bool upb_symtab_dfs(upb_def *def, upb_def **open_defs, int n,
upb_strtable *addtab) {
// This linear search makes the DFS O(n^2) in the length of the paths.
// Could make this O(n) with a hash table, but n is small.
for (int i = 0; i < n; i++) {
if (def == open_defs[i]) return false;
}
bool needcopy = false;
upb_msgdef *m = upb_dyncast_msgdef(def);
if (m) {
upb_msg_iter i;
open_defs[n++] = def;
for(i = upb_msg_begin(m); !upb_msg_done(i); i = upb_msg_next(m, i)) {
upb_fielddef *f = upb_msg_iter_field(i);
if (!upb_hassubdef(f)) continue;
needcopy |= upb_symtab_dfs(f->def, open_defs, n, addtab);
}
}
bool replacing = (upb_strtable_lookup(addtab, m->base.fqname) != NULL);
if (needcopy && !replacing) {
upb_symtab_ent e = {upb_def_dup(def)};
upb_strtable_insert(addtab, def->fqname, &e);
replacing = true;
}
return replacing;
}
bool upb_msgdef_addfields(upb_msgdef *m, upb_fielddef *const *fields, int n,
const void *ref_donor, upb_status *s) {
// Check constraints for all fields before performing any action.
for (int i = 0; i < n; i++) {
upb_fielddef *f = fields[i];
// TODO(haberman): handle the case where two fields of the input duplicate
// name or number.
if (f->msgdef != NULL) {
upb_status_seterrliteral(s, "fielddef already belongs to a message");
return false;
} else if (upb_fielddef_name(f) == NULL || upb_fielddef_number(f) == 0) {
upb_status_seterrliteral(s, "field name or number were not set");
return false;
} else if(upb_msgdef_itof(m, upb_fielddef_number(f)) ||
upb_msgdef_ntof(m, upb_fielddef_name(f))) {
upb_status_seterrliteral(s, "duplicate field name or number");
return false;
}
}
// Constraint checks ok, perform the action.
for (int i = 0; i < n; i++) {
upb_fielddef *f = fields[i];
f->msgdef = m;
upb_inttable_insert(&m->itof, upb_fielddef_number(f), upb_value_ptr(f));
upb_strtable_insert(&m->ntof, upb_fielddef_name(f), upb_value_ptr(f));
upb_ref2(f, m);
upb_ref2(m, f);
if (ref_donor) upb_fielddef_unref(f, ref_donor);
}
return true;
}
bool upb_msgdef_addfields(upb_msgdef *m, upb_fielddef *const *fields, int n,
const void *ref_donor) {
// Check constraints for all fields before performing any action.
for (int i = 0; i < n; i++) {
upb_fielddef *f = fields[i];
// TODO(haberman): handle the case where two fields of the input duplicate
// name or number.
if (f->msgdef != NULL ||
upb_fielddef_name(f) == NULL || upb_fielddef_number(f) == 0 ||
upb_msgdef_itof(m, upb_fielddef_number(f)) ||
upb_msgdef_ntof(m, upb_fielddef_name(f)))
return false;
}
// Constraint checks ok, perform the action.
for (int i = 0; i < n; i++) {
upb_fielddef *f = fields[i];
f->msgdef = m;
upb_inttable_insert(&m->itof, upb_fielddef_number(f), upb_value_ptr(f));
upb_strtable_insert(&m->ntof, upb_fielddef_name(f), upb_value_ptr(f));
upb_ref2(f, m);
upb_ref2(m, f);
if (ref_donor) upb_fielddef_unref(f, ref_donor);
}
return true;
}
bool upb_enumdef_addval(upb_enumdef *e, char *name, int32_t num) {
if (upb_enumdef_iton(e, num) || upb_enumdef_ntoi(e, name, NULL))
return false;
upb_iton_ent ent = {0, strdup(name)};
upb_strtable_insert(&e->ntoi, name, &num);
upb_inttable_insert(&e->iton, num, &ent);
return true;
}
static PyUpb_Context *get_or_create_context(struct upb_context *context)
{
PyUpb_Context *pycontext = NULL;
struct upb_string str = {.ptr = (char*)&context, .byte_len = sizeof(void*)};
struct contexttab_entry *e = upb_strtable_lookup(&contexts, &str);
if(!e) {
pycontext = (void*)PyUpb_ContextType.tp_alloc(&PyUpb_ContextType, 0);
pycontext->context = context;
struct contexttab_entry new_e = {
.e = {.key = {.ptr = (char*)&pycontext->context, .byte_len = sizeof(void*)}},
.context = pycontext
};
upb_strtable_insert(&contexts, &new_e.e);
} else {
bool upb_msgdef_addfields(upb_msgdef *m, upb_fielddef *const *fields, int n,
const void *ref_donor, upb_status *s) {
// TODO: extensions need to have a separate namespace, because proto2 allows a
// top-level extension (ie. one not in any package) to have the same name as a
// field from the message.
