diff --git a/libc/bionic/__libc_init_main_thread.cpp b/libc/bionic/__libc_init_main_thread.cpp
index 2b90c90..6279e65 100644
--- a/libc/bionic/__libc_init_main_thread.cpp
+++ b/libc/bionic/__libc_init_main_thread.cpp
@@ -126,6 +126,7 @@
   auto new_tcb = reinterpret_cast<bionic_tcb*>(mapping.static_tls + layout.offset_bionic_tcb());
   auto new_tls = reinterpret_cast<bionic_tls*>(mapping.static_tls + layout.offset_bionic_tls());
 
+  __init_static_tls(mapping.static_tls);
   new_tcb->copy_from_bootstrap(temp_tcb);
   new_tls->copy_from_bootstrap(temp_tls);
   __init_tcb(new_tcb, &main_thread);
diff --git a/libc/bionic/bionic_elf_tls.cpp b/libc/bionic/bionic_elf_tls.cpp
index 2d2cbe3..4253b97 100644
--- a/libc/bionic/bionic_elf_tls.cpp
+++ b/libc/bionic/bionic_elf_tls.cpp
@@ -29,33 +29,25 @@
 #include "private/bionic_elf_tls.h"
 
 #include <async_safe/log.h>
+#include <string.h>
 #include <sys/param.h>
 #include <unistd.h>
 
+#include "private/ScopedRWLock.h"
+#include "private/bionic_globals.h"
 #include "private/bionic_macros.h"
 #include "private/bionic_tls.h"
 
 // Search for a TLS segment in the given phdr table. Returns true if it has a
 // TLS segment and false otherwise.
 bool __bionic_get_tls_segment(const ElfW(Phdr)* phdr_table, size_t phdr_count,
-                              ElfW(Addr) load_bias, const char* mod_name,
-                              TlsSegment* out) {
+                              ElfW(Addr) load_bias, TlsSegment* out) {
   for (size_t i = 0; i < phdr_count; ++i) {
     const ElfW(Phdr)& phdr = phdr_table[i];
     if (phdr.p_type == PT_TLS) {
-      // N.B. The size does not need to be a multiple of the alignment. With
-      // ld.bfd (or after using binutils' strip), the TLS segment's size isn't
-      // rounded up.
-      size_t alignment = phdr.p_align;
-      if (alignment == 0 || !powerof2(alignment)) {
-        async_safe_fatal("error: \"%s\": TLS segment alignment is not a power of 2: %zu",
-                         mod_name, alignment);
-      }
-      // Bionic only respects TLS alignment up to one page.
-      alignment = MIN(alignment, PAGE_SIZE);
       *out = TlsSegment {
         phdr.p_memsz,
-        alignment,
+        phdr.p_align,
         reinterpret_cast<void*>(load_bias + phdr.p_vaddr),
         phdr.p_filesz,
       };
@@ -65,8 +57,75 @@
   return false;
 }
 
