Index: head/libexec/rtld-elf/aarch64/rtld_machdep.h
===================================================================
--- head/libexec/rtld-elf/aarch64/rtld_machdep.h
+++ head/libexec/rtld-elf/aarch64/rtld_machdep.h
@@ -69,6 +69,8 @@
 #define	calculate_tls_offset(prev_offset, prev_size, size, align) \
 	round(prev_offset + prev_size, align)
 #define	calculate_tls_end(off, size) 	((off) + (size))
+#define calculate_tls_post_size(align) \
+	round(TLS_TCB_SIZE, align) - TLS_TCB_SIZE
 
 #define	TLS_TCB_SIZE	16
 typedef struct {
Index: head/libexec/rtld-elf/arm/rtld_machdep.h
===================================================================
--- head/libexec/rtld-elf/arm/rtld_machdep.h
+++ head/libexec/rtld-elf/arm/rtld_machdep.h
@@ -69,6 +69,8 @@
 #define calculate_tls_offset(prev_offset, prev_size, size, align) \
     round(prev_offset + prev_size, align)
 #define calculate_tls_end(off, size)    ((off) + (size))
+#define calculate_tls_post_size(align) \
+    round(TLS_TCB_SIZE, align) - TLS_TCB_SIZE
 	
 extern void *__tls_get_addr(tls_index *ti);
 
Index: head/libexec/rtld-elf/mips/rtld_machdep.h
===================================================================
--- head/libexec/rtld-elf/mips/rtld_machdep.h
+++ head/libexec/rtld-elf/mips/rtld_machdep.h
@@ -64,10 +64,11 @@
 #define round(size, align) \
     (((size) + (align) - 1) & ~((align) - 1))
 #define calculate_first_tls_offset(size, align) \
-    round(TLS_TCB_SIZE, align)
+    TLS_TCB_SIZE
 #define calculate_tls_offset(prev_offset, prev_size, size, align) \
     round(prev_offset + prev_size, align)
 #define calculate_tls_end(off, size)    ((off) + (size))
+#define calculate_tls_post_size(align)  0
 
 extern void *__tls_get_addr(tls_index *ti);
 
Index: head/libexec/rtld-elf/powerpc/rtld_machdep.h
===================================================================
--- head/libexec/rtld-elf/powerpc/rtld_machdep.h
+++ head/libexec/rtld-elf/powerpc/rtld_machdep.h
@@ -74,10 +74,11 @@
 #define round(size, align) \
     (((size) + (align) - 1) & ~((align) - 1))
 #define calculate_first_tls_offset(size, align) \
-    round(8, align)
+    TLS_TCB_SIZE
 #define calculate_tls_offset(prev_offset, prev_size, size, align) \
     round(prev_offset + prev_size, align)
 #define calculate_tls_end(off, size)    ((off) + (size))
+#define calculate_tls_post_size(align)  0
  
 typedef struct {
 	unsigned long ti_module;
Index: head/libexec/rtld-elf/powerpc64/rtld_machdep.h
===================================================================
--- head/libexec/rtld-elf/powerpc64/rtld_machdep.h
+++ head/libexec/rtld-elf/powerpc64/rtld_machdep.h
@@ -66,10 +66,11 @@
 #define round(size, align) \
     (((size) + (align) - 1) & ~((align) - 1))
 #define calculate_first_tls_offset(size, align) \
-    round(16, align)
+    TLS_TCB_SIZE
 #define calculate_tls_offset(prev_offset, prev_size, size, align) \
     round(prev_offset + prev_size, align)
 #define calculate_tls_end(off, size)    ((off) + (size))
+#define calculate_tls_post_size(align)  0
  
 typedef struct {
 	unsigned long ti_module;
Index: head/libexec/rtld-elf/riscv/rtld_machdep.h
===================================================================
--- head/libexec/rtld-elf/riscv/rtld_machdep.h
+++ head/libexec/rtld-elf/riscv/rtld_machdep.h
@@ -89,10 +89,11 @@
 #define round(size, align) \
     (((size) + (align) - 1) & ~((align) - 1))
 #define calculate_first_tls_offset(size, align) \
-    round(16, align)
+    TLS_TCB_SIZE
 #define calculate_tls_offset(prev_offset, prev_size, size, align) \
     round(prev_offset + prev_size, align)
 #define calculate_tls_end(off, size)    ((off) + (size))
+#define calculate_tls_post_size(align)  0
 
 typedef struct {
 	unsigned long ti_module;
Index: head/libexec/rtld-elf/rtld.c
===================================================================
--- head/libexec/rtld-elf/rtld.c
+++ head/libexec/rtld-elf/rtld.c
@@ -4693,47 +4693,87 @@
     defined(__powerpc__) || defined(__riscv)
 
