dalvik/vm/AllocTracker.c

/*
 * Copyright (C) 2008 The Android Open Source Project
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *      http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

/*
 * Allocation tracking and reporting.  We maintain a circular buffer with
 * the most recent allocations.  The data can be viewed through DDMS.
 *
 * There are two basic approaches: manage the buffer with atomic updates
 * and do a system-wide suspend when DDMS requests it, or protect all
 * accesses with a mutex.  The former is potentially more efficient, but
 * the latter is much simpler and more reliable.
 *
 * Ideally we'd just use the object heap allocation mutex to guard this
 * structure, but at the point we grab that (under dvmMalloc()) we're just
 * allocating a collection of bytes and no longer have the class reference.
 * Because this is an optional feature it's best to leave the existing
 * code undisturbed and just use an additional lock.
 *
 * We don't currently track allocations of class objects.  We could, but
 * with the possible exception of Proxy objects they're not that interesting.
 *
 * TODO: if we add support for class unloading, we need to add the class
 * references here to the root set (or just disable class unloading while
 * this is active).
 *
 * TODO: consider making the parameters configurable, so DDMS can decide
 * how many allocations it wants to see and what the stack depth should be.
 * Changing the window size is easy, changing the max stack depth is harder
 * because we go from an array of fixed-size structs to variable-sized data.
 */
#include "Dalvik.h"

#define kMaxAllocRecordStackDepth   16      /* max 255 */
#define kNumAllocRecords            512     /* MUST be power of 2 */

/*
 * Record the details of an allocation.
 */
struct AllocRecord {
    ClassObject*    clazz;      /* class allocated in this block */
    u4              size;       /* total size requested */
    u2              threadId;   /* simple thread ID; could be recycled */

    /* stack trace elements; unused entries have method==NULL */
    struct {
        const Method* method;   /* which method we're executing in */
        int         pc;         /* current execution offset, in 16-bit units */
    } stackElem[kMaxAllocRecordStackDepth];

    /*
     * This was going to be either wall-clock time in seconds or monotonic
     * time in milliseconds since the VM started, to give a rough sense for
     * how long ago an allocation happened.  This adds a system call per
     * allocation, which is too much overhead.
     */
    //u4      timestamp;
};

/*
 * Initialize a few things.  This gets called early, so keep activity to
 * a minimum.
 */
bool dvmAllocTrackerStartup(void)
{
    /* prep locks */
    dvmInitMutex(&gDvm.allocTrackerLock);

    /* initialized when enabled by DDMS */
    assert(gDvm.allocRecords == NULL);

    return true;
}

/*
 * Release anything we're holding on to.
 */
void dvmAllocTrackerShutdown(void)
{
    free(gDvm.allocRecords);
    dvmDestroyMutex(&gDvm.allocTrackerLock);
}


/*
 * ===========================================================================
 *      Collection
 * ===========================================================================
 */

/*
 * Enable allocation tracking.  Does nothing if tracking is already enabled.
 *
 * Returns "true" on success.
 */
bool dvmEnableAllocTracker(void)
{
    bool result = true;
    dvmLockMutex(&gDvm.allocTrackerLock);

    if (gDvm.allocRecords == NULL) {
        LOGI("Enabling alloc tracker (%d entries, %d frames --> %d bytes)\n",
            kNumAllocRecords, kMaxAllocRecordStackDepth,
            sizeof(AllocRecord) * kNumAllocRecords);
        gDvm.allocRecordHead = gDvm.allocRecordCount = 0;
        gDvm.allocRecords =
            (AllocRecord*) malloc(sizeof(AllocRecord) * kNumAllocRecords);

        if (gDvm.allocRecords == NULL)
            result = false;
    }

    dvmUnlockMutex(&gDvm.allocTrackerLock);
    return result;
}

/*
 * Disable allocation tracking.  Does nothing if tracking is not enabled.
 */
void dvmDisableAllocTracker(void)
{
    dvmLockMutex(&gDvm.allocTrackerLock);

    if (gDvm.allocRecords != NULL) {
        free(gDvm.allocRecords);
        gDvm.allocRecords = NULL;
    }

