My new favourite AArch64 CPU instruction: rotate then merge in to flags (RMIF)

I find myself writing some CPU emulators at the moment, which has caused the AArch64 (aka. ARM64) RMIF instruction to become my new favourite instruction. It takes a 64-bit general purpose register, rotates it right by a specified number of bits, then selectively merges the low four bits into the flags register. A 6-bit field in the instruction gives the rotate amount, and a 4-bit field in the instruction gives a mask of which flag bits to overwrite versus which to leave unchanged.

One use of rmif is to emulate the x86 bt reg, imm instruction, which extracts one bit from a general purpose register, writes that bit to the C flag, and leaves other flags unchanged. Thus bt reg, imm in x86 becomes rmif reg, #((imm - 1) & 63), #2 in AArch64.

At the other end of the spectrum is the x86 inc instruction, which adds 1 to a general purpose register, and then sets most flags based on this addition, but leaves the C flag unchanged. To emulate inc reg, we can first save off the old value of the C flag (via csinc tmp, wzr, wzr, cc or adc tmp, wzr, wzr), then do adds reg, reg, #1 to perform the addition and set all the flags, then rmif tmp, #63, #2 to restore the old value of the C flag.

As another example, the AArch32 muls instruction sets the N and Z flags based on the result of the multiplication, but leaves the C and V flags unchanged. To emulate this on AArch64, we can save off all the flags (mrs tmp, NZCV), then do the multiplication, then set N and Z based on the result but also clobber C and V (ands wzr, dst, dst or adds wzr, dst, #0), then restore the old values of C and V (rmif tmp, #28, #3).

For want of a relative path

Distributing dynamically-linked ELF executables on Linux can be arduous. Some downstream effects of this include:

At first, the problem doesn't look arduous: an ELF executable can contain an rpath or runpath attribute telling the dynamic linker where to find its shared object dependencies, and if that attribute starts with the magic placeholder $ORIGIN/, then the dynamic linker will look in the directory containing the executable (or a directory nearby) for its shared object dependencies. For example, if my_executable depended upon libz.so.1, and my_executable had an rpath or runpath of $ORIGIN/libs, then the executable and the library could be distributed using the following directory structure:

my_executable
libs/
  libz.so.1

This is great, but it has one limitation: an ELF executable also contains an attribute telling the kernel where to find the dynamic linker, and that attribute has to be an absolute path (or a path relative to the current working directory); it cannot be a path relative to the executable. On contemporary x86-64 systems, that absolute path tends to be /lib64/ld-linux-x86-64.so.2. This forces ELF executables to use whatever the system provides at /lib64/ld-linux-x86-64.so.2, which is typically version N of glibc's dynamic linker, for some N. In turn, this forces the ELF executable to use version N of the rest of glibc (libc.so.6, libm.so.6, libpthread.so.0, etc).

Continuing the example, it is likely that my_executable and libz.so.1 were built against some version M of glibc. If M ≤ N, then everything will work fine, but problems often crop up when M > N. One commonly touted solution is to set up a build environment with a very old version M of glibc, build my_executable and libz.so.1 in that environment, and then distribute them and hope for M ≤ N.

The polyfill-glibc project presents another possible solution: build my_executable and libz.so.1 against whatever version of glibc is convenient, and then run polyfill-glibc --target-glibc=N my_executable libz.so.1 to make them compatible with version N of glibc.

Sometimes we don't want either of these solutions, and what we want is to distribute the required version of glibc along with the executable, as in:

my_executable
libs/
  ld-linux-x86-64.so.2
  libc.so.6
  libz.so.1

We can get close to this by adding a launcher script:

launch_my_executable
my_executable
libs/
  ld-linux-x86-64.so.2
  libc.so.6
  libz.so.1

Where launch_my_executable is something like:

#!/usr/bin/env bash

ORIGIN="$(dirname "$(readlink -f "$0")")"
exec "$ORIGIN/libs/ld-linux-x86-64.so.2" --library-path "$ORIGIN/libs:$LD_LIBRARY_PATH" "$ORIGIN/my_executable" "$@"

This will work most of the time, though comes with caveats:

As an alternative without these caveats, there's an experimental tool in the polyfill-glibc repository called set_relative_interp. For our running example, the tool would be invoked as:

$ set_relative_interp my_executable libs/ld-linux-x86-64.so.2

After running the tool as above, my_executable will use $ORIGIN/libs/ld-linux-x86-64.so.2 as its dynamic linker.

