Mutexes are used to prevent multiple threads from causing a data race by accessing shared resources at the same time. Sometimes, when locking mutexes, multiple threads hold each other's lock, and the program consequently deadlocks. Four conditions are required for deadlock to occur:

  • Mutual exclusion
  • Hold and wait
  • No preemption
  • Circular wait

Deadlock needs all four conditions, so preventing deadlock requires preventing any one of the four conditions. One simple solution is to lock the mutexes in a predefined order, which prevents circular wait.

Noncompliant Code Example

The behavior of this noncompliant code example depends on the runtime environment and the platform's scheduler. The program is susceptible to deadlock if thread thr1 attempts to lock ba2's mutex at the same time thread thr2 attempts to lock ba1's mutex in the deposit() function.

#include <stdlib.h>
#include <threads.h>
 
typedef struct {
  int balance;
  mtx_t balance_mutex;
} bank_account;

typedef struct {
  bank_account *from;
  bank_account *to;
  int amount;
} transaction;

void create_bank_account(bank_account **ba,
                         int initial_amount) {
  bank_account *nba = (bank_account *)malloc(
    sizeof(bank_account)
  );
  if (nba == NULL) {
    /* Handle error */
  }

  nba->balance = initial_amount;
  if (thrd_success
      != mtx_init(&nba->balance_mutex, mtx_plain)) {
    /* Handle error */
  }

  *ba = nba;
}

int deposit(void *ptr) {
  transaction *args = (transaction *)ptr;

  if (thrd_success != mtx_lock(&args->from->balance_mutex)) {
    /* Handle error */
  }

  /* Not enough balance to transfer */
  if (args->from->balance < args->amount) {
    if (thrd_success
        != mtx_unlock(&args->from->balance_mutex)) {
      /* Handle error */
    }
    return -1; /* Indicate error */
  }
  if (thrd_success != mtx_lock(&args->to->balance_mutex)) {
    /* Handle error */
  }

  args->from->balance -= args->amount;
  args->to->balance += args->amount;

  if (thrd_success
      != mtx_unlock(&args->from->balance_mutex)) {
    /* Handle error */
  }

  if (thrd_success
      != mtx_unlock(&args->to->balance_mutex)) {
    /* Handle error */
  }

  free(ptr);
  return 0;
}

int main(void) {
  thrd_t thr1, thr2;
  transaction *arg1;
  transaction *arg2;
  bank_account *ba1;
  bank_account *ba2;

  create_bank_account(&ba1, 1000);
  create_bank_account(&ba2, 1000);

  arg1 = (transaction *)malloc(sizeof(transaction));
  if (arg1 == NULL) {
    /* Handle error */
  }
  arg2 = (transaction *)malloc(sizeof(transaction));
  if (arg2 == NULL) {
    /* Handle error */
  }
  arg1->from = ba1;
  arg1->to = ba2;
  arg1->amount = 100;

  arg2->from = ba2;
  arg2->to = ba1;
  arg2->amount = 100;

  /* Perform the deposits */
  if (thrd_success
     != thrd_create(&thr1, deposit, (void *)arg1)) {
    /* Handle error */
  }
  if (thrd_success
      != thrd_create(&thr2, deposit, (void *)arg2)) {
    /* Handle error */
  }
  return 0;
} 

Compliant Solution

This compliant solution eliminates the circular wait condition by establishing a predefined order for locking in the deposit() function. Each thread will lock on the basis of the bank_account ID, which is set when the bank_account struct is initialized.

#include <stdlib.h>
#include <threads.h>
 
typedef struct {
  int balance;
  mtx_t balance_mutex;
 