//
// This also implies that there needs to be a separate lookup-by-name method
// for extensions. It seems desirable for iteration to return both extensions
// and non-extensions though.
//
// We also need to validate that the field number is in an extension range iff
// it is an extension.
// Check constraints for all fields before performing any action.
for (int i = 0; i < n; i++) {
upb_fielddef *f = fields[i];
// TODO(haberman): handle the case where two fields of the input duplicate
// name or number.
if (upb_fielddef_containingtype(f) != NULL) {
upb_status_seterrmsg(s, "fielddef already belongs to a message");
return false;
} else if (upb_fielddef_name(f) == NULL || upb_fielddef_number(f) == 0) {
upb_status_seterrmsg(s, "field name or number were not set");
return false;
} else if(upb_msgdef_itof(m, upb_fielddef_number(f)) ||
upb_msgdef_ntof(m, upb_fielddef_name(f))) {
upb_status_seterrmsg(s, "duplicate field name or number");
return false;
}
}
// Constraint checks ok, perform the action.
for (int i = 0; i < n; i++) {
upb_fielddef *f = fields[i];
release_containingtype(f);
f->msg.def = m;
f->msg_is_symbolic = false;
upb_inttable_insert(&m->itof, upb_fielddef_number(f), upb_value_ptr(f));
upb_strtable_insert(&m->ntof, upb_fielddef_name(f), upb_value_ptr(f));
upb_ref2(f, m);
upb_ref2(m, f);
if (ref_donor) upb_fielddef_unref(f, ref_donor);
}
return true;
}
bool upb_enumdef_addval(upb_enumdef *e, const char *name, int32_t num,
upb_status *status) {
if (!upb_isident(name, strlen(name), false, status)) {
return false;
}
if (upb_enumdef_ntoi(e, name, NULL)) {
upb_status_seterrf(status, "name '%s' is already defined", name);
return false;
}
if (!upb_strtable_insert(&e->ntoi, name, upb_value_int32(num))) {
upb_status_seterrliteral(status, "out of memory");
return false;
}
if (!upb_inttable_lookup(&e->iton, num, NULL) &&
!upb_inttable_insert(&e->iton, num, upb_value_cstr(upb_strdup(name)))) {
upb_status_seterrliteral(status, "out of memory");
upb_strtable_remove(&e->ntoi, name, NULL);
return false;
}
return true;
}
bool upb_msgdef_addfields(upb_msgdef *m, upb_fielddef *const *fields, int n) {
// Check constraints for all fields before performing any action.
for (int i = 0; i < n; i++) {
upb_fielddef *f = fields[i];
assert(upb_atomic_read(&f->refcount) > 0);
if (f->name == NULL || f->number == 0 ||
upb_msgdef_itof(m, f->number) || upb_msgdef_ntof(m, f->name))
return false;
}
// Constraint checks ok, perform the action.
for (int i = 0; i < n; i++) {
upb_fielddef *f = fields[i];
upb_msgdef_ref(m);
assert(f->msgdef == NULL);
f->msgdef = m;
upb_itof_ent itof_ent = {0, f};
upb_inttable_insert(&m->itof, f->number, &itof_ent);
upb_strtable_insert(&m->ntof, f->name, &f);
}
return true;
}
bool upb_symtab_add(upb_symtab *s, upb_def **defs, int n, upb_status *status) {
upb_rwlock_wrlock(&s->lock);
// Add all defs to a table for resolution.
upb_strtable addtab;
upb_strtable_init(&addtab, n, sizeof(upb_symtab_ent));
for (int i = 0; i < n; i++) {
upb_def *def = defs[i];
if (upb_strtable_lookup(&addtab, def->fqname)) {
upb_status_seterrf(status, "Conflicting defs named '%s'", def->fqname);
upb_strtable_free(&addtab);
return false;
}
upb_strtable_insert(&addtab, def->fqname, &def);
}
// All existing defs that can reach defs that are being replaced must
// themselves be replaced with versions that will point to the new defs.