(Ab)using gf2p8affineqb to turn indices into bits

@geofflangdale posed the question on Twitter of how to vectorise this:

__mmask64 reference_impl(__m512i indices, __mmask64 valids) {
  __mmask64 result = 0;
  for (int i = 0; i < 64; ++i) {
    if (valids.bit[i]) {
      result ^= 1ull << indices.byte[i];
    }
  }
  return result;
}

After a week of code golf also involving @HaroldAptroot, we ended up with:

__mmask64 simd_impl(__m512i indices, __mmask64 valids) {
  // Convert indices to bits within each qword lane.
  __m512i khi = _mm512_setr_epi8(
    0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
    0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
    0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04,
    0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
    0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10,
    0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
    0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40,
    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80
  );
  __m512i hi0 = _mm512_permutexvar_epi8(indices, khi);
  __m512i klo = _mm512_set1_epi64(0x0102040810204080);
  __m512i lo0 = _mm512_maskz_shuffle_epi8(valids, klo, indices);
  __m512i kid = _mm512_set1_epi64(0x8040201008040201);
  __m512i hi1 = _mm512_gf2p8affine_epi64_epi8(kid, hi0, 0);
  __m512i lo1 = _mm512_gf2p8affine_epi64_epi8(kid, lo0, 0);
  __m512i x0  = _mm512_gf2p8affine_epi64_epi8(hi1, lo1, 0);
  // Combine results from various qword lanes.
  __m512i ktr = _mm512_setr_epi8(
    0,  8, 16, 24, 32, 40, 48, 56,
    1,  9, 17, 25, 33, 41, 49, 57,
    2, 10, 18, 26, 34, 42, 50, 58,
    3, 11, 19, 27, 35, 43, 51, 59,
    4, 12, 20, 28, 36, 44, 52, 60,
    5, 13, 21, 29, 37, 45, 53, 61,
    6, 14, 22, 30, 38, 46, 54, 62,
    7, 15, 23, 31, 39, 47, 55, 63);
  __m512i x1  = _mm512_permutexvar_epi8(ktr, x0);
  __m512i x2  = _mm512_gf2p8affine_epi64_epi8(kid, x1, 0);
  // Reduce 64 bytes down to 64 bits.
  __m512i kff = _mm512_set1_epi8(0xff);
  __m512i x3  = _mm512_gf2p8affine_epi64_epi8(x2, kff, 0);
  return _mm512_movepi8_mask(x3);
}

NB: If the valid indices can be assumed to be distinct, then the final reduction from 64 bytes to 64 bits can instead be:

  return _mm512_cmpneq_epi8_mask(x2, _mm512_setzero_si512());

As is often the case, simd_impl looks nothing like reference_impl, despite doing the same thing. In particular, simd_impl contains no shifts, and instead contains alternating shuffles and invocations of the mysterious _mm512_gf2p8affine_epi64_epi8, which is the intrinsic function corresponding to the gf2p8affineqb assembly instruction. To understand how simd_impl works, we're going to have to first understand what gf2p8affineqb does.

There are various ways of understanding what gf2p8affineqb does, but for the purposes of this blog post, I think the following Python pseudo-code is most useful:

def gf2p8affineqb(src1 : vector, src2 : vector, imm8 : u8) -> vector:
  assert len(src1.byte) == len(src2.byte)
  dst = vector()
  for i in range(len(src1.byte)):
    munged_src2 = munge(src2.qword[i // 8])
    dst.byte[i] = xor_selected(src1.byte[i], munged_src2, imm8)
  return dst

def xor_selected(src1 : u8, munged_src2 : u64, imm8 : u8) -> u8:
  result = imm8
  for i in range(8):
    if src1.bit[i]:
      result ^= munged_src2.byte[i]
  return result

def munge(x : u64) -> u64:
  return transpose8x8(byte_swap(x))
  # Or equivalently:
  return bitrev_in_each_byte(transpose8x8(x))

def transpose8x8(x : u64) -> u64:
  result = 0
  for i in range(8):
    for j in range(8):
      result.byte[i].bit[j] = x.byte[j].bit[i]
  return result

def byte_swap(x : u64) -> u64:
  result = 0
  for i in range(8):
    result.byte[i] = x.byte[7 - i]
  return result

def bitrev_in_each_byte(x : u64) -> u64:
  result = 0
  for i in range(8):
    result.byte[i] = bitrev(x.byte[i])
  return result

def bitrev(x : u8) -> u8:
  result = 0
  for i in range(8):
    result.bit[i] = x.bit[7 - i]
  return result

