Solaris x86 DDI specific (Solaris x86 DDI).
A device driver uses the ddi_dmae_req structure to describe the parameters for a DMA channel. This structure contains all the information necessary to set up the channel, except for the DMA memory address and transfer count. The defaults, as specified below, support most standard devices. Other modes might be desirable for some devices, or to increase performance. The DMA engine request structure is passed to ddi_dmae_prog (9F).
The ddi_dmae_req structure contains several members, each of which controls some aspect of DMA engine operation. The structure members associated with supported DMA engine options are described here.
uchar_tder_command; /* Read / Write * /uchar_tder_bufprocess; /* Standard / Chain */ uchar_tder_path; /* 8 / 16 / 32 */ uchar_tder_cycles; /* Compat / Type A / Type B / Burst */ uchar_tder_trans; /* Single / Demand / Block */ ddi_dma_cookie_t*(*proc)(); /* address of nextcookie routine */ void*procparms; /* parameter for nextcookie call */
Specifies what DMA operation is to be performed. The value DMAE_CMD_WRITE signifies that data is to be transferred from memory to the I/O device. The value DMAE_CMD_READ signifies that data is to be transferred from the I/O device to memory. This field must be set by the driver before calling ddi_dmae_prog().
On some bus types, a driver can set der_bufprocess to the value DMAE_BUF_CHAIN to specify that multiple DMA cookies will be given to the DMA engine for a single I/O transfer. This action causes a scatter/gather operation. In this mode of operation, the driver calls ddi_dmae_prog() to give the DMA engine the DMA engine request structure and a pointer to the first cookie. The proc structure member must be set to the address of a driver nextcookie routine. This routine takes one argument, specified by the procparms structure member, and returns a pointer to a structure of type ddi_dma_cookie_t that specifies the next cookie for the I/O transfer. When the DMA engine is ready to receive an additional cookie, the bus nexus driver controlling that DMA engine calls the routine specified by the proc structure member to obtain the next cookie from the driver. The driver's nextcookie routine must then return the address of the next cookie (in static storage) to the bus nexus routine that called it. If there are no more segments in the current DMA window, then (*proc)() must return the NULL pointer.
A driver can specify the DMAE_BUF_CHAIN flag only if the particular bus architecture supports the use of multiple DMA cookies in a single I/O transfer. A bus DMA engine can support this feature either with a fixed-length scatter/gather list, or by an interrupt chaining feature. A driver must determine whether its parent bus nexus supports this feature by examining the scatter/gather list size returned in the dlim_sgllen member of the DMA limit structure returned by the driver's call to ddi_dmae_getlim(). (See ddi_dma_lim_x86(9S).) If the size of the scatter/gather list is 1, then no chaining is available. The driver must not specify the DMAE_BUF_CHAIN flag in the ddi_dmae_req structure it passes to ddi_dmae_prog(), and the driver need not provide a nextcookie routine.
If the size of the scatter/gather list is greater than 1, then DMA chaining is available, and the driver has two options. Under the first option, the driver chooses not to use the chaining feature. In this case (a) the driver must set the size of the scatter/gather list to 1 before passing it to the DMA setup routine, and (b) the driver must not set the DMAE_BUF_CHAIN flag.
Under the second option, the driver chooses to use the chaining feature, in which case, (a) it should leave the size of the scatter/gather list alone, and (b) it must set the DMAE_BUF_CHAIN flag in the ddi_dmae_req structure. Before calling ddi_dmae_prog(), the driver must prefetch cookies by repeatedly calling ddi_dma_nextseg(9F) and ddi_dma_segtocookie(9F) until either (1) the end of the DMA window is reached ( ddi_dma_nextseg(9F) returns NULL), or (2) the size of the scatter/gather list is reached, whichever occurs first. These cookies must be saved by the driver until they are requested by the nexus driver calling the driver's nextcookie routine. The driver's nextcookie routine must return the prefetched cookies in order, one cookie for each call to the nextcookie routine, until the list of prefetched cookies is exhausted. After the end of the list of cookies is reached, the nextcookie routine must return the NULL pointer.
The size of the scatter/gather list determines how many discontiguous segments of physical memory can participate in a single DMA transfer. ISA bus DMA engines have no scatter/gather capability, so their scatter/gather list sizes are 1. Other finite scatter/gather list sizes would also be possible. For performance reasons, drivers should use the chaining capability if it is available on their parent bus.
As described above, a driver making use of DMA chaining must prefetch DMA cookies before calling ddi_dmae_prog(). The reasons for this are:
First, the driver must have some way to know the total I/O count with which to program the I/O device. This I/O count must match the total size of all the DMA segments that will be chained together into one DMA operation. Depending on the size of the scatter/gather list and the memory position and alignment of the DMA object, all or just part of the current DMA window might be able to participate in a single I/O operation. The driver must compute the I/O count by adding up the sizes of the prefetched DMA cookies. The number of cookies whose sizes are to be summed is the lesser of (a) the size of the scatter/gather list, or (b) the number of segments remaining in the window.
Second, on some bus architectures, the driver's nextcookie routine can be called from a high-level interrupt routine. If the cookies were not prefetched, the nextcookie routine would have to call ddi_dma_nextseg() and ddi_dma_segtocookie() from a high-level interrupt routine, which is not recommended.
When breaking a DMA window into segments, the system arranges for the end of every segment whose number is an integral multiple of the scatter/gather list size to fall on a device-granularity boundary, as specified in the dlim_granular field in the ddi_dma_lim_x86(9S) structure.
If the scatter/gather list size is 1 (either because no chaining is available or because the driver does not want to use the chaining feature), then the total I/O count for a single DMA operation is the size of DMA segment denoted by the single DMA cookie that is passed in the call to ddi_dmae_prog(). In this case, the system arranges for each DMA segment to be a multiple of the device-granularity size.
Specifies the DMA transfer size. The default of zero (DMAE_PATH_DEF) specifies ISA compatibility mode. In that mode, channels 0, 1, 2, and 3 are programmed in 8-bit mode (DMAE_PATH_8), and channels 5, 6, and 7 are programmed in 16-bit, count-by-word mode (DMAE_PATH_16).
Specifies the timing mode to be used during DMA data transfers. The default of zero (DMAE_CYCLES_1) specifies ISA compatible timing. Drivers using this mode must also specify DMAE_TRANS_SNGL in the der_trans structure member.
Specifies the bus transfer mode that the DMA engine should expect from the device. The default value of zero (DMAE_TRANS_SNGL) specifies that the device performs one transfer for each bus arbitration cycle. Devices that use ISA compatible timing (specified by a value of zero, which is the default, in the der_cycles structure member) should use the DMAE_TRANS_SNGL mode.
See attributes(5) for descriptions of the following attributes: