// If we're using more than 256K of memory for log entries, prune
// at least 10% of the log entries.
//
// mLogElementsLock must be held when this function is called.
void LogBuffer::maybePrune(log_id_t id) {
size_t sizes = stats.sizes(id);
if (sizes > log_buffer_size(id)) {
size_t sizeOver90Percent = sizes - ((log_buffer_size(id) * 9) / 10);
size_t elements = stats.elements(id);
unsigned long pruneRows = elements * sizeOver90Percent / sizes;
elements /= 10;
if (pruneRows <= elements) {
pruneRows = elements;
}
prune(id, pruneRows);
}
}
// set the total space allocated to "id"
int LogBuffer::setSize(log_id_t id, unsigned long size) {
// Reasonable limits ...
if (!valid_size(size)) {
return -1;
}
pthread_mutex_lock(&mLogElementsLock);
log_buffer_size(id) = size;
pthread_mutex_unlock(&mLogElementsLock);
return 0;
}
// Prune at most 10% of the log entries or maxPrune, whichever is less.
//
// mLogElementsLock must be held when this function is called.
void LogBuffer::maybePrune(log_id_t id) {
size_t sizes = stats.sizes(id);
unsigned long maxSize = log_buffer_size(id);
if (sizes > maxSize) {
size_t sizeOver = sizes - ((maxSize * 9) / 10);
size_t elements = stats.realElements(id);
size_t minElements = elements / 100;
if (minElements < minPrune) {
minElements = minPrune;
}
unsigned long pruneRows = elements * sizeOver / sizes;
if (pruneRows < minElements) {
pruneRows = minElements;
}
if (pruneRows > maxPrune) {
pruneRows = maxPrune;
}
prune(id, pruneRows);
}
}
int base_get(struct base *base,
const char *key, unsigned key_sz,
char *buf, unsigned buf_sz)
{
uint128_t key_hash = md5(key, key_sz);
uint64_t log_remno;
int hpos;
int r = itree_get2(base->itree, key_hash, &log_remno, &hpos);
if (r < 0) {
return -1;
}
struct log *log = log_by_remno(base->logs, log_remno);
/* TODO: get rid of the awful malloc */
unsigned data_sz = log_buffer_size(log, hpos);
char *data = malloc(data_sz);
struct keyvalue kv;
r = log_read(log, hpos, data, data_sz, &kv);
if (r < 0) {
free(data);
return -2;
}
if (kv.value_sz > buf_sz) {
free(data);
return -3;
}
if (key_sz != kv.key_sz || memcmp(key, kv.key, key_sz) != 0) {
log_error(base->db, "Congratulations! You just found a "
"collision! Apparently key %*s has the same md5 hash as %*s!",
key_sz, key,
kv.key_sz, kv.key);
free(data);
return -1;
}
memcpy(buf, kv.value, kv.value_sz);
free(data);
return kv.value_sz;
}
// get the total space allocated to "id"
unsigned long LogBuffer::getSize(log_id_t id) {
pthread_mutex_lock(&mLogElementsLock);
size_t retval = log_buffer_size(id);
pthread_mutex_unlock(&mLogElementsLock);
return retval;
}
// prune "pruneRows" of type "id" from the buffer.
//
// mLogElementsLock must be held when this function is called.
void LogBuffer::prune(log_id_t id, unsigned long pruneRows) {
LogTimeEntry *oldest = NULL;
LogTimeEntry::lock();
// Region locked?
