Initial skeleton of in-memory locks

Signed-off-by: Matthew Heon <matthew.heon@gmail.com>

4 files changed, 634 insertions, 0 deletions

diff --git a/libpod/lock/locks.go b/libpod/lock/locks.go
new file mode 100644
index 000000000..14da7d6a7
--- /dev/null
+++ b/libpod/lock/locks.go
@@ -0,0 +1,167 @@
+package lock
+
+// #cgo LDFLAGS: -lrt -lpthread
+// #include "shm_lock.h"
+// const uint32_t bitmap_size_c = BITMAP_SIZE;
+import "C"
+
+import (
+	"syscall"
+
+	"github.com/pkg/errors"
+)
+
+var (
+	bitmapSize uint32 = uint32(C.bitmap_size_c)
+)
+
+// SHMLocks is a struct enabling POSIX semaphore locking in a shared memory
+// segment
+type SHMLocks struct {
+	lockStruct *C.shm_struct_t
+	valid      bool
+	maxLocks   uint32
+}
+
+// CreateSHMLock sets up a shared-memory segment holding a given number of POSIX
+// semaphores, and returns a struct that can be used to operate on those locks.
+// numLocks must be a multiple of the lock bitmap size (by default, 32).
+func CreateSHMLock(numLocks uint32) (*SHMLocks, error) {
+	if numLocks % bitmapSize != 0 {
+		return nil, errors.Wrapf(syscall.EINVAL, "number of locks must be a multiple of %d", C.bitmap_size_c)
+	}
+
+	locks := new(SHMLocks)
+
+	lockStruct := C.setup_lock_shm(C.uint32_t(numLocks))
+	if lockStruct == nil {
+		// We got a null pointer, so something errored
+		return nil, errors.Wrapf(syscall.ENOENT, "error creating shared memory locks")
+	}
+
+	locks.lockStruct = lockStruct
+	locks.maxLocks = numLocks
+	locks.valid = true
+
+	return locks, nil
+}
+
+// OpenSHMLock opens an existing shared-memory segment holding a given number of
+// POSIX semaphores. numLocks must match the number of locks the shared memory
+// segment was created with and be a multiple of the lock bitmap size (default
+// 32)
+func OpenSHMLock(numLocks uint32) (*SHMLocks, error) {
+	if numLocks % bitmapSize != 0 {
+		return nil, errors.Wrapf(syscall.EINVAL, "number of locks must be a multiple of %d", C.bitmap_size_c)
+	}
+
+	locks := new(SHMLocks)
+
+	lockStruct := C.open_lock_shm(C.uint32_t(numLocks))
+	if lockStruct == nil {
+		// We got a null pointer, so something errored
+		return nil, errors.Wrapf(syscall.ENOENT, "error creating shared memory locks")
+	}
+
+	locks.lockStruct = lockStruct
+	locks.maxLocks = numLocks
+	locks.valid = true
+
+	return locks, nil
+}
+
+// Close closes an existing shared-memory segment.
+// The segment will be rendered unusable after closing.
+// WARNING: If you Close() while there are still locks locked, these locks may
+// fail to release, causing a program freeze.
+// Close() is only intended to be used while testing the locks.
+func (locks *SHMLocks) Close() error {
+	if !locks.valid {
+		return errors.Wrapf(syscall.EINVAL, "locks have already been closed")
+	}
+
+	locks.valid = false
+
+	retCode := C.close_lock_shm(locks.lockStruct)
+	if retCode < 0 {
+		// Negative errno returned
+		return syscall.Errno(-1 * retCode)
+	}
+
+	return nil
+}
+
+// AllocateSemaphore allocates a semaphore from a shared-memory segment for use
+// by a container or pod.
