/**
* Let sleepers know about the wake-up condition.
*
* @param hl the list of waiting parties
* @param data waking-up data to supply to callback
*/
static void
wq_notify(hash_list_t *hl, void *data)
{
hash_list_iter_t *iter;
size_t i, count;
iter = hash_list_iterator(hl);
count = hash_list_length(hl);
i = 0;
while (hash_list_iter_has_next(iter)) {
wq_event_t *we = hash_list_iter_next(iter);
wq_status_t status;
wq_event_check(we);
/*
* Stop iteration in case callbacks have called wq_sleep() on the
* same waiting queue we're iterating on and added items to the list.
* This sanity check ensures we're not going to loop forever with a
* callback systematically appending something.
*/
if (i++ >= count) {
/* Something is odd, let them know about the calling stack */
s_critical("stopping after processing %zu item%s (list now has %u)",
count, plural(count), hash_list_length(hl));
}
status = (*we->cb)(we->arg, data);
switch (status) {
case WQ_SLEEP:
continue; /* Still sleeping, leave in the list */
case WQ_EXCLUSIVE:
case WQ_REMOVE:
goto remove;
}
s_error("invalid status %d returned by %s()",
status, stacktrace_function_name(we->cb));
remove:
hash_list_iter_remove(iter);
wq_event_free(we);
/*
* The callback may decide that we shouldn't continue notifying
* other sleepers (because it knows it grabbed a resource that others
* will need for instance). This is used as an early termination
* of the loop.
*/
if (WQ_EXCLUSIVE == status)
break;
}
hash_list_iter_release(&iter);
}
/**
* Callout queue callback fired when waiting event times out.
*/
static void
wq_timed_out(cqueue_t *cq, void *arg)
{
wq_event_t *we = arg;
hash_list_t *hl;
wq_status_t status;
wq_event_check(we);
g_assert(we->tm != NULL);
cq_zero(cq, &we->tm->timeout_ev);
hl = htable_lookup(waitqueue, we->key);
g_assert(hl != NULL);
/*
* Invoke the callback with the sentinel data signalling a timeout.
*/
status = (*we->cb)(we->arg, WQ_TIMED_OUT);
/*
* When the callback returns WQ_SLEEP, we re-instantiate the initial
* timeout.
*
* Otherwise the event is discarded (removed from the wait queue) and
* the callback will never be invoked again for this event.
*/
switch (status) {
case WQ_SLEEP:
we->tm->timeout_ev = cq_main_insert(we->tm->delay, wq_timed_out, we);
return;
case WQ_EXCLUSIVE:
s_critical("weird status WQ_EXCLUSIVE on timeout invocation of %s()",
stacktrace_function_name(we->cb));
/* FALL THROUGH */
case WQ_REMOVE:
hash_list_remove(hl, we);
/*
* Cleanup the table if it ends-up being empty.
*/
if (0 == hash_list_length(hl)) {
hash_list_free(&hl);
htable_remove(waitqueue, we->key);
}
wq_event_free(we);
return;
}
g_assert_not_reached();
}
/**
* Flush current /QH2.
*
* Depending how the QH2 builder is configured, this either sends the message
* to the target node or invokes a processing callback.
*/
static void
g2_build_qh2_flush(struct g2_qh2_builder *ctx)
{
pmsg_t *mb;
g_assert(ctx != NULL);
g_assert(ctx->t != NULL);
g_assert((ctx->n != NULL) ^ (ctx->cb != NULL));
/*
* Restore the order of children in the root packet to be the order we
* used when we added the nodes, since we prepend new children.
*/
g2_tree_reverse_children(ctx->t);
/*
* If sending over UDP, ask for reliable delivery of the query hit.
* To be able to monitor the fate of the message, we asssociate a free
* routine to it.
*/
if (ctx->to_udp) {
struct g2_qh2_pmsg_info *pmi;
WALLOC0(pmi);
pmi->magic = G2_QH2_PMI_MAGIC;
pmi->hub_id = nid_ref(NODE_ID(ctx->hub));
mb = g2_build_pmsg_extended(ctx->t, g2_qh2_pmsg_free, pmi);
pmsg_mark_reliable(mb);
} else {
mb = g2_build_pmsg(ctx->t);
}
if (GNET_PROPERTY(g2_debug) > 3) {
g_debug("%s(): flushing the following hit for "
"Q2 #%s to %s%s (%d bytes):",
G_STRFUNC, guid_hex_str(ctx->muid),
NULL == ctx->n ?
stacktrace_function_name(ctx->cb) : node_infostr(ctx->n),
NULL == ctx->n ? "()" : "", pmsg_size(mb));
g2_tfmt_tree_dump(ctx->t, stderr, G2FMT_O_PAYLOAD | G2FMT_O_PAYLEN);
}
if (ctx->n != NULL)
g2_node_send(ctx->n, mb);
else
(*ctx->cb)(mb, ctx->arg);
ctx->messages++;
ctx->current_size = 0;
g2_tree_free_null(&ctx->t);
}
/**
* Remove an event from the queue.
