C++ Core Guidelines: The Resolution of the Riddle

Contents[Show]

Today, I solve the riddle from my last post. Thanks to my readers, the analysis of the ABA problem is quite accurate.

 

park 748339 1280

Only to remind you. The rule CP.100 from the C++ core guidelines is the starting point of the riddle.

CP.100: Don’t use lock-free programming unless you absolutely have to.

The challenge in the rule states that the following code snippet has a bug. The bug should be due to the ABA problem. The post ABA - A is not the same as A gives a concise introduction into the ABA problem.

extern atomic<Link*> head;        // the shared head of a linked list

Link* nh = new Link(data, nullptr);    // make a link ready for insertion
Link* h = head.load();                 // read the shared head of the list   

do {
    if (h->data <= data) break;        // if so, insert elsewhere           
    nh->next = h;                      // next element is the previous head  
} while (!head.compare_exchange_weak(h, nh));    // write nh to head or to h 

Thanks a lot in particular to anonymous readers of my German blog, here is a runnable piece of code and a deep analysis of the issue.

#include <atomic>

class Link {
public:
    Link(int d, Link* p) : data(d), next(p) {}
    int     data;
    Link*   next;
};

void foo (int data) {
    extern std::atomic<Link*> head;

    Link* nh = new Link(data, nullptr);            // (1)
    Link* h  = head.load();                        // (2)

    do {
        if (h->data <= data) break;                // (3)
        nh->next = h;                              // (4)
    } while (!head.compare_exchange_weak(h, nh));  // (5)
}

 

First of all, what should this piece of code do? It creates a singly linked list of nodes (Link). Each node has a pointer and a data field. The pointer points to the next element (node->next), and the data field stores the value: node->data. Each new node is inserted into the singly linked list in such a way that data is ordered in ascending order.

To insert a new node into the correct position in the singly linked list, the following steps are performed.

  • Line 1: A new node is created. This is fine because the node is locally created in each thread.
  • Line 2: The pointer to the head is read. The read operation is atomic; therefore,  considered in isolation the operation is also fine. What does in isolation mean? Line 2 creates with line 5 a kind of transaction. Line 2 stores the initial state of the transaction and line 5 publish the transaction if nothing had changed in between.
  • Line 3: Correspondingly to the previous lines, this line 3 has no issue. Only a value comparison takes place which may end the function if the data of head is smaller than the new data.
  • Line 4: nh is local data; therefore, the assignment of nh->next is fine. It may happen head h was changed in the meantime and, consequently, nh->next does not refer to the head afterwards. This is only an issue if the change is committed in the next line 5.
  • Line 5: The instruction head.compare_exchange_weak(h, nh) compares head with the stored h in line 2 and exchanges h and nh in an atomic step as soon as they are the same. If head is not equal to h, h is set to head. Line 5 is the end of the atomic transaction and publishes the updated singly linked list.

What is the issue with this few lines of code? The entire transaction is based on the pointer comparison in line 5. If the pointer comparison can be fooled, the singly linked list can be broken.

There is a time window between the loading of the head (line 2) and the check if the current head is the old head (line 5). The means another thread may kick in and change in the meantime head but the first thread is not aware of it.

Let me present a buggy sequence of events.

Breaking of the Invariant

The invariant of the following singly linked list is that data is ordered in an ascending order. The blue node is the head of the list.

This is the initial structure of the list. The head has the address 0x0815.

ABAThread1

 

Thread 1

Head42

  • Wants to add the new node with data 42.
  • 42 < 47, therefore the new node should become the new head.
  • Right before the line (5), Thread 2 kicks in.

Thread 2

  • Removes the current head 47.
  • Makes the node with data 60 to the new head.

ABAThread2 60head

  • Wants to add the new node with data 30.

Head30

  • Makes 30 the new head with address 0x0815; this was the former address of 47 and will often happen because of memory reuse.

ABAThread2 30

 

 Thread 1

  • Makes the node with the data 42 to the new head; this is fine because the comparison in line 5 just compares the old with the new node and they have the same address: 0x0815.

ABAThread1 42

 Now, the singly linked list is broken because the values of the nodes are not ordered in an ascending order.

What's next?

I'm nearly done with the rules to concurrency and lock-free programming in particular. The remaining rules are about wrong assumptions to hardware/compiler combinations and the infamous double-checked locking pattern. Read about it in the next post.

 

 

Thanks a lot to my Patreon Supporters: Eric Pederson, Paul Baxter,  Sai Raghavendra Prasad Poosa, Meeting C++, Matt Braun, Avi Lachmish, Adrian Muntea, and Roman Postanciuc.

 

 

Get your e-book at leanpub:

The C++ Standard Library

 

Concurrency With Modern C++

 

Get Both as one Bundle

cover   ConcurrencyCoverFrame   bundle
With C++11, C++14, and C++17 we got a lot of new C++ libraries. In addition, the existing ones are greatly improved. The key idea of my book is to give you the necessary information to the current C++ libraries in about 200 pages.  

C++11 is the first C++ standard that deals with concurrency. The story goes on with C++17 and will continue with C++20.

I'll give you a detailed insight in the current and the upcoming concurrency in C++. This insight includes the theory and a lot of practice with more the 100 source files.

 

Get my books "The C++ Standard Library" (including C++17) and "Concurrency with Modern C++" in a bundle.

In sum, you get more than 600 pages full of modern C++ and more than 100 source files presenting concurrency in practice.

 

Add comment


Subscribe to the newsletter (+ pdf bundle)

Blog archive

Source Code

Visitors

Today 1726

All 1174178

Currently are 200 guests and no members online

Kubik-Rubik Joomla! Extensions

Latest comments