diff --git a/libc/bionic/__libc_init_main_thread.cpp b/libc/bionic/__libc_init_main_thread.cpp
index 1b539f2..0d557f1 100644
--- a/libc/bionic/__libc_init_main_thread.cpp
+++ b/libc/bionic/__libc_init_main_thread.cpp
@@ -44,7 +44,7 @@
 // Declared in "private/bionic_ssp.h".
 uintptr_t __stack_chk_guard = 0;
 
-static pthread_internal_t main_thread;
+BIONIC_USED_BEFORE_LINKER_RELOCATES static pthread_internal_t main_thread;
 
 // Setup for the main thread. For dynamic executables, this is called by the
 // linker _before_ libc is mapped in memory. This means that all writes to
diff --git a/libc/bionic/bionic_call_ifunc_resolver.cpp b/libc/bionic/bionic_call_ifunc_resolver.cpp
index e44d998..d5a812c 100644
--- a/libc/bionic/bionic_call_ifunc_resolver.cpp
+++ b/libc/bionic/bionic_call_ifunc_resolver.cpp
@@ -31,6 +31,7 @@
 #include <sys/hwprobe.h>
 #include <sys/ifunc.h>
 
+#include "bionic/macros.h"
 #include "private/bionic_auxv.h"
 
 // This code is called in the linker before it has been relocated, so minimize calls into other
@@ -40,8 +41,8 @@
 ElfW(Addr) __bionic_call_ifunc_resolver(ElfW(Addr) resolver_addr) {
 #if defined(__aarch64__)
   typedef ElfW(Addr) (*ifunc_resolver_t)(uint64_t, __ifunc_arg_t*);
-  static __ifunc_arg_t arg;
-  static bool initialized = false;
+  BIONIC_USED_BEFORE_LINKER_RELOCATES static __ifunc_arg_t arg;
+  BIONIC_USED_BEFORE_LINKER_RELOCATES static bool initialized = false;
   if (!initialized) {
     initialized = true;
     arg._size = sizeof(__ifunc_arg_t);
diff --git a/libc/bionic/libc_init_static.cpp b/libc/bionic/libc_init_static.cpp
index 7c46113..ae86da1 100644
--- a/libc/bionic/libc_init_static.cpp
+++ b/libc/bionic/libc_init_static.cpp
@@ -344,19 +344,12 @@
 // This function is called from the linker before the main executable is relocated.
 __attribute__((no_sanitize("hwaddress", "memtag"))) void __libc_init_mte(
     const memtag_dynamic_entries_t* memtag_dynamic_entries, const void* phdr_start, size_t phdr_ct,
-    uintptr_t load_bias, void* stack_top) {
+    uintptr_t load_bias) {
   bool memtag_stack = false;
   HeapTaggingLevel level =
       __get_tagging_level(memtag_dynamic_entries, phdr_start, phdr_ct, load_bias, &memtag_stack);
-  // initial_memtag_stack is used by the linker (in linker.cpp) to communicate than any library
-  // linked by this executable enables memtag-stack.
-  // memtag_stack is also set for static executables if they request memtag stack via the note,
-  // in which case it will differ from initial_memtag_stack.
-  if (__libc_shared_globals()->initial_memtag_stack || memtag_stack) {
-    memtag_stack = true;
-    __libc_shared_globals()->initial_memtag_stack_abi = true;
-    __get_bionic_tcb()->tls_slot(TLS_SLOT_STACK_MTE) = __allocate_stack_mte_ringbuffer(0, nullptr);
-  }
+  __libc_shared_globals()->initial_memtag_stack = memtag_stack;
+
   if (int64_t timed_upgrade = __get_memtag_upgrade_secs()) {
     if (level == M_HEAP_TAGGING_LEVEL_ASYNC) {
       async_safe_format_log(ANDROID_LOG_INFO, "libc",
@@ -380,15 +373,7 @@
     if (prctl(PR_SET_TAGGED_ADDR_CTRL, prctl_arg | PR_MTE_TCF_SYNC, 0, 0, 0) == 0 ||
         prctl(PR_SET_TAGGED_ADDR_CTRL, prctl_arg, 0, 0, 0) == 0) {
       __libc_shared_globals()->initial_heap_tagging_level = level;
-      __libc_shared_globals()->initial_memtag_stack = memtag_stack;
 