The mathematically inclined might notice that the above is in fact doing matrix multiplication of two 8x8 matrices of bits:

def gf2p8affineqb(src1 : vector, src2 : vector, imm8 : u8) -> vector:
  assert len(src1.byte) == len(src2.byte)
  dst = vector()
  for i in range(len(src1.qword)):
    dst.qword[i] = matmul(src1.qword[i], munge(src2.qword[i]))
  for i in range(len(src1.byte)):
    dst.byte[i] ^= imm8
  return dst

def matmul(lhs : u64, rhs : u64) -> u64:
  result = 0
  for i in range(8):
    for j in range(8):
      for k in range(8):
        b = lhs.byte[i].bit[j] * rhs.byte[j].bit[k] # * or &
        result.byte[i].bit[k] += b                  # + or ^
  return result

def munge(x : u64) -> u64:
  # Same as previously

The xor_selected view of gf2p8affineqb and the matmul view of gf2p8affineqb are complementary: I think that the xor_selected view makes it clearer what is going on, but the matmul view is useful for higher level transformations and optimisations. As a middle ground between the two views, matmul can be re-expressed as byte-level operations by unrolling the k loop:

def matmul(lhs : u64, rhs : u64) -> u64:
  result = 0
  for i in range(8):
    for j in range(8):
      if lhs.byte[i].bit[j]:
        result.byte[i] ^= rhs.byte[j]
  return result

One observation from the matmul view is that when src1.qword[i] is the identity matrix, we end up with dst.qword[i] being munge(src2.qword[i]). As a 64-bit integer, said identity matrix is 0x8040201008040201 (i.e. in byte i, just bit i is set). This explains __m512i kid = _mm512_set1_epi64(0x8040201008040201) in simd_impl (kid is just an identity matrix) and also explains __m512i hi1 = _mm512_gf2p8affine_epi64_epi8(kid, hi0, 0) and __m512i lo1 = _mm512_gf2p8affine_epi64_epi8(kid, lo0, 0) - these are just applying munge to every qword (as for what said munges are achieving, we'll get to later).

Changing tack somewhat, it is time to gradually transform reference_impl to make it look more like matmul. For this, we'll start with a simplified version of reference_impl that takes 8 indices rather than 64:

__mmask64 reference_impl_1(__m64i indices, __mmask8 valids) {
  __mmask64 result = 0;
  for (int i = 0; i < 8; ++i) {
    if (valids.bit[i]) {
      result ^= 1ull << indices.byte[i];
    }
  }
  return result;
}

The first transformation is to split each 6-bit index into its low 3 bits and high 3 bits, so that we can address bytes of result:

__mmask64 reference_impl_2(__m64i indices, __mmask8 valids) {
  __mmask64 result = 0;
  for (int i = 0; i < 8; ++i) {
    if (valids.bit[i]) {
      uint8_t b = indices.byte[i];
      uint8_t hi = b >> 3;
      uint8_t lo = b  & 7;
      result.byte[hi] ^= 1 << lo;
    }
  }
  return result;
}

Next up we perform loop fission; doing the exact same work, but using two loops rather than one (so that we can focus on the loops separately):

__mmask64 reference_impl_3(__m64i indices, __mmask8 valids) {
  __m64i hi;
  __m64i lo;
  for (int i = 0; i < 8; ++i) {
    uint8_t b = indices.byte[i];
    hi.byte[i] = b >> 3;
    lo.byte[i] = b  & 7;
  }
  __mmask64 result = 0;
  for (int i = 0; i < 8; ++i) {
    if (valids.bit[i]) {
      result.byte[hi.byte[i]] ^= 1 << lo.byte[i];
    }
  }
  return result;
}

Then the if and the 1 << can also be moved from the 2nd loop to the 1st loop:

__mmask64 reference_impl_4(__m64i indices, __mmask8 valids) {
  __m64i hi;
  __m64i lo;
  for (int i = 0; i < 8; ++i) {
    uint8_t b = indices.byte[i];
    hi.byte[i] = b >> 3;
    lo.byte[i] = valids.bit[i] ? 1 << (b & 7) : 0;
  }
  __mmask64 result = 0;
  for (int i = 0; i < 8; ++i) {
    result.byte[hi.byte[i]] ^= lo.byte[i];
  }
  return result;
}

Then a transformation that looks utterly deranged, but is key to the SIMD transformation; rather than directly indexing using hi.byte[i], we'll loop over the 8 possible values of hi.byte[i] and act when we find the right value:

__mmask64 reference_impl_5(__m64i indices, __mmask8 valids) {
  __m64i hi;
  __m64i lo;
  for (int i = 0; i < 8; ++i) {
    uint8_t b = indices.byte[i];
    hi.byte[i] = b >> 3;
    lo.byte[i] = valids.bit[i] ? 1 << (b & 7) : 0;
  }
  __mmask64 result = 0;
  for (int i = 0; i < 8; ++i) {
    for (int j = 0; j < 8; ++j) {
      if (hi.byte[i] == j) {
        result.byte[j] ^= lo.byte[i];
      }
    }
  }
  return result;
}

Next up we perform loop interchange of the two nested loops:

__mmask64 reference_impl_6(__m64i indices, __mmask8 valids) {
  __m64i hi;
  __m64i lo;
  for (int i = 0; i < 8; ++i) {
    uint8_t b = indices.byte[i];
    hi.byte[i] = b >> 3;
    lo.byte[i] = valids.bit[i] ? 1 << (b & 7) : 0;
  }
  __mmask64 result = 0;
  for (int i = 0; i < 8; ++i) {
    for (int j = 0; j < 8; ++j) {
      if (hi.byte[j] == i) {
        result.byte[i] ^= lo.byte[j];
      }
    }
  }
  return result;
}

Then another transformation that initially looks deranged; the == in hi.byte[j] == i is annoying, and can be replaced by a bit test if we one-hot encode hi:

__mmask64 reference_impl_7(__m64i indices, __mmask8 valids) {
  __m64i hi;
  __m64i lo;
  for (int i = 0; i < 8; ++i) {
    uint8_t b = indices.byte[i];
    hi.byte[i] = 1 << (b >> 3);
    lo.byte[i] = valids.bit[i] ? 1 << (b & 7) : 0;
  }
  __mmask64 result = 0;
  for (int i = 0; i < 8; ++i) {
    for (int j = 0; j < 8; ++j) {
      if (hi.byte[j].bit[i]) {
        result.byte[i] ^= lo.byte[j];
      }
    }
  }
  return result;
}

Then one final transformation to get where we want to be; apply transpose8x8 to hi, and undo it by changing .byte[j].bit[i] to .byte[i].bit[j]:

__mmask64 reference_impl_8(__m64i indices, __mmask8 valids) {
  __m64i hi;
  __m64i lo;
  for (int i = 0; i < 8; ++i) {
    uint8_t b = indices.byte[i];
    hi.byte[i] = 1 << (b >> 3);
    lo.byte[i] = valids.bit[i] ? 1 << (b & 7) : 0;
  }
  __mmask64 result = 0;
  for (int i = 0; i < 8; ++i) {
    for (int j = 0; j < 8; ++j) {
      if (transpose8x8(hi).byte[i].bit[j]) {
        result.byte[i] ^= lo.byte[j];
      }
    }
  }
  return result;
}

A number of these transformations seemed pointless or even unhelpful, but having done them all, the latter half of reference_impl_8 is exactly result = matmul(transpose8x8(hi), lo).

The expression matmul(transpose8x8(A), B) looks deceptively similar to the matmul(A, munge(B)) done by gf2p8affineqb(A, B, 0), and if munge was just transpose8x8, then gf2p8affineqb(munge(A), munge(B), 0) would be exactly matmul(transpose8x8(A), B). Unfortunately, munge also does a bit or byte reversal, causing gf2p8affineqb(munge(A), munge(B), 0) to actually be matmul(transpose8x8(A), bitrev_in_each_byte(B)) (if deriving this, note that munge(A) is bitrev_in_each_byte(transpose8x8(A)), munge(munge(B)) is byte_swap(bitrev_in_each_byte(B)), and then the bitrev_in_each_byte on A cancels out with the byte_swap on B).

