Re: [PATCH v2 1/2] tracing: ring-buffer: Have the ring buffer code do the vmap of physical memory

[Mike Rapoport <rppt@kernel.org>]

On Tue, Apr 01, 2025 at 11:11:59AM -0400, Steven Rostedt wrote:
> On Tue, 1 Apr 2025 12:56:31 +0300
> Mike Rapoport <rppt@kernel.org> wrote:
> 
> > > For example, using "mem=" on the kernel command line will literally
> > > limit the amount of RAM the kernel will use, and in doing so will
> > > limit the page allocations too.  
> > 
> > And using memmap=m$n on x86 creates a hole in System RAM that does not have
> > neither struct page nor kernel mappings and it is never considered RAM
> > anywhere in mm.
> 
> Hmm, when that is used, then we had better not "free" the buffer.
> 
> >  
> > > IOW, all of these kernel command line things are *subtle*.
> > > 
> > > Don't mis-use them by then making assumptions about how they work
> > > today (or how they will work tomorrow).  
> > 
> > I'd say it's better not to use them at all. They cause weirdness in memory
> > layout and also they are inconsistent in how architectures implement them.
> >  
> > > > Mike can correct me if I'm wrong, but the memory that was stolen was actual
> > > > memory returned by the system (E820 in x86). It reserves the memory before
> > > > the memory allocation reserves this memory. So what reserve_mem returns is
> > > > valid memory that can be used by memory allocator, but is currently just
> > > > "reserved" which means it wants to prevent the allocator from using it.  
> > > 
> > > That may indeed be true of reserve_mem.  
> > 
> > The reserve_mem behaves like any other early allocation, it has proper
> > struct pages (PG_Reserved) and it is mapped in the direct map so
> > phys_to_virt() will work on it.
> > 
> > As for mapping it to userspace, vm_iomap_memory() seems the best API to
> > use. It has all the alignment checks and will refuse to map ranges that are
> > not properly aligned and it will use vma information to create the right
> > mappings.
> >  
> 
> When using vmap() to get the virtual addresses (via the kmalloc_array() of
> struct pages), the vunmap() gives the memory back to the memory allocator:
> 
> ~# free
>                total        used        free      shared  buff/cache   available
> Mem:         8185928      296676     7840576         920      148280     7889252
> Swap:        7812092           0     7812092
> ~# rmdir /sys/kernel/tracing/instances/boot_mapped/
> ~# free
>                total        used        free      shared  buff/cache   available
> Mem:         8206404      290868     7866772         920      148384     7915536
> Swap:        7812092           0     7812092
> 
> With no issues.
> 
> But if I use vmap_page_range(), how do I give that back to the memory allocator?

But you don't need neither vmap() nor vmap_page_range() to have kernel page
tables for memory that you get from reserve_mem. It's already mapped and
plain phys_to_virt() gives you the virtual address you can use.
 
> Calling vunmap() on that memory gives me:
> 
>  1779.832484] ------------[ cut here ]------------
> [ 1779.834076] Trying to vunmap() nonexistent vm area (000000027c000000)
> [ 1779.835941] WARNING: CPU: 6 PID: 956 at mm/vmalloc.c:3413 vunmap+0x5a/0x60
> 
> What's the proper way to say: "I no longer need this physical memory I
> reserved, the kernel can now use it"?
 
free_reserved_area()

> -- Steve

-- 
Sincerely yours,
Mike.