// Do a DFS -- any path that finds a new def must replace all ancestors.
upb_strtable *symtab = &s->symtab;
upb_strtable_iter i;
upb_strtable_begin(&i, symtab);
for(; !upb_strtable_done(&i); upb_strtable_next(&i)) {
upb_def *open_defs[UPB_MAX_TYPE_DEPTH];
const upb_symtab_ent *e = upb_strtable_iter_value(&i);
upb_symtab_dfs(e->def, open_defs, 0, &addtab);
}
// Resolve all refs.
upb_strtable_begin(&i, &addtab);
for(; !upb_strtable_done(&i); upb_strtable_next(&i)) {
const upb_symtab_ent *e = upb_strtable_iter_value(&i);
upb_msgdef *m = upb_dyncast_msgdef(e->def);
if(!m) continue;
// Type names are resolved relative to the message in which they appear.
const char *base = m->base.fqname;
upb_msg_iter j;
for(j = upb_msg_begin(m); !upb_msg_done(j); j = upb_msg_next(m, j)) {
upb_fielddef *f = upb_msg_iter_field(j);
if (f->type == 0) {
upb_status_seterrf(status, "Field type was not set.");
return false;
}
if (!upb_hassubdef(f)) continue; // No resolving necessary.
upb_downcast_unresolveddef(f->def); // Type check.
const char *name = f->def->fqname;
// Resolve from either the addtab (pending adds) or symtab (existing
// defs). If both exist, prefer the pending add, because it will be
// overwriting the existing def.
upb_symtab_ent *found;
if(!(found = upb_resolve(&addtab, base, name)) &&
!(found = upb_resolve(symtab, base, name))) {
upb_status_seterrf(status, "could not resolve symbol '%s' "
"in context '%s'", name, base);
return false;
}
// Check the type of the found def.
upb_fieldtype_t expected = upb_issubmsg(f) ? UPB_DEF_MSG : UPB_DEF_ENUM;
if(found->def->type != expected) {
upb_status_seterrliteral(status, "Unexpected type");
return false;
}
if (!upb_fielddef_resolve(f, found->def, status)) return false;
}
}
// The defs in the transaction have been vetted, and can be moved to the
// symtab without causing errors.
upb_strtable_begin(&i, &addtab);
for(; !upb_strtable_done(&i); upb_strtable_next(&i)) {
const upb_symtab_ent *tmptab_e = upb_strtable_iter_value(&i);
upb_def_movetosymtab(tmptab_e->def, s);
upb_symtab_ent *symtab_e =
upb_strtable_lookup(&s->symtab, tmptab_e->def->fqname);
if(symtab_e) {
upb_deflist_push(&s->olddefs, symtab_e->def);
symtab_e->def = tmptab_e->def;
} else {
upb_strtable_insert(&s->symtab, tmptab_e->def->fqname, tmptab_e);
}
}
upb_strtable_free(&addtab);
upb_rwlock_unlock(&s->lock);
upb_symtab_gc(s);
return true;
}
/* TODO(haberman): we need a lot more testing of error conditions.
* The came_from_user stuff in particular is not tested. */
bool upb_symtab_add(upb_symtab *s, upb_def *const*defs, int n, void *ref_donor,
upb_status *status) {
int i;
upb_strtable_iter iter;
upb_def **add_defs = NULL;
upb_strtable addtab;
upb_inttable seen;
assert(!upb_symtab_isfrozen(s));
if (!upb_strtable_init(&addtab, UPB_CTYPE_PTR)) {
upb_status_seterrmsg(status, "out of memory");
return false;
}
/* Add new defs to our "add" set. */
for (i = 0; i < n; i++) {
upb_def *def = defs[i];
const char *fullname;
upb_fielddef *f;
if (upb_def_isfrozen(def)) {
upb_status_seterrmsg(status, "added defs must be mutable");
goto err;
}
assert(!upb_def_isfrozen(def));
fullname = upb_def_fullname(def);
if (!fullname) {
upb_status_seterrmsg(
status, "Anonymous defs cannot be added to a symtab");
goto err;
}
f = upb_dyncast_fielddef_mutable(def);
if (f) {
if (!upb_fielddef_containingtypename(f)) {
upb_status_seterrmsg(status,
"Standalone fielddefs must have a containing type "
"(extendee) name set");
goto err;
}
} else {
if (upb_strtable_lookup(&addtab, fullname, NULL)) {
upb_status_seterrf(status, "Conflicting defs named '%s'", fullname);
goto err;
}
/* We need this to back out properly, because if there is a failure we
* need to donate the ref back to the caller. */
def->came_from_user = true;
upb_def_donateref(def, ref_donor, s);
if (!upb_strtable_insert(&addtab, fullname, upb_value_ptr(def)))
goto oom_err;
}
}
/* Add standalone fielddefs (ie. extensions) to the appropriate messages.