The expression matmul(transpose8x8(A), bitrev_in_each_byte(B)) is very close to what we want, and the errant bitrev_in_each_byte can be cancelled out by doing another bitrev_in_each_byte on B:

__mmask64 reference_impl_9(__m64i indices, __mmask8 valids) {
  __m64i hi;
  __m64i lo;
  for (int i = 0; i < 8; ++i) {
    uint8_t b = indices.byte[i];
    hi.byte[i] = 1 << (b >> 3);
    lo.byte[i] = bitrev(valids.bit[i] ? 1 << (b & 7) : 0);
  }
  __mmask64 result = gf2p8affineqb(munge(hi), munge(lo), 0);
  return result;
}

The 1st loop is easy to express in a SIMD manner via a pair of table lookups, thereby giving us the first chunk of simd_impl:

__mmask64 simd_impl(__m512i indices, __mmask64 valids) {
  // Convert indices to bits within each qword lane.
  __m512i khi = _mm512_setr_epi8(
    0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
    0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
    0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04,
    0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
    0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10,
    0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20,
    0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40, 0x40,
    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80
  );
  __m512i hi0 = _mm512_permutexvar_epi8(indices, khi);
  __m512i klo = _mm512_set1_epi64(0x0102040810204080);
  __m512i lo0 = _mm512_maskz_shuffle_epi8(valids, klo, indices);
  __m512i kid = _mm512_set1_epi64(0x8040201008040201);
  __m512i hi1 = _mm512_gf2p8affine_epi64_epi8(kid, hi0, 0); // munge
  __m512i lo1 = _mm512_gf2p8affine_epi64_epi8(kid, lo0, 0); // munge
  __m512i x0  = _mm512_gf2p8affine_epi64_epi8(hi1, lo1, 0);
}

At this point, x0.qword[i] contains reference_impl_9(indices.qword[i], valids.word[i]). To finish up, "all" we need to do is xor together the eight qwords of x0. The traditional way of doing this would be a shuffle followed by a xor to reduce eight to four, another shuffle followed by a xor to reduce four to two, and yet another shuffle followed by a xor to reduce two to one. We can do better than the traditional approach though. The first step is to do one big shuffle rather than three sequential suffles, where the result of the big shuffle moves the eight bytes qword[i].byte[0] to be contiguous, then the eight bytes qword[i].byte[1] to be contiguous, and so on. Seen differently, the bug shuffle is a transpose on an 8x8 matrix of bytes. After this big shuffle, the remaining problem is to take each contiguous group of eight bytes and xor them together. If we wanted to add together each contiguous group of eight bytes, then _mm512_sad_epu8 against zero would be one option, but we want xor rather than add. There are a few different ways of approaching the problem, but one cute way is to apply transpose8x8 to each contiguous group of eight bytes, after which we just need to xor together each contiguous group of eight bits. Applying transpose8x8 on its own is hard, but we can apply munge fairly easily, which does transpose8x8 followed by bitrev_in_each_byte, and the bitrev_in_each_byte is harmless given that we're about to xor together the bits in each byte. This gives us the next chunk of simd_impl:

  // Combine results from various qword lanes.
  __m512i ktr = _mm512_setr_epi8(
    0,  8, 16, 24, 32, 40, 48, 56,
    1,  9, 17, 25, 33, 41, 49, 57,
    2, 10, 18, 26, 34, 42, 50, 58,
    3, 11, 19, 27, 35, 43, 51, 59,
    4, 12, 20, 28, 36, 44, 52, 60,
    5, 13, 21, 29, 37, 45, 53, 61,
    6, 14, 22, 30, 38, 46, 54, 62,
    7, 15, 23, 31, 39, 47, 55, 63);
  __m512i x1  = _mm512_permutexvar_epi8(ktr, x0); // transpose bytes
  __m512i x2  = _mm512_gf2p8affine_epi64_epi8(kid, x1, 0); // munge

If the valid indices can be assumed to be distinct, then we can or (rather than xor) together the bits in each byte, which is just _mm512_cmpneq_epi8_mask against zero.

