* [Qemu-devel] [PATCH v5 1/3] cputlb: do not evict empty entries to the vtlb
2018-12-14 17:42 [Qemu-devel] [PATCH v5 0/3] Dynamic TLB sizing Emilio G. Cota
@ 2018-12-14 17:42 ` Emilio G. Cota
2018-12-14 17:42 ` [Qemu-devel] [PATCH v5 2/3] tcg: introduce dynamic TLB sizing Emilio G. Cota
2018-12-14 17:42 ` [Qemu-devel] [PATCH v5 3/3] tcg/i386: enable " Emilio G. Cota
2 siblings, 0 replies; 6+ messages in thread
From: Emilio G. Cota @ 2018-12-14 17:42 UTC (permalink / raw)
To: qemu-devel; +Cc: Richard Henderson, Alex Bennée
Currently we evict an entry to the victim TLB when it doesn't match
the current address. But it could be that there's no match because
the current entry is empty (i.e. all -1's, for instance via tlb_flush).
Do not evict the entry to the vtlb in that case.
This change will help us keep track of the TLB's use rate, which
we'll use to implement a policy for dynamic TLB sizing.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
---
include/exec/cpu-all.h | 9 +++++++++
accel/tcg/cputlb.c | 2 +-
2 files changed, 10 insertions(+), 1 deletion(-)
diff --git a/include/exec/cpu-all.h b/include/exec/cpu-all.h
index 117d2fbbca..e21140049b 100644
--- a/include/exec/cpu-all.h
+++ b/include/exec/cpu-all.h
@@ -362,6 +362,15 @@ static inline bool tlb_hit(target_ulong tlb_addr, target_ulong addr)
return tlb_hit_page(tlb_addr, addr & TARGET_PAGE_MASK);
}
+/**
+ * tlb_entry_is_empty - return true if the entry is not in use
+ * @te: pointer to CPUTLBEntry
+ */
+static inline bool tlb_entry_is_empty(const CPUTLBEntry *te)
+{
+ return te->addr_read == -1 && te->addr_write == -1 && te->addr_code == -1;
+}
+
void dump_exec_info(FILE *f, fprintf_function cpu_fprintf);
void dump_opcount_info(FILE *f, fprintf_function cpu_fprintf);
#endif /* !CONFIG_USER_ONLY */
diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c
index af6bd8ccf9..5dc97212a9 100644
--- a/accel/tcg/cputlb.c
+++ b/accel/tcg/cputlb.c
@@ -591,7 +591,7 @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr,
* Only evict the old entry to the victim tlb if it's for a
* different page; otherwise just overwrite the stale data.
*/
- if (!tlb_hit_page_anyprot(te, vaddr_page)) {
+ if (!tlb_hit_page_anyprot(te, vaddr_page) && !tlb_entry_is_empty(te)) {
unsigned vidx = env->tlb_d[mmu_idx].vindex++ % CPU_VTLB_SIZE;
CPUTLBEntry *tv = &env->tlb_v_table[mmu_idx][vidx];
--
2.17.1
^ permalink raw reply related [flat|nested] 6+ messages in thread* [Qemu-devel] [PATCH v5 2/3] tcg: introduce dynamic TLB sizing
2018-12-14 17:42 [Qemu-devel] [PATCH v5 0/3] Dynamic TLB sizing Emilio G. Cota
2018-12-14 17:42 ` [Qemu-devel] [PATCH v5 1/3] cputlb: do not evict empty entries to the vtlb Emilio G. Cota
@ 2018-12-14 17:42 ` Emilio G. Cota
2018-12-17 22:41 ` Emilio G. Cota
2018-12-26 6:51 ` Richard Henderson
2018-12-14 17:42 ` [Qemu-devel] [PATCH v5 3/3] tcg/i386: enable " Emilio G. Cota
2 siblings, 2 replies; 6+ messages in thread
From: Emilio G. Cota @ 2018-12-14 17:42 UTC (permalink / raw)
To: qemu-devel; +Cc: Richard Henderson, Alex Bennée
Disabled in all TCG backends for now.
