Skip to contentSkip to conversation

hw/net: npcm7xx_emc: set MAC in register space

PATCH Superseded Patrick Venture 3 participants 9 messages Sep 21, 2022 – Sep 30, 2022

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 hw/net: npcm7xx_emc: set MAC in register space

The MAC address set from Qemu wasn't being saved into the register space.

---
 hw/net/npcm7xx_emc.c | 15 +++++++++++++++
 1 file changed, 15 insertions(+)
hw/net/npcm7xx_emc.c +15 -0
index 7c86bb52e5..6be1008529 100644
@@ -96,6 +96,9 @@ static const char *emc_reg_name(int regno)
 #undef REG
 }
 
+static void npcm7xx_emc_write(void *opaque, hwaddr offset,
+                              uint64_t v, unsigned size);
+
 static void emc_reset(NPCM7xxEMCState *emc)
 {
     trace_npcm7xx_emc_reset(emc->emc_num);
@@ -112,6 +115,18 @@ static void emc_reset(NPCM7xxEMCState *emc)
 
     emc->tx_active = false;
     emc->rx_active = false;
+
+    /* Set the MAC address in the register space. */
+    uint32_t value = (emc->conf.macaddr.a[0] << 24) |
+        (emc->conf.macaddr.a[1] << 16) |
+        (emc->conf.macaddr.a[2] << 8) |
+        emc->conf.macaddr.a[3];
+    npcm7xx_emc_write(emc, REG_CAMM_BASE * sizeof(uint32_t), value,
+                      sizeof(uint32_t));
+
+    value = (emc->conf.macaddr.a[4] << 24) | (emc->conf.macaddr.a[5] << 16);
+    npcm7xx_emc_write(emc, REG_CAML_BASE * sizeof(uint32_t), value,
+                      sizeof(uint32_t));
 }
 
 static void npcm7xx_emc_reset(DeviceState *dev)
Signature · 1 line
2.37.3.998.g577e59143f-goog
#1 lore raw
Quoted 12 lines · Patrick Venture

The MAC address set from Qemu wasn't being saved into the register space.

hw/net/npcm7xx_emc.c +2 -0
@@ -112,6 +115,18 @@ static void emc_reset(NPCM7xxEMCState *emc)
 
     emc->tx_active = false;
     emc->rx_active = false;
+
+    /* Set the MAC address in the register space. */

hw/net/npcm7xx_emc.c static void emc_reset(NPCM7xxEMCState *emc) · line 129

Quoted 10 lines
hw/net/npcm7xx_emc.c +10 -0
+    uint32_t value = (emc->conf.macaddr.a[0] << 24) |
+        (emc->conf.macaddr.a[1] << 16) |
+        (emc->conf.macaddr.a[2] << 8) |
+        emc->conf.macaddr.a[3];
+    npcm7xx_emc_write(emc, REG_CAMM_BASE * sizeof(uint32_t), value,
+                      sizeof(uint32_t));
+
+    value = (emc->conf.macaddr.a[4] << 24) | (emc->conf.macaddr.a[5] << 16);
+    npcm7xx_emc_write(emc, REG_CAML_BASE * sizeof(uint32_t), value,
+                      sizeof(uint32_t));

If I understand correctly, the issue here is that emc->regs[REG_CAMM_BASE] and emc->regs[REG_CAML_BASE] aren't being reset correctly. If so, I think the better approach is to simply reset them here, without going through the register-write function, the same way we already do for the handful of other registers which have non-zero reset values. That's the way other devices seem to do it.

A question to which I don't know the answer: if the guest writes to the device to change the MAC address, should that persist across reset, or should reset revert the device to the original MAC address as specified by the user on the command line or whatever ? At the moment you have the former behaviour (and end up storing the MAC address in two places as a result -- it would be neater to either keep it in only one place, or else have emc->regs[] be the current programmed MAC address and emc->conf.macaddr the value to reset to).

