Linux-specific system interfaces
29 described in this document. SVE is reported in /proc/cpuinfo as "sve".
31 * Support for the execution of SVE instructions in userspace can also be
32 detected by reading the CPU ID register ID_AA64PFR0_EL1 using an MRS
33 instruction, and checking that the value of the SVE field is nonzero. [3]
35 It does not guarantee the presence of the system interfaces described in the
36 following sections: software that needs to verify that those interfaces are
37 present must check for HWCAP_SVE instead.
39 * On hardware that supports the SVE2 extensions, HWCAP2_SVE2 will also
40 be reported in the AT_HWCAP2 aux vector entry. In addition to this,
41 optional extensions to SVE2 may be reported by the presence of:
50 This list may be extended over time as the SVE architecture evolves.
52 These extensions are also reported via the CPU ID register ID_AA64ZFR0_EL1,
53 which userspace can read using an MRS instruction. See elf_hwcaps.txt and
54 cpu-feature-registers.txt for details.
56 * Debuggers should restrict themselves to interacting with the target via the
57 NT_ARM_SVE regset. The recommended way of detecting support for this regset
58 is to connect to a target process first and then attempt a
59 ptrace(PTRACE_GETREGSET, pid, NT_ARM_SVE, &iov).
61 * Whenever SVE scalable register values (Zn, Pn, FFR) are exchanged in memory
62 between userspace and the kernel, the register value is encoded in memory in
63 an endianness-invariant layout, with bits [(8 * i + 7) : (8 * i)] encoded at
64 byte offset i from the start of the memory representation. This affects for
65 example the signal frame (struct sve_context) and ptrace interface
66 (struct user_sve_header) and associated data.
68 Beware that on big-endian systems this results in a different byte order than
69 for the FPSIMD V-registers, which are stored as single host-endian 128-bit
70 values, with bits [(127 - 8 * i) : (120 - 8 * i)] of the register encoded at
71 byte offset i. (struct fpsimd_context, struct user_fpsimd_state).
74 2. Vector length terminology
75 -----------------------------
77 The size of an SVE vector (Z) register is referred to as the "vector length".
79 To avoid confusion about the units used to express vector length, the kernel
80 adopts the following conventions:
82 * Vector length (VL) = size of a Z-register in bytes
84 * Vector quadwords (VQ) = size of a Z-register in units of 128 bits
88 The VQ convention is used where the underlying granularity is important, such
89 as in data structure definitions. In most other situations, the VL convention
90 is used. This is consistent with the meaning of the "VL" pseudo-register in
91 the SVE instruction set architecture.
94 3. System call behaviour
95 -------------------------
97 * On syscall, V0..V31 are preserved (as without SVE). Thus, bits [127:0] of
98 Z0..Z31 are preserved. All other bits of Z0..Z31, and all of P0..P15 and FFR
99 become unspecified on return from a syscall.
101 * The SVE registers are not used to pass arguments to or receive results from
104 * In practice the affected registers/bits will be preserved or will be replaced
105 with zeros on return from a syscall, but userspace should not make
106 assumptions about this. The kernel behaviour may vary on a case-by-case
109 * All other SVE state of a thread, including the currently configured vector
110 length, the state of the PR_SVE_VL_INHERIT flag, and the deferred vector
111 length (if any), is preserved across all syscalls, subject to the specific
112 exceptions for execve() described in section 6.
114 In particular, on return from a fork() or clone(), the parent and new child
115 process or thread share identical SVE configuration, matching that of the
116 parent before the call.
122 * A new signal frame record sve_context encodes the SVE registers on signal
125 * This record is supplementary to fpsimd_context. The FPSR and FPCR registers
126 are only present in fpsimd_context. For convenience, the content of V0..V31
127 is duplicated between sve_context and fpsimd_context.
129 * The signal frame record for SVE always contains basic metadata, in particular
130 the thread's vector length (in sve_context.vl).
