[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH V4 3/4] Introduce XEN scsiback module



On 08/12/2014 11:13 PM, Nicholas A. Bellinger wrote:
Hi Juergen & Co,

Finally had a chance to review this code.  Comments are inline below..

Thank you very much for your review!


On Fri, 2014-08-08 at 09:49 +0200, jgross@xxxxxxxx wrote:
From: Juergen Gross <jgross@xxxxxxxx>

Introduces the XEN pvSCSI backend. With pvSCSI it is possible for a XEN domU
to issue SCSI commands to a SCSI LUN assigned to that domU. The SCSI commands
are passed to the pvSCSI backend in a driver domain (usually Dom0) which is
owner of the physical device. This allows e.g. to use SCSI tape drives in a
XEN domU.

The code is taken from the pvSCSI implementation in XEN done by Fujitsu based
on Linux kernel 2.6.18.

Changes from the original version are:
- port to upstream kernel
- put all code in just one source file
- adapt to Linux style guide
- use target core infrastructure instead doing pure pass-through
- enable module unloading
- support SG-list in grant page(s)
- support task abort
- remove redundant struct backend
- allocate resources dynamically
- correct minor error in scsiback_fast_flush_area
- free allocated resources in case of error during I/O preparation
- remove CDB emulation, now handled by target core infrastructure

Signed-off-by: Juergen Gross <jgross@xxxxxxxx>

Xen related parts
Acked-by: David Vrabel <david.vrabel@xxxxxxxxxx>
---

<SNIP>

diff --git a/drivers/xen/xen-scsiback.c b/drivers/xen/xen-scsiback.c
new file mode 100644
index 0000000..4a0d6e3
--- /dev/null
+++ b/drivers/xen/xen-scsiback.c

<SNIP>

+struct scsiback_nacl {
+       /* Binary World Wide unique Port Name for pvscsi Initiator port */
+       u64 iport_wwpn;
+       /* ASCII formatted WWPN for Sas Initiator port */
+       char iport_name[VSCSI_NAMELEN];
+       /* Returned by scsiback_make_nodeacl() */
+       struct se_node_acl se_node_acl;
+};
+

Given that this code is similar to how loopback + vhost-scsi function,
and uses a (locally) generated nexus for each WWPN endpoint, the
scsiback_nacl and associated code will be unused should be be dropped.

Done.


<SNIP>

+static void scsiback_do_resp_with_sense(char *sense_buffer, int32_t result,
+                       uint32_t resid, struct vscsibk_pend *pending_req)
+{
+       struct vscsiif_response *ring_res;
+       struct vscsibk_info *info = pending_req->info;
+       int notify;
+       struct scsi_sense_hdr sshdr;
+       unsigned long flags;
+       unsigned len;
+
+       spin_lock_irqsave(&info->ring_lock, flags);
+
+       ring_res = RING_GET_RESPONSE(&info->ring, info->ring.rsp_prod_pvt);
+       info->ring.rsp_prod_pvt++;
+
+       ring_res->rslt   = result;
+       ring_res->rqid   = pending_req->rqid;
+
+       if (sense_buffer != NULL &&
+           scsi_normalize_sense(sense_buffer, VSCSIIF_SENSE_BUFFERSIZE,
+                                &sshdr)) {
+               len = min_t(unsigned, 8 + sense_buffer[7],
+                           VSCSIIF_SENSE_BUFFERSIZE);
+               memcpy(ring_res->sense_buffer, sense_buffer, len);
+               ring_res->sense_len = len;
+       } else {
+               ring_res->sense_len = 0;
+       }
+
+       ring_res->residual_len = resid;
+
+       RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&info->ring, notify);
+       spin_unlock_irqrestore(&info->ring_lock, flags);
+
+       if (notify)
+               notify_remote_via_irq(info->irq);
+
+       if (pending_req->v2p)
+               kref_put(&pending_req->v2p->kref,
+                        scsiback_free_translation_entry);
+
+       kmem_cache_free(scsiback_cachep, pending_req);
+}
+
+static void scsiback_cmd_done(struct vscsibk_pend *pending_req)
+{
+       struct vscsibk_info *info = pending_req->info;
+       unsigned char *sense_buffer;
+       unsigned int resid;
+       int errors;
+
+       sense_buffer = pending_req->sense_buffer;
+       resid        = pending_req->se_cmd.residual_count;
+       errors       = pending_req->result;
+
+       if (errors && log_print_stat)
+               scsiback_print_status(sense_buffer, errors, pending_req);
+
+       scsiback_fast_flush_area(pending_req);
+       scsiback_do_resp_with_sense(sense_buffer, errors, resid, pending_req);
+       scsiback_put(info);
+
+       transport_generic_free_cmd(&pending_req->se_cmd, 0);
+}
+

The usage here of scsiback_do_resp_with_sense() -> kmem_cache_free() for
*pending_req, and then invoking transport_generic_free_cmd() with
&pending_req->se_cmd is an free after use bug..

