Message ID | 56CB821E.7040809@ti.com |
---|---|
State | New |
Headers | show |
On Monday 22 February 2016 16:48:14 Murali Karicheri wrote: > Arnd, > > As promised, here is what I found wrong with the commit 899077 that introduced a > regression. With these changes, I am able to boot kernel without issues on K2 platforms. Thanks so much for looking into this! > From the commit description, it appears that you are trying to make the driver do the right > thing if compiled for a 64 bit systems. Is it mandatory for all kernel drivers to be > 64bit compliant? Similar question on supporting mixed endian in all kernel drivers. I would generally expect all device driver code to be written as architecture-independent as possible, for multiple reasons: * hardware gets reused all the time, we have plenty of drivers that started out on big-endian powerpc32 or mips32 and are now used on 64-bit little-endian arm, so you should not make any assumptions * code gets copied into other drivers, so whatever you write should be able to serve as an example to other developers > Keystone can have SoC configured to be in big endian mode for peripherals and DSP. I'm not entirely sure what this means > so that is > something we need to support if there is customer interest. Wondering why do one run BE kernel > binary on these platforms? Any reason? I saw some reference to that in past discussion on this > regression issue. The only real reason to run a big-endian ARM kernel is for compatibility with user space that has either is known to not be portable to little-endian, or that has only ever been used on big-endian machines and that might have unknown problems. This is typically the case for proprietary user space network stacks of the kind you find in commercial network infrastructure hardware, but there are a couple of other uses in enterprise systems that have source code ported from mainframes. > diff --git a/drivers/net/ethernet/ti/netcp_core.c b/drivers/net/ethernet/ti/netcp_core.c > index c61d66d..ac35161 100644 > --- a/drivers/net/ethernet/ti/netcp_core.c > +++ b/drivers/net/ethernet/ti/netcp_core.c > @@ -167,7 +167,7 @@ static void set_pad_info(u32 pad0, u32 pad1, u32 pad2, struct knav_dma_desc *des > { > desc->pad[0] = cpu_to_le32(pad0); > desc->pad[1] = cpu_to_le32(pad1); > - desc->pad[2] = cpu_to_le32(pad1); > + desc->pad[2] = cpu_to_le32(pad2); > } I had found this hunk earlier. > static void set_org_pkt_info(dma_addr_t buff, u32 buff_len, > @@ -870,8 +870,8 @@ static int netcp_allocate_rx_buf(struct netcp_intf *netcp, int fdq) > } > buf_len = PAGE_SIZE; > dma = dma_map_page(netcp->dev, page, 0, buf_len, DMA_TO_DEVICE); > - pad[0] = lower_32_bits(dma); > - pad[1] = upper_32_bits(dma); > + pad[0] = lower_32_bits((uintptr_t)page); > + pad[1] = upper_32_bits((uintptr_t)page); > pad[2] = 0; > } And this is my stupid mistake that I failed to see. > @@ -1194,9 +1194,9 @@ static int netcp_tx_submit_skb(struct netcp_intf *netcp, > } > > set_words(&tmp, 1, &desc->packet_info); > - tmp = lower_32_bits((uintptr_t)&skb); > + tmp = lower_32_bits((uintptr_t)skb); > set_words(&tmp, 1, &desc->pad[0]); > - tmp = upper_32_bits((uintptr_t)&skb); > + tmp = upper_32_bits((uintptr_t)skb); > set_words(&tmp, 1, &desc->pad[1]); > > if (tx_pipe->flags & SWITCH_TO_PORT_IN_TAGINFO) { And this is another one of the same sort. Not my best patch ever obviously, but at least I understand where I went wrong now, and see that it was only me being sloppy in the conversion rather than a more fundamental misdesign. Thanks, Arnd
Arnd, On 02/22/2016 05:13 PM, Arnd Bergmann wrote: > On Monday 22 February 2016 16:48:14 Murali Karicheri wrote: >> Arnd, >> >> As promised, here is what I found wrong with the commit 899077 that introduced a >> regression. With these changes, I am able to boot kernel without issues on K2 platforms. > > Thanks so much for looking into this! > >> From the commit description, it appears that you are trying to make the driver do the right >> thing if compiled for a 64 bit systems. Is it mandatory for all kernel drivers to be >> 64bit compliant? Similar question on supporting mixed endian in all kernel drivers. > > I would generally expect all device driver code to be written as architecture-independent > as possible, for multiple reasons: > > * hardware gets reused all the time, we have plenty of drivers that started out on > big-endian powerpc32 or mips32 and are now used on 64-bit little-endian arm, so > you should not make any assumptions > > * code gets copied into other drivers, so whatever you write should be able to serve > as an example to other developers > Ok. Got it. >> Keystone can have SoC configured to be in big endian mode for peripherals and DSP. > > I'm not entirely sure what this means This means, ARM core can be using LE/BE and rest of the system can be configured through a pin (to SOC) to operate in BE/LE. So need to take care of all mixed endian configuration properly. Refer to http://www.ti.com/lit/ds/symlink/tci6636k2h.pdf for more details if interested. > >> so that is >> something we need to support if there is customer interest. Wondering why do one run BE kernel >> binary on these platforms? Any reason? I saw some reference to that in past discussion on this >> regression issue. > > The