[Top][All Lists]

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

Re: [PATCH 4/4] hw/nvme: fix controller hot unplugging

From: Hannes Reinecke
Subject: Re: [PATCH 4/4] hw/nvme: fix controller hot unplugging
Date: Thu, 8 Jul 2021 08:15:28 +0200
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:78.0) Gecko/20100101 Thunderbird/78.11.0

On 7/7/21 6:14 PM, Stefan Hajnoczi wrote:
On Wed, Jul 07, 2021 at 12:43:56PM +0200, Hannes Reinecke wrote:
On 7/7/21 11:53 AM, Klaus Jensen wrote:
On Jul  7 09:49, Hannes Reinecke wrote:
On 7/6/21 11:33 AM, Klaus Jensen wrote:
From: Klaus Jensen <k.jensen@samsung.com>

Prior to this patch the nvme-ns devices are always children of the
NvmeBus owned by the NvmeCtrl. This causes the namespaces to be
unrealized when the parent device is removed. However, when subsystems
are involved, this is not what we want since the namespaces may be
attached to other controllers as well.

This patch adds an additional NvmeBus on the subsystem device. When
nvme-ns devices are realized, if the parent controller device is linked
to a subsystem, the parent bus is set to the subsystem one instead. This
makes sure that namespaces are kept alive and not unrealized.

Signed-off-by: Klaus Jensen <k.jensen@samsung.com>
  hw/nvme/nvme.h   | 18 ++++++++++--------
  hw/nvme/ctrl.c   |  8 +++++---
  hw/nvme/ns.c     | 32 +++++++++++++++++++++++++-------
  hw/nvme/subsys.c |  4 ++++
  4 files changed, 44 insertions(+), 18 deletions(-)

diff --git a/hw/nvme/nvme.h b/hw/nvme/nvme.h
index c4065467d877..9401e212f9f7 100644
--- a/hw/nvme/nvme.h
+++ b/hw/nvme/nvme.h
  typedef struct NvmeCtrl NvmeCtrl;
  typedef struct NvmeNamespace NvmeNamespace;
+#define TYPE_NVME_BUS "nvme-bus"
+typedef struct NvmeBus {
+    BusState parent_bus;
+    bool     is_subsys;
+} NvmeBus;
  #define TYPE_NVME_SUBSYS "nvme-subsys"
  #define NVME_SUBSYS(obj) \
      OBJECT_CHECK(NvmeSubsystem, (obj), TYPE_NVME_SUBSYS)
  typedef struct NvmeSubsystem {
      DeviceState parent_obj;
+    NvmeBus     bus;
      uint8_t     subnqn[256];
      NvmeCtrl      *ctrls[NVME_MAX_CONTROLLERS];
@@ -365,13 +374,6 @@ typedef struct NvmeCQueue {
      QTAILQ_HEAD(, NvmeRequest) req_list;
  } NvmeCQueue;
-#define TYPE_NVME_BUS "nvme-bus"
-#define NVME_BUS(obj) OBJECT_CHECK(NvmeBus, (obj), TYPE_NVME_BUS)
-typedef struct NvmeBus {
-    BusState parent_bus;
-} NvmeBus;
  #define TYPE_NVME "nvme"
  #define NVME(obj) \
          OBJECT_CHECK(NvmeCtrl, (obj), TYPE_NVME)
@@ -463,7 +465,7 @@ typedef struct NvmeCtrl {
  static inline NvmeNamespace *nvme_ns(NvmeCtrl *n, uint32_t nsid)
-    if (!nsid || nsid > NVME_MAX_NAMESPACES) {
+    if (!n || !nsid || nsid > NVME_MAX_NAMESPACES) {
          return NULL;
diff --git a/hw/nvme/ctrl.c b/hw/nvme/ctrl.c
index 90e3ee2b70ee..7c8fca36d9a5 100644
--- a/hw/nvme/ctrl.c
+++ b/hw/nvme/ctrl.c
@@ -6516,11 +6516,13 @@ static void nvme_exit(PCIDevice *pci_dev)
      for (i = 1; i <= NVME_MAX_NAMESPACES; i++) {
          ns = nvme_ns(n, i);
-        if (!ns) {
-            continue;
+        if (ns) {
+            ns->attached--;
+    }
-        nvme_ns_cleanup(ns);
+    if (n->subsys) {
+        nvme_subsys_unregister_ctrl(n->subsys, n);
      if (n->subsys) {
diff --git a/hw/nvme/ns.c b/hw/nvme/ns.c
index 3c4f5b8c714a..612a2786d75d 100644
--- a/hw/nvme/ns.c
+++ b/hw/nvme/ns.c
@@ -444,13 +444,29 @@ void nvme_ns_cleanup(NvmeNamespace *ns)
  static void nvme_ns_realize(DeviceState *dev, Error **errp)
      NvmeNamespace *ns = NVME_NS(dev);
-    BusState *s = qdev_get_parent_bus(dev);
-    NvmeCtrl *n = NVME(s->parent);
-    NvmeSubsystem *subsys = n->subsys;
+    BusState *qbus = qdev_get_parent_bus(dev);
+    NvmeBus *bus = NVME_BUS(qbus);
+    NvmeCtrl *n = NULL;
+    NvmeSubsystem *subsys = NULL;
      uint32_t nsid = ns->params.nsid;
      int i;
-    if (!n->subsys) {
+    if (bus->is_subsys) {
+        subsys = NVME_SUBSYS(qbus->parent);
+    } else {
+        n = NVME(qbus->parent);
+        subsys = n->subsys;
+    }
+    if (subsys) {
+        /*
+         * If this namespace belongs to a subsystem (through a
link on the
+         * controller device), reparent the device.
+         */
+        if (!qdev_set_parent_bus(dev, &subsys->bus.parent_bus, errp)) {
+            return;
+        }
+    } else {
          if (ns->params.detached) {
              error_setg(errp, "detached requires that the nvme
device is "
                         "linked to an nvme-subsys device");
@@ -470,7 +486,7 @@ static void nvme_ns_realize(DeviceState
*dev, Error **errp)
      if (!nsid) {
          for (i = 1; i <= NVME_MAX_NAMESPACES; i++) {
-            if (nvme_ns(n, i) || nvme_subsys_ns(subsys, i)) {
+            if (nvme_subsys_ns(subsys, i) || nvme_ns(n, i)) {
@@ -483,7 +499,7 @@ static void nvme_ns_realize(DeviceState
*dev, Error **errp)
      } else {
-        if (nvme_ns(n, nsid) || nvme_subsys_ns(subsys, nsid)) {
+        if (nvme_subsys_ns(subsys, nsid) || nvme_ns(n, nsid)) {
              error_setg(errp, "namespace id '%d' already
allocated", nsid);
@@ -509,7 +525,9 @@ static void nvme_ns_realize(DeviceState
*dev, Error **errp)
-    nvme_attach_ns(n, ns);
+    if (n) {
+        nvme_attach_ns(n, ns);
+    }
  static Property nvme_ns_props[] = {
diff --git a/hw/nvme/subsys.c b/hw/nvme/subsys.c
index 92caa604a280..fb7c3a7c55fc 100644
--- a/hw/nvme/subsys.c
+++ b/hw/nvme/subsys.c
@@ -50,6 +50,10 @@ static void nvme_subsys_realize(DeviceState
*dev, Error **errp)
      NvmeSubsystem *subsys = NVME_SUBSYS(dev);
+    qbus_create_inplace(&subsys->bus, sizeof(NvmeBus),
+                        dev->id);
+    subsys->bus.is_subsys = true;

