When a module has relocation sections with tens of thousands of
entries, counting the distinct/unique entries only (i.e. no
duplicates) at load time can take tens of seconds and up to minutes.
The sore point is the count_relocs() function which is called as part
of the architecture specific module loading processing path:
-> load_module() generic
-> module_frob_arch_sections() arch specific
-> get_plt_size() 32-bit
-> get_stubs_size() 64-bit
-> count_relocs()
Here count_relocs is being called to find out how many distinct
targets of R_PPC_REL24 relocations there are, since each distinct
target needs a PLT entry or a stub created for it.
The previous counting algorithm has O(n^2) complexity. Basically two
solutions were proposed on the e-mail list: a hash based approach and
a sort based approach.
The hash based approach is the fastest (O(n)) but the has it needs
additional memory and for certain corner cases it could take lots of
memory due to the degeneration of the hash. One such proposal was
submitted here:
http://ozlabs.org/pipermail/linuxppc-dev/2007-June/037641.html
The sort based approach is slower (O(n * log n + n)) but if the
sorting is done "in place" it doesn't need additional memory.
This has O(n + n * log n) complexity with no additional memory
requirements.
This commit implements the in-place sort option.
Signed-off-by: Emil Medve <Emilian.Medve@Freescale.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Lucas Woods <woodzy@gmail.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
We were using -mno-minimal-toc on everything in arch/powerpc/kernel,
which means that all the functions in there were putting all their
TOC entries in the top-level TOC, and it was overflowing on an
allyesconfig build. For various reasons, prom_init.c does need
-mno-minimal-toc, but the other .c files in there can use sub-TOCs
quite happily. This change is sufficient for now to stop the TOC
overflowing; other directories under arch/powerpc also use
-mno-minimal-toc and could also be changed later if necessary.
Lmbench runs with and without this patch showed no significant speed
differences.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Paul Mackerras <paulus@samba.org>
The PCI IRQ code has a fallback when the device-tree parsing fails, that
tries to map the interrupt indicated by PCI_INTERRUPT_LINE if the firmware
set something in there. This is a bit fragile but has proven useful in some
cases so far. However, it's causing us to incorrectly try to map interrupt 0
on various setups, so let's prevent that case, as none of the cases where
the fallback is legit should have an IRQ 0.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This patch changes the PowerPC PCI code to disable IO and/or Memory
decoding on a PCI device when a resource of that type failed to be
allocated. This is done to avoid having unallocated dangling BARs
enabled that might try to decode on top of other devices.
If a proper resource is assigned later on, then pci_enable_device()
will take care of re-enabling decoding.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Apple firmware has a strange way to "close" bridge resources by setting
them to some bogus values that overlap RAM (strangely, I haven't seen it
conflicting with DMA so far...). This explicitely closes them to avoid
problems. Previously, they would be closed as a consequence of failing
to be allocated, but this makes it more explicit, and thus the log
message is more explicit too.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Our implementation of pcibios_enable_device() has a couple of problems.
One is that it should not check IORESOURCE_UNSET, as this might be
left dangling after resource assignment (shouldn't but there are
bugs), but instead, we make it check resource->parent which should
be a reliable indication that the resource has been successfully
claimed (it's in the resource tree).
Then, we also need to skip ROM resources that haven't been enabled
as x86 does.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This merge the two implementations, based on the previously
fixed up 32 bits one. The pcibios_enable_device_hook in ppc_md
is now available for ppc64 use. Also remove the new unused
"initial" parameter from it and fixup users.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Our implementation of pcibios_enable_device() incorrectly ignores
the mask argument and always checks that all resources have been
allocated, which isn't the right thing to do anymore.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
The way iSeries manages PCI IO and Memory resources is a bit strange
and is based on overriding the content of those resources with home
cooked ones afterward.
This changes it a bit to better integrate with the new resource handling
so that the "virtual" tokens that iSeries replaces resources with are
done from the proper per-device fixup hook, and bridge resources are
set to enclose that token space. This fixes various things such as
the output of /proc/iomem & ioports, among others. This also fixes up
various boot messages as well.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
The 32 bits PCI code now uses the generic code for assigning unassigned
resources and an algorithm similar to x86 for claiming existing ones.
This works far better than the 64 bits code which basically can only
claim existing ones (pci_probe_only=1) or would fall apart completely.
This merges them so that the new 32 bits implementation is used for both.
