spider_net.rst (10026B)
1.. SPDX-License-Identifier: GPL-2.0 2 3=========================== 4The Spidernet Device Driver 5=========================== 6 7Written by Linas Vepstas <linas@austin.ibm.com> 8 9Version of 7 June 2007 10 11Abstract 12======== 13This document sketches the structure of portions of the spidernet 14device driver in the Linux kernel tree. The spidernet is a gigabit 15ethernet device built into the Toshiba southbridge commonly used 16in the SONY Playstation 3 and the IBM QS20 Cell blade. 17 18The Structure of the RX Ring. 19============================= 20The receive (RX) ring is a circular linked list of RX descriptors, 21together with three pointers into the ring that are used to manage its 22contents. 23 24The elements of the ring are called "descriptors" or "descrs"; they 25describe the received data. This includes a pointer to a buffer 26containing the received data, the buffer size, and various status bits. 27 28There are three primary states that a descriptor can be in: "empty", 29"full" and "not-in-use". An "empty" or "ready" descriptor is ready 30to receive data from the hardware. A "full" descriptor has data in it, 31and is waiting to be emptied and processed by the OS. A "not-in-use" 32descriptor is neither empty or full; it is simply not ready. It may 33not even have a data buffer in it, or is otherwise unusable. 34 35During normal operation, on device startup, the OS (specifically, the 36spidernet device driver) allocates a set of RX descriptors and RX 37buffers. These are all marked "empty", ready to receive data. This 38ring is handed off to the hardware, which sequentially fills in the 39buffers, and marks them "full". The OS follows up, taking the full 40buffers, processing them, and re-marking them empty. 41 42This filling and emptying is managed by three pointers, the "head" 43and "tail" pointers, managed by the OS, and a hardware current 44descriptor pointer (GDACTDPA). The GDACTDPA points at the descr 45currently being filled. When this descr is filled, the hardware 46marks it full, and advances the GDACTDPA by one. Thus, when there is 47flowing RX traffic, every descr behind it should be marked "full", 48and everything in front of it should be "empty". If the hardware 49discovers that the current descr is not empty, it will signal an 50interrupt, and halt processing. 51 52The tail pointer tails or trails the hardware pointer. When the 53hardware is ahead, the tail pointer will be pointing at a "full" 54descr. The OS will process this descr, and then mark it "not-in-use", 55and advance the tail pointer. Thus, when there is flowing RX traffic, 56all of the descrs in front of the tail pointer should be "full", and 57all of those behind it should be "not-in-use". When RX traffic is not 58flowing, then the tail pointer can catch up to the hardware pointer. 59The OS will then note that the current tail is "empty", and halt 60processing. 61 62The head pointer (somewhat mis-named) follows after the tail pointer. 63When traffic is flowing, then the head pointer will be pointing at 64a "not-in-use" descr. The OS will perform various housekeeping duties 65on this descr. This includes allocating a new data buffer and 66dma-mapping it so as to make it visible to the hardware. The OS will 67then mark the descr as "empty", ready to receive data. Thus, when there 68is flowing RX traffic, everything in front of the head pointer should 69be "not-in-use", and everything behind it should be "empty". If no 70RX traffic is flowing, then the head pointer can catch up to the tail 71pointer, at which point the OS will notice that the head descr is 72"empty", and it will halt processing. 73 74Thus, in an idle system, the GDACTDPA, tail and head pointers will 75all be pointing at the same descr, which should be "empty". All of the 76other descrs in the ring should be "empty" as well. 77 78The show_rx_chain() routine will print out the locations of the 79GDACTDPA, tail and head pointers. It will also summarize the contents 80of the ring, starting at the tail pointer, and listing the status 81of the descrs that follow. 82 83A typical example of the output, for a nearly idle system, might be:: 84 85 net eth1: Total number of descrs=256 86 net eth1: Chain tail located at descr=20 87 net eth1: Chain head is at 20 88 net eth1: HW curr desc (GDACTDPA) is at 21 89 net eth1: Have 1 descrs with stat=x40800101 90 net eth1: HW next desc (GDACNEXTDA) is at 22 91 net eth1: Last 255 descrs with stat=xa0800000 92 93In the above, the hardware has filled in one descr, number 20. Both 94head and tail are pointing at 20, because it has not yet been emptied. 95Meanwhile, hw is pointing at 21, which is free. 96 97The "Have nnn decrs" refers to the descr starting at the tail: in this 98case, nnn=1 descr, starting at descr 20. The "Last nnn descrs" refers 99to all of the rest of the descrs, from the last status change. The "nnn" 100is a count of how many descrs have exactly the same status. 