igc: Clean RX descriptor error flags
Subsystems (2)
PATCH 1/3 igc: Clean RX descriptor error flags
I225 advanced receive descriptor doesn't have the following extend error
bits: CE, SE, SEQ, CXE. In addition to that, the bit TCPE is called L4E in the datasheet.
This patch cleans up the code accordingly, and gets rid of the macro IGC_RXDEXT_ERR_FRAME_ERR_MASK since it doesn't make much sense anymore.
--- drivers/net/ethernet/intel/igc/igc_defines.h | 14 +------------- drivers/net/ethernet/intel/igc/igc_main.c | 5 ++--- 2 files changed, 3 insertions(+), 16 deletions(-)
drivers/net/ethernet/intel/igc/igc_defines.h +1 -13
@@ -324,22 +324,10 @@/* Advanced Receive Descriptor bit definitions */ #define IGC_RXDADV_STAT_TSIP 0x08000/* timestamp in packet */ -#define IGC_RXDEXT_STATERR_CE 0x01000000 -#define IGC_RXDEXT_STATERR_SE 0x02000000 -#define IGC_RXDEXT_STATERR_SEQ 0x04000000 -#define IGC_RXDEXT_STATERR_CXE 0x10000000 -#define IGC_RXDEXT_STATERR_TCPE 0x20000000 +#define IGC_RXDEXT_STATERR_L4E 0x20000000#define IGC_RXDEXT_STATERR_IPE 0x40000000#define IGC_RXDEXT_STATERR_RXE 0x80000000 -/* Same mask, but for extended and packet split descriptors */ -#define IGC_RXDEXT_ERR_FRAME_ERR_MASK ( \ - IGC_RXDEXT_STATERR_CE | \ - IGC_RXDEXT_STATERR_SE | \ - IGC_RXDEXT_STATERR_SEQ | \ - IGC_RXDEXT_STATERR_CXE | \ - IGC_RXDEXT_STATERR_RXE) -#define IGC_MRQC_RSS_FIELD_IPV4_TCP 0x00010000#define IGC_MRQC_RSS_FIELD_IPV4 0x00020000#define IGC_MRQC_RSS_FIELD_IPV6_TCP_EX 0x00040000
drivers/net/ethernet/intel/igc/igc_main.c +2 -3
@@ -1437,7 +1437,7 @@ static void igc_rx_checksum(struct igc_ring *ring,/* TCP/UDP checksum error bit is set */ if (igc_test_staterr (rx_desc, -IGC_RXDEXT_STATERR_TCPE | +IGC_RXDEXT_STATERR_L4E |IGC_RXDEXT_STATERR_IPE )) {/* work around errata with sctp packets where the TCPE aka * L4E bit is set incorrectly on 64 byte (60 byte w/o crc)
@@ -1752,8 +1752,7 @@ static bool igc_cleanup_headers(struct igc_ring *rx_ring,union igc_adv_rx_desc *rx_desc, struct sk_buff *skb) { - if (unlikely((igc_test_staterr(rx_desc, - IGC_RXDEXT_ERR_FRAME_ERR_MASK)))) { + if (unlikely(igc_test_staterr(rx_desc, IGC_RXDEXT_STATERR_RXE))) { struct net_device *netdev = rx_ring->netdev;if (!(netdev->features &NETIF_F_RXALL )) {
Signature · 1 line
2.26.2
PATCH 2/3 igc: Check descriptor's DD bit in igc_clean_rx_irq()
I225 advanced receive descriptor provides the Descriptor Done (DD) bit which indicates hardware is done with that receive descriptor and software should handle it.
This patch fixes igc_clean_rx_irq() so we check that bit to determine if we are done handling incoming packets instead of checking the packet length information. It also gets rid of rx_desc->wb.upper.length assignments spread through the code required to make the previous approach to work.
