Skip to content

Latest commit

 

History

History
370 lines (303 loc) · 12.9 KB

COMMON_ERRORS.md

File metadata and controls

370 lines (303 loc) · 12.9 KB

Commonly recurring errors in bmcweb

What follows is a list of common errors that new users to bmcweb tend to make when operating within its bounds for the first time. If this is your first time developing in bmcweb, the maintainers highly recommend reading and understanding all of common traps before continuing with any development. Every single one of the examples below compile without warnings, but are incorrect in not-always-obvious ways, or impose a pattern that tends to cause hard to find bugs, or bugs that appear later. Every one has been submitted to code review multiple times.

1. Directly dereferencing a pointer without checking for validity first

int myBadMethod(const nlohmann::json& j){
    const int* myPtr = j.get_if<int>();
    return *myPtr;
}

This pointer is not guaranteed to be filled, and could be a null dereference.

2. String views aren't null terminated

int getIntFromString(std::string_view s){
    return std::atoi(s.data());
}

This will give the right answer much of the time, but has the possibility to fail when string_view is not null terminated. Use from_chars instead, which takes both a pointer and a length

3. Not handling input errors

int getIntFromString(const std::string& s){
    return std::atoi(s.c_str());
}

In the case where the string is not representable as an int, this will trigger undefined behavior at system level. Code needs to check for validity of the string, ideally with something like from_chars, and return the appropriate error code.

4. Walking off the end of a string

std::string getFilenameFromPath(const std::string& path){
    size_t index = path.find("/");
    if (index != std::string::npos){
        // If the string ends with "/", this will walk off the end of the string.
        return path.substr(pos + 1);
    }
    return "";
}

5. Using methods that throw (or not handling bad inputs)

int myBadMethod(nlohmann::json& j){
    return j.get<int>();
}

This method throws, and bad inputs will not be handled

Commonly used methods that fall into this pattern:

  • std::variant::get
  • std::vector::at
  • std::map::at
  • std::set::at
  • std::<generic container type>::at
  • nlohmann::json::operator!=
  • nlohmann::json::operator+=
  • nlohmann::json::at
  • nlohmann::json::get
  • nlohmann::json::get_ref
  • nlohmann::json::get_to
  • nlohmann::json::items
  • nlohmann::json::operator<<
  • nlohmann::json::operator>>
  • std::filesystem::create_directory
  • std::filesystem::rename
  • std::filesystem::file_size
  • std::stoi
  • std::stol
  • std::stoll

Special note: JSON

nlohmann::json::parse by default throws on failure, but also accepts an optional argument that causes it to not throw: set the 3rd argument to false.

nlohmann::json::dump by default throws on failure, but also accepts an optional argument that causes it to not throw: set the 4th argument to replace. Although ignore preserves content 1:1, replace is preferred from a security point of view.

Special note: Boost

there is a whole class of boost asio functions that provide both a method that throws on failure, and a method that accepts and returns an error code. This is not a complete list, but users should verify in the boost docs when calling into asio methods, and prefer the one that returns an error code instead of throwing.

  • boost::asio::ip::tcp::acceptor::bind();
  • boost::asio::ip::tcp::acceptor::cancel();
  • boost::asio::ip::tcp::acceptor::close();
  • boost::asio::ip::tcp::acceptor::listen();
  • boost::asio::ip::address::make_address();

6. Blocking functions

bmcweb uses a single reactor for all operations. Blocking that reactor for any amount of time causes all other operations to stop. The common blocking functions that tend to be called incorrectly are:

  • sleep()
  • boost::asio::ip::tcp::socket::read()
  • boost::asio::ip::tcp::socket::read_some()
  • boost::asio::ip::tcp::socket::write()
  • boost::asio::ip::tcp::socket::write_some()
  • boost::asio::ip::tcp::socket::connect()
  • boost::asio::ip::tcp::socket::send()
  • boost::asio::ip::tcp::socket::wait()
  • boost::asio::steady_timer::wait()

Note: an exception is made for filesystem/disk IO read and write. This is mostly due to not having great abstractions for it that mate well with the async system, the fact that most filesystem accesses are into tmpfs (and therefore should be "fast" most of the time) and in general how little the filesystem is used in practice.

7. Lack of locking between subsequent calls

While global data structures are discouraged, they are sometimes required to store temporary state for operations that require it. Given the single threaded nature of bmcweb, they are not required to be explicitly threadsafe, but they must be always left in a valid state, and checked for other uses before occupying.

std::optional<std::string> currentOperation;
void firstCallbackInFlow(){
    currentOperation = "Foo";
}
void secondCallbackInFlow(){
    currentOperation.reset();
}

In the above case, the first callback needs a check to ensure that currentOperation is not already being used.

8. Wildcard reference captures in lambdas

std::string x; auto mylambda = [&](){
    x = "foo";
}
do_async_read(mylambda)

Numerous times, lifetime issues of const references have been injected into async bmcweb code. While capturing by reference can be useful, given how difficult these types of bugs are to triage, bmcweb explicitly requires that all code captures variables by name explicitly, and calls out each variable being captured by value or by reference. The above prototypes would change to [&x]()... Which makes clear that x is captured, and its lifetime needs tracked.

9. URLs should end in "/"

BMCWEB("/foo/bar");

Unless you explicitly have a reason not to (as there is one known exception where the behavior must differ) all URL handlers should end in "/". The bmcweb route handler will detect routes ending in slash and generate routes for both the route ending in slash and the one without. This allows both URLs to be used by users. While many specifications do not require this, it resolves a whole class of bug that we've seen in the past.

