Skip to contentSkip to conversation

hw/net: e1000: Correct the initial value of VET register

PATCH Superseded Bin Meng 3 participants 20 messages Jul 2, 2021 – Jul 14, 2021

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Bin Meng

author
#0 lore raw

PATCH v2 1/3 hw/net: e1000: Correct the initial value of VET register

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

 hw/net/e1000.c | 2 ++
 1 file changed, 2 insertions(+)
hw/net/e1000.c +2 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
 #include "hw/pci/pci.h"
 #include "hw/qdev-properties.h"
 #include "migration/vmstate.h"
+#include "net/eth.h"
 #include "net/net.h"
 #include "net/checksum.h"
 #include "sysemu/sysemu.h"
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
     [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                 E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                 E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,
 };
 
 /* Helper function, *curr == 0 means the value is not set */
Signature · 1 line
2.25.1
author
#1 lore raw

PATCH v2 2/3 hw/net: e1000e: Correct the initial value of VET register

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000e driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

Unlike e1000 in QEMU, e1000e uses a field 'vet' in "struct E1000Core" to cache the value of VET register, but the cache only gets updated when VET register is written. To always get a consistent VET value no matter VET is written or remains its reset value, drop the 'vet' field and use 'core->mac[VET]' directly.

---

Changes in v2:
- keep the 'vet' field in "struct E1000Core" for migration compatibility

 hw/net/e1000e_core.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)
hw/net/e1000e_core.c +6 -5
index b75f2ab8fc..38b3e3b784 100644
@@ -35,6 +35,7 @@
 
 #include "qemu/osdep.h"
 #include "qemu/log.h"
+#include "net/eth.h"
 #include "net/net.h"
 #include "net/tap.h"
 #include "hw/pci/msi.h"
@@ -731,7 +732,7 @@ e1000e_process_tx_desc(E1000ECore *core,
             if (e1000x_vlan_enabled(core->mac) &&
                 e1000x_is_vlan_txd(txd_lower)) {
                 net_tx_pkt_setup_vlan_header_ex(tx->tx_pkt,
-                    le16_to_cpu(dp->upper.fields.special), core->vet);
+                    le16_to_cpu(dp->upper.fields.special), core->mac[VET]);
             }
             if (e1000e_tx_pkt_send(core, tx, queue_index)) {
                 e1000e_on_tx_done_update_stats(core, tx->tx_pkt);
@@ -1012,7 +1013,7 @@ e1000e_receive_filter(E1000ECore *core, const uint8_t *buf, int size)
 {
     uint32_t rctl = core->mac[RCTL];
 
-    if (e1000x_is_vlan_packet(buf, core->vet) &&
+    if (e1000x_is_vlan_packet(buf, core->mac[VET]) &&
         e1000x_vlan_rx_filter_enabled(core->mac)) {
         uint16_t vid = lduw_be_p(buf + 14);
         uint32_t vfta = ldl_le_p((uint32_t *)(core->mac + VFTA) +
@@ -1686,7 +1687,7 @@ e1000e_receive_iov(E1000ECore *core, const struct iovec *iov, int iovcnt)
     }
 
     net_rx_pkt_attach_iovec_ex(core->rx_pkt, iov, iovcnt, iov_ofs,
-                               e1000x_vlan_enabled(core->mac), core->vet);
+                               e1000x_vlan_enabled(core->mac), core->mac[VET]);
 
     e1000e_rss_parse_packet(core, core->rx_pkt, &rss_info);
     e1000e_rx_ring_init(core, &rxr, rss_info.queue);
@@ -2397,8 +2398,7 @@ static void
 e1000e_set_vet(E1000ECore *core, int index, uint32_t val)
 {
     core->mac[VET] = val & 0xffff;
-    core->vet = le16_to_cpu(core->mac[VET]);
-    trace_e1000e_vlan_vet(core->vet);
+    trace_e1000e_vlan_vet(core->mac[VET]);
 }
 
 static void
@@ -3442,6 +3442,7 @@ static const uint32_t e1000e_mac_reg_init[] = {
     [RXCSUM]        = E1000_RXCSUM_IPOFLD | E1000_RXCSUM_TUOFLD,
     [ITR]           = E1000E_MIN_XITR,
     [EITR...EITR + E1000E_MSIX_VEC_NUM - 1] = E1000E_MIN_XITR,
+    [VET]           = ETH_P_VLAN,
 };
 
 void
Signature · 1 line
2.25.1
author
#2 lore raw

PATCH v2 3/3 hw/net: e1000e: Don't zero out the VLAN tag in the legacy RX descriptor

In reply to Bin Meng

From: Christina Wang <[email protected]>

In the legacy RX descriptor mode, VLAN tag was saved to d->special by e1000e_build_rx_metadata() in e1000e_write_lgcy_rx_descr(), but it was then zeroed out again at the end of the call, which is wrong.

Fixes c89d416a2b0f ("e1000e: Don't zero out buffer address in rx descriptor")

---

(no changes since v1)

 hw/net/e1000e_core.c | 1 -
 1 file changed, 1 deletion(-)
hw/net/e1000e_core.c +0 -1
index 38b3e3b784..738c7169e4 100644
@@ -1286,7 +1286,6 @@ e1000e_write_lgcy_rx_descr(E1000ECore *core, uint8_t *desc,
                              &d->special);
     d->errors = (uint8_t) (le32_to_cpu(status_flags) >> 24);
     d->status = (uint8_t) le32_to_cpu(status_flags);
-    d->special = 0;
 }
 
 static inline void
Signature · 1 line
2.25.1
#3 lore raw
In reply to Bin Meng
Quoted 32 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

  hw/net/e1000.c | 2 ++
  1 file changed, 2 insertions(+)
hw/net/e1000.c +1 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
  #include "hw/pci/pci.h"
  #include "hw/qdev-properties.h"
  #include "migration/vmstate.h"
+#include "net/eth.h"
  #include "net/net.h"
  #include "net/checksum.h"
  #include "sysemu/sysemu.h"

hw/net/e1000.c static const uint32_t mac_reg_init[] = { · line 258

Quoted 4 lines
hw/net/e1000.c +1 -0
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
      [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                  E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                  E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