    dvmUnlockMutex(&gDvm.allocTrackerLock);
}

/*
 * Get the last few stack frames.
 */
static void getStackFrames(Thread* self, AllocRecord* pRec)
{
    int stackDepth = 0;
    void* fp;

    fp = self->curFrame;

    while ((fp != NULL) && (stackDepth < kMaxAllocRecordStackDepth)) {
        const StackSaveArea* saveArea = SAVEAREA_FROM_FP(fp);
        const Method* method = saveArea->method;

        if (!dvmIsBreakFrame(fp)) {
            pRec->stackElem[stackDepth].method = method;
            if (dvmIsNativeMethod(method)) {
                pRec->stackElem[stackDepth].pc = 0;
            } else {
                assert(saveArea->xtra.currentPc >= method->insns &&
                        saveArea->xtra.currentPc <
                        method->insns + dvmGetMethodInsnsSize(method));
                pRec->stackElem[stackDepth].pc =
                    (int) (saveArea->xtra.currentPc - method->insns);
            }
            stackDepth++;
        }

        assert(fp != saveArea->prevFrame);
        fp = saveArea->prevFrame;
    }

    /* clear out the rest (normally there won't be any) */
    while (stackDepth < kMaxAllocRecordStackDepth) {
        pRec->stackElem[stackDepth].method = NULL;
        pRec->stackElem[stackDepth].pc = 0;
        stackDepth++;
    }
}

/*
 * Add a new allocation to the set.
 */
void dvmDoTrackAllocation(ClassObject* clazz, int size)
{
    dvmLockMutex(&gDvm.allocTrackerLock);
    if (gDvm.allocRecords == NULL)
        goto bail;

    Thread* self = dvmThreadSelf();
    if (self == NULL) {
        LOGW("alloc tracker: no thread\n");
        goto bail;
    }

    /* advance and clip */
    if (++gDvm.allocRecordHead == kNumAllocRecords)
        gDvm.allocRecordHead = 0;

    AllocRecord* pRec = &gDvm.allocRecords[gDvm.allocRecordHead];

    pRec->clazz = clazz;
    pRec->size = size;
    pRec->threadId = self->threadId;
    getStackFrames(self, pRec);

    if (gDvm.allocRecordCount < kNumAllocRecords)
        gDvm.allocRecordCount++;

bail:
    dvmUnlockMutex(&gDvm.allocTrackerLock);
}


/*
 * ===========================================================================
 *      Reporting
 * ===========================================================================
 */

/*
The data we send to DDMS contains everything we have recorded.

Message header (all values big-endian):
  (1b) message header len (to allow future expansion); includes itself
  (1b) entry header len
  (1b) stack frame len
  (2b) number of entries
  (4b) offset to string table from start of message
  (2b) number of class name strings
  (2b) number of method name strings
  (2b) number of source file name strings
  For each entry:
    (4b) total allocation size
    (2b) threadId
    (2b) allocated object's class name index
    (1b) stack depth
    For each stack frame:
      (2b) method's class name
      (2b) method name
      (2b) method source file
      (2b) line number, clipped to 32767; -2 if native; -1 if no source
  (xb) class name strings
  (xb) method name strings
  (xb) source file strings

  As with other DDM traffic, strings are sent as a 4-byte length
  followed by UTF-16 data.

We send up 16-bit unsigned indexes into string tables.  In theory there
can be (kMaxAllocRecordStackDepth * kNumAllocRecords) unique strings in
each table, but in practice there should be far fewer.

The chief reason for using a string table here is to keep the size of
the DDMS message to a minimum.  This is partly to make the protocol
efficient, but also because we have to form the whole thing up all at
once in a memory buffer.

