JetCore/__mysql.cpp

#include "__mysql.h"
#include "Exception.h"
#include <iostream>
#include <string>
#include <unordered_map>

namespace jet {

   namespace {
      struct PooledConnection {
	 MYSQL *handle;
	 std::string host, user, database;
      };
      // Persistent, per-worker-thread connection cache - reused across
      // requests instead of a fresh mysql_real_connect() every time.
      // Safe with no locking: the apiserver's EPoll worker pool is a
      // fixed set of OS threads for the process's whole life (see
      // MysqlThreadScope's own doc comment - mysql_thread_init() runs
      // once per worker thread at thread start), so a thread_local
      // instance here means a given MYSQL* handle is only ever touched
      // by the one thread that created it, never concurrently accessed.
      // Keyed by sessionid, which is how this codebase's own <database>/
      // <sql>/<ifrow> tags already address a specific connection
      // (Global::sessions is keyed the same way) - the cached host/user/
      // database are kept alongside purely as a defensive check that a
      // reused entry really matches what THIS invocation asked for,
      // not a real expectation that sessionid is ever reused across
      // genuinely different backing databases in practice.
      thread_local std::unordered_map<std::string, PooledConnection> connectionPool;
   }

   __mysql::__mysql(coreutils::ZString &in, coreutils::MString &parentOut, Global &global, Tag *parent, Tag *local) : Tag(in, parentOut, global, parent, this) {
      
      if(!keywordDefined("host"))
	throw coreutils::Exception("host must be specified for mysql tag.");
      // jet-1.4's <database> spells the same attribute "name" instead of
      // "database" (e.g. SESSIONCTL.tag's own <sessionctl> definition) -
      // accept either rather than requiring that real content to change.
      coreutils::ZString databaseKeyword = keywordDefined("database") ? "database" : "name";
      if(!keywordDefined(databaseKeyword))
	throw coreutils::Exception("database must be specified for mysql tag.");
      if(!keywordDefined("user"))
	throw coreutils::Exception("user must be specified for mysql tag.");
      if(!keywordDefined("password"))
	throw coreutils::Exception("password must be specified for mysql tag.");
            
      // Same double-indirection bug as the host/user/password/database
      // fix above - resolveKeyword("sessionid") already returns the
      // resolved VALUE (e.g. "_SESSIONCTL"), so wrapping it in keywords[...]
      // used that value as a map KEY instead, auto-vivifying an empty
      // string (no attribute is ever literally named "_SESSIONCTL").
      // sessionId was therefore always empty regardless of what
      // sessionid= actually specified - confirmed live via gdb: two
      // real <mysql>/<database> tags with genuinely different sessionid
      // values ("" and "_SESSIONCTL") both registered under "", the
      // second colliding with "sessionid already exists.".
      sessionId = resolveKeyword("sessionid");

      // mysql_real_connect() returns NULL on failure but does NOT itself
      // free the handle mysql_init() allocated - the caller is still
      // expected to mysql_close() that original handle to avoid leaking
      // it (MySQL C API docs). The previous code overwrote `mysql` with
      // the connect() return value directly, so on failure the only
      // pointer to the real handle was lost before ever being freed - a
      // real leak on every failed connection attempt (this member's
      // eventual value, NULL, is what a normal ~__mysql() would try to
      // mysql_close() anyway, which is a safe no-op, but that never even
      // runs here since the constructor throws below - see the
      // addSession() comment for why that matters more).
      // resolveKeyword() already returns the fully-resolved VALUE - the
      // previous keywords[resolveKeyword("host")] wrapped that resolved
      // VALUE ("localhost") back into a keywords[] lookup, i.e. used it
      // as a MAP KEY instead of the value itself. Since no attribute is
      // ever literally named "localhost"/the real username/etc.,
      // std::map::operator[] silently auto-vivified an empty MString for
      // every one of the four parameters - mysql_real_connect() was
      // therefore always called with host/user/password/database all
      // empty, regardless of what the tag actually specified. Confirmed
      // live: MySQL's own error ("Access denied for user 'barant'@'localhost'
      // (using password: NO)") showed the password was genuinely empty,
      // not merely wrong.
      coreutils::MString hostVal = resolveKeyword("host");
      coreutils::MString userVal = resolveKeyword("user");
      coreutils::MString passwordVal = resolveKeyword("password");
      coreutils::MString databaseVal = resolveKeyword(databaseKeyword);

