skd: Enable request tags for the block layer queue
authorBart Van Assche <bart.vanassche@wdc.com>
Thu, 17 Aug 2017 20:13:26 +0000 (13:13 -0700)
committerJens Axboe <axboe@kernel.dk>
Fri, 18 Aug 2017 14:45:29 +0000 (08:45 -0600)
Use the request tag when allocating a skd_fitmsg_context or
skd_request_context such that the lists used to track free elements
can be eliminated. Swap the skd_end_request() and skd_release_req()
calls to avoid triggering a use-after-free. Remove
skd_fitmsg_context.state and .outstanding because FIT messages are
shared among requests and because updating a FIT message after a
request has finished whould trigger a use-after-free.

Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.de>
Cc: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
drivers/block/skd_main.c

index 392c898d86e2b4ddde208cd795d5fd737f88ee33..35343fbf41440b787e02d97263f316cfc0689727 100644 (file)
@@ -16,6 +16,7 @@
 #include <linux/slab.h>
 #include <linux/spinlock.h>
 #include <linux/blkdev.h>
+#include <linux/blk-mq.h>
 #include <linux/sched.h>
 #include <linux/interrupt.h>
 #include <linux/compiler.h>
@@ -154,11 +155,6 @@ enum skd_req_state {
        SKD_REQ_STATE_TIMEOUT,
 };
 
-enum skd_fit_msg_state {
-       SKD_MSG_STATE_IDLE,
-       SKD_MSG_STATE_BUSY,
-};
-
 enum skd_check_status_action {
        SKD_CHECK_STATUS_REPORT_GOOD,
        SKD_CHECK_STATUS_REPORT_SMART_ALERT,
@@ -173,12 +169,7 @@ struct skd_msg_buf {
 };
 
 struct skd_fitmsg_context {
-       enum skd_fit_msg_state state;
-
-       struct skd_fitmsg_context *next;
-
        u32 id;
-       u16 outstanding;
 
        u32 length;
 
@@ -189,8 +180,6 @@ struct skd_fitmsg_context {
 struct skd_request_context {
        enum skd_req_state state;
 
-       struct skd_request_context *next;
-
        u16 id;
        u32 fitmsg_id;
 
@@ -264,10 +253,8 @@ struct skd_device {
 
        u32 timeout_slot[SKD_N_TIMEOUT_SLOT];
        u32 timeout_stamp;
-       struct skd_fitmsg_context *skmsg_free_list;
        struct skd_fitmsg_context *skmsg_table;
 
-       struct skd_request_context *skreq_free_list;
        struct skd_request_context *skreq_table;
 
        struct skd_special_context internal_skspcl;
@@ -387,8 +374,8 @@ static void skd_send_fitmsg(struct skd_device *skdev,
 static void skd_send_special_fitmsg(struct skd_device *skdev,
                                    struct skd_special_context *skspcl);
 static void skd_request_fn(struct request_queue *rq);
-static void skd_end_request(struct skd_device *skdev,
-               struct skd_request_context *skreq, blk_status_t status);
+static void skd_end_request(struct skd_device *skdev, struct request *req,
+                           blk_status_t status);
 static bool skd_preop_sg_list(struct skd_device *skdev,
                             struct skd_request_context *skreq);
 static void skd_postop_sg_list(struct skd_device *skdev,
@@ -405,8 +392,6 @@ static void skd_soft_reset(struct skd_device *skdev);
 const char *skd_drive_state_to_str(int state);
 const char *skd_skdev_state_to_str(enum skd_drvr_state state);
 static void skd_log_skdev(struct skd_device *skdev, const char *event);
-static void skd_log_skmsg(struct skd_device *skdev,
-                         struct skd_fitmsg_context *skmsg, const char *event);
 static void skd_log_skreq(struct skd_device *skdev,
                          struct skd_request_context *skreq, const char *event);
 