Why not always create a subsystem, and distinguish between 'shared'
and 'non-shared' subsystems instead of the ->subsys pointer?
That way all namespaces can be children of the subsystem, we won't
need any reparenting, and the whole thing will be more in-line with
qdev, no?

Hi Hannes,

Thanks for your reviews and comments!

So, I have actually considered what you suggest.

The problem is that the nvme-ns device currently expects to plug into a
bus (an 'nvme-bus') and we cannot really get rid of that 'bus' parameter
without breaking compatibility. I experimented with removing the bus
from the nvme device and instead creating it in the nvme-subsys device.
My idea here was to implicitly always create an nvme-subsys if not
explicitly created by the user, but since nvme-subsys does not plug into
any bus itself, it is not exposed in the qtree and thus not reachable
(i.e., when realizing the nvme-ns device, it will complain that there
are no 'nvme-bus' available to plug into). To make this work, I believe
we'd have to have the nvme-subsys plug into the main system bus (or some
other bus rooted at main-system-bus), and I'd prefer not to do that
since this is already a flawed design and I think it would be more
intrusive than what my patch does.


I seem to trip over issues for which no patch can possibly be accepted as
first the infrastructure has to be reworked.

... there is a reason why I'm avoiding posting patches to qemu-devel :-(

I raised these issues on the mailinglist some time ago[1]. We didn't
really find a good solution, but I think the conclusion is that the
bus/device design of subsystems and namespaces is flawed. That's why I
am working on an "objectification" of the two devices as suggested by
Stefan[2] as a proper fix for this mess.

Actually, I _do_ think that it would be good to have an nvme-subsystem bus,
providing a list of all namespaces for that subsystem.
Creating a controller would then mean that one has to create a controller
and a namespace _separately_, as then the namespace would _not_ be a child
of the controller.
But that's arguably a good thing, as the namespaces _do_ have a separate
lifetime from controllers.
And the lifetime of the subsystem could be completely self-contained; the
controller would be looking up subsystems via the subsystem NQN if present;
I guess we'll need to create ad-hoc subsystems for PCIe.
But nothing insurmountable.
And certainly nothing which require a large-scale rework of qemu internals,
I would think.

I don't remember the reason to use --object instead of --device. Maybe
the --device approach is viable and it makes some sense because this is
a "device" (like a SCSI LUN) that exposed to the guest. It would be
necessary to have a nvme-subsys device on the system bus to act as the
bus for nvme-ns devices as Klaus described.

Which is what I've had in mind.
The only issue here is that a nvme controller will also generate or attach to a nvme subsys, and the subsys will have a different lifetime than the controller. But it should be possible to key the lifetime of the subsys to the attached controllers (via refcounting, say), so that once the last controller is detached from a subsystem the subsystem itself is also detached.

But _if_ you work on the rework, please ensure that not only NVMe is
represented here. SCSI multipathing has the same issue currently; there is a
strict host->lun->block->OS device relationship currently in SCSI, making it
impossible to represent a multipathed SCSI device in qemu.
The only sneaky way is to allow for a shared OS device (ie specifying the
filename more than once), but that completely sidesteps qemu internals, too.

Do you want one SCSIDevice to be accessible via multiple SCSIBus
instances for multipathing in QEMU?

Precisely. Some things like ALUA or multipath issues in general are tricky to reproduce, and qemu would be a neat way here. But currently
we can't as we cannot emulate multipathing.

Any hints are welcome; I've tried several times to implement it but inevitably failed :-(


Dr. Hannes Reinecke                Kernel Storage Architect
hare@suse.de                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer

reply via email to

[Prev in Thread] Current Thread [Next in Thread]