bool
tls_thread_preinit()
{
#ifdef X64
/* i#3356: Write a non-zero value to the gs base to work around an AMD bug
* present on pre-4.7 Linux kernels. See the call to this in our signal
* handler for more information.
*/
if (proc_get_vendor() != VENDOR_AMD)
return true;
/* First identify a temp-native thread with a real segment in
* place but just an invalid .magic field. We do not want to clobber the
* legitimate segment base in that case.
*/
if (safe_read_tls_magic() == TLS_MAGIC_INVALID) {
os_local_state_t *tls = (os_local_state_t *)safe_read_tls_self();
if (tls != NULL &&
tls->state.spill_space.dcontext->owning_thread == get_sys_thread_id())
return true;
}
/* XXX: What about Mac on AMD? Presumably by the time anyone wants to run
* that combination the Mac kernel will have fixed this if they haven't already.
*/
/* We just don't have time to support non-arch_prctl and test it. */
if (tls_global_type != TLS_TYPE_ARCH_PRCTL) {
ASSERT_BUG_NUM(3356, tls_global_type == TLS_TYPE_ARCH_PRCTL);
return false;
}
int res = dynamorio_syscall(SYS_arch_prctl, 2, ARCH_SET_GS, NON_ZERO_UNINIT_GSBASE);
LOG(GLOBAL, LOG_THREADS, 1,
"%s: set non-zero pre-init gs base for thread " TIDFMT "\n", __FUNCTION__,
get_sys_thread_id());
return res == 0;
#else
return true;
#endif
}
void
tls_thread_init(os_local_state_t *os_tls, byte *segment)
{
/* We have four different ways to obtain TLS, each with its own limitations:
*
* 1) Piggyback on the threading system (like we do on Windows): here that would
* be pthreads, which uses a segment since at least RH9, and uses gdt-based
* segments for NPTL. The advantage is we won't run out of ldt or gdt entries
* (except when the app itself would). The disadvantage is we're stealing
* application slots and we rely on user mode interfaces.
*
* 2) Steal an ldt entry via SYS_modify_ldt. This suffers from the 8K ldt entry
* limit and requires that we update manually on a new thread. For 64-bit
* we're limited here to a 32-bit base. (Strangely, the kernel's
* include/asm-x86_64/ldt.h implies that the base is ignored: but it doesn't
* seem to be.)
*
* 3) Steal a gdt entry via SYS_set_thread_area. There is a 3rd unused entry
* (after pthreads and wine) we could use. The kernel swaps for us, and with
* CLONE_TLS the kernel will set up the entry for a new thread for us. Xref
* PR 192231 and PR 285898. This system call is disabled on 64-bit 2.6
* kernels (though the man page for arch_prctl implies it isn't for 2.5
* kernels?!?)
*
* 4) Use SYS_arch_prctl. This is only implemented on 64-bit kernels, and can
* only be used to set the gdt entries that fs and gs select for. Faster to
* use <4GB base (obtain with mmap MAP_32BIT) since can use gdt; else have to
* use wrmsr. The man pages say "ARCH_SET_GS is disabled in some kernels".
*/
uint selector;
int index = -1;
int res;
#ifdef X64
/* First choice is gdt, which means arch_prctl. Since this may fail
* on some kernels, we require -heap_in_lower_4GB so we can fall back
* on modify_ldt.
*/
byte *cur_gs;
res = dynamorio_syscall(SYS_arch_prctl, 2, ARCH_GET_GS, &cur_gs);
if (res >= 0) {
LOG(GLOBAL, LOG_THREADS, 1, "os_tls_init: cur gs base is "PFX"\n", cur_gs);
/* If we're a non-initial thread, gs will be set to the parent thread's value */
if (cur_gs == NULL || is_dynamo_address(cur_gs) ||
/* By resolving i#107, we can handle gs conflicts between app and dr. */
INTERNAL_OPTION(mangle_app_seg)) {
res = dynamorio_syscall(SYS_arch_prctl, 2, ARCH_SET_GS, segment);
if (res >= 0) {
os_tls->tls_type = TLS_TYPE_ARCH_PRCTL;
LOG(GLOBAL, LOG_THREADS, 1,
"os_tls_init: arch_prctl successful for base "PFX"\n", segment);
/* Kernel should have written %gs for us if using GDT */
if (!dynamo_initialized && read_thread_register(SEG_TLS) == 0) {
LOG(GLOBAL, LOG_THREADS, 1, "os_tls_init: using MSR\n");
tls_using_msr = true;
}
if (IF_CLIENT_INTERFACE_ELSE(INTERNAL_OPTION(private_loader), false)) {
res = dynamorio_syscall(SYS_arch_prctl, 2, ARCH_SET_FS,
os_tls->os_seg_info.priv_lib_tls_base);
/* Assuming set fs must be successful if set gs succeeded. */
ASSERT(res >= 0);
}
} else {
/* we've found a kernel where ARCH_SET_GS is disabled */
ASSERT_CURIOSITY(false && "arch_prctl failed on set but not get");
LOG(GLOBAL, LOG_THREADS, 1,
"os_tls_init: arch_prctl failed: error %d\n", res);
}
} else {
/* FIXME PR 205276: we don't currently handle it: fall back on ldt, but
* we'll have the same conflict w/ the selector...