(See e1000_properties[])

Thanks

Quoted 3 lines
hw/net/e1000.c +0 -0
  };
 
  /* Helper function, *curr == 0 means the value is not set */
author
#4 lore raw
Quoted 43 lines · Jason Wang
Quoted 32 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

  hw/net/e1000.c | 2 ++
  1 file changed, 2 insertions(+)
hw/net/e1000.c +1 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
  #include "hw/pci/pci.h"
  #include "hw/qdev-properties.h"
  #include "migration/vmstate.h"
+#include "net/eth.h"
  #include "net/net.h"
  #include "net/checksum.h"
  #include "sysemu/sysemu.h"

hw/net/e1000.c static const uint32_t mac_reg_init[] = { · line 258

Quoted 4 lines
hw/net/e1000.c +1 -0
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
      [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                  E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                  E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Regards,
Bin

author
#5 lore raw
Quoted 45 lines · Bin Meng
Quoted 43 lines · Jason Wang
Quoted 32 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

  hw/net/e1000.c | 2 ++
  1 file changed, 2 insertions(+)
hw/net/e1000.c +1 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
  #include "hw/pci/pci.h"
  #include "hw/qdev-properties.h"
  #include "migration/vmstate.h"
+#include "net/eth.h"
  #include "net/net.h"
  #include "net/checksum.h"
  #include "sysemu/sysemu.h"

hw/net/e1000.c static const uint32_t mac_reg_init[] = { · line 258

Quoted 4 lines
hw/net/e1000.c +1 -0
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
      [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                  E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                  E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

Quoted 4 lines
Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Ping?

#6 lore raw
Quoted 43 lines · Bin Meng
Quoted 42 lines · Bin Meng
Quoted 41 lines · Jason Wang
Quoted 32 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

   hw/net/e1000.c | 2 ++
   1 file changed, 2 insertions(+)
hw/net/e1000.c +1 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
   #include "hw/pci/pci.h"
   #include "hw/qdev-properties.h"
   #include "migration/vmstate.h"
+#include "net/eth.h"
   #include "net/net.h"
   #include "net/checksum.h"
   #include "sysemu/sysemu.h"

hw/net/e1000.c static const uint32_t mac_reg_init[] = { · line 258

Quoted 4 lines
hw/net/e1000.c +1 -0
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
       [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                   E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                   E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

Quoted 4 lines
Quoted 3 lines
Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Ping?

I admit migration "works" but it doesn't mean it's not broken. It changes the guest visible default value of VET register, so it may break things silently for the guest.

For old machine types, we should stick the value to the one without this fix.

Thanks

Quoted 1 line
author
#7 lore raw
Quoted 52 lines · Jason Wang
Quoted 43 lines · Bin Meng
Quoted 42 lines · Bin Meng
Quoted 41 lines · Jason Wang
Quoted 32 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

   hw/net/e1000.c | 2 ++
   1 file changed, 2 insertions(+)
hw/net/e1000.c +1 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
   #include "hw/pci/pci.h"
   #include "hw/qdev-properties.h"
   #include "migration/vmstate.h"
+#include "net/eth.h"
   #include "net/net.h"
   #include "net/checksum.h"
   #include "sysemu/sysemu.h"

hw/net/e1000.c static const uint32_t mac_reg_init[] = { · line 258

Quoted 4 lines
hw/net/e1000.c +1 -0
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
       [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                   E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                   E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

Quoted 4 lines
Quoted 3 lines
Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Ping?

Quoted 6 lines

I admit migration "works" but it doesn't mean it's not broken. It changes the guest visible default value of VET register, so it may break things silently for the guest.

For old machine types, we should stick the value to the one without this fix.

Could you please propose a solution on how to handle such a scenario in a generic way in QEMU? (+Peter)

The POR reset value is wrong in QEMU and has carried forward the wrong value for years, and correcting it to its right value needs to do what?

Regards,
Bin

#8 lore raw
Quoted 49 lines · Bin Meng
Quoted 48 lines · Jason Wang
Quoted 45 lines · Bin Meng
Quoted 40 lines · Bin Meng
Quoted 39 lines · Jason Wang
Quoted 32 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

    hw/net/e1000.c | 2 ++
    1 file changed, 2 insertions(+)
hw/net/e1000.c +1 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
    #include "hw/pci/pci.h"
    #include "hw/qdev-properties.h"
    #include "migration/vmstate.h"
+#include "net/eth.h"
    #include "net/net.h"
    #include "net/checksum.h"
    #include "sysemu/sysemu.h"

hw/net/e1000.c static const uint32_t mac_reg_init[] = { · line 258

Quoted 4 lines
hw/net/e1000.c +1 -0
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
        [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                    E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                    E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

Quoted 3 lines
Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Ping?

Quoted 8 lines
Quoted 6 lines

I admit migration "works" but it doesn't mean it's not broken. It changes the guest visible default value of VET register, so it may break things silently for the guest.

For old machine types, we should stick the value to the one without this fix.

Could you please propose a solution on how to handle such a scenario in a generic way in QEMU? (+Peter)

Well, I think I've suggested you to have a look at how things is done in for handling such compatibility in e1000_properties.

Quoted 4 lines

The POR reset value is wrong in QEMU and has carried forward the wrong value for years, and correcting it to its right value needs to do what?

We should stick to the wrong behavior for old machine types.

That's all.

Quoted 4 lines

Regards,
Bin

author
#9 lore raw
Quoted 67 lines · Jason Wang
Quoted 49 lines · Bin Meng
Quoted 48 lines · Jason Wang
Quoted 45 lines · Bin Meng
Quoted 40 lines · Bin Meng
Quoted 39 lines · Jason Wang
Quoted 32 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

    hw/net/e1000.c | 2 ++
    1 file changed, 2 insertions(+)
hw/net/e1000.c +1 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
    #include "hw/pci/pci.h"
    #include "hw/qdev-properties.h"
    #include "migration/vmstate.h"
+#include "net/eth.h"
    #include "net/net.h"
    #include "net/checksum.h"
    #include "sysemu/sysemu.h"

hw/net/e1000.c static const uint32_t mac_reg_init[] = { · line 258

Quoted 4 lines
hw/net/e1000.c +1 -0
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
        [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                    E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                    E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

Quoted 3 lines
Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Ping?