I'm not sure we're consistent between device models about that, eg the e1000 seems to reset to the initial MAC addr, but the imx_fec keeps the guest-set value over resets. Jason, is there a recommended "right way" to handle guest-programmable MAC addresses over device reset ?

thanks
-- PMM

#2 lore raw
Quoted 41 lines · Peter Maydell
Quoted 12 lines · Patrick Venture

The MAC address set from Qemu wasn't being saved into the register space.

hw/net/npcm7xx_emc.c +2 -0
@@ -112,6 +115,18 @@ static void emc_reset(NPCM7xxEMCState *emc)
 
     emc->tx_active = false;
     emc->rx_active = false;
+
+    /* Set the MAC address in the register space. */

hw/net/npcm7xx_emc.c static void emc_reset(NPCM7xxEMCState *emc) · line 129

Quoted 10 lines
hw/net/npcm7xx_emc.c +10 -0
+    uint32_t value = (emc->conf.macaddr.a[0] << 24) |
+        (emc->conf.macaddr.a[1] << 16) |
+        (emc->conf.macaddr.a[2] << 8) |
+        emc->conf.macaddr.a[3];
+    npcm7xx_emc_write(emc, REG_CAMM_BASE * sizeof(uint32_t), value,
+                      sizeof(uint32_t));
+
+    value = (emc->conf.macaddr.a[4] << 24) | (emc->conf.macaddr.a[5] << 16);
+    npcm7xx_emc_write(emc, REG_CAML_BASE * sizeof(uint32_t), value,
+                      sizeof(uint32_t));

If I understand correctly, the issue here is that emc->regs[REG_CAMM_BASE] and emc->regs[REG_CAML_BASE] aren't being reset correctly. If so, I think the better approach is to simply reset them here, without going through the register-write function, the same way we already do for the handful of other registers which have non-zero reset values. That's the way other devices seem to do it.

A question to which I don't know the answer: if the guest writes to the device to change the MAC address, should that persist across reset, or should reset revert the device to the original MAC address as specified by the user on the command line or whatever ? At the moment you have the former behaviour (and end up storing the MAC address in two places as a result -- it would be neater to either keep it in only one place, or else have emc->regs[] be the current programmed MAC address and emc->conf.macaddr the value to reset to).

Quoted 5 lines

I'm not sure we're consistent between device models about that, eg the e1000 seems to reset to the initial MAC addr, but the imx_fec keeps the guest-set value over resets. Jason, is there a recommended "right way" to handle guest-programmable MAC addresses over device reset ?

I think it depends on the NIC.

E1000 has a EEPROM interface for providing the MAC address for the ethernet controller before it can be accessed by the driver during reset. For modern Intel NICs like E810, it has similar semantics but using NVM instead of EEPROM. So the current e1000 behaviour seems to be correct (treat the initiali MAC as the one stored in the EEPROM).

I guess most NIC should behave the same as having a persistent storage for MAC for the controller during reset, but I'm not sure this is the case for imx_fec.

(Anyhow, if we change the behaviour we need keep migration compatibility)

Thanks

Quoted 4 lines

thanks
-- PMM

#3 lore raw
Quoted 61 lines · Jason Wang
Quoted 2 lines · Peter Maydell

On Thu, 22 Sept 2022 at 00:47, Patrick Venture <[email protected]>

Quoted 2 lines · wrote:
Quoted 2 lines

The MAC address set from Qemu wasn't being saved into the register

space.

Quoted 6 lines
Quoted 6 lines
(quoted diff) +0 -0
@@ -112,6 +115,18 @@ static void emc_reset(NPCM7xxEMCState *emc)
 
Quoted 12 lines

hw/net/npcm7xx_emc.c static void emc_reset(NPCM7xxEMCState *emc)

Quoted 12 lines
hw/net/npcm7xx_emc.c +10 -0
     emc->tx_active = false;
     emc->rx_active = false;
+
+    /* Set the MAC address in the register space. */
+    uint32_t value = (emc->conf.macaddr.a[0] << 24) |
+        (emc->conf.macaddr.a[1] << 16) |
+        (emc->conf.macaddr.a[2] << 8) |
+        emc->conf.macaddr.a[3];
+    npcm7xx_emc_write(emc, REG_CAMM_BASE * sizeof(uint32_t), value,
+                      sizeof(uint32_t));
+
+    value = (emc->conf.macaddr.a[4] << 24) | (emc->conf.macaddr.a[5]

<< 16);

