Skip to contentSkip to conversation

Revert "e1000/rtl8139: update HMP NIC when every bit is written"

PATCH In review Michael S. Tsirkin 3 participants 16 messages Nov 18, 2013 – Nov 21, 2013

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Revert "e1000/rtl8139: update HMP NIC when every bit is written"

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
Signature · 1 line
MST
Reviewed-by on v1 1/1
#1 lore raw
Quoted 9 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>

-vlad

Quoted 35 lines
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#2 lore raw
In reply to Michael S. Tsirkin
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#3 lore raw
Quoted 11 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

The patch that was applied was controversial and more then 1 person expressed reservations.

-vlad

Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#4 lore raw
Quoted 10 lines · Vlad Yasevich
Quoted 9 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

Quoted 8 lines
Quoted 2 lines

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

But AIUI there's also a valid bit in that high order byte on e1000, so reverting cd5be582 means we stuff a new mac into qemu less often, but it's still only accurate some of the time.

Quoted 3 lines

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

So how is RTL less broken without cd5be582? AIUI the valid bit is off in a separate register on RTL, so we have no guarantee about order of updating the mac. Without cd5be582 the info in the monitor may be permanently broken if the guest uses a write order other than what we assume.

Quoted 2 lines

The patch that was applied was controversial and more then 1 person expressed reservations.

Understood, but it also raised and addressed a shortcoming in the previous patches. If cd5be582 was controversial because the monitor was getting updated with incorrect mac addresses then this simple revert doesn't solve that problem. Thanks,

Alex

Quoted 50 lines
Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#5 lore raw
Quoted 17 lines · Alex Williamson
Quoted 16 lines · Vlad Yasevich
Quoted 11 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

Quoted 6 lines
Quoted 2 lines

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

But AIUI there's also a valid bit in that high order byte on e1000, so reverting cd5be582 means we stuff a new mac into qemu less often, but it's still only accurate some of the time.

Yes, there is a slight issue with validity of mac at the time of processing packets. I have an outstanding question on the Intel list about this behavior with real HW. But, with e1000, the validity bit provides a much higher guarantee that a guest that will be setting the mac address will write the high register second to guarantee that when the valid bit is written, the mac is fully valid. As a result we don't really need the e1000 part of the cd5be5829.

Quoted 11 lines
Quoted 4 lines

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

So how is RTL less broken without cd5be582? AIUI the valid bit is off in a separate register on RTL, so we have no guarantee about order of updating the mac. Without cd5be582 the info in the monitor may be permanently broken if the guest uses a write order other than what we assume.

This one is actually not as bad either. RTL spec requires that receive register writes happen as 32 bit word writes. This is what linux and bsd drivers do, so from driver perspective, the issue is the same. What our emulation layer does is turn these 32 bit writes into 4 8-bit writes. This is likely due to some very broken and very old drivers, but I am not sure.

So, the information in the monitor will be broken if the guest
does: write_hi(); write_lo(); A part of me would really like to see a guest that does this :)

The current code isn't perfect either. It still has a potential to show the wrong mac address in the monitor. I doesn't make a lot of sense to me to replace one sub-optimal solution with another sub-optimal solution, especially since no-one complained.

-vlad

Quoted 63 lines
Quoted 2 lines

The patch that was applied was controversial and more then 1 person expressed reservations.

Understood, but it also raised and addressed a shortcoming in the previous patches. If cd5be582 was controversial because the monitor was getting updated with incorrect mac addresses then this simple revert doesn't solve that problem. Thanks,

Alex

Quoted 50 lines
Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#6 lore raw
Quoted 25 lines · Vlad Yasevich
Quoted 17 lines · Alex Williamson
Quoted 16 lines · Vlad Yasevich
Quoted 11 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

Quoted 6 lines
Quoted 2 lines

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

But AIUI there's also a valid bit in that high order byte on e1000, so reverting cd5be582 means we stuff a new mac into qemu less often, but it's still only accurate some of the time.

Quoted 8 lines

Yes, there is a slight issue with validity of mac at the time of processing packets. I have an outstanding question on the Intel list about this behavior with real HW. But, with e1000, the validity bit provides a much higher guarantee that a guest that will be setting the mac address will write the high register second to guarantee that when the valid bit is written, the mac is fully valid. As a result we don't really need the e1000 part of the cd5be5829.

But doesn't that valid bit mean that a mac update will start and end with a write to the high order register? So we're assuming:

a) write RA + 1 (invalidate)
b) write RA (write low)
c) write RA + 1 (write high + valid)

Without cd5be5829 the only change is that we don't store a new mac into the monitor at b). The mac stored in the monitor is still wrong from a) until c). So it's ever so slightly less broken without cd5be5829.

Quoted 19 lines
Quoted 11 lines
Quoted 4 lines

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

So how is RTL less broken without cd5be582? AIUI the valid bit is off in a separate register on RTL, so we have no guarantee about order of updating the mac. Without cd5be582 the info in the monitor may be permanently broken if the guest uses a write order other than what we assume.

This one is actually not as bad either. RTL spec requires that receive register writes happen as 32 bit word writes. This is what linux and bsd drivers do, so from driver perspective, the issue is the same. What our emulation layer does is turn these 32 bit writes into 4 8-bit writes. This is likely due to some very broken and very old drivers, but I am not sure.

Quoted 3 lines

So, the information in the monitor will be broken if the guest
does: write_hi(); write_lo(); A part of me would really like to see a guest that does this :)

So the argument for reverting here is that it seems unlikely that the dwords get written in the hi->lo order and we'd rather the monitor get stuck with the wrong mac forever than it show the wrong mac address for a tiny window for everybody else? I think you say something about sub-optimal here...

Quoted 5 lines

The current code isn't perfect either. It still has a potential to show the wrong mac address in the monitor. I doesn't make a lot of sense to me to replace one sub-optimal solution with another sub-optimal solution, especially since no-one complained.

Exactly, the code isn't perfect either way and this revert is just replacing one sub-optimal solution for another. So why do it? Thanks,

Alex

Quoted 64 lines
Quoted 63 lines
Quoted 2 lines

The patch that was applied was controversial and more then 1 person expressed reservations.

Understood, but it also raised and addressed a shortcoming in the previous patches. If cd5be582 was controversial because the monitor was getting updated with incorrect mac addresses then this simple revert doesn't solve that problem. Thanks,

Alex

Quoted 50 lines
Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#7 lore raw
Quoted 38 lines · Alex Williamson
Quoted 25 lines · Vlad Yasevich
Quoted 17 lines · Alex Williamson
Quoted 16 lines · Vlad Yasevich
Quoted 11 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

Quoted 6 lines
Quoted 2 lines

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

But AIUI there's also a valid bit in that high order byte on e1000, so reverting cd5be582 means we stuff a new mac into qemu less often, but it's still only accurate some of the time.

