Skip to contentSkip to conversation

pc-bios/s390-ccw: Two fixes for the virtio initialization

PATCH In review Thomas Huth 4 participants 9 messages Jun 23, 2022

lore
Share & tools
lore
#0 lore raw

cover pc-bios/s390-ccw: Two fixes for the virtio initialization

Two minor fixes for the virtio initialization that I spotted while looking at the code. These currently do not fix any problem (since QEMU is quite forgiving here when doing things wrong), but we should anyway try to follow the virtio specification here.

Thomas Huth (2):
pc-bios/s390-ccw/virtio: Set missing status bits while initializing pc-bios/s390-ccw/virtio: Read device config after feature negotiation

 pc-bios/s390-ccw/virtio.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)
Signature · 1 line
2.31.1
#1 lore raw

PATCH 1/2 pc-bios/s390-ccw/virtio: Set missing status bits while initializing

According chapter "3.1.1 Driver Requirements: Device Initialization" of the Virtio specification (v1.1), a driver for a device has to set the ACKNOWLEDGE and DRIVER bits in the status field after resetting the device. The s390-ccw bios skipped these steps so far and seems like QEMU never cared. Anyway, it's better to follow the spec, so let's set these bits now in the right spots, too.

---
 pc-bios/s390-ccw/virtio.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)
pc-bios/s390-ccw/virtio.c +14 -4
index 5d2c6e3381..4e85a2eb82 100644
@@ -220,7 +220,7 @@ int virtio_run(VDev *vdev, int vqid, VirtioCmd *cmd)
 void virtio_setup_ccw(VDev *vdev)
 {
     int i, rc, cfg_size = 0;
-    unsigned char status = VIRTIO_CONFIG_S_DRIVER_OK;
+    uint8_t status;
     struct VirtioFeatureDesc {
         uint32_t features;
         uint8_t index;
@@ -234,6 +234,10 @@ void virtio_setup_ccw(VDev *vdev)
 
     run_ccw(vdev, CCW_CMD_VDEV_RESET, NULL, 0, false);
 
+    status = VIRTIO_CONFIG_S_ACKNOWLEDGE;
+    rc = run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status, sizeof(status), false);
+    IPL_assert(rc == 0, "Could not write ACKNOWLEDGE status to host");
+
     switch (vdev->senseid.cu_model) {
     case VIRTIO_ID_NET:
         vdev->nr_vqs = 2;
@@ -253,6 +257,11 @@ void virtio_setup_ccw(VDev *vdev)
     default:
         panic("Unsupported virtio device\n");
     }
+
+    status |= VIRTIO_CONFIG_S_DRIVER;
+    rc = run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status, sizeof(status), false);
+    IPL_assert(rc == 0, "Could not write DRIVER status to host");
+
     IPL_assert(
         run_ccw(vdev, CCW_CMD_READ_CONF, &vdev->config, cfg_size, false) == 0,
        "Could not get block device configuration");
@@ -291,9 +300,10 @@ void virtio_setup_ccw(VDev *vdev)
             run_ccw(vdev, CCW_CMD_SET_VQ, &info, sizeof(info), false) == 0,
             "Cannot set VQ info");
     }
-    IPL_assert(
-        run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status, sizeof(status), false) == 0,
-        "Could not write status to host");
+
+    status |= VIRTIO_CONFIG_S_DRIVER_OK;
+    rc = run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status, sizeof(status), false);
+    IPL_assert(rc == 0, "Could not write DRIVER_OK status to host");
 }
 
 bool virtio_is_supported(SubChannelId schid)
Signature · 1 line
2.31.1
Acked-by on v1 1/2
#2 lore raw

Re: [PATCH 1/2] pc-bios/s390-ccw/virtio: Set missing status bits while initializing

Am 23.06.22 um 09:11 schrieb Thomas Huth:

Quoted 6 lines

According chapter "3.1.1 Driver Requirements: Device Initialization" of the Virtio specification (v1.1), a driver for a device has to set the ACKNOWLEDGE and DRIVER bits in the status field after resetting the device. The s390-ccw bios skipped these steps so far and seems like QEMU never cared. Anyway, it's better to follow the spec, so let's set these bits now in the right spots, too.