+// Returns the index of the semaphore that was allocated
+func (locks *SHMLocks) AllocateSemaphore() (uint32, error) {
+	if !locks.valid {
+		return 0, errors.Wrapf(syscall.EINVAL, "locks have already been closed")
+	}
+
+	retCode := C.allocate_semaphore(locks.lockStruct)
+	if retCode < 0 {
+		// Negative errno returned
+		return 0, syscall.Errno(-1 * retCode)
+	}
+
+	return uint32(retCode), nil
+}
+
+// DeallocateSemaphore frees a semaphore in a shared-memory segment for use by
+// a container of pod
+func (locks *SHMLocks) DeallocateSemaphore(sem uint32) error {
+	if !locks.valid {
+		return errors.Wrapf(syscall.EINVAL, "locks have already been closed")
+	}
+
+	if sem > locks.maxLocks {
+		return errors.Wrapf(syscall.EINVAL, "given semaphore %d is higher than maximum locks count %d", sem, locks.maxLocks)
+	}
+
+	retCode := C.deallocate_semaphore(locks.lockStruct, C.uint32_t(sem))
+	if retCode < 0 {
+		// Negative errno returned
+		return syscall.Errno(-1 * retCode)
+	}
+
+	return nil
+}
+
+// LockSemaphore locks the given semaphore
+func (locks *SHMLocks) LockSemaphore(sem uint32) error {
+	if !locks.valid {
+		return errors.Wrapf(syscall.EINVAL, "locks have already been closed")
+	}
+
+	if sem > locks.maxLocks {
+		return errors.Wrapf(syscall.EINVAL, "given semaphore %d is higher than maximum locks count %d", sem, locks.maxLocks)
+	}
+
+	retCode := C.lock_semaphore(locks.lockStruct, C.uint32_t(sem))
+	if retCode < 0 {
+		// Negative errno returned
+		return syscall.Errno(-1 * retCode)
+	}
+
+	return nil
+}
+
+// UnlockSemaphore locks the given semaphore
+func (locks *SHMLocks) UnlockSemaphore(sem uint32) error {
+	if !locks.valid {
+		return errors.Wrapf(syscall.EINVAL, "locks have already been closed")
+	}
+
+	if sem > locks.maxLocks {
+		return errors.Wrapf(syscall.EINVAL, "given semaphore %d is higher than maximum locks count %d", sem, locks.maxLocks)
+	}
+
+	retCode := C.unlock_semaphore(locks.lockStruct, C.uint32_t(sem))
+	if retCode < 0 {
+		// Negative errno returned
+		return syscall.Errno(-1 * retCode)
+	}
+
+	return nil
+}
diff --git a/libpod/lock/locks_test.go b/libpod/lock/locks_test.go
new file mode 100644
index 000000000..b14d6e19a
--- /dev/null
+++ b/libpod/lock/locks_test.go
@@ -0,0 +1,84 @@
+package lock
+
+import (
+	"fmt"
+	"os"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// All tests here are in the same process, which somewhat limits their utility
+// The big intent of this package it multiprocess locking, which is really hard
+// to test without actually having multiple processes...
+// We can at least verify that the locks work within the local process.
+
+// 4 * BITMAP_SIZE to ensure we have to traverse bitmaps
+const numLocks = 128
+
+// We need a test main to ensure that the SHM is created before the tests run
+func TestMain(m *testing.M) {
+	shmLock, err := CreateSHMLock(numLocks)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "Error creating SHM for tests: %v\n", err)
+		os.Exit(-1)
+	}
+
+	// Close the SHM - every subsequent test will reopen
+	if err := shmLock.Close(); err != nil {
+		fmt.Fprintf(os.Stderr, "Error closing SHM locks: %v\n", err)
+		os.Exit(-1)
+	}
+
+	exitCode := m.Run()
+
+	// We need to remove the SHM segment to clean up after ourselves
+	os.RemoveAll("/dev/shm/libpod_lock")
+
+	os.Exit(exitCode)
+}
+
+
+func runLockTest(t *testing.T, testFunc func(*testing.T, *SHMLocks)) {
+	locks, err := OpenSHMLock(numLocks)
+	if err != nil {
+		t.Fatalf("Error opening locks: %v", err)
+	}
+	defer func() {
+		if err := locks.Close(); err != nil {
+			t.Fatalf("Error closing locks: %v", err)
+		}
+	}()
+
+	success := t.Run("locks", func (t *testing.T) {
+		testFunc(t, locks)
+	})
+	if !success {
+		t.Fail()
+	}
+}
+
+// Test that creating an SHM with a bad size fails
+func TestCreateNewSHMBadSize(t *testing.T) {
+	// Odd number, not a power of 2, should never be a word size on a system
+	_, err := CreateSHMLock(7)
+	assert.Error(t, err)
+}
+
+// Test allocating - lock - unlock - deallocate cycle, single lock
+func TestLockLifecycleSingleLock(t *testing.T) {
+	runLockTest(t, func(t *testing.T, locks *SHMLocks) {
+		sem, err := locks.AllocateSemaphore()
+		require.NoError(t, err)
+
+		err = locks.LockSemaphore(sem)
+		assert.NoError(t, err)
+
+		err = locks.UnlockSemaphore(sem)
+		assert.NoError(t, err)
+
+		err = locks.DeallocateSemaphore(sem)
+		assert.NoError(t, err)
+	})
+}
diff --git a/libpod/lock/shm_lock.c b/libpod/lock/shm_lock.c
new file mode 100644
index 000000000..ab715891c
--- /dev/null
+++ b/libpod/lock/shm_lock.c
@@ -0,0 +1,340 @@
+#include <errno.h>
+#include <fcntl.h>
+#include <semaphore.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#include "shm_lock.h"
+
+// Compute the size of the SHM struct
+size_t compute_shm_size(uint32_t num_bitmaps) {
+  return sizeof(shm_struct_t) + (num_bitmaps * sizeof(lock_group_t));
+}
+
+// Set up an SHM segment holding locks for libpod
+// num_locks must be a multiple of BITMAP_SIZE (32 by default)
+// Returns a valid pointer on success or NULL on error
+shm_struct_t *setup_lock_shm(uint32_t num_locks) {
+  int shm_fd, i, j, ret_code;
+  uint32_t num_bitmaps;
+  size_t shm_size;
+  shm_struct_t *shm;
+
+  // TODO maybe set errno so we can get errors back to libpod?
