On Wednesday 03 February 2016 16:21:05 Grygorii Strashko wrote:
> On 02/03/2016 04:11 PM, Franklin S Cooper Jr. wrote:
> > On 02/02/2016 07:19 PM, Franklin S Cooper Jr. wrote:
> >>>
> >> So only making this change on the latest master with no
> >> other changes I see the boot problem again.
>
> yep. I can confirm that.
>
> Also, I'm today came up with the similar fix that you've proposed before in
> this thread.
> So, Could we move forward this way?
I still think it would be good to actually understand what the actual
problem was.
> From 8280895f01b33edba303e7374431bef47630f26c Mon Sep 17 00:00:00 2001
> From: Grygorii Strashko <grygorii.stras...@ti.com>
> Date: Wed, 3 Feb 2016 15:11:40 +0200
> Subject: [PATCH] net: ti: netcp: restore get/set_pad_info() functionality
>
> The commit 899077791403 ("netcp: try to reduce type confusion in descriptors")
> introduces a regression in Kernel 4.5-rc1 as it breaks
> get/set_pad_info() functionality.
>
> The TI NETCP driver uses pad0 and pad1 fields of knav_dma_desc to
> store DMA/MEM buffer pointer and buffer size respectively. And in both
> cases the pointer type size is 32 bit regardless of LAPE enabled or
> not.
> !LAPE LPAE
> sizeof(void*) 32bit 32bit
> sizeof(dma_addr_t) 32bit 32bit
> sizeof(phys_addr_t) 32bit 64bit
>
This looks wrong: I was getting the build warnings originally
because of 64-bit dma_addr_t, and that should be the only way that
this driver can operate, because in some configurations on keystone
there is no memory below 4GB, and there is no dma-ranges property
in the DT that shifts around the start of the DMA addresses.
This doesn't all fit together yet, maybe you have a better idea
of what is going on.
I don't think we should ever have a platform that has dma_addr_t
and phys_addr_t be different.
> Unfortunately, above commit changed buffer's pointers save/restore
> code (get/set_pad_info()) and added intermediate conversation to u64
> which causes TI NETCP driver crash in RX/TX path due to "Unable to
> handle kernel NULL pointer" exception.
The conversion is not what made it crash, it must be a bug in the
conversion ;-)
> @@ -163,11 +153,10 @@ static void set_desc_info(u32 desc_info, u32 pkt_info,
> desc->packet_info = cpu_to_le32(pkt_info);
> }
>
> -static void set_pad_info(u32 pad0, u32 pad1, u32 pad2, struct knav_dma_desc
> *desc)
> +static void set_pad_info(u32 pad0, u32 pad1, struct knav_dma_desc *desc)
> {
> desc->pad[0] = cpu_to_le32(pad0);
> desc->pad[1] = cpu_to_le32(pad1);
> - desc->pad[2] = cpu_to_le32(pad1);
> }
As in my earlier patch, the line you are removing here was clearly
broken, but evidently that is not the only bug.
> static void set_org_pkt_info(dma_addr_t buff, u32 buff_len,
> @@ -581,7 +570,6 @@ static void netcp_free_rx_desc_chain(struct netcp_intf
> *netcp,
> dma_addr_t dma_desc, dma_buf;
> unsigned int buf_len, dma_sz = sizeof(*ndesc);
> void *buf_ptr;
> - u32 pad[2];
> u32 tmp;
>
> get_words(&dma_desc, 1, &desc->next_desc);
> @@ -593,14 +581,12 @@ static void netcp_free_rx_desc_chain(struct netcp_intf
> *netcp,
> break;
> }
> get_pkt_info(&dma_buf, &tmp, &dma_desc, ndesc);
> - get_pad_ptr(&buf_ptr, ndesc);
> + get_pad_info((u32 *)&buf_ptr, &buf_len, ndesc);
I'd prefer not to put code like this back, as this cannot possibly
do the right thing on a 64-bit architecture.
> @@ -1078,7 +1058,6 @@ netcp_tx_map_skb(struct sk_buff *skb, struct netcp_intf
> *netcp)
> u32 page_offset = frag->page_offset;
> u32 buf_len = skb_frag_size(frag);
> dma_addr_t desc_dma;
> - u32 desc_dma_32;
> u32 pkt_info;
>
> dma_addr = dma_map_page(dev, page, page_offset, buf_len,
> @@ -1100,8 +1079,7 @@ netcp_tx_map_skb(struct sk_buff *skb, struct netcp_intf
> *netcp)
> (netcp->tx_compl_qid & KNAV_DMA_DESC_RETQ_MASK) <<
> KNAV_DMA_DESC_RETQ_SHIFT;
> set_pkt_info(dma_addr, buf_len, 0, ndesc);
> - desc_dma_32 = (u32)desc_dma;
> - set_words(&desc_dma_32, 1, &pdesc->next_desc);
> + set_words(&desc_dma, 1, &pdesc->next_desc);
> pkt_len += buf_len;
> if (pdesc != desc)
> knav_pool_desc_map(netcp->tx_pool, pdesc,
This is clearly broken on big-endian kernels with 64-bit dma_addr_t, so even
if we revert the rest I think this part has to stay.