64 bits now gets the new PCI flags for controlling the behaviour, though
the old pci_probe_only global is still there for now to be cleared if you
want to.
I kept a pcibios_claim_one_bus() function mostly based on the old 64
bits code for use by the DLPAR hotplug. This will have to be cleaned
up, thought I hope it will work in the meantime.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
The PCI code in 32 and 64 bits fixes up resources differently.
32 bits uses a header quirk plus handles bridges in pcibios_fixup_bus()
while 64 bits does things in various places depending on whether you
are using OF probing, using PCI hotplug, etc...
This merges those by basically using the 32 bits approach for both,
with various tweaks to make 64 bits work with the new approach.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This merges the PowerPC 32 and 64 bits version of pcibios_resource_to_bus
and pcibios_bus_to_resource().
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This adds flags the platforms can use to enable domain numbers
in /proc/bus/pci.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
The 32 bits PCI code carries an old hack that was only useful for G5
machines. Nowdays, the 32 bits kernel doesn't support any of those
machines anymore so the hack is basically never used, so remove it.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This adds to the 32 bits PCI code some flags, replacing the old
pci_assign_all_busses global, that allow us to control various
aspects of the PCI probing, such as whether to re-assign all
resources or not, or to not try to assign anything at all.
This also adds the flag x86 already has to avoid ISA alignment
on bridges that don't have ISA forwarding enabled (no legacy
devices on the top level bus) and sets it for PowerMacs.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
The 32 bits PowerPC PCI code has a hack for use by some PowerMacs
to try to re-open PCI<->PCI bridge IO resources that were closed
by the firmware. This is no longer necessary as the generic code
will now do that for us.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This makes the 32 bits PowerPC PCI code use the generic code to assign
resources to devices that had unassigned or conflicting resources.
This allow us to remove the local implementation that was incomplete and
could not assign for example a PCI<->PCI bridge from scratch, which is
needed on various embedded platforms.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
There's a stale & bogus piece of code in 32 bits PCI code that
complains about ISA related alignment issues. Just remove it.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
PowerPC currently doesn't implement pci_set_dma_mask(), which means drivers
calling it will get the generic version in drivers/pci/pci.c.
The powerpc dma mapping ops include a dma_set_mask() hook, which luckily is
not implemented by anyone - so there is no bug in the fact that the hook
is currently never called.
However in future we'll add implementation(s) of dma_set_mask(), and so we
need pci_set_dma_mask() to call the hook.
To save adding a hook to the dma mapping ops, pci-set_consistent_dma_mask()
simply calls the dma_set_mask() hook and then copies the new mask into
dev.coherenet_dma_mask.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Paul Mackerras <paulus@samba.org>
We have multiple calls to has_feature being inlined, but gcc can't
be sure that the store via get_paca() doesn't alias the path to
cur_cpu_spec->feature.
Reorder to put the calls to read_purr and read_spurr adjacent to each
other. To add a sense of consistency, reorder the remaining lines to
perform parallel steps on purr and scaled purr of each line instead of
calculating and then using one value before going on to the next.
In addition, we can tell gcc that no SPURR means no PURR. The test is
completely hidden in the PURR case, and in the !PURR case the second test
is eliminated resulting in the simple register copy in the out-of-line
branch.
Further, gcc sees get_paca()->system_time referenced several times and
allocates a register to address it (shadowing r13) instead of caching its
value. Reading into a local varable saves the shadow of r13 and removes
a potentially duplicate load (between the nested if and its parent).
Signed-off-by: Milton Miller <miltonm@bga.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
If CPU_FTR_PURR is not set, we will never set cpu_purr_data->initialized.
Checking via __get_cpu_var on 64 bit avoids one dependent load compared
to cpu_has_feature in the not-present case, and is always required when
it is present. The code is under CONFIG_VIRT_CPU_ACCOUNTING so 32 bit
will not be affected.
Signed-off-by: Milton Miller <miltonm@bga.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
timer_interrupt() was calculating per_cpu_offset several times, having to
start from the toc because of potential aliasing issues.
Placing both decrementer per_cpu varables in a struct and calculating
the address once with __get_cpu_var results in better code on both 32
and 64 bit.
Signed-off-by: Milton Miller <miltonm@bga.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Use __get_cpu_var(x) instead of per_cpu(x, smp_processor_id()), as it
is optimized on ppc64 to access the current cpu's per-cpu offset directly;
it's local_paca.offset instead of TOC->paca[local_paca->processor_id].offset.
