On 08/30/2012 12:07 PM, Yehuda Sadeh wrote:
> On Fri, Aug 24, 2012 at 9:33 AM, Alex Elder <elder@xxxxxxxxxxx> wrote:
>> The rbd options don't really apply to the ceph client. So don't
>> store a pointer to it in the ceph_client structure, and put them
>> (a struct, not a pointer) into the rbd_dev structure proper.
>>
>> Pass the rbd device structure to rbd_client_create() so it can
>> assign rbd_dev->rbdc if successful, and have it return an error code
>> instead of the rbd client pointer.
>>
>> Signed-off-by: Alex Elder <elder@xxxxxxxxxxx>
>> ---
>> drivers/block/rbd.c | 49
>> ++++++++++++++++---------------------------------
>> 1 file changed, 16 insertions(+), 33 deletions(-)
>>
>> diff --git a/drivers/block/rbd.c b/drivers/block/rbd.c
>> index 130dbc2..b40f553 100644
>> --- a/drivers/block/rbd.c
>> +++ b/drivers/block/rbd.c
>> @@ -98,7 +98,6 @@ struct rbd_options {
>> */
>> struct rbd_client {
>> struct ceph_client *client;
>> - struct rbd_options *rbd_opts;
>> struct kref kref;
>> struct list_head node;
>> };
>> @@ -152,6 +151,7 @@ struct rbd_device {
>> struct gendisk *disk; /* blkdev's gendisk and rq */
>> struct request_queue *q;
>>
>> + struct rbd_options rbd_opts;
>> struct rbd_client *rbd_client;
>>
>> char name[DEV_NAME_LEN]; /* blkdev name, e.g. rbd3 */
>> @@ -273,8 +273,7 @@ static const struct block_device_operations
>> rbd_bd_ops = {
>> * Initialize an rbd client instance.
>> * We own *ceph_opts.
>> */
>> -static struct rbd_client *rbd_client_create(struct ceph_options *ceph_opts,
>> - struct rbd_options *rbd_opts)
>> +static struct rbd_client *rbd_client_create(struct ceph_options *ceph_opts)
>> {
>> struct rbd_client *rbdc;
>> int ret = -ENOMEM;
>> @@ -298,8 +297,6 @@ static struct rbd_client *rbd_client_create(struct
>> ceph_options *ceph_opts,
>> if (ret < 0)
>> goto out_err;
>>
>> - rbdc->rbd_opts = rbd_opts;
>> -
>> spin_lock(&rbd_client_list_lock);
>> list_add_tail(&rbdc->node, &rbd_client_list);
>> spin_unlock(&rbd_client_list_lock);
>> @@ -402,42 +399,33 @@ static int parse_rbd_opts_token(char *c, void
>> *private)
>> * Get a ceph client with specific addr and configuration, if one does
>> * not exist create it.
>> */
>> -static struct rbd_client *rbd_get_client(const char *mon_addr,
>> - size_t mon_addr_len,
>> - char *options)
>> +static int rbd_get_client(struct rbd_device *rbd_dev, const char *mon_addr,
>> + size_t mon_addr_len, char *options)
>> {
>> - struct rbd_client *rbdc;
>> + struct rbd_options *rbd_opts = &rbd_dev->rbd_opts;
>> struct ceph_options *ceph_opts;
>> - struct rbd_options *rbd_opts;
>> -
>> - rbd_opts = kzalloc(sizeof(*rbd_opts), GFP_KERNEL);
>> - if (!rbd_opts)
>> - return ERR_PTR(-ENOMEM);
>> + struct rbd_client *rbdc;
>>
>> rbd_opts->notify_timeout = RBD_NOTIFY_TIMEOUT_DEFAULT;
>>
>> ceph_opts = ceph_parse_options(options, mon_addr,
>> mon_addr + mon_addr_len,
>> parse_rbd_opts_token, rbd_opts);
>> - if (IS_ERR(ceph_opts)) {
>> - kfree(rbd_opts);
>> - return ERR_CAST(ceph_opts);
>> - }
>> + if (IS_ERR(ceph_opts))
>> + return PTR_ERR(ceph_opts);
>>
>> rbdc = rbd_client_find(ceph_opts);
>> if (rbdc) {
>> /* using an existing client */
>> ceph_destroy_options(ceph_opts);
>
> It'll probably be better to use a reference count to control ceph_opts lifecycle
I looked at this.
In this case, ceph_opts is directly tied to a ceph_client.
If a ceph client exists whose options match the one that
has been created here, that client will be used, and its
options get destroyed when the ceph_client gets destroyed
(and the ceph_client is refcounted). The ceph_opts
created here will never be referenced by anything else
so it's safe to just destroy it.
If no existing rbd_client has a ceph_client with matching
ceph_options, then one is created below and it will use
the ceph_opts created here and thus its life cycle will
be managed by that ceph_client's reference count.
I believe this means there is no reason to reference count
ceph_opts.
Yehuda, please let me know if you disagree with this. If
not, would you sign off on this?
There is a different problem here though, which I will
address separately. The rbd_client_find() has protection
of rbd_client_list_lock(), and inserting a new rbd
client into the list in rbd_client_create() gets the
same protection, but there's a race between the two
that could lead to the creation of multiple clients for
the same set of options. This may still produce correct
behavior but it's not the way it should work.
I've created this bug to track getting this fixed at
some (near) future date:
http://tracker.newdream.net/issues/3094
-Alex
>> - kfree(rbd_opts);
>> -
>> - return rbdc;
>> + } else {
>> + rbdc = rbd_client_create(ceph_opts);
>> + if (IS_ERR(rbdc))
>> + return PTR_ERR(rbdc);
>> }
>> + rbd_dev->rbd_client = rbdc;
>>
>> - rbdc = rbd_client_create(ceph_opts, rbd_opts);
>> - if (IS_ERR(rbdc))
>> - kfree(rbd_opts);
>> -
>> - return rbdc;
>> + return 0;
>> }
>>
>> /*
>> @@ -455,7 +443,6 @@ static void rbd_client_release(struct kref *kref)
>> spin_unlock(&rbd_client_list_lock);
>>
>> ceph_destroy_client(rbdc->client);
>> - kfree(rbdc->rbd_opts);
>> kfree(rbdc);
>> }
>>
>> @@ -2526,13 +2513,9 @@ static ssize_t rbd_add(struct bus_type *bus,
>> if (rc)
>> goto err_put_id;
>>
>> - rbd_dev->rbd_client = rbd_get_client(mon_addrs, mon_addrs_size - 1,
>> - options);
>> - if (IS_ERR(rbd_dev->rbd_client)) {
>> - rc = PTR_ERR(rbd_dev->rbd_client);
>> - rbd_dev->rbd_client = NULL;
>> + rc = rbd_get_client(rbd_dev, mon_addrs, mon_addrs_size - 1, options);
>> + if (rc < 0)
>> goto err_put_id;
>> - }
>>
>> /* pick the pool */
>> osdc = &rbd_dev->rbd_client->client->osdc;
>> --
>> 1.7.9.5
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
>> the body of a message to majordomo@xxxxxxxxxxxxxxx
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>
>
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
