10.6 Bus Memory Mapping

In many cases data is exchanged between the driver and the device through the memory. Two variants are possible:

(a) memory is located on the device card

(b) memory is the main memory of the computer

In case (a) the driver always copies the data back and forth between the on-card memory and the main memory as necessary. To map the on-card memory into the kernel virtual address space the physical address and length of the on-card memory must be defined as a SYS_RES_MEMORY resource. That resource can then be allocated and activated, and its virtual address obtained using rman_get_virtual(). The older drivers used the function pmap_mapdev() for this purpose, which should not be used directly any more. Now it is one of the internal steps of resource activation.

Most of the ISA cards will have their memory configured for physical location somewhere in range 640KB-1MB. Some of the ISA cards require larger memory ranges which should be placed somewhere under 16MB (because of the 24-bit address limitation on the ISA bus). In that case if the machine has more memory than the start address of the device memory (in other words, they overlap) a memory hole must be configured at the address range used by devices. Many BIOSes allow configuration of a memory hole of 1MB starting at 14MB or 15MB. FreeBSD can handle the memory holes properly if the BIOS reports them properly (this feature may be broken on old BIOSes).

In case (b) just the address of the data is sent to the device, and the device uses DMA to actually access the data in the main memory. Two limitations are present: First, ISA cards can only access memory below 16MB. Second, the contiguous pages in virtual address space may not be contiguous in physical address space, so the device may have to do scatter/gather operations. The bus subsystem provides ready solutions for some of these problems, the rest has to be done by the drivers themselves.

Two structures are used for DMA memory allocation, bus_dma_tag_t and bus_dmamap_t. Tag describes the properties required for the DMA memory. Map represents a memory block allocated according to these properties. Multiple maps may be associated with the same tag.

Tags are organized into a tree-like hierarchy with inheritance of the properties. A child tag inherits all the requirements of its parent tag, and may make them more strict but never more loose.

Normally one top-level tag (with no parent) is created for each device unit. If multiple memory areas with different requirements are needed for each device then a tag for each of them may be created as a child of the parent tag.

The tags can be used to create a map in two ways.

First, a chunk of contiguous memory conformant with the tag requirements may be allocated (and later may be freed). This is normally used to allocate relatively long-living areas of memory for communication with the device. Loading of such memory into a map is trivial: it is always considered as one chunk in the appropriate physical memory range.

Second, an arbitrary area of virtual memory may be loaded into a map. Each page of this memory will be checked for conformance to the map requirement. If it conforms then it is left at its original location. If it is not then a fresh conformant “bounce page” is allocated and used as intermediate storage. When writing the data from the non-conformant original pages they will be copied to their bounce pages first and then transferred from the bounce pages to the device. When reading the data would go from the device to the bounce pages and then copied to their non-conformant original pages. The process of copying between the original and bounce pages is called synchronization. This is normally used on a per-transfer basis: buffer for each transfer would be loaded, transfer done and buffer unloaded.

The functions working on the DMA memory are:

As of now PREREAD and POSTWRITE are null operations but that may change in the future, so they must not be ignored in the driver. Synchronization is not needed for the memory obtained from bus_dmamem_alloc().

Before calling the callback function from bus_dmamap_load() the segment array is stored in the stack. And it gets pre-allocated for the maximal number of segments allowed by the tag. Because of this the practical limit for the number of segments on i386 architecture is about 250-300 (the kernel stack is 4KB minus the size of the user structure, size of a segment array entry is 8 bytes, and some space must be left). Because the array is allocated based on the maximal number this value must not be set higher than really needed. Fortunately, for most of hardware the maximal supported number of segments is much lower. But if the driver wants to handle buffers with a very large number of scatter-gather segments it should do that in portions: load part of the buffer, transfer it to the device, load next part of the buffer, and so on.

Another practical consequence is that the number of segments may limit the size of the buffer. If all the pages in the buffer happen to be physically non-contiguous then the maximal supported buffer size for that fragmented case would be (nsegments * page_size). For example, if a maximal number of 10 segments is supported then on i386 maximal guaranteed supported buffer size would be 40K. If a higher size is desired then special tricks should be used in the driver.