Quoted 8 lines
Quoted 6 lines

I admit migration "works" but it doesn't mean it's not broken. It changes the guest visible default value of VET register, so it may break things silently for the guest.

For old machine types, we should stick the value to the one without this fix.

Could you please propose a solution on how to handle such a scenario in a generic way in QEMU? (+Peter)

Well, I think I've suggested you to have a look at how things is done in for handling such compatibility in e1000_properties.

Quoted 1 line
Quoted 8 lines
Quoted 3 lines

The POR reset value is wrong in QEMU and has carried forward the wrong value for years, and correcting it to its right value needs to do what?

We should stick to the wrong behavior for old machine types.

That's all.

So that means the following SD patch is also wrong (+Philippe) which changes the default value of capability register.
http://patchwork.ozlabs.org/project/qemu-devel/patch/[email protected]/

Can we get some agreement among maintainers?

Regards,
Bin

#10 lore raw
Quoted 67 lines · Bin Meng
Quoted 66 lines · Jason Wang
Quoted 47 lines · Bin Meng
Quoted 46 lines · Jason Wang
Quoted 45 lines · Bin Meng
Quoted 40 lines · Bin Meng
Quoted 39 lines · Jason Wang
Quoted 32 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

     hw/net/e1000.c | 2 ++
     1 file changed, 2 insertions(+)
hw/net/e1000.c +1 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
     #include "hw/pci/pci.h"
     #include "hw/qdev-properties.h"
     #include "migration/vmstate.h"
+#include "net/eth.h"
     #include "net/net.h"
     #include "net/checksum.h"
     #include "sysemu/sysemu.h"

hw/net/e1000.c static const uint32_t mac_reg_init[] = { · line 258

Quoted 4 lines
hw/net/e1000.c +1 -0
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
         [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                     E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                     E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

Quoted 3 lines
Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Ping?

Quoted 8 lines
Quoted 6 lines

I admit migration "works" but it doesn't mean it's not broken. It changes the guest visible default value of VET register, so it may break things silently for the guest.

For old machine types, we should stick the value to the one without this fix.

Could you please propose a solution on how to handle such a scenario in a generic way in QEMU? (+Peter)

Well, I think I've suggested you to have a look at how things is done in for handling such compatibility in e1000_properties.

Quoted 3 lines

The POR reset value is wrong in QEMU and has carried forward the wrong value for years, and correcting it to its right value needs to do what?

Quoted 6 lines
Quoted 3 lines

We should stick to the wrong behavior for old machine types.

That's all.

So that means the following SD patch is also wrong (+Philippe) which changes the default value of capability register.
http://patchwork.ozlabs.org/project/qemu-devel/patch/[email protected]/

It should compat capareg for the old value for old machine types.

Quoted 2 lines

Can we get some agreement among maintainers?

It's not about the agreement but about to have a stable ABI. I don't know the case for sd but e1000 is used in various  and we work hard to unbreak the migration compatibility among downstream versions. Git log on e1000.c will tell you more.

Thanks

Quoted 4 lines

Regards,
Bin

author
#11 lore raw
Quoted 72 lines · Jason Wang
Quoted 69 lines · Bin Meng
Quoted 68 lines · Jason Wang
Quoted 47 lines · Bin Meng
Quoted 46 lines · Jason Wang
Quoted 45 lines · Bin Meng
Quoted 40 lines · Bin Meng
Quoted 39 lines · Jason Wang
Quoted 32 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

     hw/net/e1000.c | 2 ++
     1 file changed, 2 insertions(+)
hw/net/e1000.c +1 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
     #include "hw/pci/pci.h"
     #include "hw/qdev-properties.h"
     #include "migration/vmstate.h"
+#include "net/eth.h"
     #include "net/net.h"
     #include "net/checksum.h"
     #include "sysemu/sysemu.h"

hw/net/e1000.c static const uint32_t mac_reg_init[] = { · line 258

Quoted 4 lines
hw/net/e1000.c +1 -0
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
         [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                     E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                     E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

Quoted 3 lines
Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Ping?

Quoted 8 lines
Quoted 6 lines

I admit migration "works" but it doesn't mean it's not broken. It changes the guest visible default value of VET register, so it may break things silently for the guest.

For old machine types, we should stick the value to the one without this fix.

Could you please propose a solution on how to handle such a scenario in a generic way in QEMU? (+Peter)

Well, I think I've suggested you to have a look at how things is done in for handling such compatibility in e1000_properties.

Quoted 3 lines

The POR reset value is wrong in QEMU and has carried forward the wrong value for years, and correcting it to its right value needs to do what?

We should stick to the wrong behavior for old machine types.

Quoted 7 lines
Quoted 4 lines
Quoted 1 line

That's all.

So that means the following SD patch is also wrong (+Philippe) which changes the default value of capability register.
http://patchwork.ozlabs.org/project/qemu-devel/patch/[email protected]/

It should compat capareg for the old value for old machine types.

Yeah, it's already a property for the SD controller model but someone views it as a bug because the model implements 64-bit but not reporting it in the capability register.

Quoted 10 lines
Quoted 4 lines

Can we get some agreement among maintainers?

It's not about the agreement but about to have a stable ABI. I don't know the case for sd but e1000 is used in various and we work hard to unbreak the migration compatibility among downstream versions. Git log on e1000.c will tell you more.

Agreement or stable ABI, whatever we call, but we should be in some consistency.

IMHO maintainers should reach an agreement to some extent on how compatibility should be achieved. I just found silly to add a property to fix a real bug in the model, and we preserve the bug all over releases.

