Based on kernel version 3.16. Page generated on 2014-08-06 21:40 EST.
1 2 The Spidernet Device Driver 3 =========================== 4 5 Written by Linas Vepstas <firstname.lastname@example.org> 6 7 Version of 7 June 2007 8 9 Abstract 10 ======== 11 This document sketches the structure of portions of the spidernet 12 device driver in the Linux kernel tree. The spidernet is a gigabit 13 ethernet device built into the Toshiba southbridge commonly used 14 in the SONY Playstation 3 and the IBM QS20 Cell blade. 15 16 The Structure of the RX Ring. 17 ============================= 18 The receive (RX) ring is a circular linked list of RX descriptors, 19 together with three pointers into the ring that are used to manage its 20 contents. 21 22 The elements of the ring are called "descriptors" or "descrs"; they 23 describe the received data. This includes a pointer to a buffer 24 containing the received data, the buffer size, and various status bits. 25 26 There are three primary states that a descriptor can be in: "empty", 27 "full" and "not-in-use". An "empty" or "ready" descriptor is ready 28 to receive data from the hardware. A "full" descriptor has data in it, 29 and is waiting to be emptied and processed by the OS. A "not-in-use" 30 descriptor is neither empty or full; it is simply not ready. It may 31 not even have a data buffer in it, or is otherwise unusable. 32 33 During normal operation, on device startup, the OS (specifically, the 34 spidernet device driver) allocates a set of RX descriptors and RX 35 buffers. These are all marked "empty", ready to receive data. This 36 ring is handed off to the hardware, which sequentially fills in the 37 buffers, and marks them "full". The OS follows up, taking the full 38 buffers, processing them, and re-marking them empty. 39 40 This filling and emptying is managed by three pointers, the "head" 41 and "tail" pointers, managed by the OS, and a hardware current 42 descriptor pointer (GDACTDPA). The GDACTDPA points at the descr 43 currently being filled. When this descr is filled, the hardware 44 marks it full, and advances the GDACTDPA by one. Thus, when there is 45 flowing RX traffic, every descr behind it should be marked "full", 46 and everything in front of it should be "empty". If the hardware 47 discovers that the current descr is not empty, it will signal an 48 interrupt, and halt processing. 49 50 The tail pointer tails or trails the hardware pointer. When the 51 hardware is ahead, the tail pointer will be pointing at a "full" 52 descr. The OS will process this descr, and then mark it "not-in-use", 53 and advance the tail pointer. Thus, when there is flowing RX traffic, 54 all of the descrs in front of the tail pointer should be "full", and 55 all of those behind it should be "not-in-use". When RX traffic is not 56 flowing, then the tail pointer can catch up to the hardware pointer. 57 The OS will then note that the current tail is "empty", and halt 58 processing. 59 60 The head pointer (somewhat mis-named) follows after the tail pointer. 61 When traffic is flowing, then the head pointer will be pointing at 62 a "not-in-use" descr. The OS will perform various housekeeping duties 63 on this descr. This includes allocating a new data buffer and 64 dma-mapping it so as to make it visible to the hardware. The OS will 65 then mark the descr as "empty", ready to receive data. Thus, when there 66 is flowing RX traffic, everything in front of the head pointer should 67 be "not-in-use", and everything behind it should be "empty". If no 68 RX traffic is flowing, then the head pointer can catch up to the tail 69 pointer, at which point the OS will notice that the head descr is 70 "empty", and it will halt processing. 71 72 Thus, in an idle system, the GDACTDPA, tail and head pointers will 73 all be pointing at the same descr, which should be "empty". All of the 74 other descrs in the ring should be "empty" as well. 75 76 The show_rx_chain() routine will print out the locations of the 77 GDACTDPA, tail and head pointers. It will also summarize the contents 78 of the ring, starting at the tail pointer, and listing the status 79 of the descrs that follow. 80 81 A typical example of the output, for a nearly idle system, might be 82 83 net eth1: Total number of descrs=256 84 net eth1: Chain tail located at descr=20 85 net eth1: Chain head is at 20 86 net eth1: HW curr desc (GDACTDPA) is at 21 87 net eth1: Have 1 descrs with stat=x40800101 88 net eth1: HW next desc (GDACNEXTDA) is at 22 89 net eth1: Last 255 descrs with stat=xa0800000 90 91 In the above, the hardware has filled in one descr, number 20. Both 92 head and tail are pointing at 20, because it has not yet been emptied. 93 Meanwhile, hw is pointing at 21, which is free. 94 95 The "Have nnn decrs" refers to the descr starting at the tail: in this 96 case, nnn=1 descr, starting at descr 20. The "Last nnn descrs" refers 97 to all of the rest of the descrs, from the last status change. The "nnn" 98 is a count of how many descrs have exactly the same status. 