Quoted 8 lines

Yes, there is a slight issue with validity of mac at the time of processing packets. I have an outstanding question on the Intel list about this behavior with real HW. But, with e1000, the validity bit provides a much higher guarantee that a guest that will be setting the mac address will write the high register second to guarantee that when the valid bit is written, the mac is fully valid. As a result we don't really need the e1000 part of the cd5be5829.

But doesn't that valid bit mean that a mac update will start and end with a write to the high order register? So we're assuming:

Quoted 7 lines

a) write RA + 1 (invalidate)
b) write RA (write low)
c) write RA + 1 (write high + valid)

Without cd5be5829 the only change is that we don't store a new mac into the monitor at b). The mac stored in the monitor is still wrong from a) until c). So it's ever so slightly less broken without cd5be5829.

Yes.

Quoted 23 lines
Quoted 19 lines
Quoted 11 lines
Quoted 4 lines

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

So how is RTL less broken without cd5be582? AIUI the valid bit is off in a separate register on RTL, so we have no guarantee about order of updating the mac. Without cd5be582 the info in the monitor may be permanently broken if the guest uses a write order other than what we assume.

This one is actually not as bad either. RTL spec requires that receive register writes happen as 32 bit word writes. This is what linux and bsd drivers do, so from driver perspective, the issue is the same. What our emulation layer does is turn these 32 bit writes into 4 8-bit writes. This is likely due to some very broken and very old drivers, but I am not sure.

Quoted 3 lines

So, the information in the monitor will be broken if the guest
does: write_hi(); write_lo(); A part of me would really like to see a guest that does this :)

Quoted 5 lines

So the argument for reverting here is that it seems unlikely that the dwords get written in the hi->lo order and we'd rather the monitor get stuck with the wrong mac forever than it show the wrong mac address for a tiny window for everybody else? I think you say something about sub-optimal here...

The argument is that I don't think it makes sense to maintain two differently broken implementations in 1.6 and 1.7.

It's better to have them broken in the same way and then maybe have a non broken one in 1.8.
Or just leave well alone - no one complained so far, this change is of the "I read the spec and it seems to say we should do X" variety.

Quoted 9 lines
Quoted 6 lines

The current code isn't perfect either. It still has a potential to show the wrong mac address in the monitor. I doesn't make a lot of sense to me to replace one sub-optimal solution with another sub-optimal solution, especially since no-one complained.

Exactly, the code isn't perfect either way and this revert is just replacing one sub-optimal solution for another. So why do it?

Because we had the old buggy behaviour for years and no one complained. Maybe the new buggy behaviour will also be fine. But I'm not inclined to take the chance.

Quoted 70 lines

Thanks,

Alex

Quoted 64 lines
Quoted 63 lines
Quoted 2 lines

The patch that was applied was controversial and more then 1 person expressed reservations.

Understood, but it also raised and addressed a shortcoming in the previous patches. If cd5be582 was controversial because the monitor was getting updated with incorrect mac addresses then this simple revert doesn't solve that problem. Thanks,

Alex

Quoted 50 lines
Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#8 lore raw
Quoted 72 lines · Michael S. Tsirkin
Quoted 38 lines · Alex Williamson
Quoted 25 lines · Vlad Yasevich
Quoted 17 lines · Alex Williamson
Quoted 16 lines · Vlad Yasevich
Quoted 11 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

Quoted 6 lines
Quoted 2 lines

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

But AIUI there's also a valid bit in that high order byte on e1000, so reverting cd5be582 means we stuff a new mac into qemu less often, but it's still only accurate some of the time.

Quoted 8 lines

Yes, there is a slight issue with validity of mac at the time of processing packets. I have an outstanding question on the Intel list about this behavior with real HW. But, with e1000, the validity bit provides a much higher guarantee that a guest that will be setting the mac address will write the high register second to guarantee that when the valid bit is written, the mac is fully valid. As a result we don't really need the e1000 part of the cd5be5829.

But doesn't that valid bit mean that a mac update will start and end with a write to the high order register? So we're assuming:

Quoted 7 lines

a) write RA + 1 (invalidate)
b) write RA (write low)
c) write RA + 1 (write high + valid)

Without cd5be5829 the only change is that we don't store a new mac into the monitor at b). The mac stored in the monitor is still wrong from a) until c). So it's ever so slightly less broken without cd5be5829.

Yes.

Quoted 23 lines
Quoted 22 lines
Quoted 11 lines
Quoted 4 lines

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

So how is RTL less broken without cd5be582? AIUI the valid bit is off in a separate register on RTL, so we have no guarantee about order of updating the mac. Without cd5be582 the info in the monitor may be permanently broken if the guest uses a write order other than what we assume.

This one is actually not as bad either. RTL spec requires that receive register writes happen as 32 bit word writes. This is what linux and bsd drivers do, so from driver perspective, the issue is the same. What our emulation layer does is turn these 32 bit writes into 4 8-bit writes. This is likely due to some very broken and very old drivers, but I am not sure.

So, the information in the monitor will be broken if the guest
does: write_hi(); write_lo(); A part of me would really like to see a guest that does this :)

Quoted 8 lines
Quoted 5 lines

So the argument for reverting here is that it seems unlikely that the dwords get written in the hi->lo order and we'd rather the monitor get stuck with the wrong mac forever than it show the wrong mac address for a tiny window for everybody else? I think you say something about sub-optimal here...

The argument is that I don't think it makes sense to maintain two differently broken implementations in 1.6 and 1.7.

Maintain 1.6 support after 1.7 is released... is this some new policy?

Quoted 2 lines

It's better to have them broken in the same way and then maybe have a non broken one in 1.8.

So 1.6 is broken because rtl/e1000 never informs the monitor about mac updates. QEMU 1.7 adds the following (maybe others, these are just the ones I know of):

655d3b63 - Update rtl/e1000 on reset
7c36507c - Update e1000 on high dword mac write
23c37c37 - Update rtl on high byte mac write
cd5be582 - Update rtl/e1000 on any mac write

And you're suggesting that reverting just that last commit makes a significant enough difference to make it worthwhile to do so? I'm going to go back to Vlad's replacing sub-optimal with sub-optimal comment.

Quoted 3 lines

Or just leave well alone - no one complained so far, this change is of the "I read the spec and it seems to say we should do X" variety.