  /* Should not change after initialization */
  unsigned int id;
} bank_account;

typedef struct {
  bank_account *from;
  bank_account *to;
  int amount;
} transaction;

unsigned int global_id = 1;

void create_bank_account(bank_account **ba,
                         int initial_amount) {
  bank_account *nba = (bank_account *)malloc(
    sizeof(bank_account)
  );
  if (nba == NULL) {
    /* Handle error */
  }

  nba->balance = initial_amount;
  if (thrd_success
      != mtx_init(&nba->balance_mutex, mtx_plain)) {
    /* Handle error */
  }

  nba->id = global_id++;
  *ba = nba;
}

int deposit(void *ptr) {
  transaction *args = (transaction *)ptr;
  int result = -1;
  mtx_t *first;
  mtx_t *second;

  if (args->from->id == args->to->id) {
    return -1; /* Indicate error */
  }

  /* Ensure proper ordering for locking */
  if (args->from->id < args->to->id) {
    first = &args->from->balance_mutex;
    second = &args->to->balance_mutex;
  } else {
    first = &args->to->balance_mutex;
    second = &args->from->balance_mutex;
  }
  if (thrd_success != mtx_lock(first)) {
    /* Handle error */
  }
  if (thrd_success != mtx_lock(second)) {
    /* Handle error */
  }

  /* Not enough balance to transfer */
  if (args->from->balance >= args->amount) {
    args->from->balance -= args->amount;
    args->to->balance += args->amount;
    result = 0;
  }

  if (thrd_success != mtx_unlock(second)) {
    /* Handle error */
  }
  if (thrd_success != mtx_unlock(first)) {
    /* Handle error */
  }
  free(ptr);
  return result;
} 

Risk Assessment

Deadlock prevents multiple threads from progressing, halting program execution. A denial-of-service attack is possible if the attacker can create the conditions for deadlock.

Rule

Severity

Likelihood

Remediation Cost

Priority

Level

CON35-C

Low

Probable

Medium

P4

L3

Related Vulnerabilities

Search for vulnerabilities resulting from the violation of this rule on the CERT website.

Automated Detection

ToolVersionCheckerDescription
Astrée
24.04
deadlockSupported by sound analysis (deadlock alarm)
CodeSonar
8.1p0
CONCURRENCY.LOCK.ORDERConflicting lock order
Coverity
2017.07
ORDER_REVERSALFully implemented
Cppcheck Premium

24.9.0

premium-cert-con35-cPartially implemented
Helix QAC

2024.3

C1772, C1773
Klocwork
2024.3

CONC.DL
CONC.NO_UNLOCK


Parasoft C/C++test
2023.1
CERT_C-CON35-a

Do not acquire locks in different order

PC-lint Plus

1.4

2462

Fully supported

Polyspace Bug Finder

R2024a

CERT C: Rule CON35-C

Checks for deadlock (rule partially covered)

Related Guidelines

Key here (explains table format and definitions)

Taxonomy

Taxonomy item

Relationship

CERT Oracle Secure Coding Standard for JavaLCK07-J. Avoid deadlock by requesting and releasing locks in the same orderPrior to 2018-01-12: CERT: Unspecified Relationship

  



6 Comments

  1. 1. i don't think thread execution is random, exactly. can you state this more precisely?
    2. you don't need an introduction header, and the introductory section should state precisely what the guideline is: the developer should do something or should not do something. The guideline should be written to be enforceable. The title and the intro is the statement of the guideline; everything else is an example.
    3. The references are not correctly formatted.
    4. I don't think that 4 is a level. Take another look at how priority and level are calculated.
    5. I'll try to look at the technical content in depth later.

    1. In addition to Rob's comments, I have one:

      • The code snippets are technically correct, but they are clearly abstract. Also they should probably have a main() method that starts the threads. You can say something like "with the right timing, this code could deadlock." But I would like to also see a more practical example of deadlock...consider the AccountBalance class from CON12-J, which you helpfully linked to, thanks.
      1. I have updated my code to have a more concrete example, and I believe all the problems stated has been addressed.

        1. The example is better. More comments:

          • The references section should be last in the order
          • Please use code tags when discussing code snippets in prose, eg: the foo() function.
          • need to check for malloc failure on arg1 and arg2 in NCCE. Also need to check for errors in the various pthread routines.
          • In the CS the id field should be declared const.
          • Given the length of both code examples, the CS should not repeat any functions that are identical to those in the CCE. (such as main()).
          1. I was thinking about having id as const, but I don't think there is a way to dynamically allocate a struct and set its const variable. But I might be wrong.

            1. As far as struct initialization goes, I think you're right. The way to do it in C++ is to use a class (rather than a struct), and then your id can be const, and initialized by the constructor. Using a class is better than a struct, since you can make the fields private, improving encapsulation.