Update per CR feedback.

Update per CR feedback.

Add #ifdef INVARIRANTS and NULL check.

+1 to advise caution on this design. Sounds like you want to roll logs more often.

• Incorporate CR feedback.

Remove change to separate fix accidentally introduced.

• Use patch as suggested by cem instead. Move roundup within vfs_nonvmio_* instead of using check in allocbuf().

In D16573#356822, @kib wrote:

Why did you moved the check for mallocspace back into the want_malloc_buf() ?

Incorporated feedback.

• CR feedback.

• CR feedback.

• CR feedback.

Don't round new buffer size up to page size in buffer cache in cases where new buffer would be malloc buffer.

In D16573#352448, @kib wrote:

Hm, I am not sure. I agree that due to the bug, B_MALLOC buffers are never created by the current code. But, do we want to waste that many memory for small non-vmio buffers ? Isn't it better to fix creation of B_MALLOC buffers instead of removing them ? If a filesystem uses 512 or 1024 byte buffers, we do not want to multiply its memory use by a factor of 8 or 4.

It looks simple enough to set B_MALLOC if b_bufsize is zero and newbsize is less than half of the page size.

• Fix whitespace.

• CR feedback.

• CR feedback.

• Remove B_MALLOC flag entirely.

• Incorporate CR feedback.

Incorporate CR feedback.

Fix build break in i386 build.

Make memory mapped via pmap_qenter() non-executable.

Upon inspection of callers of pmap_qenter, I was unable to find any that required X. Most callers either made a temporary mapping for memcpy, or were allocating objects to be used by the vmem system itself. The only one raised concern was the ELF header load called as part of the exec path. From what I can tell only the elf header is mapped, none of the code within that gets executed. I will resubmit the original diff.

Per CR feedback localized the changes so that only vm_thread_new() is affected.