I think it's great that Vlad did the research and we now know how to fix these correctly. I suggest we leave an incremental improvement alone in 1.7 and think about implementing it correctly in 1.8. Reading the spec and figuring out how it's supposed to work seems like a better idea than guessing which data element gets written last.

Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

The current code isn't perfect either. It still has a potential to show the wrong mac address in the monitor. I doesn't make a lot of sense to me to replace one sub-optimal solution with another sub-optimal solution, especially since no-one complained.

Quoted 6 lines
Quoted 2 lines

Exactly, the code isn't perfect either way and this revert is just replacing one sub-optimal solution for another. So why do it?

Because we had the old buggy behaviour for years and no one complained. Maybe the new buggy behaviour will also be fine. But I'm not inclined to take the chance.

Except somebody did complain and filed a bug that Amos was trying to fix. This one revert does not bring us back to 1.6 behavior and does not make a significant improvement in the current behavior, so I just don't see the point. Thanks,

Alex

Quoted 66 lines
Quoted 66 lines
Quoted 64 lines
Quoted 63 lines
Quoted 2 lines

The patch that was applied was controversial and more then 1 person expressed reservations.

Understood, but it also raised and addressed a shortcoming in the previous patches. If cd5be582 was controversial because the monitor was getting updated with incorrect mac addresses then this simple revert doesn't solve that problem. Thanks,

Alex

Quoted 50 lines
Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#9 lore raw
Quoted 79 lines · Alex Williamson
Quoted 78 lines · Michael S. Tsirkin
Quoted 38 lines · Alex Williamson
Quoted 25 lines · Vlad Yasevich
Quoted 17 lines · Alex Williamson
Quoted 16 lines · Vlad Yasevich
Quoted 11 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

Quoted 6 lines
Quoted 2 lines

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

But AIUI there's also a valid bit in that high order byte on e1000, so reverting cd5be582 means we stuff a new mac into qemu less often, but it's still only accurate some of the time.

Quoted 8 lines

Yes, there is a slight issue with validity of mac at the time of processing packets. I have an outstanding question on the Intel list about this behavior with real HW. But, with e1000, the validity bit provides a much higher guarantee that a guest that will be setting the mac address will write the high register second to guarantee that when the valid bit is written, the mac is fully valid. As a result we don't really need the e1000 part of the cd5be5829.

But doesn't that valid bit mean that a mac update will start and end with a write to the high order register? So we're assuming:

Quoted 7 lines

a) write RA + 1 (invalidate)
b) write RA (write low)
c) write RA + 1 (write high + valid)

Without cd5be5829 the only change is that we don't store a new mac into the monitor at b). The mac stored in the monitor is still wrong from a) until c). So it's ever so slightly less broken without cd5be5829.

Yes.

Quoted 28 lines
Quoted 19 lines
Quoted 11 lines
Quoted 4 lines

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

So how is RTL less broken without cd5be582? AIUI the valid bit is off in a separate register on RTL, so we have no guarantee about order of updating the mac. Without cd5be582 the info in the monitor may be permanently broken if the guest uses a write order other than what we assume.

This one is actually not as bad either. RTL spec requires that receive register writes happen as 32 bit word writes. This is what linux and bsd drivers do, so from driver perspective, the issue is the same. What our emulation layer does is turn these 32 bit writes into 4 8-bit writes. This is likely due to some very broken and very old drivers, but I am not sure.

Quoted 3 lines

So, the information in the monitor will be broken if the guest
does: write_hi(); write_lo(); A part of me would really like to see a guest that does this :)

So the argument for reverting here is that it seems unlikely that the dwords get written in the hi->lo order and we'd rather the monitor get stuck with the wrong mac forever than it show the wrong mac address for a tiny window for everybody else? I think you say something about sub-optimal here...

Quoted 4 lines
Quoted 2 lines

The argument is that I don't think it makes sense to maintain two differently broken implementations in 1.6 and 1.7.

Maintain 1.6 support after 1.7 is released... is this some new policy?

I happen to care about maintaining it downstream too.

Quoted 7 lines
Quoted 2 lines

It's better to have them broken in the same way and then maybe have a non broken one in 1.8.

So 1.6 is broken because rtl/e1000 never informs the monitor about mac updates. QEMU 1.7 adds the following (maybe others, these are just the ones I know of):

Quoted 7 lines

655d3b63 - Update rtl/e1000 on reset
7c36507c - Update e1000 on high dword mac write
23c37c37 - Update rtl on high byte mac write
cd5be582 - Update rtl/e1000 on any mac write

And you're suggesting that reverting just that last commit makes a significant enough difference to make it worthwhile to do so?

Yes. I think the rest of the commits are not controversial. They are not bugfixes - they are required for the mac change event feature.
Without them libvirt can not use the mac change event.

Quoted 10 lines

I'm going
to go back to Vlad's replacing sub-optimal with sub-optimal comment.

Quoted 3 lines

Or just leave well alone - no one complained so far, this change is of the "I read the spec and it seems to say we should do X" variety.

I think it's great that Vlad did the research and we now know how to fix these correctly. I suggest we leave an incremental improvement alone in 1.7 and think about implementing it correctly in 1.8.

Why is this an improvement?

Quoted 3 lines

Reading the spec
and figuring out how it's supposed to work seems like a better idea than guessing which data element gets written last.

Maybe it would be if the spec was complete or correct.

Quoted 9 lines
Quoted 9 lines
Quoted 8 lines
Quoted 5 lines

The current code isn't perfect either. It still has a potential to show the wrong mac address in the monitor. I doesn't make a lot of sense to me to replace one sub-optimal solution with another sub-optimal solution, especially since no-one complained.

Exactly, the code isn't perfect either way and this revert is just replacing one sub-optimal solution for another. So why do it?

Quoted 6 lines
Quoted 3 lines

Because we had the old buggy behaviour for years and no one complained. Maybe the new buggy behaviour will also be fine. But I'm not inclined to take the chance.

Except somebody did complain and filed a bug that Amos was trying to fix.

Really? Try to find this bug. It was all based on misreading the spec.

Quoted 1 line

This one revert does not bring us back to 1.6 behavior

I judge that with this revert it's close enough.

Quoted 5 lines

and does
not make a significant improvement in the current behavior, so I just don't see the point. Thanks,

Alex

The point is to avoid introducing random noise in a complex system. Reading guest driver code to figure out what it does is very relevant and making changes like this one based on just reading the spec just does not work.

Quoted 68 lines
Quoted 66 lines
Quoted 66 lines
Quoted 64 lines
Quoted 63 lines
Quoted 2 lines

The patch that was applied was controversial and more then 1 person expressed reservations.