We use separate string tables for class names, method names, and source
files to keep the indexes small.  There will generally be no overlap
between the contents of these tables.
*/
const int kMessageHeaderLen = 15;
const int kEntryHeaderLen = 9;
const int kStackFrameLen = 8;

/*
 * Return the index of the head element.
 *
 * We point at the most-recently-written record, so if allocRecordCount is 1
 * we want to use the current element.  Take "head+1" and subtract count
 * from it.
 *
 * We need to handle underflow in our circular buffer, so we add
 * kNumAllocRecords and then mask it back down.
 */
inline static int headIndex(void)
{
    return (gDvm.allocRecordHead+1 + kNumAllocRecords - gDvm.allocRecordCount)
        & (kNumAllocRecords-1);
}

/*
 * Dump the contents of a PointerSet full of character pointers.
 */
static void dumpStringTable(PointerSet* strings)
{
    int count = dvmPointerSetGetCount(strings);
    int i;

    for (i = 0; i < count; i++)
        printf("  %s\n", (const char*) dvmPointerSetGetEntry(strings, i));
}

/*
 * Get the method's source file.  If we don't know it, return "" instead
 * of a NULL pointer.
 */
static const char* getMethodSourceFile(const Method* method)
{
    const char* fileName = dvmGetMethodSourceFile(method);
    if (fileName == NULL)
        fileName = "";
    return fileName;
}

/*
 * Generate string tables.
 *
 * Our source material is UTF-8 string constants from DEX files.  If we
 * want to be thorough we can generate a hash value for each string and
 * use the VM hash table implementation, or we can do a quick & dirty job
 * by just maintaining a list of unique pointers.  If the same string
 * constant appears in multiple DEX files we'll end up with duplicates,
 * but in practice this shouldn't matter (and if it does, we can uniq-sort
 * the result in a second pass).
 */
static bool populateStringTables(PointerSet* classNames,
    PointerSet* methodNames, PointerSet* fileNames)
{
    int count = gDvm.allocRecordCount;
    int idx = headIndex();
    int classCount, methodCount, fileCount;         /* debug stats */

    classCount = methodCount = fileCount = 0;

    while (count--) {
        AllocRecord* pRec = &gDvm.allocRecords[idx];

        dvmPointerSetAddEntry(classNames, pRec->clazz->descriptor);
        classCount++;

        int i;
        for (i = 0; i < kMaxAllocRecordStackDepth; i++) {
            if (pRec->stackElem[i].method == NULL)
                break;

            const Method* method = pRec->stackElem[i].method;
            dvmPointerSetAddEntry(classNames, method->clazz->descriptor);
            classCount++;
            dvmPointerSetAddEntry(methodNames, method->name);
            methodCount++;
            dvmPointerSetAddEntry(fileNames, getMethodSourceFile(method));
            fileCount++;
        }

        idx = (idx + 1) & (kNumAllocRecords-1);
    }

    LOGI("class %d/%d, method %d/%d, file %d/%d\n",
        dvmPointerSetGetCount(classNames), classCount,
        dvmPointerSetGetCount(methodNames), methodCount,
        dvmPointerSetGetCount(fileNames), fileCount);

    return true;
}

/*
 * Generate the base info (i.e. everything but the string tables).
 *
 * This should be called twice.  On the first call, "ptr" is NULL and
 * "baseLen" is zero.  The return value is used to allocate a buffer.
 * On the second call, "ptr" points to a data buffer, and "baseLen"
 * holds the value from the result of the first call.
 *
 * The size of the output data is returned.
 */
static size_t generateBaseOutput(u1* ptr, size_t baseLen,
    const PointerSet* classNames, const PointerSet* methodNames,
    const PointerSet* fileNames)
{
    u1* origPtr = ptr;
    int count = gDvm.allocRecordCount;
    int idx = headIndex();

    if (origPtr != NULL) {
        set1(&ptr[0], kMessageHeaderLen);
        set1(&ptr[1], kEntryHeaderLen);
        set1(&ptr[2], kStackFrameLen);
        set2BE(&ptr[3], count);
        set4BE(&ptr[5], baseLen);
        set2BE(&ptr[9], dvmPointerSetGetCount(classNames));
        set2BE(&ptr[11], dvmPointerSetGetCount(methodNames));
        set2BE(&ptr[13], dvmPointerSetGetCount(fileNames));
    }
    ptr += kMessageHeaderLen;

    while (count--) {
        AllocRecord* pRec = &gDvm.allocRecords[idx];

