kmalloc before kmem_cache_init
I am kind of lost trying to figure out how can kmallocs work if they are called before kmem_cache_init (e.g. such a case occurs when in start_kernel() we call parse_early_param() (which in turn might call early_serial8250_setup() if earlycon= is used in kernel command-line) before mm_init() (which later calls kmem_cache_init() )
From what I understood, kmalloc() (SLUB implementation) first tries to find out the slab fit for the size being kmalloced and slabs are nothing but kmalloc_caches[]. Now these kmalloc_caches get initialized in kmem_cache_init().
What am I missing here ? Regards, Sukanto Ghosh -- Regards, Sukanto Ghosh
Hi Sukanto On Sat, Jan 28, 2012 at 10:02 PM, Sukanto Ghosh <sukanto.cse.iitb@gmail.com> wrote:
I am kind of lost trying to figure out how can kmallocs work if they are called before kmem_cache_init
(e.g. such a case occurs when in start_kernel() we call parse_early_param() (which in turn might call early_serial8250_setup() if earlycon= is used in kernel command-line) before mm_init() (which later calls kmem_cache_init() )
I don't think that you're allowed to call kmalloc during early_setup. I don't see where early_serial8250_setup calls kmalloc. -- Dave Hylands Shuswap, BC, Canada http://www.davehylands.com
Hi Dave, If you look into start_kernel() the call to parse_early_param() precedes mm_init(). parse_early_param() eventually calls do_early_param(). do_early_param() parses for "earlycon" in kernel commandline and then calls the setup_function associated with earlycon. I have in my commandline: earlycon=uart8250,mmio32,0x10000000,9600 Call flow starting from drivers/tty/serial/8250_early.c will lead to kmalloc early_param("early_con", setup_early_serial8250_console) setup_early_serial8250_console() early_serial8250_setup() parse_options() ioremap_nocache() ... arch-specific-ioremap -- some form of arch specific __ioremap_caller --- get_vm_area_caller() __get_vm_area_node kzalloc_node kmalloc_node kmalloc Regards, Sukanto On Jan 29, 2012 12:28 PM, "Dave Hylands" <dhylands@gmail.com> wrote:
Hi Sukanto
On Sat, Jan 28, 2012 at 10:02 PM, Sukanto Ghosh <sukanto.cse.iitb@gmail.com> wrote:
I am kind of lost trying to figure out how can kmallocs work if they are called before kmem_cache_init
(e.g. such a case occurs when in start_kernel() we call parse_early_param() (which in turn might call early_serial8250_setup() if earlycon= is used in kernel command-line) before mm_init() (which later calls kmem_cache_init() )
I don't think that you're allowed to call kmalloc during early_setup. I don't see where early_serial8250_setup calls kmalloc.
-- Dave Hylands Shuswap, BC, Canada http://www.davehylands.com
Hi Sukanto, On Sat, Jan 28, 2012 at 11:25 PM, Sukanto Ghosh <sukanto.cse.iitb@gmail.com> wrote:
Hi Dave,
If you look into start_kernel() the call to parse_early_param() precedes mm_init(). parse_early_param() eventually calls do_early_param(). do_early_param() parses for "earlycon" in kernel commandline and then calls the setup_function associated with earlycon.