Understood, but it also raised and addressed a shortcoming in the previous patches. If cd5be582 was controversial because the monitor was getting updated with incorrect mac addresses then this simple revert doesn't solve that problem. Thanks,

Alex

Quoted 50 lines
Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#10 lore raw
In reply to Alex Williamson
Quoted 35 lines · Alex Williamson
Quoted 25 lines · Vlad Yasevich
Quoted 17 lines · Alex Williamson
Quoted 16 lines · Vlad Yasevich
Quoted 11 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

Quoted 6 lines
Quoted 2 lines

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

But AIUI there's also a valid bit in that high order byte on e1000, so reverting cd5be582 means we stuff a new mac into qemu less often, but it's still only accurate some of the time.

Quoted 8 lines

Yes, there is a slight issue with validity of mac at the time of processing packets. I have an outstanding question on the Intel list about this behavior with real HW. But, with e1000, the validity bit provides a much higher guarantee that a guest that will be setting the mac address will write the high register second to guarantee that when the valid bit is written, the mac is fully valid. As a result we don't really need the e1000 part of the cd5be5829.

Quoted 7 lines

But doesn't that valid bit mean that a mac update will start and end with a write to the high order register? So we're assuming:

a) write RA + 1 (invalidate)
b) write RA (write low)
c) write RA + 1 (write high + valid)

No. On update, only steps b and c typically happen. Thus my question to the on the intel list.

Quoted 3 lines

Without cd5be5829 the only change is that we don't store a new mac into the monitor at b). The mac stored in the monitor is still wrong from a) until c). So it's ever so slightly less broken without cd5be5829.

Since there is really no a), the mac in the monitor is only different after step b). since it's is incomplete and we expect step c), there is really no point in updating it.

Quoted 20 lines
Quoted 20 lines
Quoted 6 lines
Quoted 5 lines

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

Quoted 6 lines

So how is RTL less broken without cd5be582? AIUI the valid bit is off in a separate register on RTL, so we have no guarantee about order of updating the mac. Without cd5be582 the info in the monitor may be permanently broken if the guest uses a write order other than what we assume.

This one is actually not as bad either. RTL spec requires that receive register writes happen as 32 bit word writes. This is what linux and bsd drivers do, so from driver perspective, the issue is the same. What our emulation layer does is turn these 32 bit writes into 4 8-bit writes. This is likely due to some very broken and very old drivers, but I am not sure.

Quoted 7 lines
Quoted 3 lines

So, the information in the monitor will be broken if the guest
does: write_hi(); write_lo(); A part of me would really like to see a guest that does this :)

So the argument for reverting here is that it seems unlikely that the dwords get written in the hi->lo order and we'd rather the monitor get stuck with the wrong mac forever

For how many/which guests? I know it's not linux or BSD. I need to boot windows to see what it does, but I think it does the right thing.

Quoted 2 lines

than it show the wrong mac address for
a tiny window for everybody else?

Yes, this would happen for everybody. If you are querying the output, you might see this and it will show up as 2 changes.

We are talking about 2 "tiny" amounts here:
On the one hand, we __might__ have guests that write mac and reverse order thus showing wrong address.
On the other hand we have all the guests who will show the wrong address for a _short_ time.

I have a hard time deciding, but have a slight preference for a small uncertainty (the # of backward writers is very small), rather then a small certainty (everyone will be effected for a small period of time).

Quoted 11 lines

I think you say something about
sub-optimal here...

Quoted 5 lines

The current code isn't perfect either. It still has a potential to show the wrong mac address in the monitor. I doesn't make a lot of sense to me to replace one sub-optimal solution with another sub-optimal solution, especially since no-one complained.

Exactly, the code isn't perfect either way and this revert is just replacing one sub-optimal solution for another. So why do it?

Another part of this, for me, is that a change was made that we had a disagreement on and a different approach was being discussed. The revert is to bring us back to before this change.

-vlad

Quoted 71 lines

Thanks,

Alex

Quoted 64 lines
Quoted 63 lines
Quoted 2 lines

The patch that was applied was controversial and more then 1 person expressed reservations.

Understood, but it also raised and addressed a shortcoming in the previous patches. If cd5be582 was controversial because the monitor was getting updated with incorrect mac addresses then this simple revert doesn't solve that problem. Thanks,

Alex

Quoted 50 lines
Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#11 lore raw
Quoted 39 lines · Vlad Yasevich
Quoted 38 lines · Alex Williamson
Quoted 25 lines · Vlad Yasevich
Quoted 17 lines · Alex Williamson
Quoted 16 lines · Vlad Yasevich
Quoted 11 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

Quoted 6 lines
Quoted 2 lines

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

But AIUI there's also a valid bit in that high order byte on e1000, so reverting cd5be582 means we stuff a new mac into qemu less often, but it's still only accurate some of the time.

Quoted 8 lines

Yes, there is a slight issue with validity of mac at the time of processing packets. I have an outstanding question on the Intel list about this behavior with real HW. But, with e1000, the validity bit provides a much higher guarantee that a guest that will be setting the mac address will write the high register second to guarantee that when the valid bit is written, the mac is fully valid. As a result we don't really need the e1000 part of the cd5be5829.

But doesn't that valid bit mean that a mac update will start and end with a write to the high order register? So we're assuming:

Quoted 7 lines
Quoted 4 lines

a) write RA + 1 (invalidate)
b) write RA (write low)
c) write RA + 1 (write high + valid)

No. On update, only steps b and c typically happen. Thus my question to the on the intel list.

So perhaps the bit is some kind of data latch bit and the mac address fields within those registers are effectively scratch until that bit is written?

Quoted 7 lines
Quoted 3 lines

Without cd5be5829 the only change is that we don't store a new mac into the monitor at b). The mac stored in the monitor is still wrong from a) until c). So it's ever so slightly less broken without cd5be5829.

Since there is really no a), the mac in the monitor is only different after step b). since it's is incomplete and we expect step c), there is really no point in updating it.

Great, so I have no argument against reverting, or just fixing, that chunk of cd5be5829. Let's implement the latch bit too.

Quoted 23 lines
Quoted 23 lines
Quoted 22 lines
Quoted 11 lines
Quoted 4 lines

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

So how is RTL less broken without cd5be582? AIUI the valid bit is off in a separate register on RTL, so we have no guarantee about order of updating the mac. Without cd5be582 the info in the monitor may be permanently broken if the guest uses a write order other than what we assume.

This one is actually not as bad either. RTL spec requires that receive register writes happen as 32 bit word writes. This is what linux and bsd drivers do, so from driver perspective, the issue is the same. What our emulation layer does is turn these 32 bit writes into 4 8-bit writes. This is likely due to some very broken and very old drivers, but I am not sure.