LastLogTimes::iterator t = mTimes.begin();
while(t != mTimes.end()) {
LogTimeEntry *entry = (*t);
if (entry->owned_Locked()
&& (!oldest || (oldest->mStart > entry->mStart))) {
oldest = entry;
}
t++;
}
LogBufferElementCollection::iterator it;
// prune by worst offender by uid
while (pruneRows > 0) {
// recalculate the worst offender on every batched pass
uid_t worst = (uid_t) -1;
size_t worst_sizes = 0;
size_t second_worst_sizes = 0;
if ((id != LOG_ID_CRASH) && mPrune.worstUidEnabled()) {
LidStatistics &l = stats.id(id);
l.sort();
UidStatisticsCollection::iterator iu = l.begin();
if (iu != l.end()) {
UidStatistics *u = *iu;
worst = u->getUid();
worst_sizes = u->sizes();
if (++iu != l.end()) {
second_worst_sizes = (*iu)->sizes();
}
}
}
bool kick = false;
for(it = mLogElements.begin(); it != mLogElements.end();) {
LogBufferElement *e = *it;
if (oldest && (oldest->mStart <= e->getMonotonicTime())) {
break;
}
if (e->getLogId() != id) {
++it;
continue;
}
uid_t uid = e->getUid();
if (uid == worst) {
it = mLogElements.erase(it);
unsigned short len = e->getMsgLen();
stats.subtract(len, id, worst, e->getPid());
delete e;
kick = true;
pruneRows--;
if ((pruneRows == 0) || (worst_sizes < second_worst_sizes)) {
break;
}
worst_sizes -= len;
} else if (mPrune.naughty(e)) { // BlackListed
it = mLogElements.erase(it);
stats.subtract(e->getMsgLen(), id, uid, e->getPid());
delete e;
pruneRows--;
if (pruneRows == 0) {
break;
}
} else {
++it;
}
}
if (!kick || !mPrune.worstUidEnabled()) {
break; // the following loop will ask bad clients to skip/drop
}
}
bool whitelist = false;
it = mLogElements.begin();
while((pruneRows > 0) && (it != mLogElements.end())) {
LogBufferElement *e = *it;
if (e->getLogId() == id) {
if (oldest && (oldest->mStart <= e->getMonotonicTime())) {
if (!whitelist) {
if (stats.sizes(id) > (2 * log_buffer_size(id))) {
// kick a misbehaving log reader client off the island
oldest->release_Locked();
} else {
oldest->triggerSkip_Locked(pruneRows);
}
}
break;
}
if (mPrune.nice(e)) { // WhiteListed
whitelist = true;
it++;
continue;
}
it = mLogElements.erase(it);
stats.subtract(e->getMsgLen(), id, e->getUid(), e->getPid());
delete e;
pruneRows--;
} else {
it++;
}
}
if (whitelist && (pruneRows > 0)) {
it = mLogElements.begin();
while((it != mLogElements.end()) && (pruneRows > 0)) {
LogBufferElement *e = *it;
if (e->getLogId() == id) {
if (oldest && (oldest->mStart <= e->getMonotonicTime())) {
if (stats.sizes(id) > (2 * log_buffer_size(id))) {
// kick a misbehaving log reader client off the island
oldest->release_Locked();
} else {
oldest->triggerSkip_Locked(pruneRows);
}
break;
}
it = mLogElements.erase(it);
stats.subtract(e->getMsgLen(), id, e->getUid(), e->getPid());
delete e;
pruneRows--;
} else {
it++;
}
}
}
LogTimeEntry::unlock();
}
// prune "pruneRows" of type "id" from the buffer.
//
// This garbage collection task is used to expire log entries. It is called to
// remove all logs (clear), all UID logs (unprivileged clear), or every
// 256 or 10% of the total logs (whichever is less) to prune the logs.
//
// First there is a prep phase where we discover the reader region lock that
// acts as a backstop to any pruning activity to stop there and go no further.
//
// There are three major pruning loops that follow. All expire from the oldest
// entries. Since there are multiple log buffers, the Android logging facility
// will appear to drop entries 'in the middle' when looking at multiple log
// sources and buffers. This effect is slightly more prominent when we prune
// the worst offender by logging source. Thus the logs slowly loose content
// and value as you move back in time. This is preferred since chatty sources
// invariably move the logs value down faster as less chatty sources would be
// expired in the noise.