*/
static void
wq_remove(wq_event_t *we)
{
hash_list_t *hl;
wq_event_check(we);
hl = htable_lookup(waitqueue, we->key);
if (NULL == hl) {
s_critical("attempt to remove event %s() on unknown key %p",
stacktrace_function_name(we->cb), we->key);
} if (NULL == hash_list_remove(hl, we)) {
s_critical("attempt to remove unknown event %s() on %p",
stacktrace_function_name(we->cb), we->key);
} else if (0 == hash_list_length(hl)) {
hash_list_free(&hl);
htable_remove(waitqueue, we->key);
}
wq_event_free(we);
}
/**
* Hash list iterator callback to free and remove waiting events.
*/
static bool
wq_free_waiting(void *key, void *unused_data)
{
wq_event_t *we = key;
wq_event_check(we);
(void) unused_data;
s_warning("leaked waiting event %s() on %p",
stacktrace_function_name(we->cb), we->key);
wq_event_free(we);
return TRUE;
}
/**
* RPC timeout callback.
*/
static void
g2_rpc_timeout(cqueue_t *cq, void *obj)
{
struct g2_rpc *gr = obj;
g2_rpc_check(gr);
if (GNET_PROPERTY(g2_rpc_debug) > 1) {
g_debug("%s(): /%s RPC to %s timed out, calling %s()",
G_STRFUNC, g2_msg_type_name(gr->key.type),
host_addr_to_string(gr->key.addr),
stacktrace_function_name(gr->cb));
}
cq_zero(cq, &gr->timeout_ev);
(*gr->cb)(NULL, NULL, gr->arg);
g2_rpc_free(gr, FALSE);
}
/*
* Ensure we do not have a check/hook installed already, otherwise loudly
* warn them once, as this is probably not intended!
*/
static void
pmsg_no_presend_check(const pmsg_t * const mb, const char *caller)
{
/*
* Because m_check and m_hook are in the same union and have the same
* memory size, it is sufficient to check for one field being non-NULL.
*/
if G_LIKELY(NULL == mb->m_u.m_check)
return;
s_carp_once("%s(): mb=%p (%d byte%s, prio=%u, refcnt=%u, flags=0x%x)"
" already has %s %s()",
caller, mb, pmsg_size(mb), plural(pmsg_size(mb)),
mb->m_prio, mb->m_refcnt, mb->m_flags,
(mb->m_flags & PMSG_PF_HOOK) ? "transmit hook" : "can-send callback",
stacktrace_function_name(mb->m_u.m_check));
}
/**
* Trigger callback and then put the watchdog to sleep, ignoring any desire
* from the callback to re-arm the watchdog.
*
* @return TRUE if we stopped the watchdog, FALSE if it was already aslept,
* in which case the trigger was not invoked.
*/
bool
wd_expire(watchdog_t *wd)
{
watchdog_check(wd);
if (NULL == wd->ev)
return FALSE;
cq_cancel(&wd->ev);
(*wd->trigger)(wd, wd->arg);
if (wd->ev != NULL) {
g_critical("%s(): "
"watchdog \"%s\" re-armed within %s() callback, turning it off",
G_STRFUNC, wd_name(wd), stacktrace_function_name(wd->trigger));
}
return TRUE;
}
/**
* Notification that a message was received that could be the answer to
* a pending RPC.
*
* @param n the node from which we got the message
* @param t the received message tree
*
* @return TRUE if the message was indeed an RPC reply, FALSE otherwise.
*/
bool
g2_rpc_answer(const gnutella_node_t *n, const g2_tree_t *t)
{
struct g2_rpc *gr;
struct g2_rpc_key key;
enum g2_msg type;
type = g2_msg_name_type(g2_tree_name(t));
key.type = g2_rpc_send_type(type);
key.addr = n->addr;
gr = hevset_lookup(g2_rpc_pending, &key);
if (NULL == gr) {
/*
* No known RPC, but wait... we can receive a /QKA when we issue a /Q2
* and the query key we knew for the remote host has expired, hence
* we must look whether there is not a /Q2 pending as well in that
* case. Once again, the lack of MUID in these messages is a handicap.
