linux/arch/arc/mm/tlb.c

781 lines
23 KiB
C
Raw Normal View History

/*
* TLB Management (flush/create/diagnostics) for ARC700
*
* Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 as
* published by the Free Software Foundation.
*
* vineetg: Aug 2011
* -Reintroduce duplicate PD fixup - some customer chips still have the issue
*
* vineetg: May 2011
* -No need to flush_cache_page( ) for each call to update_mmu_cache()
* some of the LMBench tests improved amazingly
* = page-fault thrice as fast (75 usec to 28 usec)
* = mmap twice as fast (9.6 msec to 4.6 msec),
* = fork (5.3 msec to 3.7 msec)
*
* vineetg: April 2011 :
* -MMU v3: PD{0,1} bits layout changed: They don't overlap anymore,
* helps avoid a shift when preparing PD0 from PTE
*
* vineetg: April 2011 : Preparing for MMU V3
* -MMU v2/v3 BCRs decoded differently
* -Remove TLB_SIZE hardcoding as it's variable now: 256 or 512
* -tlb_entry_erase( ) can be void
* -local_flush_tlb_range( ):
* = need not "ceil" @end
* = walks MMU only if range spans < 32 entries, as opposed to 256
*
* Vineetg: Sept 10th 2008
* -Changes related to MMU v2 (Rel 4.8)
*
* Vineetg: Aug 29th 2008
* -In TLB Flush operations (Metal Fix MMU) there is a explict command to
* flush Micro-TLBS. If TLB Index Reg is invalid prior to TLBIVUTLB cmd,
* it fails. Thus need to load it with ANY valid value before invoking
* TLBIVUTLB cmd
*
* Vineetg: Aug 21th 2008:
* -Reduced the duration of IRQ lockouts in TLB Flush routines
* -Multiple copies of TLB erase code seperated into a "single" function
* -In TLB Flush routines, interrupt disabling moved UP to retrieve ASID
* in interrupt-safe region.
*
* Vineetg: April 23rd Bug #93131
* Problem: tlb_flush_kernel_range() doesnt do anything if the range to
* flush is more than the size of TLB itself.
*
* Rahul Trivedi : Codito Technologies 2004
*/
#include <linux/module.h>
#include <linux/bug.h>
#include <asm/arcregs.h>
#include <asm/setup.h>
#include <asm/mmu_context.h>
#include <asm/mmu.h>
/* Need for ARC MMU v2
*
* ARC700 MMU-v1 had a Joint-TLB for Code and Data and is 2 way set-assoc.
* For a memcpy operation with 3 players (src/dst/code) such that all 3 pages
* map into same set, there would be contention for the 2 ways causing severe
* Thrashing.
*
* Although J-TLB is 2 way set assoc, ARC700 caches J-TLB into uTLBS which has
* much higher associativity. u-D-TLB is 8 ways, u-I-TLB is 4 ways.
* Given this, the thrasing problem should never happen because once the 3
* J-TLB entries are created (even though 3rd will knock out one of the prev
* two), the u-D-TLB and u-I-TLB will have what is required to accomplish memcpy
*
* Yet we still see the Thrashing because a J-TLB Write cause flush of u-TLBs.
* This is a simple design for keeping them in sync. So what do we do?
* The solution which James came up was pretty neat. It utilised the assoc
* of uTLBs by not invalidating always but only when absolutely necessary.
*
* - Existing TLB commands work as before
* - New command (TLBWriteNI) for TLB write without clearing uTLBs
* - New command (TLBIVUTLB) to invalidate uTLBs.
*
* The uTLBs need only be invalidated when pages are being removed from the
* OS page table. If a 'victim' TLB entry is being overwritten in the main TLB
* as a result of a miss, the removed entry is still allowed to exist in the
* uTLBs as it is still valid and present in the OS page table. This allows the
* full associativity of the uTLBs to hide the limited associativity of the main
* TLB.
*
* During a miss handler, the new "TLBWriteNI" command is used to load
* entries without clearing the uTLBs.