--- drivers/net/ethernet/intel/igc/igc_defines.h | 1 + drivers/net/ethernet/intel/igc/igc_main.c | 14 ++++++++------ 2 files changed, 9 insertions(+), 6 deletions(-)
drivers/net/ethernet/intel/igc/igc_defines.h +1 -0
@@ -316,6 +316,7 @@#define IGC_SRRCTL_TIMER0SEL (timer) (((timer) & 0x3) << 17)/* Receive Descriptor bit definitions */ +#define IGC_RXD_STAT_DD 0x01/* Descriptor Done */ #define IGC_RXD_STAT_EOP 0x02/* End of Packet */ #define IGC_RXD_STAT_IXSM 0x04/* Ignore checksum */ #define IGC_RXD_STAT_UDPCS 0x10/* UDP xsum calculated */
drivers/net/ethernet/intel/igc/igc_main.c +8 -6
@@ -551,7 +551,6 @@ static void igc_configure_rx_ring(struct igc_adapter *adapter,/* initialize Rx descriptor 0 */ rx_desc =IGC_RX_DESC (ring,0 ); - rx_desc->wb.upper.length =0 ;/* enable receive descriptor fetching */ rxdctl |=IGC_RXDCTL_QUEUE_ENABLE ;
@@ -1880,9 +1879,6 @@ static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) i -= rx_ring->count; } -/* clear the length for the next_to_use descriptor */ - rx_desc->wb.upper.length =0 ; - cleaned_count--; }while (cleaned_count);
@@ -1924,8 +1920,12 @@ static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget) } rx_desc =IGC_RX_DESC (rx_ring, rx_ring->next_to_clean); - size =le16_to_cpu (rx_desc->wb.upper.length); -if (!size) + +/* If we reached a descriptor with 'Descriptor Done' bit not +* set, it means we have handled all descriptors owned by +* software already so we should prematurely break the loop. +*/ +if (!igc_test_staterr (rx_desc,IGC_RXD_STAT_DD ))break ;/* This memory barrier is needed to keep us from reading
@@ -1934,6 +1934,8 @@ static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget)*/ dma_rmb (); + size =le16_to_cpu (rx_desc->wb.upper.length); + rx_buffer =igc_get_rx_buffer (rx_ring, size);/* retrieve a buffer from the ring */
Signature · 1 line
2.26.2
PATCH 3/3 igc: Fix SRRCTL register setup
In reply to Andre GuedesSRRCTL register is set with 'one buffer descriptor' option (see DESCTYPE setting a few lines below) so setting BSIZEHEADER bits is pointless. They should be zero. Also, since there is no header buffer we should set the header buffer address field from the receive descriptor to zero for the sake of consistency.
--- drivers/net/ethernet/intel/igc/igc_main.c | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/net/ethernet/intel/igc/igc_main.c +1 -3
@@ -531,14 +531,11 @@ static void igc_configure_rx_ring(struct igc_adapter *adapter, ring->next_to_clean =0 ; ring->next_to_use =0 ; -/* set descriptor configuration */ - srrctl =IGC_RX_HDR_LEN <<IGC_SRRCTL_BSIZEHDRSIZE_SHIFT ;if (ring_uses_large_buffer (ring)) srrctl |=IGC_RXBUFFER_3072 >>IGC_SRRCTL_BSIZEPKT_SHIFT ;else srrctl |=IGC_RXBUFFER_2048 >>IGC_SRRCTL_BSIZEPKT_SHIFT ; srrctl |=IGC_SRRCTL_DESCTYPE_ADV_ONEBUF ; -wr32 (IGC_SRRCTL (reg_idx), srrctl); rxdctl |=IGC_RX_PTHRESH ;
@@ -1869,6 +1866,7 @@ static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count)* because each write-back erases this info. */ rx_desc->read.pkt_addr =cpu_to_le64 (bi->dma + bi->page_offset); + rx_desc->read.hdr_addr =0 ; rx_desc++; bi++;
Signature · 1 line
2.26.2
PATCH 3/3 igc: Fix SRRCTL register setup
Quoted 21 lines · Andre Guedes
SRRCTL register is set with 'one buffer descriptor' option (see DESCTYPE setting a few lines below) so setting BSIZEHEADER bits is pointless. They should be zero. Also, since there is no header buffer we should set the header buffer address field from the receive descriptor to zero for the sake of consistency.
Signed-off-by Andre Guedes <[email protected]>--- drivers/net/ethernet/intel/igc/igc_main.c | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-)drivers/net/ethernet/intel/igc/igc_main.c +0
-1@@ -531,14 +531,11 @@ static void igc_configure_rx_ring(struct igc_adapter *adapter, ring->next_to_clean =0 ; ring->next_to_use =0 ; -/* set descriptor configuration */
drivers/net/ethernet/intel/igc/igc_main.c static void igc_configure_rx_ring(struct igc_adapter *adapter, · line 541
Quoted 10 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-2- srrctl =IGC_RX_HDR_LEN <<IGC_SRRCTL_BSIZEHDRSIZE_SHIFT ;if (ring_uses_large_buffer (ring)) srrctl |=IGC_RXBUFFER_3072 >>IGC_SRRCTL_BSIZEPKT_SHIFT ;else srrctl |=IGC_RXBUFFER_2048 >>IGC_SRRCTL_BSIZEPKT_SHIFT ; srrctl |=IGC_SRRCTL_DESCTYPE_ADV_ONEBUF ; -wr32 (IGC_SRRCTL (reg_idx), srrctl); rxdctl |=IGC_RX_PTHRESH ;
Some of this was left in place to leave parity with the ixgbe driver which was required to populate that field in order to enable RSC/LRO.
