[lkml]   [2019]   [Mar]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [RFC dma-buf 0/3] Improve the dma-buf tracking
Hello Chenbo,Thank you for your RFC series.

On Wed, 27 Feb 2019 at 09:24, Chenbo Feng <> wrote:
> Currently, all dma-bufs share the same anonymous inode. While we can count
> how many dma-buf fds or mappings a process has, we can't get the size of
> the backing buffers or tell if two entries point to the same dma-buf. And
> in debugfs, we can get a per-buffer breakdown of size and reference count,
> but can't tell which processes are actually holding the references to each
> buffer.
> To resolve the issue above and provide better method for userspace to track
> the dma-buf usage across different processes, the following changes are
> proposed in dma-buf kernel side. First of all, replace the singleton inode
> inside the dma-buf subsystem with a mini-filesystem, and assign each
> dma-buf a unique inode out of this filesystem. With this change, calling
> stat(2) on each entry gives the caller a unique ID (st_ino), the buffer's
> size (st_size), and even the number of pages assigned to each dma-buffer.
> Secoundly, add the inode information to /sys/kernel/debug/dma_buf/bufinfo
> so in the case where a buffer is mmap()ed into a process’s address space
> but all remaining fds have been closed, we can still get the dma-buf
> information and try to accociate it with the process by searching the
> proc/pid/maps and looking for the corresponding inode number exposed in
> dma-buf debug fs. Thirdly, created an ioctl to assign names to dma-bufs
> which lets userspace assign short names (e.g., "CAMERA") to buffers. This
> information can be extremely helpful for tracking and accounting shared
> buffers based on their usage and original purpose. Last but not least, add
> dma-buf information to /proc/pid/fdinfo by adding a show_fdinfo() handler
> to dma_buf_file_operations. The handler will print the file_count and name
> of each buffer.

In general, I think I like the idea as it contributes to a much more
relevant usage analysis of dma-buf backed buffers.
I will get to doing a more detailed review soon, but immediately, we
might want to think a bit about the get/set_name IOCTLS - do we need
to think of disallowing multiple renaming of buffers once they start
getting used? It could otherwise make the whole metrics a lot

> Greg Hackmann (3):
> dma-buf: give each buffer a full-fledged inode
> dma-buf: add DMA_BUF_{GET,SET}_NAME ioctls
> dma-buf: add show_fdinfo handler
> drivers/dma-buf/dma-buf.c | 121 ++++++++++++++++++++++++++++++++---
> include/linux/dma-buf.h | 5 +-
> include/uapi/linux/dma-buf.h | 4 ++
> include/uapi/linux/magic.h | 1 +
> 4 files changed, 122 insertions(+), 9 deletions(-)
> --
> 2.21.0.rc2.261.ga7da99ff1b-goog


 \ /
  Last update: 2019-03-14 18:50    [W:0.185 / U:8.948 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site