-void StaticTlsLayout::reserve_tcb() {
-  offset_bionic_tcb_ = reserve_type<bionic_tcb>();
+// Return true if the alignment of a TLS segment is a valid power-of-two. Also
+// cap the alignment if it's too high.
+bool __bionic_check_tls_alignment(size_t* alignment) {
+  // N.B. The size does not need to be a multiple of the alignment. With
+  // ld.bfd (or after using binutils' strip), the TLS segment's size isn't
+  // rounded up.
+  if (*alignment == 0 || !powerof2(*alignment)) {
+    return false;
+  }
+  // Bionic only respects TLS alignment up to one page.
+  *alignment = MIN(*alignment, PAGE_SIZE);
+  return true;
+}
+
+size_t StaticTlsLayout::offset_thread_pointer() const {
+  return offset_bionic_tcb_ + (-MIN_TLS_SLOT * sizeof(void*));
+}
+
+// Reserves space for the Bionic TCB and the executable's TLS segment. Returns
+// the offset of the executable's TLS segment.
+size_t StaticTlsLayout::reserve_exe_segment_and_tcb(const TlsSegment* exe_segment,
+                                                    const char* progname __attribute__((unused))) {
+  // Special case: if the executable has no TLS segment, then just allocate a
+  // TCB and skip the minimum alignment check on ARM.
+  if (exe_segment == nullptr) {
+    offset_bionic_tcb_ = reserve_type<bionic_tcb>();
+    return 0;
+  }
+
+#if defined(__arm__) || defined(__aarch64__)
+
+  // First reserve enough space for the TCB before the executable segment.
+  reserve(sizeof(bionic_tcb), 1);
+
+  // Then reserve the segment itself.
+  const size_t result = reserve(exe_segment->size, exe_segment->alignment);
+
+  // The variant 1 ABI that ARM linkers follow specifies a 2-word TCB between
+  // the thread pointer and the start of the executable's TLS segment, but both
+  // the thread pointer and the TLS segment are aligned appropriately for the
+  // TLS segment. Calculate the distance between the thread pointer and the
+  // EXE's segment.
+  const size_t exe_tpoff = __BIONIC_ALIGN(sizeof(void*) * 2, exe_segment->alignment);
+
+  const size_t min_bionic_alignment = BIONIC_ROUND_UP_POWER_OF_2(MAX_TLS_SLOT) * sizeof(void*);
+  if (exe_tpoff < min_bionic_alignment) {
+    async_safe_fatal("error: \"%s\": executable's TLS segment is underaligned: "
+                     "alignment is %zu, needs to be at least %zu for %s Bionic",
+                     progname, exe_segment->alignment, min_bionic_alignment,
+                     (sizeof(void*) == 4 ? "ARM" : "ARM64"));
+  }
+
+  offset_bionic_tcb_ = result - exe_tpoff - (-MIN_TLS_SLOT * sizeof(void*));
+  return result;
+
+#elif defined(__i386__) || defined(__x86_64__)
+
+  // x86 uses variant 2 TLS layout. The executable's segment is located just
+  // before the TCB.
+  static_assert(MIN_TLS_SLOT == 0, "First slot of bionic_tcb must be slot #0 on x86");
+  const size_t exe_size = round_up_with_overflow_check(exe_segment->size, exe_segment->alignment);
+  reserve(exe_size, 1);
+  const size_t max_align = MAX(alignof(bionic_tcb), exe_segment->alignment);
+  offset_bionic_tcb_ = reserve(sizeof(bionic_tcb), max_align);
+  return offset_bionic_tcb_ - exe_size;
+
+#else
+#error "Unrecognized architecture"
+#endif
 }
 
 void StaticTlsLayout::reserve_bionic_tls() {
@@ -76,6 +135,10 @@
 void StaticTlsLayout::finish_layout() {
   // Round the offset up to the alignment.
   offset_ = round_up_with_overflow_check(offset_, alignment_);
+
+  if (overflowed_) {
+    async_safe_fatal("error: TLS segments in static TLS overflowed");
+  }
 }
 
 // The size is not required to be a multiple of the alignment. The alignment
@@ -94,3 +157,33 @@
   if (value < old_value) overflowed_ = true;
   return value;
 }
+
+// Copy each TLS module's initialization image into a newly-allocated block of
+// static TLS memory. To reduce dirty pages, this function only writes to pages
+// within the static TLS that need initialization. The memory should already be
+// zero-initialized on entry.
+void __init_static_tls(void* static_tls) {
+  // The part of the table we care about (i.e. static TLS modules) never changes
+  // after startup, but we still need the mutex because the table could grow,
+  // moving the initial part. If this locking is too slow, we can duplicate the
+  // static part of the table.
+  TlsModules& modules = __libc_shared_globals()->tls_modules;
+  ScopedReadLock locker(&modules.rwlock);
+
+  for (size_t i = 0; i < modules.module_count; ++i) {
+    TlsModule& module = modules.module_table[i];
+    if (module.static_offset == SIZE_MAX) {
+      // All of the static modules come before all of the dynamic modules, so
+      // once we see the first dynamic module, we're done.
+      break;
+    }
+    if (module.segment.init_size == 0) {
+      // Skip the memcpy call for TLS segments with no initializer, which is
+      // common.
+      continue;
+    }
+    memcpy(static_cast<char*>(static_tls) + module.static_offset,
+           module.segment.init_ptr,
+           module.segment.init_size);
+  }
+}
diff --git a/libc/bionic/libc_init_static.cpp b/libc/bionic/libc_init_static.cpp
index 68650ed..8fbc20e 100644
--- a/libc/bionic/libc_init_static.cpp
+++ b/libc/bionic/libc_init_static.cpp
@@ -83,10 +83,29 @@
   }
 }
 
-static void layout_static_tls() {
+static void layout_static_tls(KernelArgumentBlock& args) {
   StaticTlsLayout& layout = __libc_shared_globals()->static_tls_layout;
   layout.reserve_bionic_tls();
-  layout.reserve_tcb();
+
+  const char* progname = args.argv[0];
+  ElfW(Phdr)* phdr_start = reinterpret_cast<ElfW(Phdr)*>(getauxval(AT_PHDR));
+  size_t phdr_ct = getauxval(AT_PHNUM);
+
+  static TlsModule mod;
+  if (__bionic_get_tls_segment(phdr_start, phdr_ct, 0, &mod.segment)) {
+    if (!__bionic_check_tls_alignment(&mod.segment.alignment)) {
+      async_safe_fatal("error: TLS segment alignment in \"%s\" is not a power of 2: %zu\n",
+                       progname, mod.segment.alignment);
+    }
+    mod.static_offset = layout.reserve_exe_segment_and_tcb(&mod.segment, progname);
+    mod.first_generation = 1;
+    __libc_shared_globals()->tls_modules.generation = 1;
+    __libc_shared_globals()->tls_modules.module_count = 1;
+    __libc_shared_globals()->tls_modules.module_table = &mod;
+  } else {
+    layout.reserve_exe_segment_and_tcb(nullptr, progname);
+  }
+
   layout.finish_layout();
 }
 
@@ -111,7 +130,7 @@
   __libc_init_globals();
   __libc_shared_globals()->init_progname = args.argv[0];
   __libc_init_AT_SECURE(args.envp);
-  layout_static_tls();
+  layout_static_tls(args);
   __libc_init_main_thread_final();
   __libc_init_common();
 
diff --git a/libc/bionic/pthread_create.cpp b/libc/bionic/pthread_create.cpp
index 8676a45..31e0378 100644
--- a/libc/bionic/pthread_create.cpp
+++ b/libc/bionic/pthread_create.cpp
@@ -288,7 +288,8 @@
   auto tcb = reinterpret_cast<bionic_tcb*>(mapping.static_tls + layout.offset_bionic_tcb());
   auto tls = reinterpret_cast<bionic_tls*>(mapping.static_tls + layout.offset_bionic_tls());
 
-  // (Re)initialize TLS pointers.
+  // Initialize TLS memory.
+  __init_static_tls(mapping.static_tls);
   __init_tcb(tcb, thread);
   __init_tcb_stack_guard(tcb);
   __init_bionic_tls_ptrs(tcb, tls);
diff --git a/libc/private/bionic_elf_tls.h b/libc/private/bionic_elf_tls.h
index 48a4d25..09e1958 100644
--- a/libc/private/bionic_elf_tls.h
+++ b/libc/private/bionic_elf_tls.h
@@ -29,6 +29,8 @@
 #pragma once
 
 #include <link.h>
+#include <pthread.h>
+#include <stdatomic.h>
 #include <stdint.h>
 #include <sys/cdefs.h>
 
@@ -40,8 +42,9 @@
 };
 
 __LIBC_HIDDEN__ bool __bionic_get_tls_segment(const ElfW(Phdr)* phdr_table, size_t phdr_count,
-                                              ElfW(Addr) load_bias, const char* mod_name,
-                                              TlsSegment* out);
+                                              ElfW(Addr) load_bias, TlsSegment* out);
+
+__LIBC_HIDDEN__ bool __bionic_check_tls_alignment(size_t* alignment);
 
 struct StaticTlsLayout {
   constexpr StaticTlsLayout() {}
@@ -58,13 +61,17 @@
 public:
   size_t offset_bionic_tcb() const { return offset_bionic_tcb_; }
   size_t offset_bionic_tls() const { return offset_bionic_tls_; }
+  size_t offset_thread_pointer() const;
 
   size_t size() const { return offset_; }
   size_t alignment() const { return alignment_; }
   bool overflowed() const { return overflowed_; }
 