*
* When the OS page table is updated, TLB entries that may be associated with a
* removed page are removed (flushed) from the TLB using TLBWrite. In this
* circumstance, the uTLBs must also be cleared. This is done by using the
* existing TLBWrite command. An explicit IVUTLB is also required for those
* corner cases when TLBWrite was not executed at all because the corresp
* J-TLB entry got evicted/replaced.
*/
/* A copy of the ASID from the PID reg is kept in asid_cache */
DEFINE_PER_CPU(unsigned int, asid_cache) = MM_CTXT_FIRST_CYCLE;
/*
* Utility Routine to erase a J-TLB entry
* Caller needs to setup Index Reg (manually or via getIndex)
*/
static inline void __tlb_entry_erase(void)
{
write_aux_reg(ARC_REG_TLBPD1, 0);
write_aux_reg(ARC_REG_TLBPD0, 0);
write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
}
static inline unsigned int tlb_entry_lkup(unsigned long vaddr_n_asid)
{
unsigned int idx;
write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid);
write_aux_reg(ARC_REG_TLBCOMMAND, TLBProbe);
idx = read_aux_reg(ARC_REG_TLBINDEX);
return idx;
}
static void tlb_entry_erase(unsigned int vaddr_n_asid)
{
unsigned int idx;
/* Locate the TLB entry for this vaddr + ASID */
idx = tlb_entry_lkup(vaddr_n_asid);
/* No error means entry found, zero it out */
if (likely(!(idx & TLB_LKUP_ERR))) {
__tlb_entry_erase();
} else {
/* Duplicate entry error */
WARN(idx == TLB_DUP_ERR, "Probe returned Dup PD for %x\n",
vaddr_n_asid);
}
}
/****************************************************************************
* ARC700 MMU caches recently used J-TLB entries (RAM) as uTLBs (FLOPs)
*
* New IVUTLB cmd in MMU v2 explictly invalidates the uTLB
*
* utlb_invalidate ( )
* -For v2 MMU calls Flush uTLB Cmd
* -For v1 MMU does nothing (except for Metal Fix v1 MMU)
* This is because in v1 TLBWrite itself invalidate uTLBs
***************************************************************************/
static void utlb_invalidate(void)
{
#if (CONFIG_ARC_MMU_VER >= 2)
#if (CONFIG_ARC_MMU_VER == 2)
/* MMU v2 introduced the uTLB Flush command.
* There was however an obscure hardware bug, where uTLB flush would
* fail when a prior probe for J-TLB (both totally unrelated) would
* return lkup err - because the entry didnt exist in MMU.
* The Workround was to set Index reg with some valid value, prior to
* flush. This was fixed in MMU v3 hence not needed any more
*/
unsigned int idx;
/* make sure INDEX Reg is valid */
idx = read_aux_reg(ARC_REG_TLBINDEX);
/* If not write some dummy val */
if (unlikely(idx & TLB_LKUP_ERR))
write_aux_reg(ARC_REG_TLBINDEX, 0xa);
#endif
write_aux_reg(ARC_REG_TLBCOMMAND, TLBIVUTLB);
#endif
}
static void tlb_entry_insert(unsigned int pd0, unsigned int pd1)
{
unsigned int idx;
/*
* First verify if entry for this vaddr+ASID already exists
* This also sets up PD0 (vaddr, ASID..) for final commit
*/
idx = tlb_entry_lkup(pd0);
/*
* If Not already present get a free slot from MMU.
* Otherwise, Probe would have located the entry and set INDEX Reg
* with existing location. This will cause Write CMD to over-write
* existing entry with new PD0 and PD1
*/
if (likely(idx & TLB_LKUP_ERR))
write_aux_reg(ARC_REG_TLBCOMMAND, TLBGetIndex);
/* setup the other half of TLB entry (pfn, rwx..) */
write_aux_reg(ARC_REG_TLBPD1, pd1);
/*
* Commit the Entry to MMU
* It doesnt sound safe to use the TLBWriteNI cmd here
* which doesn't flush uTLBs. I'd rather be safe than sorry.