        /* compute depth */
        int  depth;
        for (depth = 0; depth < kMaxAllocRecordStackDepth; depth++) {
            if (pRec->stackElem[depth].method == NULL)
                break;
        }

        /* output header */
        if (origPtr != NULL) {
            set4BE(&ptr[0], pRec->size);
            set2BE(&ptr[4], pRec->threadId);
            set2BE(&ptr[6],
                dvmPointerSetFind(classNames, pRec->clazz->descriptor));
            set1(&ptr[8], depth);
        }
        ptr += kEntryHeaderLen;

        /* convert stack frames */
        int i;
        for (i = 0; i < depth; i++) {
            if (origPtr != NULL) {
                const Method* method = pRec->stackElem[i].method;
                int lineNum;

                lineNum = dvmLineNumFromPC(method, pRec->stackElem[i].pc);
                if (lineNum > 32767)
                    lineNum = 32767;

                set2BE(&ptr[0], dvmPointerSetFind(classNames,
                        method->clazz->descriptor));
                set2BE(&ptr[2], dvmPointerSetFind(methodNames,
                        method->name));
                set2BE(&ptr[4], dvmPointerSetFind(fileNames,
                        getMethodSourceFile(method)));
                set2BE(&ptr[6], (u2)lineNum);
            }
            ptr += kStackFrameLen;
        }

        idx = (idx + 1) & (kNumAllocRecords-1);
    }

    return ptr - origPtr;
}

/*
 * Compute the size required to store a string table.  Includes the length
 * word and conversion to UTF-16.
 */
static size_t computeStringTableSize(PointerSet* strings)
{
    int count = dvmPointerSetGetCount(strings);
    size_t size = 0;
    int i;

    for (i = 0; i < count; i++) {
        const char* str = (const char*) dvmPointerSetGetEntry(strings, i);

        size += 4 + dvmUtf8Len(str) * 2;
    }

    return size;
}

/*
 * Convert a UTF-8 string to UTF-16.  We also need to byte-swap the values
 * to big-endian, and we can't assume even alignment on the target.
 *
 * Returns the string's length, in characters.
 */
int convertUtf8ToUtf16BEUA(u1* utf16Str, const char* utf8Str)
{
    u1* origUtf16Str = utf16Str;

    while (*utf8Str != '\0') {
        u2 utf16 = dexGetUtf16FromUtf8(&utf8Str);       /* advances utf8Str */
        set2BE(utf16Str, utf16);
        utf16Str += 2;
    }

    return (utf16Str - origUtf16Str) / 2;
}

/*
 * Output a string table serially.
 */
static size_t outputStringTable(PointerSet* strings, u1* ptr)
{
    int count = dvmPointerSetGetCount(strings);
    u1* origPtr = ptr;
    int i;

    for (i = 0; i < count; i++) {
        const char* str = (const char*) dvmPointerSetGetEntry(strings, i);
        int charLen;

        /* copy UTF-8 string to big-endian unaligned UTF-16 */
        charLen = convertUtf8ToUtf16BEUA(&ptr[4], str);
        set4BE(&ptr[0], charLen);

        ptr += 4 + charLen * 2;
    }

    return ptr - origPtr;
}

/*
 * Generate a DDM packet with all of the tracked allocation data.
 *
 * On success, returns "true" with "*pData" and "*pDataLen" set.
 */
bool dvmGenerateTrackedAllocationReport(u1** pData, size_t* pDataLen)
{
    bool result = false;
    u1* buffer = NULL;

    dvmLockMutex(&gDvm.allocTrackerLock);

    /*
     * Part 1: generate string tables.
     */
    PointerSet* classNames = NULL;
    PointerSet* methodNames = NULL;
    PointerSet* fileNames = NULL;