I can find plenty of examples in the current QEMU tree that were accepted that changed the bugous register behavior, but it was not asked to add new properties to keep the bugos behavior.

e.g.: commit ce8e43760e8e ("hw/net: fsl_etsec: Reverse the RCTRL.RSF logic")

Regards,
Bin

#12 lore raw
Quoted 86 lines · Bin Meng
Quoted 66 lines · Jason Wang
Quoted 64 lines · Bin Meng
Quoted 63 lines · Jason Wang
Quoted 47 lines · Bin Meng
Quoted 46 lines · Jason Wang
Quoted 45 lines · Bin Meng
Quoted 40 lines · Bin Meng
Quoted 39 lines · Jason Wang
Quoted 32 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

      hw/net/e1000.c | 2 ++
      1 file changed, 2 insertions(+)
hw/net/e1000.c +1 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
      #include "hw/pci/pci.h"
      #include "hw/qdev-properties.h"
      #include "migration/vmstate.h"
+#include "net/eth.h"
      #include "net/net.h"
      #include "net/checksum.h"
      #include "sysemu/sysemu.h"

hw/net/e1000.c static const uint32_t mac_reg_init[] = { · line 258

Quoted 4 lines
hw/net/e1000.c +1 -0
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
          [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                      E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                      E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

Quoted 3 lines
Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Ping?

Quoted 8 lines
Quoted 6 lines

I admit migration "works" but it doesn't mean it's not broken. It changes the guest visible default value of VET register, so it may break things silently for the guest.

For old machine types, we should stick the value to the one without this fix.

Could you please propose a solution on how to handle such a scenario in a generic way in QEMU? (+Peter)

Well, I think I've suggested you to have a look at how things is done in for handling such compatibility in e1000_properties.

Quoted 3 lines

The POR reset value is wrong in QEMU and has carried forward the wrong value for years, and correcting it to its right value needs to do what?

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

We should stick to the wrong behavior for old machine types.

That's all.

So that means the following SD patch is also wrong (+Philippe) which changes the default value of capability register.
http://patchwork.ozlabs.org/project/qemu-devel/patch/[email protected]/

It should compat capareg for the old value for old machine types.

Yeah, it's already a property for the SD controller model but someone views it as a bug because the model implements 64-bit but not reporting it in the capability register.

Quoted 7 lines
Quoted 2 lines

Can we get some agreement among maintainers?

It's not about the agreement but about to have a stable ABI. I don't know the case for sd but e1000 is used in various and we work hard to unbreak the migration compatibility among downstream versions. Git log on e1000.c will tell you more.

Quoted 6 lines

Agreement or stable ABI, whatever we call, but we should be in some consistency.

IMHO maintainers should reach an agreement to some extent on how compatibility should be achieved. I just found silly to add a property to fix a real bug in the model, and we preserve the bug all over releases.

That's the price for the stable ABI. See one of my recent fix - d83f46d189 virtio-pci: compat page aligned ATS. It keeps the "buggy" behavior to unbreak the migration.

Quoted 6 lines

I can find plenty of examples in the current QEMU tree that were accepted that changed the bugous register behavior, but it was not asked to add new properties to keep the bugos behavior.

e.g.: commit ce8e43760e8e ("hw/net: fsl_etsec: Reverse the RCTRL.RSF logic")

I guess it's simply because fsl_etsec is not used in any distributions/production environments or the maintainer may just not notice things like this.

But for e1000(e), we should stick to a stable ABI for consistency. Otherwise it would be very tricky to fix them after we saw real issues. We had learnt a lot during the past decade.

Thanks

Quoted 4 lines

Regards,
Bin

author
#13 lore raw
Quoted 97 lines · Jason Wang
Quoted 86 lines · Bin Meng
Quoted 66 lines · Jason Wang
Quoted 64 lines · Bin Meng
Quoted 63 lines · Jason Wang
Quoted 47 lines · Bin Meng
Quoted 46 lines · Jason Wang
Quoted 45 lines · Bin Meng
Quoted 40 lines · Bin Meng
Quoted 39 lines · Jason Wang
Quoted 32 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

      hw/net/e1000.c | 2 ++
      1 file changed, 2 insertions(+)
hw/net/e1000.c +1 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
      #include "hw/pci/pci.h"
      #include "hw/qdev-properties.h"
      #include "migration/vmstate.h"
+#include "net/eth.h"
      #include "net/net.h"
      #include "net/checksum.h"
      #include "sysemu/sysemu.h"

hw/net/e1000.c static const uint32_t mac_reg_init[] = { · line 258

Quoted 4 lines
hw/net/e1000.c +1 -0
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
          [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                      E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                      E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

Quoted 3 lines
Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Ping?

Quoted 8 lines
Quoted 6 lines

I admit migration "works" but it doesn't mean it's not broken. It changes the guest visible default value of VET register, so it may break things silently for the guest.

For old machine types, we should stick the value to the one without this fix.

Could you please propose a solution on how to handle such a scenario in a generic way in QEMU? (+Peter)

Well, I think I've suggested you to have a look at how things is done in for handling such compatibility in e1000_properties.

Quoted 3 lines

The POR reset value is wrong in QEMU and has carried forward the wrong value for years, and correcting it to its right value needs to do what?

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

We should stick to the wrong behavior for old machine types.

That's all.

So that means the following SD patch is also wrong (+Philippe) which changes the default value of capability register.
http://patchwork.ozlabs.org/project/qemu-devel/patch/[email protected]/

It should compat capareg for the old value for old machine types.

Yeah, it's already a property for the SD controller model but someone views it as a bug because the model implements 64-bit but not reporting it in the capability register.

Quoted 7 lines
Quoted 2 lines

Can we get some agreement among maintainers?

It's not about the agreement but about to have a stable ABI. I don't know the case for sd but e1000 is used in various and we work hard to unbreak the migration compatibility among downstream versions. Git log on e1000.c will tell you more.

Quoted 6 lines

Agreement or stable ABI, whatever we call, but we should be in some consistency.

IMHO maintainers should reach an agreement to some extent on how compatibility should be achieved. I just found silly to add a property to fix a real bug in the model, and we preserve the bug all over releases.

Quoted 4 lines

That's the price for the stable ABI. See one of my recent fix - d83f46d189 virtio-pci: compat page aligned ATS. It keeps the "buggy" behavior to unbreak the migration.