@@ -424,7 +409,7 @@ static void skd_fail_all_pending(struct skd_device *skdev)
                req = blk_peek_request(q);
                if (req == NULL)
                        break;
-               blk_start_request(req);
+               WARN_ON_ONCE(blk_queue_start_tag(q, req));
                __blk_end_request_all(req, BLK_STS_IOERR);
        }
 }
@@ -523,6 +508,7 @@ static void skd_request_fn(struct request_queue *q)
        u64 cmdctxt;
        u32 timo_slot;
        int flush, fua;
+       u32 tag;
 
        if (skdev->state != SKD_DRVR_STATE_ONLINE) {
                if (skd_fail_all(q))
@@ -531,9 +517,7 @@ static void skd_request_fn(struct request_queue *q)
        }
 
        if (blk_queue_stopped(skdev->queue)) {
-               if (skdev->skmsg_free_list == NULL ||
-                   skdev->skreq_free_list == NULL ||
-                   skdev->in_flight >= skdev->queue_low_water_mark)
+               if (skdev->in_flight >= skdev->queue_low_water_mark)
                        /* There is still some kind of shortage */
                        return;
 
@@ -581,27 +565,6 @@ static void skd_request_fn(struct request_queue *q)
                        break;
                }
 
-               /* Is a skd_request_context available? */
-               skreq = skdev->skreq_free_list;
-               if (skreq == NULL) {
-                       dev_dbg(&skdev->pdev->dev, "Out of req=%p\n", q);
-                       break;
-               }
-               SKD_ASSERT(skreq->state == SKD_REQ_STATE_IDLE);
-               SKD_ASSERT((skreq->id & SKD_ID_INCR) == 0);
-
-               /* Now we check to see if we can get a fit msg */
-               if (skmsg == NULL) {
-                       if (skdev->skmsg_free_list == NULL) {
-                               dev_dbg(&skdev->pdev->dev, "Out of msg\n");
-                               break;
-                       }
-               }
-
-               skreq->flush_cmd = 0;
-               skreq->n_sg = 0;
-               skreq->sg_byte_count = 0;
-
                /*
                 * OK to now dequeue request from q.
                 *
@@ -609,7 +572,22 @@ static void skd_request_fn(struct request_queue *q)
                 * the native request. Note that skd_request_context is
                 * available but is still at the head of the free list.
                 */
-               blk_start_request(req);
+               WARN_ON_ONCE(blk_queue_start_tag(q, req));
+
+               tag = blk_mq_unique_tag(req);
+               WARN_ONCE(tag >= skd_max_queue_depth,
+                         "%#x > %#x (nr_requests = %lu)\n", tag,
+                         skd_max_queue_depth, q->nr_requests);
+
+               skreq = &skdev->skreq_table[tag];
+               SKD_ASSERT(skreq->state == SKD_REQ_STATE_IDLE);
+               SKD_ASSERT((skreq->id & SKD_ID_INCR) == 0);
+
+               skreq->id = tag + SKD_ID_RW_REQUEST;
+               skreq->flush_cmd = 0;
+               skreq->n_sg = 0;
+               skreq->sg_byte_count = 0;
+
                skreq->req = req;
                skreq->fitmsg_id = 0;
 
@@ -618,27 +596,13 @@ static void skd_request_fn(struct request_queue *q)
 
                if (req->bio && !skd_preop_sg_list(skdev, skreq)) {
                        dev_dbg(&skdev->pdev->dev, "error Out\n");
-                       skd_end_request(skdev, skreq, BLK_STS_RESOURCE);
+                       skd_end_request(skdev, skreq->req, BLK_STS_RESOURCE);
                        continue;
                }
 
                /* Either a FIT msg is in progress or we have to start one. */
                if (skmsg == NULL) {
-                       /* Are there any FIT msg buffers available? */
-                       skmsg = skdev->skmsg_free_list;
-                       if (skmsg == NULL) {
-                               dev_dbg(&skdev->pdev->dev,
-                                       "Out of msg skdev=%p\n",
-                                       skdev);
-                               break;
-                       }
-                       SKD_ASSERT(skmsg->state == SKD_MSG_STATE_IDLE);
-                       SKD_ASSERT((skmsg->id & SKD_ID_INCR) == 0);
-
-                       skdev->skmsg_free_list = skmsg->next;
-
-                       skmsg->state = SKD_MSG_STATE_BUSY;
-                       skmsg->id += SKD_ID_INCR;
+                       skmsg = &skdev->skmsg_table[tag];
 