I have another version for testing below. That removes the logic that
splits and reassembles the 64-bit values, but leaves the other changes
in place. Can you try this?
Arnd
diff --git a/drivers/net/ethernet/ti/netcp_core.c
b/drivers/net/ethernet/ti/netcp_core.c
index c61d66d38634..cda19f2401c1 100644
--- a/drivers/net/ethernet/ti/netcp_core.c
+++ b/drivers/net/ethernet/ti/netcp_core.c
@@ -120,16 +120,14 @@ static void get_pkt_info(dma_addr_t *buff, u32 *buff_len,
dma_addr_t *ndesc,
static void get_pad_info(u32 *pad0, u32 *pad1, u32 *pad2, struct knav_dma_desc
*desc)
{
*pad0 = le32_to_cpu(desc->pad[0]);
- *pad1 = le32_to_cpu(desc->pad[1]);
- *pad2 = le32_to_cpu(desc->pad[2]);
+ *pad2 = le32_to_cpu(desc->pad[1]);
}
static void get_pad_ptr(void **padptr, struct knav_dma_desc *desc)
{
u64 pad64;
- pad64 = le32_to_cpu(desc->pad[0]) +
- ((u64)le32_to_cpu(desc->pad[1]) << 32);
+ pad64 = le32_to_cpu(desc->pad[0]);
*padptr = (void *)(uintptr_t)pad64;
}
@@ -166,8 +164,7 @@ static void set_desc_info(u32 desc_info, u32 pkt_info,
static void set_pad_info(u32 pad0, u32 pad1, u32 pad2, struct knav_dma_desc
*desc)
{
desc->pad[0] = cpu_to_le32(pad0);
- desc->pad[1] = cpu_to_le32(pad1);
- desc->pad[2] = cpu_to_le32(pad1);
+ desc->pad[1] = cpu_to_le32(pad2);
}
static void set_org_pkt_info(dma_addr_t buff, u32 buff_len,
@@ -581,7 +578,7 @@ static void netcp_free_rx_desc_chain(struct netcp_intf
*netcp,
dma_addr_t dma_desc, dma_buf;
unsigned int buf_len, dma_sz = sizeof(*ndesc);
void *buf_ptr;
- u32 pad[2];
+ u32 pad;
u32 tmp;
get_words(&dma_desc, 1, &desc->next_desc);
@@ -599,8 +596,8 @@ static void netcp_free_rx_desc_chain(struct netcp_intf
*netcp,
knav_pool_desc_put(netcp->rx_pool, desc);
}
- get_pad_info(&pad[0], &pad[1], &buf_len, desc);
- buf_ptr = (void *)(uintptr_t)(pad[0] + ((u64)pad[1] << 32));
+ get_pad_info(&pad, NULL, &buf_len, desc);
+ buf_ptr = (void *)(uintptr_t)(pad);
if (buf_ptr)
netcp_frag_free(buf_len <= PAGE_SIZE, buf_ptr);
@@ -653,8 +650,8 @@ static int netcp_process_one_rx_packet(struct netcp_intf
*netcp)
}
get_pkt_info(&dma_buff, &buf_len, &dma_desc, desc);
- get_pad_info(&pad[0], &pad[1], &org_buf_len, desc);
- org_buf_ptr = (void *)(uintptr_t)(pad[0] + ((u64)pad[1] << 32));
+ get_pad_info(&pad[0], NULL, &org_buf_len, desc);
+ org_buf_ptr = (void *)(uintptr_t)(pad[0]);
if (unlikely(!org_buf_ptr)) {
dev_err(netcp->ndev_dev, "NULL bufptr in desc\n");
@@ -858,8 +855,7 @@ static int netcp_allocate_rx_buf(struct netcp_intf *netcp,
int fdq)
if (unlikely(dma_mapping_error(netcp->dev, dma)))
goto fail;
- pad[0] = lower_32_bits((uintptr_t)bufptr);
- pad[1] = upper_32_bits((uintptr_t)bufptr);
+ pad[0] = (uintptr_t)bufptr;
} else {
/* Allocate a secondary receive queue entry */
@@ -870,8 +866,7 @@ 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] = (u32)(dma);
pad[2] = 0;
}
@@ -882,7 +877,7 @@ static int netcp_allocate_rx_buf(struct netcp_intf *netcp,
int fdq)
pkt_info |= (netcp->rx_queue_id & KNAV_DMA_DESC_RETQ_MASK) <<
KNAV_DMA_DESC_RETQ_SHIFT;
set_org_pkt_info(dma, buf_len, hwdesc);
- set_pad_info(pad[0], pad[1], pad[2], hwdesc);
+ set_pad_info(pad[0], 0, pad[2], hwdesc);
set_desc_info(desc_info, pkt_info, hwdesc);
/* Push to FDQs */
@@ -1194,10 +1189,8 @@ 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 = (uintptr_t)&skb;
set_words(&tmp, 1, &desc->pad[0]);
- tmp = upper_32_bits((uintptr_t)&skb);
- set_words(&tmp, 1, &desc->pad[1]);
if (tx_pipe->flags & SWITCH_TO_PORT_IN_TAGINFO) {
tmp = tx_pipe->switch_to_port;
