Hello.
Thank you for your comment.
James Morris wrote:
I don't need to use spinlocks here.
What I need to do here is avoid read/write reordering,
so mb() will be appropriate here.
struct something {
struct something *next;
void *data;
};
struct something *prev_ptr = ...;
struct something *new_ptr = kmalloc(sizeof(*new_ptr), GFP_KERNEL);
new_ptr->next = NULL;
new_ptr->data = some_value;
mb();
prev_ptr->next = new_ptr;
TOMOYO Linux doesn't use locks for reading singly-linked list.
Thus, new_ptr->data has to be made visible to other CPUs
before new_ptr is appended at the tail of singly-linked list.
Otherwise, other CPU may read undefined new_ptr->data values.
Performance is not critical because this mb() is called
only when appending new entry to the singly-linked list.
Since new_ptr never be removed, new_ptr needn't to know
it's previous element's address, thus having only ->next pointer is enough.
new_ptr->next is assigned *only once*
(initialized with NULL, and assigned non-NULL later),
indicating "if ptr->next == NULL, ptr is the last element" and
"if ptr->next != NULL, ptr is not the last element".
If I use "struct hlist_node" which has two pointers,
it needs more memory than "struct something" which has one pointer.
It is possible to use API defined in include/linux/list.h ,
but to reduce memory usage, I dare not use these API.
Singly-linked list's rule in TOMOYO Linux is quite simple,
"ptr->next is NULL if the last element" and "non-NULL if not".
Regards.
-