So the way this should work is similar to how loopback currently does
things:

   - Move the kmem_cache_free() for pending_req from
     scsiback_do_resp_with_sense() to scsiback_release_cmd()
   - Remove the transport_generic_free_cmd() from scsiback_cmd_done()
   - Copy what tcm_loop_check_stop_free() does into
     scsiback_check_stop_free(), and remove target_put_sess_cmd()

Done.


+static void scsiback_cmd_exec(struct vscsibk_pend *pending_req)
+{
+       struct se_cmd *se_cmd = &pending_req->se_cmd;
+       struct se_session *sess = pending_req->v2p->tpg->tpg_nexus->tvn_se_sess;
+       int rc;
+
+       memset(pending_req->sense_buffer, 0, VSCSIIF_SENSE_BUFFERSIZE);
+
+       memset(se_cmd, 0, sizeof(*se_cmd));
+       se_cmd->prot_pto = true;
+

No need to set prot_pto = true, as T10_PI support is not enabled..

Removed.


+       scsiback_get(pending_req->info);
+       rc = target_submit_cmd_map_sgls(se_cmd, sess, pending_req->cmnd,
+                       pending_req->sense_buffer, pending_req->v2p->lun,
+                       pending_req->data_len, 0,
+                       pending_req->sc_data_direction, 0,
+                       pending_req->sgl, pending_req->n_sg,
+                       NULL, 0, NULL, 0);
+       if (rc < 0) {
+               transport_send_check_condition_and_sense(se_cmd,
+                               TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE, 0);
+               transport_generic_free_cmd(se_cmd, 0);
+       }
+}
+

<SNIP>

+
+static void scsiback_device_action(struct vscsibk_pend *pending_req,
+       enum tcm_tmreq_table act, int tag)
+{
+       int rc, err = FAILED;
+       struct scsiback_tpg *tpg = pending_req->v2p->tpg;
+       struct se_cmd *se_cmd = &pending_req->se_cmd;
+       struct scsiback_tmr *tmr;
+
+       tmr = kzalloc(sizeof(struct scsiback_tmr), GFP_KERNEL);
+       if (!tmr) {
+               pr_err("xen-pvscsi: %s: kmalloc() error\n", __func__);
+               goto out;
+       }
+       init_waitqueue_head(&tmr->tmr_wait);
+
+       transport_init_se_cmd(se_cmd, tpg->se_tpg.se_tpg_tfo,
+               tpg->tpg_nexus->tvn_se_sess, 0, DMA_NONE, MSG_SIMPLE_TAG,
+               &pending_req->sense_buffer[0]);
+
+       rc = core_tmr_alloc_req(se_cmd, tmr, act, GFP_KERNEL);
+       if (rc < 0)
+               goto out;
+
+       se_cmd->se_tmr_req->ref_task_tag = tag;
+
+       if (transport_lookup_tmr_lun(se_cmd, pending_req->v2p->lun) < 0)
+               goto out;
+
+       transport_generic_handle_tmr(se_cmd);
+       wait_event(tmr->tmr_wait, atomic_read(&tmr->tmr_complete));
+
+       err = (se_cmd->se_tmr_req->response == TMR_FUNCTION_COMPLETE) ?
+               SUCCESS : FAILED;
+
+out:
+       if (tmr) {
+               transport_generic_free_cmd(&pending_req->se_cmd, 1);
+               kfree(tmr);
+       }
+
+       scsiback_do_resp_with_sense(NULL, err, 0, pending_req);
+}

With scsiback_do_resp_with_sense() no longer freeing pending_req
directly, this special case will need a kmem_cache_free()

Added.