I have in my commandline: earlycon=uart8250,mmio32,0x10000000,9600
Call flow starting from drivers/tty/serial/8250_early.c will lead to kmalloc
early_param("early_con", setup_early_serial8250_console) setup_early_serial8250_console() early_serial8250_setup() parse_options() ioremap_nocache() ... arch-specific-ioremap -- some form of arch specific __ioremap_caller --- get_vm_area_caller() __get_vm_area_node kzalloc_node kmalloc_node kmalloc
It looks like CONFIG_FIX_EARLYCON_MEM defaults to y (on x86 anyways), so this would cause the path that doesn't call ioremam_nocache to be taken. I'm guessing that if you specify earycon, then you also need to ensure that CONFIG_FIX_EARLYCON is set. You didn't mention which architecture you were using. ARM has an early_printk which is sort of like an early console. -- Dave Hylands Shuswap, BC, Canada http://www.davehylands.com
So to use a MMIO based 8250 as "earlycon", either we need to have some sort of fixmap (as in case of x86) or otherwise we need to have a ioremap_nocache() version that works even before kmem_cache_init() is called. I am not sure how many architectures will have such form of ioremap other than maybe nommu variants. Greg, Has anyone used the ioremap path so far ? If not, then why have it there as usual ioremap can be used only after mm_init(), but this driver whenever used will get called earlier than that ? Regards, Sukanto On Sun, Jan 29, 2012 at 3:03 PM, Dave Hylands <dhylands@gmail.com> wrote:
Hi Sukanto,
On Sat, Jan 28, 2012 at 11:25 PM, Sukanto Ghosh <sukanto.cse.iitb@gmail.com> wrote:
Hi Dave,
If you look into start_kernel() the call to parse_early_param() precedes mm_init(). parse_early_param() eventually calls do_early_param(). do_early_param() parses for "earlycon" in kernel commandline and then calls the setup_function associated with earlycon.
I have in my commandline: earlycon=uart8250,mmio32,0x10000000,9600
Call flow starting from drivers/tty/serial/8250_early.c will lead to kmalloc
early_param("early_con", setup_early_serial8250_console) setup_early_serial8250_console() early_serial8250_setup() parse_options() ioremap_nocache() ... arch-specific-ioremap -- some form of arch specific __ioremap_caller --- get_vm_area_caller() __get_vm_area_node kzalloc_node kmalloc_node kmalloc
It looks like CONFIG_FIX_EARLYCON_MEM defaults to y (on x86 anyways), so this would cause the path that doesn't call ioremam_nocache to be taken.
I'm guessing that if you specify earycon, then you also need to ensure that CONFIG_FIX_EARLYCON is set.
You didn't mention which architecture you were using. ARM has an early_printk which is sort of like an early console.
-- Dave Hylands Shuswap, BC, Canada http://www.davehylands.com
-- Regards, Sukanto Ghosh
On 01/29/2012 04:29 AM, Sukanto Ghosh wrote:
So to use a MMIO based 8250 as "earlycon", either we need to have some sort of fixmap (as in case of x86) or otherwise we need to have a ioremap_nocache() version that works even before kmem_cache_init() is called. I am not sure how many architectures will have such form of ioremap other than maybe nommu variants.
Some architectures check for mem_init_done. Probably every architecture must provide an "early_ioremap" function to be used in early _code_.
Greg,
Has anyone used the ioremap path so far ? If not, then why have it there as usual ioremap can be used only after mm_init(), but this driver whenever used will get called earlier than that ?
Regards, Sukanto
On Sun, Jan 29, 2012 at 3:03 PM, Dave Hylands<dhylands@gmail.com> wrote:
Hi Sukanto,
On Sat, Jan 28, 2012 at 11:25 PM, Sukanto Ghosh <sukanto.cse.iitb@gmail.com> wrote:
Hi Dave,
If you look into start_kernel() the call to parse_early_param() precedes mm_init(). parse_early_param() eventually calls do_early_param(). do_early_param() parses for "earlycon" in kernel commandline and then calls the setup_function associated with earlycon.
I have in my commandline: earlycon=uart8250,mmio32,0x10000000,9600
Call flow starting from drivers/tty/serial/8250_early.c will lead to kmalloc
early_param("early_con", setup_early_serial8250_console) setup_early_serial8250_console() early_serial8250_setup() parse_options() ioremap_nocache() ... arch-specific-ioremap -- some form of arch specific __ioremap_caller --- get_vm_area_caller() __get_vm_area_node kzalloc_node kmalloc_node kmalloc
It looks like CONFIG_FIX_EARLYCON_MEM defaults to y (on x86 anyways), so this would cause the path that doesn't call ioremam_nocache to be taken.
I'm guessing that if you specify earycon, then you also need to ensure that CONFIG_FIX_EARLYCON is set.
You didn't mention which architecture you were using. ARM has an early_printk which is sort of like an early console.
-- Dave Hylands Shuswap, BC, Canada http://www.davehylands.com
-Fredrick
participants (3)
-
Dave Hylands -
Fredrick -
Sukanto Ghosh