132 * The SVE registers may or may not be included in the record, depending on
133 whether the registers are live for the thread. The registers are present if
135 sve_context.head.size >= SVE_SIG_CONTEXT_SIZE(sve_vq_from_vl(sve_context.vl)).
137 * If the registers are present, the remainder of the record has a vl-dependent
138 size and layout. Macros SVE_SIG_* are defined [1] to facilitate access to
141 * Each scalable register (Zn, Pn, FFR) is stored in an endianness-invariant
142 layout, with bits [(8 * i + 7) : (8 * i)] stored at byte offset i from the
143 start of the register's representation in memory.
145 * If the SVE context is too big to fit in sigcontext.__reserved[], then extra
146 space is allocated on the stack, an extra_context record is written in
147 __reserved[] referencing this space. sve_context is then written in the
148 extra space. Refer to [1] for further details about this mechanism.
154 When returning from a signal handler:
156 * If there is no sve_context record in the signal frame, or if the record is
157 present but contains no register data as desribed in the previous section,
158 then the SVE registers/bits become non-live and take unspecified values.
160 * If sve_context is present in the signal frame and contains full register
161 data, the SVE registers become live and are populated with the specified
162 data. However, for backward compatibility reasons, bits [127:0] of Z0..Z31
163 are always restored from the corresponding members of fpsimd_context.vregs[]
164 and not from sve_context. The remaining bits are restored from sve_context.
166 * Inclusion of fpsimd_context in the signal frame remains mandatory,
167 irrespective of whether sve_context is present or not.
169 * The vector length cannot be changed via signal return. If sve_context.vl in
170 the signal frame does not match the current vector length, the signal return
171 attempt is treated as illegal, resulting in a forced SIGSEGV.
177 Some new prctl() calls are added to allow programs to manage the SVE vector
180 prctl(PR_SVE_SET_VL, unsigned long arg)
182 Sets the vector length of the calling thread and related flags, where
183 arg == vl | flags. Other threads of the calling process are unaffected.
185 vl is the desired vector length, where sve_vl_valid(vl) must be true.
191 Inherit the current vector length across execve(). Otherwise, the
192 vector length is reset to the system default at execve(). (See
197 Defer the requested vector length change until the next execve()
198 performed by this thread.
200 The effect is equivalent to implicit exceution of the following
201 call immediately after the next execve() (if any) by the thread:
203 prctl(PR_SVE_SET_VL, arg & ~PR_SVE_SET_VL_ONEXEC)
205 This allows launching of a new program with a different vector
206 length, while avoiding runtime side effects in the caller.
209 Without PR_SVE_SET_VL_ONEXEC, the requested change takes effect
213 Return value: a nonnegative on success, or a negative value on error:
214 EINVAL: SVE not supported, invalid vector length requested, or
220 * Either the calling thread's vector length or the deferred vector length
221 to be applied at the next execve() by the thread (dependent on whether
222 PR_SVE_SET_VL_ONEXEC is present in arg), is set to the largest value
223 supported by the system that is less than or equal to vl. If vl ==
224 SVE_VL_MAX, the value set will be the largest value supported by the
227 * Any previously outstanding deferred vector length change in the calling
230 * The returned value describes the resulting configuration, encoded as for
231 PR_SVE_GET_VL. The vector length reported in this value is the new
232 current vector length for this thread if PR_SVE_SET_VL_ONEXEC was not
233 present in arg; otherwise, the reported vector length is the deferred
234 vector length that will be applied at the next execve() by the calling
237 * Changing the vector length causes all of P0..P15, FFR and all bits of
238 Z0..Z31 except for Z0 bits [127:0] .. Z31 bits [127:0] to become
239 unspecified. Calling PR_SVE_SET_VL with vl equal to the thread's current
240 vector length, or calling PR_SVE_SET_VL with the PR_SVE_SET_VL_ONEXEC
241 flag, does not constitute a change to the vector length for this purpose.