But this series does not break the migration, as we discussed in the previous thread.

Quoted 8 lines
Quoted 6 lines

I can find plenty of examples in the current QEMU tree that were accepted that changed the bugous register behavior, but it was not asked to add new properties to keep the bugos behavior.

e.g.: commit ce8e43760e8e ("hw/net: fsl_etsec: Reverse the RCTRL.RSF logic")

Quoted 8 lines

I guess it's simply because fsl_etsec is not used in any distributions/production environments or the maintainer may just not notice things like this.

But for e1000(e), we should stick to a stable ABI for consistency. Otherwise it would be very tricky to fix them after we saw real issues. We had learnt a lot during the past decade.

Okay, do we have such a kind of widely used device model list? And we should document such a process that we should keep compatibility on these devices as well.

Regarding this VET register, do you know what guest relies on the POR value which is zero? Zero is not a valid ethernet VLAN type. I don't think changing this will break any guests. The commit message says Linux e1000(e) driver just rewrites this register by itself. Given that's probably the most commonly used OS, it's not strange this bug gets unnoticed for years.

Regards,
Bin

#14 lore raw
Quoted 92 lines · Bin Meng
Quoted 91 lines · Jason Wang
Quoted 88 lines · Bin Meng
Quoted 65 lines · Jason Wang
Quoted 64 lines · Bin Meng
Quoted 63 lines · Jason Wang
Quoted 47 lines · Bin Meng
Quoted 46 lines · Jason Wang
Quoted 45 lines · Bin Meng
Quoted 40 lines · Bin Meng
Quoted 39 lines · Jason Wang
Quoted 32 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is not always the case for everyone. Drivers relying on the reset value of VET won't be able to transmit and receive VLAN frames in QEMU.

---

(no changes since v1)

       hw/net/e1000.c | 2 ++
       1 file changed, 2 insertions(+)
hw/net/e1000.c +1 -0
index 4f75b44cfc..20cbba6411 100644
@@ -29,6 +29,7 @@
       #include "hw/pci/pci.h"
       #include "hw/qdev-properties.h"
       #include "migration/vmstate.h"
+#include "net/eth.h"
       #include "net/net.h"
       #include "net/checksum.h"
       #include "sysemu/sysemu.h"

hw/net/e1000.c static const uint32_t mac_reg_init[] = { · line 258

Quoted 4 lines
hw/net/e1000.c +1 -0
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
           [MANC]    = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
                       E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
                       E1000_MANC_RMCP_EN,
+    [VET]     = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

Quoted 3 lines
Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Ping?

Quoted 8 lines
Quoted 6 lines

I admit migration "works" but it doesn't mean it's not broken. It changes the guest visible default value of VET register, so it may break things silently for the guest.

For old machine types, we should stick the value to the one without this fix.

Could you please propose a solution on how to handle such a scenario in a generic way in QEMU? (+Peter)

Well, I think I've suggested you to have a look at how things is done in for handling such compatibility in e1000_properties.

Quoted 3 lines

The POR reset value is wrong in QEMU and has carried forward the wrong value for years, and correcting it to its right value needs to do what?

Quoted 7 lines
Quoted 6 lines
Quoted 3 lines

We should stick to the wrong behavior for old machine types.

That's all.

So that means the following SD patch is also wrong (+Philippe) which changes the default value of capability register.
http://patchwork.ozlabs.org/project/qemu-devel/patch/[email protected]/

It should compat capareg for the old value for old machine types.

Yeah, it's already a property for the SD controller model but someone views it as a bug because the model implements 64-bit but not reporting it in the capability register.

Quoted 5 lines
Quoted 1 line

Can we get some agreement among maintainers?

It's not about the agreement but about to have a stable ABI. I don't know the case for sd but e1000 is used in various and we work hard to unbreak the migration compatibility among downstream versions. Git log on e1000.c will tell you more.

Agreement or stable ABI, whatever we call, but we should be in some consistency.

IMHO maintainers should reach an agreement to some extent on how compatibility should be achieved. I just found silly to add a property to fix a real bug in the model, and we preserve the bug all over releases.

Quoted 6 lines
Quoted 4 lines

That's the price for the stable ABI. See one of my recent fix - d83f46d189 virtio-pci: compat page aligned ATS. It keeps the "buggy" behavior to unbreak the migration.

But this series does not break the migration, as we discussed in the previous thread.

It actually did,

(qemu) qemu-kvm: get_pci_config_device: Bad config data: i=0x104 read:
    0 device: 20 cmask: ff wmask: 0 w1cmask:0

Since the register is RO.

Quoted 11 lines
Quoted 11 lines
Quoted 6 lines

I can find plenty of examples in the current QEMU tree that were accepted that changed the bugous register behavior, but it was not asked to add new properties to keep the bugos behavior.

e.g.: commit ce8e43760e8e ("hw/net: fsl_etsec: Reverse the RCTRL.RSF logic")

I guess it's simply because fsl_etsec is not used in any distributions/production environments or the maintainer may just not notice things like this.

Quoted 7 lines
Quoted 4 lines

But for e1000(e), we should stick to a stable ABI for consistency. Otherwise it would be very tricky to fix them after we saw real issues. We had learnt a lot during the past decade.

Okay, do we have such a kind of widely used device model list? And we should document such a process that we should keep compatibility on these devices as well.

I can only say for networking devices:

- e1000(e)
- rtl8139
- virtio-net

Quoted 3 lines

Regarding this VET register, do you know what guest relies on the POR value which is zero?

I don't know.

Quoted 2 lines

Zero is not a valid ethernet VLAN type. I don't
think changing this will break any guests.

You might be right here, but it would be late if we find it breaks any one.

If it's not a lot of work, I tend to bother with compat stuffs for this.

Thanks

Quoted 8 lines

The commit message says
Linux e1000(e) driver just rewrites this register by itself. Given that's probably the most commonly used OS, it's not strange this bug gets unnoticed for years.

