On Tue, Jun 04, 2024 at 10:33:49AM +0100, Simon Horman wrote:
> On Fri, May 31, 2024 at 08:37:41AM -0700, Shradha Gupta wrote:
> > Allow variable size indirection table allocation in MANA instead
> > of using a constant value MANA_INDIRECT_TABLE_SIZE.
> > The size is now derived from the MANA_QUERY_VPORT_CONFIG and the
> > indirection table is allocated dynamically.
> >
> > Signed-off-by: Shradha Gupta <shradhagupta@xxxxxxxxxxxxxxxxxxx>
> > Reviewed-by: Dexuan Cui <decui@xxxxxxxxxxxxx>
> > Reviewed-by: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>
>
> ...
>
> > diff --git a/drivers/net/ethernet/microsoft/mana/mana_en.c b/drivers/net/ethernet/microsoft/mana/mana_en.c
>
> ...
>
> > @@ -2344,11 +2352,33 @@ static int mana_create_vport(struct mana_port_context *apc,
> > return mana_create_txq(apc, net);
> > }
> >
> > +static int mana_rss_table_alloc(struct mana_port_context *apc)
> > +{
> > + if (!apc->indir_table_sz) {
> > + netdev_err(apc->ndev,
> > + "Indirection table size not set for vPort %d\n",
> > + apc->port_idx);
> > + return -EINVAL;
> > + }
> > +
> > + apc->indir_table = kcalloc(apc->indir_table_sz, sizeof(u32), GFP_KERNEL);
> > + if (!apc->indir_table)
> > + return -ENOMEM;
> > +
> > + apc->rxobj_table = kcalloc(apc->indir_table_sz, sizeof(mana_handle_t), GFP_KERNEL);
> > + if (!apc->rxobj_table) {
> > + kfree(apc->indir_table);
>
> Hi, Shradha
>
> Perhaps I am on the wrong track here, but I have some concerns
> about clean-up paths.
>
> Firstly. I think that apc->indir_table should be to NULL here for
> consistency with other clean-up paths. Or alternatively, fields of apc
> should not set to NULL elsewhere after being freed.
Hi Simon,
Thanks for the comments. This makes sense, I am planning of consistently removing
the NULLify from other places too as per Leon's comments.
>
> In looking into this I noticed that mana_probe() does not call
> mana_remove() or return an error in the cases where mana_probe_port() or
> mana_attach() fail unless add_adev also fails. If so, is that intentional?
Right, so most calls like mana_probe_port(), mana_attach() cleanup after themselves
in the code if there is any error. So, not having to call mana_remove() in these
cases in mana_probe() is intentional. But I do agree that an error is returned in
mana_probe() only if add_adev also fails. I'll fix that too in the next version
>
> In any case, I would suggest as a follow-up, arranging things so that when
> an error occurs in a function, anything that was allocated is unwound
> before returning an error.
>
> I think this would make allocation/deallocation easier to reason with.
> And I suspect it would avoid both the need for fields of structures to be
> zeroed after being freed, and the need to call mana_remove() from
> mana_probe().
Agreed
>
> > + return -ENOMEM;
> > + }
> > +
> > + return 0;
> > +}
> > +
> > static void mana_rss_table_init(struct mana_port_context *apc)
> > {
> > int i;
> >
> > - for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++)
> > + for (i = 0; i < apc->indir_table_sz; i++)
> > apc->indir_table[i] =
> > ethtool_rxfh_indir_default(i, apc->num_queues);
> > }
>
> ...
>
> > @@ -2739,11 +2772,17 @@ static int mana_probe_port(struct mana_context *ac, int port_idx,
> > err = register_netdev(ndev);
> > if (err) {
> > netdev_err(ndev, "Unable to register netdev.\n");
> > - goto reset_apc;
> > + goto free_indir;
> > }
> >
> > return 0;
> >
> > +free_indir:
> > + apc->indir_table_sz = 0;
> > + kfree(apc->indir_table);
> > + apc->indir_table = NULL;
> > + kfree(apc->rxobj_table);
> > + apc->rxobj_table = NULL;
> > reset_apc:
> > kfree(apc->rxqs);
> > apc->rxqs = NULL;
>
> nit: Not strictly related to this patch, but the reset_apc code should
> probably be a call to mana_cleanup_port_context() as it is the dual of
> mana_init_port_context() which is called earlier in mana_probe_port()
Sure, let me do that too.
>
> ...
>
> > @@ -2931,6 +2972,11 @@ void mana_remove(struct gdma_dev *gd, bool suspending)
> > }
> >
> > unregister_netdevice(ndev);
> > + apc->indir_table_sz = 0;
> > + kfree(apc->indir_table);
> > + apc->indir_table = NULL;
> > + kfree(apc->rxobj_table);
> > + apc->rxobj_table = NULL;
>
> The code to free and zero indir_table_sz and indir_table appears twice
> in this patch. Perhaps a helper to do this, which would be the dual
> of mana_rss_table_alloc is in order.
Makes sense, will change this too.
Thanks,
Shradha.
>
> >
> > rtnl_unlock();
> >
>
> ...