drivers/net/ethernet/intel/igc/igc_main.c static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) · line 1872
Quoted 8 lines
drivers/net/ethernet/intel/igc/igc_main.c +1
-0@@ -1869,6 +1866,7 @@ static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count)* because each write-back erases this info. */ rx_desc->read.pkt_addr =cpu_to_le64 (bi->dma + bi->page_offset); + rx_desc->read.hdr_addr =0 ; rx_desc++; bi++;
If you are going to do this it would be better to replace the line that is setting the length to zero instead of just adding this line. That way you can avoid having to rewrite it. I only had bothered with clearing the length field as it was a 32b field, however if you are wanting to flush the full 64b then I would recommend doing it there rather than here.
PATCH 3/3 igc: Fix SRRCTL register setup
Quoting Alexander Duyck (2020-08-10 15:56:31)
Quoted 6 lines
Quoted 6 lines
drivers/net/ethernet/intel/igc/igc_main.c +1
-0@@ -1869,6 +1866,7 @@ static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count)* because each write-back erases this info. */ rx_desc->read.pkt_addr =cpu_to_le64 (bi->dma + bi->page_offset); + rx_desc->read.hdr_addr =0 ;
Quoted 9 lines
drivers/net/ethernet/intel/igc/igc_main.c static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) · line 1872
Quoted 2 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-0rx_desc++; bi++;If you are going to do this it would be better to replace the line that is setting the length to zero instead of just adding this line. That way you can avoid having to rewrite it. I only had bothered with clearing the length field as it was a 32b field, however if you are wanting to flush the full 64b then I would recommend doing it there rather than here.
Just to make sure I'm on the same page, do you mean to move this line to patch 2/3, right?
PATCH 2/3 igc: Check descriptor's DD bit in igc_clean_rx_irq()
In reply to Andre GuedesQuoted 5 lines · Andre Guedes
I225 advanced receive descriptor provides the Descriptor Done (DD) bit which indicates hardware is done with that receive descriptor and software should handle it.
Quoted 7 lines
This patch fixes
igc_clean_rx_irq()so we check that bit to determine if we are done handling incoming packets instead of checking the packet length information. It also gets rid ofrx_desc->wb.upper.lengthassignments spread through the code required to make the previous approach to work.Signed-off-by Andre Guedes <[email protected]>
I do not agree with this patch. The patch itself will break a number of things.
Quoted 22 lines
--- drivers/net/ethernet/intel/igc/igc_defines.h | 1 + drivers/net/ethernet/intel/igc/igc_main.c | 14 ++++++++------ 2 files changed, 9 insertions(+), 6 deletions(-)drivers/net/ethernet/intel/igc/igc_defines.h +1
-0@@ -316,6 +316,7 @@#define IGC_SRRCTL_TIMER0SEL (timer) (((timer) & 0x3) << 17)/* Receive Descriptor bit definitions */ +#define IGC_RXD_STAT_DD 0x01/* Descriptor Done */ #define IGC_RXD_STAT_EOP 0x02/* End of Packet */ #define IGC_RXD_STAT_IXSM 0x04/* Ignore checksum */ #define IGC_RXD_STAT_UDPCS 0x10/* UDP xsum calculated */ drivers/net/ethernet/intel/igc/igc_main.c +0
-0
drivers/net/ethernet/intel/igc/igc_main.c static void igc_configure_rx_ring(struct igc_adapter *adapter, · line 556
Quoted 7 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-1@@ -551,7 +551,6 @@ static void igc_configure_rx_ring(struct igc_adapter *adapter,/* initialize Rx descriptor 0 */ rx_desc =IGC_RX_DESC (ring,0 ); - rx_desc->wb.upper.length =0 ;/* enable receive descriptor fetching */ rxdctl |=IGC_RXDCTL_QUEUE_ENABLE ;
This is initializing the ring for the first descriptor to 0. Without this line you break the driver. Without this you need to memset the entire descriptor ring to 0.
drivers/net/ethernet/intel/igc/igc_main.c static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) · line 1884
Quoted 10 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-3@@ -1880,9 +1879,6 @@ static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) i -= rx_ring->count; } -/* clear the length for the next_to_use descriptor */ - rx_desc->wb.upper.length =0 ; - cleaned_count--; }while (cleaned_count);
Same here. Without doing this you can potentially hang as you need to make sure the status bits are cleared before releasing a descriptor to the device.
