Ethernet i210 (e1000 driver) on tegra K1
Hi,
I'm using a ethernet controller intel i210
(http://www.commell.com.tw/product/Surveillance/MPX-210.htm) on my nvidia tegra k1 jetson.
I not an expert with pci, but the only way to make it working in u-boot (upstream) is with the workaround below.
E1000 is very common, so finding a critical bug in this driver seems weird...
Do you think there is a bug in e1000.c or in tegra pci layer?
drivers/net/e1000.c +0 -0
@@ -5186,7 +5186,7 @@ static int _e1000_transmit(struct e1000_hw *hw,
void *txpacket, int length)
txp = tx_base + tx_tail;
tx_tail = (tx_tail + 1) % 8;- txp->buffer_addr = cpu_to_le64(virt_to_bus(hw->pdev, nv_packet));
+ txp->buffer_addr = cpu_to_le64((unsigned long) nv_packet);
txp->lower.data = cpu_to_le32(hw->txd_cmd | length);
txp->upper.data = 0;thank you
Signature · 3 lines
Best regards / Bien cordialement Ivan Mercier
Quoted 5 lines · Ivan Mercier
Hi,
I'm using a ethernet controller intel i210
(http://www.commell.com.tw/product/Surveillance/MPX-210.htm) on my nvidia tegra k1 jetson.
(You didn't actually CC anyone involved with Tegra, so I only accidentally noticed this while I was looking at my mailing list folder)
Quoted 10 lines
I not an expert with pci, but the only way to make it working in u-boot (upstream) is with the workaround below.
E1000 is very common, so finding a critical bug in this driver seems weird...
Do you think there is a bug ine1000.cor in tegra pci layer?
drivers/net/e1000.c static int _e1000_transmit(struct e1000_hw *hw, · line 5192
Quoted 8 lines
drivers/net/e1000.c +1
-1@@ -5186,7 +5186,7 @@ static int _e1000_transmit(struct e1000_hw *hw, void *txpacket, int length) txp = tx_base + tx_tail; tx_tail = (tx_tail +1 ) %8 ; - txp->buffer_addr =cpu_to_le64 (virt_to_bus (hw->pdev, nv_packet)); + txp->buffer_addr =cpu_to_le64 ((unsigned long ) nv_packet); txp->lower.data =cpu_to_le32 (hw->txd_cmd | length); txp->upper.data =0 ;
In order to work out what's going on, perhaps you could print out the values of nv_packet and virt_to_bus(hw->pdev, nv_packet).
It's not terribly surprising that removing the call to virt_to_bus works, since IIRC U-Boot on Tegra uses the same address setup for the PCIe bus as for CPU physical addresses as for CPU virtual addresses.
So, the question is: what is virt_to_bus() doing, and is it the right API to call?
I see that virt_to_bus() is defined as:
e1000.c:
#define virt_to_bus(devno, v) pci_virt_to_mem(devno, (void *) (v))
pci.h:
#define pci_virt_to_mem(dev, addr) \
pci_virt_to_bus((dev), (addr), PCI_REGION_MEM)#define pci_virt_to_bus(dev, addr, flags) \
pci_hose_phys_to_bus(pci_bus_to_hose(PCI_BUS(dev)), \
(virt_to_phys(addr)), (flags))I know that the RTL8169 driver works on the same board (it's soldered down and attached to the other PCIe root port on the SoC). For what looks like the same "use case", it seems to call pci_mem_to_phys() which is:
pci.h:
#define pci_mem_to_phys(dev, addr) \
pci_bus_to_phys((dev), (addr), PCI_REGION_MEM)#define pci_bus_to_phys(dev, addr, flags) \
pci_hose_bus_to_phys(pci_bus_to_hose(PCI_BUS(dev)), (addr), \
(flags))That's odd, since one of those does a phys -> bus translation and the other does a bus -> phys translation. That's the opposite direction, so both can't possibly be right. I wonder if those mapping functions are no-ops on whatever architectures the e1000 driver has been used on (and hence have caused no issues), but fail for some reason on ARM?
(The other difference is the call to virt_to_phys(), but that's a no-op on ARM as far as I can tell).
Quoted 37 lines · Stephen Warren
Quoted 6 lines · Ivan Mercier
Hi,
I'm using a ethernet controller intel i210
(http://www.commell.com.tw/product/Surveillance/MPX-210.htm) on my nvidia tegra k1 jetson.(You didn't actually CC anyone involved with Tegra, so I only accidentally noticed this while I was looking at my mailing list folder)
Quoted 12 lines
I not an expert with pci, but the only way to make it working in u-boot (upstream) is with the workaround below.
E1000 is very common, so finding a critical bug in this driver seems weird...
Do you think there is a bug ine1000.cor in tegra pci layer?drivers/net/e1000.c static int _e1000_transmit(struct e1000_hw *hw, · line 5192
Quoted 8 lines
drivers/net/e1000.c +1
-1@@ -5186,7 +5186,7 @@ static int _e1000_transmit(struct e1000_hw *hw, void *txpacket, int length) txp = tx_base + tx_tail; tx_tail = (tx_tail +1 ) %8 ; - txp->buffer_addr =cpu_to_le64 (virt_to_bus (hw->pdev, nv_packet)); + txp->buffer_addr =cpu_to_le64 ((unsigned long ) nv_packet); txp->lower.data =cpu_to_le32 (hw->txd_cmd | length); txp->upper.data =0 ;In order to work out what's going on, perhaps you could print out the values of
nv_packetandvirt_to_bus(hw->pdev, nv_packet).
Quoted 6 lines
It's not terribly surprising that removing the call to
virt_to_busworks, since IIRC U-Boot on Tegra uses the same address setup for the PCIe bus as for CPU physical addresses as for CPU virtual addresses.So, the question is: what is
virt_to_bus()doing, and is it the right API to call?
virt_to_bus() is to translate cpu virtual address to pci bus physical address. Basically two levels of translation.
Quoted 25 lines
I see that
virt_to_bus()is defined as:
e1000.c:#define
virt_to_bus(devno, v)pci_virt_to_mem(devno, (void *) (v))
pci.h:#define pci_virt_to_mem(dev, addr) \ pci_virt_to_bus((dev), (addr), PCI_REGION_MEM)#define pci_virt_to_bus(dev, addr, flags) \ pci_hose_phys_to_bus(pci_bus_to_hose(PCI_BUS(dev)), \ (virt_to_phys(addr)), (flags))I know that the RTL8169 driver works on the same board (it's soldered down and attached to the other PCIe root port on the SoC). For what looks like the same "use case", it seems to call
pci_mem_to_phys()which is:
pci.h:#define pci_mem_to_phys(dev, addr) \ pci_bus_to_phys((dev), (addr), PCI_REGION_MEM)
Quoted 7 lines
#define pci_bus_to_phys(dev, addr, flags) \ pci_hose_bus_to_phys(pci_bus_to_hose(PCI_BUS(dev)), (addr), \ (flags))That's odd, since one of those does a phys -> bus translation and the other does a bus -> phys translation. That's the opposite direction, so both can't possibly be right. I wonder if those mapping functions are no-ops on
Indeed these address translation macros in pci.h are confusing. But most SoC defines 1:1:1 mapping so these are not needed. The reason why it fails on Tegra is probably due to the addresses are not 1:1:1 mapped?
Quoted 5 lines
whatever architectures the e1000 driver has been used on (and hence have caused no issues), but fail for some reason on ARM?
(The other difference is the call to
virt_to_phys(), but that's a no-op on ARM as far as I can tell).
Regards,
Bin
Quoted 76 lines · Bin Meng
Quoted 37 lines · Stephen Warren
Quoted 6 lines · Ivan Mercier
Hi,
I'm using a ethernet controller intel i210
(http://www.commell.com.tw/product/Surveillance/MPX-210.htm) on my nvidia tegra k1 jetson.(You didn't actually CC anyone involved with Tegra, so I only accidentally noticed this while I was looking at my mailing list folder)
Quoted 12 lines
I not an expert with pci, but the only way to make it working in u-boot (upstream) is with the workaround below.
E1000 is very common, so finding a critical bug in this driver seems weird...