-  void reserve_tcb();
+  size_t reserve_exe_segment_and_tcb(const TlsSegment* exe_segment, const char* progname);
   void reserve_bionic_tls();
+  size_t reserve_solib_segment(const TlsSegment& segment) {
+    return reserve(segment.size, segment.alignment);
+  }
   void finish_layout();
 
 private:
@@ -76,3 +83,39 @@
 
   size_t round_up_with_overflow_check(size_t value, size_t alignment);
 };
+
+// A descriptor for a single ELF TLS module.
+struct TlsModule {
+  TlsSegment segment;
+
+  // Offset into the static TLS block or SIZE_MAX for a dynamic module.
+  size_t static_offset = SIZE_MAX;
+
+  // The generation in which this module was loaded. Dynamic TLS lookups use
+  // this field to detect when a module has been unloaded.
+  size_t first_generation = 0;
+
+  // Used by the dynamic linker to track the associated soinfo* object.
+  void* soinfo_ptr = nullptr;
+};
+
+// Table of the ELF TLS modules. Either the dynamic linker or the static
+// initialization code prepares this table, and it's then used during thread
+// creation and for dynamic TLS lookups.
+struct TlsModules {
+  constexpr TlsModules() {}
+
+  // A generation counter. The value is incremented each time an solib is loaded
+  // or unloaded.
+  _Atomic(size_t) generation = 0;
+
+  // Access to the TlsModule[] table requires taking this lock.
+  pthread_rwlock_t rwlock = PTHREAD_RWLOCK_INITIALIZER;
+
+  // Pointer to a block of TlsModule objects. The first module has ID 1 and
+  // is stored at index 0 in this table.
+  size_t module_count = 0;
+  TlsModule* module_table = nullptr;
+};
+
+void __init_static_tls(void* static_tls);
diff --git a/libc/private/bionic_globals.h b/libc/private/bionic_globals.h
index b5e677e..4d40476 100644
--- a/libc/private/bionic_globals.h
+++ b/libc/private/bionic_globals.h
@@ -69,6 +69,7 @@
   abort_msg_t* abort_msg = nullptr;
 
   StaticTlsLayout static_tls_layout;
+  TlsModules tls_modules;
 
   // Values passed from the linker to libc.so.
   const char* init_progname = nullptr;
diff --git a/linker/linker.cpp b/linker/linker.cpp
index c5701dd..6687d23 100644
--- a/linker/linker.cpp
+++ b/linker/linker.cpp
@@ -65,8 +65,10 @@
 #include "linker_phdr.h"
 #include "linker_relocs.h"
 #include "linker_reloc_iterators.h"
+#include "linker_tls.h"
 #include "linker_utils.h"
 
+#include "private/bionic_globals.h"
 #include "android-base/macros.h"
 #include "android-base/strings.h"
 #include "android-base/stringprintf.h"
@@ -1655,6 +1657,7 @@
     if (!si->is_linked() && !si->prelink_image()) {
       return false;
     }
+    register_soinfo_tls(si);
   }
 
   // Step 4: Construct the global group. Note: DF_1_GLOBAL bit of a library is
@@ -1890,6 +1893,7 @@
            si->get_realpath(),
            si);
     notify_gdb_of_unload(si);
+    unregister_soinfo_tls(si);
     get_cfi_shadow()->BeforeUnload(si);
     soinfo_free(si);
   }
@@ -2669,16 +2673,32 @@
 #else
 static ElfW(Addr) get_addend(ElfW(Rel)* rel, ElfW(Addr) reloc_addr) {
   if (ELFW(R_TYPE)(rel->r_info) == R_GENERIC_RELATIVE ||
-      ELFW(R_TYPE)(rel->r_info) == R_GENERIC_IRELATIVE) {
+      ELFW(R_TYPE)(rel->r_info) == R_GENERIC_IRELATIVE ||
+      ELFW(R_TYPE)(rel->r_info) == R_GENERIC_TLS_DTPREL ||
+      ELFW(R_TYPE)(rel->r_info) == R_GENERIC_TLS_TPREL) {
     return *reinterpret_cast<ElfW(Addr)*>(reloc_addr);
   }
   return 0;
 }
 #endif
 