-      if (memtag_stack) {
-        void* pg_start =
-            reinterpret_cast<void*>(page_start(reinterpret_cast<uintptr_t>(stack_top)));
-        if (mprotect(pg_start, page_size(), PROT_READ | PROT_WRITE | PROT_MTE | PROT_GROWSDOWN)) {
-          async_safe_fatal("error: failed to set PROT_MTE on main thread stack: %m");
-        }
-      }
       struct sigaction action = {};
       action.sa_flags = SA_SIGINFO | SA_RESTART;
       action.sa_sigaction = __enable_mte_signal_handler;
@@ -404,11 +389,36 @@
   }
   // We did not enable MTE, so we do not need to arm the upgrade timer.
   __libc_shared_globals()->heap_tagging_upgrade_timer_sec = 0;
-  // We also didn't enable memtag_stack.
-  __libc_shared_globals()->initial_memtag_stack = false;
 }
+
+// Figure out whether we need to map the stack as PROT_MTE.
+// For dynamic executables, this has to be called after loading all
+// DT_NEEDED libraries, in case one of them needs stack MTE.
+__attribute__((no_sanitize("hwaddress", "memtag"))) void __libc_init_mte_stack(void* stack_top) {
+  if (!__libc_shared_globals()->initial_memtag_stack) {
+    return;
+  }
+
+  // Even if the device doesn't support MTE, we have to allocate stack
+  // history buffers for code compiled for stack MTE. That is because the
+  // codegen expects a buffer to be present in TLS_SLOT_STACK_MTE either
+  // way.
+  __libc_shared_globals()->initial_memtag_stack_abi = true;
+  __get_bionic_tcb()->tls_slot(TLS_SLOT_STACK_MTE) = __allocate_stack_mte_ringbuffer(0, nullptr);
+
+  if (!__libc_mte_enabled()) {
+    __libc_shared_globals()->initial_memtag_stack = false;
+    return;
+  }
+  void* pg_start = reinterpret_cast<void*>(page_start(reinterpret_cast<uintptr_t>(stack_top)));
+  if (mprotect(pg_start, page_size(), PROT_READ | PROT_WRITE | PROT_MTE | PROT_GROWSDOWN)) {
+    async_safe_fatal("error: failed to set PROT_MTE on main thread stack: %m");
+  }
+}
+
 #else   // __aarch64__
-void __libc_init_mte(const memtag_dynamic_entries_t*, const void*, size_t, uintptr_t, void*) {}
+void __libc_init_mte(const memtag_dynamic_entries_t*, const void*, size_t, uintptr_t) {}
+void __libc_init_mte_stack(void*) {}
 #endif  // __aarch64__
 
 void __libc_init_profiling_handlers() {
@@ -436,7 +446,8 @@
   __libc_init_common();
   __libc_init_mte(/*memtag_dynamic_entries=*/nullptr,
                   reinterpret_cast<ElfW(Phdr)*>(getauxval(AT_PHDR)), getauxval(AT_PHNUM),
-                  /*load_bias = */ 0, /*stack_top = */ raw_args);
+                  /*load_bias = */ 0);
+  __libc_init_mte_stack(/*stack_top = */ raw_args);
   __libc_init_scudo();
   __libc_init_profiling_handlers();
   __libc_init_fork_handler();
@@ -508,6 +519,11 @@
 // compiled with -ffreestanding to avoid implicit string.h function calls. (It shouldn't strictly
 // be necessary, though.)
 __LIBC_HIDDEN__ libc_shared_globals* __libc_shared_globals() {
-  static libc_shared_globals globals;
+  BIONIC_USED_BEFORE_LINKER_RELOCATES static libc_shared_globals globals;
   return &globals;
 }
+
+__LIBC_HIDDEN__ bool __libc_mte_enabled() {
+  HeapTaggingLevel lvl = __libc_shared_globals()->initial_heap_tagging_level;
+  return lvl == M_HEAP_TAGGING_LEVEL_SYNC || lvl == M_HEAP_TAGGING_LEVEL_ASYNC;
+}
diff --git a/libc/platform/bionic/macros.h b/libc/platform/bionic/macros.h
index 93268c1..837583e 100644
--- a/libc/platform/bionic/macros.h
+++ b/libc/platform/bionic/macros.h
@@ -97,3 +97,26 @@
 static inline T* _Nonnull untag_address(T* _Nonnull p) {
   return reinterpret_cast<T*>(untag_address(reinterpret_cast<uintptr_t>(p)));
 }
+
+// MTE globals protects internal and external global variables. One of the main
+// things that MTE globals does is force all global variable accesses to go
+// through the GOT. In the linker though, some global variables are accessed (or
+// address-taken) prior to relocations being processed. Because relocations
+// haven't run yet, the GOT entry hasn't been populated, and this leads to
+// crashes. Thus, any globals used by the linker prior to relocation should be
+// annotated with this attribute, which suppresses tagging of this global
+// variable, restoring the pc-relative address computation.
+//
+// A way to find global variables that need this attribute is to build the
+// linker/libc with `SANITIZE_TARGET=memtag_globals`, push them onto a device
+// (it doesn't have to be MTE capable), and then run an executable using
+// LD_LIBRARY_PATH and using the linker in interpreter mode (e.g.
+// `LD_LIBRARY_PATH=/data/tmp/ /data/tmp/linker64 /data/tmp/my_binary`). A
+// good heuristic is that the global variable is in a file that should be
+// compiled with `-ffreestanding` (but there are global variables there that
+// don't need thisattribute).
+#if __has_feature(memtag_globals)
+#define BIONIC_USED_BEFORE_LINKER_RELOCATES __attribute__((no_sanitize("memtag")))
+#else  // __has_feature(memtag_globals)
+#define BIONIC_USED_BEFORE_LINKER_RELOCATES
+#endif  // __has_feature(memtag_globals)
diff --git a/libc/platform/bionic/mte.h b/libc/platform/bionic/mte.h
index 98b3d27..423ed0f 100644
--- a/libc/platform/bionic/mte.h
+++ b/libc/platform/bionic/mte.h
@@ -28,6 +28,7 @@
 
 #pragma once
 
+#include <stddef.h>
 #include <sys/auxv.h>
 #include <sys/mman.h>
 #include <sys/prctl.h>
@@ -49,6 +50,36 @@
   return supported;
 }
 