So, the information in the monitor will be broken if the guest
does: write_hi(); write_lo(); A part of me would really like to see a guest that does this :)

Quoted 6 lines
Quoted 3 lines

So the argument for reverting here is that it seems unlikely that the dwords get written in the hi->lo order and we'd rather the monitor get stuck with the wrong mac forever

For how many/which guests? I know it's not linux or BSD. I need to boot windows to see what it does, but I think it does the right thing.

How many guests do you plan to test?

Quoted 12 lines
Quoted 2 lines

than it show the wrong mac address for
a tiny window for everybody else?

Yes, this would happen for everybody. If you are querying the output, you might see this and it will show up as 2 changes.

We are talking about 2 "tiny" amounts here:
On the one hand, we __might__ have guests that write mac and reverse order thus showing wrong address.
On the other hand we have all the guests who will show the wrong address for a _short_ time.

Quoted 3 lines

I have a hard time deciding, but have a slight preference for a small uncertainty (the # of backward writers is very small), rather then a small certainty (everyone will be effected for a small period of time).

Why compromise? Why not implement the register that handles whether the mac is valid on RTL?

Quoted 9 lines
Quoted 9 lines

I think you say something about
sub-optimal here...

Quoted 5 lines

The current code isn't perfect either. It still has a potential to show the wrong mac address in the monitor. I doesn't make a lot of sense to me to replace one sub-optimal solution with another sub-optimal solution, especially since no-one complained.

Quoted 6 lines
Quoted 2 lines

Exactly, the code isn't perfect either way and this revert is just replacing one sub-optimal solution for another. So why do it?

Another part of this, for me, is that a change was made that we had a disagreement on and a different approach was being discussed. The revert is to bring us back to before this change.

I was surprised to see it had been committed too, but that's a different argument for revert than what's being presented here. Thanks,

Alex

Quoted 68 lines
Quoted 67 lines
Quoted 64 lines
Quoted 63 lines
Quoted 2 lines

The patch that was applied was controversial and more then 1 person expressed reservations.

Understood, but it also raised and addressed a shortcoming in the previous patches. If cd5be582 was controversial because the monitor was getting updated with incorrect mac addresses then this simple revert doesn't solve that problem. Thanks,

Alex

Quoted 50 lines
Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#12 lore raw
Quoted 56 lines · Alex Williamson
Quoted 39 lines · Vlad Yasevich
Quoted 38 lines · Alex Williamson
Quoted 25 lines · Vlad Yasevich
Quoted 17 lines · Alex Williamson
Quoted 16 lines · Vlad Yasevich
Quoted 11 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

Quoted 6 lines
Quoted 2 lines

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

But AIUI there's also a valid bit in that high order byte on e1000, so reverting cd5be582 means we stuff a new mac into qemu less often, but it's still only accurate some of the time.

Quoted 8 lines

Yes, there is a slight issue with validity of mac at the time of processing packets. I have an outstanding question on the Intel list about this behavior with real HW. But, with e1000, the validity bit provides a much higher guarantee that a guest that will be setting the mac address will write the high register second to guarantee that when the valid bit is written, the mac is fully valid. As a result we don't really need the e1000 part of the cd5be5829.

But doesn't that valid bit mean that a mac update will start and end with a write to the high order register? So we're assuming:

Quoted 7 lines
Quoted 4 lines

a) write RA + 1 (invalidate)
b) write RA (write low)
c) write RA + 1 (write high + valid)

No. On update, only steps b and c typically happen. Thus my question to the on the intel list.

So perhaps the bit is some kind of data latch bit and the mac address fields within those registers are effectively scratch until that bit is written?

Quoted 4 lines
Quoted 3 lines

Without cd5be5829 the only change is that we don't store a new mac into the monitor at b). The mac stored in the monitor is still wrong from a) until c). So it's ever so slightly less broken without cd5be5829.

Quoted 6 lines
Quoted 3 lines

Since there is really no a), the mac in the monitor is only different after step b). since it's is incomplete and we expect step c), there is really no point in updating it.

Great, so I have no argument against reverting, or just fixing, that chunk of cd5be5829. Let's implement the latch bit too.

Right. So for 1.7 revert is the only realistic thing. For 1.8 we'll revisit.

Quoted 63 lines
Quoted 23 lines
Quoted 23 lines
Quoted 22 lines
Quoted 11 lines
Quoted 4 lines

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

So how is RTL less broken without cd5be582? AIUI the valid bit is off in a separate register on RTL, so we have no guarantee about order of updating the mac. Without cd5be582 the info in the monitor may be permanently broken if the guest uses a write order other than what we assume.

This one is actually not as bad either. RTL spec requires that receive register writes happen as 32 bit word writes. This is what linux and bsd drivers do, so from driver perspective, the issue is the same. What our emulation layer does is turn these 32 bit writes into 4 8-bit writes. This is likely due to some very broken and very old drivers, but I am not sure.

So, the information in the monitor will be broken if the guest
does: write_hi(); write_lo(); A part of me would really like to see a guest that does this :)

Quoted 6 lines
Quoted 3 lines

So the argument for reverting here is that it seems unlikely that the dwords get written in the hi->lo order and we'd rather the monitor get stuck with the wrong mac forever

For how many/which guests? I know it's not linux or BSD. I need to boot windows to see what it does, but I think it does the right thing.

How many guests do you plan to test?

Quoted 12 lines
Quoted 2 lines

than it show the wrong mac address for
a tiny window for everybody else?

Yes, this would happen for everybody. If you are querying the output, you might see this and it will show up as 2 changes.

We are talking about 2 "tiny" amounts here:
On the one hand, we __might__ have guests that write mac and reverse order thus showing wrong address.
On the other hand we have all the guests who will show the wrong address for a _short_ time.

Quoted 3 lines

I have a hard time deciding, but have a slight preference for a small uncertainty (the # of backward writers is very small), rather then a small certainty (everyone will be effected for a small period of time).

Why compromise? Why not implement the register that handles whether the mac is valid on RTL?

Quoted 12 lines
Quoted 11 lines

I think you say something about
sub-optimal here...

Quoted 5 lines

The current code isn't perfect either. It still has a potential to show the wrong mac address in the monitor. I doesn't make a lot of sense to me to replace one sub-optimal solution with another sub-optimal solution, especially since no-one complained.

Exactly, the code isn't perfect either way and this revert is just replacing one sub-optimal solution for another. So why do it?

Quoted 8 lines
Quoted 3 lines

Another part of this, for me, is that a change was made that we had a disagreement on and a different approach was being discussed. The revert is to bring us back to before this change.

