lkml.org 
[lkml]   [2015]   [May]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH-v2 1/9] target: Convert se_node_acl->device_list[] to RCU hlist
From
Date
(resending)

On Fri, 2015-05-22 at 01:24 -0700, Christoph Hellwig wrote:
> > - spin_lock_irqsave(&se_sess->se_node_acl->device_list_lock, flags);
> > - se_cmd->se_deve = se_sess->se_node_acl->device_list[unpacked_lun];
> > - if (se_cmd->se_deve->lun_flags & TRANSPORT_LUNFLAGS_INITIATOR_ACCESS) {
> > - struct se_dev_entry *deve = se_cmd->se_deve;
> > -
> > + rcu_read_lock();
> > + deve = target_nacl_find_deve(nacl, unpacked_lun);
> > + if (deve) {
> > deve->total_cmds++;
>
> This update will now be racy, ditto for the read/write_bytes update
> later.

This should become an atomic_long_t increment, yes..?

>
> > +bool target_lun_is_rdonly(struct se_cmd *cmd)
> > +{
> > + struct se_session *se_sess = cmd->se_sess;
> > + struct se_dev_entry *deve;
> > + bool ret;
> > +
> > + if (cmd->se_lun->lun_access & TRANSPORT_LUNFLAGS_READ_ONLY)
> > + return true;
> > +
> > + rcu_read_lock();
> > + deve = target_nacl_find_deve(se_sess->se_node_acl, cmd->orig_fe_lun);
> > + ret = (deve && deve->lun_flags & TRANSPORT_LUNFLAGS_READ_ONLY);
> > + rcu_read_unlock();
> > +
> > + return ret;
> > +}
> > +EXPORT_SYMBOL(target_lun_is_rdonly);
>
> This should be a separate prep patch like, like it was in my original
> version. I also still think you want this whole patch:
>
> http://git.infradead.org/users/hch/scsi.git/commitdiff/e9a71bda1a120e0488c5c4e4b2f17f14333e2dc6
>
> as storing a pointer to the dev entry without a refcount is bound to
> cause trouble. I don't have a tree with all the patches applied
> available, but I doubt it fully gets that right.
>

Yes, this helper is from your patch above.

Considering there is a single user of it here, and complexities involved
for a RCU conversion + bisect, is it really work adding as a separate
patch ahead of this one..?

> > +void target_pr_kref_release(struct kref *kref)
> > +{
> > + struct se_dev_entry *deve = container_of(kref, struct se_dev_entry,
> > + pr_kref);
> > + complete(&deve->pr_comp);
> > }
> >
> > /* core_enable_device_list_for_node():
>
> > + kref_put(&orig->pr_kref, target_pr_kref_release);
> > + wait_for_completion(&orig->pr_comp);
> >
>
> > + kref_put(&orig->pr_kref, target_pr_kref_release);
> > /*
> > - * Disable struct se_dev_entry LUN ACL mapping
> > + * Before fireing off RCU callback, wait for any in process SPEC_I_PT=1
> > + * or REGISTER_AND_MOVE PR operation to complete.
> > */
> > + wait_for_completion(&orig->pr_comp);
> > + kfree_rcu(orig, rcu_head);
>
> The release callback should just call kfree_rcu, no need to wait for the
> release in the caller.
>

Why doesn't se_dev_entry release this need to wait for the special case
references to drop..?

> Also can you drop the _pr from the name? It's a generic refcount now
> even if the PR code is the only consumer so far.
>
> > +void target_pr_kref_release(struct kref *);
>
> Instead of exporting the release function it would be much more obvious
> to have a
>
> void target_deve_put(struct se_dev_entry *deve)
> {
> kref_put(&deve->pr_kref, target_deve_release);
> }
>
> helper. Probably paired with one for the get side.
>

Sure. Adding this now.

> > static void core_scsi3_lunacl_undepend_item(struct se_dev_entry *se_deve)
> > {
> > - struct se_lun_acl *lun_acl = se_deve->se_lun_acl;
> > + struct se_lun_acl *lun_acl;
> > struct se_node_acl *nacl;
> > struct se_portal_group *tpg;
> > +
> > + if (!se_deve) {
> > + pr_err("core_scsi3_lunacl_undepend_item passed NULL se_deve\n");
> > + dump_stack();
> > + return;
> > + }
>
> How could this happen and how is it related to this patch?
>

Dropped.

> > - if (!deve->se_lun || !deve->se_lun_acl) {
> > - spin_unlock_irq(&nacl->device_list_lock);
> > + rcu_read_lock();
> > + deve = target_nacl_find_deve(nacl, lacl->mapped_lun);
> > + if (!deve) {
> > + rcu_read_unlock();
> > return -ENODEV;
>
> So previously a lot of these files returned -ENODEV when not having
> an explicit node ACL, and now they don't. If that was intentional
> it should be documented in the changelog, or preferably moved into
> a preparation patch of its own.
>

Ok, will update the changelog for this.

> > + struct se_node_acl *se_node_acl;
>
> Where is this field coming from? It's not documented in the changelog
> and doesn't seem to be actually used either.
>

Nice catch. Dropping this unused pointer now.



\
 
 \ /
  Last update: 2015-05-22 11:21    [W:0.069 / U:1.472 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site