Adding upstream version 18.2.2.upstream/18.2.2

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

1 files changed, 892 insertions, 0 deletions

diff --git a/src/erasure-code/clay/ErasureCodeClay.cc b/src/erasure-code/clay/ErasureCodeClay.cc
new file mode 100644
index 000000000..1e5a517de
--- /dev/null
+++ b/src/erasure-code/clay/ErasureCodeClay.cc
@@ -0,0 +1,892 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2018 Indian Institute of Science <office.ece@iisc.ac.in>
+ *
+ * Author: Myna Vajha <mynaramana@gmail.com>
+ *
+ *  This library is free software; you can redistribute it and/or
+ *  modify it under the terms of the GNU Lesser General Public
+ *  License as published by the Free Software Foundation; either
+ *  version 2.1 of the License, or (at your option) any later version.
+ *
+ */
+
+#include <errno.h>
+#include <algorithm>
+
+#include "ErasureCodeClay.h"
+
+#include "common/debug.h"
+#include "erasure-code/ErasureCodePlugin.h"
+#include "include/ceph_assert.h"
+#include "include/str_map.h"
+#include "include/stringify.h"
+#include "osd/osd_types.h"
+
+
+#define dout_context g_ceph_context
+#define dout_subsys ceph_subsys_osd
+#undef dout_prefix
+#define dout_prefix _prefix(_dout)
+
+#define LARGEST_VECTOR_WORDSIZE 16
+#define talloc(type, num) (type *) malloc(sizeof(type)*(num))
+
+using namespace std;
+using namespace ceph;
+static ostream& _prefix(std::ostream* _dout)
+{
+  return *_dout << "ErasureCodeClay: ";
+}
+
+static int pow_int(int a, int x) {
+  int power = 1;
+  while (x) {
+    if (x & 1) power *= a;
+    x /= 2;
+    a *= a;
+  }
+  return power;
+}
+
+ErasureCodeClay::~ErasureCodeClay()
+{
+  for (int i = 0; i < q*t; i++) {
+    if (U_buf[i].length() != 0) U_buf[i].clear();
+  }
+}
+
+int ErasureCodeClay::init(ErasureCodeProfile &profile,
+			  ostream *ss)
+{
+  int r;
+  r = parse(profile, ss);
+  if (r)
+    return r;
+
+  r = ErasureCode::init(profile, ss);
+  if (r)
+    return r;
+  ErasureCodePluginRegistry &registry = ErasureCodePluginRegistry::instance();
+  r = registry.factory(mds.profile["plugin"],
+		       directory,
+		       mds.profile,
+		       &mds.erasure_code,
+		       ss);
+  if (r)
+    return r;
+  r = registry.factory(pft.profile["plugin"],
+		       directory,
+		       pft.profile,
+		       &pft.erasure_code,
+		       ss);
+  return r;
+
+}
+
+unsigned int ErasureCodeClay::get_chunk_size(unsigned int object_size) const
+{
+  unsigned int alignment_scalar_code = pft.erasure_code->get_chunk_size(1);
+  unsigned int alignment = sub_chunk_no * k * alignment_scalar_code;
+  
+  return round_up_to(object_size, alignment) / k;
+}
+
+int ErasureCodeClay::minimum_to_decode(const set<int> &want_to_read,
+				       const set<int> &available,
+				       map<int, vector<pair<int, int>>> *minimum)
+{
+  if (is_repair(want_to_read, available)) {
+    return minimum_to_repair(want_to_read, available, minimum);
+  } else {
+    return ErasureCode::minimum_to_decode(want_to_read, available, minimum);
+  }
+}
+
+int ErasureCodeClay::decode(const set<int> &want_to_read,
+			    const map<int, bufferlist> &chunks,
+			    map<int, bufferlist> *decoded, int chunk_size)
+{
+  set<int> avail;
+  for ([[maybe_unused]] auto& [node, bl] : chunks) {
+    avail.insert(node);
+    (void)bl;  // silence -Wunused-variable
+  }
+
+  if (is_repair(want_to_read, avail) && 
+         ((unsigned int)chunk_size > chunks.begin()->second.length())) {
+    return repair(want_to_read, chunks, decoded, chunk_size);
+  } else {
+    return ErasureCode::_decode(want_to_read, chunks, decoded);
+  }
+}
+
+void p(const set<int> &s) { cerr << s; } // for gdb
+
+int ErasureCodeClay::encode_chunks(const set<int> &want_to_encode,
+				   map<int, bufferlist> *encoded)
+{
+  map<int, bufferlist> chunks;
+  set<int> parity_chunks;
+  int chunk_size = (*encoded)[0].length();
+
+  for (int i = 0; i < k + m; i++) {
+    if (i < k) {
+      chunks[i] = (*encoded)[i];
+    } else {
+      chunks[i+nu] = (*encoded)[i];
+      parity_chunks.insert(i+nu);
+    }
+  }
+
+  for (int i = k; i < k + nu; i++) {
+    bufferptr buf(buffer::create_aligned(chunk_size, SIMD_ALIGN));
+    buf.zero();
+    chunks[i].push_back(std::move(buf));
+  }
+
+  int res = decode_layered(parity_chunks, &chunks);
+  for (int i = k ; i < k + nu; i++) {
+    // need to clean some of the intermediate chunks here!!