*/
write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
}
/*
* Un-conditionally (without lookup) erase the entire MMU contents
*/
noinline void local_flush_tlb_all(void)
{
unsigned long flags;
unsigned int entry;
struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
local_irq_save(flags);
/* Load PD0 and PD1 with template for a Blank Entry */
write_aux_reg(ARC_REG_TLBPD1, 0);
write_aux_reg(ARC_REG_TLBPD0, 0);
for (entry = 0; entry < mmu->num_tlb; entry++) {
/* write this entry to the TLB */
write_aux_reg(ARC_REG_TLBINDEX, entry);
write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
}
utlb_invalidate();
local_irq_restore(flags);
}
/*
* Flush the entrie MM for userland. The fastest way is to move to Next ASID
*/
noinline void local_flush_tlb_mm(struct mm_struct *mm)
{
/*
* Small optimisation courtesy IA64
* flush_mm called during fork,exit,munmap etc, multiple times as well.
* Only for fork( ) do we need to move parent to a new MMU ctxt,
* all other cases are NOPs, hence this check.
*/
if (atomic_read(&mm->mm_users) == 0)
return;
/*
* - Move to a new ASID, but only if the mm is still wired in
* (Android Binder ended up calling this for vma->mm != tsk->mm,
* causing h/w - s/w ASID to get out of sync)
* - Also get_new_mmu_context() new implementation allocates a new
* ASID only if it is not allocated already - so unallocate first
*/
destroy_context(mm);
if (current->mm == mm)
get_new_mmu_context(mm);
}
/*
* Flush a Range of TLB entries for userland.
* @start is inclusive, while @end is exclusive
* Difference between this and Kernel Range Flush is
* -Here the fastest way (if range is too large) is to move to next ASID
* without doing any explicit Shootdown
* -In case of kernel Flush, entry has to be shot down explictly
*/
void local_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
unsigned long end)
{
const unsigned int cpu = smp_processor_id();
unsigned long flags;
/* If range @start to @end is more than 32 TLB entries deep,
* its better to move to a new ASID rather than searching for
* individual entries and then shooting them down
*
* The calc above is rough, doesn't account for unaligned parts,
* since this is heuristics based anyways
*/
if (unlikely((end - start) >= PAGE_SIZE * 32)) {
local_flush_tlb_mm(vma->vm_mm);
return;
}
/*
* @start moved to page start: this alone suffices for checking
* loop end condition below, w/o need for aligning @end to end
* e.g. 2000 to 4001 will anyhow loop twice
*/
start &= PAGE_MASK;
local_irq_save(flags);
if (asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID) {
while (start < end) {
tlb_entry_erase(start | hw_pid(vma->vm_mm, cpu));
start += PAGE_SIZE;
}
}
utlb_invalidate();
local_irq_restore(flags);
}
/* Flush the kernel TLB entries - vmalloc/modules (Global from MMU perspective)
* @start, @end interpreted as kvaddr
* Interestingly, shared TLB entries can also be flushed using just
* @start,@end alone (interpreted as user vaddr), although technically SASID
* is also needed. However our smart TLbProbe lookup takes care of that.