Quoted 5 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-1@@ -1924,8 +1920,12 @@ static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget) } rx_desc =IGC_RX_DESC (rx_ring, rx_ring->next_to_clean); - size =le16_to_cpu (rx_desc->wb.upper.length);
drivers/net/ethernet/intel/igc/igc_main.c static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget) · line 1931
Quoted 10 lines
drivers/net/ethernet/intel/igc/igc_main.c +6
-1-if (!size) + +/* If we reached a descriptor with 'Descriptor Done' bit not +* set, it means we have handled all descriptors owned by +* software already so we should prematurely break the loop. +*/ +if (!igc_test_staterr (rx_desc,IGC_RXD_STAT_DD ))break ;/* This memory barrier is needed to keep us from reading
Why add an extra check when you don't need to? This doesn't make sense. The DD bit tells you that the descriptor was written back but you can do the same thing by reading the size field.
drivers/net/ethernet/intel/igc/igc_main.c static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget) · line 1941
Quoted 9 lines
drivers/net/ethernet/intel/igc/igc_main.c +2
-0@@ -1934,6 +1934,8 @@ static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget)*/ dma_rmb (); + size =le16_to_cpu (rx_desc->wb.upper.length); + rx_buffer =igc_get_rx_buffer (rx_ring, size);/* retrieve a buffer from the ring */
This should be fine since the dma_rmb() will prevent the reads from being reordered. However it is really redundant. For a bit of history the reason for reading the size as the first field is because previously we had bugs on PowerPC where the length field was being read first, and then the DD bit. As such if the length is 0 before the writeback, and non-zero after then you can spare yourself some cycles by reading the size first and if it is non-zero then you know the descriptor has valid data to be read.
PATCH 3/3 igc: Fix SRRCTL register setup
In reply to Andre GuedesQuoted 18 lines · Andre Guedes
Quoting Alexander Duyck (2020-08-10 15:56:31)
Quoted 6 lines
Quoted 6 lines
drivers/net/ethernet/intel/igc/igc_main.c +1
-0@@ -1869,6 +1866,7 @@ static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count)* because each write-back erases this info. */ rx_desc->read.pkt_addr =cpu_to_le64 (bi->dma + bi->page_offset); + rx_desc->read.hdr_addr =0 ;Quoted 9 lines
drivers/net/ethernet/intel/igc/igc_main.c static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) · line 1872
Quoted 2 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-0rx_desc++; bi++;If you are going to do this it would be better to replace the line that is setting the length to zero instead of just adding this line. That way you can avoid having to rewrite it. I only had bothered with clearing the length field as it was a 32b field, however if you are wanting to flush the full 64b then I would recommend doing it there rather than here.
Quoted 2 lines
Just to make sure I'm on the same page, do you mean to move this line to patch 2/3, right?
No, I hadn't had a chance to take a look at patch 2 yet. I think patch 2 is ill advised as the patch is currently broken, and even if done correctly you don't get any benefit out of it that I can see. From what I can tell this patch was likely covering up some of the errors introduced in patch 2. Now that I see this in conjunction with patch 2 I would say you should probably just drop patch 2 and this one as well since they aren't adding any real value other than removing a redundant write which was just there to keep this mostly in sync with how we did it for ixgbe.
The reason the driver path was coded the way it is in order to get away from having to clear the entire descriptor after processing it and to avoid having to explicitly track next_to_use and next_to_clean. Instead we leave the descriptor as mostly read-only until we reallocate the buffer and give it back to the device. All we have to do is clear the length field of the next_to_use descriptor when we are done allocating so that we do not overrun the descriptors when cleaning. It makes it much easier to debug when the descriptors are left in place as long as possible since we can then come back and look at the memory and generally I have found performance is improved as we are not having to dirty cachelines prematurely.
PATCH 2/3 igc: Check descriptor's DD bit in igc_clean_rx_irq()
In reply to Alexander DuyckHi Alexander,
Quoting Alexander Duyck (2020-08-10 18:25:12)
Quoted 19 lines
drivers/net/ethernet/intel/igc/igc_main.c static void igc_configure_rx_ring(struct igc_adapter *adapter, · line 556
Quoted 10 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-1@@ -551,7 +551,6 @@ static void igc_configure_rx_ring(struct igc_adapter *adapter,/* initialize Rx descriptor 0 */ rx_desc =IGC_RX_DESC (ring,0 ); - rx_desc->wb.upper.length =0 ;/* enable receive descriptor fetching */ rxdctl |=IGC_RXDCTL_QUEUE_ENABLE ;This is initializing the ring for the first descriptor to 0. Without this line you break the driver. Without this you need to memset the entire descriptor ring to 0.