If we really do need to xor the bits together, then what we want is this function applied to every byte:

def xor_together_bits(x : u8) -> u8:
  result = 0
  for i in range(8):
    if x.bit[i]:
      result ^= 0xff
  return result

If you're thinking that xor_together_bits looks very similar to xor_selected, then you'd be right: xor_together_bits is just xor_selected where every byte of munged_src2 is 0xff, and it so happens that if every byte of src2 is 0xff, then the same is true for munged_src2. This gives the final chunk of simd_impl:

  // Reduce 64 bytes down to 64 bits.
  __m512i kff = _mm512_set1_epi8(0xff);
  __m512i x3  = _mm512_gf2p8affine_epi64_epi8(x2, kff, 0);
  return _mm512_movepi8_mask(x3);

What even is a pidfd anyway?

In recent versions of the Linux kernel, a pidfd is a special type of file that holds a reference to a process. Notably, a pidfd allows for certain process-related operations to be performed in a race-free manner, and it allows poll / select / epoll to be used to detect process termination.

Before you get too excited:

There are various ways of obtaining a pidfd:

Kernel versionglibc versionFunction
5.22.2.5 / 2.31clone with CLONE_PIDFD flag
5.3N/Aclone3 with CLONE_PIDFD flag
5.3 / 5.102.36pidfd_open
5.42.39pidfd_spawn / pidfd_spawnp
6.52.2.5 / N/Agetsockopt with SO_PEERPIDFD optname
6.52.2.5 / 2.39recvmsg with SCM_PIDFD cmsg_type

Once you have a pidfd, there are a bunch of things you can do with it:

Kernel versionglibc versionFunction
5.12.36pidfd_send_signal
5.2 / 5.52.39pidfd_getpid
5.32.2.5 / 2.3.2poll / select / epoll
5.42.2.5 / 2.36waitid with P_PIDFD mode
5.62.36pidfd_getfd
5.82.14setns
5.10 / 5.122.36process_madvise
5.152.36process_mrelease
6.92.2.5 / 2.28fstat / statx for meaningful stx_ino

Some of the subsequent text refers to a process being alive or zombie or dead. These terms come from the usual lifecycle of a unix process: it is initially alive, then transitions to zombie when it terminates, and then transitions to dead once it is waited upon. As a quick summary of the states:

AliveZombieDead
Can execute code and receive signals
Has pid number
Exit code / status retrievable
pidfd polls as readable
Cleaned up by kernel

clone with CLONE_PIDFD flag

Available since: kernel 5.2, glibc 2.31 (or glibc 2.2.5 if you provide your own definition of CLONE_PIDFD; its value is 0x1000).

If the CLONE_PIDFD flag is specified, then clone returns a freshly allocated pidfd referring to the child (in addition to returning the pid number of the child). The O_CLOEXEC flag is automatically set on the returned pidfd. Note that if CLONE_PIDFD is specified, then CLONE_THREAD cannot be specified, nor can CLONE_DETACHED. Furthermore, if CLONE_PIDFD is specified, then CLONE_PARENT_SETTID cannot be specified (unless using clone3).

One of the arguments to clone is the signal number that the child will send to its parent when the child terminates. Setting this to anything other than SIGCHLD has several consequences:

Note that if the child calls execve (or a similar exec function), then the termination signal number is reset to SIGCHLD, and the above points stop applying.

clone3 with CLONE_PIDFD flag

Available since: kernel 5.3, no glibc wrapper.

This function is just a more extensible version of clone; everything written above about clone applies equally to clone3.

pidfd_open

Available since: kernel 5.3, glibc 2.36.

This function takes a pid number (in the pid namespace of the caller), and returns a freshly allocated pidfd refering to said process (or an error if said process does not exist). It is inherently racy, unless the pid number being passed is the result of getpid (i.e. creating a pidfd referring to your own process).

Since kernel 5.10, the PIDFD_NONBLOCK flag can be passed to pidfd_open, which affects subsequent waitid calls. No other flags are valid to pass. The O_CLOEXEC flag is automatically set on the returned pidfd.

pidfd_spawn / pidfd_spawnp

Available since: kernel 5.4, glibc 2.39.

These functions are like posix_spawn / posix_spawnp, except that they have an int* output parameter for a freshly allocated pidfd instead of a pid_t* output parameter for a pid number. The O_CLOEXEC flag is automatically set on the returned pidfd.

In glibc 2.39, bug BZ#31695 causes these functions to leak a file descriptor in some error scenarios. This will hopefully be fixed in 2.40.

getsockopt with SO_PEERPIDFD optname

Available since: kernel 6.5, glibc 2.2.5 for getsockopt. The definition of SO_PEERPIDFD is not tied to a particular glibc version; its value is 77 should you need to provide your own definition of it.