Do you think there is a bug ine1000.cor in tegra pci layer?drivers/net/e1000.c static int _e1000_transmit(struct e1000_hw *hw, · line 5192
Quoted 8 lines
drivers/net/e1000.c +1
-1@@ -5186,7 +5186,7 @@ static int _e1000_transmit(struct e1000_hw *hw, void *txpacket, int length) txp = tx_base + tx_tail; tx_tail = (tx_tail +1 ) %8 ; - txp->buffer_addr =cpu_to_le64 (virt_to_bus (hw->pdev, nv_packet)); + txp->buffer_addr =cpu_to_le64 ((unsigned long ) nv_packet); txp->lower.data =cpu_to_le32 (hw->txd_cmd | length); txp->upper.data =0 ;In order to work out what's going on, perhaps you could print out the values of
nv_packetandvirt_to_bus(hw->pdev, nv_packet).Quoted 6 lines
It's not terribly surprising that removing the call to
virt_to_busworks, since IIRC U-Boot on Tegra uses the same address setup for the PCIe bus as for CPU physical addresses as for CPU virtual addresses.So, the question is: what is
virt_to_bus()doing, and is it the right API to call?
virt_to_bus()is to translate cpu virtual address to pci bus physical address. Basically two levels of translation.Quoted 28 lines
I see that
virt_to_bus()is defined as:
e1000.c:#define
virt_to_bus(devno, v)pci_virt_to_mem(devno, (void *) (v))
pci.h:#define pci_virt_to_mem(dev, addr) \ pci_virt_to_bus((dev), (addr), PCI_REGION_MEM)#define pci_virt_to_bus(dev, addr, flags) \ pci_hose_phys_to_bus(pci_bus_to_hose(PCI_BUS(dev)), \ (virt_to_phys(addr)), (flags))I know that the RTL8169 driver works on the same board (it's soldered down and attached to the other PCIe root port on the SoC). For what looks like the same "use case", it seems to call
pci_mem_to_phys()which is:
pci.h:#define pci_mem_to_phys(dev, addr) \ pci_bus_to_phys((dev), (addr), PCI_REGION_MEM)#define pci_bus_to_phys(dev, addr, flags) \ pci_hose_bus_to_phys(pci_bus_to_hose(PCI_BUS(dev)), (addr), \ (flags))
Quoted 8 lines
Quoted 3 lines
That's odd, since one of those does a phys -> bus translation and the other does a bus -> phys translation. That's the opposite direction, so both can't possibly be right. I wonder if those mapping functions are no-ops on
Indeed these address translation macros in
pci.hare confusing. But most SoC defines 1:1:1 mapping so these are not needed. The reason why it fails on Tegra is probably due to the addresses are not 1:1:1 mapped?
I'm pretty sure everything is 1:1. That's why removing the call to any translation macro made the e1000 driver work for Ivan.
Hi Stephen
By accident I stumbled over this post and by incident now that we at Toradex are working on an Apalis TK1 module design using the same i210/i211 gigabit Ethernet chip as already widely deployed on Apalis T30 we are facing this as well.
Quoted 13 lines · Stephen Warren
Quoted 12 lines · Bin Meng
Quoted 10 lines · Stephen Warren
Quoted 7 lines · Ivan Mercier
Hi,
I'm using a ethernet controller intel i210
(http://www.commell.com.tw/product/Surveillance/MPX-210.htm) on my
nvidia tegra k1 jetson.
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
(You didn't actually CC anyone involved with Tegra, so I only accidentally
noticed this while I was looking at my mailing list folder)
Exactly. Looking through the e1000 git log one should also have seen that me and others actually worked on it and that we are indeed successfully using it on Tegra albeit T30 only so far.
Quoted 19 lines
Quoted 19 lines
Quoted 19 lines
Quoted 19 lines
I not an expert with pci, but the only way to make it working in u-boot
(upstream) is with the workaround below.E1000 is very common, so finding a critical bug in this driver seems
weird...
Do you think there is a bug ine1000.cor in tegra pci layer?drivers/net/e1000.c +0
-0@@ -5186,7 +5186,7 @@ static int _e1000_transmit(struct e1000_hw *hw, void *txpacket, int length) ???????txp = tx_base + tx_tail; ???????tx_tail = (tx_tail + 1) % 8;
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 5 lines
drivers/net/e1000.c +1
-1-????txp->buffer_addr = cpu_to_le64(virt_to_bus(hw->pdev, nv_packet)); +????txp->buffer_addr = cpu_to_le64((unsigned long) nv_packet);???????txp->lower.data =
cpu_to_le32(hw->txd_cmd | length); ???????txp->upper.data = 0;In order to work out what's going on, perhaps you could print out the values
ofnv_packetandvirt_to_bus(hw->pdev, nv_packet).