Signed-off-by: Emilio G. Cota <cota@braap.org>
---
include/exec/cpu-defs.h | 48 +++++++++-
include/exec/cpu_ldst.h | 21 +++++
tcg/aarch64/tcg-target.h | 1 +
tcg/arm/tcg-target.h | 1 +
tcg/i386/tcg-target.h | 1 +
tcg/mips/tcg-target.h | 1 +
tcg/ppc/tcg-target.h | 1 +
tcg/s390/tcg-target.h | 1 +
tcg/sparc/tcg-target.h | 1 +
tcg/tci/tcg-target.h | 1 +
accel/tcg/cputlb.c | 184 +++++++++++++++++++++++++++++++++++++--
11 files changed, 254 insertions(+), 7 deletions(-)
diff --git a/include/exec/cpu-defs.h b/include/exec/cpu-defs.h
index 6a60f94a41..b14a00b027 100644
--- a/include/exec/cpu-defs.h
+++ b/include/exec/cpu-defs.h
@@ -67,6 +67,19 @@ typedef uint64_t target_ulong;
#define CPU_TLB_ENTRY_BITS 5
#endif
+#if TCG_TARGET_IMPLEMENTS_DYN_TLB
+#define CPU_TLB_DYN_MIN_BITS 6
+#define CPU_TLB_DYN_DEFAULT_BITS 8
+/*
+ * Assuming TARGET_PAGE_BITS==12, with 2**22 entries we can cover 2**(22+12) ==
+ * 2**34 == 16G of address space. This is roughly what one would expect a
+ * TLB to cover in a modern (as of 2018) x86_64 CPU. For instance, Intel
+ * Skylake's Level-2 STLB has 16 1G entries.
+ */
+#define CPU_TLB_DYN_MAX_BITS 22
+
+#else /* !TCG_TARGET_IMPLEMENTS_DYN_TLB */
+
/* TCG_TARGET_TLB_DISPLACEMENT_BITS is used in CPU_TLB_BITS to ensure that
* the TLB is not unnecessarily small, but still small enough for the
* TLB lookup instruction sequence used by the TCG target.
@@ -98,6 +111,7 @@ typedef uint64_t target_ulong;
NB_MMU_MODES <= 8 ? 3 : 4))
#define CPU_TLB_SIZE (1 << CPU_TLB_BITS)
+#endif /* TCG_TARGET_IMPLEMENTS_DYN_TLB */
typedef struct CPUTLBEntry {
/* bit TARGET_LONG_BITS to TARGET_PAGE_BITS : virtual address
@@ -141,6 +155,18 @@ typedef struct CPUIOTLBEntry {
MemTxAttrs attrs;
} CPUIOTLBEntry;
+/**
+ * struct CPUTLBWindow
+ * @begin_ns: host time (in ns) at the beginning of the time window
+ * @max_entries: maximum number of entries observed in the window
+ *
+ * See also: tlb_mmu_resize_locked()
+ */
+typedef struct CPUTLBWindow {
+ int64_t begin_ns;
+ size_t max_entries;
+} CPUTLBWindow;
+
typedef struct CPUTLBDesc {
/*
* Describe a region covering all of the large pages allocated
@@ -152,6 +178,10 @@ typedef struct CPUTLBDesc {
target_ulong large_page_mask;
/* The next index to use in the tlb victim table. */
size_t vindex;
+#if TCG_TARGET_IMPLEMENTS_DYN_TLB
+ CPUTLBWindow window;
+ size_t n_used_entries;
+#endif
} CPUTLBDesc;
/*
@@ -176,6 +206,20 @@ typedef struct CPUTLBCommon {
size_t elide_flush_count;
} CPUTLBCommon;
+#if TCG_TARGET_IMPLEMENTS_DYN_TLB
+# define CPU_TLB \
+ /* tlb_mask[i] contains (n_entries - 1) << CPU_TLB_ENTRY_BITS */ \
+ uintptr_t tlb_mask[NB_MMU_MODES]; \
+ CPUTLBEntry *tlb_table[NB_MMU_MODES];
+# define CPU_IOTLB \
+ CPUIOTLBEntry *iotlb[NB_MMU_MODES];
+#else
+# define CPU_TLB \
+ CPUTLBEntry tlb_table[NB_MMU_MODES][CPU_TLB_SIZE];
+# define CPU_IOTLB \
+ CPUIOTLBEntry iotlb[NB_MMU_MODES][CPU_TLB_SIZE];
+#endif
+
/*
* The meaning of each of the MMU modes is defined in the target code.
* Note that NB_MMU_MODES is not yet defined; we can only reference it
@@ -184,9 +228,9 @@ typedef struct CPUTLBCommon {
#define CPU_COMMON_TLB \
CPUTLBCommon tlb_c; \
CPUTLBDesc tlb_d[NB_MMU_MODES]; \
- CPUTLBEntry tlb_table[NB_MMU_MODES][CPU_TLB_SIZE]; \
+ CPU_TLB \
CPUTLBEntry tlb_v_table[NB_MMU_MODES][CPU_VTLB_SIZE]; \
- CPUIOTLBEntry iotlb[NB_MMU_MODES][CPU_TLB_SIZE]; \
+ CPU_IOTLB \
CPUIOTLBEntry iotlb_v[NB_MMU_MODES][CPU_VTLB_SIZE];
#else
diff --git a/include/exec/cpu_ldst.h b/include/exec/cpu_ldst.h
index 959068495a..83b2907d86 100644
--- a/include/exec/cpu_ldst.h
+++ b/include/exec/cpu_ldst.h
@@ -135,6 +135,21 @@ static inline target_ulong tlb_addr_write(const CPUTLBEntry *entry)
#endif
}
+#if TCG_TARGET_IMPLEMENTS_DYN_TLB
+/* Find the TLB index corresponding to the mmu_idx + address pair. */
+static inline uintptr_t tlb_index(CPUArchState *env, uintptr_t mmu_idx,
+ target_ulong addr)
+{
+ uintptr_t size_mask = env->tlb_mask[mmu_idx] >> CPU_TLB_ENTRY_BITS;
+
+ return (addr >> TARGET_PAGE_BITS) & size_mask;
+}
+
+static inline size_t tlb_n_entries(CPUArchState *env, uintptr_t mmu_idx)
+{
+ return (env->tlb_mask[mmu_idx] >> CPU_TLB_ENTRY_BITS) + 1;
+}
+#else
/* Find the TLB index corresponding to the mmu_idx + address pair. */
static inline uintptr_t tlb_index(CPUArchState *env, uintptr_t mmu_idx,
target_ulong addr)
@@ -142,6 +157,12 @@ static inline uintptr_t tlb_index(CPUArchState *env, uintptr_t mmu_idx,
return (addr >> TARGET_PAGE_BITS) & (CPU_TLB_SIZE - 1);
}
+static inline size_t tlb_n_entries(CPUArchState *env, uintptr_t mmu_idx)
+{
+ return CPU_TLB_SIZE;
+}
+#endif /* TCG_TARGET_IMPLEMENTS_DYN_TLB */
+
/* Find the TLB entry corresponding to the mmu_idx + address pair. */
static inline CPUTLBEntry *tlb_entry(CPUArchState *env, uintptr_t mmu_idx,
target_ulong addr)
diff --git a/tcg/aarch64/tcg-target.h b/tcg/aarch64/tcg-target.h
index 9aea1d1771..3060d83d14 100644
--- a/tcg/aarch64/tcg-target.h
+++ b/tcg/aarch64/tcg-target.h
@@ -15,6 +15,7 @@
#define TCG_TARGET_INSN_UNIT_SIZE 4
#define TCG_TARGET_TLB_DISPLACEMENT_BITS 24
+#define TCG_TARGET_IMPLEMENTS_DYN_TLB 0
#undef TCG_TARGET_STACK_GROWSUP
typedef enum {
diff --git a/tcg/arm/tcg-target.h b/tcg/arm/tcg-target.h
index 94b3578c55..0e8b79d20f 100644
--- a/tcg/arm/tcg-target.h
+++ b/tcg/arm/tcg-target.h
@@ -60,6 +60,7 @@ extern int arm_arch;
#undef TCG_TARGET_STACK_GROWSUP
#define TCG_TARGET_INSN_UNIT_SIZE 4
#define TCG_TARGET_TLB_DISPLACEMENT_BITS 16
+#define TCG_TARGET_IMPLEMENTS_DYN_TLB 0
typedef enum {
TCG_REG_R0 = 0,
diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index 9fdf37f23c..9e4bfa90d1 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -27,6 +27,7 @@
#define TCG_TARGET_INSN_UNIT_SIZE 1
#define TCG_TARGET_TLB_DISPLACEMENT_BITS 31
+#define TCG_TARGET_IMPLEMENTS_DYN_TLB 0
#ifdef __x86_64__
# define TCG_TARGET_REG_BITS 64
diff --git a/tcg/mips/tcg-target.h b/tcg/mips/tcg-target.h
index a8222476f0..a97f31113e 100644
--- a/tcg/mips/tcg-target.h
+++ b/tcg/mips/tcg-target.h
@@ -37,6 +37,7 @@
#define TCG_TARGET_INSN_UNIT_SIZE 4
#define TCG_TARGET_TLB_DISPLACEMENT_BITS 16
+#define TCG_TARGET_IMPLEMENTS_DYN_TLB 0
#define TCG_TARGET_NB_REGS 32
typedef enum {
diff --git a/tcg/ppc/tcg-target.h b/tcg/ppc/tcg-target.h
index be52ad1d2e..8f03328af4 100644
--- a/tcg/ppc/tcg-target.h
+++ b/tcg/ppc/tcg-target.h
@@ -34,6 +34,7 @@
#define TCG_TARGET_NB_REGS 32
#define TCG_TARGET_INSN_UNIT_SIZE 4
#define TCG_TARGET_TLB_DISPLACEMENT_BITS 16
+#define TCG_TARGET_IMPLEMENTS_DYN_TLB 0
typedef enum {
TCG_REG_R0, TCG_REG_R1, TCG_REG_R2, TCG_REG_R3,
diff --git a/tcg/s390/tcg-target.h b/tcg/s390/tcg-target.h
index 6f2b06a7d1..df92f3065a 100644
--- a/tcg/s390/tcg-target.h
+++ b/tcg/s390/tcg-target.h
@@ -27,6 +27,7 @@
#define TCG_TARGET_INSN_UNIT_SIZE 2
#define TCG_TARGET_TLB_DISPLACEMENT_BITS 19
+#define TCG_TARGET_IMPLEMENTS_DYN_TLB 0
typedef enum TCGReg {
TCG_REG_R0 = 0,
diff --git a/tcg/sparc/tcg-target.h b/tcg/sparc/tcg-target.h
index d8339bf010..975ddc7b0d 100644
--- a/tcg/sparc/tcg-target.h
+++ b/tcg/sparc/tcg-target.h
@@ -29,6 +29,7 @@
#define TCG_TARGET_INSN_UNIT_SIZE 4
#define TCG_TARGET_TLB_DISPLACEMENT_BITS 32
+#define TCG_TARGET_IMPLEMENTS_DYN_TLB 0
#define TCG_TARGET_NB_REGS 32
typedef enum {
diff --git a/tcg/tci/tcg-target.h b/tcg/tci/tcg-target.h
index 26140d78cb..bcfd8d69e6 100644
--- a/tcg/tci/tcg-target.h
+++ b/tcg/tci/tcg-target.h
@@ -43,6 +43,7 @@
#define TCG_TARGET_INTERPRETER 1
#define TCG_TARGET_INSN_UNIT_SIZE 1
#define TCG_TARGET_TLB_DISPLACEMENT_BITS 32
+#define TCG_TARGET_IMPLEMENTS_DYN_TLB 0
#if UINTPTR_MAX == UINT32_MAX
# define TCG_TARGET_REG_BITS 32
diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c
index 5dc97212a9..b5af36ad1a 100644
--- a/accel/tcg/cputlb.c
+++ b/accel/tcg/cputlb.c
@@ -74,6 +74,169 @@ QEMU_BUILD_BUG_ON(sizeof(target_ulong) > sizeof(run_on_cpu_data));
QEMU_BUILD_BUG_ON(NB_MMU_MODES > 16);
#define ALL_MMUIDX_BITS ((1 << NB_MMU_MODES) - 1)
+#if TCG_TARGET_IMPLEMENTS_DYN_TLB
+static inline size_t sizeof_tlb(CPUArchState *env, uintptr_t mmu_idx)
+{
+ return env->tlb_mask[mmu_idx] + (1 << CPU_TLB_ENTRY_BITS);
+}
+
+static void tlb_window_reset(CPUTLBWindow *window, int64_t ns,
+ size_t max_entries)
+{
+ window->begin_ns = ns;
+ window->max_entries = max_entries;
+}
+
+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);
+ }
+}
+
+/**
+ * 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 == 100) {
+ new_size = MIN(old_size << 2, 1 << CPU_TLB_DYN_MAX_BITS);
+ } else 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_new(CPUTLBEntry, new_size);
+ env->iotlb[mmu_idx] = g_new(CPUIOTLBEntry, new_size);
+}
+
+static inline void tlb_table_flush_by_mmuidx(CPUArchState *env, int mmu_idx)
+{
+ tlb_mmu_resize_locked(env, mmu_idx);
+ memset(env->tlb_table[mmu_idx], -1, sizeof_tlb(env, mmu_idx));
+ env->tlb_d[mmu_idx].n_used_entries = 0;
+}
+
+static inline void tlb_n_used_entries_inc(CPUArchState *env, uintptr_t mmu_idx)
+{
+ env->tlb_d[mmu_idx].n_used_entries++;
+}
+
+static inline void tlb_n_used_entries_dec(CPUArchState *env, uintptr_t mmu_idx)
+{
+ env->tlb_d[mmu_idx].n_used_entries--;
+}
+
+#else /* !TCG_TARGET_IMPLEMENTS_DYN_TLB */
+
+static inline void tlb_dyn_init(CPUArchState *env)
+{
+}
+
+static inline void tlb_table_flush_by_mmuidx(CPUArchState *env, int mmu_idx)
+{
+ memset(env->tlb_table[mmu_idx], -1, sizeof(env->tlb_table[0]));
+}
+
+static inline void tlb_n_used_entries_inc(CPUArchState *env, uintptr_t mmu_idx)
+{
+}
+
+static inline void tlb_n_used_entries_dec(CPUArchState *env, uintptr_t mmu_idx)
+{
+}
+#endif /* TCG_TARGET_IMPLEMENTS_DYN_TLB */
+
void tlb_init(CPUState *cpu)
{
CPUArchState *env = cpu->env_ptr;
@@ -82,6 +245,8 @@ void tlb_init(CPUState *cpu)
/* Ensure that cpu_reset performs a full flush. */
env->tlb_c.dirty = ALL_MMUIDX_BITS;
+
+ tlb_dyn_init(env);
}
/* flush_all_helper: run fn across all cpus
@@ -122,7 +287,7 @@ void tlb_flush_counts(size_t *pfull, size_t *ppart, size_t *pelide)
static void tlb_flush_one_mmuidx_locked(CPUArchState *env, int mmu_idx)
{
- memset(env->tlb_table[mmu_idx], -1, sizeof(env->tlb_table[0]));
+ tlb_table_flush_by_mmuidx(env, mmu_idx);
memset(env->tlb_v_table[mmu_idx], -1, sizeof(env->tlb_v_table[0]));
env->tlb_d[mmu_idx].large_page_addr = -1;
env->tlb_d[mmu_idx].large_page_mask = -1;
@@ -225,12 +390,14 @@ static inline bool tlb_hit_page_anyprot(CPUTLBEntry *tlb_entry,
}
/* Called with tlb_c.lock held */
-static inline void tlb_flush_entry_locked(CPUTLBEntry *tlb_entry,
+static inline bool tlb_flush_entry_locked(CPUTLBEntry *tlb_entry,
target_ulong page)
{
if (tlb_hit_page_anyprot(tlb_entry, page)) {
memset(tlb_entry, -1, sizeof(*tlb_entry));
+ return true;
}
+ return false;
}
/* Called with tlb_c.lock held */
@@ -241,7 +408,9 @@ static inline void tlb_flush_vtlb_page_locked(CPUArchState *env, int mmu_idx,
assert_cpu_is_self(ENV_GET_CPU(env));
for (k = 0; k < CPU_VTLB_SIZE; k++) {
- tlb_flush_entry_locked(&env->tlb_v_table[mmu_idx][k], page);
+ if (tlb_flush_entry_locked(&env->tlb_v_table[mmu_idx][k], page)) {
+ tlb_n_used_entries_dec(env, mmu_idx);
+ }
}
}
@@ -258,7 +427,9 @@ static void tlb_flush_page_locked(CPUArchState *env, int midx,
midx, lp_addr, lp_mask);
tlb_flush_one_mmuidx_locked(env, midx);
} else {
- tlb_flush_entry_locked(tlb_entry(env, midx, page), page);
+ if (tlb_flush_entry_locked(tlb_entry(env, midx, page), page)) {
+ tlb_n_used_entries_dec(env, midx);
+ }
tlb_flush_vtlb_page_locked(env, midx, page);
}
}
@@ -435,8 +606,9 @@ void tlb_reset_dirty(CPUState *cpu, ram_addr_t start1, ram_addr_t length)
qemu_spin_lock(&env->tlb_c.lock);
for (mmu_idx = 0; mmu_idx < NB_MMU_MODES; mmu_idx++) {
unsigned int i;
+ unsigned int n = tlb_n_entries(env, mmu_idx);
- for (i = 0; i < CPU_TLB_SIZE; i++) {
+ for (i = 0; i < n; i++) {
tlb_reset_dirty_range_locked(&env->tlb_table[mmu_idx][i], start1,
length);
}
@@ -598,6 +770,7 @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr,
/* Evict the old entry into the victim tlb. */
copy_tlb_helper_locked(tv, te);
env->iotlb_v[mmu_idx][vidx] = env->iotlb[mmu_idx][index];
+ tlb_n_used_entries_dec(env, mmu_idx);
}
/* refill the tlb */
@@ -649,6 +822,7 @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr,
}
copy_tlb_helper_locked(te, &tn);
+ tlb_n_used_entries_inc(env, mmu_idx);
qemu_spin_unlock(&env->tlb_c.lock);
}
--
2.17.1
^ permalink raw reply related [flat|nested] 6+ messages in thread* Re: [Qemu-devel] [PATCH v5 2/3] tcg: introduce dynamic TLB sizing
2018-12-14 17:42 ` [Qemu-devel] [PATCH v5 2/3] tcg: introduce dynamic TLB sizing Emilio G. Cota
@ 2018-12-17 22:41 ` Emilio G. Cota
2018-12-26 6:51 ` Richard Henderson
1 sibling, 0 replies; 6+ messages in thread
From: Emilio G. Cota @ 2018-12-17 22:41 UTC (permalink / raw)
To: qemu-devel; +Cc: Richard Henderson, Alex Bennée
On Fri, Dec 14, 2018 at 12:42:32 -0500, Emilio G. Cota wrote:
> +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;
(snip)
> + 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 == 100) {
> + new_size = MIN(old_size << 2, 1 << CPU_TLB_DYN_MAX_BITS);
> + } else if (rate > 70) {
> + new_size = MIN(old_size << 1, 1 << CPU_TLB_DYN_MAX_BITS);
I won't resend a v6 until v5 gets any comments, but I just
want to point out that after doing some more benchmarking,
quadrupling the size here is unnecessary; in fact, perf is 1% better
on average if we just double the TLB size when rate > 70.
Emilio
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: [Qemu-devel] [PATCH v5 2/3] tcg: introduce dynamic TLB sizing
2018-12-14 17:42 ` [Qemu-devel] [PATCH v5 2/3] tcg: introduce dynamic TLB sizing Emilio G. Cota
2018-12-17 22:41 ` Emilio G. Cota
@ 2018-12-26 6:51 ` Richard Henderson
1 sibling, 0 replies; 6+ messages in thread
From: Richard Henderson @ 2018-12-26 6:51 UTC (permalink / raw)
To: Emilio G. Cota, qemu-devel; +Cc: Alex Bennée
On 12/15/18 4:42 AM, Emilio G. Cota wrote:
> +#if TCG_TARGET_IMPLEMENTS_DYN_TLB
> +#define CPU_TLB_DYN_MIN_BITS 6
> +#define CPU_TLB_DYN_DEFAULT_BITS 8
> +/*
> + * Assuming TARGET_PAGE_BITS==12, with 2**22 entries we can cover 2**(22+12) ==
> + * 2**34 == 16G of address space. This is roughly what one would expect a
> + * TLB to cover in a modern (as of 2018) x86_64 CPU. For instance, Intel
> + * Skylake's Level-2 STLB has 16 1G entries.
> + */
> +#define CPU_TLB_DYN_MAX_BITS 22
For 32-bit hosts, we need to limit this to (32 - TARGET_PAGE_BITS) so that we
do not require a double-word shift when implementing the tlb load. We probably
want to restrict it even further because a 32-bit host may well not have 2**27
bytes of available memory.
For 64-bit hosts, we should limit this to (TARGET_VIRT_ADDR_SPACE_BITS -
TARGET_PAGE_BITS) so that we do not grow the tlb past the guest's address space.
> + env->tlb_table[mmu_idx] = g_new(CPUTLBEntry, new_size);
> + env->iotlb[mmu_idx] = g_new(CPUIOTLBEntry, new_size);
For 32-bit hosts, we should probably be prepared for this to fail; see above re
large tlb sizes. If it fails, we should be able to go back to the previous tlb
size, since we just freed that memory.
r~
^ permalink raw reply [flat|nested] 6+ messages in thread
* [Qemu-devel] [PATCH v5 3/3] tcg/i386: enable dynamic TLB sizing
2018-12-14 17:42 [Qemu-devel] [PATCH v5 0/3] Dynamic TLB sizing Emilio G. Cota
2018-12-14 17:42 ` [Qemu-devel] [PATCH v5 1/3] cputlb: do not evict empty entries to the vtlb Emilio G. Cota
2018-12-14 17:42 ` [Qemu-devel] [PATCH v5 2/3] tcg: introduce dynamic TLB sizing Emilio G. Cota
@ 2018-12-14 17:42 ` Emilio G. Cota
2 siblings, 0 replies; 6+ messages in thread
From: Emilio G. Cota @ 2018-12-14 17:42 UTC (permalink / raw)
To: qemu-devel; +Cc: Richard Henderson, Alex Bennée
As the following experiments show, this series is a net perf gain,
particularly for memory-heavy workloads. Experiments are run on an
Intel(R) Xeon(R) Gold 6142 CPU @ 2.60GHz.
1. System boot + shudown, debian aarch64:
- Before (v3.1.0):
Performance counter stats for './die.sh v3.1.0' (10 runs):
9019.797015 task-clock (msec) # 0.993 CPUs utilized ( +- 0.23% )
29,910,312,379 cycles # 3.316 GHz ( +- 0.14% )
54,699,252,014 instructions # 1.83 insn per cycle ( +- 0.08% )
10,061,951,686 branches # 1115.541 M/sec ( +- 0.08% )
172,966,530 branch-misses # 1.72% of all branches ( +- 0.07% )
9.084039051 seconds time elapsed ( +- 0.23% )
- After:
Performance counter stats for './die.sh tlb-dyn-v5' (10 runs):
8624.084842 task-clock (msec) # 0.993 CPUs utilized ( +- 0.23% )
28,556,123,404 cycles # 3.311 GHz ( +- 0.13% )
51,755,089,512 instructions # 1.81 insn per cycle ( +- 0.05% )
9,526,513,946 branches # 1104.641 M/sec ( +- 0.05% )
166,578,509 branch-misses # 1.75% of all branches ( +- 0.19% )
8.680540350 seconds time elapsed ( +- 0.24% )
That is, a 4.4% perf increase.
2. System boot + shutdown, ubuntu 18.04 x86_64:
- Before (v3.1.0):
56100.574751 task-clock (msec) # 1.016 CPUs utilized ( +- 4.81% )
200,745,466,128 cycles # 3.578 GHz ( +- 5.24% )
431,949,100,608 instructions # 2.15 insn per cycle ( +- 5.65% )
77,502,383,330 branches # 1381.490 M/sec ( +- 6.18% )
844,681,191 branch-misses # 1.09% of all branches ( +- 3.82% )
55.221556378 seconds time elapsed ( +- 5.01% )
- After:
56603.419540 task-clock (msec) # 1.019 CPUs utilized ( +- 10.19% )
202,217,930,479 cycles # 3.573 GHz ( +- 10.69% )
439,336,291,626 instructions # 2.17 insn per cycle ( +- 14.14% )
80,538,357,447 branches # 1422.853 M/sec ( +- 16.09% )
776,321,622 branch-misses # 0.96% of all branches ( +- 3.77% )
55.549661409 seconds time elapsed ( +- 10.44% )
No improvement (within noise range). Note that for this workload,
increasing the time window too much can lead to perf degradation,
since it flushes the TLB *very* frequently.
3. x86_64 SPEC06int:
x86_64-softmmu speedup vs. v3.1.0 for SPEC06int (test set)
Host: Intel(R) Xeon(R) Gold 6142 CPU @ 2.60GHz (Skylake)
5.5 +------------------------------------------------------------------------+
| +-+ |
5 |-+.................+-+...............................tlb-dyn-v5.......+-|
| * * |
4.5 |-+.................*.*................................................+-|
| * * |
4 |-+.................*.*................................................+-|
| * * |
3.5 |-+.................*.*................................................+-|
| * * |
3 |-+......+-+*.......*.*................................................+-|
| * * * * |
2.5 |-+......*..*.......*.*.................................+-+*...........+-|
| * * * * * * |
2 |-+......*..*.......*.*.................................*..*...........+-|
| * * * * * * +-+ |
1.5 |-+......*..*.......*.*.................................*..*.*+-+.*+-+.+-|
| * * *+-+ * * +-+ *+-+ +-+ +-+ * * * * * * |
1 |++++-+*+*++*+*++*++*+*++*+*+++-+*+*+-++*+-++++-++++-+++*++*+*++*+*++*+++|
| * * * * * * * * * * * * * * * * * * * * * * * * * * |
0.5 +------------------------------------------------------------------------+
400.perlb401.bzip403.g429445.g456.hm462.libq464.h471.omn47483.xalancbgeomean
png: https://imgur.com/YRF90f7
That is, a 1.51x average speedup over the baseline, with a max speedup
of 5.17x.
Here's a different look at the SPEC06int results, using KVM as the baseline:
x86_64-softmmu slowdown vs. KVM for SPEC06int (test set)
Host: Intel(R) Xeon(R) Gold 6142 CPU @ 2.60GHz (Skylake)
25 +---------------------------------------------------------------------------+
| +-+ +-+ |
| * * +-+ v3.1.0 |
| * * +-+ tlb-dyn-v5 |
| * * * * +-+ |
20 |-+.................*.*.............................*.+-+......*.*........+-|
| * * * # # * * |
| +-+ * * * # # * * |
| * * * * * # # * * |
15 |-+......*.*........*.*.............................*.#.#......*.+-+......+-|
| * * * * * # # * #|# |
| * * * * +-+ * # # * +-+ |
| * * +-+ * * ++-+ +-+ * # # * # # +-+ |
| * * +-+ * * * ## *| +-+ * # # * # # +-+ |
10 |-+......*.*..*.+-+.*.*........*.##.......++-+.*.+-+*.#.#......*.#.#.*.*..+-|
| * * * +-+ * * * ## +-+ *# # * # #* # # +-+ * # # * * |
| * * * # # * * +-+ * ## * +-+ *# # * # #* # # * * * # # *+-+ |
| * * * # # * * * +-+ * ## * # # *# # * # #* # # * * * # # * ## |
5 |-+......*.+-+*.#.#.*.*..*.#.#.*.##.*.#.#.*#.#.*.#.#*.#.#.*.*..*.#.#.*.##.+-|
| * # #* # # * +-+* # # * ## * # # *# # * # #* # # * * * # # * ## |
| * # #* # # * # #* # # * ## * # # *# # * # #* # # * +-+* # # * ## |
| ++-+ * # #* # # * # #* # # * ## * # # *# # * # #* # # * # #* # # * ## |
|+++*#+#+*+#+#*+#+#+*+#+#*+#+#+*+##+*+#+#+*#+#+*+#+#*+#+#+*+#+#*+#+#+*+##+++|
0 +---------------------------------------------------------------------------+
400.perlbe401.bzi403.gc429445.go456.h462.libqu464.h471.omne4483.xalancbmgeomean
png: https://imgur.com/YzAMNEV
After this series, we bring down the average SPEC06int slowdown vs KVM
from 11.47x to 7.58x.