 /*
+ * Return pointer to allocated TLS block
+ */
+static void *
+get_tls_block_ptr(void *tcb, size_t tcbsize)
+{
+    size_t extra_size, post_size, pre_size, tls_block_size;
+    size_t tls_init_align;
+
+    tls_init_align = MAX(obj_main->tlsalign, 1);
+
+    /* Compute fragments sizes. */
+    extra_size = tcbsize - TLS_TCB_SIZE;
+    post_size = calculate_tls_post_size(tls_init_align);
+    tls_block_size = tcbsize + post_size;
+    pre_size = roundup2(tls_block_size, tls_init_align) - tls_block_size;
+
+    return ((char *)tcb - pre_size - extra_size);
+}
+
+/*
  * Allocate Static TLS using the Variant I method.
+ *
+ * For details on the layout, see lib/libc/gen/tls.c.
+ *
+ * NB: rtld's tls_static_space variable includes TLS_TCB_SIZE and post_size as
+ *     it is based on tls_last_offset, and TLS offsets here are really TCB
+ *     offsets, whereas libc's tls_static_space is just the executable's static
+ *     TLS segment.
  */
 void *
 allocate_tls(Obj_Entry *objs, void *oldtcb, size_t tcbsize, size_t tcbalign)
 {
     Obj_Entry *obj;
-    char *tcb;
-    Elf_Addr **tls;
-    Elf_Addr *dtv;
+    char *tls_block;
+    Elf_Addr *dtv, **tcb;
     Elf_Addr addr;
     int i;
+    size_t extra_size, maxalign, post_size, pre_size, tls_block_size;
+    size_t tls_init_align;
 
     if (oldtcb != NULL && tcbsize == TLS_TCB_SIZE)
 	return (oldtcb);
 
     assert(tcbsize >= TLS_TCB_SIZE);
-    tcb = xcalloc(1, tls_static_space - TLS_TCB_SIZE + tcbsize);
-    tls = (Elf_Addr **)(tcb + tcbsize - TLS_TCB_SIZE);
+    maxalign = MAX(tcbalign, tls_static_max_align);
+    tls_init_align = MAX(obj_main->tlsalign, 1);
 
+    /* Compute fragmets sizes. */
+    extra_size = tcbsize - TLS_TCB_SIZE;
+    post_size = calculate_tls_post_size(tls_init_align);
+    tls_block_size = tcbsize + post_size;
+    pre_size = roundup2(tls_block_size, tls_init_align) - tls_block_size;
+    tls_block_size += pre_size + tls_static_space - TLS_TCB_SIZE - post_size;
+
+    /* Allocate whole TLS block */
+    tls_block = malloc_aligned(tls_block_size, maxalign);
+    tcb = (Elf_Addr **)(tls_block + pre_size + extra_size);
+
     if (oldtcb != NULL) {
-	memcpy(tls, oldtcb, tls_static_space);
-	free(oldtcb);
+	memcpy(tls_block, get_tls_block_ptr(oldtcb, tcbsize),
+	    tls_static_space);
+	free_aligned(get_tls_block_ptr(oldtcb, tcbsize));
 
 	/* Adjust the DTV. */
-	dtv = tls[0];
+	dtv = tcb[0];
 	for (i = 0; i < dtv[1]; i++) {
 	    if (dtv[i+2] >= (Elf_Addr)oldtcb &&
 		dtv[i+2] < (Elf_Addr)oldtcb + tls_static_space) {
-		dtv[i+2] = dtv[i+2] - (Elf_Addr)oldtcb + (Elf_Addr)tls;
+		dtv[i+2] = dtv[i+2] - (Elf_Addr)oldtcb + (Elf_Addr)tcb;
 	    }
 	}
     } else {
 	dtv = xcalloc(tls_max_index + 2, sizeof(Elf_Addr));
-	tls[0] = dtv;
+	tcb[0] = dtv;
 	dtv[0] = tls_dtv_generation;
 	dtv[1] = tls_max_index;
 
 	for (obj = globallist_curr(objs); obj != NULL;
 	  obj = globallist_next(obj)) {
 	    if (obj->tlsoffset > 0) {
-		addr = (Elf_Addr)tls + obj->tlsoffset;
+		addr = (Elf_Addr)tcb + obj->tlsoffset;
 		if (obj->tlsinitsize > 0)
 		    memcpy((void*) addr, obj->tlsinit, obj->tlsinitsize);
 		if (obj->tlssize > obj->tlsinitsize)
@@ -4752,14 +4792,19 @@
 {
     Elf_Addr *dtv;
     Elf_Addr tlsstart, tlsend;
-    int dtvsize, i;
+    size_t post_size;
+    size_t dtvsize, i, tls_init_align;
 
     assert(tcbsize >= TLS_TCB_SIZE);
+    tls_init_align = MAX(obj_main->tlsalign, 1);
 
-    tlsstart = (Elf_Addr)tcb + tcbsize - TLS_TCB_SIZE;
-    tlsend = tlsstart + tls_static_space;
+    /* Compute fragments sizes. */
+    post_size = calculate_tls_post_size(tls_init_align);
 
-    dtv = *(Elf_Addr **)tlsstart;
+    tlsstart = (Elf_Addr)tcb + TLS_TCB_SIZE + post_size;
+    tlsend = (Elf_Addr)tcb + tls_static_space;
+
+    dtv = *(Elf_Addr **)tcb;
     dtvsize = dtv[1];
     for (i = 0; i < dtvsize; i++) {
 	if (dtv[i+2] && (dtv[i+2] < tlsstart || dtv[i+2] >= tlsend)) {
@@ -4767,7 +4812,7 @@
 	}
     }
     free(dtv);
-    free(tcb);
+    free_aligned(get_tls_block_ptr(tcb, tcbsize));
 }
 
 #endif