+    chunks[i].clear();
+  }
+  return res;
+}
+
+int ErasureCodeClay::decode_chunks(const set<int> &want_to_read,
+				   const map<int, bufferlist> &chunks,
+				   map<int, bufferlist> *decoded)
+{
+  set<int> erasures;
+  map<int, bufferlist> coded_chunks;
+
+  for (int i = 0; i < k + m; i++) {
+    if (chunks.count(i) == 0) {
+      erasures.insert(i < k ? i : i+nu);
+    }
+    ceph_assert(decoded->count(i) > 0);
+    coded_chunks[i < k ? i : i+nu] = (*decoded)[i];
+  }
+  int chunk_size = coded_chunks[0].length();
+
+  for (int i = k; i < k+nu; i++) {
+    bufferptr buf(buffer::create_aligned(chunk_size, SIMD_ALIGN));
+    buf.zero();
+    coded_chunks[i].push_back(std::move(buf));
+  }
+
+  int res = decode_layered(erasures, &coded_chunks);
+  for (int i = k; i < k+nu; i++) {
+    coded_chunks[i].clear();
+  }
+  return res;
+}
+
+int ErasureCodeClay::parse(ErasureCodeProfile &profile,
+			   ostream *ss)
+{
+  int err = 0;
+  err = ErasureCode::parse(profile, ss);
+  err |= to_int("k", profile, &k, DEFAULT_K, ss);
+  err |= to_int("m", profile, &m, DEFAULT_M, ss);
+
+  err |= sanity_check_k_m(k, m, ss);
+
+  err |= to_int("d", profile, &d, std::to_string(k+m-1), ss);
+
+  // check for scalar_mds in profile input
+  if (profile.find("scalar_mds") == profile.end() ||
+      profile.find("scalar_mds")->second.empty()) {
+    mds.profile["plugin"] = "jerasure";
+    pft.profile["plugin"] = "jerasure";
+  } else {
+    std::string p = profile.find("scalar_mds")->second;
+    if ((p == "jerasure") || (p == "isa") || (p == "shec")) {
+      mds.profile["plugin"] = p;
+      pft.profile["plugin"] = p;
+    } else {
+        *ss << "scalar_mds " << mds.profile["plugin"] <<
+               "is not currently supported, use one of 'jerasure',"<<
+               " 'isa', 'shec'" << std::endl;
+        err = -EINVAL;
+        return err;
+    }
+  }
+
+  if (profile.find("technique") == profile.end() ||
+      profile.find("technique")->second.empty()) {
+    if ((mds.profile["plugin"]=="jerasure") || (mds.profile["plugin"]=="isa") ) {
+      mds.profile["technique"] = "reed_sol_van";
+      pft.profile["technique"] = "reed_sol_van";
+    } else {
+      mds.profile["technique"] = "single";
+      pft.profile["technique"] = "single";
+    }
+  } else {
+    std::string p = profile.find("technique")->second;
+    if (mds.profile["plugin"] == "jerasure") {
+      if ( (p == "reed_sol_van") || (p == "reed_sol_r6_op") || (p == "cauchy_orig")
+           || (p == "cauchy_good") || (p == "liber8tion")) {
+        mds.profile["technique"] = p;
+        pft.profile["technique"] = p;
+      } else {
+        *ss << "technique " << p << "is not currently supported, use one of "
+	    << "reed_sol_van', 'reed_sol_r6_op','cauchy_orig',"
+	    << "'cauchy_good','liber8tion'"<< std::endl;
+        err = -EINVAL;
+        return err;
+      }
+    } else if (mds.profile["plugin"] == "isa") {
+      if ( (p == "reed_sol_van") || (p == "cauchy")) {
+        mds.profile["technique"] = p;
+        pft.profile["technique"] = p;
+      } else {
+        *ss << "technique " << p << "is not currently supported, use one of"
+	    << "'reed_sol_van','cauchy'"<< std::endl;
+        err = -EINVAL;
+        return err;
+      }
+    } else {
+      if ( (p == "single") || (p == "multiple")) {
+        mds.profile["technique"] = p;