SO_PEERPIDFD is the pidfd version of SO_PEERCRED. For a unix socket created via socketpair, SO_PEERPIDFD gives a pidfd referring to the process that called socketpair, meanwhile for a connected unix stream socket, SO_PEERPIDFD gives a pidfd referring to the process that called connect (if called on the server end of the socket) or the process that called listen (if called on the client end of the socket). The O_CLOEXEC flag is automatically set on the returned pidfd.

recvmsg with SCM_PIDFD cmsg_type

Available since: kernel 6.5, glibc 2.39 (or glibc 2.2.5 if you provide your own definition of SCM_PIDFD; its value is 0x04).

SCM_PIDFD is the pidfd version of (the pid part of) SCM_CREDENTIALS. If the receivier sets SO_PASSPIDFD on a unix socket (c.f. setting SO_PASSCRED), then it'll receive a SCM_PIDFD cmsg as part of receiving a message, with the associated cmsg data being a freshly allocated pidfd referring to the process of the sender of the message (or some other process if the sender has CAP_SYS_ADMIN and specifies a pid number other than itself as part of its SCM_CREDENTIALS). The O_CLOEXEC flag is automatically set on the pidfd.

pidfd_send_signal

Available since: kernel 5.1, glibc 2.36.

This function is similar to kill / rt_sigqueueinfo: it sends a signal to a process. It differs from these functions in that the destination is given as a pidfd rather than as a pid number.

This function also accepts the result of open("/proc/$pid") as an fd, though it is the only function to do so: open("/proc/$pid") does not give a pidfd, and no other functions accept the result of open("/proc/$pid") in place of a pidfd.

pidfd_getpid

Available since: kernel 5.2, glibc 2.39.

This function is the inverse of pidfd_open: given a pidfd, it returns the pid number associated with the underlying process. This function requires that /proc be mounted, and returns the pid number in the pid namespace associated with the mounted /proc. Note that the pid number can be reused for a different process once the underlying process is dead.

Changed in kernel 5.5: if the process referenced by the pidfd is dead, this function returns -1 (prior to 5.5, it returned whatever pid number the process had prior to its death).

Note that this is not a direct system call; instead it opens /proc/self/fdinfo/$pidfd and parses the Pid: line therein.

poll / select / epoll

Available since: kernel 5.3, glibc 2.2.5 (poll / select) or glibc 2.3.2 (epoll).

These functions can be used to asynchronously monitor a pidfd. They will report the pidfd as readable iff the underlying process is a zombie or is dead. Note however that read on a pidfd always fails; to get the exit code / status of the process, use waitid (possibly with WNOHANG).

waitid with P_PIDFD mode

Available since: kernel 5.4, glibc 2.36 (or glibc 2.2.5 if you provide your own definition of P_PIDFD; its value is 3).

waitid(P_PIDFD, fd, infop, options) is identical to waitid(P_PID, pidfd_getpid(fd), infop, options), except for the following:

In particular, note that:

The above points are true for all waitid calls, including P_PIDFD calls. The first time a zombie is waited upon (by any kind of wait / waitpid / waitid call), then the exit code / status is retreived, and subsequent attempts to wait upon it (again by any kind of wait / waitpid / waitid call) will fail.

When a process transitions from alive to zombie, if that process's parent's SIGCHLD handler is SIG_IGN or has SA_NOCLDWAIT, then the kernel does an automatic wait call on behalf of the parent and discards the result, thereby transitioning the child onward from zombie to dead. This causes all attempts to wait upon the child (including via P_PIDFD) to fail. The only exception to this is if the child was created with clone or clone3, and the termination signal was specified as something other than SIGCHLD, and the child has not called execve or similar: given this combination of circumstances, the automatic wait call will not recognise the child.

pidfd_getfd

Available since: kernel 5.6, glibc 2.36.

This function takes a pidfd, along with an fd number in the file table of the process referenced by the pidfd, creates a duplicate of that file descriptor in the file table of the calling process, and returns the new fd number. The effect is similar to what would happen if the referenced process used an SCM_RIGHTS message to send a file descriptor to the calling process. The O_CLOEXEC flag is automatically set on the new fd.

Calling this function incurs a PTRACE_MODE_ATTACH_REALCREDS security check.

setns

Available since: kernel 5.8, glibc 2.14.