+
+  // We need a nonzero number of locks
+  if (num_locks == 0) {
+    return NULL;
+  }
+
+  // Calculate the number of bitmaps required
+  if (num_locks % BITMAP_SIZE != 0) {
+    // Number of locks not a multiple of BITMAP_SIZE
+    return NULL;
+  }
+  num_bitmaps = num_locks / BITMAP_SIZE;
+
+  // Calculate size of the shm segment
+  shm_size = compute_shm_size(num_bitmaps);
+
+  // Create a new SHM segment for us
+  shm_fd = shm_open(SHM_NAME, O_RDWR | O_CREAT | O_EXCL, 0600);
+  if (shm_fd < 0) {
+    return NULL;
+  }
+
+  // Increase its size to what we need
+  ret_code = ftruncate(shm_fd, shm_size);
+  if (ret_code < 0) {
+    goto CLEANUP_UNLINK;
+  }
+
+  // Map the shared memory in
+  shm = mmap(NULL, shm_size, PROT_READ | PROT_WRITE, MAP_SHARED, shm_fd, 0);
+  if (shm == MAP_FAILED) {
+    goto CLEANUP_UNLINK;
+  }
+
+  // We have successfully mapped the memory, now initialize the region
+  shm->magic = MAGIC;
+  shm->num_locks = num_locks;
+  shm->num_bitmaps = num_bitmaps;
+
+  // Initialize the semaphore that protects the bitmaps
+  ret_code = sem_init(&(shm->segment_lock), 1, 1);
+  if (ret_code < 0) {
+    goto CLEANUP_UNMAP;
+  }
+
+  // Initialize all bitmaps to 0 initially
+  // And initialize all semaphores they use
+  for (i = 0; i < num_bitmaps; i++) {
+    shm->locks[i].bitmap = 0;
+    for (j = 0; j < BITMAP_SIZE; j++) {
+      ret_code = sem_init(&(shm->locks[i].locks[j]), 1, 1);
+      if (ret_code < 0) {
+	goto CLEANUP_UNMAP;
+      }
+    }
+  }
+
+  // Close the file descriptor, we're done with it
+  // Ignore errors, it's ok if we leak a single FD and this should only run once
+  close(shm_fd);
+
+  return shm;
+
+  // Cleanup after an error
+ CLEANUP_UNMAP:
+  munmap(shm, shm_size);
+ CLEANUP_UNLINK:
+  close(shm_fd);
+  shm_unlink(SHM_NAME);
+  return NULL;
+}
+
+// Open an existing SHM segment holding libpod locks
+// num_locks is the number of locks that will be configured in the SHM segment
+// num_locks must be a multiple of BITMAP_SIZE (32 by default)
+// Returns a valid pointer on success or NULL on error
+shm_struct_t *open_lock_shm(uint32_t num_locks) {
+  int shm_fd;
+  shm_struct_t *shm;
+  size_t shm_size;
+  uint32_t num_bitmaps;
+
+  // We need a nonzero number of locks
+  if (num_locks == 0) {
+    return NULL;
+  }
+
+  // Calculate the number of bitmaps required
+  if (num_locks % BITMAP_SIZE != 0) {
+    // Number of locks not a multiple of BITMAP_SIZE
+    return NULL;
+  }
+  num_bitmaps = num_locks / BITMAP_SIZE;
+
+  // Calculate size of the shm segment
+  shm_size = compute_shm_size(num_bitmaps);
+
+  shm_fd = shm_open(SHM_NAME, O_RDWR, 0600);
+  if (shm_fd < 0) {
+    return NULL;
+  }
+
+  // Map the shared memory in
+  shm = mmap(NULL, shm_size, PROT_READ | PROT_WRITE, MAP_SHARED, shm_fd, 0);
+
+  // Ignore errors, it's ok if we leak a single FD since this only runs once
+  close(shm_fd);
+
+  // Check if we successfully mmap'd
+  if (shm == MAP_FAILED) {
+    return NULL;
+  }
+
+  // Need to check the SHM to see if it's actually our locks
+  if (shm->magic != MAGIC) {
+    goto CLEANUP;
+  }
+  if (shm->num_locks != num_locks) {
+    goto CLEANUP;
+  }
+
+  return shm;
+
+ CLEANUP:
+  munmap(shm, shm_size);
+  return NULL;
+}
+
+// Close an open SHM lock struct, unmapping the backing memory.