* If the appropriate message only exists in the existing symtab, duplicate
* it so we have a mutable copy we can add the fields to. */
for (i = 0; i < n; i++) {
upb_def *def = defs[i];
upb_fielddef *f = upb_dyncast_fielddef_mutable(def);
const char *msgname;
upb_value v;
upb_msgdef *m;
if (!f) continue;
msgname = upb_fielddef_containingtypename(f);
/* We validated this earlier in this function. */
assert(msgname);
/* If the extendee name is absolutely qualified, move past the initial ".".
* TODO(haberman): it is not obvious what it would mean if this was not
* absolutely qualified. */
if (msgname[0] == '.') {
msgname++;
}
if (upb_strtable_lookup(&addtab, msgname, &v)) {
/* Extendee is in the set of defs the user asked us to add. */
m = upb_value_getptr(v);
} else {
/* Need to find and dup the extendee from the existing symtab. */
const upb_msgdef *frozen_m = upb_symtab_lookupmsg(s, msgname);
if (!frozen_m) {
upb_status_seterrf(status,
"Tried to extend message %s that does not exist "
"in this SymbolTable.",
msgname);
goto err;
}
m = upb_msgdef_dup(frozen_m, s);
if (!m) goto oom_err;
if (!upb_strtable_insert(&addtab, msgname, upb_value_ptr(m))) {
upb_msgdef_unref(m, s);
goto oom_err;
}
}
if (!upb_msgdef_addfield(m, f, ref_donor, status)) {
goto err;
}
}
/* Add dups of any existing def that can reach a def with the same name as
* anything in our "add" set. */
if (!upb_inttable_init(&seen, UPB_CTYPE_BOOL)) goto oom_err;
upb_strtable_begin(&iter, &s->symtab);
for (; !upb_strtable_done(&iter); upb_strtable_next(&iter)) {
upb_def *def = upb_value_getptr(upb_strtable_iter_value(&iter));
upb_resolve_dfs(def, &addtab, s, &seen, status);
if (!upb_ok(status)) goto err;
}
upb_inttable_uninit(&seen);
/* Now using the table, resolve symbolic references for subdefs. */
upb_strtable_begin(&iter, &addtab);
for (; !upb_strtable_done(&iter); upb_strtable_next(&iter)) {
const char *base;
upb_def *def = upb_value_getptr(upb_strtable_iter_value(&iter));
upb_msgdef *m = upb_dyncast_msgdef_mutable(def);
upb_msg_field_iter j;
if (!m) continue;
/* Type names are resolved relative to the message in which they appear. */
base = upb_msgdef_fullname(m);
for(upb_msg_field_begin(&j, m);
!upb_msg_field_done(&j);
upb_msg_field_next(&j)) {
upb_fielddef *f = upb_msg_iter_field(&j);
const char *name = upb_fielddef_subdefname(f);
if (name && !upb_fielddef_subdef(f)) {
/* Try the lookup in the current set of to-be-added defs first. If not
* there, try existing defs. */
upb_def *subdef = upb_resolvename(&addtab, base, name);
if (subdef == NULL) {
subdef = upb_resolvename(&s->symtab, base, name);
}
if (subdef == NULL) {
upb_status_seterrf(
status, "couldn't resolve name '%s' in message '%s'", name, base);
goto err;
} else if (!upb_fielddef_setsubdef(f, subdef, status)) {
goto err;
}
}
}
}
/* We need an array of the defs in addtab, for passing to upb_def_freeze. */
add_defs = malloc(sizeof(void*) * upb_strtable_count(&addtab));
if (add_defs == NULL) goto oom_err;
upb_strtable_begin(&iter, &addtab);
for (n = 0; !upb_strtable_done(&iter); upb_strtable_next(&iter)) {
add_defs[n++] = upb_value_getptr(upb_strtable_iter_value(&iter));
}
if (!upb_def_freeze(add_defs, n, status)) goto err;
/* This must be delayed until all errors have been detected, since error
* recovery code uses this table to cleanup defs. */
upb_strtable_uninit(&addtab);
/* TODO(haberman) we don't properly handle errors after this point (like
* OOM in upb_strtable_insert() below). */
for (i = 0; i < n; i++) {
upb_def *def = add_defs[i];
const char *name = upb_def_fullname(def);
upb_value v;
bool success;
if (upb_strtable_remove(&s->symtab, name, &v)) {
const upb_def *def = upb_value_getptr(v);
upb_def_unref(def, s);
}
success = upb_strtable_insert(&s->symtab, name, upb_value_ptr(def));
UPB_ASSERT_VAR(success, success == true);
}
free(add_defs);
return true;
oom_err:
upb_status_seterrmsg(status, "out of memory");
err: {
/* For defs the user passed in, we need to donate the refs back. For defs
* we dup'd, we need to just unref them. */
upb_strtable_begin(&iter, &addtab);
for (; !upb_strtable_done(&iter); upb_strtable_next(&iter)) {
upb_def *def = upb_value_getptr(upb_strtable_iter_value(&iter));
bool came_from_user = def->came_from_user;
def->came_from_user = false;
if (came_from_user) {
upb_def_donateref(def, s, ref_donor);
} else {
upb_def_unref(def, s);
}
}
}
upb_strtable_uninit(&addtab);
free(add_defs);
assert(!upb_ok(status));
return false;
}
/* Starts a depth-first traversal at "def", recursing into any subdefs
* (ie. submessage types). Adds duplicates of existing defs to addtab
* wherever necessary, so that the resulting symtab will be consistent once
* addtab is added.