+static bool is_tls_reloc(ElfW(Word) type) {
+  switch (type) {
+    case R_GENERIC_TLS_DTPMOD:
+    case R_GENERIC_TLS_DTPREL:
+    case R_GENERIC_TLS_TPREL:
+    case R_GENERIC_TLSDESC:
+      return true;
+    default:
+      return false;
+  }
+}
+
 template<typename ElfRelIteratorT>
 bool soinfo::relocate(const VersionTracker& version_tracker, ElfRelIteratorT&& rel_iterator,
                       const soinfo_list_t& global_group, const soinfo_list_t& local_group) {
+  const size_t tls_tp_base = __libc_shared_globals()->static_tls_layout.offset_thread_pointer();
+
   for (size_t idx = 0; rel_iterator.has_next(); ++idx) {
     const auto rel = rel_iterator.next();
     if (rel == nullptr) {
@@ -2701,7 +2721,22 @@
     const ElfW(Sym)* s = nullptr;
     soinfo* lsi = nullptr;
 
-    if (sym != 0) {
+    if (sym == 0) {
+      // Do nothing.
+    } else if (ELF_ST_BIND(symtab_[sym].st_info) == STB_LOCAL && is_tls_reloc(type)) {
+      // In certain situations, the Gold linker accesses a TLS symbol using a
+      // relocation to an STB_LOCAL symbol in .dynsym of either STT_SECTION or
+      // STT_TLS type. Bionic doesn't support these relocations, so issue an
+      // error. References:
+      //  - https://groups.google.com/d/topic/generic-abi/dJ4_Y78aQ2M/discussion
+      //  - https://sourceware.org/bugzilla/show_bug.cgi?id=17699
+      s = &symtab_[sym];
+      sym_name = get_string(s->st_name);
+      DL_ERR("unexpected TLS reference to local symbol \"%s\": "
+             "sym type %d, rel type %u (idx %zu of \"%s\")",
+             sym_name, ELF_ST_TYPE(s->st_info), type, idx, get_realpath());
+      return false;
+    } else {
       sym_name = get_string(symtab_[sym].st_name);
       const version_info* vi = nullptr;
 
@@ -2738,6 +2773,10 @@
           case R_GENERIC_GLOB_DAT:
           case R_GENERIC_RELATIVE:
           case R_GENERIC_IRELATIVE:
+          case R_GENERIC_TLS_DTPMOD:
+          case R_GENERIC_TLS_DTPREL:
+          case R_GENERIC_TLS_TPREL:
+          case R_GENERIC_TLSDESC:
 #if defined(__aarch64__)
           case R_AARCH64_ABS64:
           case R_AARCH64_ABS32:
@@ -2785,12 +2824,21 @@
           }
         }
 #endif
-        if (ELF_ST_TYPE(s->st_info) == STT_TLS) {
-          DL_ERR("unsupported ELF TLS symbol \"%s\" referenced by \"%s\"",
-                 sym_name, get_realpath());
-          return false;
+        if (is_tls_reloc(type)) {
+          if (ELF_ST_TYPE(s->st_info) != STT_TLS) {
+            DL_ERR("reference to non-TLS symbol \"%s\" from TLS relocation in \"%s\"",
+                   sym_name, get_realpath());
+            return false;
+          }
+          sym_addr = s->st_value;
+        } else {
+          if (ELF_ST_TYPE(s->st_info) == STT_TLS) {
+            DL_ERR("reference to TLS symbol \"%s\" from non-TLS relocation in \"%s\"",
+                   sym_name, get_realpath());
+            return false;
+          }
+          sym_addr = lsi->resolve_symbol_address(s);
         }
-        sym_addr = lsi->resolve_symbol_address(s);
 #if !defined(__LP64__)
         if (protect_segments) {
           if (phdr_table_unprotect_segments(phdr, phnum, load_bias) < 0) {
@@ -2863,6 +2911,39 @@
           *reinterpret_cast<ElfW(Addr)*>(reloc) = ifunc_addr;
         }
         break;
+      case R_GENERIC_TLS_TPREL:
+        count_relocation(kRelocRelative);
+        MARK(rel->r_offset);
+        {
+          ElfW(Addr) tpoff = 0;
+          if (sym == 0) {
+            // By convention in ld.bfd and lld, an omitted symbol
+            // (ELFW(R_SYM) == 0) refers to the local module.
+            lsi = this;
+          }
+          if (lsi == nullptr) {
+            // Unresolved weak relocation. Leave tpoff at 0 to resolve
+            // &weak_tls_symbol to __get_tls().
+          } else if (soinfo_tls* lsi_tls = lsi->get_tls()) {
+            if (lsi_tls->module->static_offset != SIZE_MAX) {
+              tpoff += lsi_tls->module->static_offset - tls_tp_base;
+            } else {
+              DL_ERR("TLS symbol \"%s\" in dlopened \"%s\" referenced from \"%s\" using IE access model",
+                     sym_name, lsi->get_realpath(), get_realpath());
+              return false;
+            }
+          } else {
+            DL_ERR("TLS relocation refers to symbol \"%s\" in solib \"%s\" with no TLS segment",
+                   sym_name, lsi->get_realpath());
+            return false;
+          }
+          tpoff += sym_addr + addend;
+          TRACE_TYPE(RELO, "RELO TLS_TPREL %16p <- %16p %s\n",
+                     reinterpret_cast<void*>(reloc),
+                     reinterpret_cast<void*>(tpoff), sym_name);
+          *reinterpret_cast<ElfW(Addr)*>(reloc) = tpoff;
+        }
+        break;
 