+inline void* get_tagged_address(const void* ptr) {
+#if defined(__aarch64__)
+  if (mte_supported()) {
+    __asm__ __volatile__(".arch_extension mte; ldg %0, [%0]" : "+r"(ptr));
+  }
+#endif  // aarch64
+  return const_cast<void*>(ptr);
+}
+
+// Inserts a random tag tag to `ptr`, using any of the set lower 16 bits in
+// `mask` to exclude the corresponding tag from being generated. Note: This does
+// not tag memory. This generates a pointer to be used with set_memory_tag.
+inline void* insert_random_tag(const void* ptr, __attribute__((unused)) uint64_t mask = 0) {
+#if defined(__aarch64__)
+  if (mte_supported() && ptr) {
+    __asm__ __volatile__(".arch_extension mte; irg %0, %0, %1" : "+r"(ptr) : "r"(mask));
+  }
+#endif  // aarch64
+  return const_cast<void*>(ptr);
+}
+
+// Stores the address tag in `ptr` to memory, at `ptr`.
+inline void set_memory_tag(__attribute__((unused)) void* ptr) {
+#if defined(__aarch64__)
+  if (mte_supported()) {
+    __asm__ __volatile__(".arch_extension mte; stg %0, [%0]" : "+r"(ptr));
+  }
+#endif  // aarch64
+}
+
 #ifdef __aarch64__
 class ScopedDisableMTE {
   size_t prev_tco_;
diff --git a/libc/private/bionic_globals.h b/libc/private/bionic_globals.h
index a1bebda..02713fc 100644
--- a/libc/private/bionic_globals.h
+++ b/libc/private/bionic_globals.h
@@ -157,6 +157,7 @@
 };
 
 __LIBC_HIDDEN__ libc_shared_globals* __libc_shared_globals();
+__LIBC_HIDDEN__ bool __libc_mte_enabled();
 __LIBC_HIDDEN__ void __libc_init_fdsan();
 __LIBC_HIDDEN__ void __libc_init_fdtrack();
 __LIBC_HIDDEN__ void __libc_init_profiling_handlers();
diff --git a/linker/Android.bp b/linker/Android.bp
index a06ca29..847a9b2 100644
--- a/linker/Android.bp
+++ b/linker/Android.bp
@@ -108,6 +108,12 @@
 
     // We need to access Bionic private headers in the linker.
     include_dirs: ["bionic/libc"],
+
+    sanitize: {
+        // Supporting memtag_globals in the linker would be tricky,
+        // because it relocates itself very early.
+        memtag_globals: false,
+    },
 }
 
 // ========================================================
diff --git a/linker/dlfcn.cpp b/linker/dlfcn.cpp
index fee19f4..82f2728 100644
--- a/linker/dlfcn.cpp
+++ b/linker/dlfcn.cpp
@@ -331,6 +331,7 @@
     __libdl_info->gnu_bloom_filter_ = linker_si.gnu_bloom_filter_;
     __libdl_info->gnu_bucket_ = linker_si.gnu_bucket_;
     __libdl_info->gnu_chain_ = linker_si.gnu_chain_;
+    __libdl_info->memtag_dynamic_entries_ = linker_si.memtag_dynamic_entries_;
 
     __libdl_info->ref_count_ = 1;
     __libdl_info->strtab_size_ = linker_si.strtab_size_;
diff --git a/linker/linker.cpp b/linker/linker.cpp
index 88d02dc..517950c 100644
--- a/linker/linker.cpp
+++ b/linker/linker.cpp
@@ -51,6 +51,7 @@
 #include <android-base/scopeguard.h>
 #include <async_safe/log.h>
 #include <bionic/pthread_internal.h>
+#include <platform/bionic/mte.h>
 
 // Private C library headers.
 
@@ -2361,7 +2362,7 @@
         void* tls_block = get_tls_block_for_this_thread(tls_module, /*should_alloc=*/true);
         *symbol = static_cast<char*>(tls_block) + sym->st_value;
       } else {
-        *symbol = reinterpret_cast<void*>(found->resolve_symbol_address(sym));
+        *symbol = get_tagged_address(reinterpret_cast<void*>(found->resolve_symbol_address(sym)));
       }
       failure_guard.Disable();
       LD_LOG(kLogDlsym,
@@ -2791,15 +2792,25 @@
   return true;
 }
 
-static void apply_relr_reloc(ElfW(Addr) offset, ElfW(Addr) load_bias) {
-  ElfW(Addr) address = offset + load_bias;
-  *reinterpret_cast<ElfW(Addr)*>(address) += load_bias;
+static void apply_relr_reloc(ElfW(Addr) offset, ElfW(Addr) load_bias, bool has_memtag_globals) {
+  ElfW(Addr) destination = offset + load_bias;
+  if (!has_memtag_globals) {
+    *reinterpret_cast<ElfW(Addr)*>(destination) += load_bias;
+    return;
+  }
+
+  ElfW(Addr)* tagged_destination =
+      reinterpret_cast<ElfW(Addr)*>(get_tagged_address(reinterpret_cast<void*>(destination)));
+  ElfW(Addr) tagged_value = reinterpret_cast<ElfW(Addr)>(
+      get_tagged_address(reinterpret_cast<void*>(*tagged_destination + load_bias)));
+  *tagged_destination = tagged_value;
 }
 
 // Process relocations in SHT_RELR section (experimental).
 // Details of the encoding are described in this post:
 //   https://groups.google.com/d/msg/generic-abi/bX460iggiKg/Pi9aSwwABgAJ
-bool relocate_relr(const ElfW(Relr)* begin, const ElfW(Relr)* end, ElfW(Addr) load_bias) {
+bool relocate_relr(const ElfW(Relr) * begin, const ElfW(Relr) * end, ElfW(Addr) load_bias,
+                   bool has_memtag_globals) {
   constexpr size_t wordsize = sizeof(ElfW(Addr));
 