I was surprised to see it had been committed too, but that's a different argument for revert than what's being presented here. Thanks,

Alex

Looks like the same argument.
So people that have experience maintaining this device expressed reservations and unsurprisingly digging into it we find issues.

Quoted 70 lines
Quoted 68 lines
Quoted 67 lines
Quoted 64 lines
Quoted 63 lines
Quoted 2 lines

The patch that was applied was controversial and more then 1 person expressed reservations.

Understood, but it also raised and addressed a shortcoming in the previous patches. If cd5be582 was controversial because the monitor was getting updated with incorrect mac addresses then this simple revert doesn't solve that problem. Thanks,

Alex

Quoted 50 lines
Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#13 lore raw
In reply to Alex Williamson
Quoted 86 lines · Alex Williamson
Quoted 39 lines · Vlad Yasevich
Quoted 38 lines · Alex Williamson
Quoted 25 lines · Vlad Yasevich
Quoted 17 lines · Alex Williamson
Quoted 16 lines · Vlad Yasevich
Quoted 11 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

Quoted 6 lines
Quoted 2 lines

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

But AIUI there's also a valid bit in that high order byte on e1000, so reverting cd5be582 means we stuff a new mac into qemu less often, but it's still only accurate some of the time.

Quoted 8 lines

Yes, there is a slight issue with validity of mac at the time of processing packets. I have an outstanding question on the Intel list about this behavior with real HW. But, with e1000, the validity bit provides a much higher guarantee that a guest that will be setting the mac address will write the high register second to guarantee that when the valid bit is written, the mac is fully valid. As a result we don't really need the e1000 part of the cd5be5829.

But doesn't that valid bit mean that a mac update will start and end with a write to the high order register? So we're assuming:

Quoted 7 lines
Quoted 4 lines

a) write RA + 1 (invalidate)
b) write RA (write low)
c) write RA + 1 (write high + valid)

No. On update, only steps b and c typically happen. Thus my question to the on the intel list.

So perhaps the bit is some kind of data latch bit and the mac address fields within those registers are effectively scratch until that bit is written?

Quoted 7 lines
Quoted 3 lines

Without cd5be5829 the only change is that we don't store a new mac into the monitor at b). The mac stored in the monitor is still wrong from a) until c). So it's ever so slightly less broken without cd5be5829.

Since there is really no a), the mac in the monitor is only different after step b). since it's is incomplete and we expect step c), there is really no point in updating it.

Great, so I have no argument against reverting, or just fixing, that chunk of cd5be5829. Let's implement the latch bit too.

Quoted 23 lines
Quoted 23 lines
Quoted 22 lines
Quoted 11 lines
Quoted 4 lines

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

So how is RTL less broken without cd5be582? AIUI the valid bit is off in a separate register on RTL, so we have no guarantee about order of updating the mac. Without cd5be582 the info in the monitor may be permanently broken if the guest uses a write order other than what we assume.

This one is actually not as bad either. RTL spec requires that receive register writes happen as 32 bit word writes. This is what linux and bsd drivers do, so from driver perspective, the issue is the same. What our emulation layer does is turn these 32 bit writes into 4 8-bit writes. This is likely due to some very broken and very old drivers, but I am not sure.

So, the information in the monitor will be broken if the guest
does: write_hi(); write_lo(); A part of me would really like to see a guest that does this :)

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

So the argument for reverting here is that it seems unlikely that the dwords get written in the hi->lo order and we'd rather the monitor get stuck with the wrong mac forever

For how many/which guests? I know it's not linux or BSD. I need to boot windows to see what it does, but I think it does the right thing.

How many guests do you plan to test?

I think the proposal was to see if anyone reports an issue :)

Quoted 13 lines
Quoted 13 lines
Quoted 3 lines

than it show the wrong mac address for
a tiny window for everybody else?

Yes, this would happen for everybody. If you are querying the output, you might see this and it will show up as 2 changes.

We are talking about 2 "tiny" amounts here:
On the one hand, we __might__ have guests that write mac and reverse order thus showing wrong address.
On the other hand we have all the guests who will show the wrong address for a _short_ time.

Quoted 7 lines
Quoted 3 lines

I have a hard time deciding, but have a slight preference for a small uncertainty (the # of backward writers is very small), rather then a small certainty (everyone will be effected for a small period of time).

Why compromise? Why not implement the register that handles whether the mac is valid on RTL?

I am working on this but it's too late for 1.7.

For 1.7, we have a question: impact a very small number (# -> 0) of guests for a long time, or impact a everyone for a very short time.

Original code did the former and fixed the original issue reported. The commit in question that we are talking about reverting did the later simply because we didn't know any better.

Quoted 12 lines
Quoted 12 lines
Quoted 11 lines

I think you say something about
sub-optimal here...

Quoted 5 lines

The current code isn't perfect either. It still has a potential to show the wrong mac address in the monitor. I doesn't make a lot of sense to me to replace one sub-optimal solution with another sub-optimal solution, especially since no-one complained.

Exactly, the code isn't perfect either way and this revert is just replacing one sub-optimal solution for another. So why do it?

Quoted 6 lines
Quoted 3 lines

Another part of this, for me, is that a change was made that we had a disagreement on and a different approach was being discussed. The revert is to bring us back to before this change.

I was surprised to see it had been committed too, but that's a different argument for revert than what's being presented here. Thanks,

It is another argument for the revert non the less.

Thanks
-vlad

Quoted 74 lines

Alex

Quoted 68 lines
Quoted 67 lines
Quoted 64 lines
Quoted 63 lines
Quoted 2 lines

The patch that was applied was controversial and more then 1 person expressed reservations.

Understood, but it also raised and addressed a shortcoming in the previous patches. If cd5be582 was controversial because the monitor was getting updated with incorrect mac addresses then this simple revert doesn't solve that problem. Thanks,

Alex

Quoted 50 lines
Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#14 lore raw
Quoted 119 lines · Vlad Yasevich
Quoted 86 lines · Alex Williamson
Quoted 39 lines · Vlad Yasevich
Quoted 38 lines · Alex Williamson
Quoted 25 lines · Vlad Yasevich
Quoted 17 lines · Alex Williamson
Quoted 16 lines · Vlad Yasevich
Quoted 11 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

Quoted 6 lines
Quoted 2 lines

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

But AIUI there's also a valid bit in that high order byte on e1000, so reverting cd5be582 means we stuff a new mac into qemu less often, but it's still only accurate some of the time.