*/
ASSERT_BUG_NUM(205276, cur_gs == NULL);
}
}
#endif
if (os_tls->tls_type == TLS_TYPE_NONE) {
/* Second choice is set_thread_area */
/* PR 285898: if we added CLONE_SETTLS to all clone calls (and emulated vfork
* with clone) we could avoid having to set tls up for each thread (as well
* as solve race PR 207903), at least for kernel 2.5.32+. For now we stick
* w/ manual setup.
*/
our_modify_ldt_t desc;
/* Pick which GDT slots we'll use for DR TLS and for library TLS if
* using the private loader.
*/
choose_gdt_slots(os_tls);
if (tls_gdt_index > -1) {
/* Now that we know which GDT slot to use, install the per-thread base
* into it.
*/
/* Base here must be 32-bit */
IF_X64(ASSERT(DYNAMO_OPTION(heap_in_lower_4GB) &&
segment <= (byte*)UINT_MAX));
initialize_ldt_struct(&desc, segment, PAGE_SIZE, tls_gdt_index);
res = dynamorio_syscall(SYS_set_thread_area, 1, &desc);
LOG(GLOBAL, LOG_THREADS, 3,
"%s: set_thread_area %d => %d res, %d index\n",
__FUNCTION__, tls_gdt_index, res, desc.entry_number);
ASSERT(res < 0 || desc.entry_number == tls_gdt_index);
} else {
res = -1; /* fall back on LDT */
}
if (res >= 0) {
LOG(GLOBAL, LOG_THREADS, 1,
"os_tls_init: set_thread_area successful for base "PFX" @index %d\n",
segment, tls_gdt_index);
os_tls->tls_type = TLS_TYPE_GDT;
index = tls_gdt_index;
selector = GDT_SELECTOR(index);
WRITE_DR_SEG(selector); /* macro needs lvalue! */
} else {
IF_VMX86(ASSERT_NOT_REACHED()); /* since no modify_ldt */
LOG(GLOBAL, LOG_THREADS, 1,
"os_tls_init: set_thread_area failed: error %d\n", res);
}
#ifdef CLIENT_INTERFACE
/* Install the library TLS base. */
if (INTERNAL_OPTION(private_loader) && res >= 0) {
app_pc base = os_tls->os_seg_info.priv_lib_tls_base;
/* lib_tls_gdt_index is picked in choose_gdt_slots. */
ASSERT(lib_tls_gdt_index >= gdt_entry_tls_min);
initialize_ldt_struct(&desc, base, GDT_NO_SIZE_LIMIT,
lib_tls_gdt_index);
res = dynamorio_syscall(SYS_set_thread_area, 1, &desc);
LOG(GLOBAL, LOG_THREADS, 3,
"%s: set_thread_area %d => %d res, %d index\n",
__FUNCTION__, lib_tls_gdt_index, res, desc.entry_number);
if (res >= 0) {
/* i558 update lib seg reg to enforce the segment changes */
selector = GDT_SELECTOR(lib_tls_gdt_index);
LOG(GLOBAL, LOG_THREADS, 2, "%s: setting %s to selector 0x%x\n",
__FUNCTION__, reg_names[LIB_SEG_TLS], selector);
WRITE_LIB_SEG(selector);
}
}
#endif
}
if (os_tls->tls_type == TLS_TYPE_NONE) {
/* Third choice: modify_ldt, which should be available on kernel 2.3.99+ */
/* Base here must be 32-bit */
IF_X64(ASSERT(DYNAMO_OPTION(heap_in_lower_4GB) && segment <= (byte*)UINT_MAX));
/* we have the thread_initexit_lock so no race here */
index = find_unused_ldt_index();
selector = LDT_SELECTOR(index);
ASSERT(index != -1);
create_ldt_entry((void *)segment, PAGE_SIZE, index);
os_tls->tls_type = TLS_TYPE_LDT;
WRITE_DR_SEG(selector); /* macro needs lvalue! */
LOG(GLOBAL, LOG_THREADS, 1,
"os_tls_init: modify_ldt successful for base "PFX" w/ index %d\n",
segment, index);
}
os_tls->ldt_index = index;
}