   ElfW(Addr) base = 0;
@@ -2810,7 +2821,7 @@
     if ((entry&1) == 0) {
       // Even entry: encodes the offset for next relocation.
       offset = static_cast<ElfW(Addr)>(entry);
-      apply_relr_reloc(offset, load_bias);
+      apply_relr_reloc(offset, load_bias, has_memtag_globals);
       // Set base offset for subsequent bitmap entries.
       base = offset + wordsize;
       continue;
@@ -2821,7 +2832,7 @@
     while (entry != 0) {
       entry >>= 1;
       if ((entry&1) != 0) {
-        apply_relr_reloc(offset, load_bias);
+        apply_relr_reloc(offset, load_bias, has_memtag_globals);
       }
       offset += wordsize;
     }
@@ -3325,6 +3336,18 @@
   // it each time we look up a symbol with a version.
   if (!validate_verdef_section(this)) return false;
 
+  // MTE globals requires remapping data segments with PROT_MTE as anonymous mappings, because file
+  // based mappings may not be backed by tag-capable memory (see "MAP_ANONYMOUS" on
+  // https://www.kernel.org/doc/html/latest/arch/arm64/memory-tagging-extension.html). This is only
+  // done if the binary has MTE globals (evidenced by the dynamic table entries), as it destroys
+  // page sharing. It's also only done on devices that support MTE, because the act of remapping
+  // pages is unnecessary on non-MTE devices (where we might still run MTE-globals enabled code).
+  if (should_tag_memtag_globals() &&
+      remap_memtag_globals_segments(phdr, phnum, base) == 0) {
+    tag_globals();
+    protect_memtag_globals_ro_segments(phdr, phnum, base);
+  }
+
   flags_ |= FLAG_PRELINKED;
   return true;
 }
@@ -3397,6 +3420,10 @@
     return false;
   }
 
+  if (should_tag_memtag_globals()) {
+    name_memtag_globals_segments(phdr, phnum, base, get_realpath(), vma_names_);
+  }
+
   /* Handle serializing/sharing the RELRO segment */
   if (extinfo && (extinfo->flags & ANDROID_DLEXT_WRITE_RELRO)) {
     if (phdr_table_serialize_gnu_relro(phdr, phnum, load_bias,
@@ -3435,6 +3462,48 @@
   return true;
 }
 
+// https://github.com/ARM-software/abi-aa/blob/main/memtagabielf64/memtagabielf64.rst#global-variable-tagging
+void soinfo::tag_globals() {
+  if (is_linked()) return;
+  if (flags_ & FLAG_GLOBALS_TAGGED) return;
+  flags_ |= FLAG_GLOBALS_TAGGED;
+
+  constexpr size_t kTagGranuleSize = 16;
+  const uint8_t* descriptor_stream = reinterpret_cast<const uint8_t*>(memtag_globals());
+
+  if (memtag_globalssz() == 0) {
+    DL_ERR("Invalid memtag descriptor pool size: %zu", memtag_globalssz());
+  }
+
+  uint64_t addr = load_bias;
+  uleb128_decoder decoder(descriptor_stream, memtag_globalssz());
+  // Don't ever generate tag zero, to easily distinguish between tagged and
+  // untagged globals in register/tag dumps.
+  uint64_t last_tag_mask = 1;
+  constexpr uint64_t kDistanceReservedBits = 3;
+
+  while (decoder.has_bytes()) {
+    uint64_t value = decoder.pop_front();
+    uint64_t distance = (value >> kDistanceReservedBits) * kTagGranuleSize;
+    uint64_t ngranules = value & ((1 << kDistanceReservedBits) - 1);
+    if (ngranules == 0) {
+      ngranules = decoder.pop_front() + 1;
+    }
+
+    addr += distance;
+    void* tagged_addr;
+    tagged_addr = insert_random_tag(reinterpret_cast<void*>(addr), last_tag_mask);
+    uint64_t tag = (reinterpret_cast<uint64_t>(tagged_addr) >> 56) & 0x0f;
+    last_tag_mask = 1 | (1 << tag);
+
+    for (size_t k = 0; k < ngranules; k++) {
+      auto* granule = static_cast<uint8_t*>(tagged_addr) + k * kTagGranuleSize;
+      set_memory_tag(static_cast<void*>(granule));
+    }
+    addr += ngranules * kTagGranuleSize;
+  }
+}
+
 static std::vector<android_namespace_t*> init_default_namespace_no_config(bool is_asan, bool is_hwasan) {
   g_default_namespace.set_isolated(false);
   auto default_ld_paths = is_asan ? kAsanDefaultLdPaths : (
diff --git a/linker/linker.h b/linker/linker.h
index ac2222d..b696fd9 100644
--- a/linker/linker.h
+++ b/linker/linker.h
@@ -179,7 +179,8 @@
 int get_application_target_sdk_version();
 ElfW(Versym) find_verdef_version_index(const soinfo* si, const version_info* vi);
 bool validate_verdef_section(const soinfo* si);
-bool relocate_relr(const ElfW(Relr)* begin, const ElfW(Relr)* end, ElfW(Addr) load_bias);
+bool relocate_relr(const ElfW(Relr) * begin, const ElfW(Relr) * end, ElfW(Addr) load_bias,
+                   bool has_memtag_globals);
 
 struct platform_properties {
 #if defined(__aarch64__)
diff --git a/linker/linker_main.cpp b/linker/linker_main.cpp
index 48ed723..f65f82d 100644
--- a/linker/linker_main.cpp
+++ b/linker/linker_main.cpp
@@ -46,6 +46,7 @@
 #include "linker_tls.h"
 #include "linker_utils.h"
 