I have not tested that but I agree with this

Quoted 55 lines
---
  pc-bios/s390-ccw/virtio.c | 18 ++++++++++++++----
  1 file changed, 14 insertions(+), 4 deletions(-)
pc-bios/s390-ccw/virtio.c +14 -4
index 5d2c6e3381..4e85a2eb82 100644
@@ -220,7 +220,7 @@ int virtio_run(VDev *vdev, int vqid, VirtioCmd *cmd)
  void virtio_setup_ccw(VDev *vdev)
  {
      int i, rc, cfg_size = 0;
-    unsigned char status = VIRTIO_CONFIG_S_DRIVER_OK;
+    uint8_t status;
      struct VirtioFeatureDesc {
          uint32_t features;
          uint8_t index;
@@ -234,6 +234,10 @@ void virtio_setup_ccw(VDev *vdev)
 
      run_ccw(vdev, CCW_CMD_VDEV_RESET, NULL, 0, false);
 
+    status = VIRTIO_CONFIG_S_ACKNOWLEDGE;
+    rc = run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status, sizeof(status), false);
+    IPL_assert(rc == 0, "Could not write ACKNOWLEDGE status to host");
+
      switch (vdev->senseid.cu_model) {
      case VIRTIO_ID_NET:
          vdev->nr_vqs = 2;
@@ -253,6 +257,11 @@ void virtio_setup_ccw(VDev *vdev)
      default:
          panic("Unsupported virtio device\n");
      }
+
+    status |= VIRTIO_CONFIG_S_DRIVER;
+    rc = run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status, sizeof(status), false);
+    IPL_assert(rc == 0, "Could not write DRIVER status to host");
+
      IPL_assert(
          run_ccw(vdev, CCW_CMD_READ_CONF, &vdev->config, cfg_size, false) == 0,
         "Could not get block device configuration");
@@ -291,9 +300,10 @@ void virtio_setup_ccw(VDev *vdev)
              run_ccw(vdev, CCW_CMD_SET_VQ, &info, sizeof(info), false) == 0,
              "Cannot set VQ info");
      }
-    IPL_assert(
-        run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status, sizeof(status), false) == 0,
-        "Could not write status to host");
+
+    status |= VIRTIO_CONFIG_S_DRIVER_OK;
+    rc = run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status, sizeof(status), false);
+    IPL_assert(rc == 0, "Could not write DRIVER_OK status to host");
  }
 
  bool virtio_is_supported(SubChannelId schid)
Reviewed-by on v1 1/2
#3 lore raw

Re: [PATCH 1/2] pc-bios/s390-ccw/virtio: Set missing status bits while initializing

In reply to Thomas Huth
Quoted 11 lines · Thomas Huth

According chapter "3.1.1 Driver Requirements: Device Initialization" of the Virtio specification (v1.1), a driver for a device has to set the ACKNOWLEDGE and DRIVER bits in the status field after resetting the device. The s390-ccw bios skipped these steps so far and seems like QEMU never cared. Anyway, it's better to follow the spec, so let's set these bits now in the right spots, too.

---
 pc-bios/s390-ccw/virtio.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)

The QEMU implementation seems to be pretty tolerant :)

Reviewed-by on v1 1/2
#4 lore raw

Re: [PATCH 1/2] pc-bios/s390-ccw/virtio: Set missing status bits while initializing

In reply to Thomas Huth
Quoted 8 lines · Thomas Huth

According chapter "3.1.1 Driver Requirements: Device Initialization" of the Virtio specification (v1.1), a driver for a device has to set the ACKNOWLEDGE and DRIVER bits in the status field after resetting the device. The s390-ccw bios skipped these steps so far and seems like QEMU never cared. Anyway, it's better to follow the spec, so let's set these bits now in the right spots, too.

Quoted 62 lines
---
 pc-bios/s390-ccw/virtio.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)
pc-bios/s390-ccw/virtio.c +1 -1
index 5d2c6e3381..4e85a2eb82 100644
@@ -220,7 +220,7 @@ int virtio_run(VDev *vdev, int vqid, VirtioCmd
 *cmd)
 void virtio_setup_ccw(VDev *vdev)
 {
     int i, rc, cfg_size = 0;
-    unsigned char status = VIRTIO_CONFIG_S_DRIVER_OK;
+    uint8_t status;
     struct VirtioFeatureDesc {
         uint32_t features;
         uint8_t index;
pc-bios/s390-ccw/virtio.c +4 -0
@@ -234,6 +234,10 @@ void virtio_setup_ccw(VDev *vdev)
 
     run_ccw(vdev, CCW_CMD_VDEV_RESET, NULL, 0, false);
 