 #if defined(__aarch64__)
       case R_AARCH64_ABS64:
@@ -2964,14 +3045,6 @@
          */
         DL_ERR("%s R_AARCH64_COPY relocations are not supported", get_realpath());
         return false;
-      case R_AARCH64_TLS_TPREL64:
-        TRACE_TYPE(RELO, "RELO TLS_TPREL64 *** %16llx <- %16llx - %16llx\n",
-                   reloc, (sym_addr + addend), rel->r_offset);
-        break;
-      case R_AARCH64_TLSDESC:
-        TRACE_TYPE(RELO, "RELO TLSDESC *** %16llx <- %16llx - %16llx\n",
-                   reloc, (sym_addr + addend), rel->r_offset);
-        break;
 #elif defined(__x86_64__)
       case R_X86_64_32:
         count_relocation(kRelocRelative);
@@ -3076,6 +3149,19 @@
                                   &ARM_exidx, &ARM_exidx_count);
 #endif
 
+  TlsSegment tls_segment;
+  if (__bionic_get_tls_segment(phdr, phnum, load_bias, &tls_segment)) {
+    if (!__bionic_check_tls_alignment(&tls_segment.alignment)) {
+      if (!relocating_linker) {
+        DL_ERR("TLS segment alignment in \"%s\" is not a power of 2: %zu",
+               get_realpath(), tls_segment.alignment);
+      }
+      return false;
+    }
+    tls_ = std::make_unique<soinfo_tls>();
+    tls_->segment = tls_segment;
+  }
+
   // Extract useful information from dynamic section.
   // Note that: "Except for the DT_NULL element at the end of the array,
   // and the relative order of DT_NEEDED elements, entries may appear in any order."
diff --git a/linker/linker_main.cpp b/linker/linker_main.cpp
index 9b4ce47..b0c27dc 100644
--- a/linker/linker_main.cpp
+++ b/linker/linker_main.cpp
@@ -415,6 +415,8 @@
     }
   }
 
+  linker_setup_exe_static_tls(g_argv[0]);
+
   // Load ld_preloads and dependencies.
   std::vector<const char*> needed_library_name_list;
   size_t ld_preloads_count = 0;
@@ -452,8 +454,7 @@
     si->increment_ref_count();
   }
 
-  layout_linker_static_tls();
-
+  linker_finalize_static_tls();
   __libc_init_main_thread_final();
 
   if (!get_cfi_shadow()->InitialLinkDone(solist)) __linker_cannot_link(g_argv[0]);
diff --git a/linker/linker_soinfo.cpp b/linker/linker_soinfo.cpp
index 93079ca..dcc6bf3 100644
--- a/linker/linker_soinfo.cpp
+++ b/linker/linker_soinfo.cpp
@@ -628,6 +628,10 @@
   return secondary_namespaces_;
 }
 