Quoted 4 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-0@@ -1880,9 +1879,6 @@ static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) i -= rx_ring->count; }
Quoted 10 lines
drivers/net/ethernet/intel/igc/igc_main.c static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) · line 1884
Quoted 6 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-3-/* clear the length for the next_to_use descriptor */ - rx_desc->wb.upper.length =0 ; - cleaned_count--; }while (cleaned_count);Same here. Without doing this you can potentially hang as you need to make sure the status bits are cleared before releasing a descriptor to the device.
Yes, after your first comment on patch 3/3 I realized this was bogus. Moving the 'read.hdr_addr = 0' to this patch should fix it. That's why I thought you were suggesting that.
Quoted 14 lines
Quoted 14 lines
drivers/net/ethernet/intel/igc/igc_main.c +6
-2@@ -1924,8 +1920,12 @@ static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget) } rx_desc =IGC_RX_DESC (rx_ring, rx_ring->next_to_clean); - size =le16_to_cpu (rx_desc->wb.upper.length); -if (!size) + +/* If we reached a descriptor with 'Descriptor Done' bit not +* set, it means we have handled all descriptors owned by +* software already so we should prematurely break the loop. +*/ +if (!igc_test_staterr (rx_desc,IGC_RXD_STAT_DD ))break ;
Quoted 5 lines
drivers/net/ethernet/intel/igc/igc_main.c static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget) · line 1931
Quoted 1 line
drivers/net/ethernet/intel/igc/igc_main.c +0
-0/* This memory barrier is needed to keep us from reading Why add an extra check when you don't need to? This doesn't make sense. The DD bit tells you that the descriptor was written back but you can do the same thing by reading the size field.
I was using i40e and ice drivers as reference. If I'm reading it correctly, they check the DD bit first (or similar) and then read the length information. I don't see a strong reason besides making the code a bit more readable.
Quoted 8 lines
Quoted 8 lines
drivers/net/ethernet/intel/igc/igc_main.c +2
-0@@ -1934,6 +1934,8 @@ static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget)*/ dma_rmb (); + size =le16_to_cpu (rx_desc->wb.upper.length); + rx_buffer =igc_get_rx_buffer (rx_ring, size);
Quoted 10 lines
drivers/net/ethernet/intel/igc/igc_main.c static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget) · line 1941
Quoted 1 line
drivers/net/ethernet/intel/igc/igc_main.c +0
-0/* retrieve a buffer from the ring */ This should be fine since the
dma_rmb()will prevent the reads from being reordered. However it is really redundant. For a bit of history the reason for reading the size as the first field is because previously we had bugs on PowerPC where the length field was being read first, and then the DD bit. As such if the length is 0 before the writeback, and non-zero after then you can spare yourself some cycles by reading the size first and if it is non-zero then you know the descriptor has valid data to be read.
Thanks for the history context.
Cheers,
Andre
PATCH 3/3 igc: Fix SRRCTL register setup
In reply to Alexander DuyckQuoting Alexander Duyck (2020-08-10 18:41:41)
Quoted 22 lines · Andre Guedes
Quoted 18 lines
Quoting Alexander Duyck (2020-08-10 15:56:31)
Quoted 6 lines
Quoted 6 lines
drivers/net/ethernet/intel/igc/igc_main.c +1
-0@@ -1869,6 +1866,7 @@ static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count)* because each write-back erases this info. */ rx_desc->read.pkt_addr =cpu_to_le64 (bi->dma + bi->page_offset); + rx_desc->read.hdr_addr =0 ;Quoted 9 lines
drivers/net/ethernet/intel/igc/igc_main.c static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) · line 1872
Quoted 2 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-0rx_desc++; bi++;If you are going to do this it would be better to replace the line that is setting the length to zero instead of just adding this line. That way you can avoid having to rewrite it. I only had bothered with clearing the length field as it was a 32b field, however if you are wanting to flush the full 64b then I would recommend doing it there rather than here.
Quoted 2 lines
Just to make sure I'm on the same page, do you mean to move this line to patch 2/3, right?
Quoted 9 lines
No, I hadn't had a chance to take a look at patch 2 yet. I think patch 2 is ill advised as the patch is currently broken, and even if done correctly you don't get any benefit out of it that I can see. From what I can tell this patch was likely covering up some of the errors introduced in patch 2. Now that I see this in conjunction with patch 2 I would say you should probably just drop patch 2 and this one as well since they aren't adding any real value other than removing a redundant write which was just there to keep this mostly in sync with how we did it for ixgbe.