      // Reuse a live connection this same worker thread already holds
      // for this sessionid, rather than paying mysql_real_connect()'s
      // full handshake cost on every single request. mysql_ping() is
      // what makes this safe - it's a genuine round trip to the server,
      // so a connection MySQL itself already dropped (its own
      // wait_timeout, or an admin-initiated restart like this project
      // already lived through once) is detected and discarded here
      // rather than handed to processContainer() to fail on later.
      std::string poolKey = sessionId.str();
      auto pooled = connectionPool.find(poolKey);
      if(pooled != connectionPool.end() &&
	 pooled->second.host == hostVal.str() &&
	 pooled->second.user == userVal.str() &&
	 pooled->second.database == databaseVal.str()) {
	 if(mysql_ping(pooled->second.handle) == 0) {
	    mysql = pooled->second.handle;
	 } else {
	    mysql_close(pooled->second.handle);
	    connectionPool.erase(pooled);
	 }
      }

      if(!mysql) {
	 MYSQL *handle = mysql_init(NULL);
	 mysql = mysql_real_connect(handle,
				    hostVal.c_str(),
				    userVal.c_str(),
				    passwordVal.c_str(),
				    databaseVal.c_str(), 0, NULL, 0);

	 if(!mysql) {
	    std::string err = mysql_error(handle);
	    mysql_close(handle);
	    throw coreutils::Exception("database connection failed: " + err);
	 }

	 connectionPool[poolKey] = PooledConnection{mysql, hostVal.str(), userVal.str(), databaseVal.str()};
      }

      // Registered only once the connection is confirmed live - a real,
      // reproducible crash (heap corruption - "free(): invalid pointer"
      // seen live, then two independent server-hang repros) traced back
      // to this: addSession() used to run BEFORE the connect attempt, so
      // a FAILED connection (which throws, right above) left a dangling
      // `this` pointer registered under sessionId in Global's session
      // map forever - ~__mysql() (which would normally call
      // removeSession()) never runs for a constructor that itself threw,
      // per ordinary C++ rules (only already-completed base-class
      // subobjects get destroyed during exception unwinding, not the
      // derived object whose constructor is still executing). The very
      // next <mysql> tag construction using the same sessionId (e.g. two
      // product pages with an empty sessionid="" back to back, both
      // failing the same way) then collided with that stale entry.
      // Confirmed via a real, deterministic two-request repro before this
      // fix, and confirmed gone after it.
      global.addSession(sessionId, this);

      processContainer(container);
      
   }

   __mysql::~__mysql() {
      global.removeSession(sessionId);
      mysql_free_result(result);
      // mysql itself is deliberately NOT closed here anymore - see the
      // connectionPool doc comment above the constructor. It's owned by
      // this worker thread's persistent pool now and gets reused by
      // whatever request this thread handles next, instead of being
      // torn down and fully reconnected from scratch every time.
   }