*/
void local_flush_tlb_kernel_range(unsigned long start, unsigned long end)
{
unsigned long flags;
/* exactly same as above, except for TLB entry not taking ASID */
if (unlikely((end - start) >= PAGE_SIZE * 32)) {
local_flush_tlb_all();
return;
}
start &= PAGE_MASK;
local_irq_save(flags);
while (start < end) {
tlb_entry_erase(start);
start += PAGE_SIZE;
}
utlb_invalidate();
local_irq_restore(flags);
}
/*
* Delete TLB entry in MMU for a given page (??? address)
* NOTE One TLB entry contains translation for single PAGE
*/
void local_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
{
const unsigned int cpu = smp_processor_id();
unsigned long flags;
/* Note that it is critical that interrupts are DISABLED between
* checking the ASID and using it flush the TLB entry
*/
local_irq_save(flags);
if (asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID) {
tlb_entry_erase((page & PAGE_MASK) | hw_pid(vma->vm_mm, cpu));
utlb_invalidate();
}
local_irq_restore(flags);
}
#ifdef CONFIG_SMP
struct tlb_args {
struct vm_area_struct *ta_vma;
unsigned long ta_start;
unsigned long ta_end;
};
static inline void ipi_flush_tlb_page(void *arg)
{
struct tlb_args *ta = arg;
local_flush_tlb_page(ta->ta_vma, ta->ta_start);
}
static inline void ipi_flush_tlb_range(void *arg)
{
struct tlb_args *ta = arg;
local_flush_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
}
static inline void ipi_flush_tlb_kernel_range(void *arg)
{
struct tlb_args *ta = (struct tlb_args *)arg;
local_flush_tlb_kernel_range(ta->ta_start, ta->ta_end);
}
void flush_tlb_all(void)
{
on_each_cpu((smp_call_func_t)local_flush_tlb_all, NULL, 1);
}
void flush_tlb_mm(struct mm_struct *mm)
{
on_each_cpu_mask(mm_cpumask(mm), (smp_call_func_t)local_flush_tlb_mm,
mm, 1);
}
void flush_tlb_page(struct vm_area_struct *vma, unsigned long uaddr)
{
struct tlb_args ta = {
.ta_vma = vma,
.ta_start = uaddr
};
on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_tlb_page, &ta, 1);
}
void flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
unsigned long end)
{
struct tlb_args ta = {
.ta_vma = vma,
.ta_start = start,
.ta_end = end
};
on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_tlb_range, &ta, 1);
}
void flush_tlb_kernel_range(unsigned long start, unsigned long end)
{
struct tlb_args ta = {
.ta_start = start,
.ta_end = end
};
on_each_cpu(ipi_flush_tlb_kernel_range, &ta, 1);
}
#endif
/*
* Routine to create a TLB entry
*/
void create_tlb(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
{
unsigned long flags;
unsigned int asid_or_sasid, rwx;
unsigned long pd0, pd1;
/*
* create_tlb() assumes that current->mm == vma->mm, since
* -it ASID for TLB entry is fetched from MMU ASID reg (valid for curr)
* -completes the lazy write to SASID reg (again valid for curr tsk)
*
* Removing the assumption involves
* -Using vma->mm->context{ASID,SASID}, as opposed to MMU reg.
* -Fix the TLB paranoid debug code to not trigger false negatives.
* -More importantly it makes this handler inconsistent with fast-path
* TLB Refill handler which always deals with "current"
*
* Lets see the use cases when current->mm != vma->mm and we land here
* 1. execve->copy_strings()->__get_user_pages->handle_mm_fault
* Here VM wants to pre-install a TLB entry for user stack while
* current->mm still points to pre-execve mm (hence the condition).
* However the stack vaddr is soon relocated (randomization) and
* move_page_tables() tries to undo that TLB entry.
* Thus not creating TLB entry is not any worse.
*
* 2. ptrace(POKETEXT) causes a CoW - debugger(current) inserting a
* breakpoint in debugged task. Not creating a TLB now is not
* performance critical.
*
* Both the cases above are not good enough for code churn.