+soinfo_tls* soinfo::get_tls() const {
+  return has_min_version(5) ? tls_.get() : nullptr;
+}
+
 ElfW(Addr) soinfo::resolve_symbol_address(const ElfW(Sym)* s) const {
   if (ELF_ST_TYPE(s->st_info) == STT_GNU_IFUNC) {
     return call_ifunc_resolver(s->st_value + load_bias);
diff --git a/linker/linker_soinfo.h b/linker/linker_soinfo.h
index 44bff28..87e385d 100644
--- a/linker/linker_soinfo.h
+++ b/linker/linker_soinfo.h
@@ -30,8 +30,10 @@
 
 #include <link.h>
 
+#include <memory>
 #include <string>
 
+#include "private/bionic_elf_tls.h"
 #include "linker_namespaces.h"
 
 #define FLAG_LINKED           0x00000001
@@ -61,7 +63,7 @@
                                          // unset.
 #define FLAG_NEW_SOINFO       0x40000000 // new soinfo format
 
-#define SOINFO_VERSION 4
+#define SOINFO_VERSION 5
 
 typedef void (*linker_dtor_function_t)();
 typedef void (*linker_ctor_function_t)(int, char**, char**);
@@ -100,6 +102,17 @@
 // TODO(dimitry): remove reference from soinfo member functions to this class.
 class VersionTracker;
 
+// The first ELF TLS module has ID 1. Zero is reserved for the first word of
+// the DTV, a generation count, and unresolved weak symbols also use module
+// ID 0.
+static constexpr size_t kUninitializedModuleId = 0;
+
+struct soinfo_tls {
+  TlsSegment segment;
+  size_t module_id = kUninitializedModuleId;
+  TlsModule* module = nullptr;
+};
+
 #if defined(__work_around_b_24465209__)
 #define SOINFO_NAME_LEN 128
 #endif
@@ -284,6 +297,8 @@
   void add_secondary_namespace(android_namespace_t* secondary_ns);
   android_namespace_list_t& get_secondary_namespaces();
 
+  soinfo_tls* get_tls() const;
+
   void set_mapped_by_caller(bool reserved_map);
   bool is_mapped_by_caller() const;
 
@@ -366,6 +381,9 @@
   // version >= 4
   ElfW(Relr)* relr_;
   size_t relr_count_;
+
+  // version >= 5
+  std::unique_ptr<soinfo_tls> tls_;
 };
 
 // This function is used by dlvsym() to calculate hash of sym_ver
diff --git a/linker/linker_tls.cpp b/linker/linker_tls.cpp
index 3327453..a50d926 100644
--- a/linker/linker_tls.cpp
+++ b/linker/linker_tls.cpp
@@ -28,20 +28,75 @@
 
 #include "linker_tls.h"
 
+#include <vector>
+
+#include "private/ScopedRWLock.h"
 #include "private/bionic_defs.h"
 #include "private/bionic_elf_tls.h"
 #include "private/bionic_globals.h"
 #include "private/linker_native_bridge.h"
+#include "linker_main.h"
+#include "linker_soinfo.h"
+
+static bool g_static_tls_finished;
+static std::vector<TlsModule> g_tls_modules;
+
+static size_t get_unused_module_index() {
+  for (size_t i = 0; i < g_tls_modules.size(); ++i) {
+    if (g_tls_modules[i].soinfo_ptr == nullptr) {
+      return i;
+    }
+  }
+  g_tls_modules.push_back({});
+  __libc_shared_globals()->tls_modules.module_count = g_tls_modules.size();
+  __libc_shared_globals()->tls_modules.module_table = g_tls_modules.data();
+  return g_tls_modules.size() - 1;
+}
+
+static void register_tls_module(soinfo* si, size_t static_offset) {
+  // The global TLS module table points at the std::vector of modules declared
+  // in this file, so acquire a write lock before modifying the std::vector.
+  ScopedWriteLock locker(&__libc_shared_globals()->tls_modules.rwlock);
+
+  size_t module_idx = get_unused_module_index();
+  TlsModule* module = &g_tls_modules[module_idx];
+
+  soinfo_tls* si_tls = si->get_tls();
+  si_tls->module_id = module_idx + 1;
+  si_tls->module = module;
+
+  *module = {
+    .segment = si_tls->segment,
+    .static_offset = static_offset,
+    .first_generation = ++__libc_shared_globals()->tls_modules.generation,
+    .soinfo_ptr = si,
+  };
+}
+
+static void unregister_tls_module(soinfo* si) {
+  ScopedWriteLock locker(&__libc_shared_globals()->tls_modules.rwlock);
+
+  soinfo_tls* si_tls = si->get_tls();
+  CHECK(si_tls->module->static_offset == SIZE_MAX);
+  CHECK(si_tls->module->soinfo_ptr == si);
+  *si_tls->module = {};
+  si_tls->module_id = kUninitializedModuleId;
+  si_tls->module = nullptr;
+}
 
