1 ===================================================
2 Scalable Matrix Extension support for AArch64 Linux
3 ===================================================
5 This document outlines briefly the interface provided to userspace by Linux in
6 order to support use of the ARM Scalable Matrix Extension (SME).
8 This is an outline of the most important features and issues only and not
9 intended to be exhaustive. It should be read in conjunction with the SVE
10 documentation in sve.rst which provides details on the Streaming SVE mode
13 This document does not aim to describe the SME architecture or programmer's
14 model. To aid understanding, a minimal description of relevant programmer's
15 model features for SME is included in Appendix A.
21 * PSTATE.SM, PSTATE.ZA, the streaming mode vector length, the ZA
22 register state and TPIDR2_EL0 are tracked per thread.
24 * The presence of SME is reported to userspace via HWCAP2_SME in the aux vector
25 AT_HWCAP2 entry. Presence of this flag implies the presence of the SME
26 instructions and registers, and the Linux-specific system interfaces
27 described in this document. SME is reported in /proc/cpuinfo as "sme".
29 * Support for the execution of SME instructions in userspace can also be
30 detected by reading the CPU ID register ID_AA64PFR1_EL1 using an MRS
31 instruction, and checking that the value of the SME 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 HWCAP2_SME instead.
37 * There are a number of optional SME features, presence of these is reported
38 through AT_HWCAP2 through:
48 This list may be extended over time as the SME architecture evolves.
50 These extensions are also reported via the CPU ID register ID_AA64SMFR0_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, NT_ARM_SSVE and NT_ARM_ZA regsets. The recommended way
56 of detecting support for these regsets is to connect to a target process
57 first and then attempt a
59 ptrace(PTRACE_GETREGSET, pid, NT_ARM_<regset>, &iov).
61 * Whenever ZA register values are exchanged in memory between userspace and
62 the kernel, the register value is encoded in memory as a series of horizontal
63 vectors from 0 to VL/8-1 stored in the same endianness invariant format as is
66 * On thread creation TPIDR2_EL0 is preserved unless CLONE_SETTLS is specified,
67 in which case it is set to 0.
72 SME defines a second vector length similar to the SVE vector length which is
73 controls the size of the streaming mode SVE vectors and the ZA matrix array.
74 The ZA matrix is square with each side having as many bytes as a streaming
78 3. Sharing of streaming and non-streaming mode SVE state
79 ---------------------------------------------------------
81 It is implementation defined which if any parts of the SVE state are shared
82 between streaming and non-streaming modes. When switching between modes
83 via software interfaces such as ptrace if no register content is provided as
84 part of switching no state will be assumed to be shared and everything will
88 4. System call behaviour
89 -------------------------
91 * On syscall PSTATE.ZA is preserved, if PSTATE.ZA==1 then the contents of the
92 ZA matrix are preserved.
94 * On syscall PSTATE.SM will be cleared and the SVE registers will be handled
95 as per the standard SVE ABI.
97 * Neither the SVE registers nor ZA are used to pass arguments to or receive
98 results from any syscall.
100 * On process creation (eg, clone()) the newly created process will have
103 * All other SME state of a thread, including the currently configured vector
104 length, the state of the PR_SME_VL_INHERIT flag, and the deferred vector
105 length (if any), is preserved across all syscalls, subject to the specific
106 exceptions for execve() described in section 6.
112 * Signal handlers are invoked with streaming mode and ZA disabled.
114 * A new signal frame record za_context encodes the ZA register contents on
117 * The signal frame record for ZA always contains basic metadata, in particular
118 the thread's vector length (in za_context.vl).
120 * The ZA matrix may or may not be included in the record, depending on
121 the value of PSTATE.ZA. The registers are present if and only if:
122 za_context.head.size >= ZA_SIG_CONTEXT_SIZE(sve_vq_from_vl(za_context.vl))
123 in which case PSTATE.ZA == 1.
125 * If matrix data is present, the remainder of the record has a vl-dependent
126 size and layout. Macros ZA_SIG_* are defined [1] to facilitate access to
129 * The matrix is stored as a series of horizontal vectors in the same format as
130 is used for SVE vectors.