+        pft.profile["technique"] = p;
+      } else {
+        *ss << "technique " << p << "is not currently supported, use one of"<<
+               "'single','multiple'"<< std::endl;
+        err = -EINVAL;
+        return err;
+      }
+    }
+  }
+  if ((d < k) || (d > k + m - 1)) {
+    *ss << "value of d " << d
+        << " must be within [ " << k << "," << k+m-1 << "]" << std::endl;
+    err = -EINVAL;
+    return err;
+  }
+
+  q = d - k + 1;
+  if ((k + m) % q) {
+    nu = q - (k + m) % q;
+  } else {
+    nu = 0;
+  }
+
+  if (k+m+nu > 254) {
+    err = -EINVAL;
+    return err;
+  }
+
+  if (mds.profile["plugin"] == "shec") {
+    mds.profile["c"] = '2';
+    pft.profile["c"] = '2';
+  }
+  mds.profile["k"] = std::to_string(k+nu);
+  mds.profile["m"] = std::to_string(m);
+  mds.profile["w"] = '8';
+
+  pft.profile["k"] = '2';
+  pft.profile["m"] = '2';
+  pft.profile["w"] = '8';
+
+  t = (k + m + nu) / q;
+  sub_chunk_no = pow_int(q, t);
+
+  dout(10) << __func__
+	   << " (q,t,nu)=(" << q << "," << t << "," << nu <<")" << dendl;
+
+  return err;
+}
+
+int ErasureCodeClay::is_repair(const set<int> &want_to_read,
+			       const set<int> &available_chunks) {
+
+  if (includes(available_chunks.begin(), available_chunks.end(),
+               want_to_read.begin(), want_to_read.end())) return 0;
+  if (want_to_read.size() > 1) return 0;
+
+  int i = *want_to_read.begin();
+  int lost_node_id = (i < k) ? i: i+nu;
+  for (int x = 0; x < q; x++) {
+    int node = (lost_node_id/q)*q+x;
+    node = (node < k) ? node : node-nu;
+    if (node != i) { // node in the same group other than erased node
+      if (available_chunks.count(node) == 0) return 0;
+    }
+  }
+
+  if (available_chunks.size() < (unsigned)d) return 0;
+  return 1;
+}
+
+int ErasureCodeClay::minimum_to_repair(const set<int> &want_to_read,
+				       const set<int> &available_chunks,
+				       map<int, vector<pair<int, int>>> *minimum)
+{
+  int i = *want_to_read.begin();
+  int lost_node_index = (i < k) ? i : i+nu;
+  int rep_node_index = 0;
+
+  // add all the nodes in lost node's y column.
+  vector<pair<int, int>> sub_chunk_ind;
+  get_repair_subchunks(lost_node_index, sub_chunk_ind);
+  if ((available_chunks.size() >= (unsigned)d)) {
+    for (int j = 0; j < q; j++) {
+      if (j != lost_node_index%q) {
+        rep_node_index = (lost_node_index/q)*q+j;
+        if (rep_node_index < k) {
+          minimum->insert(make_pair(rep_node_index, sub_chunk_ind));
+        } else if (rep_node_index >= k+nu) {
+          minimum->insert(make_pair(rep_node_index-nu, sub_chunk_ind));
+        }
+      }
+    }
+    for (auto chunk : available_chunks) {
+      if (minimum->size() >= (unsigned)d) {
+	break;
+      }
+      if (!minimum->count(chunk)) {
+	minimum->emplace(chunk, sub_chunk_ind);
+      }
+    }
+  } else {
+    dout(0) << "minimum_to_repair: shouldn't have come here" << dendl;
+    ceph_assert(0);
+  }
+  ceph_assert(minimum->size() == (unsigned)d);
+  return 0;
+}
+
+void ErasureCodeClay::get_repair_subchunks(const int &lost_node,
+					   vector<pair<int, int>> &repair_sub_chunks_ind)
+{
+  const int y_lost = lost_node / q;
+  const int x_lost = lost_node % q;
+
+  const int seq_sc_count = pow_int(q, t-1-y_lost);
+  const int num_seq = pow_int(q, y_lost);
+
+  int index = x_lost * seq_sc_count;