   void __mysql::query(coreutils::MString query) {
      int rc = mysql_real_query(mysql, query.getData(), query.getLength());
      result = mysql_store_result(mysql);
      if(result) {
	 row = mysql_fetch_row(result);
	 fieldLength = mysql_fetch_lengths(result);
	 qFields = mysql_num_fields(result);
	 // Built once per query, not once per column LOOKUP - see the
	 // matching comment on getColumnValue() and on fieldIndex's own
	 // declaration in __mysql.h. mysql_fetch_fields() returns the
	 // whole field array in one call (mysql_fetch_field_direct()
	 // looping ix=0..qFields-1, as getColumnValue() used to do on
	 // every single named access, does the exact same work over and
	 // over for a result set whose schema never changes after this
	 // point).
	 fieldIndex.clear();
	 MYSQL_FIELD *fields = mysql_fetch_fields(result);
	 for(unsigned int ix = 0; ix < qFields; ++ix)
	   fieldIndex[std::string(fields[ix].name)] = (int)ix;
      } else {
	 // A failed/errored query (real content builds one from an unset
	 // $[...] parameter, e.g. "where tasks.id=" with no value - genuine
	 // MySQL syntax error) previously left row/qFields holding whatever
	 // a PRIOR successful query on this same session object had set,
	 // since only the success branch above ever touched them. A later
	 // getColumnValue() call then looped 0..qFields (stale, nonzero)
	 // calling mysql_fetch_field_direct() against this NULL `result` -
	 // segfaulting (confirmed via gdb: SIGSEGV in
	 // mysql_fetch_field_direct, called from getColumnValue) instead of
	 // reaching the "column does not exist in session result." the
	 // caller already throws/catches for a genuinely missing column.
	 // Resetting both here makes a failed query present as "no row, no
	 // columns" - clean and consistent regardless of what ran before it
	 // on this session.
	 row = NULL;
	 qFields = 0;
	 fieldIndex.clear();
      }
   }

   void __mysql::nextRow() {
      row = mysql_fetch_row(result);
      fieldLength = mysql_fetch_lengths(result);
   }

   bool __mysql::hasRow() {
      return row != NULL;
   }
   
   coreutils::ZString __mysql::getColumnValue(coreutils::ZString column) {
      MYSQL_FIELD *field;      
      if(column == "?#") {
	 nbrOfColumns = (int)qFields;
	 return nbrOfColumns;
      } else if(column.ifNext("#") || column.ifNext("?")) {
	 // "#N"/"?N" (column name at 1-based position N) are the same
	 // jet-1.4 syntax under two different prefixes - real, unmodified
	 // 1.4 content uses both (barant.tag's own <showtable> definition
	 // writes $[?$[index]], never $[#$[index]]) - confirmed via a real
	 // repro: admin.page's own Users/Products tables rendered every
	 // column header and cell completely blank, since "?N" fell all
	 // the way through to the ordinary column-VALUE lookup below
	 // (nothing named "?3" exists in the row) and read as empty rather
	 // than as "the name of column 3", per the jet-1.4-fallback #2
	 // empty-string-on-missing-column behavior elsewhere in this
	 // codebase (see Tag.cpp).
	 if(column.eod()) {
	    nbrOfRows = (int)mysql_num_rows(result);
	    return nbrOfRows;
	 } else {
	    int index = column.asInteger();
	    field = mysql_fetch_field_direct(result, index - 1);
	    return coreutils::ZString(field->name);
	 }
      }

      // A query that ran successfully but matched zero rows (a real,
      // valid query - e.g. cart.page's own aggregate/JOIN query against
      // an empty cart session - not a failed one, see query()'s own
      // row=NULL/qFields=0 reset above for that separate case) leaves
      // qFields correctly reflecting the real schema but `row` NULL.
      // The loop below finds a genuine matching column name and then
      // dereferences row[ix] - segfaulting (confirmed via gdb: SIGSEGV
      // right here) instead of reaching a catchable error. Reusing the
      // same "column does not exist" exception the loop's own fallthrough
      // already throws - no current row means no column value is
      // obtainable either way, and callers already expect and catch
      // this exact exception.
      if(!row)
	throw coreutils::Exception("column does not exist in session result.");

      // O(1) against the map query() already built for this result set,
      // instead of re-deriving field metadata (mysql_fetch_field_direct,
      // a real MySQL C API call) and string-comparing column names one
      // by one on every single named access - see fieldIndex's own doc
      // comment in __mysql.h for the full story.
      auto it = fieldIndex.find(column.str());
      if(it != fieldIndex.end())
	return coreutils::ZString(row[it->second], fieldLength[it->second]);
      throw coreutils::Exception("column does not exist in session result.");
   }
   
}