gh-149584: Fix excessive overhead in the Tachyon profiler regarding the cache behavior (#149649)

Use exact remote reads for interpreter state, thread state, and
interpreter frame structs instead of pulling full remote pages into the
profiler page cache. This matches the core change from
#149585.

The profiler clears the page cache between samples, so live entries are
always packed at the front. Track the live count and only clear/search
that prefix instead of scanning all 1024 slots on the hot path.

Use the frame cache to predict the next thread state and top frame
address, then batch interpreter/thread/frame reads with process_vm_readv
when profiling a Linux target. Reuse prefetched frame buffers in the
frame walker when the prediction is valid.

Cache the last FrameInfo tuple per code object/instruction offset, reuse
cached thread id objects, and append cached parent frames directly on
full frame-cache hits. This cuts Python allocation churn in the
steady-state profiler path.

Author: pablogsal

Lib/profiling/sampling/sample.py

@@ -327,6 +327,33 @@ def _print_unwinder_stats(self):
         print(f"    Hits:             {code_hits:n} ({ANSIColors.GREEN}{fmt(code_hits_pct)}%{ANSIColors.RESET})")
         print(f"    Misses:           {code_misses:n} ({ANSIColors.RED}{fmt(code_misses_pct)}%{ANSIColors.RESET})")
 
+        batched_attempts = stats.get('batched_read_attempts', 0)
+        batched_successes = stats.get('batched_read_successes', 0)
+        batched_misses = stats.get('batched_read_misses', 0)
+        segments_requested = stats.get('batched_read_segments_requested', 0)
+        segments_completed = stats.get('batched_read_segments_completed', 0)
+        if batched_attempts > 0:
+            batched_success_rate = stats.get('batched_read_success_rate', 0.0)
+            batched_miss_rate = 100.0 - batched_success_rate
+            segment_completion_rate = stats.get(
+                'batched_read_segment_completion_rate', 0.0
+            )
+
+            print(f"  {ANSIColors.CYAN}Batched Reads:{ANSIColors.RESET}")
+            print(f"    Attempts:         {batched_attempts:n}")
+            print(
+                f"    Successes:        {batched_successes:n} "
+                f"({ANSIColors.GREEN}{fmt(batched_success_rate)}%{ANSIColors.RESET})"
+            )
+            print(
+                f"    Misses:           {batched_misses:n} "
+                f"({ANSIColors.RED}{fmt(batched_miss_rate)}%{ANSIColors.RESET})"
+            )
+            print(
+                f"    Segments read:    {segments_completed:n}/{segments_requested:n} "
+                f"({ANSIColors.GREEN}{fmt(segment_completion_rate)}%{ANSIColors.RESET})"
+            )
+
         # Memory operations
         memory_reads = stats.get('memory_reads', 0)
         memory_bytes = stats.get('memory_bytes_read', 0)

Lib/test/test_external_inspection.py

@@ -3767,6 +3767,13 @@ def test_get_stats(self):
             "frames_read_from_cache",
             "frames_read_from_memory",
             "frame_cache_hit_rate",
+            "batched_read_attempts",
+            "batched_read_successes",
+            "batched_read_misses",
+            "batched_read_segments_requested",
+            "batched_read_segments_completed",
+            "batched_read_success_rate",
+            "batched_read_segment_completion_rate",
         ]
         for key in expected_keys:
             self.assertIn(key, stats)

Misc/NEWS.d/next/Library/2026-05-10-19-26-50.gh-issue-149584.x7Qm9A.rst

@@ -0,0 +1,4 @@
+Fix excessive overhead in the Tachyon profiler when inspecting a remote
+process by avoiding repeated remote page-cache scans, batching predicted
+remote reads, and reusing cached profiler result objects. Patch by Pablo
+Galindo and Maurycy Pawłowski-Wieroński.

Modules/_remote_debugging/_remote_debugging.h

@@ -30,6 +30,7 @@ extern "C" {
 #include "internal/pycore_llist.h"          // struct llist_node
 #include "internal/pycore_long.h"           // _PyLong_GetZero
 #include "internal/pycore_pyerrors.h"       // _PyErr_FormatFromCause
+#include "internal/pycore_pyhash.h"        // _Py_HashPointerRaw
 #include "internal/pycore_stackref.h"       // Py_TAG_BITS
 #include "../../Python/remote_debug.h"
 
@@ -215,6 +216,8 @@ typedef struct {
     PyObject *file_name;
     int first_lineno;
     PyObject *linetable;  // bytes
+    PyObject *last_frame_info;
+    ptrdiff_t last_addrq;
     uintptr_t addr_code_adaptive;
 } CachedCodeMetadata;
 
@@ -224,11 +227,41 @@ typedef struct {
 
 typedef struct {
     uint64_t thread_id;                      // 0 = empty slot
+    uintptr_t thread_state_addr;
     uintptr_t addrs[FRAME_CACHE_MAX_FRAMES];
     Py_ssize_t num_addrs;
+    PyObject *thread_id_obj;                 // owned reference, NULL if empty
     PyObject *frame_list;                    // owned reference, NULL if empty
 } FrameCacheEntry;
 
+#define INTERPRETER_THREAD_CACHE_SIZE 32
+#if (INTERPRETER_THREAD_CACHE_SIZE & (INTERPRETER_THREAD_CACHE_SIZE - 1)) != 0
+#  error "INTERPRETER_THREAD_CACHE_SIZE must be a power of two"
+#endif
+
+// The two per-interpreter L2 caches below are split into per-field tables so
+// that a writer rebinding one slot cannot leave stale data in a field owned by
+// the other when the slot is reused across interpreters.
+typedef struct {
+    uintptr_t interpreter_addr;
+    uintptr_t thread_state_addr;
+} InterpreterTstateCacheEntry;
+typedef struct {
+    uintptr_t interpreter_addr;
+    uint64_t code_object_generation;
+} InterpreterGenerationCacheEntry;
+
+// Carries already-read thread state and/or frame buffers across helpers so the
+// downstream callee can skip a remote read. Address fields are caller-supplied
+// inputs; buffer pointers (tstate, frame) are NULL unless a prior batched read
+// successfully populated them.
+typedef struct {
+    const char *tstate;
+    uintptr_t tstate_addr;
+    const char *frame;
+    uintptr_t frame_addr;
+} RemoteReadPrefetch;
+
 /* Statistics for profiling performance analysis */
 typedef struct {
     uint64_t total_samples;                  // Total number of get_stack_trace calls
@@ -242,14 +275,44 @@ typedef struct {
     uint64_t code_object_cache_hits;         // Code object cache hits
     uint64_t code_object_cache_misses;       // Code object cache misses
     uint64_t stale_cache_invalidations;      // Times stale entries were cleared
+    uint64_t batched_read_attempts;          // Batched remote-read attempts
+    uint64_t batched_read_successes;         // Attempts that read all requested segments
+    uint64_t batched_read_misses;            // Attempts that fell back or partially read
+    uint64_t batched_read_segments_requested; // Segments requested by batched reads
+    uint64_t batched_read_segments_completed; // Segments completed by batched reads
 } UnwinderStats;
 
+#if defined(__GNUC__) || defined(__clang__)
+#  define REMOTE_DEBUG_UNLIKELY(value) __builtin_expect(!!(value), 0)
+#else
+#  define REMOTE_DEBUG_UNLIKELY(value) (value)
+#endif
+
 /* Stats tracking macros - no-op when stats collection is disabled */
 #define STATS_INC(unwinder, field) \
-    do { if ((unwinder)->collect_stats) (unwinder)->stats.field++; } while(0)
+    do { if (REMOTE_DEBUG_UNLIKELY((unwinder)->collect_stats)) (unwinder)->stats.field++; } while(0)
 
 #define STATS_ADD(unwinder, field, val) \
-    do { if ((unwinder)->collect_stats) (unwinder)->stats.field += (val); } while(0)
+    do { if (REMOTE_DEBUG_UNLIKELY((unwinder)->collect_stats)) (unwinder)->stats.field += (val); } while(0)
+
+#if HAVE_PROCESS_VM_READV
+#  define STATS_BATCHED_READ(unwinder, requested, completed) \
+    do { \
+        if (REMOTE_DEBUG_UNLIKELY((unwinder)->collect_stats)) { \
+            (unwinder)->stats.batched_read_attempts++; \
+            (unwinder)->stats.batched_read_segments_requested += (uint64_t)(requested); \
+            (unwinder)->stats.batched_read_segments_completed += (uint64_t)(completed); \
+            if ((completed) == (requested)) { \
+                (unwinder)->stats.batched_read_successes++; \
+            } \
+            else { \
+                (unwinder)->stats.batched_read_misses++; \
+            } \
+        } \
+    } while(0)
+#else
+#  define STATS_BATCHED_READ(unwinder, requested, completed) ((void)0)
+#endif
 
 typedef struct {
     PyTypeObject *RemoteDebugging_Type;
@@ -290,7 +353,6 @@ typedef struct {
     struct _Py_AsyncioModuleDebugOffsets async_debug_offsets;
     uintptr_t interpreter_addr;
     uintptr_t tstate_addr;
-    uint64_t code_object_generation;
     _Py_hashtable_t *code_object_cache;
     int debug;
     int only_active_thread;
@@ -302,9 +364,17 @@ typedef struct {
     int cache_frames;
     int collect_stats;  // whether to collect statistics
     uint32_t stale_invalidation_counter;  // counter for throttling frame_cache_invalidate_stale
+    // L1 single-entry shortcut over cached_tstates[]: most workloads sample one
+    // interpreter, so check these pairs before hashing into the table below.
+    uintptr_t cached_tstate_interpreter_addr;
+    uintptr_t cached_tstate_addr;
+    uintptr_t cached_generation_interpreter_addr;
+    uint64_t cached_code_object_generation;
     RemoteDebuggingState *cached_state;
     FrameCacheEntry *frame_cache;  // preallocated array of FRAME_CACHE_MAX_THREADS entries
     UnwinderStats stats;  // statistics for performance analysis
+    InterpreterTstateCacheEntry cached_tstates[INTERPRETER_THREAD_CACHE_SIZE];
+    InterpreterGenerationCacheEntry cached_generations[INTERPRETER_THREAD_CACHE_SIZE];
 #ifdef Py_GIL_DISABLED
     uint32_t tlbc_generation;
     _Py_hashtable_t *tlbc_cache;
@@ -361,11 +431,13 @@ typedef struct {
 typedef struct {
     /* Inputs */
     uintptr_t frame_addr;           // Starting frame address
+    uintptr_t thread_state_addr;    // Owning thread state address
     uintptr_t base_frame_addr;      // Sentinel at bottom (for validation)
     uintptr_t gc_frame;             // GC frame address (0 if not tracking)
     uintptr_t last_profiled_frame;  // Last cached frame (0 if no cache)
     StackChunkList *chunks;         // Pre-copied stack chunks
     int skip_first_frame;           // Skip frame_addr itself (continue from its caller)
+    RemoteReadPrefetch prefetch;     // Optional already-read thread/frame buffers
 
     /* Outputs */
     PyObject *frame_info;           // List to append FrameInfo objects
@@ -548,6 +620,7 @@ extern int process_frame_chain(
 extern int frame_cache_init(RemoteUnwinderObject *unwinder);
 extern void frame_cache_cleanup(RemoteUnwinderObject *unwinder);
 extern FrameCacheEntry *frame_cache_find(RemoteUnwinderObject *unwinder, uint64_t thread_id);
+extern FrameCacheEntry *frame_cache_find_by_tstate(RemoteUnwinderObject *unwinder, uintptr_t tstate_addr);
 extern int clear_last_profiled_frames(RemoteUnwinderObject *unwinder);
 extern void frame_cache_invalidate_stale(RemoteUnwinderObject *unwinder, PyObject *result);
 extern int frame_cache_lookup_and_extend(
@@ -566,6 +639,7 @@ extern int frame_cache_store(
     PyObject *frame_list,
     const uintptr_t *addrs,
     Py_ssize_t num_addrs,
+    uintptr_t thread_state_addr,
     uintptr_t base_frame_addr,
     uintptr_t last_frame_visited);
 
@@ -605,7 +679,8 @@ extern PyObject* unwind_stack_for_thread(
     uintptr_t *current_tstate,
     uintptr_t gil_holder_tstate,
     uintptr_t gc_frame,
-    uintptr_t main_thread_tstate
+    uintptr_t main_thread_tstate,
+    const RemoteReadPrefetch *prefetch
 );
 
 /* Thread stopping functions (for blocking mode) */

Modules/_remote_debugging/clinic/module.c.h

@@ -405,6 +405,8 @@ parse_code_object(RemoteUnwinderObject *unwinder,
         meta->func_name = func;
         meta->file_name = file;
         meta->linetable = linetable;
+        meta->last_frame_info = NULL;
+        meta->last_addrq = -1;
         meta->first_lineno = GET_MEMBER(int, code_object, unwinder->debug_offsets.code_object.firstlineno);
         meta->addr_code_adaptive = real_address + (uintptr_t)unwinder->debug_offsets.code_object.co_code_adaptive;
 
@@ -482,6 +484,12 @@ parse_code_object(RemoteUnwinderObject *unwinder,
     addrq = (uint16_t *)ip - (uint16_t *)meta->addr_code_adaptive;
 #endif
     ;  // Empty statement to avoid C23 extension warning
+
+    if (!unwinder->opcodes && meta->last_frame_info != NULL && meta->last_addrq == addrq) {
+        *result = Py_NewRef(meta->last_frame_info);
+        return 0;
+    }
+
     LocationInfo info = {0};
     bool ok = parse_linetable(addrq, PyBytes_AS_STRING(meta->linetable),
                               PyBytes_GET_SIZE(meta->linetable),
@@ -529,6 +537,11 @@ parse_code_object(RemoteUnwinderObject *unwinder,
         goto error;
     }
 
+    if (!unwinder->opcodes) {
+        Py_XSETREF(meta->last_frame_info, Py_NewRef(tuple));
+        meta->last_addrq = addrq;
+    }
+
     *result = tuple;
     return 0;
 

Modules/_remote_debugging/code_objects.c

@@ -30,6 +30,7 @@ frame_cache_cleanup(RemoteUnwinderObject *unwinder)
         return;
     }
     for (int i = 0; i < FRAME_CACHE_MAX_THREADS; i++) {
+        Py_CLEAR(unwinder->frame_cache[i].thread_id_obj);
         Py_CLEAR(unwinder->frame_cache[i].frame_list);
     }
     PyMem_Free(unwinder->frame_cache);
@@ -53,6 +54,21 @@ frame_cache_find(RemoteUnwinderObject *unwinder, uint64_t thread_id)
     return NULL;
 }
 
+FrameCacheEntry *
+frame_cache_find_by_tstate(RemoteUnwinderObject *unwinder, uintptr_t tstate_addr)
+{
+    if (!unwinder->frame_cache || tstate_addr == 0) {
+        return NULL;
+    }
+    for (int i = 0; i < FRAME_CACHE_MAX_THREADS; i++) {
+        if (unwinder->frame_cache[i].thread_state_addr == tstate_addr) {
+            assert(unwinder->frame_cache[i].num_addrs <= FRAME_CACHE_MAX_FRAMES);
+            return &unwinder->frame_cache[i];
+        }
+    }
+    return NULL;
+}
+
 // Allocate a cache slot for a thread
 // Returns NULL if cache is full (graceful degradation)
 static FrameCacheEntry *
@@ -127,8 +143,10 @@ frame_cache_invalidate_stale(RemoteUnwinderObject *unwinder, PyObject *result)
         }
         if (!found) {
             // Clear this entry
+            Py_CLEAR(unwinder->frame_cache[i].thread_id_obj);
             Py_CLEAR(unwinder->frame_cache[i].frame_list);
             unwinder->frame_cache[i].thread_id = 0;
+            unwinder->frame_cache[i].thread_state_addr = 0;
             unwinder->frame_cache[i].num_addrs = 0;
             STATS_INC(unwinder, stale_cache_invalidations);
         }
@@ -216,6 +234,7 @@ frame_cache_store(
     PyObject *frame_list,
     const uintptr_t *addrs,
     Py_ssize_t num_addrs,
+    uintptr_t thread_state_addr,
     uintptr_t base_frame_addr,
     uintptr_t last_frame_visited)
 {
@@ -257,6 +276,13 @@ frame_cache_store(
         return -1;
     }
     entry->thread_id = thread_id;
+    entry->thread_state_addr = thread_state_addr;
+    if (entry->thread_id_obj == NULL) {
+        entry->thread_id_obj = PyLong_FromUnsignedLongLong(thread_id);
+        if (entry->thread_id_obj == NULL) {
+            return -1;
+        }
+    }
     memcpy(entry->addrs, addrs, num_addrs * sizeof(uintptr_t));
     entry->num_addrs = num_addrs;
     assert(entry->num_addrs == num_addrs);