hw/nvme: reimplement flush to allow cancellation

Prior to this patch, a broadcast flush would result in submitting
multiple "fire and forget" aios (no reference saved to the aiocbs
returned from the blk_aio_flush calls).

Fix this by issuing the flushes one after another.

Signed-off-by: Klaus Jensen <k.jensen@samsung.com>
Reviewed-by: Keith Busch <kbusch@kernel.org>

1 files changed, 124 insertions, 80 deletions

diff --git a/hw/nvme/ctrl.c b/hw/nvme/ctrl.c
index 762bb82e3c..26c65a12e8 100644
--- a/hw/nvme/ctrl.c
+++ b/hw/nvme/ctrl.c
@@ -1788,22 +1788,19 @@ static inline bool nvme_is_write(NvmeRequest *req)
            rw->opcode == NVME_CMD_WRITE_ZEROES;
 }
 
+static AioContext *nvme_get_aio_context(BlockAIOCB *acb)
+{
+    return qemu_get_aio_context();
+}
+
 static void nvme_misc_cb(void *opaque, int ret)
 {
     NvmeRequest *req = opaque;
-    NvmeNamespace *ns = req->ns;
 
-    BlockBackend *blk = ns->blkconf.blk;
-    BlockAcctCookie *acct = &req->acct;
-    BlockAcctStats *stats = blk_get_stats(blk);
-
-    trace_pci_nvme_misc_cb(nvme_cid(req), blk_name(blk));
+    trace_pci_nvme_misc_cb(nvme_cid(req));
 
     if (ret) {
-        block_acct_failed(stats, acct);
         nvme_aio_err(req, ret);
-    } else {
-        block_acct_done(stats, acct);
     }
 
     nvme_enqueue_req_completion(nvme_cq(req), req);
@@ -1919,41 +1916,6 @@ static void nvme_aio_format_cb(void *opaque, int ret)
     nvme_enqueue_req_completion(nvme_cq(req), req);
 }
 
-struct nvme_aio_flush_ctx {
-    NvmeRequest     *req;
-    NvmeNamespace   *ns;
-    BlockAcctCookie acct;
-};
-
-static void nvme_aio_flush_cb(void *opaque, int ret)
-{
-    struct nvme_aio_flush_ctx *ctx = opaque;
-    NvmeRequest *req = ctx->req;
-    uintptr_t *num_flushes = (uintptr_t *)&req->opaque;
-
-    BlockBackend *blk = ctx->ns->blkconf.blk;
-    BlockAcctCookie *acct = &ctx->acct;
-    BlockAcctStats *stats = blk_get_stats(blk);
-
-    trace_pci_nvme_aio_flush_cb(nvme_cid(req), blk_name(blk));
-
-    if (!ret) {
-        block_acct_done(stats, acct);
-    } else {
-        block_acct_failed(stats, acct);
-        nvme_aio_err(req, ret);
-    }
-
-    (*num_flushes)--;
-    g_free(ctx);
-
-    if (*num_flushes) {
-        return;
-    }
-
-    nvme_enqueue_req_completion(nvme_cq(req), req);
-}
-
 static void nvme_verify_cb(void *opaque, int ret)
 {
     NvmeBounceContext *ctx = opaque;
@@ -2868,57 +2830,139 @@ static uint16_t nvme_compare(NvmeCtrl *n, NvmeRequest *req)
     return NVME_NO_COMPLETE;
 }
 
-static uint16_t nvme_flush(NvmeCtrl *n, NvmeRequest *req)
-{
-    uint32_t nsid = le32_to_cpu(req->cmd.nsid);
-    uintptr_t *num_flushes = (uintptr_t *)&req->opaque;
-    uint16_t status;
-    struct nvme_aio_flush_ctx *ctx;
+typedef struct NvmeFlushAIOCB {
+    BlockAIOCB common;
+    BlockAIOCB *aiocb;
+    NvmeRequest *req;
+    QEMUBH *bh;
+    int ret;
+
     NvmeNamespace *ns;
+    uint32_t nsid;
+    bool broadcast;
+} NvmeFlushAIOCB;
 
-    trace_pci_nvme_flush(nvme_cid(req), nsid);
+static void nvme_flush_cancel(BlockAIOCB *acb)
+{
+    NvmeFlushAIOCB *iocb = container_of(acb, NvmeFlushAIOCB, common);
 
-    if (nsid != NVME_NSID_BROADCAST) {
-        req->ns = nvme_ns(n, nsid);
-        if (unlikely(!req->ns)) {
-            return NVME_INVALID_FIELD | NVME_DNR;
-        }
+    iocb->ret = -ECANCELED;
 