10. URLs constructed in aggregate

std::string routeStart = "/redfish/v1";

BMCWEB_ROUTE(routestart + "/SessionService/")

Very commonly, bmcweb maintainers and contributors alike have to do audits of all routes that are available, to verify things like security and documentation accuracy. While these processes are largely manual, they can mostly be conducted by a simple grep statement to search for urls in question. Doing the above makes the route handlers no longer greppable, and complicates bmcweb patchsets as a whole.

11. Not responding to 404

BMCWEB_ROUTE("/myendpoint/<str>",
    [](Request& req, Response& res, const std::string& id){
     crow::connections::systemBus->async_method_call(
          [asyncResp](const boost::system::error_code& ec,
                      const std::string& myProperty) {
              if (ec)
              {
                  messages::internalError(asyncResp->res);
                  return;
              }
              ... handle code
          },
          "xyz.openbmc_project.Logging",
          "/xyz/openbmc_project/mypath/" + id,
          "xyz.MyInterface", "GetAll", "");
});

All bmcweb routes should handle 404 (not found) properly, and return it where appropriate. 500 internal error is not a substitute for this, and should be only used if there isn't a more appropriate error code that can be returned. This is important, because a number of vulnerability scanners attempt injection attacks in the form of /myendpoint/foobar, or /myendpoint/#$*(%)&#%$)(\*& in an attempt to circumvent security. If the server returns 500 to any of these requests, the security scanner logs it as an error for followup. While in general these errors are benign, and not actually a real security threat, having a clean security run allows maintainers to minimize the amount of time spent triaging issues reported from these scanning tools.

An implementation of the above that handles 404 would look like:

BMCWEB_ROUTE("/myendpoint/<str>",
    [](Request& req, Response& res, const std::string& id){
     crow::connections::systemBus->async_method_call(
          [asyncResp](const boost::system::error_code& ec,
                      const std::string& myProperty) {
              if (ec == <error code that gets returned by not found>){
                  messages::resourceNotFound(res);
                  return;
              }
              if (ec)
              {
                  messages::internalError(asyncResp->res);
                  return;
              }
              ... handle code
          },
          "xyz.openbmc_project.Logging",
          "/xyz/openbmc_project/mypath/" + id,
          "xyz.MyInterface", "GetAll", "");
});

Note: A more general form of this rule is that no handler should ever return 500 on a working system, and any cases where 500 is found, can immediately be assumed to be a bug in either the system, or bmcweb.

12. Imprecise matching

void isInventoryPath(const std::string& path){
    if (path.find("inventory")){
        return true;
    }
    return false;
}

When matching dbus paths, HTTP fields, interface names, care should be taken to avoid doing direct string containment matching. Doing so can lead to errors where fan1 and fan11 both report to the same object, and cause behavior breaks in subtle ways.

When using dbus paths, rely on the methods on sdbusplus::message::object_path. When parsing HTTP field and lists, use the RFC7230 implementations from boost::beast.

Other commonly misused methods are: boost::iequals. Unless the standard you're implementing (as is the case in some HTTP fields) requires case insensitive comparisons, casing should be obeyed, especially when relying on user-driven data.

  • boost::starts_with
  • boost::ends_with
  • std::string::starts_with
  • std::string::ends_with
  • std::string::rfind

The above methods tend to be misused to accept user data and parse various fields from it. In practice, there tends to be better, purpose built methods for removing just the field you need.

13. Complete replacement of the response object

void getMembers(crow::Response& res){
  res.jsonValue = {{"Value", 2}};
}

In many cases, bmcweb is doing multiple async actions in parallel, and orthogonal keys within the Response object might be filled in from another task. Completely replacing the json object can lead to convoluted situations where the output of the response is dependent on the order of the asynchronous actions completing, which cannot be guaranteed, and has many time caused bugs.

void getMembers(crow::Response& res){
  res.jsonValue["Value"] = 2;
}

As an added benefit, this code is also more efficient, as it avoids the intermediate object construction and the move, and as a result, produces smaller binaries.

Note, another form of this error involves calling nlohmann::json::reset(), to clear an object that's already been filled in. This has the potential to clear correct data that was already filled in from other sources.

14. Very long lambda callbacks

dbus::utility::getSubTree("/", interfaces,
                         [asyncResp](boost::system::error_code& ec,
                                     MapperGetSubTreeResult& res){
                            <many lines of code>
                         })

Inline lambdas, while useful in some contexts, are difficult to read, and have inconsistent formatting with tools like clang-format, which causes significant problems in review, and in applying patchsets that might have minor conflicts. In addition, because they are declared in a function scope, they are difficult to unit test, and produce log messages that are difficult to read given their unnamed nature.

Prefer to either use std::bind_front, and a normal method to handle the return, or a lambda that is less than 10 lines of code to handle an error inline. In cases where std::bind_front cannot be used, such as in sdbusplus::asio::connection::async_method_call, keep the lambda length less than 10 lines, and call the appropriate function for handling non-trivial transforms.

void afterGetSubTree(std::shared_ptr<bmcweb::AsyncResp>& asyncResp,
                     boost::system::error_code& ec,
                     MapperGetSubTreeResult& res){
   <many lines of code>
}

dbus::utility::getSubTree("/xyz/openbmc_project/inventory", interfaces,
                         std::bind_front(afterGetSubTree, asyncResp));

See also the Cpp Core Guidelines for generalized guidelines on when lambdas are appropriate. The above recommendation is aligned with the C++ Core Guidelines.