+  for (int ind_seq = 0; ind_seq < num_seq; ind_seq++) {
+    repair_sub_chunks_ind.push_back(make_pair(index, seq_sc_count));
+    index += q * seq_sc_count;
+  }
+}
+
+int ErasureCodeClay::get_repair_sub_chunk_count(const set<int> &want_to_read)
+{
+  int weight_vector[t];
+  std::fill(weight_vector, weight_vector + t, 0);
+  for (auto to_read : want_to_read) {
+    weight_vector[to_read / q]++;
+  }
+
+  int repair_subchunks_count = 1;
+  for (int y = 0; y < t; y++) {
+    repair_subchunks_count = repair_subchunks_count*(q-weight_vector[y]);
+  }
+
+  return sub_chunk_no - repair_subchunks_count;
+}
+
+int ErasureCodeClay::repair(const set<int> &want_to_read,
+			    const map<int, bufferlist> &chunks,
+			    map<int, bufferlist> *repaired, int chunk_size)
+{
+
+  ceph_assert((want_to_read.size() == 1) && (chunks.size() == (unsigned)d));
+
+  int repair_sub_chunk_no = get_repair_sub_chunk_count(want_to_read);
+  vector<pair<int, int>> repair_sub_chunks_ind;
+
+  unsigned repair_blocksize = chunks.begin()->second.length();
+  assert(repair_blocksize%repair_sub_chunk_no == 0);
+
+  unsigned sub_chunksize = repair_blocksize/repair_sub_chunk_no;
+  unsigned chunksize = sub_chunk_no*sub_chunksize;
+
+  ceph_assert(chunksize == (unsigned)chunk_size);
+
+  map<int, bufferlist> recovered_data;
+  map<int, bufferlist> helper_data;
+  set<int> aloof_nodes;
+
+  for (int i =  0; i < k + m; i++) {
+    // included helper data only for d+nu nodes.
+    if (auto found = chunks.find(i); found != chunks.end()) { // i is a helper
+      if (i<k) {
+	helper_data[i] = found->second;
+      } else {
+	helper_data[i+nu] = found->second;
+      }
+    } else {
+      if (i != *want_to_read.begin()) { // aloof node case.
+        int aloof_node_id = (i < k) ? i: i+nu;
+        aloof_nodes.insert(aloof_node_id);
+      } else {
+        bufferptr ptr(buffer::create_aligned(chunksize, SIMD_ALIGN));
+	ptr.zero();
+        int lost_node_id = (i < k) ? i : i+nu;
+        (*repaired)[i].push_back(ptr);
+        recovered_data[lost_node_id] = (*repaired)[i];
+        get_repair_subchunks(lost_node_id, repair_sub_chunks_ind);
+      }
+    }
+  }
+
+  // this is for shortened codes i.e., when nu > 0
+  for (int i=k; i < k+nu; i++) {
+    bufferptr ptr(buffer::create_aligned(repair_blocksize, SIMD_ALIGN));
+    ptr.zero();
+    helper_data[i].push_back(ptr);
+  }
+
+  ceph_assert(helper_data.size()+aloof_nodes.size()+recovered_data.size() ==
+	      (unsigned) q*t);
+
+  int r = repair_one_lost_chunk(recovered_data, aloof_nodes,
+				helper_data, repair_blocksize,
+				repair_sub_chunks_ind);
+
+  // clear buffers created for the purpose of shortening
+  for (int i = k; i < k+nu; i++) {
+    helper_data[i].clear();
+  }
+
+  return r;
+}
+
+int ErasureCodeClay::repair_one_lost_chunk(map<int, bufferlist> &recovered_data,
+					   set<int> &aloof_nodes,
+					   map<int, bufferlist> &helper_data,
+					   int repair_blocksize,
+					   vector<pair<int,int>> &repair_sub_chunks_ind)
+{
+  unsigned repair_subchunks = (unsigned)sub_chunk_no / q;
+  unsigned sub_chunksize = repair_blocksize / repair_subchunks;
+
+  int z_vec[t];
+  map<int, set<int> > ordered_planes;
+  map<int, int> repair_plane_to_ind;
+  int count_retrieved_sub_chunks = 0;
+  int plane_ind = 0;
+
+  bufferptr buf(buffer::create_aligned(sub_chunksize, SIMD_ALIGN));
+  bufferlist temp_buf;