-        block_acct_start(blk_get_stats(req->ns->blkconf.blk), &req->acct, 0,
-                         BLOCK_ACCT_FLUSH);
-        req->aiocb = blk_aio_flush(req->ns->blkconf.blk, nvme_misc_cb, req);
-        return NVME_NO_COMPLETE;
+    if (iocb->aiocb) {
+        blk_aio_cancel_async(iocb->aiocb);
     }
+}
 
-    /* 1-initialize; see comment in nvme_dsm */
-    *num_flushes = 1;
+static const AIOCBInfo nvme_flush_aiocb_info = {
+    .aiocb_size = sizeof(NvmeFlushAIOCB),
+    .cancel_async = nvme_flush_cancel,
+    .get_aio_context = nvme_get_aio_context,
+};
 
-    for (int i = 1; i <= NVME_MAX_NAMESPACES; i++) {
-        ns = nvme_ns(n, i);
-        if (!ns) {
-            continue;
-        }
+static void nvme_flush_ns_cb(void *opaque, int ret)
+{
+    NvmeFlushAIOCB *iocb = opaque;
+    NvmeNamespace *ns = iocb->ns;
 
-        ctx = g_new(struct nvme_aio_flush_ctx, 1);
-        ctx->req = req;
-        ctx->ns = ns;
+    if (ret < 0) {
+        iocb->ret = ret;
+        goto out;
+    } else if (iocb->ret < 0) {
+        goto out;
+    }
 
-        (*num_flushes)++;
+    if (ns) {
+        trace_pci_nvme_flush_ns(iocb->nsid);
 
-        block_acct_start(blk_get_stats(ns->blkconf.blk), &ctx->acct, 0,
-                         BLOCK_ACCT_FLUSH);
-        blk_aio_flush(ns->blkconf.blk, nvme_aio_flush_cb, ctx);
+        iocb->ns = NULL;
+        iocb->aiocb = blk_aio_flush(ns->blkconf.blk, nvme_flush_ns_cb, iocb);
+        return;
     }
 
-    /* account for the 1-initialization */
-    (*num_flushes)--;
+out:
+    iocb->aiocb = NULL;
+    qemu_bh_schedule(iocb->bh);
+}
 
-    if (*num_flushes) {
-        status = NVME_NO_COMPLETE;
-    } else {
-        status = req->status;
+static void nvme_flush_bh(void *opaque)
+{
+    NvmeFlushAIOCB *iocb = opaque;
+    NvmeRequest *req = iocb->req;
+    NvmeCtrl *n = nvme_ctrl(req);
+    int i;
+
+    if (iocb->ret < 0) {
+        goto done;
+    }
+
+    if (iocb->broadcast) {
+        for (i = iocb->nsid + 1; i <= NVME_MAX_NAMESPACES; i++) {
+            iocb->ns = nvme_ns(n, i);
+            if (iocb->ns) {
+                iocb->nsid = i;
+                break;
+            }
+        }
+    }
+
+    if (!iocb->ns) {
+        goto done;
+    }
+
+    nvme_flush_ns_cb(iocb, 0);
+    return;
+
+done:
+    qemu_bh_delete(iocb->bh);
+    iocb->bh = NULL;
+
+    iocb->common.cb(iocb->common.opaque, iocb->ret);
+
+    qemu_aio_unref(iocb);
+
+    return;
+}
+
+static uint16_t nvme_flush(NvmeCtrl *n, NvmeRequest *req)
+{
+    NvmeFlushAIOCB *iocb;
+    uint32_t nsid = le32_to_cpu(req->cmd.nsid);
+    uint16_t status;
+
+    iocb = qemu_aio_get(&nvme_flush_aiocb_info, NULL, nvme_misc_cb, req);
+
+    iocb->req = req;
+    iocb->bh = qemu_bh_new(nvme_flush_bh, iocb);
+    iocb->ret = 0;
+    iocb->ns = NULL;
+    iocb->nsid = 0;
+    iocb->broadcast = (nsid == NVME_NSID_BROADCAST);
+
+    if (!iocb->broadcast) {
+        if (!nvme_nsid_valid(n, nsid)) {
+            status = NVME_INVALID_NSID | NVME_DNR;
+            goto out;
+        }
+
+        iocb->ns = nvme_ns(n, nsid);
+        if (!iocb->ns) {
+            status = NVME_INVALID_FIELD | NVME_DNR;
+            goto out;
+        }
+
+        iocb->nsid = nsid;
     }
 
+    req->aiocb = &iocb->common;
+    qemu_bh_schedule(iocb->bh);
+
+    return NVME_NO_COMPLETE;
+
+out:
+    qemu_bh_delete(iocb->bh);
+    iocb->bh = NULL;
+    qemu_aio_unref(iocb);
+
     return status;
 }