+    status = VIRTIO_CONFIG_S_ACKNOWLEDGE;
+    rc = run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status,
 sizeof(status), false);
+    IPL_assert(rc == 0, "Could not write ACKNOWLEDGE status to
 host");
+
     switch (vdev->senseid.cu_model) {
     case VIRTIO_ID_NET:
         vdev->nr_vqs = 2;
pc-bios/s390-ccw/virtio.c +5 -0
@@ -253,6 +257,11 @@ void virtio_setup_ccw(VDev *vdev)
     default:
         panic("Unsupported virtio device\n");
     }
+
+    status |= VIRTIO_CONFIG_S_DRIVER;
+    rc = run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status,
 sizeof(status), false);
+    IPL_assert(rc == 0, "Could not write DRIVER status to host");
+
     IPL_assert(
         run_ccw(vdev, CCW_CMD_READ_CONF, &vdev->config, cfg_size,

false) == 0,

        "Could not get block device configuration");
pc-bios/s390-ccw/virtio.c +4 -3
@@ -291,9 +300,10 @@ void virtio_setup_ccw(VDev *vdev)
             run_ccw(vdev, CCW_CMD_SET_VQ, &info, sizeof(info),
 false) == 0,
             "Cannot set VQ info");
     }
-    IPL_assert(
-        run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status, sizeof(status),
 false) == 0,
-        "Could not write status to host");
+
+    status |= VIRTIO_CONFIG_S_DRIVER_OK;
+    rc = run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status,
 sizeof(status), false);
+    IPL_assert(rc == 0, "Could not write DRIVER_OK status to host");

}

bool virtio_is_supported(SubChannelId schid)

#5 lore raw

PATCH 2/2 pc-bios/s390-ccw/virtio: Read device config after feature negotiation

In reply to Thomas Huth

Feature negotiation should be done first, since some fields in the config area can depend on the negotiated features and thus should rather be read afterwards.

---
 pc-bios/s390-ccw/virtio.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)
pc-bios/s390-ccw/virtio.c +3 -4
index 4e85a2eb82..0e92e994df 100644
@@ -262,10 +262,6 @@ void virtio_setup_ccw(VDev *vdev)
     rc = run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status, sizeof(status), false);
     IPL_assert(rc == 0, "Could not write DRIVER status to host");
 
-    IPL_assert(
-        run_ccw(vdev, CCW_CMD_READ_CONF, &vdev->config, cfg_size, false) == 0,
-       "Could not get block device configuration");
-
     /* Feature negotiation */
     for (i = 0; i < ARRAY_SIZE(vdev->guest_features); i++) {
         feats.features = 0;
@@ -278,6 +274,9 @@ void virtio_setup_ccw(VDev *vdev)
         IPL_assert(rc == 0, "Could not set features bits");
     }
 
+    rc = run_ccw(vdev, CCW_CMD_READ_CONF, &vdev->config, cfg_size, false);
+    IPL_assert(rc == 0, "Could not get block device configuration");
+
     for (i = 0; i < vdev->nr_vqs; i++) {
         VqInfo info = {
             .queue = (unsigned long long) ring_area + (i * VIRTIO_RING_SIZE),
Signature · 1 line
2.31.1
#6 lore raw

Re: [PATCH 2/2] pc-bios/s390-ccw/virtio: Read device config after feature negotiation

Quoted 3 lines · Thomas Huth

Feature negotiation should be done first, since some fields in the config area can depend on the negotiated features and thus should rather be read afterwards.

I suppose we don't negotiate any features that might affect the size of the config space? Anyway, restricting ourselves to the minimum length should be fine.

Quoted 22 lines
---
 pc-bios/s390-ccw/virtio.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)
pc-bios/s390-ccw/virtio.c +0 -4
index 4e85a2eb82..0e92e994df 100644
@@ -262,10 +262,6 @@ void virtio_setup_ccw(VDev *vdev)
     rc = run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status, sizeof(status), false);
     IPL_assert(rc == 0, "Could not write DRIVER status to host");
 
-    IPL_assert(
-        run_ccw(vdev, CCW_CMD_READ_CONF, &vdev->config, cfg_size, false) == 0,
-       "Could not get block device configuration");
-
     /* Feature negotiation */
     for (i = 0; i < ARRAY_SIZE(vdev->guest_features); i++) {
         feats.features = 0;

pc-bios/s390-ccw/virtio.c void virtio_setup_ccw(VDev *vdev) · line 278

Quoted 5 lines
pc-bios/s390-ccw/virtio.c +2 -0
@@ -278,6 +274,9 @@ void virtio_setup_ccw(VDev *vdev)
         IPL_assert(rc == 0, "Could not set features bits");
     }
 
+    rc = run_ccw(vdev, CCW_CMD_READ_CONF, &vdev->config, cfg_size, false);
+    IPL_assert(rc == 0, "Could not get block device configuration");

Since you move this anyway: s/block device/boot device/ ?

pc-bios/s390-ccw/virtio.c void virtio_setup_ccw(VDev *vdev) · line 282

Quoted 4 lines
pc-bios/s390-ccw/virtio.c +1 -0
+
     for (i = 0; i < vdev->nr_vqs; i++) {
         VqInfo info = {
             .queue = (unsigned long long) ring_area + (i * VIRTIO_RING_SIZE),

Related to feature negotiation: It seems that the bios currently supports legacy virtio only, doesn't it? It's probably fine for now, but there might be a way in the future to disable legacy virtio for all devices in QEMU, so we'll probably want to add virtio-1 support some day.