What about not setting BSIZEHEADER bits since 'one buffer descriptor' option is used in SRRCTL?
Quoted 11 lines
The reason the driver path was coded the way it is in order to get away from having to clear the entire descriptor after processing it and to avoid having to explicitly track
next_to_useandnext_to_clean. Instead we leave the descriptor as mostly read-only until we reallocate the buffer and give it back to the device. All we have to do is clear the length field of thenext_to_usedescriptor when we are done allocating so that we do not overrun the descriptors when cleaning. It makes it much easier to debug when the descriptors are left in place as long as possible since we can then come back and look at the memory and generally I have found performance is improved as we are not having to dirty cachelines prematurely.
Thanks for the context.
- Andre
PATCH 2/3 igc: Check descriptor's DD bit in igc_clean_rx_irq()
In reply to Andre GuedesQuoted 58 lines · Andre Guedes
Hi Alexander,
Quoting Alexander Duyck (2020-08-10 18:25:12)
Quoted 19 lines
drivers/net/ethernet/intel/igc/igc_main.c static void igc_configure_rx_ring(struct igc_adapter *adapter, · line 556
Quoted 10 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-1@@ -551,7 +551,6 @@ static void igc_configure_rx_ring(struct igc_adapter *adapter,/* initialize Rx descriptor 0 */ rx_desc =IGC_RX_DESC (ring,0 ); - rx_desc->wb.upper.length =0 ;/* enable receive descriptor fetching */ rxdctl |=IGC_RXDCTL_QUEUE_ENABLE ;This is initializing the ring for the first descriptor to 0. Without this line you break the driver. Without this you need to memset the entire descriptor ring to 0.
Quoted 4 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-0@@ -1880,9 +1879,6 @@ static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) i -= rx_ring->count; }Quoted 10 lines
drivers/net/ethernet/intel/igc/igc_main.c static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) · line 1884
Quoted 6 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-3-/* clear the length for the next_to_use descriptor */ - rx_desc->wb.upper.length =0 ; - cleaned_count--; }while (cleaned_count);Same here. Without doing this you can potentially hang as you need to make sure the status bits are cleared before releasing a descriptor to the device.
Yes, after your first comment on patch 3/3 I realized this was bogus. Moving the 'read.hdr_addr = 0' to this patch should fix it. That's why I thought you were suggesting that.
Quoted 14 lines
Quoted 14 lines
drivers/net/ethernet/intel/igc/igc_main.c +6
-2@@ -1924,8 +1920,12 @@ static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget) } rx_desc =IGC_RX_DESC (rx_ring, rx_ring->next_to_clean); - size =le16_to_cpu (rx_desc->wb.upper.length); -if (!size) + +/* If we reached a descriptor with 'Descriptor Done' bit not +* set, it means we have handled all descriptors owned by +* software already so we should prematurely break the loop. +*/ +if (!igc_test_staterr (rx_desc,IGC_RXD_STAT_DD ))break ;Quoted 5 lines
drivers/net/ethernet/intel/igc/igc_main.c static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget) · line 1931
Quoted 1 line
drivers/net/ethernet/intel/igc/igc_main.c +0
-0/* This memory barrier is needed to keep us from reading Why add an extra check when you don't need to? This doesn't make sense. The DD bit tells you that the descriptor was written back but you can do the same thing by reading the size field.
Quoted 3 lines
I was using i40e and ice drivers as reference. If I'm reading it correctly, they check the DD bit first (or similar) and then read the length information. I don't see a strong reason besides making the code a bit more readable.
The i40e driver does the same thing as igb and ixgbe. The ice driver doesn't but in the case of that driver the pkt_len field resides in the first qword which is the same as the pkt_addr so they overlap and the same approach cannot be used.
Quoted 19 lines
Quoted 8 lines
Quoted 8 lines
drivers/net/ethernet/intel/igc/igc_main.c +2
-0@@ -1934,6 +1934,8 @@ static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget)*/ dma_rmb (); + size =le16_to_cpu (rx_desc->wb.upper.length); + rx_buffer =igc_get_rx_buffer (rx_ring, size);Quoted 10 lines
drivers/net/ethernet/intel/igc/igc_main.c static int igc_clean_rx_irq(struct igc_q_vector *q_vector, const int budget) · line 1941
Quoted 1 line
drivers/net/ethernet/intel/igc/igc_main.c +0
-0/* retrieve a buffer from the ring */ This should be fine since the
dma_rmb()will prevent the reads from being reordered. However it is really redundant. For a bit of history the reason for reading the size as the first field is because previously we had bugs on PowerPC where the length field was being read first, and then the DD bit. As such if the length is 0 before the writeback, and non-zero after then you can spare yourself some cycles by reading the size first and if it is non-zero then you know the descriptor has valid data to be read.