Regards,
Bin

author
#15 lore raw
Quoted 103 lines · Jason Wang
Quoted 92 lines · Bin Meng
Quoted 91 lines · Jason Wang
Quoted 88 lines · Bin Meng
Quoted 65 lines · Jason Wang
Quoted 64 lines · Bin Meng
Quoted 63 lines · Jason Wang
Quoted 47 lines · Bin Meng
Quoted 46 lines · Jason Wang
Quoted 45 lines · Bin Meng
Quoted 40 lines · Bin Meng
Quoted 39 lines · Jason Wang
Quoted 36 lines · Bin Meng

From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per
the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is
not always the case for everyone. Drivers relying on the reset value
of VET won't be able to transmit and receive VLAN frames in QEMU.

Reported-by: Markus Carlstedt <[email protected]>
Signed-off-by: Christina Wang <[email protected]>
Signed-off-by: Bin Meng <[email protected]>
---

(no changes since v1)

hw/net/e1000.c | 2 ++
1 file changed, 2 insertions(+)

diff --git a/hw/net/e1000.c b/hw/net/e1000.c
index 4f75b44cfc..20cbba6411 100644
--- a/hw/net/e1000.c
+++ b/hw/net/e1000.c
@@ -29,6 +29,7 @@
#include "hw/pci/pci.h"
#include "hw/qdev-properties.h"
#include "migration/vmstate.h"
+#include "net/eth.h"
#include "net/net.h"
#include "net/checksum.h"
#include "sysemu/sysemu.h"
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
[MANC] = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
E1000_MANC_RMCP_EN,
+ [VET] = ETH_P_VLAN,

I wonder if we need a compat flag for this, since we change the behavior.

Quoted 3 lines
Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Ping?

Quoted 8 lines
Quoted 6 lines

I admit migration "works" but it doesn't mean it's not broken. It changes the guest visible default value of VET register, so it may break things silently for the guest.

For old machine types, we should stick the value to the one without this fix.

Could you please propose a solution on how to handle such a scenario in a generic way in QEMU? (+Peter)

Well, I think I've suggested you to have a look at how things is done in for handling such compatibility in e1000_properties.

Quoted 3 lines

The POR reset value is wrong in QEMU and has carried forward the wrong value for years, and correcting it to its right value needs to do what?

Quoted 7 lines
Quoted 6 lines
Quoted 3 lines

We should stick to the wrong behavior for old machine types.

That's all.

So that means the following SD patch is also wrong (+Philippe) which changes the default value of capability register.
http://patchwork.ozlabs.org/project/qemu-devel/patch/[email protected]/

It should compat capareg for the old value for old machine types.

Yeah, it's already a property for the SD controller model but someone views it as a bug because the model implements 64-bit but not reporting it in the capability register.

Quoted 5 lines
Quoted 1 line

Can we get some agreement among maintainers?

It's not about the agreement but about to have a stable ABI. I don't know the case for sd but e1000 is used in various and we work hard to unbreak the migration compatibility among downstream versions. Git log on e1000.c will tell you more.

Agreement or stable ABI, whatever we call, but we should be in some consistency.

IMHO maintainers should reach an agreement to some extent on how compatibility should be achieved. I just found silly to add a property to fix a real bug in the model, and we preserve the bug all over releases.

Quoted 6 lines
Quoted 4 lines

That's the price for the stable ABI. See one of my recent fix - d83f46d189 virtio-pci: compat page aligned ATS. It keeps the "buggy" behavior to unbreak the migration.

But this series does not break the migration, as we discussed in the previous thread.

Quoted 6 lines

It actually did,

(qemu) qemu-kvm: get_pci_config_device: Bad config data: i=0x104 read:
0 device: 20 cmask: ff wmask: 0 w1cmask:0

Since the register is RO.

Did you mean the VET register is read-only? I don't understand this as the manual says it's RW and Linux driver programs it to 0x8100.

Quoted 22 lines
Quoted 13 lines
Quoted 13 lines
Quoted 8 lines

I can find plenty of examples in the current QEMU tree that were accepted that changed the bugous register behavior, but it was not asked to add new properties to keep the bugos behavior.

e.g.: commit ce8e43760e8e ("hw/net: fsl_etsec: Reverse the RCTRL.RSF logic")

I guess it's simply because fsl_etsec is not used in any distributions/production environments or the maintainer may just not notice things like this.

Quoted 7 lines
Quoted 4 lines

But for e1000(e), we should stick to a stable ABI for consistency. Otherwise it would be very tricky to fix them after we saw real issues. We had learnt a lot during the past decade.

Okay, do we have such a kind of widely used device model list? And we should document such a process that we should keep compatibility on these devices as well.

Quoted 6 lines

I can only say for networking devices:

- e1000(e)
- rtl8139
- virtio-net

Please, at least have these documented as a development process.

Quoted 8 lines
Quoted 3 lines

Regarding this VET register, do you know what guest relies on the POR value which is zero?

I don't know.

Quoted 8 lines
Quoted 2 lines

Zero is not a valid ethernet VLAN type. I don't
think changing this will break any guests.

You might be right here, but it would be late if we find it breaks any one.

If it's not a lot of work, I tend to bother with compat stuffs for this.

Regards,
Bin

#16 lore raw
Quoted 100 lines · Bin Meng
Quoted 99 lines · Jason Wang
Quoted 96 lines · Bin Meng
Quoted 85 lines · Jason Wang
Quoted 84 lines · Bin Meng
Quoted 65 lines · Jason Wang
Quoted 64 lines · Bin Meng
Quoted 63 lines · Jason Wang
Quoted 47 lines · Bin Meng
Quoted 46 lines · Jason Wang
Quoted 45 lines · Bin Meng
Quoted 40 lines · Bin Meng
Quoted 39 lines · Jason Wang

在 2021/7/2 下午5:24, Bin Meng 写道:
From: Christina Wang <[email protected]>

The initial value of VLAN Ether Type (VET) register is 0x8100, as per
the manual and real hardware.

While Linux e1000 driver always writes VET register to 0x8100, it is
not always the case for everyone. Drivers relying on the reset value
of VET won't be able to transmit and receive VLAN frames in QEMU.