246 Gets the vector length of the calling thread.
248 The following flag may be OR-ed into the result:
252 Vector length will be inherited across execve().
254 There is no way to determine whether there is an outstanding deferred
255 vector length change (which would only normally be the case between a
256 fork() or vfork() and the corresponding execve() in typical use).
258 To extract the vector length from the result, and it with
261 Return value: a nonnegative value on success, or a negative value on error:
262 EINVAL: SVE not supported.
266 ---------------------
268 * A new regset NT_ARM_SVE is defined for use with PTRACE_GETREGSET and
271 Refer to [2] for definitions.
273 The regset data starts with struct user_sve_header, containing:
277 Size of the complete regset, in bytes.
278 This depends on vl and possibly on other things in the future.
280 If a call to PTRACE_GETREGSET requests less data than the value of
281 size, the caller can allocate a larger buffer and retry in order to
282 read the complete regset.
286 Maximum size in bytes that the regset can grow to for the target
287 thread. The regset won't grow bigger than this even if the target
288 thread changes its vector length etc.
292 Target thread's current vector length, in bytes.
296 Maximum possible vector length for the target thread.
304 SVE registers are not live (GETREGSET) or are to be made
305 non-live (SETREGSET).
307 The payload is of type struct user_fpsimd_state, with the same
308 meaning as for NT_PRFPREG, starting at offset
309 SVE_PT_FPSIMD_OFFSET from the start of user_sve_header.
311 Extra data might be appended in the future: the size of the
312 payload should be obtained using SVE_PT_FPSIMD_SIZE(vq, flags).
314 vq should be obtained using sve_vq_from_vl(vl).
320 SVE registers are live (GETREGSET) or are to be made live
323 The payload contains the SVE register data, starting at offset
324 SVE_PT_SVE_OFFSET from the start of user_sve_header, and with
325 size SVE_PT_SVE_SIZE(vq, flags);
327 ... OR-ed with zero or more of the following flags, which have the same
328 meaning and behaviour as the corresponding PR_SET_VL_* flags:
332 SVE_PT_VL_ONEXEC (SETREGSET only).
334 * The effects of changing the vector length and/or flags are equivalent to
335 those documented for PR_SVE_SET_VL.
337 The caller must make a further GETREGSET call if it needs to know what VL is
338 actually set by SETREGSET, unless is it known in advance that the requested
341 * In the SVE_PT_REGS_SVE case, the size and layout of the payload depends on
342 the header fields. The SVE_PT_SVE_*() macros are provided to facilitate
343 access to the members.
345 * In either case, for SETREGSET it is permissible to omit the payload, in which
346 case only the vector length and flags are changed (along with any
347 consequences of those changes).
349 * For SETREGSET, if an SVE_PT_REGS_SVE payload is present and the
350 requested VL is not supported, the effect will be the same as if the
351 payload were omitted, except that an EIO error is reported. No
352 attempt is made to translate the payload data to the correct layout
353 for the vector length actually set. The thread's FPSIMD state is
354 preserved, but the remaining bits of the SVE registers become
355 unspecified. It is up to the caller to translate the payload layout
356 for the actual VL and retry.
358 * The effect of writing a partial, incomplete payload is unspecified.
361 8. ELF coredump extensions
362 ---------------------------
364 * A NT_ARM_SVE note will be added to each coredump for each thread of the
365 dumped process. The contents will be equivalent to the data that would have
366 been read if a PTRACE_GETREGSET of NT_ARM_SVE were executed for each thread
367 when the coredump was generated.
370 9. System runtime configuration
371 --------------------------------
373 * To mitigate the ABI impact of expansion of the signal frame, a policy
374 mechanism is provided for administrators, distro maintainers and developers
375 to set the default vector length for userspace processes:
377 /proc/sys/abi/sve_default_vector_length
379 Writing the text representation of an integer to this file sets the system
380 default vector length to the specified value, unless the value is greater
381 than the maximum vector length supported by the system in which case the
382 default vector length is set to that maximum.