*
* More specifically, if any def D is found in the DFS that:
*
* 1. can reach a def that is being replaced by something in addtab, AND
*
* 2. is not itself being replaced already (ie. this name doesn't already
* exist in addtab)
*
* ...then a duplicate (new copy) of D will be added to addtab.
*
* Returns true if this happened for any def reachable from "def."
*
* It is slightly tricky to do this correctly in the presence of cycles. If we
* detect that our DFS has hit a cycle, we might not yet know if any SCCs on
* our stack can reach a def in addtab or not. Once we figure this out, that
* answer needs to apply to *all* defs in these SCCs, even if we visited them
* already. So a straight up one-pass cycle-detecting DFS won't work.
*
* To work around this problem, we traverse each SCC (which we already
* computed, since these defs are frozen) as a single node. We first compute
* whether the SCC as a whole can reach any def in addtab, then we dup (or not)
* the entire SCC. This requires breaking the encapsulation of upb_refcounted,
* since that is where we get the data about what SCC we are in. */
static bool upb_resolve_dfs(const upb_def *def, upb_strtable *addtab,
const void *new_owner, upb_inttable *seen,
upb_status *s) {
upb_value v;
bool need_dup;
const upb_def *base;
const void* memoize_key;
/* Memoize results of this function for efficiency (since we're traversing a
* DAG this is not needed to limit the depth of the search).
*
* We memoize by SCC instead of by individual def. */
memoize_key = def->base.group;
if (upb_inttable_lookupptr(seen, memoize_key, &v))
return upb_value_getbool(v);
/* Visit submessages for all messages in the SCC. */
need_dup = false;
base = def;
do {
upb_value v;
const upb_msgdef *m;
assert(upb_def_isfrozen(def));
if (def->type == UPB_DEF_FIELD) continue;
if (upb_strtable_lookup(addtab, upb_def_fullname(def), &v)) {
need_dup = true;
}
/* For messages, continue the recursion by visiting all subdefs, but only
* ones in different SCCs. */
m = upb_dyncast_msgdef(def);
if (m) {
upb_msg_field_iter i;
for(upb_msg_field_begin(&i, m);
!upb_msg_field_done(&i);
upb_msg_field_next(&i)) {
upb_fielddef *f = upb_msg_iter_field(&i);
const upb_def *subdef;
if (!upb_fielddef_hassubdef(f)) continue;
subdef = upb_fielddef_subdef(f);
/* Skip subdefs in this SCC. */
if (def->base.group == subdef->base.group) continue;
/* |= to avoid short-circuit; we need its side-effects. */
need_dup |= upb_resolve_dfs(subdef, addtab, new_owner, seen, s);
if (!upb_ok(s)) return false;
}
}
} while ((def = (upb_def*)def->base.next) != base);
if (need_dup) {
/* Dup all defs in this SCC that don't already have entries in addtab. */
def = base;
do {
const char *name;
if (def->type == UPB_DEF_FIELD) continue;
name = upb_def_fullname(def);
if (!upb_strtable_lookup(addtab, name, NULL)) {
upb_def *newdef = upb_def_dup(def, new_owner);
if (!newdef) goto oom;
newdef->came_from_user = false;
if (!upb_strtable_insert(addtab, name, upb_value_ptr(newdef)))
goto oom;
}
} while ((def = (upb_def*)def->base.next) != base);
}
upb_inttable_insertptr(seen, memoize_key, upb_value_bool(need_dup));
return need_dup;
oom:
upb_status_seterrmsg(s, "out of memory");
return false;
}