Quoted 8 lines

Yes, there is a slight issue with validity of mac at the time of processing packets. I have an outstanding question on the Intel list about this behavior with real HW. But, with e1000, the validity bit provides a much higher guarantee that a guest that will be setting the mac address will write the high register second to guarantee that when the valid bit is written, the mac is fully valid. As a result we don't really need the e1000 part of the cd5be5829.

But doesn't that valid bit mean that a mac update will start and end with a write to the high order register? So we're assuming:

Quoted 7 lines
Quoted 4 lines

a) write RA + 1 (invalidate)
b) write RA (write low)
c) write RA + 1 (write high + valid)

No. On update, only steps b and c typically happen. Thus my question to the on the intel list.

So perhaps the bit is some kind of data latch bit and the mac address fields within those registers are effectively scratch until that bit is written?

Quoted 7 lines
Quoted 3 lines

Without cd5be5829 the only change is that we don't store a new mac into the monitor at b). The mac stored in the monitor is still wrong from a) until c). So it's ever so slightly less broken without cd5be5829.

Since there is really no a), the mac in the monitor is only different after step b). since it's is incomplete and we expect step c), there is really no point in updating it.

Great, so I have no argument against reverting, or just fixing, that chunk of cd5be5829. Let's implement the latch bit too.

Quoted 23 lines
Quoted 23 lines
Quoted 22 lines
Quoted 11 lines
Quoted 4 lines

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

So how is RTL less broken without cd5be582? AIUI the valid bit is off in a separate register on RTL, so we have no guarantee about order of updating the mac. Without cd5be582 the info in the monitor may be permanently broken if the guest uses a write order other than what we assume.

This one is actually not as bad either. RTL spec requires that receive register writes happen as 32 bit word writes. This is what linux and bsd drivers do, so from driver perspective, the issue is the same. What our emulation layer does is turn these 32 bit writes into 4 8-bit writes. This is likely due to some very broken and very old drivers, but I am not sure.

So, the information in the monitor will be broken if the guest
does: write_hi(); write_lo(); A part of me would really like to see a guest that does this :)

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

So the argument for reverting here is that it seems unlikely that the dwords get written in the hi->lo order and we'd rather the monitor get stuck with the wrong mac forever

For how many/which guests? I know it's not linux or BSD. I need to boot windows to see what it does, but I think it does the right thing.

How many guests do you plan to test?

I think the proposal was to see if anyone reports an issue :)

Quoted 13 lines
Quoted 13 lines
Quoted 3 lines

than it show the wrong mac address for
a tiny window for everybody else?

Yes, this would happen for everybody. If you are querying the output, you might see this and it will show up as 2 changes.

We are talking about 2 "tiny" amounts here:
On the one hand, we __might__ have guests that write mac and reverse order thus showing wrong address.
On the other hand we have all the guests who will show the wrong address for a _short_ time.

Quoted 7 lines
Quoted 3 lines

I have a hard time deciding, but have a slight preference for a small uncertainty (the # of backward writers is very small), rather then a small certainty (everyone will be effected for a small period of time).

Why compromise? Why not implement the register that handles whether the mac is valid on RTL?

Quoted 8 lines

I am working on this but it's too late for 1.7.

For 1.7, we have a question: impact a very small number (# -> 0) of guests for a long time, or impact a everyone for a very short time.

Original code did the former and fixed the original issue reported. The commit in question that we are talking about reverting did the later simply because we didn't know any better.

This thread is getting too long. What I've learned from it is that there are valid arguments backed by the spec to indicate why the version of e1000 prior to cd5be582 was more correct and why we think it does not update the monitor with incorrect mac information. rtl8139 is less convincing, but all the drivers we know about behave in one way that allows us to make an assumption about write order and avoid spurious, incorrect mac address updates to the monitor. Had that information been provided in the commit log we could have all spent the day doing relevant work. Please send a v2. Thanks,

Alex

Quoted 99 lines
Quoted 12 lines
Quoted 12 lines
Quoted 11 lines

I think you say something about
sub-optimal here...

Quoted 5 lines

The current code isn't perfect either. It still has a potential to show the wrong mac address in the monitor. I doesn't make a lot of sense to me to replace one sub-optimal solution with another sub-optimal solution, especially since no-one complained.

Exactly, the code isn't perfect either way and this revert is just replacing one sub-optimal solution for another. So why do it?

Quoted 6 lines
Quoted 3 lines

Another part of this, for me, is that a change was made that we had a disagreement on and a different approach was being discussed. The revert is to bring us back to before this change.

I was surprised to see it had been committed too, but that's a different argument for revert than what's being presented here. Thanks,

It is another argument for the revert non the less.

Thanks
-vlad

Quoted 74 lines

Alex

Quoted 68 lines
Quoted 67 lines
Quoted 64 lines
Quoted 63 lines
Quoted 2 lines

The patch that was applied was controversial and more then 1 person expressed reservations.

Understood, but it also raised and addressed a shortcoming in the previous patches. If cd5be582 was controversial because the monitor was getting updated with incorrect mac addresses then this simple revert doesn't solve that problem. Thanks,

Alex

Quoted 50 lines
Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;
#15 lore raw
Quoted 139 lines · Alex Williamson
Quoted 119 lines · Vlad Yasevich
Quoted 86 lines · Alex Williamson
Quoted 39 lines · Vlad Yasevich
Quoted 38 lines · Alex Williamson
Quoted 25 lines · Vlad Yasevich
Quoted 17 lines · Alex Williamson
Quoted 16 lines · Vlad Yasevich
Quoted 11 lines · Alex Williamson
Quoted 5 lines · Michael S. Tsirkin

This reverts commit cd5be5829c1ce87aa6b3a7806524fac07ac9a757. Digging into hardware specs shows this does not actually make QEMU behave more like hardware.
Let's stick to the tried heuristic for 1.7 and possibly revisit for 1.8.

If this is broken, then so are these:

23c37c37f0280761072c23bf67d3a4f3c0ff25aa
7c36507c2b8776266f50c5e2739bd18279953b93

These aren't really broken. They just assume that the high order writes will happen after the low order writes.

Quoted 6 lines
Quoted 2 lines

In the case of e1000, this is a little more then an assumption (particularly in the case of nic initilization).

But AIUI there's also a valid bit in that high order byte on e1000, so reverting cd5be582 means we stuff a new mac into qemu less often, but it's still only accurate some of the time.

Quoted 8 lines

Yes, there is a slight issue with validity of mac at the time of processing packets. I have an outstanding question on the Intel list about this behavior with real HW. But, with e1000, the validity bit provides a much higher guarantee that a guest that will be setting the mac address will write the high register second to guarantee that when the valid bit is written, the mac is fully valid. As a result we don't really need the e1000 part of the cd5be5829.

