next up previous
Next: User-level DMA without Kernel Up: User-Level DMA - Early Previous: The FLASH solution

The PAL Code approach

We have seen so far that the mechanism of shadow addressing is a fast and reliable method to pass physical addresses to a DMA engine from user-space. What is difficult however, is to achieve atomicity of a user-level DMA operation, that is, to pass both physical addresses to hardware, without any danger of mixing physical addresses of different processes, and create race conditions. All previous approaches solve the problem of atomicity, by changing the context switch code to either abort semi-initiated DMA operations (e.g. in SHRIMP), or explicitly tell the DMA engine that a context switch has happened (e.g. in FLASH). If only there were a way to execute two assembly instructions uninterrupted from user-space, then the atomicity problem would have been solved. Unfortunately, traditional systems do not allow user-level processes to execute uninterrupted code because malicious users many monopolize the computing system. A recent processor, however, the DEC Alpha processor, provides a special mode of execution, the PAL mode, which allows uninterrupted execution [13]. PAL code is organized in 16-instruction long PAL calls. A PAL call is executed uninterrupted. To ensure protection, only super-users are allowed to write and install PAL functions. However, once a PAL function is installed, any ordinary user is allowed to invoke it.

User-level DMA atomicity can be achieved by turning the two instructions needed to start a DMA operation into a PAL call. Thus, the pseudo-code to start a DMA operation would look as follows:

  
     DMA(vsource, vdestination, size)
         call_pal user_level_dma(vsource, 
                  vdestination, size)
and the user_level_dma PAL call would be implemented as:
 DMA(vsource, vdestination, size)

STORE size TO shadow(vdestination)

LOAD return_status FROM shadow(vsource)

This solution achieves user-level DMA without any changes to the operating system kernel code. We believe that systems equipped with the Alpha processor should use this method of user-level DMA. The PAL code solution to user-level DMA has been incorporated into the Telegraphos I network interface [9].


next up previous
Next: User-level DMA without Kernel Up: User-Level DMA - Early Previous: The FLASH solution

Evangelos Markatos
Fri Nov 8 12:27:33 EET 1996