int64_t qemu_get_clock_ns(QEMUClock *clock)
{
switch(clock->type) {
case QEMU_CLOCK_REALTIME:
return get_clock();
default:
case QEMU_CLOCK_VIRTUAL:
return cpu_get_clock();
case QEMU_CLOCK_HOST:
return get_clock_realtime();
}
}
static QEMUClock *qemu_new_clock(int type)
{
QEMUClock *clock;
clock = qemu_mallocz(sizeof(QEMUClock));
clock->type = type;
clock->enabled = 1;
notifier_list_init(&clock->reset_notifiers);
/* required to detect & report backward jumps */
if (type == QEMU_CLOCK_HOST) {
clock->last = get_clock_realtime();
}
return clock;
}
static void tlb_dyn_init(CPUArchState *env)
{
int i;
for (i = 0; i < NB_MMU_MODES; i++) {
CPUTLBDesc *desc = &env->tlb_d[i];
size_t n_entries = 1 << CPU_TLB_DYN_DEFAULT_BITS;
tlb_window_reset(&desc->window, get_clock_realtime(), 0);
desc->n_used_entries = 0;
env->tlb_mask[i] = (n_entries - 1) << CPU_TLB_ENTRY_BITS;
env->tlb_table[i] = g_new(CPUTLBEntry, n_entries);
env->iotlb[i] = g_new(CPUIOTLBEntry, n_entries);
}
}
int64_t qemu_get_clock(QEMUClock *clock)
{
switch(clock->type) {
case QEMU_CLOCK_REALTIME:
return get_clock() / 1000000;
default:
case QEMU_CLOCK_VIRTUAL:
if (use_icount) {
return cpu_get_icount();
} else {
return cpu_get_clock();
}
case QEMU_CLOCK_HOST:
return get_clock_realtime();
}
}
static int64_t get_clock(void)
{
#if defined(__linux__) || (defined(__FreeBSD__) && __FreeBSD_version >= 500000) \
|| defined(__DragonFly__) || defined(__FreeBSD_kernel__)
if (use_rt_clock) {
struct timespec ts;
clock_gettime(CLOCK_MONOTONIC, &ts);
return ts.tv_sec * 1000000000LL + ts.tv_nsec;
} else
#endif
{
/* XXX: using gettimeofday leads to problems if the date
changes, so it should be avoided. */
return get_clock_realtime();
}
}
int64_t qemu_get_clock(QEMUClock *clock)
{
switch(clock->type) {
case QEMU_CLOCK_REALTIME:
return get_clock() / 1000000;
default:
case QEMU_CLOCK_VIRTUAL:
#ifdef MARSS_QEMU
if (in_simulation) {
return cpu_get_sim_clock();
}
#endif
if (use_icount) {
return cpu_get_icount();
} else {
return cpu_get_clock();
}
case QEMU_CLOCK_HOST:
return get_clock_realtime();
}
}
int64_t qemu_get_clock_ns(QEMUClock *clock)
{
int64_t now, last;
switch(clock->type) {
case QEMU_CLOCK_REALTIME:
return get_clock();
default:
case QEMU_CLOCK_VIRTUAL:
if (use_icount) {
return cpu_get_icount();
} else {
return cpu_get_clock();
}
case QEMU_CLOCK_HOST:
now = get_clock_realtime();
last = clock->last;
clock->last = now;
if (now < last) {
notifier_list_notify(&clock->reset_notifiers, &now);
}
return now;
}
}
/**
* tlb_mmu_resize_locked() - perform TLB resize bookkeeping; resize if necessary
* @env: CPU that owns the TLB
* @mmu_idx: MMU index of the TLB
*
* Called with tlb_lock_held.
*
* We have two main constraints when resizing a TLB: (1) we only resize it
* on a TLB flush (otherwise we'd have to take a perf hit by either rehashing
* the array or unnecessarily flushing it), which means we do not control how
* frequently the resizing can occur; (2) we don't have access to the guest's
* future scheduling decisions, and therefore have to decide the magnitude of
* the resize based on past observations.
*
* In general, a memory-hungry process can benefit greatly from an appropriately
* sized TLB, since a guest TLB miss is very expensive. This doesn't mean that
* we just have to make the TLB as large as possible; while an oversized TLB
* results in minimal TLB miss rates, it also takes longer to be flushed
* (flushes can be _very_ frequent), and the reduced locality can also hurt
* performance.