*/
if (current->active_mm != vma->vm_mm)
return;
local_irq_save(flags);
tlb_paranoid_check(asid_mm(vma->vm_mm, smp_processor_id()), address);
address &= PAGE_MASK;
/* update this PTE credentials */
pte_val(*ptep) |= (_PAGE_PRESENT | _PAGE_ACCESSED);
/* Create HW TLB(PD0,PD1) from PTE */
/* ASID for this task */
asid_or_sasid = read_aux_reg(ARC_REG_PID) & 0xff;
pd0 = address | asid_or_sasid | (pte_val(*ptep) & PTE_BITS_IN_PD0);
ARC: MMUv4 preps/1 - Fold PTE K/U access flags The current ARC VM code has 13 flags in Page Table entry: some software (accesed/dirty/non-linear-maps) and rest hardware specific. With 8k MMU page, we need 19 bits for addressing page frame so remaining 13 bits is just about enough to accomodate the current flags. In MMUv4 there are 2 additional flags, SZ (normal or super page) and WT (cache access mode write-thru) - and additionally PFN is 20 bits (vs. 19 before for 8k). Thus these can't be held in current PTE w/o making each entry 64bit wide. It seems there is some scope of compressing the current PTE flags (and freeing up a few bits). Currently PTE contains fully orthogonal distinct access permissions for kernel and user mode (Kr, Kw, Kx; Ur, Uw, Ux) which can be folded into one set (R, W, X). The translation of 3 PTE bits into 6 TLB bits (when programming the MMU) can be done based on following pre-requites/assumptions: 1. For kernel-mode-only translations (vmalloc: 0x7000_0000 to 0x7FFF_FFFF), PTE additionally has PAGE_GLOBAL flag set (and user space entries can never be global). Thus such a PTE can translate to Kr, Kw, Kx (as appropriate) and zero for User mode counterparts. 2. For non global entries, the PTE flags can be used to create mirrored K and U TLB bits. This is true after commit a950549c675f2c8c504 "ARC: copy_(to|from)_user() to honor usermode-access permissions" which ensured that user-space translations _MUST_ have same access permissions for both U/K mode accesses so that copy_{to,from}_user() play fair with fault based CoW break and such... There is no such thing as free lunch - the cost is slightly infalted TLB-Miss Handlers. Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
2013-06-17 20:42:13 +08:00
/*
* ARC MMU provides fully orthogonal access bits for K/U mode,
* however Linux only saves 1 set to save PTE real-estate
* Here we convert 3 PTE bits into 6 MMU bits:
* -Kernel only entries have Kr Kw Kx 0 0 0
* -User entries have mirrored K and U bits
*/
rwx = pte_val(*ptep) & PTE_BITS_RWX;
if (pte_val(*ptep) & _PAGE_GLOBAL)
rwx <<= 3; /* r w x => Kr Kw Kx 0 0 0 */
else
rwx |= (rwx << 3); /* r w x => Kr Kw Kx Ur Uw Ux */
pd1 = rwx | (pte_val(*ptep) & PTE_BITS_NON_RWX_IN_PD1);
tlb_entry_insert(pd0, pd1);
local_irq_restore(flags);
}
ARC: [mm] Lazy D-cache flush (non aliasing VIPT) flush_dcache_page( ) is MM hook to ensure that a page has consistent views between kernel and userspace. Thus it is called when * kernel writes to a page which at some later point could get mapped to userspace (so kernel mapping needs to be flushed-n-inv) * kernel is about to read from a page with possible userspace mappings (so userspace mappings needs to be made coherent with kernel ones) However for Non aliasing VIPT dcache, any userspace mapping will always be congruent to kernel mapping. Thus d-cache need need not be flushed at all (or delayed indefinitely). The only reason it does need to be flushed is when mapping code pages. Since icache doesn't snoop dcache, those dirty dcache lines need to be written back to memory and icache line invalidated so that icache lines fetch will get the right data. Decent gains on LMBench fork/exec/sh and File I/O micro-benchmarks. (1) FPGA @ 80 MHZ Processor, Processes - times in microseconds - smaller is better ------------------------------------------------------------------------------ Host OS Mhz null null open slct sig sig fork exec sh call I/O stat clos TCP inst hndl proc proc proc --------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- 3.9-rc6-a Linux 3.9.0-r 80 4.79 8.72 66.7 116. 