Quoted 4 lines

hw/net/npcm7xx_emc.c static void emc_reset(NPCM7xxEMCState *emc) · line 129

Quoted 2 lines
hw/net/npcm7xx_emc.c +2 -0
+    npcm7xx_emc_write(emc, REG_CAML_BASE * sizeof(uint32_t), value,
+                      sizeof(uint32_t));

If I understand correctly, the issue here is that

emc->regs[REG_CAMM_BASE]

Quoted 15 lines

and emc->regs[REG_CAML_BASE] aren't being reset correctly. If so, I think the better approach is to simply reset them here, without going through the register-write function, the same way we already do for the handful of other registers which have non-zero reset values. That's the way other devices seem to do it.

A question to which I don't know the answer: if the guest writes to the device to change the MAC address, should that persist across reset, or should reset revert the device to the original MAC address as specified by the user on the command line or whatever ? At the moment you have the former behaviour (and end up storing the MAC address in two places as a result -- it would be neater to either keep it in only one place, or else have emc->regs[] be the current programmed MAC address and emc->conf.macaddr the value to reset to).

Quoted 5 lines

I'm not sure we're consistent between device models about that, eg the e1000 seems to reset to the initial MAC addr, but the imx_fec keeps the guest-set value over resets. Jason, is there a recommended "right way" to handle guest-programmable MAC addresses over device reset ?

I think it depends on the NIC.

E1000 has a EEPROM interface for providing the MAC address for the ethernet controller before it can be accessed by the driver during reset. For modern Intel NICs like E810, it has similar semantics but using NVM instead of EEPROM. So the current e1000 behaviour seems to be correct (treat the initiali MAC as the one stored in the EEPROM).

Quoted 4 lines

I guess most NIC should behave the same as having a persistent storage for MAC for the controller during reset, but I'm not sure this is the case for imx_fec.

So the first time the NIC is realized, it should take the value from the command line. Then later if the guest OS updates it, it should always on reset use that provided value?

Quoted 11 lines

(Anyhow, if we change the behaviour we need keep migration compatibility)

Thanks

Quoted 4 lines

thanks
-- PMM

#4 lore raw
Quoted 26 lines · Patrick Venture
Quoted 25 lines · Jason Wang
Quoted 9 lines · Peter Maydell

A question to which I don't know the answer: if the guest writes to the device to change the MAC address, should that persist across reset, or should reset revert the device to the original MAC address as specified by the user on the command line or whatever ? At the moment you have the former behaviour (and end up storing the MAC address in two places as a result -- it would be neater to either keep it in only one place, or else have emc->regs[] be the current programmed MAC address and emc->conf.macaddr the value to reset to).

Quoted 5 lines

I'm not sure we're consistent between device models about that, eg the e1000 seems to reset to the initial MAC addr, but the imx_fec keeps the guest-set value over resets. Jason, is there a recommended "right way" to handle guest-programmable MAC addresses over device reset ?

I think it depends on the NIC.

E1000 has a EEPROM interface for providing the MAC address for the ethernet controller before it can be accessed by the driver during reset. For modern Intel NICs like E810, it has similar semantics but using NVM instead of EEPROM. So the current e1000 behaviour seems to be correct (treat the initiali MAC as the one stored in the EEPROM).

Quoted 7 lines
Quoted 3 lines

I guess most NIC should behave the same as having a persistent storage for MAC for the controller during reset, but I'm not sure this is the case for imx_fec.

So the first time the NIC is realized, it should take the value from the command line. Then later if the guest OS updates it, it should always on reset use that provided value?

I think what Jason is suggesting is that that should depend on what the real hardware does. On a physical board, if the guest sets the MAC address, and then you power-cycle the hardware, does the MAC that it set still persist after powercycle ? Does the guest writing to these MAC registers correspond to writing to an EEPROM ?

thanks
-- PMM

#5 lore raw
Quoted 36 lines · Peter Maydell
Quoted 3 lines · Patrick Venture
Quoted 2 lines · Jason Wang

On Thu, Sep 22, 2022 at 8:35 PM Peter Maydell <[email protected]>

Quoted 23 lines · wrote:
Quoted 23 lines
Quoted 9 lines

A question to which I don't know the answer: if the guest writes to the device to change the MAC address, should that persist across reset, or should reset revert the device to the original MAC address as specified by the user on the command line or whatever ? At the moment you have the former behaviour (and end up storing the MAC address in two places as a result -- it would be neater to either keep it in only one place, or else have emc->regs[] be the current programmed MAC address and emc->conf.macaddr the value to reset to).