                        /* Initialize the FIT msg header */
                        fmh = &skmsg->msg_buf->fmh;
@@ -673,7 +637,6 @@ static void skd_request_fn(struct request_queue *q)
                        cpu_to_be32(skreq->sg_byte_count);
 
                /* Complete resource allocations. */
-               skdev->skreq_free_list = skreq->next;
                skreq->state = SKD_REQ_STATE_BUSY;
                skreq->id += SKD_ID_INCR;
 
@@ -717,23 +680,22 @@ static void skd_request_fn(struct request_queue *q)
                blk_stop_queue(skdev->queue);
 }
 
-static void skd_end_request(struct skd_device *skdev,
-               struct skd_request_context *skreq, blk_status_t error)
+static void skd_end_request(struct skd_device *skdev, struct request *req,
+                           blk_status_t error)
 {
        if (unlikely(error)) {
-               struct request *req = skreq->req;
                char *cmd = (rq_data_dir(req) == READ) ? "read" : "write";
                u32 lba = (u32)blk_rq_pos(req);
                u32 count = blk_rq_sectors(req);
 
                dev_err(&skdev->pdev->dev,
                        "Error cmd=%s sect=%u count=%u id=0x%x\n", cmd, lba,
-                       count, skreq->id);
+                       count, req->tag);
        } else
-               dev_dbg(&skdev->pdev->dev, "id=0x%x error=%d\n", skreq->id,
+               dev_dbg(&skdev->pdev->dev, "id=0x%x error=%d\n", req->tag,
                        error);
 
-       __blk_end_request_all(skreq->req, error);
+       __blk_end_request_all(req, error);
 }
 
 static bool skd_preop_sg_list(struct skd_device *skdev,
@@ -1346,7 +1308,6 @@ static void skd_send_fitmsg(struct skd_device *skdev,
                            struct skd_fitmsg_context *skmsg)
 {
        u64 qcmd;
-       struct fit_msg_hdr *fmh;
 
        dev_dbg(&skdev->pdev->dev, "dma address 0x%llx, busy=%d\n",
                skmsg->mb_dma_address, skdev->in_flight);
@@ -1355,9 +1316,6 @@ static void skd_send_fitmsg(struct skd_device *skdev,
        qcmd = skmsg->mb_dma_address;
        qcmd |= FIT_QCMD_QID_NORMAL;
 
-       fmh = &skmsg->msg_buf->fmh;
-       skmsg->outstanding = fmh->num_protocol_cmds_coalesced;
-
        if (unlikely(skdev->dbg_level > 1)) {
                u8 *bp = (u8 *)skmsg->msg_buf;
                int i;
@@ -1547,19 +1505,20 @@ skd_check_status(struct skd_device *skdev,
 }
 
 static void skd_resolve_req_exception(struct skd_device *skdev,
-                                     struct skd_request_context *skreq)
+                                     struct skd_request_context *skreq,
+                                     struct request *req)
 {
        u8 cmp_status = skreq->completion.status;
 
        switch (skd_check_status(skdev, cmp_status, &skreq->err_info)) {
        case SKD_CHECK_STATUS_REPORT_GOOD:
        case SKD_CHECK_STATUS_REPORT_SMART_ALERT:
-               skd_end_request(skdev, skreq, BLK_STS_OK);
+               skd_end_request(skdev, req, BLK_STS_OK);
                break;
 
        case SKD_CHECK_STATUS_BUSY_IMMINENT:
                skd_log_skreq(skdev, skreq, "retry(busy)");
-               blk_requeue_request(skdev->queue, skreq->req);
+               blk_requeue_request(skdev->queue, req);
                dev_info(&skdev->pdev->dev, "drive BUSY imminent\n");
                skdev->state = SKD_DRVR_STATE_BUSY_IMMINENT;
                skdev->timer_countdown = SKD_TIMER_MINUTES(20);
@@ -1567,16 +1526,16 @@ static void skd_resolve_req_exception(struct skd_device *skdev,
                break;
 