101 102The status x4... corresponds to "full" and status xa... corresponds 103to "empty". The actual value printed is RXCOMST_A. 104 105In the device driver source code, a different set of names are 106used for these same concepts, so that:: 107 108 "empty" == SPIDER_NET_DESCR_CARDOWNED == 0xa 109 "full" == SPIDER_NET_DESCR_FRAME_END == 0x4 110 "not in use" == SPIDER_NET_DESCR_NOT_IN_USE == 0xf 111 112 113The RX RAM full bug/feature 114=========================== 115 116As long as the OS can empty out the RX buffers at a rate faster than 117the hardware can fill them, there is no problem. If, for some reason, 118the OS fails to empty the RX ring fast enough, the hardware GDACTDPA 119pointer will catch up to the head, notice the not-empty condition, 120ad stop. However, RX packets may still continue arriving on the wire. 121The spidernet chip can save some limited number of these in local RAM. 122When this local ram fills up, the spider chip will issue an interrupt 123indicating this (GHIINT0STS will show ERRINT, and the GRMFLLINT bit 124will be set in GHIINT1STS). When the RX ram full condition occurs, 125a certain bug/feature is triggered that has to be specially handled. 126This section describes the special handling for this condition. 127 128When the OS finally has a chance to run, it will empty out the RX ring. 129In particular, it will clear the descriptor on which the hardware had 130stopped. However, once the hardware has decided that a certain 131descriptor is invalid, it will not restart at that descriptor; instead 132it will restart at the next descr. This potentially will lead to a 133deadlock condition, as the tail pointer will be pointing at this descr, 134which, from the OS point of view, is empty; the OS will be waiting for 135this descr to be filled. However, the hardware has skipped this descr, 136and is filling the next descrs. Since the OS doesn't see this, there 137is a potential deadlock, with the OS waiting for one descr to fill, 138while the hardware is waiting for a different set of descrs to become 139empty. 140 141A call to show_rx_chain() at this point indicates the nature of the 142problem. A typical print when the network is hung shows the following:: 143 144 net eth1: Spider RX RAM full, incoming packets might be discarded! 145 net eth1: Total number of descrs=256 146 net eth1: Chain tail located at descr=255 147 net eth1: Chain head is at 255 148 net eth1: HW curr desc (GDACTDPA) is at 0 149 net eth1: Have 1 descrs with stat=xa0800000 150 net eth1: HW next desc (GDACNEXTDA) is at 1 151 net eth1: Have 127 descrs with stat=x40800101 152 net eth1: Have 1 descrs with stat=x40800001 153 net eth1: Have 126 descrs with stat=x40800101 154 net eth1: Last 1 descrs with stat=xa0800000 155 156Both the tail and head pointers are pointing at descr 255, which is 157marked xa... which is "empty". Thus, from the OS point of view, there 158is nothing to be done. In particular, there is the implicit assumption 159that everything in front of the "empty" descr must surely also be empty, 160as explained in the last section. The OS is waiting for descr 255 to 161become non-empty, which, in this case, will never happen. 162 163The HW pointer is at descr 0. This descr is marked 0x4.. or "full". 164Since its already full, the hardware can do nothing more, and thus has 165halted processing. Notice that descrs 0 through 254 are all marked 166"full", while descr 254 and 255 are empty. (The "Last 1 descrs" is 167descr 254, since tail was at 255.) Thus, the system is deadlocked, 168and there can be no forward progress; the OS thinks there's nothing 169to do, and the hardware has nowhere to put incoming data. 170 171This bug/feature is worked around with the spider_net_resync_head_ptr() 172routine. When the driver receives RX interrupts, but an examination 173of the RX chain seems to show it is empty, then it is probable that 174the hardware has skipped a descr or two (sometimes dozens under heavy 175network conditions). The spider_net_resync_head_ptr() subroutine will 176search the ring for the next full descr, and the driver will resume 177operations there. Since this will leave "holes" in the ring, there 178is also a spider_net_resync_tail_ptr() that will skip over such holes. 179 180As of this writing, the spider_net_resync() strategy seems to work very 181well, even under heavy network loads. 182 183 184The TX ring 185=========== 186The TX ring uses a low-watermark interrupt scheme to make sure that 187the TX queue is appropriately serviced for large packet sizes. 188 189For packet sizes greater than about 1KBytes, the kernel can fill 190the TX ring quicker than the device can drain it. Once the ring 191is full, the netdev is stopped. When there is room in the ring, 192the netdev needs to be reawakened, so that more TX packets are placed 193in the ring. The hardware can empty the ring about four times per jiffy, 194so its not appropriate to wait for the poll routine to refill, since 195the poll routine runs only once per jiffy. The low-watermark mechanism 196marks a descr about 1/4th of the way from the bottom of the queue, so 197that an interrupt is generated when the descr is processed. This 198interrupt wakes up the netdev, which can then refill the queue. 199For large packets, this mechanism generates a relatively small number 200of interrupts, about 1K/sec. For smaller packets, this will drop to zero 201interrupts, as the hardware can empty the queue faster than the kernel 202can fill it.