*/
if (G2_MSG_QKA == type) {
key.type = G2_MSG_Q2;
gr = hevset_lookup(g2_rpc_pending, &key);
if (gr != NULL)
goto found; /* Sent a /Q2, got a /QKA back */
}
if (GNET_PROPERTY(g2_rpc_debug) > 1) {
g_debug("%s(): unexpected /%s RPC reply from %s",
G_STRFUNC, g2_msg_type_name(key.type), node_infostr(n));
}
return FALSE;
}
found:
/*
* Got a reply for an RPC we sent, based solely on message type and
* source address of the message. This is weak, but G2 works like that.
*
* Weakness comes from the fact that we cannot have multiple RPCs with
* a same IP address but towards different ports, nor have concurrent
* RPCs with the same host for several different by similar requests
* (although this can be viewed as anti-hammering, servers should protect
* against that in different ways, a crippled protocol not being an answer).
*
* The only transaction where we could use the MUID is /Q2 -> /QA but we
* leave that check to the GUESS layer and make sure here that we have only
* one single RPC transaction at a time with a given IP address.
*/
if (GNET_PROPERTY(g2_rpc_debug) > 2) {
g_debug("%s(): /%s RPC to %s got a /%s reply, calling %s()",
G_STRFUNC, g2_msg_type_name(gr->key.type),
host_addr_to_string(gr->key.addr), g2_tree_name(t),
stacktrace_function_name(gr->cb));
}
(*gr->cb)(n, t, gr->arg);
g2_rpc_free(gr, FALSE);
return TRUE;
}
/**
* Read value from database file, returning a pointer to the allocated
* deserialized data. These data can be modified freely and stored back,
* but their lifetime will not exceed that of the next call to a dbmw
* operation on the same descriptor.
*
* User code does not need to bother with freeing the allocated data, this
* is managed directly by the DBM wrapper.
*
* @param dw the DBM wrapper
* @param key the key (constant-width, determined at open time)
* @param lenptr if non-NULL, writes length of (deserialized) value
*
* @return pointer to value, or NULL if it was either not found or the
* deserialization failed.
*/
G_GNUC_HOT void *
dbmw_read(dbmw_t *dw, const void *key, size_t *lenptr)
{
struct cached *entry;
dbmap_datum_t dval;
dbmw_check(dw);
g_assert(key);
dw->r_access++;
entry = map_lookup(dw->values, key);
if (entry) {
if (dbg_ds_debugging(dw->dbg, 5, DBG_DSF_CACHING | DBG_DSF_ACCESS)) {
dbg_ds_log(dw->dbg, dw, "%s: read cache hit on %s key=%s%s",
G_STRFUNC, entry->dirty ? "dirty" : "clean",
dbg_ds_keystr(dw->dbg, key, (size_t) -1),
entry->absent ? " (absent)" : "");
}
dw->r_hits++;
if (lenptr)
*lenptr = entry->len;
return entry->data;
}
/*
* Not cached, must read from DB.
*/
dw->ioerr = FALSE;
dval = dbmap_lookup(dw->dm, key);
if (dbmap_has_ioerr(dw->dm)) {
dw->ioerr = TRUE;
dw->error = errno;
s_warning_once_per(LOG_PERIOD_SECOND,
"DBMW \"%s\" I/O error whilst reading entry: %s",
dw->name, dbmap_strerror(dw->dm));
return NULL;
} else if (NULL == dval.data)
return NULL; /* Not found in DB */
/*
* Value was found, allocate a cache entry object for it.
*/
WALLOC0(entry);
/*
* Deserialize data if needed.
*/
if (dw->unpack) {
/*
* Allocate cache entry arena to hold the deserialized version.
*/
entry->data = walloc(dw->value_size);
entry->len = dw->value_size;
bstr_reset(dw->bs, dval.data, dval.len, BSTR_F_ERROR);
if (!dbmw_deserialize(dw, dw->bs, entry->data, dw->value_size)) {
s_critical("DBMW \"%s\" deserialization error in %s(): %s",
dw->name, stacktrace_function_name(dw->unpack),
bstr_error(dw->bs));
/* Not calling value free routine on deserialization failures */
wfree(entry->data, dw->value_size);
WFREE(entry);
return NULL;
}
if (lenptr)
*lenptr = dw->value_size;
} else {
g_assert(dw->value_size >= dval.len);
if (dval.len) {
entry->len = dval.len;
entry->data = wcopy(dval.data, dval.len);
} else {
entry->data = NULL;
entry->len = 0;
}
if (lenptr)
*lenptr = dval.len;
}
g_assert((entry->len != 0) == (entry->data != NULL));
/*
* Insert into cache.