+// The given shm_struct_t will be rendered unusable as a result.
+// On success, 0 is returned. On failure, negative ERRNO values are returned.
+int32_t close_lock_shm(shm_struct_t *shm) {
+  int ret_code;
+  size_t shm_size;
+
+  // We can't unmap null...
+  if (shm == NULL) {
+    return -1 * EINVAL;
+  }
+
+  shm_size = compute_shm_size(shm->num_bitmaps);
+
+  ret_code = munmap(shm, shm_size);
+
+  if (ret_code != 0) {
+    return -1 * errno;
+  }
+
+  return 0;
+}
+
+// Allocate the first available semaphore
+// Returns a positive integer guaranteed to be less than UINT32_MAX on success,
+// or negative errno values on failure
+// On sucess, the returned integer is the number of the semaphore allocated
+int64_t allocate_semaphore(shm_struct_t *shm) {
+  int ret_code, i;
+  bitmap_t test_map;
+  int64_t sem_number, num_within_bitmap;
+
+  if (shm == NULL) {
+    return -1 * EINVAL;
+  }
+
+  // Lock the semaphore controlling access to our shared memory
+  do {
+    ret_code = sem_wait(&(shm->segment_lock));
+  } while(ret_code == EINTR);
+  if (ret_code != 0) {
+    return -1 * errno;
+  }
+
+  // Loop through our bitmaps to search for one that is not full
+  for (i = 0; i < shm->num_bitmaps; i++) {
+    if (shm->locks[i].bitmap != 0xFFFFFFFF) {
+      test_map = 0x1;
+      num_within_bitmap = 0;
+      while (test_map != 0) {
+	if ((test_map & shm->locks[i].bitmap) == 0) {
+	  // Compute the number of the semaphore we are allocating
+	  sem_number = (BITMAP_SIZE * i) + num_within_bitmap;
+	  // OR in the bitmap
+	  shm->locks[i].bitmap = shm->locks[i].bitmap | test_map;
+	  // Clear the semaphore
+	  sem_post(&(shm->segment_lock));
+	  // Return the semaphore we've allocated
+	  return sem_number;
+	}
+	test_map = test_map << 1;
+	num_within_bitmap++;
+      }
+      // We should never fall through this loop
+      // TODO maybe an assert() here to panic if we do?
+    }
+  }
+
+  // Post to the semaphore to clear the lock
+  sem_post(&(shm->segment_lock));
+
+  // All bitmaps are full
+  // We have no available semaphores, report allocation failure
+  return -1 * ENOSPC;
+}
+
+// Deallocate a given semaphore
+// Returns 0 on success, negative ERRNO values on failure
+int32_t deallocate_semaphore(shm_struct_t *shm, uint32_t sem_index) {
+  bitmap_t test_map;
+  int bitmap_index, index_in_bitmap, ret_code, i;
+
+  if (shm == NULL) {
+    return -1 * EINVAL;
+  }
+
+  // Check if the lock index is valid
+  if (sem_index >= shm->num_locks) {
+    return -1 * EINVAL;
+  }
+
+  bitmap_index = sem_index / BITMAP_SIZE;
+  index_in_bitmap = sem_index % BITMAP_SIZE;
+
+  // This should never happen if the sem_index test above succeeded, but better
+  // safe than sorry
+  if (bitmap_index >= shm->num_bitmaps) {
+    return -1 * EFAULT;
+  }
+
+  test_map = 0x1;
+  for (i = 0; i < index_in_bitmap; i++) {
+    test_map = test_map << 1;
+  }
+
+  // Lock the semaphore controlling access to our shared memory
+  do {
+    ret_code = sem_wait(&(shm->segment_lock));
+  } while(ret_code == EINTR);
+  if (ret_code != 0) {
+    return -1 * errno;
+  }
+
+  // Check if the semaphore is allocated
+  if ((test_map & shm->locks[bitmap_index].bitmap) == 0) {
+    // Post to the semaphore to clear the lock
+    sem_post(&(shm->segment_lock));
+
+    return -1 * ENOENT;
+  }
+
+  // The semaphore is allocated, clear it
+  // Invert the bitmask we used to test to clear the bit
+  test_map = ~test_map;
+  shm->locks[bitmap_index].bitmap = shm->locks[bitmap_index].bitmap & test_map;
+
+  // Post to the semaphore to clear the lock
+  sem_post(&(shm->segment_lock));
+
+  return 0;
+}
+
+// Lock a given semaphore
+// Does not check if the semaphore is allocated - this ensures that, even for
+// removed containers, we can still successfully lock to check status (and
+// subsequently realize they have been removed).