Signed-off-by: Emilio G. Cota <cota@braap.org>
---
tcg/i386/tcg-target.h | 2 +-
tcg/i386/tcg-target.inc.c | 28 ++++++++++++++--------------
2 files changed, 15 insertions(+), 15 deletions(-)
diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h
index 9e4bfa90d1..8b6475d786 100644
--- a/tcg/i386/tcg-target.h
+++ b/tcg/i386/tcg-target.h
@@ -27,7 +27,7 @@
#define TCG_TARGET_INSN_UNIT_SIZE 1
#define TCG_TARGET_TLB_DISPLACEMENT_BITS 31
-#define TCG_TARGET_IMPLEMENTS_DYN_TLB 0
+#define TCG_TARGET_IMPLEMENTS_DYN_TLB 1
#ifdef __x86_64__
# define TCG_TARGET_REG_BITS 64
diff --git a/tcg/i386/tcg-target.inc.c b/tcg/i386/tcg-target.inc.c
index 436195894b..5cbb07deab 100644
--- a/tcg/i386/tcg-target.inc.c
+++ b/tcg/i386/tcg-target.inc.c
@@ -330,6 +330,7 @@ static inline int tcg_target_const_match(tcg_target_long val, TCGType type,
#define OPC_ARITH_GvEv (0x03) /* ... plus (ARITH_FOO << 3) */
#define OPC_ANDN (0xf2 | P_EXT38)
#define OPC_ADD_GvEv (OPC_ARITH_GvEv | (ARITH_ADD << 3))
+#define OPC_AND_GvEv (OPC_ARITH_GvEv | (ARITH_AND << 3))
#define OPC_BLENDPS (0x0c | P_EXT3A | P_DATA16)
#define OPC_BSF (0xbc | P_EXT)
#define OPC_BSR (0xbd | P_EXT)
@@ -1625,7 +1626,7 @@ static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
}
if (TCG_TYPE_PTR == TCG_TYPE_I64) {
hrexw = P_REXW;
- if (TARGET_PAGE_BITS + CPU_TLB_BITS > 32) {
+ if (TARGET_PAGE_BITS + CPU_TLB_DYN_MAX_BITS > 32) {
tlbtype = TCG_TYPE_I64;
tlbrexw = P_REXW;
}
@@ -1633,6 +1634,15 @@ static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
}
tcg_out_mov(s, tlbtype, r0, addrlo);
+ tcg_out_shifti(s, SHIFT_SHR + tlbrexw, r0,
+ TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
+
+ tcg_out_modrm_offset(s, OPC_AND_GvEv + trexw, r0, TCG_AREG0,
+ offsetof(CPUArchState, tlb_mask[mem_index]));
+
+ tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r0, TCG_AREG0,
+ offsetof(CPUArchState, tlb_table[mem_index]));
+
/* If the required alignment is at least as large as the access, simply
copy the address and mask. For lesser alignments, check that we don't
cross pages for the complete access. */
@@ -1642,20 +1652,10 @@ static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
tcg_out_modrm_offset(s, OPC_LEA + trexw, r1, addrlo, s_mask - a_mask);
}
tlb_mask = (target_ulong)TARGET_PAGE_MASK | a_mask;
-
- tcg_out_shifti(s, SHIFT_SHR + tlbrexw, r0,
- TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS);
-
tgen_arithi(s, ARITH_AND + trexw, r1, tlb_mask, 0);
- tgen_arithi(s, ARITH_AND + tlbrexw, r0,
- (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS, 0);
-
- tcg_out_modrm_sib_offset(s, OPC_LEA + hrexw, r0, TCG_AREG0, r0, 0,
- offsetof(CPUArchState, tlb_table[mem_index][0])
- + which);
/* cmp 0(r0), r1 */
- tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw, r1, r0, 0);
+ tcg_out_modrm_offset(s, OPC_CMP_GvEv + trexw, r1, r0, which);
/* Prepare for both the fast path add of the tlb addend, and the slow
path function argument setup. There are two cases worth note:
@@ -1672,7 +1672,7 @@ static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
if (TARGET_LONG_BITS > TCG_TARGET_REG_BITS) {
/* cmp 4(r0), addrhi */
- tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi, r0, 4);
+ tcg_out_modrm_offset(s, OPC_CMP_GvEv, addrhi, r0, which + 4);
/* jne slow_path */
tcg_out_opc(s, OPC_JCC_long + JCC_JNE, 0, 0, 0);
@@ -1684,7 +1684,7 @@ static inline void tcg_out_tlb_load(TCGContext *s, TCGReg addrlo, TCGReg addrhi,
/* add addend(r0), r1 */
tcg_out_modrm_offset(s, OPC_ADD_GvEv + hrexw, r1, r0,
- offsetof(CPUTLBEntry, addend) - which);
+ offsetof(CPUTLBEntry, addend));
}
/*
--
2.17.1
^ permalink raw reply related [flat|nested] 6+ messages in thread