132 * If the ZA context is too big to fit in sigcontext.__reserved[], then extra
133 space is allocated on the stack, an extra_context record is written in
134 __reserved[] referencing this space. za_context is then written in the
135 extra space. Refer to [1] for further details about this mechanism.
141 When returning from a signal handler:
143 * If there is no za_context record in the signal frame, or if the record is
144 present but contains no register data as described in the previous section,
147 * If za_context is present in the signal frame and contains matrix data then
148 PSTATE.ZA is set to 1 and ZA is populated with the specified data.
150 * The vector length cannot be changed via signal return. If za_context.vl in
151 the signal frame does not match the current vector length, the signal return
152 attempt is treated as illegal, resulting in a forced SIGSEGV.
158 Some new prctl() calls are added to allow programs to manage the SME vector
161 prctl(PR_SME_SET_VL, unsigned long arg)
163 Sets the vector length of the calling thread and related flags, where
164 arg == vl | flags. Other threads of the calling process are unaffected.
166 vl is the desired vector length, where sve_vl_valid(vl) must be true.
172 Inherit the current vector length across execve(). Otherwise, the
173 vector length is reset to the system default at execve(). (See
178 Defer the requested vector length change until the next execve()
179 performed by this thread.
181 The effect is equivalent to implicit execution of the following
182 call immediately after the next execve() (if any) by the thread:
184 prctl(PR_SME_SET_VL, arg & ~PR_SME_SET_VL_ONEXEC)
186 This allows launching of a new program with a different vector
187 length, while avoiding runtime side effects in the caller.
189 Without PR_SME_SET_VL_ONEXEC, the requested change takes effect
193 Return value: a nonnegative on success, or a negative value on error:
194 EINVAL: SME not supported, invalid vector length requested, or
200 * Either the calling thread's vector length or the deferred vector length
201 to be applied at the next execve() by the thread (dependent on whether
202 PR_SME_SET_VL_ONEXEC is present in arg), is set to the largest value
203 supported by the system that is less than or equal to vl. If vl ==
204 SVE_VL_MAX, the value set will be the largest value supported by the
207 * Any previously outstanding deferred vector length change in the calling
210 * The returned value describes the resulting configuration, encoded as for
211 PR_SME_GET_VL. The vector length reported in this value is the new
212 current vector length for this thread if PR_SME_SET_VL_ONEXEC was not
213 present in arg; otherwise, the reported vector length is the deferred
214 vector length that will be applied at the next execve() by the calling
217 * Changing the vector length causes all of ZA, P0..P15, FFR and all bits of
218 Z0..Z31 except for Z0 bits [127:0] .. Z31 bits [127:0] to become
219 unspecified, including both streaming and non-streaming SVE state.
220 Calling PR_SME_SET_VL with vl equal to the thread's current vector
221 length, or calling PR_SME_SET_VL with the PR_SVE_SET_VL_ONEXEC flag,
222 does not constitute a change to the vector length for this purpose.
224 * Changing the vector length causes PSTATE.ZA and PSTATE.SM to be cleared.
225 Calling PR_SME_SET_VL with vl equal to the thread's current vector
226 length, or calling PR_SME_SET_VL with the PR_SVE_SET_VL_ONEXEC flag,
227 does not constitute a change to the vector length for this purpose.
232 Gets the vector length of the calling thread.
234 The following flag may be OR-ed into the result:
238 Vector length will be inherited across execve().
240 There is no way to determine whether there is an outstanding deferred
241 vector length change (which would only normally be the case between a
242 fork() or vfork() and the corresponding execve() in typical use).
244 To extract the vector length from the result, bitwise and it with
247 Return value: a nonnegative value on success, or a negative value on error:
248 EINVAL: SME not supported.
252 ---------------------
254 * A new regset NT_ARM_SSVE is defined for access to streaming mode SVE
255 state via PTRACE_GETREGSET and PTRACE_SETREGSET, this is documented in
258 * A new regset NT_ARM_ZA is defined for ZA state for access to ZA state via
259 PTRACE_GETREGSET and PTRACE_SETREGSET.
261 Refer to [2] for definitions.
263 The regset data starts with struct user_za_header, containing:
267 Size of the complete regset, in bytes.
268 This depends on vl and possibly on other things in the future.
270 If a call to PTRACE_GETREGSET requests less data than the value of
271 size, the caller can allocate a larger buffer and retry in order to
272 read the complete regset.