+// Returns 0 on success, -1 on failure
+int32_t lock_semaphore(shm_struct_t *shm, uint32_t sem_index) {
+  int bitmap_index, index_in_bitmap, ret_code;
+
+  if (shm == NULL) {
+    return -1 * EINVAL;
+  }
+
+  if (sem_index >= shm->num_locks) {
+    return -1 * EINVAL;
+  }
+
+  bitmap_index = sem_index / BITMAP_SIZE;
+  index_in_bitmap = sem_index % BITMAP_SIZE;
+
+  // Lock the semaphore controlling access to our shared memory
+  do {
+    ret_code = sem_wait(&(shm->locks[bitmap_index].locks[index_in_bitmap]));
+  } while(ret_code == EINTR);
+  if (ret_code != 0) {
+    return -1 * errno;
+  }
+
+  return 0;
+}
+
+// Unlock a given semaphore
+// Does not check if the semaphore is allocated - this ensures that, even for
+// removed containers, we can still successfully lock to check status (and
+// subsequently realize they have been removed).
+// Returns 0 on success, -1 on failure
+int32_t unlock_semaphore(shm_struct_t *shm, uint32_t sem_index) {
+  int bitmap_index, index_in_bitmap;
+
+  if (shm == NULL) {
+    return -1 * EINVAL;
+  }
+
+  if (sem_index >= shm->num_locks) {
+    return -1 * EINVAL;
+  }
+
+  bitmap_index = sem_index / BITMAP_SIZE;
+  index_in_bitmap = sem_index % BITMAP_SIZE;
+
+  sem_post(&(shm->locks[bitmap_index].locks[index_in_bitmap]));
+
+  return 0;
+}
diff --git a/libpod/lock/shm_lock.h b/libpod/lock/shm_lock.h
new file mode 100644
index 000000000..6988789dd
--- /dev/null
+++ b/libpod/lock/shm_lock.h
@@ -0,0 +1,43 @@
+#ifndef shm_locks_h_
+#define shm_locks_h_
+
+#include <semaphore.h>
+#include <stdint.h>
+
+// Magic number to ensure we open the right SHM segment
+#define MAGIC 0xA5A5
+
+// Name of the SHM
+#define SHM_NAME "/libpod_lock"
+
+// Type for our bitmaps
+typedef uint32_t bitmap_t;
+
+// bitmap size
+#define BITMAP_SIZE (sizeof(bitmap_t) * 8)
+
+// Struct to hold a single bitmap and associated locks
+typedef struct lock_group {
+  bitmap_t bitmap;
+  sem_t    locks[BITMAP_SIZE];
+} lock_group_t;
+
+// Struct to hold our SHM locks
+typedef struct shm_struct {
+  uint16_t     magic;
+  sem_t        segment_lock;
+  uint32_t     num_bitmaps;
+  uint32_t     num_locks;
+  lock_group_t locks[];
+} shm_struct_t;
+
+size_t compute_shm_size(uint32_t num_bitmaps);
+shm_struct_t *setup_lock_shm(uint32_t num_locks);
+shm_struct_t *open_lock_shm(uint32_t num_locks);
+int32_t close_lock_shm(shm_struct_t *shm);
+int64_t allocate_semaphore(shm_struct_t *shm);
+int32_t deallocate_semaphore(shm_struct_t *shm, uint32_t sem_index);
+int32_t lock_semaphore(shm_struct_t *shm, uint32_t sem_index);
+int32_t unlock_semaphore(shm_struct_t *shm, uint32_t sem_index);
+
+#endif