diff mbox series

[PULL,for-6.0,4/4] hw/block/nvme: drain namespaces on sq deletion

Message ID 20210412070141.521593-5-its@irrelevant.dk (mailing list archive)
State New, archived
Headers show
Series [PULL,for-6.0,1/4] docs: add nvme emulation documentation | expand

Commit Message

Klaus Jensen April 12, 2021, 7:01 a.m. UTC
From: Klaus Jensen <k.jensen@samsung.com>

For most commands, when issuing an AIO, the BlockAIOCB is stored in the
NvmeRequest aiocb pointer when the AIO is issued. The main use of this
is cancelling AIOs when deleting submission queues (it is currently not
used for Abort).

However, some commands like Dataset Management Zone Management Send
(zone reset) may involve more than one AIO and here the AIOs are issued
without saving a reference to the BlockAIOCB. This is a problem since
nvme_del_sq() will attempt to cancel outstanding AIOs, potentially with
an invalid BlockAIOCB since the aiocb pointer is not NULL'ed when the
request structure is recycled.

Fix this by

  1. making sure the aiocb pointer is NULL'ed when requests are recycled
  2. only attempt to cancel the AIO if the aiocb is non-NULL
  3. if any AIOs could not be cancelled, drain all aio as a last resort.

Fixes: dc04d25e2f3f ("hw/block/nvme: add support for the format nvm command")
Fixes: c94973288cd9 ("hw/block/nvme: add broadcast nsid support flush command")
Fixes: e4e430b3d6ba ("hw/block/nvme: add simple copy command")
Fixes: 5f5dc4c6a942 ("hw/block/nvme: zero out zones on reset")
Fixes: 2605257a26b8 ("hw/block/nvme: add the dataset management command")
Cc: Gollu Appalanaidu <anaidu.gollu@samsung.com>
Cc: Minwoo Im <minwoo.im@samsung.com>
Signed-off-by: Klaus Jensen <k.jensen@samsung.com>
Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
---
 hw/block/nvme.c | 23 +++++++++++++++++++++--
 1 file changed, 21 insertions(+), 2 deletions(-)
diff mbox series

Patch

diff --git a/hw/block/nvme.c b/hw/block/nvme.c
index 65bb2cfc21d7..624a1431d072 100644
--- a/hw/block/nvme.c
+++ b/hw/block/nvme.c
@@ -470,6 +470,7 @@  static void nvme_req_clear(NvmeRequest *req)
 {
     req->ns = NULL;
     req->opaque = NULL;
+    req->aiocb = NULL;
     memset(&req->cqe, 0x0, sizeof(req->cqe));
     req->status = NVME_SUCCESS;
 }
@@ -3687,6 +3688,7 @@  static uint16_t nvme_del_sq(NvmeCtrl *n, NvmeRequest *req)
     NvmeSQueue *sq;
     NvmeCQueue *cq;
     uint16_t qid = le16_to_cpu(c->qid);
+    uint32_t nsid;
 
     if (unlikely(!qid || nvme_check_sqid(n, qid))) {
         trace_pci_nvme_err_invalid_del_sq(qid);
@@ -3698,9 +3700,26 @@  static uint16_t nvme_del_sq(NvmeCtrl *n, NvmeRequest *req)
     sq = n->sq[qid];
     while (!QTAILQ_EMPTY(&sq->out_req_list)) {
         r = QTAILQ_FIRST(&sq->out_req_list);
-        assert(r->aiocb);
-        blk_aio_cancel(r->aiocb);
+        if (r->aiocb) {
+            blk_aio_cancel(r->aiocb);
+        }
     }
+
+    /*
+     * Drain all namespaces if there are still outstanding requests that we
+     * could not cancel explicitly.
+     */
+    if (!QTAILQ_EMPTY(&sq->out_req_list)) {
+        for (nsid = 1; nsid <= NVME_MAX_NAMESPACES; nsid++) {
+            NvmeNamespace *ns = nvme_ns(n, nsid);
+            if (ns) {
+                nvme_ns_drain(ns);
+            }
+        }
+    }
+
+    assert(QTAILQ_EMPTY(&sq->out_req_list));
+
     if (!nvme_check_cqid(n, sq->cqid)) {
         cq = n->cq[sq->cqid];
         QTAILQ_REMOVE(&cq->sq_list, sq, entry);