But doesn't that valid bit mean that a mac update will start and end with a write to the high order register? So we're assuming:

Quoted 7 lines
Quoted 4 lines

a) write RA + 1 (invalidate)
b) write RA (write low)
c) write RA + 1 (write high + valid)

No. On update, only steps b and c typically happen. Thus my question to the on the intel list.

So perhaps the bit is some kind of data latch bit and the mac address fields within those registers are effectively scratch until that bit is written?

Quoted 7 lines
Quoted 3 lines

Without cd5be5829 the only change is that we don't store a new mac into the monitor at b). The mac stored in the monitor is still wrong from a) until c). So it's ever so slightly less broken without cd5be5829.

Since there is really no a), the mac in the monitor is only different after step b). since it's is incomplete and we expect step c), there is really no point in updating it.

Great, so I have no argument against reverting, or just fixing, that chunk of cd5be5829. Let's implement the latch bit too.

Quoted 23 lines
Quoted 23 lines
Quoted 22 lines
Quoted 11 lines
Quoted 4 lines

In the case of RTL nic, it is just an assumption, but it hasn't been shown faulty yet. We do plan to address this a bit more thoroughly.

So how is RTL less broken without cd5be582? AIUI the valid bit is off in a separate register on RTL, so we have no guarantee about order of updating the mac. Without cd5be582 the info in the monitor may be permanently broken if the guest uses a write order other than what we assume.

This one is actually not as bad either. RTL spec requires that receive register writes happen as 32 bit word writes. This is what linux and bsd drivers do, so from driver perspective, the issue is the same. What our emulation layer does is turn these 32 bit writes into 4 8-bit writes. This is likely due to some very broken and very old drivers, but I am not sure.

So, the information in the monitor will be broken if the guest
does: write_hi(); write_lo(); A part of me would really like to see a guest that does this :)

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

So the argument for reverting here is that it seems unlikely that the dwords get written in the hi->lo order and we'd rather the monitor get stuck with the wrong mac forever

For how many/which guests? I know it's not linux or BSD. I need to boot windows to see what it does, but I think it does the right thing.

How many guests do you plan to test?

I think the proposal was to see if anyone reports an issue :)

Quoted 13 lines
Quoted 13 lines
Quoted 3 lines

than it show the wrong mac address for
a tiny window for everybody else?

Yes, this would happen for everybody. If you are querying the output, you might see this and it will show up as 2 changes.

We are talking about 2 "tiny" amounts here:
On the one hand, we __might__ have guests that write mac and reverse order thus showing wrong address.
On the other hand we have all the guests who will show the wrong address for a _short_ time.

Quoted 7 lines
Quoted 3 lines

I have a hard time deciding, but have a slight preference for a small uncertainty (the # of backward writers is very small), rather then a small certainty (everyone will be effected for a small period of time).

Why compromise? Why not implement the register that handles whether the mac is valid on RTL?

Quoted 8 lines

I am working on this but it's too late for 1.7.

For 1.7, we have a question: impact a very small number (# -> 0) of guests for a long time, or impact a everyone for a very short time.

Original code did the former and fixed the original issue reported. The commit in question that we are talking about reverting did the later simply because we didn't know any better.

This thread is getting too long. What I've learned from it is that there are valid arguments backed by the spec to indicate why the version of e1000 prior to cd5be582 was more correct and why we think it does not update the monitor with incorrect mac information. rtl8139 is less convincing, but all the drivers we know about behave in one way that allows us to make an assumption about write order and avoid spurious, incorrect mac address updates to the monitor. Had that information been provided in the commit log we could have all spent the day doing relevant work. Please send a v2. Thanks,

Quoted 1 line

Alex

I tweaked commit log before sending pull request.

Quoted 101 lines
Quoted 99 lines
Quoted 12 lines
Quoted 12 lines
Quoted 11 lines

I think you say something about
sub-optimal here...

Quoted 5 lines

The current code isn't perfect either. It still has a potential to show the wrong mac address in the monitor. I doesn't make a lot of sense to me to replace one sub-optimal solution with another sub-optimal solution, especially since no-one complained.

Exactly, the code isn't perfect either way and this revert is just replacing one sub-optimal solution for another. So why do it?

Quoted 6 lines
Quoted 3 lines

Another part of this, for me, is that a change was made that we had a disagreement on and a different approach was being discussed. The revert is to bring us back to before this change.

I was surprised to see it had been committed too, but that's a different argument for revert than what's being presented here. Thanks,

It is another argument for the revert non the less.

Thanks
-vlad

Quoted 74 lines

Alex

Quoted 68 lines
Quoted 67 lines
Quoted 64 lines
Quoted 63 lines
Quoted 2 lines

The patch that was applied was controversial and more then 1 person expressed reservations.

Understood, but it also raised and addressed a shortcoming in the previous patches. If cd5be582 was controversial because the monitor was getting updated with incorrect mac addresses then this simple revert doesn't solve that problem. Thanks,

Alex

Quoted 50 lines
Quoted 49 lines

None of these change the behavior of hardware, they only change when the monitor gets told about mac address changes. I'd suggest either add the emulation described in each spec or revert all of them. A partial revert is just noise. Thanks,

Alex

Quoted 38 lines
Cc Amos Kong <[email protected]>
Cc Alex Williamson <[email protected]>
---
 hw/net/e1000.c   | 2 +-
 hw/net/rtl8139.c | 5 ++++-
 2 files changed, 5 insertions(+), 2 deletions(-)
hw/net/e1000.c +1 -1
index ae63591..8387443 100644
@@ -1106,7 +1106,7 @@ mac_writereg(E1000State *s, int index, uint32_t val)
 
     s->mac_reg[index] = val;
 
-    if (index == RA || index == RA + 1) {
+    if (index == RA + 1) {
         macaddr[0] = cpu_to_le32(s->mac_reg[RA]);
         macaddr[1] = cpu_to_le32(s->mac_reg[RA + 1]);
         qemu_format_nic_info_str(qemu_get_queue(s->nic), (uint8_t *)macaddr);
hw/net/rtl8139.c +4 -1
index 7f2b4db..5329f44 100644
@@ -2741,7 +2741,10 @@ static void rtl8139_io_writeb(void *opaque, uint8_t addr, uint32_t val)
 
     switch (addr)
     {
-        case MAC0 ... MAC0+5:
+        case MAC0 ... MAC0+4:
+            s->phys[addr - MAC0] = val;
+            break;
+        case MAC0+5:
             s->phys[addr - MAC0] = val;
             qemu_format_nic_info_str(qemu_get_queue(s->nic), s->phys);
             break;