tcm_fc: Convert to per-cpu command map pre-allocation of ft_cmd

This patch converts tcm_fc to use transport_init_session_tags()
pre-allocation logic for struct ft_cmd descriptors using per-cpu
session tag pooling in order to effectively avoid memory allocation
+ release for each received I/O.

It adds percpu_ida_alloc() in ft_recv_cmd() to obtain an tag and
locate ft_cmd from se_sess->sess_cmd_map[], and percpu_ida_free()
in ft_free_cmd() to release the tag based upon se_cmd->map_tag id.

It also uses a TCM_FC_DEFAULT_TAGS value of 512, that puts the
per se_sess->sess_cmd_map allocation at ~360K on 64-bit.

v2 changes:

  - Handle possible tag < 0 failure with GFP_ATOMIC

Cc: Mark Rustad <mark.d.rustad@intel.com>
Cc: Robert Love <robert.w.love@intel.com>
Cc: Kent Overstreet <kmo@daterainc.com>
Signed-off-by: Nicholas Bellinger <nab@daterainc.com>

drivers/target/tcm_fc/tfc_sess.c

@@ -210,7 +210,8 @@ static struct ft_sess *ft_sess_create(struct ft_tport *tport, u32 port_id,
 	if (!sess)
 		return NULL;
 
-	sess->se_sess = transport_init_session();
+	sess->se_sess = transport_init_session_tags(TCM_FC_DEFAULT_TAGS,
+						    sizeof(struct ft_cmd));
 	if (IS_ERR(sess->se_sess)) {
 		kfree(sess);
 		return NULL;