only real reason to run a big-endian ARM kernel is for compatibility with user space > that has either is known to not be portable to little-endian, or that has only ever been > used on big-endian machines and that might have unknown problems. > > This is typically the case for proprietary user space network stacks of the kind you > find in commercial network infrastructure hardware, but there are a couple of other > uses in enterprise systems that have source code ported from mainframes. > >> diff --git a/drivers/net/ethernet/ti/netcp_core.c b/drivers/net/ethernet/ti/netcp_core.c >> index c61d66d..ac35161 100644 >> --- a/drivers/net/ethernet/ti/netcp_core.c >> +++ b/drivers/net/ethernet/ti/netcp_core.c >> @@ -167,7 +167,7 @@ static void set_pad_info(u32 pad0, u32 pad1, u32 pad2, struct knav_dma_desc *des >> { >> desc->pad[0] = cpu_to_le32(pad0); >> desc->pad[1] = cpu_to_le32(pad1); >> - desc->pad[2] = cpu_to_le32(pad1); >> + desc->pad[2] = cpu_to_le32(pad2); >> } > > I had found this hunk earlier. > >> static void set_org_pkt_info(dma_addr_t buff, u32 buff_len, >> @@ -870,8 +870,8 @@ static int netcp_allocate_rx_buf(struct netcp_intf *netcp, int fdq) >> } >> buf_len = PAGE_SIZE; >> dma = dma_map_page(netcp->dev, page, 0, buf_len, DMA_TO_DEVICE); >> - pad[0] = lower_32_bits(dma); >> - pad[1] = upper_32_bits(dma); >> + pad[0] = lower_32_bits((uintptr_t)page); >> + pad[1] = upper_32_bits((uintptr_t)page); >> pad[2] = 0; >> } > > And this is my stupid mistake that I failed to see. > >> @@ -1194,9 +1194,9 @@ static int netcp_tx_submit_skb(struct netcp_intf *netcp, >> } >> >> set_words(&tmp, 1, &desc->packet_info); >> - tmp = lower_32_bits((uintptr_t)&skb); >> + tmp = lower_32_bits((uintptr_t)skb); >> set_words(&tmp, 1, &desc->pad[0]); >> - tmp = upper_32_bits((uintptr_t)&skb); >> + tmp = upper_32_bits((uintptr_t)skb); >> set_words(&tmp, 1, &desc->pad[1]); >> >> if (tx_pipe->flags & SWITCH_TO_PORT_IN_TAGINFO) { > > And this is another one of the same sort. > > Not my best patch ever obviously, but at least I understand where I went > wrong now, and see that it was only me being sloppy in the conversion rather > than a more fundamental misdesign. > So do you plan to re-spin the patch again with the above change? Murali > Thanks, > > Arnd > -- Murali Karicheri Linux Kernel, Keystone
On Monday 22 February 2016 18:09:45 Murali Karicheri wrote: > On 02/22/2016 05:13 PM, Arnd Bergmann wrote: > > On Monday 22 February 2016 16:48:14 Murali Karicheri wrote: > >> Keystone can have SoC configured to be in big endian mode for peripherals and DSP. > > > > I'm not entirely sure what this means > > This means, ARM core can be using LE/BE and rest of the system can be configured through a pin > (to SOC) to operate in BE/LE. So need to take care of all mixed endian configuration > properly. Refer to http://www.ti.com/lit/ds/symlink/tci6636k2h.pdf for more details if interested. Ah, that is unfortunate. For the moment, let's hope that nobody ever uses that pin, so we can support both big-endian and little-endian kernels without having to do any runtime detection of the board settings. If anyone is misguided enough to flip that configuration pin to the opposite setting, then we will require extra DT properties to list for each device what endianess the hardware uses, and check that flag at runtime, which causes a little extra overhead. If you are able to provide any feedback to the hardware designers, please ask them to never do configurable endianess again: it solves no actual problems but creates endless nightmares if anyone ever uses that. > > And this is another one of the same sort. > > > > Not my best patch ever obviously, but at least I understand where I went > > wrong now, and see that it was only me being sloppy in the conversion rather > > than a more fundamental misdesign. > > > > So do you plan to re-spin the patch again with the above change? I probably won't bother at this point, but if you could send a patch to get back to my version with your fixes, I'll send an Ack. Arnd
diff --git a/drivers/net/ethernet/ti/netcp_core.c b/drivers/net/ethernet/ti/netcp_core.c index c61d66d..ac35161 100644 --- a/drivers/net/ethernet/ti/netcp_core.c +++ b/drivers/net/ethernet/ti/netcp_core.c @@ -167,7 +167,7 @@ static void set_pad_info(u32 pad0, u32 pad1, u32 pad2, struct knav_dma_desc *des { desc->pad[0] = cpu_to_le32(pad0); desc->pad[1] = cpu_to_le32(pad1); - desc->pad[2] = cpu_to_le32(pad1); + desc->pad[2] = cpu_to_le32(pad2); } static void set_org_pkt_info(dma_addr_t buff, u32 buff_len, @@ -870,8 +870,8 @@ static int netcp_allocate_rx_buf(struct netcp_intf *netcp, int fdq) } buf_len = PAGE_SIZE; dma = dma_map_page(netcp->dev, page, 0, buf_len, DMA_TO_DEVICE); - pad[0] = lower_32_bits(dma); - pad[1] = upper_32_bits(dma); + pad[0] = lower_32_bits((uintptr_t)page); + pad[1] = upper_32_bits((uintptr_t)page); pad[2] = 0; } @@ -1194,9 +1194,9 @@ static int netcp_tx_submit_skb(struct netcp_intf *netcp, } set_words(&tmp, 1, &desc->packet_info); - tmp = lower_32_bits((uintptr_t)&skb); + tmp = lower_32_bits((uintptr_t)skb); set_words(&tmp, 1, &desc->pad[0]); - tmp = upper_32_bits((uintptr_t)&skb); + tmp = upper_32_bits((uintptr_t)skb); set_words(&tmp, 1, &desc->pad[1]); if (tx_pipe->flags & SWITCH_TO_PORT_IN_TAGINFO) {