Reported-by: Markus Carlstedt <[email protected]>
Signed-off-by: Christina Wang <[email protected]>
Signed-off-by: Bin Meng <[email protected]>
---

(no changes since v1)

hw/net/e1000.c | 2 ++
1 file changed, 2 insertions(+)

diff --git a/hw/net/e1000.c b/hw/net/e1000.c
index 4f75b44cfc..20cbba6411 100644
--- a/hw/net/e1000.c
+++ b/hw/net/e1000.c
@@ -29,6 +29,7 @@
#include "hw/pci/pci.h"
#include "hw/qdev-properties.h"
#include "migration/vmstate.h"
+#include "net/eth.h"
#include "net/net.h"
#include "net/checksum.h"
#include "sysemu/sysemu.h"
@@ -254,6 +255,7 @@ static const uint32_t mac_reg_init[] = {
[MANC] = E1000_MANC_EN_MNG2HOST | E1000_MANC_RCV_TCO_EN |
E1000_MANC_ARP_EN | E1000_MANC_0298_EN |
E1000_MANC_RMCP_EN,
+ [VET] = ETH_P_VLAN,
I wonder if we need a compat flag for this, since we change the behavior.

Quoted 3 lines
Quoted 2 lines

(See e1000_properties[])

No we don't need to since it does not break migration.

Ping?

Quoted 8 lines
Quoted 6 lines

I admit migration "works" but it doesn't mean it's not broken. It changes the guest visible default value of VET register, so it may break things silently for the guest.

For old machine types, we should stick the value to the one without this fix.

Could you please propose a solution on how to handle such a scenario in a generic way in QEMU? (+Peter)

Well, I think I've suggested you to have a look at how things is done in for handling such compatibility in e1000_properties.

Quoted 3 lines

The POR reset value is wrong in QEMU and has carried forward the wrong value for years, and correcting it to its right value needs to do what?

Quoted 7 lines
Quoted 6 lines
Quoted 3 lines

We should stick to the wrong behavior for old machine types.

That's all.

So that means the following SD patch is also wrong (+Philippe) which changes the default value of capability register.
http://patchwork.ozlabs.org/project/qemu-devel/patch/[email protected]/

It should compat capareg for the old value for old machine types.

Yeah, it's already a property for the SD controller model but someone views it as a bug because the model implements 64-bit but not reporting it in the capability register.

Quoted 5 lines
Quoted 1 line

Can we get some agreement among maintainers?

It's not about the agreement but about to have a stable ABI. I don't know the case for sd but e1000 is used in various and we work hard to unbreak the migration compatibility among downstream versions. Git log on e1000.c will tell you more.

Agreement or stable ABI, whatever we call, but we should be in some consistency.

Quoted 8 lines
Quoted 4 lines

IMHO maintainers should reach an agreement to some extent on how compatibility should be achieved. I just found silly to add a property to fix a real bug in the model, and we preserve the bug all over releases.

That's the price for the stable ABI. See one of my recent fix - d83f46d189 virtio-pci: compat page aligned ATS. It keeps the "buggy" behavior to unbreak the migration.

But this series does not break the migration, as we discussed in the previous thread.

Quoted 7 lines
Quoted 6 lines

It actually did,

(qemu) qemu-kvm: get_pci_config_device: Bad config data: i=0x104 read:
0 device: 20 cmask: ff wmask: 0 w1cmask:0

Since the register is RO.

Did you mean the VET register is read-only?

No, I meant for ATS.

Quoted 2 lines

I don't understand this as
the manual says it's RW and Linux driver programs it to 0x8100.

So ATS cap is read only so it can be validated by qemu.

VET is RW so qemu can't do similar checks.

Quoted 19 lines
Quoted 19 lines
Quoted 18 lines
Quoted 6 lines
Quoted 6 lines

I can find plenty of examples in the current QEMU tree that were accepted that changed the bugous register behavior, but it was not asked to add new properties to keep the bugos behavior.

Quoted 9 lines
Quoted 1 line

e.g.: commit ce8e43760e8e ("hw/net: fsl_etsec: Reverse the RCTRL.RSF logic")

I guess it's simply because fsl_etsec is not used in any distributions/production environments or the maintainer may just not notice things like this.

But for e1000(e), we should stick to a stable ABI for consistency. Otherwise it would be very tricky to fix them after we saw real issues. We had learnt a lot during the past decade.

Okay, do we have such a kind of widely used device model list? And we should document such a process that we should keep compatibility on these devices as well.

Quoted 7 lines
Quoted 6 lines

I can only say for networking devices:

- e1000(e)
- rtl8139
- virtio-net

Please, at least have these documented as a development process.

See Peter's reply. We should take care of the devices that is involved in at least one versioned machine types.

If you wish, you can post a patch to clarify things in the doc.

Thanks

Quoted 17 lines
Quoted 7 lines
Quoted 3 lines

Regarding this VET register, do you know what guest relies on the POR value which is zero?

I don't know.

Quoted 7 lines
Quoted 2 lines

Zero is not a valid ethernet VLAN type. I don't
think changing this will break any guests.

You might be right here, but it would be late if we find it breaks any one.

If it's not a lot of work, I tend to bother with compat stuffs for this.

Regards,
Bin

#17 lore raw
In reply to Bin Meng
Quoted 20 lines · Bin Meng
Quoted 10 lines · Jason Wang
Quoted 1 line · Bin Meng

Can we get some agreement among maintainers?

It's not about the agreement but about to have a stable ABI. I don't know the case for sd but e1000 is used in various and we work hard to unbreak the migration compatibility among downstream versions. Git log on e1000.c will tell you more.

Agreement or stable ABI, whatever we call, but we should be in some consistency.

IMHO maintainers should reach an agreement to some extent on how compatibility should be achieved. I just found silly to add a property to fix a real bug in the model, and we preserve the bug all over releases.

Quoted 5 lines

I can find plenty of examples in the current QEMU tree that were accepted that changed the bugous register behavior, but it was not asked to add new properties to keep the bugos behavior.

e.g.: commit ce8e43760e8e ("hw/net: fsl_etsec: Reverse the RCTRL.RSF logic")

There is basically a judgement call going on here about whether the device is "significant enough" that it's worth the effort of preserving back-compatibility of bugs.