I did exactly that which gives:
Loading: nv_packet = 0xfffecb00pci_hose_phys_to_bus: invalid physical addressvirt_to_bus(hw->pdev, nv_packet) = 0x00000000
So that looks OK but does not yet give us much further clues.
Looking at where the actual translation happens
in?drivers/pci/pci_common.c:
int __pci_hose_phys_to_bus(struct pci_controller *hose,
...
if (bus_addr >= res->bus_start &&
????bus_addr < res->bus_start + res->size) {
*ba = bus_addr;...
So I added some more prints there giving:
bus_addr = 0xfffecb00phys_addr = 0xfffecb00res->phys_start = 0x80000000res->bus_start = 0x80000000res->size = 0x80000000
Ah, OK. That makes sense now. Just a classical overflow happening in the address range check (e.g. 0x80000000 + 0x80000000 giving 0x0 in 32- bits).
Strange that this works for others but e.g. the Realtek driver seems to do the inverse conversion:
int __pci_hose_bus_to_phys(struct pci_controller *hose,
...
if (bus_addr >= res->bus_start &&
????(bus_addr - res->bus_start) < res->size) {
*pa = (bus_addr - res->bus_start + res-Quoted 1 line
phys_start);
Ah, they worked around it by changing the order of the addition vs. subtraction in the range check. That will work.
I will be sending out a proper patch to fix this shortly.
But why does it work on T30?
bus_addr = 0xffee92e0phys_addr = 0xffee92e0res->phys_start = 0x80000000res->bus_start = 0x80000000res->size = 0x7ff00000
Ah funny, ooks like we got lucky due to the size being slightly below 0x80000000 (;-p).
Quoted 59 lines
Quoted 53 lines
Quoted 9 lines
It's not terribly surprising that removing the call to
virt_to_busworks,
since IIRC U-Boot on Tegra uses the same address setup for the PCIe bus as
for CPU physical addresses as for CPU virtual addresses.So, the question is: what is
virt_to_bus()doing, and is it the right API to
call?
virt_to_bus()is to translate cpu virtual address to pci bus physical
address. Basically two levels of translation.Quoted 32 lines
I see that
virt_to_bus()is defined as:
e1000.c:#define
virt_to_bus(devno, v)???pci_virt_to_mem(devno, (void *) (v))
pci.h:#define
pci_virt_to_mem(dev, addr)\
?????????pci_virt_to_bus((dev), (addr),PCI_REGION_MEM)#define
pci_virt_to_bus(dev, addr, flags)\
?????????pci_hose_phys_to_bus(pci_bus_to_hose(PCI_BUS(dev)), \
??????????????????????????????(virt_to_phys(addr)), (flags))I know that the RTL8169 driver works on the same board (it's soldered down
and attached to the other PCIe root port on the SoC). For what looks like
the same "use case", it seems to callpci_mem_to_phys()which is:
pci.h:#define
pci_mem_to_phys(dev, addr)\
?????????pci_bus_to_phys((dev), (addr),PCI_REGION_MEM)#define
pci_bus_to_phys(dev, addr, flags)\
?????????pci_hose_bus_to_phys(pci_bus_to_hose(PCI_BUS(dev)), (addr), \
?????????????????????????????????(flags))Quoted 6 lines
That's odd, since one of those does a phys -> bus translation and the other
does a bus -> phys translation. That's the opposite direction, so both can't
possibly be right. I wonder if those mapping functions are no-ops onQuoted 5 lines
Indeed these address translation macros in
pci.hare confusing. But most SoC defines 1:1:1 mapping so these are not needed. The reason why
it fails on Tegra is probably due to the addresses are not 1:1:1 mapped?
Quoted 3 lines
I'm pretty sure everything is 1:1. That's why removing the call to any
translation macro made the e1000 driver work for Ivan.
Yes, it's just 1:1:1 mapped.
@Ivan: Thank you very much for having reported this. Saves me some nightmare on bringing up our Apalis TK1 when time will be short next year couple weeks before the Embedded World show (;-p).
Cheers
Marcel