+  temp_buf.push_back(buf);
+
+  for (auto [index,count] : repair_sub_chunks_ind) {
+    for (int j = index; j < index + count; j++) {
+      get_plane_vector(j, z_vec);
+      int order = 0;
+      // check across all erasures and aloof nodes
+      for ([[maybe_unused]] auto& [node, bl] : recovered_data) {
+        if (node % q == z_vec[node / q]) order++;
+        (void)bl;  // silence -Wunused-variable
+      }
+      for (auto node : aloof_nodes) {
+        if (node % q == z_vec[node / q]) order++;
+      }
+      ceph_assert(order > 0);
+      ordered_planes[order].insert(j);
+      // to keep track of a sub chunk within helper buffer recieved
+      repair_plane_to_ind[j] = plane_ind;
+      plane_ind++;
+    }
+  }
+  assert((unsigned)plane_ind == repair_subchunks);
+
+  for (int i = 0; i < q*t; i++) {
+    if (U_buf[i].length() == 0) {
+      bufferptr buf(buffer::create_aligned(sub_chunk_no*sub_chunksize, SIMD_ALIGN));
+      buf.zero();
+      U_buf[i].push_back(std::move(buf));
+    }
+  }
+
+  int lost_chunk;
+  int count = 0;
+  for ([[maybe_unused]] auto& [node, bl] : recovered_data) {
+    lost_chunk = node;
+    count++;
+    (void)bl;  // silence -Wunused-variable
+  }
+  ceph_assert(count == 1);
+
+  set<int> erasures;
+  for (int i = 0; i < q; i++) {
+    erasures.insert(lost_chunk - lost_chunk % q + i);
+  }
+  for (auto node : aloof_nodes) {
+    erasures.insert(node);
+  }
+
+  for (int order = 1; ;order++) {
+    if (ordered_planes.count(order) == 0) {
+      break;
+    }
+    for (auto z : ordered_planes[order]) {
+      get_plane_vector(z, z_vec);
+
+      for (int y = 0; y < t; y++) {
+	for (int x = 0; x < q; x++) {
+	  int node_xy = y*q + x;
+	  map<int, bufferlist> known_subchunks;
+	  map<int, bufferlist> pftsubchunks;
+	  set<int> pft_erasures;
+	  if (erasures.count(node_xy) == 0) {
+	    assert(helper_data.count(node_xy) > 0);
+	    int z_sw = z + (x - z_vec[y])*pow_int(q,t-1-y);
+	    int node_sw = y*q + z_vec[y];
+	    int i0 = 0, i1 = 1, i2 = 2, i3 = 3;
+	    if (z_vec[y] > x) {
+	      i0 = 1;
+	      i1 = 0;
+	      i2 = 3;
+	      i3 = 2;
+	    }
+	    if (aloof_nodes.count(node_sw) > 0) {
+	      assert(repair_plane_to_ind.count(z) > 0);
+	      assert(repair_plane_to_ind.count(z_sw) > 0);
+	      pft_erasures.insert(i2);
+	      known_subchunks[i0].substr_of(helper_data[node_xy], repair_plane_to_ind[z]*sub_chunksize, sub_chunksize);
+	      known_subchunks[i3].substr_of(U_buf[node_sw], z_sw*sub_chunksize, sub_chunksize);
+	      pftsubchunks[i0] = known_subchunks[i0];
+	      pftsubchunks[i1] = temp_buf;
+	      pftsubchunks[i2].substr_of(U_buf[node_xy], z*sub_chunksize, sub_chunksize);
+	      pftsubchunks[i3] = known_subchunks[i3];
+	      for (int i=0; i<3; i++) {
+		pftsubchunks[i].rebuild_aligned(SIMD_ALIGN);
+	      }
+	      pft.erasure_code->decode_chunks(pft_erasures, known_subchunks, &pftsubchunks);
+	    } else {
+	      ceph_assert(helper_data.count(node_sw) > 0);
+	      ceph_assert(repair_plane_to_ind.count(z) > 0);
+	      if (z_vec[y] != x){
+		pft_erasures.insert(i2);
+		ceph_assert(repair_plane_to_ind.count(z_sw) > 0);
+		known_subchunks[i0].substr_of(helper_data[node_xy], repair_plane_to_ind[z]*sub_chunksize, sub_chunksize);
+		known_subchunks[i1].substr_of(helper_data[node_sw], repair_plane_to_ind[z_sw]*sub_chunksize, sub_chunksize);