Quoted 1 line
Thanks for the history context.
No problem, just glad I saw this before it went too far down the path of undoing the work that was done in the past.
Thanks.
- Alex
PATCH 3/3 igc: Fix SRRCTL register setup
In reply to Andre GuedesQuoted 34 lines · Andre Guedes
Quoting Alexander Duyck (2020-08-10 18:41:41)
Quoted 22 lines · Andre Guedes
Quoted 18 lines
Quoting Alexander Duyck (2020-08-10 15:56:31)
Quoted 6 lines
Quoted 6 lines
drivers/net/ethernet/intel/igc/igc_main.c +1
-0@@ -1869,6 +1866,7 @@ static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count)* because each write-back erases this info. */ rx_desc->read.pkt_addr =cpu_to_le64 (bi->dma + bi->page_offset); + rx_desc->read.hdr_addr =0 ;Quoted 9 lines
drivers/net/ethernet/intel/igc/igc_main.c static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) · line 1872
Quoted 2 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-0rx_desc++; bi++;If you are going to do this it would be better to replace the line that is setting the length to zero instead of just adding this line. That way you can avoid having to rewrite it. I only had bothered with clearing the length field as it was a 32b field, however if you are wanting to flush the full 64b then I would recommend doing it there rather than here.
Quoted 2 lines
Just to make sure I'm on the same page, do you mean to move this line to patch 2/3, right?
Quoted 9 lines
No, I hadn't had a chance to take a look at patch 2 yet. I think patch 2 is ill advised as the patch is currently broken, and even if done correctly you don't get any benefit out of it that I can see. From what I can tell this patch was likely covering up some of the errors introduced in patch 2. Now that I see this in conjunction with patch 2 I would say you should probably just drop patch 2 and this one as well since they aren't adding any real value other than removing a redundant write which was just there to keep this mostly in sync with how we did it for ixgbe.
Quoted 2 lines
What about not setting BSIZEHEADER bits since 'one buffer descriptor' option is used in SRRCTL?
Does it cause some sort of problem to be populating it? If not I would say leave it. It isn't too different from just populating the field with 0 and should have no effect since the field is unused when in one buffer mode.
Quoted 15 lines
Quoted 11 lines
The reason the driver path was coded the way it is in order to get away from having to clear the entire descriptor after processing it and to avoid having to explicitly track
next_to_useandnext_to_clean. Instead we leave the descriptor as mostly read-only until we reallocate the buffer and give it back to the device. All we have to do is clear the length field of thenext_to_usedescriptor when we are done allocating so that we do not overrun the descriptors when cleaning. It makes it much easier to debug when the descriptors are left in place as long as possible since we can then come back and look at the memory and generally I have found performance is improved as we are not having to dirty cachelines prematurely.Thanks for the context.
- Andre
PATCH 3/3 igc: Fix SRRCTL register setup
Quoting Alexander Duyck (2020-08-11 11:04:52)
Quoted 35 lines · Andre Guedes
Quoted 34 lines
Quoting Alexander Duyck (2020-08-10 18:41:41)
Quoted 31 lines · Andre Guedes
Quoted 18 lines
Quoting Alexander Duyck (2020-08-10 15:56:31)
Quoted 6 lines
Quoted 6 lines
drivers/net/ethernet/intel/igc/igc_main.c +1
-0@@ -1869,6 +1866,7 @@ static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count)* because each write-back erases this info. */ rx_desc->read.pkt_addr =cpu_to_le64 (bi->dma + bi->page_offset); + rx_desc->read.hdr_addr =0 ;Quoted 9 lines
drivers/net/ethernet/intel/igc/igc_main.c static void igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count) · line 1872
Quoted 2 lines
drivers/net/ethernet/intel/igc/igc_main.c +0
-0rx_desc++; bi++;If you are going to do this it would be better to replace the line that is setting the length to zero instead of just adding this line. That way you can avoid having to rewrite it. I only had bothered with clearing the length field as it was a 32b field, however if you are wanting to flush the full 64b then I would recommend doing it there rather than here.
Quoted 2 lines
Just to make sure I'm on the same page, do you mean to move this line to patch 2/3, right?