+static int _scsiback_do_cmd_fn(struct vscsibk_info *info)
+{
+       struct vscsiif_back_ring *ring = &info->ring;
+       struct vscsiif_request *ring_req;
+       struct vscsibk_pend *pending_req;
+       RING_IDX rc, rp;
+       int err, more_to_do = 0;
+       uint32_t result;
+       uint8_t act;
+
+       rc = ring->req_cons;
+       rp = ring->sring->req_prod;
+       rmb();  /* guest system is accessing ring, too */
+
+       if (RING_REQUEST_PROD_OVERFLOW(ring, rp)) {
+               rc = ring->rsp_prod_pvt;
+               pr_warn("xen-pvscsi: Dom%d provided bogus ring requests (%#x - %#x = 
%u). Halting ring processing\n",
+                          info->domid, rp, rc, rp - rc);
+               return -EACCES;
+       }
+
+       while ((rc != rp)) {
+               if (RING_REQUEST_CONS_OVERFLOW(ring, rc))
+                       break;
+               pending_req = kmem_cache_alloc(scsiback_cachep, GFP_KERNEL);
+               if (NULL == pending_req) {
+                       more_to_do = 1;
+                       break;
+               }
+

Ideally this will end up using percpu_ida descriptor pre-allocation via
se_sess->sess_tag_poll + se_sess->sess_cmd_map in order to avoid the
fast-path memory allocation.

See commit 4824d3bfb909 for an example of how this was done for
vhost-scsi, but should be considered a post-merge optimization..

Okay, put on my todo list.


+               ring_req = RING_GET_REQUEST(ring, rc);
+               ring->req_cons = ++rc;
+
+               act = ring_req->act;
+               err = prepare_pending_reqs(info, ring_req, pending_req);
+               if (err) {
+                       switch (err) {
+                       case -ENODEV:
+                               result = DID_NO_CONNECT;
+                               break;
+                       default:
+                               result = DRIVER_ERROR;
+                               break;
+                       }
+                       scsiback_do_resp_with_sense(NULL, result << 24, 0,
+                                                   pending_req);

kmem_cache_free() added.

+                       more_to_do = 1;
+                       break;
+               }
+
+               switch (act) {
+               case VSCSIIF_ACT_SCSI_CDB:
+                       if (scsiback_gnttab_data_map(ring_req, pending_req)) {
+                               scsiback_fast_flush_area(pending_req);
+                               scsiback_do_resp_with_sense(NULL,
+                                       DRIVER_ERROR << 24, 0, pending_req);

Another special case for scsiback_do_resp_with_sense() that with the
recommended changes require a kmem_cache_free() call.

And here.


+                       } else {
+                               scsiback_cmd_exec(pending_req);
+                       }
+                       break;
+               case VSCSIIF_ACT_SCSI_ABORT:
+                       scsiback_device_action(pending_req, TMR_ABORT_TASK,
+                               ring_req->ref_rqid);
+                       break;
+               case VSCSIIF_ACT_SCSI_RESET:
+                       scsiback_device_action(pending_req, TMR_LUN_RESET, 0);
+                       break;
+               default:
+                       pr_err_ratelimited("xen-pvscsi: invalid request\n");
+                       scsiback_do_resp_with_sense(NULL, DRIVER_ERROR << 24,
+                                                   0, pending_req);

Ditto here.

And here.


+                       break;
+               }
+
+               /* Yield point for this unbounded loop. */
+               cond_resched();
+       }
+
+       if (RING_HAS_UNCONSUMED_REQUESTS(ring))
+               more_to_do = 1;
+
+       return more_to_do;
+}
+

<SNIP>

+static struct se_node_acl *
+scsiback_alloc_fabric_acl(struct se_portal_group *se_tpg)
+{
+       struct scsiback_nacl *nacl;
+
+       nacl = kzalloc(sizeof(struct scsiback_nacl), GFP_KERNEL);
+       if (!nacl) {
+               pr_err("Unable to allocate struct scsiback_nacl\n");
+               return NULL;
+       }
+
+       return &nacl->se_node_acl;
+}
+
+static void
+scsiback_release_fabric_acl(struct se_portal_group *se_tpg,
+                            struct se_node_acl *se_nacl)
+{
+       struct scsiback_nacl *nacl = container_of(se_nacl,
+                       struct scsiback_nacl, se_node_acl);
+       kfree(nacl);
+}
+
+static u32 scsiback_tpg_get_inst_index(struct se_portal_group *se_tpg)
+{
+       return 1;
+}
+
+static struct se_node_acl *
+scsiback_make_nodeacl(struct se_portal_group *se_tpg,
+                      struct config_group *group,
+                      const char *name)
+{
+       struct se_node_acl *se_nacl, *se_nacl_new;
+       struct scsiback_nacl *nacl;
+       u64 wwpn = 0;
+       u32 nexus_depth;
+
+       se_nacl_new = scsiback_alloc_fabric_acl(se_tpg);
+       if (!se_nacl_new)
+               return ERR_PTR(-ENOMEM);
+
+       nexus_depth = 1;
+       /*
+        * se_nacl_new may be released by core_tpg_add_initiator_node_acl()
+        * when converting a NodeACL from demo mode -> explict
+        */
+       se_nacl = core_tpg_add_initiator_node_acl(se_tpg, se_nacl_new,
+                               name, nexus_depth);
+       if (IS_ERR(se_nacl)) {
+               scsiback_release_fabric_acl(se_tpg, se_nacl_new);
+               return se_nacl;
+       }
+       /*
+        * Locate our struct scsiback_nacl and set the FC Nport WWPN
+        */
+       nacl = container_of(se_nacl, struct scsiback_nacl, se_node_acl);
+       nacl->iport_wwpn = wwpn;
+
+       return se_nacl;
+}
+
+static void scsiback_drop_nodeacl(struct se_node_acl *se_acl)
+{
+       struct scsiback_nacl *nacl = container_of(se_acl,
+                               struct scsiback_nacl, se_node_acl);
+       core_tpg_del_initiator_node_acl(se_acl->se_tpg, se_acl, 1);
+       kfree(nacl);
+}
+