99 100 The status x4... corresponds to "full" and status xa... corresponds 101 to "empty". The actual value printed is RXCOMST_A. 102 103 In the device driver source code, a different set of names are 104 used for these same concepts, so that 105 106 "empty" == SPIDER_NET_DESCR_CARDOWNED == 0xa 107 "full" == SPIDER_NET_DESCR_FRAME_END == 0x4 108 "not in use" == SPIDER_NET_DESCR_NOT_IN_USE == 0xf 109 110 111 The RX RAM full bug/feature 112 =========================== 113 114 As long as the OS can empty out the RX buffers at a rate faster than 115 the hardware can fill them, there is no problem. If, for some reason, 116 the OS fails to empty the RX ring fast enough, the hardware GDACTDPA 117 pointer will catch up to the head, notice the not-empty condition, 118 ad stop. However, RX packets may still continue arriving on the wire. 119 The spidernet chip can save some limited number of these in local RAM. 120 When this local ram fills up, the spider chip will issue an interrupt 121 indicating this (GHIINT0STS will show ERRINT, and the GRMFLLINT bit 122 will be set in GHIINT1STS). When the RX ram full condition occurs, 123 a certain bug/feature is triggered that has to be specially handled. 124 This section describes the special handling for this condition. 125 126 When the OS finally has a chance to run, it will empty out the RX ring. 127 In particular, it will clear the descriptor on which the hardware had 128 stopped. However, once the hardware has decided that a certain 129 descriptor is invalid, it will not restart at that descriptor; instead 130 it will restart at the next descr. This potentially will lead to a 131 deadlock condition, as the tail pointer will be pointing at this descr, 132 which, from the OS point of view, is empty; the OS will be waiting for 133 this descr to be filled. However, the hardware has skipped this descr, 134 and is filling the next descrs. Since the OS doesn't see this, there 135 is a potential deadlock, with the OS waiting for one descr to fill, 136 while the hardware is waiting for a different set of descrs to become 137 empty. 138 139 A call to show_rx_chain() at this point indicates the nature of the 140 problem. A typical print when the network is hung shows the following: 141 142 net eth1: Spider RX RAM full, incoming packets might be discarded! 143 net eth1: Total number of descrs=256 144 net eth1: Chain tail located at descr=255 145 net eth1: Chain head is at 255 146 net eth1: HW curr desc (GDACTDPA) is at 0 147 net eth1: Have 1 descrs with stat=xa0800000 148 net eth1: HW next desc (GDACNEXTDA) is at 1 149 net eth1: Have 127 descrs with stat=x40800101 150 net eth1: Have 1 descrs with stat=x40800001 151 net eth1: Have 126 descrs with stat=x40800101 152 net eth1: Last 1 descrs with stat=xa0800000 153 154 Both the tail and head pointers are pointing at descr 255, which is 155 marked xa... which is "empty". Thus, from the OS point of view, there 156 is nothing to be done. In particular, there is the implicit assumption 157 that everything in front of the "empty" descr must surely also be empty, 158 as explained in the last section. The OS is waiting for descr 255 to 159 become non-empty, which, in this case, will never happen. 160 161 The HW pointer is at descr 0. This descr is marked 0x4.. or "full". 162 Since its already full, the hardware can do nothing more, and thus has 163 halted processing. Notice that descrs 0 through 254 are all marked 164 "full", while descr 254 and 255 are empty. (The "Last 1 descrs" is 165 descr 254, since tail was at 255.) Thus, the system is deadlocked, 166 and there can be no forward progress; the OS thinks there's nothing 167 to do, and the hardware has nowhere to put incoming data. 168 169 This bug/feature is worked around with the spider_net_resync_head_ptr() 170 routine. When the driver receives RX interrupts, but an examination 171 of the RX chain seems to show it is empty, then it is probable that 172 the hardware has skipped a descr or two (sometimes dozens under heavy 173 network conditions). The spider_net_resync_head_ptr() subroutine will 174 search the ring for the next full descr, and the driver will resume 175 operations there. Since this will leave "holes" in the ring, there 176 is also a spider_net_resync_tail_ptr() that will skip over such holes. 177 178 As of this writing, the spider_net_resync() strategy seems to work very 179 well, even under heavy network loads. 180 181 182 The TX ring 183 =========== 184 The TX ring uses a low-watermark interrupt scheme to make sure that 185 the TX queue is appropriately serviced for large packet sizes. 186 187 For packet sizes greater than about 1KBytes, the kernel can fill 188 the TX ring quicker than the device can drain it. Once the ring 189 is full, the netdev is stopped. When there is room in the ring, 190 the netdev needs to be reawakened, so that more TX packets are placed 191 in the ring. The hardware can empty the ring about four times per jiffy, 192 so its not appropriate to wait for the poll routine to refill, since 193 the poll routine runs only once per jiffy. The low-watermark mechanism 194 marks a descr about 1/4th of the way from the bottom of the queue, so 195 that an interrupt is generated when the descr is processed. This 196 interrupt wakes up the netdev, which can then refill the queue. 197 For large packets, this mechanism generates a relatively small number 198 of interrupts, about 1K/sec. For smaller packets, this will drop to zero 199 interrupts, as the hardware can empty the queue faster than the kernel 200 can fill it. 201 202 203 ======= END OF DOCUMENT ========