Receiver

1
Receiver packet-splitting

• A look at how a driver can cause the 82573L NIC
  to separate a packets headers from its data

2
NIC can do packet-parsing

• Intels newest gigabit Ethernet controllers offer
  an enhancement to the extended Receive
  Descriptor, called packet-split format, which
  empowers the hardware to recognize the packet
  headers used with the most common network
  protocols and to automatically separate those
  headers from their accompanying packet data

3
Extended RX-Descriptors
CPU writes this, NIC reads it
NIC writes this, CPU reads it
Base-address (64-bits)
MRQ (multiple receive queues)
Packet- checksum
IP identification
reserved (0)
Extended status
Packet- length
VLAN tag
Extended errors
The device-driver initializes the
base-address field with the physical address
of a packet-buffer, and it initializes the
reserved field with a zero-value the
network hardware will later modify both fields
The network controller will
write-back the values for these fields
when it has transferred a received packets
data into the packet-buffer
4
Packet-Split RX-Descriptors
CPU writes this, NIC reads it
NIC writes this, CPU reads it
Base-address 0 (64-bits)
MRQ (multiple receive queues)
Packet- checksum
IP identification
Extended status
Packet- Length 0
VLAN tag
Extended errors
Base-address 1 (64-bits)
Base-address 2 (64-bits)
Packet Length 3
Packet Length 2
Packet Length 1
Header Length
S P
Base-address 3 (64-bits)
reserved
1
The network controller will write-back
values to these fields when it has transferred
a received packets data into those
packet-buffers
The device-driver initializes four
base-address fields (even-numbered addresses)
5
Same Extended Status/Errors
19 18 17 16 15 14 13 12
11 10 9 8 7 6 5 4
3 2 1 0
0
0
0
0
A C K
0
0
0
0
U D P V
I P I V
0
P I F
I P C S
T C P C S
U D P C S
V P
I X S M
E O P
D D
11 10 9 8 7
6 5 4 3 2
1 0
RXE
IPE
TCPE
0
0
SEQ
SE
CE
0
0
0
0
6
Syntax modifications for fetch
typedef struct unsigned long
long base_addr0 unsigned long
long base_addr1 unsigned long
long base_addr2 unsigned long
long base_addr3 RX_DESC_FETCH
7
Syntax modifications for store
typedef struct unsigned int mrq unsigned
short ip_identification unsigned
short packet_chksum unsigned
int desc_status20 unsigned
int desc_errors12 unsigned
short packet_length0 unsigned
short vlan_tag unsigned short header_length
unsigned short packet_length1 unsigned
short packet_length2 unsigned
short packet_length3 unsigned long
long reserved RX_DESC_STORE
8
Same syntax for the union
typedef union RX_DESC_FETCH rxf RX_DESC
_STORE rxs RX_DESCRIPTOR
9
NIC Registers involved
31
10
0
D T Y P
RCTL
Device Control register
31
15
0
Reserved (0)
E X T E N
RFCTL
Receive Filter Control register
31 24 23
16 15 8
7 0
LEN3 (1KB)
LEN0 (128B)
LEN1 (1KB)
LEN2 (1KB)
PSRCTL
Packet-Split Receive Control register
10
Each descriptor has four buffers
Packet-Split Rx-descriptor
base_addr0
base_addr1
base_addr2
base_addr3
buffer0
buffer1
buffer2
buffer3
Four buffers are allocated for receiving one
packet
11
Refresh for reuse

• As with the extended receive-descriptors, it is
  necessary for a device-driver to setup each
  packet-split receive-descriptor any time it is
  going to be reused, since prior
  buffer-addresses get overwritten during a
  packet-reception by the network controller
• So driver needs a formula for recalculating
  buffer-addresses, or use a shadow array

12
Kernel-memory layout
Sixteen Rx-descriptors (32-bytes each)
Sixty-four receive-buffers (1024-bytes each)
512 bytes
65536 bytes
KMEM_SIZE ( 66048 bytes)
kmem
13
Caveats

• Short packets are not always split
• Unrecognized packet-headers may not be separated
  from accompanying packet-data
• Demonstrating the packet-split capability will
  require us to devise a way to transmit packets
  which have the TCP/UDP and IP packet-headers that
  the NIC recognizes

14
Our pktsplit.c demo

• We created a minimal kernel-module for
  demonstrating the NICs packet-splitting
  capabilities

TIMEOUT for an in-class demonstration
15
In-class exercise

• Can you enhance our pktsplit.c module so that
  its Receive-Descriptor Queue will function
  automatically as a ring-buffer (as happens in our
  extended.c example)?
• Your best option for this is to install an ISR
  which will reinitialize some Rx-Descriptors (and
  advance the RDT index) each time an RXDMT0
  interrupt gets triggered