        case SKD_CHECK_STATUS_REQUEUE_REQUEST:
-               if ((unsigned long) ++skreq->req->special < SKD_MAX_RETRIES) {
+               if ((unsigned long) ++req->special < SKD_MAX_RETRIES) {
                        skd_log_skreq(skdev, skreq, "retry");
-                       blk_requeue_request(skdev->queue, skreq->req);
+                       blk_requeue_request(skdev->queue, req);
                        break;
                }
                /* fall through */
 
        case SKD_CHECK_STATUS_REPORT_ERROR:
        default:
-               skd_end_request(skdev, skreq, BLK_STS_IOERR);
+               skd_end_request(skdev, req, BLK_STS_IOERR);
                break;
        }
 }
@@ -1585,44 +1544,8 @@ static void skd_resolve_req_exception(struct skd_device *skdev,
 static void skd_release_skreq(struct skd_device *skdev,
                              struct skd_request_context *skreq)
 {
-       u32 msg_slot;
-       struct skd_fitmsg_context *skmsg;
-
        u32 timo_slot;
 
-       /*
-        * Reclaim the FIT msg buffer if this is
-        * the first of the requests it carried to
-        * be completed. The FIT msg buffer used to
-        * send this request cannot be reused until
-        * we are sure the s1120 card has copied
-        * it to its memory. The FIT msg might have
-        * contained several requests. As soon as
-        * any of them are completed we know that
-        * the entire FIT msg was transferred.
-        * Only the first completed request will
-        * match the FIT msg buffer id. The FIT
-        * msg buffer id is immediately updated.
-        * When subsequent requests complete the FIT
-        * msg buffer id won't match, so we know
-        * quite cheaply that it is already done.
-        */
-       msg_slot = skreq->fitmsg_id & SKD_ID_SLOT_MASK;
-       SKD_ASSERT(msg_slot < skdev->num_fitmsg_context);
-
-       skmsg = &skdev->skmsg_table[msg_slot];
-       if (skmsg->id == skreq->fitmsg_id) {
-               SKD_ASSERT(skmsg->state == SKD_MSG_STATE_BUSY);
-               SKD_ASSERT(skmsg->outstanding > 0);
-               skmsg->outstanding--;
-               if (skmsg->outstanding == 0) {
-                       skmsg->state = SKD_MSG_STATE_IDLE;
-                       skmsg->id += SKD_ID_INCR;
-                       skmsg->next = skdev->skmsg_free_list;
-                       skdev->skmsg_free_list = skmsg;
-               }
-       }
-
        /*
         * Decrease the number of active requests.
         * Also decrements the count in the timeout slot.
@@ -1644,8 +1567,20 @@ static void skd_release_skreq(struct skd_device *skdev,
         */
        skreq->state = SKD_REQ_STATE_IDLE;
        skreq->id += SKD_ID_INCR;
-       skreq->next = skdev->skreq_free_list;
-       skdev->skreq_free_list = skreq;
+}
+
+static struct skd_request_context *skd_skreq_from_rq(struct skd_device *skdev,
+                                                    struct request *rq)
+{
+       struct skd_request_context *skreq;
+       int i;
+
+       for (i = 0, skreq = skdev->skreq_table; i < skdev->num_fitmsg_context;
+            i++, skreq++)
+               if (skreq->req == rq)
+                       return skreq;
+
+       return NULL;
 }
 
 static int skd_isr_completion_posted(struct skd_device *skdev,
@@ -1654,7 +1589,8 @@ static int skd_isr_completion_posted(struct skd_device *skdev,
        struct fit_completion_entry_v1 *skcmp;
        struct fit_comp_error_info *skerr;
        u16 req_id;
-       u32 req_slot;
+       u32 tag;
+       struct request *rq;
        struct skd_request_context *skreq;
        u16 cmp_cntxt;
        u8 cmp_status;
@@ -1702,18 +1638,24 @@ static int skd_isr_completion_posted(struct skd_device *skdev,
                 * r/w request (see skd_start() above) or a special request.
                 */
                req_id = cmp_cntxt;
-               req_slot = req_id & SKD_ID_SLOT_AND_TABLE_MASK;
+               tag = req_id & SKD_ID_SLOT_AND_TABLE_MASK;
 