*
* To achieve near-optimal performance for all kinds of workloads, we:
*
* 1. Aggressively increase the size of the TLB when the use rate of the
* TLB being flushed is high, since it is likely that in the near future this
* memory-hungry process will execute again, and its memory hungriness will
* probably be similar.
*
* 2. Slowly reduce the size of the TLB as the use rate declines over a
* reasonably large time window. The rationale is that if in such a time window
* we have not observed a high TLB use rate, it is likely that we won't observe
* it in the near future. In that case, once a time window expires we downsize
* the TLB to match the maximum use rate observed in the window.
*
* 3. Try to keep the maximum use rate in a time window in the 30-70% range,
* since in that range performance is likely near-optimal. Recall that the TLB
* is direct mapped, so we want the use rate to be low (or at least not too
* high), since otherwise we are likely to have a significant amount of
* conflict misses.
*/
static void tlb_mmu_resize_locked(CPUArchState *env, int mmu_idx)
{
CPUTLBDesc *desc = &env->tlb_d[mmu_idx];
size_t old_size = tlb_n_entries(env, mmu_idx);
size_t rate;
size_t new_size = old_size;
int64_t now = get_clock_realtime();
int64_t window_len_ms = 100;
int64_t window_len_ns = window_len_ms * 1000 * 1000;
bool window_expired = now > desc->window.begin_ns + window_len_ns;
if (desc->n_used_entries > desc->window.max_entries) {
desc->window.max_entries = desc->n_used_entries;
}
rate = desc->window.max_entries * 100 / old_size;
if (rate > 70) {
new_size = MIN(old_size << 1, 1 << CPU_TLB_DYN_MAX_BITS);
} else if (rate < 30 && window_expired) {
size_t ceil = pow2ceil(desc->window.max_entries);
size_t expected_rate = desc->window.max_entries * 100 / ceil;
/*
* Avoid undersizing when the max number of entries seen is just below
* a pow2. For instance, if max_entries == 1025, the expected use rate
* would be 1025/2048==50%. However, if max_entries == 1023, we'd get
* 1023/1024==99.9% use rate, so we'd likely end up doubling the size
* later. Thus, make sure that the expected use rate remains below 70%.
* (and since we double the size, that means the lowest rate we'd
* expect to get is 35%, which is still in the 30-70% range where
* we consider that the size is appropriate.)
*/
if (expected_rate > 70) {
ceil *= 2;
}
new_size = MAX(ceil, 1 << CPU_TLB_DYN_MIN_BITS);
}
if (new_size == old_size) {
if (window_expired) {
tlb_window_reset(&desc->window, now, desc->n_used_entries);
}
return;
}
g_free(env->tlb_table[mmu_idx]);
g_free(env->iotlb[mmu_idx]);
tlb_window_reset(&desc->window, now, 0);
/* desc->n_used_entries is cleared by the caller */
env->tlb_mask[mmu_idx] = (new_size - 1) << CPU_TLB_ENTRY_BITS;
env->tlb_table[mmu_idx] = g_try_new(CPUTLBEntry, new_size);
env->iotlb[mmu_idx] = g_try_new(CPUIOTLBEntry, new_size);
/*
* If the allocations fail, try smaller sizes. We just freed some
* memory, so going back to half of new_size has a good chance of working.
* Increased memory pressure elsewhere in the system might cause the
* allocations to fail though, so we progressively reduce the allocation
* size, aborting if we cannot even allocate the smallest TLB we support.
*/
while (env->tlb_table[mmu_idx] == NULL || env->iotlb[mmu_idx] == NULL) {
if (new_size == (1 << CPU_TLB_DYN_MIN_BITS)) {
error_report("%s: %s", __func__, strerror(errno));
abort();
}
new_size = MAX(new_size >> 1, 1 << CPU_TLB_DYN_MIN_BITS);
env->tlb_mask[mmu_idx] = (new_size - 1) << CPU_TLB_ENTRY_BITS;
g_free(env->tlb_table[mmu_idx]);
g_free(env->iotlb[mmu_idx]);
env->tlb_table[mmu_idx] = g_try_new(CPUTLBEntry, new_size);
env->iotlb[mmu_idx] = g_try_new(CPUIOTLBEntry, new_size);
}
}