No, I hadn't had a chance to take a look at patch 2 yet. I think patch 2 is ill advised as the patch is currently broken, and even if done correctly you don't get any benefit out of it that I can see. From what I can tell this patch was likely covering up some of the errors introduced in patch 2. Now that I see this in conjunction with patch 2 I would say you should probably just drop patch 2 and this one as well since they aren't adding any real value other than removing a redundant write which was just there to keep this mostly in sync with how we did it for ixgbe.
Quoted 7 lines
Quoted 2 lines
What about not setting BSIZEHEADER bits since 'one buffer descriptor' option is used in SRRCTL?
Does it cause some sort of problem to be populating it? If not I would say leave it. It isn't too different from just populating the field with 0 and should have no effect since the field is unused when in one buffer mode.
No problem that I'm aware of. When I first came across that code I found it a bit misleading since no header buffer is actually allocated by the driver. That was the only motivation for this patch.
So Jeff/Tony please disregard patches 2 and 3 from this series. Let's move forward with patch 1 only.
Thanks for the review, Alexander.
- Andre
Nguyen
PATCH 3/3 igc: Fix SRRCTL register setup
Quoted 72 lines · Andre Guedes
Quoting Alexander Duyck (2020-08-11 11:04:52)
Quoted 58 lines
Quoted 52 lines · Andre Guedes
Quoting Alexander Duyck (2020-08-10 18:41:41)
Quoted 33 lines
Quoted 27 lines · Andre Guedes
Quoting Alexander Duyck (2020-08-10 15:56:31)
Quoted 13 lines
Quoted 5 lines
(quoted diff) +0
-0@@ -1869,6 +1866,7 @@ static void
igc_alloc_rx_buffers(struct igc_ring *rx_ring, u16 cleaned_count)
* because each write-back erases this
info.Quoted 7 lines
*/ rx_desc->read.pkt_addr = cpu_to_le64(bi-Quoted 1 line
dma +
bi->page_offset);+
rx_desc->read.hdr_addr= 0;rx_desc++; bi++;Quoted 11 lines
If you are going to do this it would be better to replace the line
that is setting the length to zero instead of just adding this line.
That way you can avoid having to rewrite it. I only had bothered with
clearing the length field as it was a 32b field, however if you are
wanting to flush the full 64b then I would recommend doing it there
rather than here.Quoted 3 lines
Just to make sure I'm on the same page, do you mean to move this line to
patch 2/3, right?Quoted 16 lines
No, I hadn't had a chance to take a look at patch 2 yet. I think patch
2 is ill advised as the patch is currently broken, and even if done
correctly you don't get any benefit out of it that I can see. From
what I can tell this patch was likely covering up some of the errors
introduced in patch 2. Now that I see this in conjunction with patch 2
I would say you should probably just drop patch 2 and this one as well
since they aren't adding any real value other than removing a redundant write which was just there to keep this mostly in sync with
how we did it for ixgbe.Quoted 3 lines
What about not setting BSIZEHEADER bits since 'one buffer descriptor' option
is used in SRRCTL?Quoted 6 lines
Does it cause some sort of problem to be populating it? If not I would
say leave it. It isn't too different from just populating the field with 0 and should have no effect since the field is unused when in one
buffer mode.No problem that I'm aware of. When I first came across that code I found it a
bit misleading since no header buffer is actually allocated by the driver. That
was the only motivation for this patch.
Quoted 3 lines
So Jeff/Tony please disregard patches 2 and 3 from this series. Let's move forward with
patch 1 only.
I will apply patch 1 and drop the other 2.
Thanks,
Tony
Brown
PATCH 1/3 igc: Clean RX descriptor error flags
In reply to Andre GuedesQuoted 14 lines
From: Intel-wired-lan <[email protected]> On Behalf Of Andre Guedes
Sent: Monday, August 10, 2020 2:09 PM
To: intel-wired-lan at lists.osuosl.org
Subject: [Intel-wired-lan] [PATCH 1/3] igc: Clean RX descriptor error flagsI225 advanced receive descriptor doesn't have the following extend error
bits: CE, SE, SEQ, CXE. In addition to that, the bit TCPE is called L4E in the datasheet.This patch cleans up the code accordingly, and gets rid of the macro
IGC_RXDEXT_ERR_FRAME_ERR_MASKsince it doesn't make much sense anymore.
Quoted 6 lines
Signed-off-by Andre Guedes <[email protected]>--- drivers/net/ethernet/intel/igc/igc_defines.h | 14 +------------- drivers/net/ethernet/intel/igc/igc_main.c | 5 ++--- 2 files changed, 3 insertions(+), 16 deletions(-)