                /* Is this other than a r/w request? */
-               if (req_slot >= skdev->num_req_context) {
+               if (tag >= skdev->num_req_context) {
                        /*
                         * This is not a completion for a r/w request.
                         */
+                       WARN_ON_ONCE(blk_map_queue_find_tag(skdev->queue->
+                                                           queue_tags, tag));
                        skd_complete_other(skdev, skcmp, skerr);
                        continue;
                }
 
-               skreq = &skdev->skreq_table[req_slot];
+               rq = blk_map_queue_find_tag(skdev->queue->queue_tags, tag);
+               if (WARN(!rq, "No request for tag %#x -> %#x\n", cmp_cntxt,
+                        tag))
+                       continue;
+               skreq = skd_skreq_from_rq(skdev, rq);
 
                /*
                 * Make sure the request ID for the slot matches.
@@ -1745,26 +1687,16 @@ static int skd_isr_completion_posted(struct skd_device *skdev,
                if (skreq->n_sg > 0)
                        skd_postop_sg_list(skdev, skreq);
 
-               if (!skreq->req) {
-                       dev_dbg(&skdev->pdev->dev,
-                               "NULL backptr skdreq %p, req=0x%x req_id=0x%x\n",
-                               skreq, skreq->id, req_id);
-               } else {
-                       /*
-                        * Capture the outcome and post it back to the
-                        * native request.
-                        */
-                       if (likely(cmp_status == SAM_STAT_GOOD))
-                               skd_end_request(skdev, skreq, BLK_STS_OK);
-                       else
-                               skd_resolve_req_exception(skdev, skreq);
-               }
+               /* Mark the FIT msg and timeout slot as free. */
+               skd_release_skreq(skdev, skreq);
 
                /*
-                * Release the skreq, its FIT msg (if one), timeout slot,
-                * and queue depth.
+                * Capture the outcome and post it back to the native request.
                 */
-               skd_release_skreq(skdev, skreq);
+               if (likely(cmp_status == SAM_STAT_GOOD))
+                       skd_end_request(skdev, rq, BLK_STS_OK);
+               else
+                       skd_resolve_req_exception(skdev, skreq, rq);
 
                /* skd_isr_comp_limit equal zero means no limit */
                if (limit) {
@@ -2099,44 +2031,26 @@ static void skd_recover_requests(struct skd_device *skdev)
 
        for (i = 0; i < skdev->num_req_context; i++) {
                struct skd_request_context *skreq = &skdev->skreq_table[i];
+               struct request *req = skreq->req;
 
                if (skreq->state == SKD_REQ_STATE_BUSY) {
                        skd_log_skreq(skdev, skreq, "recover");
 
                        SKD_ASSERT((skreq->id & SKD_ID_INCR) != 0);
-                       SKD_ASSERT(skreq->req != NULL);
+                       SKD_ASSERT(req != NULL);
 
                        /* Release DMA resources for the request. */
                        if (skreq->n_sg > 0)
                                skd_postop_sg_list(skdev, skreq);
 
-                       skd_end_request(skdev, skreq, BLK_STS_IOERR);
-
                        skreq->req = NULL;
 
                        skreq->state = SKD_REQ_STATE_IDLE;
                        skreq->id += SKD_ID_INCR;
-               }
-               if (i > 0)
-                       skreq[-1].next = skreq;
-               skreq->next = NULL;
-       }
-       skdev->skreq_free_list = skdev->skreq_table;
-
-       for (i = 0; i < skdev->num_fitmsg_context; i++) {
-               struct skd_fitmsg_context *skmsg = &skdev->skmsg_table[i];
 
-               if (skmsg->state == SKD_MSG_STATE_BUSY) {
-                       skd_log_skmsg(skdev, skmsg, "salvaged");
-                       SKD_ASSERT((skmsg->id & SKD_ID_INCR) != 0);
-                       skmsg->state = SKD_MSG_STATE_IDLE;
-                       skmsg->id += SKD_ID_INCR;
+                       skd_end_request(skdev, req, BLK_STS_IOERR);
                }
-               if (i > 0)
-                       skmsg[-1].next = skmsg;
-               skmsg->next = NULL;
        }
-       skdev->skmsg_free_list = skdev->skmsg_table;
 
        for (i = 0; i < SKD_N_TIMEOUT_SLOT; i++)
                skdev->timeout_slot[i] = 0;
@@ -2876,7 +2790,6 @@ static int skd_cons_skmsg(struct skd_device *skdev)
 
                skmsg->id = i + SKD_ID_FIT_MSG;
 