*/
(void) allocate_entry(dw, key, entry);
if (dbg_ds_debugging(dw->dbg, 4, DBG_DSF_CACHING)) {
dbg_ds_log(dw->dbg, dw, "%s: cached %s key=%s%s",
G_STRFUNC, entry->dirty ? "dirty" : "clean",
dbg_ds_keystr(dw->dbg, key, (size_t) -1),
entry->absent ? " (absent)" : "");
}
return entry->data;
}
/**
* Write back cached value to disk.
* @return TRUE on success
*/
static bool
write_back(dbmw_t *dw, const void *key, struct cached *value)
{
dbmap_datum_t dval;
bool ok;
g_assert(value->dirty);
if (value->absent) {
/* Key not present, value is null item */
dval.data = NULL;
dval.len = 0;
} else {
/*
* Serialize value into our reused message block if a
* serialization routine was provided.
*/
if (dw->pack) {
pmsg_reset(dw->mb);
(*dw->pack)(dw->mb, value->data);
dval.data = pmsg_start(dw->mb);
dval.len = pmsg_size(dw->mb);
/*
* We allocated the message block one byte larger than the
* maximum size, in order to detect unexpected serialization
* overflows.
*/
if (dval.len > dw->value_data_size) {
/* Don't s_carp() as this is asynchronous wrt data change */
s_critical("DBMW \"%s\" serialization overflow in %s() "
"whilst flushing dirty entry",
dw->name, stacktrace_function_name(dw->pack));
return FALSE;
}
} else {
dval.data = value->data;
dval.len = value->len;
}
}
/*
* If cached entry is absent, delete the key.
* Otherwise store the serialized value.
*
* Dirty bit is cleared on success.
*/
if (
dbg_ds_debugging(dw->dbg, 1,
DBG_DSF_CACHING | DBG_DSF_UPDATE | DBG_DSF_INSERT | DBG_DSF_DELETE)
) {
dbg_ds_log(dw->dbg, dw, "%s: %s dirty value (%zu byte%s) key=%s",
G_STRFUNC, value->absent ? "deleting" : "flushing",
dval.len, plural(dval.len),
dbg_ds_keystr(dw->dbg, key, (size_t) -1));
}
dw->ioerr = FALSE;
ok = value->absent ?
dbmap_remove(dw->dm, key) : dbmap_insert(dw->dm, key, dval);
if (ok) {
value->dirty = FALSE;
} else if (dbmap_has_ioerr(dw->dm)) {
dw->ioerr = TRUE;
dw->error = errno;
s_warning("DBMW \"%s\" I/O error whilst %s dirty entry: %s",
dw->name, value->absent ? "deleting" : "flushing",
dbmap_strerror(dw->dm));
} else {
s_warning("DBMW \"%s\" error whilst %s dirty entry: %s",
dw->name, value->absent ? "deleting" : "flushing",
dbmap_strerror(dw->dm));
}
return ok;
}
/**
* Iterate over the DB, invoking the callback on each item along with the
* supplied argument and removing the item when the callback returns TRUE.
*
* @return the amount of removed entries.
*/
size_t
dbmw_foreach_remove(dbmw_t *dw, dbmw_cbr_t cbr, void *arg)
{
struct foreach_ctx ctx;
struct cache_foreach_ctx fctx;
size_t pruned;
dbmw_check(dw);
if (dbg_ds_debugging(dw->dbg, 1, DBG_DSF_ITERATOR)) {
dbg_ds_log(dw->dbg, dw, "%s: starting with %s(%p)", G_STRFUNC,
stacktrace_function_name(cbr), arg);
}
/*
* Before iterating we flush the deleted keys we know about in the cache
* and whose deletion was deferred, so that the underlying map will
* not have to iterate on them.
*/
dbmw_sync(dw, DBMW_SYNC_CACHE | DBMW_DELETED_ONLY);
/*
* Some values may be present only in the cache. Hence we clear all
* marks in the cache and each traversed value that happens to be
* present in the cache will be marked as "traversed".