#7 lore raw

Re: [PATCH 2/2] pc-bios/s390-ccw/virtio: Read device config after feature negotiation

Quoted 9 lines · Cornelia Huck
Quoted 3 lines · Thomas Huth

Feature negotiation should be done first, since some fields in the config area can depend on the negotiated features and thus should rather be read afterwards.

I suppose we don't negotiate any features that might affect the size of the config space? Anyway, restricting ourselves to the minimum length should be fine.

Actually, even the virtio spec 0.9.5 already talks about VIRTIO_BLK_F_BLK_SIZE and VIRTIO_BLK_F_GEOMETRY being necessary to get the corresponding values in the config space ... so we're currently depending on the good will of QEMU to also provide the values without these feature bits. I'm already thinking about providing a patch to properly request these feature bits in the s390-ccw bios ... but the whole code there is so ugly that I need some time to think about the right steps to clean it up first.

Quoted 24 lines
Quoted 24 lines
---
  pc-bios/s390-ccw/virtio.c | 7 +++----
  1 file changed, 3 insertions(+), 4 deletions(-)
pc-bios/s390-ccw/virtio.c +0 -4
index 4e85a2eb82..0e92e994df 100644
@@ -262,10 +262,6 @@ void virtio_setup_ccw(VDev *vdev)
      rc = run_ccw(vdev, CCW_CMD_WRITE_STATUS, &status, sizeof(status), false);
      IPL_assert(rc == 0, "Could not write DRIVER status to host");
 
-    IPL_assert(
-        run_ccw(vdev, CCW_CMD_READ_CONF, &vdev->config, cfg_size, false) == 0,
-       "Could not get block device configuration");
-
      /* Feature negotiation */
      for (i = 0; i < ARRAY_SIZE(vdev->guest_features); i++) {
          feats.features = 0;
@@ -278,6 +274,9 @@ void virtio_setup_ccw(VDev *vdev)
          IPL_assert(rc == 0, "Could not set features bits");
      }
 
Quoted 4 lines

pc-bios/s390-ccw/virtio.c void virtio_setup_ccw(VDev *vdev) · line 278

Quoted 2 lines
pc-bios/s390-ccw/virtio.c +2 -0
+    rc = run_ccw(vdev, CCW_CMD_READ_CONF, &vdev->config, cfg_size, false);
+    IPL_assert(rc == 0, "Could not get block device configuration");

Since you move this anyway: s/block device/boot device/ ?

Good idea, thanks, I'll clean that up!

Quoted 7 lines

pc-bios/s390-ccw/virtio.c void virtio_setup_ccw(VDev *vdev) · line 282

Quoted 4 lines
pc-bios/s390-ccw/virtio.c +1 -0
+
      for (i = 0; i < vdev->nr_vqs; i++) {
          VqInfo info = {
              .queue = (unsigned long long) ring_area + (i * VIRTIO_RING_SIZE),

Related to feature negotiation: It seems that the bios currently supports legacy virtio only, doesn't it?

Right.

Quoted 4 lines

It's probably fine for now, but
there might be a way in the future to disable legacy virtio for all devices in QEMU, so we'll probably want to add virtio-1 support some day.

Yes, that's certainly one thing we have to look into at one point in time ...

Thomas

#8 lore raw

Re: [PATCH 2/2] pc-bios/s390-ccw/virtio: Read device config after feature negotiation

Quoted 7 lines · Thomas Huth
Quoted 6 lines · Cornelia Huck
Quoted 3 lines · Thomas Huth

Feature negotiation should be done first, since some fields in the config area can depend on the negotiated features and thus should rather be read afterwards.

Quoted 8 lines
Quoted 3 lines

I suppose we don't negotiate any features that might affect the size of the config space? Anyway, restricting ourselves to the minimum length should be fine.

Actually, even the virtio spec 0.9.5 already talks about VIRTIO_BLK_F_BLK_SIZE and VIRTIO_BLK_F_GEOMETRY being necessary to get the corresponding values in the config space ... so we're currently depending on the good will of QEMU to also provide the values without these feature bits.

Eww. The bad thing about QEMU being so relaxed is that you don't catch issues like that... but if it works for now, there's at least no pressure.

Quoted 3 lines

I'm already thinking about providing a patch to properly request these feature bits in the s390-ccw bios ... but the whole code there is so ugly that I need some time to think about the right steps to clean it up first.

It's not ugly, it has grown organically :) (just like that overgrown spot in my garden ;)