As mentioned above, the NodeACL use is unnecessary for this driver so
you can safely drop scsiback_make_node_acl() +
scsiback_alloc_fabric_acl() + scsiback_drop_nodeacl() +
scsiback_release_fabric_acl().

Deleted.


+static int scsiback_check_stop_free(struct se_cmd *se_cmd)
+{
+       return target_put_sess_cmd(se_cmd->se_sess, se_cmd);
+}
+

As mentioned above, scsiback_check_stop_free() should follow what
tcm_loop_check_stop_free() does here..

Yep.


+static void scsiback_release_cmd(struct se_cmd *se_cmd)
+{
+}
+

Move the kmem_cache_free() for pending_req here..

Done.


+static int scsiback_shutdown_session(struct se_session *se_sess)
+{
+       return 0;
+}
+
+static void scsiback_close_session(struct se_session *se_sess)
+{
+}
+
+static u32 scsiback_sess_get_index(struct se_session *se_sess)
+{
+       return 0;
+}
+
+static int scsiback_write_pending(struct se_cmd *se_cmd)
+{
+       /* Go ahead and process the write immediately */
+       target_execute_cmd(se_cmd);
+
+       return 0;
+}
+
+static int scsiback_write_pending_status(struct se_cmd *se_cmd)
+{
+       return 0;
+}
+
+static void scsiback_set_default_node_attrs(struct se_node_acl *nacl)
+{
+}
+

Safe to drop this no-op too.

Okay.


+static u32 scsiback_get_task_tag(struct se_cmd *se_cmd)
+{
+       struct vscsibk_pend *pending_req = container_of(se_cmd,
+                               struct vscsibk_pend, se_cmd);
+
+       return pending_req->rqid;
+}
+
+static int scsiback_get_cmd_state(struct se_cmd *se_cmd)
+{
+       return 0;
+}
+
+static int scsiback_queue_data_in(struct se_cmd *se_cmd)
+{
+       struct vscsibk_pend *pending_req = container_of(se_cmd,
+                               struct vscsibk_pend, se_cmd);
+
+       pending_req->result = SAM_STAT_GOOD;
+       scsiback_cmd_done(pending_req);
+       return 0;
+}
+
+static int scsiback_queue_status(struct se_cmd *se_cmd)
+{
+       struct vscsibk_pend *pending_req = container_of(se_cmd,
+                               struct vscsibk_pend, se_cmd);
+
+       if (se_cmd->sense_buffer &&
+           ((se_cmd->se_cmd_flags & SCF_TRANSPORT_TASK_SENSE) ||
+            (se_cmd->se_cmd_flags & SCF_EMULATED_TASK_SENSE)))
+               pending_req->result = (DRIVER_SENSE << 24) |
+                                     SAM_STAT_CHECK_CONDITION;
+       else
+               pending_req->result = se_cmd->scsi_status;
+
+       scsiback_cmd_done(pending_req);
+       return 0;
+}
+
+static void scsiback_queue_tm_rsp(struct se_cmd *se_cmd)
+{
+       struct se_tmr_req *se_tmr = se_cmd->se_tmr_req;
+       struct scsiback_tmr *tmr = se_tmr->fabric_tmr_ptr;
+
+       atomic_set(&tmr->tmr_complete, 1);
+       wake_up(&tmr->tmr_wait);
+}
+
+static void scsiback_aborted_task(struct se_cmd *se_cmd)
+{
+       struct se_tmr_req *se_tmr = se_cmd->se_tmr_req;
+       struct scsiback_tmr *tmr = se_tmr->fabric_tmr_ptr;
+
+       atomic_set(&tmr->tmr_complete, 1);
+       wake_up(&tmr->tmr_wait);
+}
+

scsiback_aborted_task() should just return here, instead of the
atomic_set + wake_up().  Eg: ABORT_TASK is already calling ->
queue_tm_rsp(), and the ->aborted_task() callback is specifically for
cleaning up any fabric specific descriptor resources, before
->release_cmd() gets called.

