|
>Can anyone tell me the performance effect in
the following case:
> 1) copying data directly to mmapped memory
usingmemcpy() in user space .
> 2) copying the data in the IOCTL call which in
turn copies from user space to the kernel space
buffer(copy_from_user).
Both perform similarly for low bandwidth of data xfer.
But if u r using a high data xfer appl, such as a disk read,
or typically video streaming..( thats why u hv framebuffers ) mmap
scores.
If u see mmap, uve these operations for data xfer:
1, mmap the buffer. preferably u allocate large enough buffer
and do
remap_page_range() { there is a limitation of 128k bytes though }
Else u do through nopage handlers { which adds some overhead for
page faults }
2. Then u do direct read/writes to the buffer.
In case if copy_*_user, u do
1. fill the buffer in appl space ( which may involve swapping..
)
2. do verify_area() in copy_*_user()
3. do actual copy.
Almost 2 times data xfer ( or more if u hv swapping )
More over u can do non blocking read/writes more effectively
with mmaped memory.
*** CORRECT ME IF IM WRONG ANYWHERE ****
more comments.. ?
-nagaraj
\(*_*)/
|