Passing a pidfd to this function moves the caller into one or more of the namespaces that the process referenced by the pidfd is in. Note that this function can also be passed the result of open("/proc/$pid/ns/$name") as an fd.

process_madvise

Available since: kernel 5.10, glibc 2.36.

This function is similar to madvise, except that it operates on an arbitrary process (specified via a pidfd) rather than on the calling process.

Since 5.12, calling this function incurs PTRACE_MODE_READ_FSCREDS and CAP_SYS_NICE security checks. In 5.10 and 5.11, it incurred a PTRACE_MODE_ATTACH_FSCREDS security check.

process_mrelease

Available since: kernel 5.15, glibc 2.36.

This is a relatively niche function, which you are unlikely to ever need unless writing a userspace OOM killer. It can be called against a process which is no longer alive, but hasn't yet had its virtual memory released up by the kernel, to cause the kernel to release said virtual memory faster.

fstat / statx for meaningful stx_ino

Available since: kernel 6.9, glibc 2.2.5 (fstat) or glibc 2.28 (statx).

It has always been possible to call fstat or statx on a pidfd, but prior to kernel 6.9, it was not useful to do so. Since 6.9, calling statx on a pidfd gives a meaningful stx_ino: the 64-bit inode number of a pidfd uniquely identifies a process, so two pidfds referencing the same process will have identical stx_ino values, while two pidfds referencing different processes will have different stx_ino values. The same is true for fstat, provided that st_ino is 64 bits wide. In other words, since 6.9, a process's inode number (as observed via a pidfd) is a unique 64-bit identifier for the process, which is never reused (until the system is restarted), and is unique even across different pid namespaces.


It is likely that future kernel versions will add more things that can be done with (or to) a pidfd. As for the existing functionality, if you find yourself constrained by glibc version rather than kernel version, one option is to compile against a very recent glibc, then use polyfill-glibc to restore runtime compatibility with an older version of glibc.

In terms of future directions, some of the things that I'd like to see are:

C23 stdbit.h quick reference

MacrosImplementation
#define __STDC_ENDIAN_LITTLE__Some integer constant
#define __STDC_ENDIAN_BIG__Some integer constant
#define __STDC_ENDIAN_NATIVE____STDC_ENDIAN_LITTLE__ or
__STDC_ENDIAN_BIG__ (†)
Regular functionsImplementation
unsigned stdc_leading_zeros(T x)lzcnt(x)
unsigned stdc_first_leading_one(T x)x ? lzcnt(x) + 1 : 0
unsigned stdc_trailing_zeros(T x)tzcnt(x)
unsigned stdc_first_trailing_one(T x)x ? tzcnt(x) + 1 : 0
unsigned stdc_count_ones(T x)popcnt(x)
bool stdc_has_single_bit(T x)popcnt(x) == 1
unsigned stdc_bit_width(T x)x ? floor(log2(x)) + 1 : 0
T stdc_bit_floor(T x)x ? (T)1 << floor(log2(x)) : 0
T stdc_bit_ceil(T x)x ? (T)1 << ceil(log2(x)) : 1 (‡)
Inverted functionsImplementation
unsigned stdc_leading_ones(T x)lzcnt((T)~x)
unsigned stdc_first_leading_zero(T x)(T)~x ? lzcnt((T)~x) + 1 : 0
unsigned stdc_trailing_ones(T x)tzcnt((T)~x)
unsigned stdc_first_trailing_zero(T x)(T)~x ? tzcnt((T)~x) + 1 : 0
unsigned stdc_count_zeros(T x)popcnt((T)~x)

(†) Or some third value if the execution environment is neither little endian nor big endian.

(‡) Undefined if the << overflows or if the result does not fit in T.

Where:

The inverted functions can all be implemented by inverting the input and then passing it to one of the regular functions.

The stdc_first_leading_ functions are slightly slippery, and require a careful reading of the standard. For example, stdc_first_leading_one is defined as:

Returns the most significant index of the first 1 bit in value, plus 1. If it is not found, this function returns 0.

In turn, most significant index has the following unintuitive definition:

The most significant index is the 0-based index counting from the most significant bit, 0, to the least significant bit, w − 1, where w is the width of the type that is having its most significant index computed.

The initial patches for these functions in musl got this wrong, instead using the more intuitive definition of most significant index.

page: 1 2 3