Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | |
| 2 | Below is what the bt878 data book says about the PCI bug compatibility |
| 3 | modes of the bt878 chip. |
| 4 | |
| 5 | The triton1 insmod option sets the EN_TBFX bit in the control register. |
| 6 | The vsfx insmod option does the same for EN_VSFX bit. If you have |
| 7 | stability problems you can try if one of these options makes your box |
| 8 | work solid. |
| 9 | |
| 10 | drivers/pci/quirks.c knows about these issues, this way these bits are |
| 11 | enabled automagically for known-buggy chipsets (look at the kernel |
| 12 | messages, bttv tells you). |
| 13 | |
| 14 | HTH, |
| 15 | |
| 16 | Gerd |
| 17 | |
| 18 | ---------------------------- cut here -------------------------- |
| 19 | |
| 20 | Normal PCI Mode |
| 21 | --------------- |
| 22 | |
| 23 | The PCI REQ signal is the logical-or of the incoming function requests. |
| 24 | The inter-nal GNT[0:1] signals are gated asynchronously with GNT and |
| 25 | demultiplexed by the audio request signal. Thus the arbiter defaults to |
| 26 | the video function at power-up and parks there during no requests for |
| 27 | bus access. This is desirable since the video will request the bus more |
| 28 | often. However, the audio will have highest bus access priority. Thus |
| 29 | the audio will have first access to the bus even when issuing a request |
| 30 | after the video request but before the PCI external arbiter has granted |
| 31 | access to the Bt879. Neither function can preempt the other once on the |
| 32 | bus. The duration to empty the entire video PCI FIFO onto the PCI bus is |
| 33 | very short compared to the bus access latency the audio PCI FIFO can |
| 34 | tolerate. |
| 35 | |
| 36 | |
| 37 | 430FX Compatibility Mode |
| 38 | ------------------------ |
| 39 | |
| 40 | When using the 430FX PCI, the following rules will ensure |
Mauro Carvalho Chehab | 48773e6 | 2006-03-25 09:21:43 -0300 | [diff] [blame^] | 41 | compatibility: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 42 | |
Mauro Carvalho Chehab | 48773e6 | 2006-03-25 09:21:43 -0300 | [diff] [blame^] | 43 | (1) Deassert REQ at the same time as asserting FRAME. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 44 | (2) Do not reassert REQ to request another bus transaction until after |
| 45 | finish-ing the previous transaction. |
| 46 | |
| 47 | Since the individual bus masters do not have direct control of REQ, a |
| 48 | simple logical-or of video and audio requests would violate the rules. |
| 49 | Thus, both the arbiter and the initiator contain 430FX compatibility |
| 50 | mode logic. To enable 430FX mode, set the EN_TBFX bit as indicated in |
| 51 | Device Control Register on page 104. |
| 52 | |
| 53 | When EN_TBFX is enabled, the arbiter ensures that the two compatibility |
| 54 | rules are satisfied. Before GNT is asserted by the PCI arbiter, this |
| 55 | internal arbiter may still logical-or the two requests. However, once |
| 56 | the GNT is issued, this arbiter must lock in its decision and now route |
| 57 | only the granted request to the REQ pin. The arbiter decision lock |
| 58 | happens regardless of the state of FRAME because it does not know when |
| 59 | FRAME will be asserted (typically - each initiator will assert FRAME on |
| 60 | the cycle following GNT). When FRAME is asserted, it is the initiator s |
| 61 | responsibility to remove its request at the same time. It is the |
| 62 | arbiters responsibility to allow this request to flow through to REQ and |
| 63 | not allow the other request to hold REQ asserted. The decision lock may |
| 64 | be removed at the end of the transaction: for example, when the bus is |
| 65 | idle (FRAME and IRDY). The arbiter decision may then continue |
| 66 | asynchronously until GNT is again asserted. |
| 67 | |
| 68 | |
| 69 | Interfacing with Non-PCI 2.1 Compliant Core Logic |
| 70 | ------------------------------------------------- |
| 71 | |
| 72 | A small percentage of core logic devices may start a bus transaction |
| 73 | during the same cycle that GNT is de-asserted. This is non PCI 2.1 |
| 74 | compliant. To ensure compatibility when using PCs with these PCI |
| 75 | controllers, the EN_VSFX bit must be enabled (refer to Device Control |
| 76 | Register on page 104). When in this mode, the arbiter does not pass GNT |
| 77 | to the internal functions unless REQ is asserted. This prevents a bus |
| 78 | transaction from starting the same cycle as GNT is de-asserted. This |
| 79 | also has the side effect of not being able to take advantage of bus |
| 80 | parking, thus lowering arbitration performance. The Bt879 drivers must |
| 81 | query for these non-compliant devices, and set the EN_VSFX bit only if |
| 82 | required. |
| 83 | |