Warning about add_wired_entry

Warning about add_wired_entry

[Ralf Baechle]

The earlier posting wrt. add_wired_entry reminded my of this long standing
issue.

The kernel does not reserve any address space for use by add_wired_entry().
In other words there is the possibility that vmalloc, ioremap or other
kernel APIs will use the same address space resulting in a crash or worse.

Currently there are three users of add_wired_entry:

 - Alchemy
 - Cavium
 - Jazz

The interface itself also is hostile as it directly exposes the MIPS TLB
registers to the C programmer so should die.  And often wiring a TLB entry
is a bad idea.  TLB entries are a scarce resource and wasting them unwisely
may impact performance.  Especially on the lowest-end CPUs with just
16 TLB entries this can be significant but even on the high-end TLBs with
64 entries it's not always clear if wiring entries is a good idea.  And
there are a few core types that don't even have the capability of wiring
entries.

  Ralf

Re: Warning about add_wired_entry

[pascal]

Ralf Baechle wrote:
> The kernel does not reserve any address space for use by add_wired_entry().
> In other words there is the possibility that vmalloc, ioremap or other
> kernel APIs will use the same address space resulting in a crash or worse.

Thanks, I was not aware of that.

But what is the alternative ?  I have an Alchemy 1550 board with
an ethernet chip at 0xD-1000-0000, and last time I tried, enabling
  CONFIG_64BIT_PHYS_ADDR,
  CONFIG_ARCH_PHYS_ADDR_T_64BIT and
  CONFIG_PHYS_ADDR_T_64BIT
did not seem to allow access to the 36-bit I/O address space
through the platform_get_resource/request_mem_region machinery.
Is this supposed to work now ?

Pascal