There is at least one clear rule: if the device isn't used by any machine with a versioned machine type, then there is no need to provide back-compatibility of old buggy behaviour. (There would be no way for the user to select the old behaviour by choosing a -4.2 machine type.) This is why the fsl_etsec device doesn't need to handle that.

For PCI devices it's a bit fuzzier because in theory you can plug any PCI card into a versioned x86 PC machine.

We don't want to preserve bug-compatibility for absolutely everything, because the codebase would quickly clog up with weird behaviour that we never test and which is not of any use to users either. So you have to look at:
* how easy is providing the back-compat?
* how commonly used in production is the device?
* how likely is it that guests might care?
* would the behaviour change cause odd behaviour across a cross-version migration from the old QEMU?

Migration compat is similar, but not quite the same because the compatibility handling tends to be less invasive, so we take the "provide compat" choice more often. For non-versioned machine types, again, we're OK with breaking back-compat as long as we bump a migration version number so the user gets an error message rather than weird behaviour.

thanks
-- PMM

author
#18 lore raw
Quoted 52 lines · Peter Maydell
Quoted 20 lines · Bin Meng
Quoted 10 lines · Jason Wang
Quoted 1 line · Bin Meng

Can we get some agreement among maintainers?

It's not about the agreement but about to have a stable ABI. I don't know the case for sd but e1000 is used in various and we work hard to unbreak the migration compatibility among downstream versions. Git log on e1000.c will tell you more.

Agreement or stable ABI, whatever we call, but we should be in some consistency.

IMHO maintainers should reach an agreement to some extent on how compatibility should be achieved. I just found silly to add a property to fix a real bug in the model, and we preserve the bug all over releases.

Quoted 5 lines

I can find plenty of examples in the current QEMU tree that were accepted that changed the bugous register behavior, but it was not asked to add new properties to keep the bugos behavior.

e.g.: commit ce8e43760e8e ("hw/net: fsl_etsec: Reverse the RCTRL.RSF logic")

There is basically a judgement call going on here about whether the device is "significant enough" that it's worth the effort of preserving back-compatibility of bugs.

There is at least one clear rule: if the device isn't used by any machine with a versioned machine type, then there is no need to provide back-compatibility of old buggy behaviour. (There would be no way for the user to select the old behaviour by choosing a -4.2 machine type.) This is why the fsl_etsec device doesn't need to handle that.

For PCI devices it's a bit fuzzier because in theory you can plug any PCI card into a versioned x86 PC machine.

We don't want to preserve bug-compatibility for absolutely everything, because the codebase would quickly clog up with weird behaviour that we never test and which is not of any use to users either. So you have to look at:
* how easy is providing the back-compat?
* how commonly used in production is the device?
* how likely is it that guests might care?
* would the behaviour change cause odd behaviour across a cross-version migration from the old QEMU?

Quoted 6 lines

Migration compat is similar, but not quite the same because the compatibility handling tends to be less invasive, so we take the "provide compat" choice more often. For non-versioned machine types, again, we're OK with breaking back-compat as long as we bump a migration version number so the user gets an error message rather than weird behaviour.

Thanks Peter. I think the above information can be put in a doc in docs/devel, and with some real examples to help developers on how to achieve backward compatibility.

Regarding this series, as I mentioned in an earlier thread, I believe the possibility of breaking a guest is very low. Adding a back-compat check is not that hard either. Just not sure which factor weighs more.

Regards,
Bin

#19 lore raw
Quoted 53 lines · Bin Meng
Quoted 46 lines · Peter Maydell
Quoted 16 lines · Bin Meng
Quoted 8 lines · Jason Wang
Quoted 1 line · Bin Meng

Can we get some agreement among maintainers?

It's not about the agreement but about to have a stable ABI. I don't know the case for sd but e1000 is used in various and we work hard to unbreak the migration compatibility among downstream versions. Git log on e1000.c will tell you more.

Agreement or stable ABI, whatever we call, but we should be in some consistency.

IMHO maintainers should reach an agreement to some extent on how compatibility should be achieved. I just found silly to add a property to fix a real bug in the model, and we preserve the bug all over releases.

Quoted 5 lines

I can find plenty of examples in the current QEMU tree that were accepted that changed the bugous register behavior, but it was not asked to add new properties to keep the bugos behavior.

e.g.: commit ce8e43760e8e ("hw/net: fsl_etsec: Reverse the RCTRL.RSF logic")

There is basically a judgement call going on here about whether the device is "significant enough" that it's worth the effort of preserving back-compatibility of bugs.

There is at least one clear rule: if the device isn't used by any machine with a versioned machine type, then there is no need to provide back-compatibility of old buggy behaviour. (There would be no way for the user to select the old behaviour by choosing a -4.2 machine type.) This is why the fsl_etsec device doesn't need to handle that.

For PCI devices it's a bit fuzzier because in theory you can plug any PCI card into a versioned x86 PC machine.

We don't want to preserve bug-compatibility for absolutely everything, because the codebase would quickly clog up with weird behaviour that we never test and which is not of any use to users either. So you have to look at:
* how easy is providing the back-compat?
* how commonly used in production is the device?
* how likely is it that guests might care?
* would the behaviour change cause odd behaviour across a cross-version migration from the old QEMU?

Quoted 6 lines

Migration compat is similar, but not quite the same because the compatibility handling tends to be less invasive, so we take the "provide compat" choice more often. For non-versioned machine types, again, we're OK with breaking back-compat as long as we bump a migration version number so the user gets an error message rather than weird behaviour.

Quoted 7 lines

Thanks Peter. I think the above information can be put in a doc in docs/devel, and with some real examples to help developers on how to achieve backward compatibility.

Regarding this series, as I mentioned in an earlier thread, I believe the possibility of breaking a guest is very low. Adding a back-compat check is not that hard either. Just not sure which factor weighs more.

Just few lines of code for the compatibility. Let's do that.

Thanks

Quoted 4 lines

Regards,
Bin