384 The result can be determined by reopening the file and reading its
387 At boot, the default vector length is initially set to 64 or the maximum
388 supported vector length, whichever is smaller. This determines the initial
389 vector length of the init process (PID 1).
391 Reading this file returns the current system default vector length.
393 * At every execve() call, the new vector length of the new process is set to
394 the system default vector length, unless
396 * PR_SVE_VL_INHERIT (or equivalently SVE_PT_VL_INHERIT) is set for the
399 * a deferred vector length change is pending, established via the
400 PR_SVE_SET_VL_ONEXEC flag (or SVE_PT_VL_ONEXEC).
402 * Modifying the system default vector length does not affect the vector length
403 of any existing process or thread that does not make an execve() call.
406 Appendix A. SVE programmer's model (informative)
407 =================================================
409 This section provides a minimal description of the additions made by SVE to the
410 ARMv8-A programmer's model that are relevant to this document.
412 Note: This section is for information only and not intended to be complete or
413 to replace any architectural specification.
418 In A64 state, SVE adds the following:
420 * 32 8VL-bit vector registers Z0..Z31
421 For each Zn, Zn bits [127:0] alias the ARMv8-A vector register Vn.
423 A register write using a Vn register name zeros all bits of the corresponding
424 Zn except for bits [127:0].
426 * 16 VL-bit predicate registers P0..P15
428 * 1 VL-bit special-purpose predicate register FFR (the "first-fault register")
430 * a VL "pseudo-register" that determines the size of each vector register
432 The SVE instruction set architecture provides no way to write VL directly.
433 Instead, it can be modified only by EL1 and above, by writing appropriate
436 * The value of VL can be configured at runtime by EL1 and above:
437 16 <= VL <= VLmax, where VL must be a multiple of 16.
439 * The maximum vector length is determined by the hardware:
442 (The SVE architecture specifies 256, but permits future architecture
443 revisions to raise this limit.)
445 * FPSR and FPCR are retained from ARMv8-A, and interact with SVE floating-point
446 operations in a similar way to the way in which they interact with ARMv8
447 floating-point operations::
449 8VL-1 128 0 bit index
450 +---- //// -----------------+
460 +---- //// -----------------+
463 +---- //// --+ FPSR | |
469 +---- //// --+ VL | |
473 This only applies to bits [63:0] of Z-/V-registers.
474 FPCR contains callee-save and caller-save bits. See [4] for details.
477 A.2. Procedure call standard
478 -----------------------------
480 The ARMv8-A base procedure call standard is extended as follows with respect to
481 the additional SVE register state:
483 * All SVE register bits that are not shared with FP/SIMD are caller-save.
485 * Z8 bits [63:0] .. Z15 bits [63:0] are callee-save.
487 This follows from the way these bits are mapped to V8..V15, which are caller-
488 save in the base procedure call standard.
491 Appendix B. ARMv8-A FP/SIMD programmer's model
492 ===============================================
494 Note: This section is for information only and not intended to be complete or
495 to replace any architectural specification.
497 Refer to [4] for more information.
499 ARMv8-A defines the following floating-point / SIMD register state:
501 * 32 128-bit vector registers V0..V31
502 * 2 32-bit status/control registers FPSR, FPCR
527 This only applies to bits [63:0] of V-registers.
528 FPCR contains a mixture of callee-save and caller-save bits.
534 [1] arch/arm64/include/uapi/asm/sigcontext.h
535 AArch64 Linux signal ABI definitions
537 [2] arch/arm64/include/uapi/asm/ptrace.h
538 AArch64 Linux ptrace ABI definitions
540 [3] Documentation/arm64/cpu-feature-registers.rst
543 http://infocenter.arm.com/help/topic/com.arm.doc.ihi0055c/IHI0055C_beta_aapcs64.pdf
544 http://infocenter.arm.com/help/topic/com.arm.doc.subset.swdev.abi/index.html
545 Procedure Call Standard for the ARM 64-bit Architecture (AArch64)