This is the trivial portion, two functions with one use each.
Signed-off-by: Milton Miller <miltonm@bga.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
as its only called from time_init, which is __init.
Also remove unneeded forward declaration.
Signed-off-by: Milton Miller <miltonm@bga.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
isel (Integer Select) is a new user space instruction in the
PowerISA 2.04 spec. Not all processors implement it so lets emulate
to ensure code built with isel will run everywhere.
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
This makes the early debug option force the console loglevel
to the max. The early debug option is meant to catch messages very
early in the kernel boot process, in many cases, before the kernel
has a chance to parse the "debug" command line argument. Thus it
makes sense when CONFIG_PPC_EARLY_DEBUG is set, to force the console
log level to the max at boot time.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This adds a variant of of_translate_address that uses the dma-ranges
property instead of "ranges", it's to be used by PCI code in parsing
the dma-ranges property.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This removes "volatile" from the MMIO pointer udbg_comport
in udbg_16550.c driver, it's useless and makes checkpatch.pl
complain when adding things to this file.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
The 32 bits PCI code will display a rather scary error message
PCI: Cannot allocate resource region N of device XXX
at boot when the existing setup of a device as left by the
firmware doesn't match the kernel needs and the device needs
to be moved. This is often not an error at all, as the kernel
will generally easily reallocate the device elsewhere.
This changes the message to something less scary and lowers
its level from error to warning.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
The 32-bit powerpc resource fixup code uses unsigned longs to do the
offsetting of resources which overflows on platforms such as 4xx where
resources can be 64 bits.
This fixes it by using resource_size_t instead.
However, the IO stuff does rely on some 32 bits arithmetic, so we hack
by cropping the result of the fixups for IO resources with a 32 bits
mask.
This isn't the prettiest but should work for now until we change the
32 bits PCI code to do IO mappings like 64 bits does, within a reserved
are of the kernel address space.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This merges the 32-bit and 64-bit implementations of
pci_process_bridge_OF_ranges(). The new function is cleaner than both
the old ones, and supports 64 bits ranges on ppc32 which is necessary
for the 4xx port.
It also adds some better (hopefully) output to the kernel log which
should help diagnose problems and makes better use of existing OF
parsing helpers (avoiding a few bugs of both implementations along
the way).
There are still a few unfortunate ifdef's but there is no way around
these for now at least not until some other bits of the PCI code are
made common.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This defines isa_mem_base on both 32 and 64 bits (it used to be 32 bits
only). This avoids a few ifdef's in later patches and potentially can
allow support for VGA text mode on 64 bits powerpc.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Currently we hardwire the number of SLBs to 64, but PAPR says we
should use the ibm,slb-size property to obtain the number of SLB
entries. This uses this property instead of assuming 64. If no
property is found, we assume 64 entries as before.
This soft patches the SLB handler, so it shouldn't change performance
at all.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
and it becomes clear that we should use zalloc_maybe_bootmem.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Paul Mackerras <paulus@samba.org>
It only needs the iommu_table address. It also makes use of the node
name to print error messages. So just pass it the things it needs.
This reduces the places that know about the pci_dn by one.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Paul Mackerras <paulus@samba.org>
The 'data' member of proc_ppc64_lparcfg is unused, but the lparcfg
module's init routine allocates 4K for it.
Remove the code which allocates and frees this buffer.
Signed-off-by: Nathan Lynch <ntl@pobox.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Also use of_unregister_driver to implement of_unregister_platform_driver.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Paul Mackerras <paulus@samba.org>
The size of swapper_pg_dir is 8k instead of 4k when using 64-bit PTEs
(CONFIG_PTE_64BIT).
This was reported by Cedric Hombourger <chombourger@gmail.com>
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
The commit fa13a5a1f2 (sched: restore
deterministic CPU accounting on powerpc), unconditionally calls
update_process_tick() in system context. In the deterministic
accounting case this is the correct thing to do. However, in the
non-deterministic accounting case we need to not do this, since doing
this results in the time accounted as hardware irq time being
artificially elevated.
Also this collapses 2 consecutive '#ifdef CONFIG_VIRT_CPU_ACCOUNTING'
checks in time.h into one for neatness.
Signed-off-by: Tony Breeds <tony@bakeyournoodle.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