//
// The first loop performs blacklisting and worst offender pruning. Falling
// through when there are no notable worst offenders and have not hit the
// region lock preventing further worst offender pruning. This loop also looks
// after managing the chatty log entries and merging to help provide
// statistical basis for blame. The chatty entries are not a notification of
// how much logs you may have, but instead represent how much logs you would
// have had in a virtual log buffer that is extended to cover all the in-memory
// logs without loss. They last much longer than the represented pruned logs
// since they get multiplied by the gains in the non-chatty log sources.
//
// The second loop get complicated because an algorithm of watermarks and
// history is maintained to reduce the order and keep processing time
// down to a minimum at scale. These algorithms can be costly in the face
// of larger log buffers, or severly limited processing time granted to a
// background task at lowest priority.
//
// This second loop does straight-up expiration from the end of the logs
// (again, remember for the specified log buffer id) but does some whitelist
// preservation. Thus whitelist is a Hail Mary low priority, blacklists and
// spam filtration all take priority. This second loop also checks if a region
// lock is causing us to buffer too much in the logs to help the reader(s),
// and will tell the slowest reader thread to skip log entries, and if
// persistent and hits a further threshold, kill the reader thread.
//
// The third thread is optional, and only gets hit if there was a whitelist
// and more needs to be pruned against the backstop of the region lock.
//
// mLogElementsLock must be held when this function is called.
//
bool LogBuffer::prune(log_id_t id, unsigned long pruneRows, uid_t caller_uid) {
LogTimeEntry *oldest = NULL;
bool busy = false;
bool clearAll = pruneRows == ULONG_MAX;
LogTimeEntry::lock();
// Region locked?
LastLogTimes::iterator t = mTimes.begin();
while(t != mTimes.end()) {
LogTimeEntry *entry = (*t);
if (entry->owned_Locked() && entry->isWatching(id)
&& (!oldest || (oldest->mStart > entry->mStart))) {
oldest = entry;
}
t++;
}
LogBufferElementCollection::iterator it;
if (caller_uid != AID_ROOT) {
// Only here if clearAll condition (pruneRows == ULONG_MAX)
for(it = mLogElements.begin(); it != mLogElements.end();) {
LogBufferElement *e = *it;
if ((e->getLogId() != id) || (e->getUid() != caller_uid)) {
++it;
continue;
}
if (oldest && (oldest->mStart <= e->getSequence())) {
oldest->triggerSkip_Locked(id, pruneRows);
busy = true;
break;
}
it = erase(it);
pruneRows--;
}
LogTimeEntry::unlock();
return busy;
}
// prune by worst offender by uid
bool hasBlacklist = mPrune.naughty();
while (!clearAll && (pruneRows > 0)) {
// recalculate the worst offender on every batched pass
uid_t worst = (uid_t) -1;
size_t worst_sizes = 0;
size_t second_worst_sizes = 0;
if (worstUidEnabledForLogid(id) && mPrune.worstUidEnabled()) {
std::unique_ptr<const UidEntry *[]> sorted = stats.sort(2, id);
if (sorted.get()) {
if (sorted[0] && sorted[1]) {
worst_sizes = sorted[0]->getSizes();
// Calculate threshold as 12.5% of available storage
size_t threshold = log_buffer_size(id) / 8;
if (worst_sizes > threshold) {
worst = sorted[0]->getKey();
second_worst_sizes = sorted[1]->getSizes();
if (second_worst_sizes < threshold) {
second_worst_sizes = threshold;
}
}
}
}
}
// skip if we have neither worst nor naughty filters
if ((worst == (uid_t) -1) && !hasBlacklist) {
break;
}
bool kick = false;
bool leading = true;
it = mLogElements.begin();
// Perform at least one mandatory garbage collection cycle in following
// - clear leading chatty tags
// - coalesce chatty tags
// - check age-out of preserved logs
bool gc = pruneRows <= 1;
if (!gc && (worst != (uid_t) -1)) {
LogBufferIteratorMap::iterator f = mLastWorstUid[id].find(worst);
if ((f != mLastWorstUid[id].end())
&& (f->second != mLogElements.end())) {
leading = false;
it = f->second;
}
}
static const timespec too_old = {
EXPIRE_HOUR_THRESHOLD * 60 * 60, 0
};
LogBufferElementCollection::iterator lastt;
lastt = mLogElements.end();
--lastt;
LogBufferElementLast last;
while (it != mLogElements.end()) {
LogBufferElement *e = *it;
if (oldest && (oldest->mStart <= e->getSequence())) {
busy = true;
break;
}
if (e->getLogId() != id) {
++it;
continue;
}
unsigned short dropped = e->getDropped();
// remove any leading drops
if (leading && dropped) {
it = erase(it);
continue;
}
if (dropped && last.coalesce(e, dropped)) {
it = erase(it, true);
continue;
}
if (hasBlacklist && mPrune.naughty(e)) {
last.clear(e);
it = erase(it);
if (dropped) {
continue;
}
pruneRows--;
if (pruneRows == 0) {
break;
}
if (e->getUid() == worst) {
kick = true;
if (worst_sizes < second_worst_sizes) {
break;
}
worst_sizes -= e->getMsgLen();
}
continue;
}
if ((e->getRealTime() < ((*lastt)->getRealTime() - too_old))
|| (e->getRealTime() > (*lastt)->getRealTime())) {
break;
}
if (dropped) {
last.add(e);
if ((!gc && (e->getUid() == worst))
|| (mLastWorstUid[id].find(e->getUid())
== mLastWorstUid[id].end())) {
mLastWorstUid[id][e->getUid()] = it;
}
++it;
continue;
}
if (e->getUid() != worst) {
leading = false;
last.clear(e);
++it;
continue;
}
pruneRows--;
if (pruneRows == 0) {
break;
}
kick = true;
unsigned short len = e->getMsgLen();
// do not create any leading drops
if (leading) {
it = erase(it);
} else {
stats.drop(e);
e->setDropped(1);
if (last.coalesce(e, 1)) {
it = erase(it, true);
} else {
last.add(e);
if (!gc || (mLastWorstUid[id].find(worst)
== mLastWorstUid[id].end())) {
mLastWorstUid[id][worst] = it;
}
++it;
}
}
if (worst_sizes < second_worst_sizes) {
break;
}
worst_sizes -= len;
}
last.clear();
if (!kick || !mPrune.worstUidEnabled()) {
break; // the following loop will ask bad clients to skip/drop
}
}
bool whitelist = false;
bool hasWhitelist = mPrune.nice() && !clearAll;
it = mLogElements.begin();
while((pruneRows > 0) && (it != mLogElements.end())) {
LogBufferElement *e = *it;
if (e->getLogId() != id) {
it++;
continue;
}
if (oldest && (oldest->mStart <= e->getSequence())) {
busy = true;
if (whitelist) {
break;
}
if (stats.sizes(id) > (2 * log_buffer_size(id))) {
// kick a misbehaving log reader client off the island
oldest->release_Locked();
} else {
oldest->triggerSkip_Locked(id, pruneRows);
}
break;
}
if (hasWhitelist && !e->getDropped() && mPrune.nice(e)) { // WhiteListed
whitelist = true;
it++;
continue;
}
it = erase(it);
pruneRows--;
}
// Do not save the whitelist if we are reader range limited
if (whitelist && (pruneRows > 0)) {
it = mLogElements.begin();
while((it != mLogElements.end()) && (pruneRows > 0)) {
LogBufferElement *e = *it;
if (e->getLogId() != id) {
++it;
continue;
}
if (oldest && (oldest->mStart <= e->getSequence())) {
busy = true;
if (stats.sizes(id) > (2 * log_buffer_size(id))) {
// kick a misbehaving log reader client off the island
oldest->release_Locked();
} else {
oldest->triggerSkip_Locked(id, pruneRows);
}
break;
}
it = erase(it);
pruneRows--;
}
}
LogTimeEntry::unlock();
return (pruneRows > 0) && busy;
}