+		pftsubchunks[i0] = known_subchunks[i0];
+		pftsubchunks[i1] = known_subchunks[i1];
+		pftsubchunks[i2].substr_of(U_buf[node_xy], z*sub_chunksize, sub_chunksize);
+		pftsubchunks[i3].substr_of(temp_buf, 0, sub_chunksize);
+		for (int i=0; i<3; i++) {
+		  pftsubchunks[i].rebuild_aligned(SIMD_ALIGN);
+		}
+		pft.erasure_code->decode_chunks(pft_erasures, known_subchunks, &pftsubchunks);
+	      } else {
+		char* uncoupled_chunk = U_buf[node_xy].c_str();
+		char* coupled_chunk = helper_data[node_xy].c_str();
+		memcpy(&uncoupled_chunk[z*sub_chunksize],
+		       &coupled_chunk[repair_plane_to_ind[z]*sub_chunksize],
+		       sub_chunksize);
+	      }
+	    }
+	  }
+	} // x
+      } // y
+      ceph_assert(erasures.size() <= (unsigned)m);
+      decode_uncoupled(erasures, z, sub_chunksize);
+
+      for (auto i : erasures) {
+	int x = i % q;
+	int y = i / q;
+	int node_sw = y*q+z_vec[y];
+	int z_sw = z + (x - z_vec[y]) * pow_int(q,t-1-y);
+	set<int> pft_erasures;
+	map<int, bufferlist> known_subchunks;
+	map<int, bufferlist> pftsubchunks;
+	int i0 = 0, i1 = 1, i2 = 2, i3 = 3;
+	if (z_vec[y] > x) {
+	  i0 = 1;
+	  i1 = 0;
+	  i2 = 3;
+	  i3 = 2;
+	}
+	// make sure it is not an aloof node before you retrieve repaired_data
+	if (aloof_nodes.count(i) == 0) {
+	  if (x == z_vec[y]) { // hole-dot pair (type 0)
+	    char* coupled_chunk = recovered_data[i].c_str();
+	    char* uncoupled_chunk = U_buf[i].c_str();
+	    memcpy(&coupled_chunk[z*sub_chunksize],
+		   &uncoupled_chunk[z*sub_chunksize],
+		   sub_chunksize);
+	    count_retrieved_sub_chunks++;
+	  } else {
+	    ceph_assert(y == lost_chunk / q);
+	    ceph_assert(node_sw == lost_chunk);
+	    ceph_assert(helper_data.count(i) > 0);
+	    pft_erasures.insert(i1);
+	    known_subchunks[i0].substr_of(helper_data[i], repair_plane_to_ind[z]*sub_chunksize, sub_chunksize);
+	    known_subchunks[i2].substr_of(U_buf[i], z*sub_chunksize, sub_chunksize);
+
+	    pftsubchunks[i0] = known_subchunks[i0];
+	    pftsubchunks[i1].substr_of(recovered_data[node_sw], z_sw*sub_chunksize, sub_chunksize);
+	    pftsubchunks[i2] = known_subchunks[i2];
+	    pftsubchunks[i3] = temp_buf;
+	    for (int i=0; i<3; i++) {
+	      pftsubchunks[i].rebuild_aligned(SIMD_ALIGN);
+	    }
+	    pft.erasure_code->decode_chunks(pft_erasures, known_subchunks, &pftsubchunks);
+	  }
+	}
+      } // recover all erasures
+    } // planes of particular order
+  } // order
+
+  return 0;
+}
+
+
+int ErasureCodeClay::decode_layered(set<int> &erased_chunks,
+                                    map<int, bufferlist> *chunks)
+{
+  int num_erasures = erased_chunks.size();
+
+  int size = (*chunks)[0].length();
+  ceph_assert(size%sub_chunk_no == 0);
+  int sc_size = size / sub_chunk_no;
+
+  ceph_assert(num_erasures > 0);
+
+  for (int i = k+nu; (num_erasures < m) && (i < q*t); i++) {
+    if ([[maybe_unused]] auto [it, added] = erased_chunks.emplace(i); added) {
+      num_erasures++;
+      (void)it;  // silence -Wunused-variable
+    }
+  }
+  ceph_assert(num_erasures == m);
+
+  int max_iscore = get_max_iscore(erased_chunks);
+  int order[sub_chunk_no];
+  int z_vec[t];
+  for (int i = 0; i < q*t; i++) {
+    if (U_buf[i].length() == 0) {
+      bufferptr buf(buffer::create_aligned(size, SIMD_ALIGN));
+      buf.zero();
+      U_buf[i].push_back(std::move(buf));
+    }
+  }
+
+  set_planes_sequential_decoding_order(order, erased_chunks);