+#include "platform/bionic/macros.h"
 #include "private/KernelArgumentBlock.h"
 #include "private/bionic_call_ifunc_resolver.h"
 #include "private/bionic_globals.h"
@@ -71,7 +72,9 @@
 static void set_bss_vma_name(soinfo* si);
 
 void __libc_init_mte(const memtag_dynamic_entries_t* memtag_dynamic_entries, const void* phdr_start,
-                     size_t phdr_count, uintptr_t load_bias, void* stack_top);
+                     size_t phdr_count, uintptr_t load_bias);
+
+void __libc_init_mte_stack(void* stack_top);
 
 static void __linker_cannot_link(const char* argv0) {
   __linker_error("CANNOT LINK EXECUTABLE \"%s\": %s", argv0, linker_get_error_buffer());
@@ -365,6 +368,8 @@
   init_link_map_head(*solinker);
 
 #if defined(__aarch64__)
+  __libc_init_mte(somain->memtag_dynamic_entries(), somain->phdr, somain->phnum, somain->load_bias);
+
   if (exe_to_load == nullptr) {
     // Kernel does not add PROT_BTI to executable pages of the loaded ELF.
     // Apply appropriate protections here if it is needed.
@@ -465,8 +470,7 @@
 #if defined(__aarch64__)
   // This has to happen after the find_libraries, which will have collected any possible
   // libraries that request memtag_stack in the dynamic section.
-  __libc_init_mte(somain->memtag_dynamic_entries(), somain->phdr, somain->phnum, somain->load_bias,
-                  args.argv);
+  __libc_init_mte_stack(args.argv);
 #endif
 
   linker_finalize_static_tls();
@@ -625,8 +629,13 @@
     // Apply RELR relocations first so that the GOT is initialized for ifunc
     // resolvers.
     if (relr && relrsz) {
+      // Nothing has tagged the memtag globals here, so it is pointless either
+      // way to handle them, the tags will be zero anyway.
+      // That is moot though, because the linker does not use memtag_globals
+      // in the first place.
       relocate_relr(reinterpret_cast<ElfW(Relr*)>(ehdr + relr),
-                    reinterpret_cast<ElfW(Relr*)>(ehdr + relr + relrsz), ehdr);
+                    reinterpret_cast<ElfW(Relr*)>(ehdr + relr + relrsz), ehdr,
+                    /*has_memtag_globals=*/ false);
     }
     if (pltrel && pltrelsz) {
       call_ifunc_resolvers_for_section(reinterpret_cast<RelType*>(ehdr + pltrel),
@@ -646,6 +655,16 @@
   }
 }
 
+// Remapping MTE globals segments happens before the linker relocates itself, and so can't use
+// memcpy() from string.h. This function is compiled with -ffreestanding.
+void linker_memcpy(void* dst, const void* src, size_t n) {
+  char* dst_bytes = reinterpret_cast<char*>(dst);
+  const char* src_bytes = reinterpret_cast<const char*>(src);
+  for (size_t i = 0; i < n; ++i) {
+    dst_bytes[i] = src_bytes[i];
+  }
+}
+
 // Detect an attempt to run the linker on itself. e.g.:
 //   /system/bin/linker64 /system/bin/linker64
 // Use priority-1 to run this constructor before other constructors.
diff --git a/linker/linker_main.h b/linker/linker_main.h
index 724f43c..ffbcf0f 100644
--- a/linker/linker_main.h
+++ b/linker/linker_main.h
@@ -70,3 +70,5 @@
 soinfo* solist_get_head();
 soinfo* solist_get_somain();
 soinfo* solist_get_vdso();
+
+void linker_memcpy(void* dst, const void* src, size_t n);
diff --git a/linker/linker_phdr.cpp b/linker/linker_phdr.cpp
index 7691031..2bdd7f8 100644
--- a/linker/linker_phdr.cpp
+++ b/linker/linker_phdr.cpp
@@ -37,9 +37,12 @@
 #include <unistd.h>
 
 #include "linker.h"
+#include "linker_debug.h"
 #include "linker_dlwarning.h"
 #include "linker_globals.h"
-#include "linker_debug.h"
+#include "linker_logger.h"
+#include "linker_main.h"
+#include "linker_soinfo.h"
 #include "linker_utils.h"
 
 #include "private/bionic_asm_note.h"
@@ -1172,6 +1175,125 @@
                                    should_use_16kib_app_compat);
 }
 