Okay.


+static ssize_t scsiback_tpg_param_show_alias(struct se_portal_group *se_tpg,
+                                            char *page)
+{
+       struct scsiback_tpg *tpg = container_of(se_tpg, struct scsiback_tpg,
+                                               se_tpg);
+       ssize_t rb;
+
+       mutex_lock(&tpg->tv_tpg_mutex);
+       rb = snprintf(page, PAGE_SIZE, "%s\n", tpg->param_alias);
+       mutex_unlock(&tpg->tv_tpg_mutex);
+
+       return rb;
+}
+
+static ssize_t scsiback_tpg_param_store_alias(struct se_portal_group *se_tpg,
+                                             const char *page, size_t count)
+{
+       struct scsiback_tpg *tpg = container_of(se_tpg, struct scsiback_tpg,
+                                               se_tpg);
+       int len;
+
+       if (strlen(page) >= VSCSI_NAMELEN) {
+               pr_err("param alias: %s, exceeds max: %d\n", page,
+                       VSCSI_NAMELEN);
+               return -EINVAL;
+       }
+
+       mutex_lock(&tpg->tv_tpg_mutex);
+       len = snprintf(tpg->param_alias, VSCSI_NAMELEN, "%s", page);
+       if (tpg->param_alias[len - 1] == '\n')
+               tpg->param_alias[len - 1] = '\0';
+       mutex_unlock(&tpg->tv_tpg_mutex);
+
+       return count;
+}
+
+TF_TPG_PARAM_ATTR(scsiback, alias, S_IRUGO | S_IWUSR);
+
+static struct configfs_attribute *scsiback_param_attrs[] = {
+       &scsiback_tpg_param_alias.attr,
+       NULL,
+};
+
+static int scsiback_make_nexus(struct scsiback_tpg *tpg,
+                               const char *name)
+{
+       struct se_portal_group *se_tpg;
+       struct se_session *se_sess;
+       struct scsiback_nexus *tv_nexus;
+
+       mutex_lock(&tpg->tv_tpg_mutex);
+       if (tpg->tpg_nexus) {
+               mutex_unlock(&tpg->tv_tpg_mutex);
+               pr_debug("tpg->tpg_nexus already exists\n");
+               return -EEXIST;
+       }
+       se_tpg = &tpg->se_tpg;
+
+       tv_nexus = kzalloc(sizeof(struct scsiback_nexus), GFP_KERNEL);
+       if (!tv_nexus) {
+               mutex_unlock(&tpg->tv_tpg_mutex);
+               pr_err("Unable to allocate struct scsiback_nexus\n");
+               return -ENOMEM;
+       }
+       /*
+        *  Initialize the struct se_session pointer
+        */
+       tv_nexus->tvn_se_sess = transport_init_session(TARGET_PROT_DIN_PASS |
+                                                      TARGET_PROT_DOUT_PASS);

TARGET_PROT_D*_PASS incorrectly declares T10 PI support here, which
should be TARGET_PROT_NORMAL instead.

Okay.


<SNIP>

+
+static int scsiback_port_link(struct se_portal_group *se_tpg,
+                              struct se_lun *lun)
+{
+       struct scsiback_tpg *tpg = container_of(se_tpg,
+                               struct scsiback_tpg, se_tpg);
+
+       mutex_lock(&scsiback_mutex);
+
+       mutex_lock(&tpg->tv_tpg_mutex);
+       tpg->tv_tpg_port_count++;
+       mutex_unlock(&tpg->tv_tpg_mutex);
+
+       mutex_unlock(&scsiback_mutex);
+
+       return 0;
+}
+

AFAICT, no need to hold scsiback_mutex while incrementing
tpg->tv_tpg_port_count.

So there is a guarantee that port_link and port_unlink are never
called in parallel?


+static void scsiback_port_unlink(struct se_portal_group *se_tpg,
+                                 struct se_lun *lun)
+{
+       struct scsiback_tpg *tpg = container_of(se_tpg,
+                               struct scsiback_tpg, se_tpg);
+
+       mutex_lock(&scsiback_mutex);
+
+       mutex_lock(&tpg->tv_tpg_mutex);
+       tpg->tv_tpg_port_count--;
+       mutex_unlock(&tpg->tv_tpg_mutex);
+
+       mutex_unlock(&scsiback_mutex);
+}
+

Ditto here.


Juergen


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.