| RFC-0014: Relax zx_fifo_create constraints | |
|---|---|
| Status | Accepted |
| Areas |
|
| Description | Permit `zx_fifo_create` calls with element counts that are not powers of two. |
| Issues | |
| Gerrit change | |
| Authors | |
| Reviewers | |
| Date submitted (year-month-day) | 2020-12-02 |
| Date reviewed (year-month-day) | 2020-12-09 |
Summary
zx_fifo_create currently requires the caller to pass an element count that is a power of two.
This allows the kernel to make a small optimization on the read and write paths by using a
bit-mask instead of the modulus operator to determine the offset of a wrapped record in a buffer.
This constraint means that users cannot effectively use the FIFO to its full capacity if their
element sizes do not happen to be 4096/(2**n) bytes; in these cases the FIFO capacity
(element_size * element_count) will be less than possible full capacity of the FIFO (which
internally uses a 4096-byte buffer).
The proposal is to allow FIFOs to be created with arbitrary element counts (up to the FIFO size limit) so the full FIFO capacity can be used for FIFO records of arbitrary sizes.
Motivation
This proposal is motivated by the desire to add a field to an existing FIFO record used in the
block stack; the field in question is a trace identifier, which will be used to support
cross-process tracing of block IO. Adding this field changes the size of the FIFO records from
32 bytes to 40 bytes.
If FIFO records are 40 bytes long, then the maximum value that element_count could currently
be set to is 64, which results in 40 * 64 = 2560 bytes used in the 4096-byte buffer.
This RFC would permit an element_count value of 102, which results in 40 * 102 = 4080 bytes
used (almost doubling the usable capacity of the FIFO).
Design
This is a trivial change that is implemented in https://fuchsia-review.googlesource.com/c/fuchsia/+/409498.
Implementation
This is a trivial change that is implemented in https://fuchsia-review.googlesource.com/c/fuchsia/+/409498.
Performance
The kernel can currently make a small performance optimization since it knows the element count is a power of two; particularly, the kernel can use a bitwise-AND to determine the position of an element in the internal buffer, instead of a modulo.
That said, this performance optimization washes out given the expense of jumping back and forth between the kernel and userspace.
Concretely: performance testing in https://fuchsia-review.googlesource.com/c/fuchsia/+/409498 indicates that there are no measurable performance costs for eliminating this optimization.
Security considerations
None.
Privacy considerations
None.
Testing
This is a trivial change for which tests are added in https://fuchsia-review.googlesource.com/c/fuchsia/+/409498.
Documentation
The zx_fifo_create documentation is adjusted in
https://fuchsia-review.googlesource.com/c/fuchsia/+/409498 based on the new relaxed constraints.
Drawbacks, alternatives, and unknowns
None.
Prior art and references
None.