+static bool segment_needs_memtag_globals_remapping(const ElfW(Phdr) * phdr) {
+  // For now, MTE globals is only supported on writeable data segments.
+  return phdr->p_type == PT_LOAD && !(phdr->p_flags & PF_X) && (phdr->p_flags & PF_W);
+}
+
+/* When MTE globals are requested by the binary, and when the hardware supports
+ * it, remap the executable's PT_LOAD data pages to have PROT_MTE.
+ *
+ * Returns 0 on success, -1 on failure (error code in errno).
+ */
+int remap_memtag_globals_segments(const ElfW(Phdr) * phdr_table __unused,
+                                  size_t phdr_count __unused, ElfW(Addr) load_bias __unused) {
+#if defined(__aarch64__)
+  for (const ElfW(Phdr)* phdr = phdr_table; phdr < phdr_table + phdr_count; phdr++) {
+    if (!segment_needs_memtag_globals_remapping(phdr)) {
+      continue;
+    }
+
+    uintptr_t seg_page_start = page_start(phdr->p_vaddr) + load_bias;
+    uintptr_t seg_page_end = page_end(phdr->p_vaddr + phdr->p_memsz) + load_bias;
+    size_t seg_page_aligned_size = seg_page_end - seg_page_start;
+
+    int prot = PFLAGS_TO_PROT(phdr->p_flags);
+    // For anonymous private mappings, it may be possible to simply mprotect()
+    // the PROT_MTE flag over the top. For file-based mappings, this will fail,
+    // and we'll need to fall back. We also allow PROT_WRITE here to allow
+    // writing memory tags (in `soinfo::tag_globals()`), and set these sections
+    // back to read-only after tags are applied (similar to RELRO).
+    prot |= PROT_MTE;
+    if (mprotect(reinterpret_cast<void*>(seg_page_start), seg_page_aligned_size,
+                 prot | PROT_WRITE) == 0) {
+      continue;
+    }
+
+    void* mapping_copy = mmap(nullptr, seg_page_aligned_size, PROT_READ | PROT_WRITE,
+                              MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+    linker_memcpy(mapping_copy, reinterpret_cast<void*>(seg_page_start), seg_page_aligned_size);
+
+    void* seg_addr = mmap(reinterpret_cast<void*>(seg_page_start), seg_page_aligned_size,
+                          prot | PROT_WRITE, MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+    if (seg_addr == MAP_FAILED) return -1;
+
+    linker_memcpy(seg_addr, mapping_copy, seg_page_aligned_size);
+    munmap(mapping_copy, seg_page_aligned_size);
+  }
+#endif  // defined(__aarch64__)
+  return 0;
+}
+
+void protect_memtag_globals_ro_segments(const ElfW(Phdr) * phdr_table __unused,
+                                        size_t phdr_count __unused, ElfW(Addr) load_bias __unused) {
+#if defined(__aarch64__)
+  for (const ElfW(Phdr)* phdr = phdr_table; phdr < phdr_table + phdr_count; phdr++) {
+    int prot = PFLAGS_TO_PROT(phdr->p_flags);
+    if (!segment_needs_memtag_globals_remapping(phdr) || (prot & PROT_WRITE)) {
+      continue;
+    }
+
+    prot |= PROT_MTE;
+
+    uintptr_t seg_page_start = page_start(phdr->p_vaddr) + load_bias;
+    uintptr_t seg_page_end = page_end(phdr->p_vaddr + phdr->p_memsz) + load_bias;
+    size_t seg_page_aligned_size = seg_page_end - seg_page_start;
+    mprotect(reinterpret_cast<void*>(seg_page_start), seg_page_aligned_size, prot);
+  }
+#endif  // defined(__aarch64__)
+}
+
+void name_memtag_globals_segments(const ElfW(Phdr) * phdr_table, size_t phdr_count,
+                                  ElfW(Addr) load_bias, const char* soname,
+                                  std::list<std::string>& vma_names) {
+  for (const ElfW(Phdr)* phdr = phdr_table; phdr < phdr_table + phdr_count; phdr++) {
+    if (!segment_needs_memtag_globals_remapping(phdr)) {
+      continue;
+    }
+
+    uintptr_t seg_page_start = page_start(phdr->p_vaddr) + load_bias;
+    uintptr_t seg_page_end = page_end(phdr->p_vaddr + phdr->p_memsz) + load_bias;
+    size_t seg_page_aligned_size = seg_page_end - seg_page_start;
+
+    // For file-based mappings that we're now forcing to be anonymous mappings, set the VMA name to
+    // make debugging easier.
+    // Once we are targeting only devices that run kernel 5.10 or newer (and thus include
+    // https://android-review.git.corp.google.com/c/kernel/common/+/1934723 which causes the
+    // VMA_ANON_NAME to be copied into the kernel), we can get rid of the storage here.
+    // For now, that is not the case:
+    // https://source.android.com/docs/core/architecture/kernel/android-common#compatibility-matrix
+    constexpr int kVmaNameLimit = 80;
+    std::string& vma_name = vma_names.emplace_back('\0', kVmaNameLimit);
+    int full_vma_length =
+        async_safe_format_buffer(vma_name.data(), kVmaNameLimit, "mt:%s+%" PRIxPTR, soname,
+                                 page_start(phdr->p_vaddr)) +
+        /* include the null terminator */ 1;
+    // There's an upper limit of 80 characters, including the null terminator, in the anonymous VMA
+    // name. If we run over that limit, we end up truncating the segment offset and parts of the
+    // DSO's name, starting on the right hand side of the basename. Because the basename is the most
+    // important thing, chop off the soname from the left hand side first.
+    //
+    // Example (with '#' as the null terminator):
+    //   - "mt:/data/nativetest64/bionic-unit-tests/bionic-loader-test-libs/libdlext_test.so+e000#"
+    //     is a `full_vma_length` == 86.
+    //
+    // We need to left-truncate (86 - 80) 6 characters from the soname, plus the
+    // `vma_truncation_prefix`, so 9 characters total.
+    if (full_vma_length > kVmaNameLimit) {
+      const char vma_truncation_prefix[] = "...";
+      int soname_truncated_bytes =
+          full_vma_length - kVmaNameLimit + sizeof(vma_truncation_prefix) - 1;
+      async_safe_format_buffer(vma_name.data(), kVmaNameLimit, "mt:%s%s+%" PRIxPTR,
+                               vma_truncation_prefix, soname + soname_truncated_bytes,
+                               page_start(phdr->p_vaddr));
+    }
+    if (prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, reinterpret_cast<void*>(seg_page_start),
+              seg_page_aligned_size, vma_name.data()) != 0) {
+      DL_WARN("Failed to re-name memtag global segment.");
+    }
+  }
+}
+
 /* Change the protection of all loaded segments in memory to writable.
  * This is useful before performing relocations. Once completed, you
  * will have to call phdr_table_protect_segments to restore the original
diff --git a/linker/linker_phdr.h b/linker/linker_phdr.h
index 2f159f3..353970c 100644
--- a/linker/linker_phdr.h
+++ b/linker/linker_phdr.h
@@ -39,6 +39,8 @@
 #include "linker_mapped_file_fragment.h"
 #include "linker_note_gnu_property.h"
 