+
+  for (int iscore = 0; iscore <= max_iscore; iscore++) {
+   for (int z = 0; z < sub_chunk_no; z++) {
+      if (order[z] == iscore) {
+        decode_erasures(erased_chunks, z, chunks, sc_size);
+      }
+    }
+
+    for (int z = 0; z < sub_chunk_no; z++) {
+      if (order[z] == iscore) {
+	get_plane_vector(z, z_vec);
+        for (auto node_xy : erased_chunks) {
+          int x = node_xy % q;
+          int y = node_xy / q;
+	  int node_sw = y*q+z_vec[y];
+          if (z_vec[y] != x) {
+            if (erased_chunks.count(node_sw) == 0) {
+	      recover_type1_erasure(chunks, x, y, z, z_vec, sc_size);
+	    } else if (z_vec[y] < x){
+              ceph_assert(erased_chunks.count(node_sw) > 0);
+              ceph_assert(z_vec[y] != x);
+              get_coupled_from_uncoupled(chunks, x, y, z, z_vec, sc_size);
+	    }
+	  } else {
+	    char* C = (*chunks)[node_xy].c_str();
+            char* U = U_buf[node_xy].c_str();
+            memcpy(&C[z*sc_size], &U[z*sc_size], sc_size);
+          }
+        }
+      }
+    } // plane
+  } // iscore, order
+
+  return 0;
+}
+
+int ErasureCodeClay::decode_erasures(const set<int>& erased_chunks, int z,
+				     map<int, bufferlist>* chunks, int sc_size)
+{
+  int z_vec[t];
+
+  get_plane_vector(z, z_vec);
+
+  for (int x = 0; x < q; x++) {
+    for (int y = 0; y < t; y++) {
+      int node_xy = q*y+x;
+      int node_sw = q*y+z_vec[y];
+      if (erased_chunks.count(node_xy) == 0) {
+	if (z_vec[y] < x) {
+	  get_uncoupled_from_coupled(chunks, x, y, z, z_vec, sc_size);
+	} else if (z_vec[y] == x) {
+	  char* uncoupled_chunk = U_buf[node_xy].c_str();
+	  char* coupled_chunk = (*chunks)[node_xy].c_str();
+          memcpy(&uncoupled_chunk[z*sc_size], &coupled_chunk[z*sc_size], sc_size);
+        } else {
+          if (erased_chunks.count(node_sw) > 0) {
+            get_uncoupled_from_coupled(chunks, x, y, z, z_vec, sc_size);
+          }
+        }
+      }
+    }
+  }
+  return decode_uncoupled(erased_chunks, z, sc_size);
+}
+
+int ErasureCodeClay::decode_uncoupled(const set<int>& erased_chunks, int z, int sc_size)
+{
+  map<int, bufferlist> known_subchunks;
+  map<int, bufferlist> all_subchunks;
+
+  for (int i = 0; i < q*t; i++) {
+    if (erased_chunks.count(i) == 0) {
+      known_subchunks[i].substr_of(U_buf[i], z*sc_size, sc_size);
+      all_subchunks[i] = known_subchunks[i];
+    } else {
+      all_subchunks[i].substr_of(U_buf[i], z*sc_size, sc_size);
+    }
+    all_subchunks[i].rebuild_aligned_size_and_memory(sc_size, SIMD_ALIGN);
+    assert(all_subchunks[i].is_contiguous());
+  }
+
+  mds.erasure_code->decode_chunks(erased_chunks, known_subchunks, &all_subchunks);
+  return 0;
+}
+
+void ErasureCodeClay::set_planes_sequential_decoding_order(int* order, set<int>& erasures) {
+  int z_vec[t];
+  for (int z = 0; z < sub_chunk_no; z++) {
+    get_plane_vector(z,z_vec);
+    order[z] = 0;
+    for (auto i : erasures) {
+      if (i % q == z_vec[i / q]) {
+	order[z] = order[z] + 1;
+      }
+    }
+  }
+}
+
+void ErasureCodeClay::recover_type1_erasure(map<int, bufferlist>* chunks,
+					    int x, int y, int z,
+					    int* z_vec, int sc_size)
+{
+  set<int> erased_chunks;
+
+  int node_xy = y*q+x;
+  int node_sw = y*q+z_vec[y];
+  int z_sw = z + (x - z_vec[y]) * pow_int(q,t-1-y);
+
+  map<int, bufferlist> known_subchunks;