Quoted 5 lines

I'm not sure we're consistent between device models about that, eg the e1000 seems to reset to the initial MAC addr, but the imx_fec keeps the guest-set value over resets. Jason, is there a recommended "right way" to handle guest-programmable MAC addresses over device reset ?

I think it depends on the NIC.

E1000 has a EEPROM interface for providing the MAC address for the ethernet controller before it can be accessed by the driver during reset. For modern Intel NICs like E810, it has similar semantics but using NVM instead of EEPROM. So the current e1000 behaviour seems to be correct (treat the initiali MAC as the one stored in the EEPROM).

Quoted 7 lines
Quoted 3 lines

I guess most NIC should behave the same as having a persistent storage for MAC for the controller during reset, but I'm not sure this is the case for imx_fec.

So the first time the NIC is realized, it should take the value from the command line. Then later if the guest OS updates it, it should always on reset use that provided value?

Quoted 6 lines

I think what Jason is suggesting is that that should depend on what the real hardware does. On a physical board, if the guest sets the MAC address, and then you power-cycle the hardware, does the MAC that it set still persist after powercycle ? Does the guest writing to these MAC registers correspond to writing to an EEPROM ?

Thanks, Peter - we've reached out to the vendor off-list to seek out some details, I'll update this with a v2 when I get an answer.

Quoted 4 lines

thanks
-- PMM

#6 lore raw
Quoted 49 lines · Patrick Venture
Quoted 37 lines · Peter Maydell
Quoted 3 lines · Patrick Venture
Quoted 2 lines · Jason Wang

On Thu, Sep 22, 2022 at 8:35 PM Peter Maydell <

Quoted 23 lines · [email protected]> wrote:
Quoted 23 lines
Quoted 9 lines

A question to which I don't know the answer: if the guest writes to the device to change the MAC address, should that persist across reset, or should reset revert the device to the original MAC address as specified by the user on the command line or whatever ? At the moment you have the former behaviour (and end up storing the MAC address in two places as a result -- it would be neater to either keep it in only one place, or else have emc->regs[] be the current programmed MAC address and emc->conf.macaddr the value to reset to).

Quoted 5 lines

I'm not sure we're consistent between device models about that, eg the e1000 seems to reset to the initial MAC addr, but the imx_fec keeps the guest-set value over resets. Jason, is there a recommended "right way" to handle guest-programmable MAC addresses over device reset ?

I think it depends on the NIC.

E1000 has a EEPROM interface for providing the MAC address for the ethernet controller before it can be accessed by the driver during reset. For modern Intel NICs like E810, it has similar semantics but using NVM instead of EEPROM. So the current e1000 behaviour seems to be correct (treat the initiali MAC as the one stored in the EEPROM).

Quoted 7 lines
Quoted 3 lines

I guess most NIC should behave the same as having a persistent storage for MAC for the controller during reset, but I'm not sure this is the case for imx_fec.

So the first time the NIC is realized, it should take the value from the command line. Then later if the guest OS updates it, it should always on reset use that provided value?

Quoted 6 lines

I think what Jason is suggesting is that that should depend on what the real hardware does. On a physical board, if the guest sets the MAC address, and then you power-cycle the hardware, does the MAC that it set still persist after powercycle ? Does the guest writing to these MAC registers correspond to writing to an EEPROM ?

Quoted 3 lines

Thanks, Peter - we've reached out to the vendor off-list to seek out some details, I'll update this with a v2 when I get an answer.

"No, The EMC driver reset the MAC address registers during boot cycle/reset."

So in that case, we should disregard the value the user sets in reset and use the value provided through Qemu. Or, should we just not allow Qemu to set the MAC address at all?