If the hardware does not support scatter-gather at all or the driver wants to support some buffer size even if it is heavily fragmented then the solution is to allocate a contiguous buffer in the driver and use it as intermediate storage if the original buffer does not fit.

Below are the typical call sequences when using a map depend on the use of the map. The characters -> are used to show the flow of time.

For a buffer which stays practically fixed during all the time between attachment and detachment of a device:

bus_dmamem_alloc -> bus_dmamap_load -> ...use buffer... -> -> bus_dmamap_unload -> bus_dmamem_free

For a buffer that changes frequently and is passed from outside the driver:

          bus_dmamap_create ->
          -> bus_dmamap_load -> bus_dmamap_sync(PRE...) -> do transfer ->
          -> bus_dmamap_sync(POST...) -> bus_dmamap_unload ->
          ...
          -> bus_dmamap_load -> bus_dmamap_sync(PRE...) -> do transfer ->
          -> bus_dmamap_sync(POST...) -> bus_dmamap_unload ->
          -> bus_dmamap_destroy       


When loading a map created by bus_dmamem_alloc() the passed address and size of the buffer must be the same as used in bus_dmamem_alloc(). In this case it is guaranteed that the whole buffer will be mapped as one segment (so the callback may be based on this assumption) and the request will be executed immediately (EINPROGRESS will never be returned). All the callback needs to do in this case is to save the physical address.

A typical example would be:

          static void
        alloc_callback(void *arg, bus_dma_segment_t *seg, int nseg, int error)
        {
          *(bus_addr_t *)arg = seg[0].ds_addr;
        }

          ...
          int error;
          struct somedata {
            ....
          };
          struct somedata *vsomedata; /* virtual address */
          bus_addr_t psomedata; /* physical bus-relative address */
          bus_dma_tag_t tag_somedata;
          bus_dmamap_t map_somedata;
          ...

          error=bus_dma_tag_create(parent_tag, alignment,
           boundary, lowaddr, highaddr, /*filter*/ NULL, /*filterarg*/ NULL,
           /*maxsize*/ sizeof(struct somedata), /*nsegments*/ 1,
           /*maxsegsz*/ sizeof(struct somedata), /*flags*/ 0,
           &tag_somedata);
          if(error)
          return error;

          error = bus_dmamem_alloc(tag_somedata, &vsomedata, /* flags*/ 0,
             &map_somedata);
          if(error)
             return error;

          bus_dmamap_load(tag_somedata, map_somedata, (void *)vsomedata,
             sizeof (struct somedata), alloc_callback,
             (void *) &psomedata, /*flags*/0);       

Looks a bit long and complicated but that is the way to do it. The practical consequence is: if multiple memory areas are allocated always together it would be a really good idea to combine them all into one structure and allocate as one (if the alignment and boundary limitations permit).

When loading an arbitrary buffer into the map created by bus_dmamap_create() special measures must be taken to synchronize with the callback in case it would be delayed. The code would look like:

          {
           int s;
           int error;

           s = splsoftvm();
           error = bus_dmamap_load(
               dmat,
               dmamap,
               buffer_ptr,
               buffer_len,
               callback,
               /*callback_arg*/ buffer_descriptor,
               /*flags*/0);
           if (error == EINPROGRESS) {
               /*
                * Do whatever is needed to ensure synchronization
                * with callback. Callback is guaranteed not to be started
                * until we do splx() or tsleep().
                */
              }
           splx(s);
          }       

Two possible approaches for the processing of requests are:

1. If requests are completed by marking them explicitly as done (such as the CAM requests) then it would be simpler to put all the further processing into the callback driver which would mark the request when it is done. Then not much extra synchronization is needed. For the flow control reasons it may be a good idea to freeze the request queue until this request gets completed.

2. If requests are completed when the function returns (such as classic read or write requests on character devices) then a synchronization flag should be set in the buffer descriptor and tsleep() called. Later when the callback gets called it will do its processing and check this synchronization flag. If it is set then the callback should issue a wakeup. In this approach the callback function could either do all the needed processing (just like the previous case) or simply save the segments array in the buffer descriptor. Then after callback completes the calling function could use this saved segments array and do all the processing.