*
* We flushed deleted keys above, but that does not remove them from
* the cache structure. We don't need to traverse these after iterating
* on the map, so we make sure they are artifically set to "traversed".
*/
ctx.u.cbr = cbr;
ctx.arg = arg;
ctx.dw = dw;
map_foreach(dw->values, cache_reset_before_traversal, NULL);
pruned = dbmap_foreach_remove(dw->dm, dbmw_foreach_remove_trampoline, &ctx);
ZERO(&fctx);
fctx.removing = TRUE;
fctx.foreach = &ctx;
fctx.u.cbr = dbmw_foreach_remove_trampoline;
/*
* Continue traversal with all the cached entries that were not traversed
* already because they do not exist in the underlying map.
*
* We count these and remember how many there are so that we can determine
* the correct overall item count after an iteration without having to
* flush all the dirty values!
*
* Any cached entry that needs to be removed will be marked as such
* and we'll complete processing by discarding from the cache all
* the entries that have been marked as "removable" during the traversal.
*/
map_foreach(dw->values, cache_finish_traversal, &fctx);
map_foreach_remove(dw->values, cache_free_removable, dw);
dw->cached = fctx.cached;
dw->count_needs_sync = FALSE; /* We just counted items the slow way! */
if (dbg_ds_debugging(dw->dbg, 1, DBG_DSF_ITERATOR)) {
dbg_ds_log(dw->dbg, dw, "%s: done with %s(%p): "
"pruned %zu from dbmap, %zu from cache (%zu total), "
"has %zu unflushed entr%s in cache",
G_STRFUNC,
stacktrace_function_name(cbr), arg,
pruned, fctx.removed, pruned + fctx.removed,
dw->cached, plural_y(dw->cached));
}
return pruned + fctx.removed;
}
/**
* Common code for dbmw_foreach_trampoline() and
* dbmw_foreach_remove_trampoline().
*/
static bool
dbmw_foreach_common(bool removing, void *key, dbmap_datum_t *d, void *arg)
{
struct foreach_ctx *ctx = arg;
dbmw_t *dw = ctx->dw;
struct cached *entry;
dbmw_check(dw);
entry = map_lookup(dw->values, key);
if (entry != NULL) {
/*
* Key / value pair is present in the cache.
*
* This affects us in two ways:
*
* - We may already know that the key was deleted, in which case
* that entry is just skipped: no further access is possible
* through DBMW until that key is recreated. We still return
* TRUE to make sure the lower layers will delete the entry
* physically, since deletion has not been flushed yet (that's
* the reason we're still iterating on it).
*
* - Should the cached key need to be deleted (as determined by
* the user callback, we make sure we delete the entry in the
* cache upon callback return).
*
* Because we sync the cache (the dirty deleted items only), we should
* normally never iterate on a deleted entry, coming from the
* underlying database, hence we loudly complain!
*/
entry->traversed = TRUE; /* Signal we iterated on cached value */
if (entry->absent) {
s_carp("%s(): DBMW \"%s\" iterating over a %s absent key in cache!",
G_STRFUNC, dw->name, entry->dirty ? "dirty" : "clean");
return TRUE; /* Key was already deleted, info cached */
}
if (removing) {
bool status;
status = (*ctx->u.cbr)(key, entry->data, entry->len, ctx->arg);
if (status) {
entry->removable = TRUE; /* Discard it after traversal */
}
return status;
} else {
(*ctx->u.cb)(key, entry->data, entry->len, ctx->arg);
return FALSE;
}
} else {
bool status = FALSE;
void *data = d->data;
size_t len = d->len;
/*
* Deserialize data if needed, but do not cache this value.
* Iterating over the map must not disrupt the cache.
*/
if (dw->unpack) {
len = dw->value_size;
data = walloc(len);
bstr_reset(dw->bs, d->data, d->len, BSTR_F_ERROR);
if (!dbmw_deserialize(dw, dw->bs, data, len)) {
s_critical("DBMW \"%s\" deserialization error in %s(): %s",
dw->name,
stacktrace_function_name(dw->unpack),
bstr_error(dw->bs));
/* Not calling value free routine on deserialization failures */
wfree(data, len);
return FALSE;
}
}
if (removing) {
status = (*ctx->u.cbr)(key, data, len, ctx->arg);
} else {
(*ctx->u.cb)(key, data, len, ctx->arg);
}
if (dw->unpack) {
if (dw->valfree)
(*dw->valfree)(data, len);
wfree(data, len);
}
return status;
}
}