Quoted 6 lines
Quoted 5 lines

thanks
-- PMM

#7 lore raw
Quoted 11 lines · Patrick Venture
Quoted 8 lines · Patrick Venture
Quoted 5 lines

I think what Jason is suggesting is that that should depend on what the real hardware does. On a physical board, if the guest sets the MAC address, and then you power-cycle the hardware, does the MAC that it set still persist after powercycle ? Does the guest writing to these MAC registers correspond to writing to an EEPROM ?

Thanks, Peter - we've reached out to the vendor off-list to seek out some details, I'll update this with a v2 when I get an answer.

"No, The EMC driver reset the MAC address registers during boot cycle/reset."

OK, I guess that's clear enough. In a real full-software-stack setup is the MAC address setup usually done by a BIOS/firmware, or is it done by Linux ?

Quoted 3 lines

So in that case, we should disregard the value the user sets in reset and use the value provided through Qemu. Or, should we just not allow Qemu to set the MAC address at all?

I think that the behaviour for QEMU's model should be that on reset we should reset the MAC address registers to the user-provided value. That is, if the guest writes to the MAC address registers then that does have an effect, but only until the next reset.

That gives you reasonably plausible behaviour for both:
(1) you're emulating some guest that always sets up its own MAC address when it starts up (eg if it's done by some BIOS-level code that you're running in the guest)
(2) you're booting a guest kernel directly that expects that the firmware/BIOS/whatever has already set up a MAC address -- then the MAC address provided by QEMU/the user fills that role

More concretely:
* on reset, set the emc->regs[] fields from emc->conf.macaddr
* when using the MAC address, always use emc->regs[], never emc->conf.macaddr
* to handle the guest writes to the MAC registers, set emc->regs[], but not emc->conf.macaddr

Assuming that doesn't break your existing booting workloads, of course :-)

thanks
-- PMM

#8 lore raw
Quoted 11 lines · Peter Maydell
Quoted 1 line · Patrick Venture

On Mon, Sep 26, 2022 at 8:45 AM Patrick Venture <[email protected]>

Quoted 8 lines · wrote:
Quoted 8 lines
Quoted 5 lines

I think what Jason is suggesting is that that should depend on what the real hardware does. On a physical board, if the guest sets the MAC address, and then you power-cycle the hardware, does the MAC that it set still persist after powercycle ? Does the guest writing to these MAC registers correspond to writing to an EEPROM ?

Thanks, Peter - we've reached out to the vendor off-list to seek out

Quoted 9 lines

some details, I'll update this with a v2 when I get an answer.

Quoted 1 line

"No, The EMC driver reset the MAC address registers during boot

cycle/reset."

OK, I guess that's clear enough. In a real full-software-stack setup is the MAC address setup usually done by a BIOS/firmware, or is it done by Linux ?

The MAC address is usually set up in u-boot or by linux. The openbmc stack currently doesn't actually read what the device thinks its mac address is, but it could :) It just sets it blindly. We found this register issue because we were debugging why the MAC address we were setting in the qemu command line had no effect. Turns out, per this patch, we weren't using that value anywhere. But even with that fixed, the firmware didn't care. :)

Quoted 20 lines
Quoted 4 lines

So in that case, we should disregard the value the user sets in reset and use the value provided through Qemu. Or, should we just not allow Qemu to set the MAC address at all?

I think that the behaviour for QEMU's model should be that on reset we should reset the MAC address registers to the user-provided value. That is, if the guest writes to the MAC address registers then that does have an effect, but only until the next reset.

That gives you reasonably plausible behaviour for both:
(1) you're emulating some guest that always sets up its own MAC address when it starts up (eg if it's done by some BIOS-level code that you're running in the guest)
(2) you're booting a guest kernel directly that expects that the firmware/BIOS/whatever has already set up a MAC address -- then the MAC address provided by QEMU/the user fills that role

Quoted 8 lines

More concretely:
* on reset, set the emc->regs[] fields from emc->conf.macaddr
* when using the MAC address, always use emc->regs[], never emc->conf.macaddr
* to handle the guest writes to the MAC registers, set emc->regs[], but not emc->conf.macaddr

Thanks! That actually made it much easier to parse what you wanted to

match the behavior :)

Quoted 6 lines

Assuming that doesn't break your existing booting workloads, of course :-)

thanks
-- PMM