-               skmsg->state = SKD_MSG_STATE_IDLE;
                skmsg->msg_buf = pci_alloc_consistent(skdev->pdev,
                                                      SKD_N_FITMSG_BYTES,
                                                      &skmsg->mb_dma_address);
@@ -2891,14 +2804,8 @@ static int skd_cons_skmsg(struct skd_device *skdev)
                     "not aligned: msg_buf %p mb_dma_address %#llx\n",
                     skmsg->msg_buf, skmsg->mb_dma_address);
                memset(skmsg->msg_buf, 0, SKD_N_FITMSG_BYTES);
-
-               skmsg->next = &skmsg[1];
        }
 
-       /* Free list is in order starting with the 0th entry. */
-       skdev->skmsg_table[i - 1].next = NULL;
-       skdev->skmsg_free_list = skdev->skmsg_table;
-
 err_out:
        return rc;
 }
@@ -2958,10 +2865,7 @@ static int skd_cons_skreq(struct skd_device *skdev)
                struct skd_request_context *skreq;
 
                skreq = &skdev->skreq_table[i];
-
-               skreq->id = i + SKD_ID_RW_REQUEST;
                skreq->state = SKD_REQ_STATE_IDLE;
-
                skreq->sg = kcalloc(skdev->sgs_per_request,
                                    sizeof(struct scatterlist), GFP_KERNEL);
                if (skreq->sg == NULL) {
@@ -2978,14 +2882,8 @@ static int skd_cons_skreq(struct skd_device *skdev)
                        rc = -ENOMEM;
                        goto err_out;
                }
-
-               skreq->next = &skreq[1];
        }
 
-       /* Free list is in order starting with the 0th entry. */
-       skdev->skreq_table[i - 1].next = NULL;
-       skdev->skreq_free_list = skdev->skreq_table;
-
 err_out:
        return rc;
 }
@@ -3061,6 +2959,8 @@ static int skd_cons_disk(struct skd_device *skdev)
                goto err_out;
        }
        blk_queue_bounce_limit(q, BLK_BOUNCE_HIGH);
+       q->nr_requests = skd_max_queue_depth / 2;
+       blk_queue_init_tags(q, skd_max_queue_depth, NULL, BLK_TAG_ALLOC_FIFO);
 
        skdev->queue = q;
        disk->queue = q;
@@ -3789,18 +3689,6 @@ const char *skd_skdev_state_to_str(enum skd_drvr_state state)
        }
 }
 
-static const char *skd_skmsg_state_to_str(enum skd_fit_msg_state state)
-{
-       switch (state) {
-       case SKD_MSG_STATE_IDLE:
-               return "IDLE";
-       case SKD_MSG_STATE_BUSY:
-               return "BUSY";
-       default:
-               return "???";
-       }
-}
-
 static const char *skd_skreq_state_to_str(enum skd_req_state state)
 {
        switch (state) {
@@ -3832,15 +3720,6 @@ static void skd_log_skdev(struct skd_device *skdev, const char *event)
                skdev->timeout_stamp, skdev->skcomp_cycle, skdev->skcomp_ix);
 }
 
-static void skd_log_skmsg(struct skd_device *skdev,
-                         struct skd_fitmsg_context *skmsg, const char *event)
-{
-       dev_dbg(&skdev->pdev->dev, "skmsg=%p event='%s'\n", skmsg, event);
-       dev_dbg(&skdev->pdev->dev, "  state=%s(%d) id=0x%04x length=%d\n",
-               skd_skmsg_state_to_str(skmsg->state), skmsg->state, skmsg->id,
-               skmsg->length);
-}
-
 static void skd_log_skreq(struct skd_device *skdev,
                          struct skd_request_context *skreq, const char *event)
 {