 __BIONIC_WEAK_FOR_NATIVE_BRIDGE
 extern "C" void __linker_reserve_bionic_tls_in_static_tls() {
   __libc_shared_globals()->static_tls_layout.reserve_bionic_tls();
 }
 
-// Stub for linker static TLS layout.
-void layout_linker_static_tls() {
+void linker_setup_exe_static_tls(const char* progname) {
+  soinfo* somain = solist_get_somain();
   StaticTlsLayout& layout = __libc_shared_globals()->static_tls_layout;
-  layout.reserve_tcb();
+  if (somain->get_tls() == nullptr) {
+    layout.reserve_exe_segment_and_tcb(nullptr, progname);
+  } else {
+    register_tls_module(somain, layout.reserve_exe_segment_and_tcb(&somain->get_tls()->segment, progname));
+  }
 
   // The pthread key data is located at the very front of bionic_tls. As a
   // temporary workaround, allocate bionic_tls just after the thread pointer so
@@ -49,8 +104,32 @@
   // small enough. Specifically, Golang scans forward 384 words from the TP on
   // ARM.
   //  - http://b/118381796
-  //  - https://groups.google.com/d/msg/golang-dev/yVrkFnYrYPE/2G3aFzYqBgAJ
+  //  - https://github.com/golang/go/issues/29674
   __linker_reserve_bionic_tls_in_static_tls();
+}
 
-  layout.finish_layout();
+void linker_finalize_static_tls() {
+  g_static_tls_finished = true;
+  __libc_shared_globals()->static_tls_layout.finish_layout();
+}
+
+void register_soinfo_tls(soinfo* si) {
+  soinfo_tls* si_tls = si->get_tls();
+  if (si_tls == nullptr || si_tls->module_id != kUninitializedModuleId) {
+    return;
+  }
+  size_t static_offset = SIZE_MAX;
+  if (!g_static_tls_finished) {
+    StaticTlsLayout& layout = __libc_shared_globals()->static_tls_layout;
+    static_offset = layout.reserve_solib_segment(si_tls->segment);
+  }
+  register_tls_module(si, static_offset);
+}
+
+void unregister_soinfo_tls(soinfo* si) {
+  soinfo_tls* si_tls = si->get_tls();
+  if (si_tls == nullptr || si_tls->module_id == kUninitializedModuleId) {
+    return;
+  }
+  return unregister_tls_module(si);
 }
diff --git a/linker/linker_tls.h b/linker/linker_tls.h
index 2f0a57d..4f48337 100644
--- a/linker/linker_tls.h
+++ b/linker/linker_tls.h
@@ -28,4 +28,10 @@
 
 #pragma once
 
-void layout_linker_static_tls();
+struct soinfo;
+
+void linker_setup_exe_static_tls(const char* progname);
+void linker_finalize_static_tls();
+
+void register_soinfo_tls(soinfo* si);
+void unregister_soinfo_tls(soinfo* si);
diff --git a/tests/dlfcn_test.cpp b/tests/dlfcn_test.cpp
index 5f48e67..176a6f8 100644
--- a/tests/dlfcn_test.cpp
+++ b/tests/dlfcn_test.cpp
@@ -1086,7 +1086,7 @@
   dlerror(); // Clear any pending errors.
   void* handle = dlopen("libelf-tls-library.so", RTLD_NOW);
   ASSERT_TRUE(handle == nullptr);
-  ASSERT_SUBSTR("unsupported ELF TLS", dlerror());
+  ASSERT_SUBSTR("unknown reloc type ", dlerror());
 }
 
 TEST(dlfcn, dlopen_bad_flags) {