276 Maximum size in bytes that the regset can grow to for the target
277 thread. The regset won't grow bigger than this even if the target
278 thread changes its vector length etc.
282 Target thread's current streaming vector length, in bytes.
286 Maximum possible streaming vector length for the target thread.
290 Zero or more of the following flags, which have the same
291 meaning and behaviour as the corresponding PR_SET_VL_* flags:
295 SME_PT_VL_ONEXEC (SETREGSET only).
297 * The effects of changing the vector length and/or flags are equivalent to
298 those documented for PR_SME_SET_VL.
300 The caller must make a further GETREGSET call if it needs to know what VL is
301 actually set by SETREGSET, unless is it known in advance that the requested
304 * The size and layout of the payload depends on the header fields. The
305 SME_PT_ZA_*() macros are provided to facilitate access to the data.
307 * In either case, for SETREGSET it is permissible to omit the payload, in which
308 case the vector length and flags are changed and PSTATE.ZA is set to 0
309 (along with any consequences of those changes). If a payload is provided
310 then PSTATE.ZA will be set to 1.
312 * For SETREGSET, if the requested VL is not supported, the effect will be the
313 same as if the payload were omitted, except that an EIO error is reported.
314 No attempt is made to translate the payload data to the correct layout
315 for the vector length actually set. It is up to the caller to translate the
316 payload layout for the actual VL and retry.
318 * The effect of writing a partial, incomplete payload is unspecified.
321 8. ELF coredump extensions
322 ---------------------------
324 * NT_ARM_SSVE notes will be added to each coredump for
325 each thread of the dumped process. The contents will be equivalent to the
326 data that would have been read if a PTRACE_GETREGSET of the corresponding
327 type were executed for each thread when the coredump was generated.
329 * A NT_ARM_ZA 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_ZA 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/sme_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 32 or the maximum
353 supported vector length, whichever is smaller and supported. This
354 determines the initial 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_SME_VL_INHERIT (or equivalently SME_PT_VL_INHERIT) is set for the
364 * a deferred vector length change is pending, established via the
365 PR_SME_SET_VL_ONEXEC flag (or SME_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. SME 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, SME adds the following:
385 * A new mode, streaming mode, in which a subset of the normal FPSIMD and SVE
386 features are available. When supported EL0 software may enter and leave
387 streaming mode at any time.
389 For best system performance it is strongly encouraged for software to enable
390 streaming mode only when it is actively being used.
392 * A new vector length controlling the size of ZA and the Z registers when in
393 streaming mode, separately to the vector length used for SVE when not in
394 streaming mode. There is no requirement that either the currently selected
395 vector length or the set of vector lengths supported for the two modes in
396 a given system have any relationship. The streaming mode vector length
397 is referred to as SVL.
399 * A new ZA matrix register. This is a square matrix of SVLxSVL bits. Most
400 operations on ZA require that streaming mode be enabled but ZA can be
401 enabled without streaming mode in order to load, save and retain data.
403 For best system performance it is strongly encouraged for software to enable
404 ZA only when it is actively being used.
406 * Two new 1 bit fields in PSTATE which may be controlled via the SMSTART and
407 SMSTOP instructions or by access to the SVCR system register:
409 * PSTATE.ZA, if this is 1 then the ZA matrix is accessible and has valid
410 data while if it is 0 then ZA can not be accessed. When PSTATE.ZA is
411 changed from 0 to 1 all bits in ZA are cleared.
413 * PSTATE.SM, if this is 1 then the PE is in streaming mode. When the value
414 of PSTATE.SM is changed then it is implementation defined if the subset
415 of the floating point register bits valid in both modes may be retained.
416 Any other bits will be cleared.
422 [1] arch/arm64/include/uapi/asm/sigcontext.h
423 AArch64 Linux signal ABI definitions
425 [2] arch/arm64/include/uapi/asm/ptrace.h
426 AArch64 Linux ptrace ABI definitions
428 [3] Documentation/arm64/cpu-feature-registers.rst