Re: [Gluster-devel] rpc problems when using syncops in callbacks

[Anand Avati]

On Mon, Apr 29, 2013 at 4:33 AM, fog - <address@hidden> wrote:

Hello Krish,

yes, no deadlock occurs without blocking (... somewhat obviously). However, if I can't block I do not gain anything regarding code readability. It makes more sense to use the standard STACK_WIND / UNWIND pairs instead of creating a syncthread with a callback function.

I could use this approach if I start the syncthread in the FOP instead of the CBK function (and use syncops for everything). The problem is that in my scenario only the return of the FOP will tell me if additional FOPs need to be executed (and 99%+ of the time this won't be the case). This makes spawning a syncthread every time sound like a bad idea.

That is correct. I have a half-implemented "syncop xlator" support where the framework prepares the environment before calling in your "sync" FOP methods, but paused the work on it after realizing most of the syncop functions need extra paramters for doing proper unwind of the fops. For now the use case for syncop is if you are having a long sequence of complex fop branching, then you can encapsulate all that complexity into a single synctask_new(cbk), and deal with only one asynchronous callback instead of the entire mess.

Avati

I think you identified the occurring deadlock in your other reply correctly by the way. Seems it's a bit more complicated to use syncops correctly than I originally assumed, I'll probably go back to STACK_WIND / UNWIND chains even if the resulting code is quite messy.

Thanks for your insight
 ~fog

---

Date: Mon, 29 Apr 2013 14:30:50 +0530
From: address@hidden
To: address@hidden
CC: address@hidden

Subject: Re: [Gluster-devel] rpc problems when using syncops in callbacks

Fog,

On 04/29/2013 01:57 PM, fog - wrote:

Hello Avati,

I am wrapping the syncop call in a synctask_new (otherwise glusterFS will run into a null pointer @ synctask_get in the SYNCOP macro & crash). Below is some code to show how I do it currently to test the syncops.

typedef struct{
    xlator_t *this;    loc_t *loc;    dict_t *dic;
}syncstore_args;

int32_t __xattr_store_sync(void* data)
{
    syncstore_args *args = (syncstore_args*)data;
    return syncop_setxattr(FIRST_CHILD(args->this), args->loc, args->dic, 0);
}

int32_t xattr_store_sync(xlator_t *this, call_frame_t *frame, loc_t *loc, dict_t *dic)
{
    syncstore_args args = {this, loc, dic};
    return synctask_new(this->ctx->env, __xattr_store_sync, NULL, NULL, &args);

If you don't provide a synctask_cbk_t to synctask_new, you are using synctask in a 'blocking' mode.
That is, the thread calling synctask_new would block until the synctask_fn_t function (ie, __xattr_store_sync) returns.
An alternative way to do this would be,

int32_t xattr_store_sync(xlator_t *this, call_frame_t *frame, loc_t *loc, dict_t *dic)
{
    syncstore_args args = {this, loc, dic};
    return synctask_new(this->ctx->env, __xattr_store_sync, __xattr_store_sync_cbk, NULL, &args);
}

int32_t __xattr_store_sync_cbk (int ret, /*and the other args*/)
{
    // Your code goes here
    return ret;
}

Now, all file operations performed using syncop_* inside __xattr_store_sync would have the synchronous flavour, while leaving the calling thread (thread calling xattr_store_sync fn) 'free'. This should avoid the hang issue.

HTH,
krish

}

---

Date: Mon, 29 Apr 2013 00:19:11 -0700
Subject: Re: [Gluster-devel] rpc problems when using syncops in callbacks
From: address@hidden
To: address@hidden
CC: address@hidden

Note that you need to place your syncop code in a synctask function strictly within a syncenv (by calling synctask_new(). You're probably calling syncop_XXX() directly in your xlator code?

Avati

On Fri, Apr 26, 2013 at 2:40 AM, fog - <address@hidden> wrote:

Hello everyone,

I am trying to use syncops in a custom translator to keep my code at least borderline readable, but I am having limited success.

Problem Symptoms:
Using a syncop in a regular fop is fine. However, in a callback it causes a 'freeze' (synctask_yield called by the SYNCOP macro doesn't return).

What seems to be the Problem:

Looking at the traces, there is no corresponding trace from rpc_clnt_reply_init on the client to the trace from rpcsvc_submit_generic on the server. In other words, the rpc reply gets sent but isn't correctly received. Obviously this is not really a networking problem but something else... I'd guess it's a deadlock somewhere on the client?

From the point of the syncop call onwards the client doesn't 'get' any rpc replies any more (the next GlusterFS Handshake sent by the client, which is received by the server and replied to, leads to a disconnection accordingly).

Again: This problem is only occurring when calling a syncop from a callback function inside my translator, if I call the same syncop in a fop call it completes fine.

I hope you can make sense out of the above problem description.

Thanks for your time ~

_______________________________________________
Gluster-devel mailing list
address@hidden
https://lists.nongnu.org/mailman/listinfo/gluster-devel

_______________________________________________
Gluster-devel mailing list
address@hidden
https://lists.nongnu.org/mailman/listinfo/gluster-devel

_______________________________________________
Gluster-devel mailing list
address@hidden
https://lists.nongnu.org/mailman/listinfo/gluster-devel