    /*
     * Allocate storage.  Usually there's 60-120 of each thing (sampled
     * when max=512), but it varies widely and isn't closely bound to
     * the number of allocations we've captured.  The sets expand quickly
     * if needed.
     */
    classNames = dvmPointerSetAlloc(128);
    methodNames = dvmPointerSetAlloc(128);
    fileNames = dvmPointerSetAlloc(128);
    if (classNames == NULL || methodNames == NULL || fileNames == NULL) {
        LOGE("Failed allocating pointer sets\n");
        goto bail;
    }

    if (!populateStringTables(classNames, methodNames, fileNames))
        goto bail;

    if (false) {
        printf("Classes:\n");
        dumpStringTable(classNames);
        printf("Methods:\n");
        dumpStringTable(methodNames);
        printf("Files:\n");
        dumpStringTable(fileNames);
    }

    /*
     * Part 2: compute the size of the output.
     *
     * (Could also just write to an expanding buffer.)
     */
    size_t baseSize, totalSize;
    baseSize = generateBaseOutput(NULL, 0, classNames, methodNames, fileNames);
    assert(baseSize > 0);
    totalSize = baseSize;
    totalSize += computeStringTableSize(classNames);
    totalSize += computeStringTableSize(methodNames);
    totalSize += computeStringTableSize(fileNames);
    LOGI("Generated AT, size is %zd/%zd\n", baseSize, totalSize);

    /*
     * Part 3: allocate a buffer and generate the output.
     */
    u1* strPtr;

    buffer = (u1*) malloc(totalSize);
    strPtr = buffer + baseSize;
    generateBaseOutput(buffer, baseSize, classNames, methodNames, fileNames);
    strPtr += outputStringTable(classNames, strPtr);
    strPtr += outputStringTable(methodNames, strPtr);
    strPtr += outputStringTable(fileNames, strPtr);
    if (strPtr - buffer != (int)totalSize) {
        LOGE("size mismatch (%d vs %zd)\n", strPtr - buffer, totalSize);
        dvmAbort();
    }
    //dvmPrintHexDump(buffer, totalSize);

    *pData = buffer;
    *pDataLen = totalSize;
    buffer = NULL;          // don't free -- caller will own
    result = true;

bail:
    dvmPointerSetFree(classNames);
    dvmPointerSetFree(methodNames);
    dvmPointerSetFree(fileNames);
    free(buffer);
    dvmUnlockMutex(&gDvm.allocTrackerLock);
    //dvmDumpTrackedAllocations(false);
    return result;
}

/*
 * Dump the tracked allocations to the log file.
 *
 * If "enable" is set, we try to enable the feature if it's not already
 * active.
 */
void dvmDumpTrackedAllocations(bool enable)
{
    if (enable)
        dvmEnableAllocTracker();

    dvmLockMutex(&gDvm.allocTrackerLock);
    if (gDvm.allocRecords == NULL)
        goto bail;

    /*
     * "idx" is the head of the list.  We want to start at the end of the
     * list and move forward to the tail.
     */
    int idx = headIndex();
    int count = gDvm.allocRecordCount;

    LOGI("Tracked allocations, (head=%d count=%d)\n",
        gDvm.allocRecordHead, count);
    while (count--) {
        AllocRecord* pRec = &gDvm.allocRecords[idx];
        LOGI(" T=%-2d %6d %s\n",
            pRec->threadId, pRec->size, pRec->clazz->descriptor);

        if (true) {
            int i;
            for (i = 0; i < kMaxAllocRecordStackDepth; i++) {
                if (pRec->stackElem[i].method == NULL)
                    break;

                const Method* method = pRec->stackElem[i].method;
                if (dvmIsNativeMethod(method)) {
                    LOGI("    %s.%s (Native)\n",
                        method->clazz->descriptor, method->name);
                } else {
                    LOGI("    %s.%s +%d\n",
                        method->clazz->descriptor, method->name,
                        pRec->stackElem[i].pc);
                }
            }
        }

        /* pause periodically to help logcat catch up */
        if ((count % 5) == 0)
            usleep(40000);

        idx = (idx + 1) & (kNumAllocRecords-1);
    }

bail:
    dvmUnlockMutex(&gDvm.allocTrackerLock);
    if (false) {
        u1* data;
        size_t dataLen;
        if (dvmGenerateTrackedAllocationReport(&data, &dataLen))
            free(data);
    }
}