+  map<int, bufferlist> pftsubchunks;
+  bufferptr ptr(buffer::create_aligned(sc_size, SIMD_ALIGN));
+  ptr.zero();
+
+  int i0 = 0, i1 = 1, i2 = 2, i3 = 3;
+  if (z_vec[y] > x) {
+    i0 = 1;
+    i1 = 0;
+    i2 = 3;
+    i3 = 2;
+  }
+
+  erased_chunks.insert(i0);
+  pftsubchunks[i0].substr_of((*chunks)[node_xy], z * sc_size, sc_size);
+  known_subchunks[i1].substr_of((*chunks)[node_sw], z_sw * sc_size, sc_size);
+  known_subchunks[i2].substr_of(U_buf[node_xy], z * sc_size, sc_size);
+  pftsubchunks[i1] = known_subchunks[i1];
+  pftsubchunks[i2] = known_subchunks[i2];
+  pftsubchunks[i3].push_back(ptr);
+
+  for (int i=0; i<3; i++) {
+    pftsubchunks[i].rebuild_aligned_size_and_memory(sc_size, SIMD_ALIGN);
+  }
+
+  pft.erasure_code->decode_chunks(erased_chunks, known_subchunks, &pftsubchunks);
+}
+
+void ErasureCodeClay::get_coupled_from_uncoupled(map<int, bufferlist>* chunks,
+						 int x, int y, int z,
+						 int* z_vec, int sc_size)
+{
+  set<int> erased_chunks = {0, 1};
+
+  int node_xy = y*q+x;
+  int node_sw = y*q+z_vec[y];
+  int z_sw = z + (x - z_vec[y]) * pow_int(q,t-1-y);
+
+  ceph_assert(z_vec[y] < x);
+  map<int, bufferlist> uncoupled_subchunks;
+  uncoupled_subchunks[2].substr_of(U_buf[node_xy], z * sc_size, sc_size);
+  uncoupled_subchunks[3].substr_of(U_buf[node_sw], z_sw * sc_size, sc_size);
+
+  map<int, bufferlist> pftsubchunks;
+  pftsubchunks[0].substr_of((*chunks)[node_xy], z * sc_size, sc_size);
+  pftsubchunks[1].substr_of((*chunks)[node_sw], z_sw * sc_size, sc_size);
+  pftsubchunks[2] = uncoupled_subchunks[2];
+  pftsubchunks[3] = uncoupled_subchunks[3];
+
+  for (int i=0; i<3; i++) {
+    pftsubchunks[i].rebuild_aligned_size_and_memory(sc_size, SIMD_ALIGN);
+  }
+  pft.erasure_code->decode_chunks(erased_chunks, uncoupled_subchunks, &pftsubchunks);
+}
+
+void ErasureCodeClay::get_uncoupled_from_coupled(map<int, bufferlist>* chunks,
+						 int x, int y, int z,
+						 int* z_vec, int sc_size)
+{
+  set<int> erased_chunks = {2, 3};
+
+  int node_xy = y*q+x;
+  int node_sw = y*q+z_vec[y];
+  int z_sw = z + (x - z_vec[y]) * pow_int(q,t-1-y);
+
+  int i0 = 0, i1 = 1, i2 = 2, i3 = 3;
+  if (z_vec[y] > x) {
+    i0 = 1;
+    i1 = 0;
+    i2 = 3;
+    i3 = 2;
+  }
+  map<int, bufferlist> coupled_subchunks;
+  coupled_subchunks[i0].substr_of((*chunks)[node_xy], z * sc_size, sc_size);
+  coupled_subchunks[i1].substr_of((*chunks)[node_sw], z_sw * sc_size, sc_size);
+
+  map<int, bufferlist> pftsubchunks;
+  pftsubchunks[0] = coupled_subchunks[0];
+  pftsubchunks[1] = coupled_subchunks[1];
+  pftsubchunks[i2].substr_of(U_buf[node_xy], z * sc_size, sc_size);
+  pftsubchunks[i3].substr_of(U_buf[node_sw], z_sw * sc_size, sc_size);
+  for (int i=0; i<3; i++) {
+    pftsubchunks[i].rebuild_aligned_size_and_memory(sc_size, SIMD_ALIGN);
+  }
+  pft.erasure_code->decode_chunks(erased_chunks, coupled_subchunks, &pftsubchunks);
+}
+
+int ErasureCodeClay::get_max_iscore(set<int>& erased_chunks)
+{
+  int weight_vec[t];
+  int iscore = 0;
+  memset(weight_vec, 0, sizeof(int)*t);
+
+  for (auto i : erased_chunks) {
+    if (weight_vec[i / q] == 0) {
+      weight_vec[i / q] = 1;
+      iscore++;
+    }
+  }
+  return iscore;
+}
+
+void ErasureCodeClay::get_plane_vector(int z, int* z_vec)
+{
+  for (int i = 0; i < t; i++) {
+    z_vec[t-1-i] = z % q;
+    z = (z - z_vec[t-1-i]) / q;
+  }
+}