+#include <list>
+
 #define MAYBE_MAP_FLAG(x, from, to)  (((x) & (from)) ? (to) : 0)
 #define PFLAGS_TO_PROT(x)            (MAYBE_MAP_FLAG((x), PF_X, PROT_EXEC) | \
                                       MAYBE_MAP_FLAG((x), PF_R, PROT_READ) | \
@@ -188,3 +190,13 @@
                                             ElfW(Addr) load_bias);
 
 bool page_size_migration_supported();
+
+int remap_memtag_globals_segments(const ElfW(Phdr) * phdr_table, size_t phdr_count,
+                                  ElfW(Addr) load_bias);
+
+void protect_memtag_globals_ro_segments(const ElfW(Phdr) * phdr_table, size_t phdr_count,
+                                        ElfW(Addr) load_bias);
+
+void name_memtag_globals_segments(const ElfW(Phdr) * phdr_table, size_t phdr_count,
+                                  ElfW(Addr) load_bias, const char* soname,
+                                  std::list<std::string>& vma_names);
diff --git a/linker/linker_relocate.cpp b/linker/linker_relocate.cpp
index 0470f87..bbf8359 100644
--- a/linker/linker_relocate.cpp
+++ b/linker/linker_relocate.cpp
@@ -44,6 +44,8 @@
 #include "linker_soinfo.h"
 #include "private/bionic_globals.h"
 
+#include <platform/bionic/mte.h>
+
 static bool is_tls_reloc(ElfW(Word) type) {
   switch (type) {
     case R_GENERIC_TLS_DTPMOD:
@@ -163,7 +165,8 @@
 static bool process_relocation_impl(Relocator& relocator, const rel_t& reloc) {
   constexpr bool IsGeneral = Mode == RelocMode::General;
 
-  void* const rel_target = reinterpret_cast<void*>(reloc.r_offset + relocator.si->load_bias);
+  void* const rel_target = reinterpret_cast<void*>(
+      relocator.si->apply_memtag_if_mte_globals(reloc.r_offset + relocator.si->load_bias));
   const uint32_t r_type = ELFW(R_TYPE)(reloc.r_info);
   const uint32_t r_sym = ELFW(R_SYM)(reloc.r_info);
 
@@ -316,6 +319,7 @@
     // common in non-platform binaries.
     if (r_type == R_GENERIC_ABSOLUTE) {
       count_relocation_if<IsGeneral>(kRelocAbsolute);
+      if (found_in) sym_addr = found_in->apply_memtag_if_mte_globals(sym_addr);
       const ElfW(Addr) result = sym_addr + get_addend_rel();
       LD_DEBUG(reloc && IsGeneral, "RELO ABSOLUTE %16p <- %16p %s",
                rel_target, reinterpret_cast<void*>(result), sym_name);
@@ -326,6 +330,7 @@
       // document (IHI0044F) specifies that R_ARM_GLOB_DAT has an addend, but Bionic isn't adding
       // it.
       count_relocation_if<IsGeneral>(kRelocAbsolute);
+      if (found_in) sym_addr = found_in->apply_memtag_if_mte_globals(sym_addr);
       const ElfW(Addr) result = sym_addr + get_addend_norel();
       LD_DEBUG(reloc && IsGeneral, "RELO GLOB_DAT %16p <- %16p %s",
                rel_target, reinterpret_cast<void*>(result), sym_name);
@@ -335,7 +340,18 @@
       // In practice, r_sym is always zero, but if it weren't, the linker would still look up the
       // referenced symbol (and abort if the symbol isn't found), even though it isn't used.
       count_relocation_if<IsGeneral>(kRelocRelative);
-      const ElfW(Addr) result = relocator.si->load_bias + get_addend_rel();
+      ElfW(Addr) result = relocator.si->load_bias + get_addend_rel();
+      // MTE globals reuses the place bits for additional tag-derivation metadata for
+      // R_AARCH64_RELATIVE relocations, which makes it incompatible with
+      // `-Wl,--apply-dynamic-relocs`. This is enforced by lld, however there's nothing stopping
+      // Android binaries (particularly prebuilts) from building with this linker flag if they're
+      // not built with MTE globals. Thus, don't use the new relocation semantics if this DSO
+      // doesn't have MTE globals.
+      if (relocator.si->should_tag_memtag_globals()) {
+        int64_t* place = static_cast<int64_t*>(rel_target);
+        int64_t offset = *place;
+        result = relocator.si->apply_memtag_if_mte_globals(result + offset) - offset;
+      }
       LD_DEBUG(reloc && IsGeneral, "RELO RELATIVE %16p <- %16p",
                rel_target, reinterpret_cast<void*>(result));
       *static_cast<ElfW(Addr)*>(rel_target) = result;
@@ -600,7 +616,7 @@
     LD_DEBUG(reloc, "[ relocating %s relr ]", get_realpath());
     const ElfW(Relr)* begin = relr_;
     const ElfW(Relr)* end = relr_ + relr_count_;
-    if (!relocate_relr(begin, end, load_bias)) {
+    if (!relocate_relr(begin, end, load_bias, should_tag_memtag_globals())) {
       return false;
     }
   }
diff --git a/linker/linker_sleb128.h b/linker/linker_sleb128.h
index 6bb3199..f48fda8 100644
--- a/linker/linker_sleb128.h
+++ b/linker/linker_sleb128.h
@@ -69,3 +69,32 @@
   const uint8_t* current_;
   const uint8_t* const end_;
 };
+
+class uleb128_decoder {
+ public:
+  uleb128_decoder(const uint8_t* buffer, size_t count) : current_(buffer), end_(buffer + count) {}
+
+  uint64_t pop_front() {
+    uint64_t value = 0;
+
+    size_t shift = 0;
+    uint8_t byte;
+
+    do {
+      if (current_ >= end_) {
+        async_safe_fatal("uleb128_decoder ran out of bounds");
+      }
+      byte = *current_++;
+      value |= (static_cast<size_t>(byte & 127) << shift);
+      shift += 7;
+    } while (byte & 128);
+
+    return value;
+  }
+
+  bool has_bytes() { return current_ < end_; }
+
+ private:
+  const uint8_t* current_;
+  const uint8_t* const end_;
+};
diff --git a/linker/linker_soinfo.cpp b/linker/linker_soinfo.cpp
index 0549d36..176c133 100644
--- a/linker/linker_soinfo.cpp
+++ b/linker/linker_soinfo.cpp
@@ -44,6 +44,8 @@
 #include "linker_logger.h"
 #include "linker_relocate.h"
 #include "linker_utils.h"
+#include "platform/bionic/mte.h"
+#include "private/bionic_globals.h"
 
 SymbolLookupList::SymbolLookupList(soinfo* si)
     : sole_lib_(si->get_lookup_lib()), begin_(&sole_lib_), end_(&sole_lib_ + 1) {
@@ -304,6 +306,12 @@
   return is_gnu_hash() ? gnu_lookup(symbol_name, vi) : elf_lookup(symbol_name, vi);
 }
 
+ElfW(Addr) soinfo::apply_memtag_if_mte_globals(ElfW(Addr) sym_addr) const {
+  if (!should_tag_memtag_globals()) return sym_addr;
+  if (sym_addr == 0) return sym_addr;  // Handle undefined weak symbols.
+  return reinterpret_cast<ElfW(Addr)>(get_tagged_address(reinterpret_cast<void*>(sym_addr)));
+}
+
 const ElfW(Sym)* soinfo::gnu_lookup(SymbolName& symbol_name, const version_info* vi) const {
   const uint32_t hash = symbol_name.gnu_hash();
 
diff --git a/linker/linker_soinfo.h b/linker/linker_soinfo.h
index a776c1f..6b81f5f 100644
--- a/linker/linker_soinfo.h
+++ b/linker/linker_soinfo.h
@@ -30,6 +30,7 @@
 
 #include <link.h>
 
+#include <list>
 #include <memory>
 #include <string>
 #include <vector>
@@ -66,9 +67,10 @@
                                          // soinfo is executed and this flag is
                                          // unset.
 #define FLAG_PRELINKED        0x00000400 // prelink_image has successfully processed this soinfo
+#define FLAG_GLOBALS_TAGGED   0x00000800 // globals have been tagged by MTE.
 #define FLAG_NEW_SOINFO       0x40000000 // new soinfo format
 
-#define SOINFO_VERSION 6
+#define SOINFO_VERSION 7
 
 ElfW(Addr) call_ifunc_resolver(ElfW(Addr) resolver_addr);
 
@@ -257,6 +259,9 @@
                   const android_dlextinfo* extinfo, size_t* relro_fd_offset);
   bool protect_relro();
 
+  void tag_globals();
+  ElfW(Addr) apply_memtag_if_mte_globals(ElfW(Addr) sym_addr) const;
+
   void add_child(soinfo* child);
   void remove_all_links();
 
@@ -368,6 +373,10 @@
    should_pad_segments_ = should_pad_segments;
   }
   bool should_pad_segments() const { return should_pad_segments_; }
+  bool should_tag_memtag_globals() const {
+    return !is_linker() && memtag_globals() && memtag_globalssz() > 0 && __libc_mte_enabled();
+  }
+  std::list<std::string>& vma_names() { return vma_names_; };
 
   void set_should_use_16kib_app_compat(bool should_use_16kib_app_compat) {
     should_use_16kib_app_compat_ = should_use_16kib_app_compat;
@@ -464,6 +473,8 @@
   // version >= 7
   memtag_dynamic_entries_t memtag_dynamic_entries_;
 
+  std::list<std::string> vma_names_;
+
   // Pad gaps between segments when memory mapping?
   bool should_pad_segments_ = false;