239. 8.39 30.4 4798 14.K 34.K 3.9-rc6-b Linux 3.9.0-r 80 4.79 8.62 65.4 111. 239. 8.35 29.0 3995 12.K 30.K 3.9-rc7-c Linux 3.9.0-r 80 4.79 9.00 66.1 106. 239. 8.61 30.4 2858 10.K 24.K ^^^^ ^^^^ ^^^ File & VM system latencies in microseconds - smaller is better ------------------------------------------------------------------------------- Host OS 0K File 10K File Mmap Prot Page 100fd Create Delete Create Delete Latency Fault Fault selct --------- ------------- ------ ------ ------ ------ ------- ----- ------- ----- 3.9-rc6-a Linux 3.9.0-r 317.8 204.2 1122.3 375.1 3522.0 4.288 20.7 126.8 3.9-rc6-b Linux 3.9.0-r 298.7 223.0 1141.6 367.8 3531.0 4.866 20.9 126.4 3.9-rc7-c Linux 3.9.0-r 278.4 179.2 862.1 339.3 3705.0 3.223 20.3 126.6 ^^^^^ ^^^^^ ^^^^^ ^^^^ (2) Customer Silicon @ 500 MHz (166 MHz mem) ------------------------------------------------------------------------------ Host OS Mhz null null open slct sig sig fork exec sh call I/O stat clos TCP inst hndl proc proc proc --------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- abilis-ba Linux 3.9.0-r 497 0.71 1.38 4.58 12.0 35.5 1.40 3.89 2070 5525 13.K abilis-ca Linux 3.9.0-r 497 0.71 1.40 4.61 11.8 35.6 1.37 3.92 1411 4317 10.K ^^^^ ^^^^ ^^^ Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
2013-04-16 16:40:48 +08:00
/*
* Called at the end of pagefault, for a userspace mapped page
* -pre-install the corresponding TLB entry into MMU
* -Finalize the delayed D-cache flush of kernel mapping of page due to
* flush_dcache_page(), copy_user_page()
*
* Note that flush (when done) involves both WBACK - so physical page is
* in sync as well as INV - so any non-congruent aliases don't remain
*/
void update_mmu_cache(struct vm_area_struct *vma, unsigned long vaddr_unaligned,
pte_t *ptep)
{
unsigned long vaddr = vaddr_unaligned & PAGE_MASK;
unsigned long paddr = pte_val(*ptep) & PAGE_MASK;
struct page *page = pfn_to_page(pte_pfn(*ptep));
create_tlb(vma, vaddr, ptep);
if (page == ZERO_PAGE(0)) {
return;
}
/*
* Exec page : Independent of aliasing/page-color considerations,
* since icache doesn't snoop dcache on ARC, any dirty
* K-mapping of a code page needs to be wback+inv so that
* icache fetch by userspace sees code correctly.
* !EXEC page: If K-mapping is NOT congruent to U-mapping, flush it
* so userspace sees the right data.
* (Avoids the flush for Non-exec + congruent mapping case)
*/
if ((vma->vm_flags & VM_EXEC) ||
addr_not_cache_congruent(paddr, vaddr)) {
ARC: [mm] Lazy D-cache flush (non aliasing VIPT) flush_dcache_page( ) is MM hook to ensure that a page has consistent views between kernel and userspace. Thus it is called when * kernel writes to a page which at some later point could get mapped to userspace (so kernel mapping needs to be flushed-n-inv) * kernel is about to read from a page with possible userspace mappings (so userspace mappings needs to be made coherent with kernel ones) However for Non aliasing VIPT dcache, any userspace mapping will always be congruent to kernel mapping. Thus d-cache need need not be flushed at all (or delayed indefinitely). The only reason it does need to be flushed is when mapping code pages. Since icache doesn't snoop dcache, those dirty dcache lines need to be written back to memory and icache line invalidated so that icache lines fetch will get the right data. Decent gains on LMBench fork/exec/sh and File I/O micro-benchmarks. (1) FPGA @ 80 MHZ Processor, Processes - times in microseconds - smaller is better ------------------------------------------------------------------------------ Host OS Mhz null null open slct sig sig fork exec sh call I/O stat clos TCP inst hndl proc proc proc --------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- 3.9-rc6-a Linux 3.9.0-r 80 4.79 8.72 66.7 116. 239. 8.39 30.4 4798 14.K 34.K 3.9-rc6-b Linux 3.9.0-r 80 4.79 8.62 65.4 111. 239. 8.35 29.0 3995 12.K 30.K 3.9-rc7-c Linux 3.9.0-r 80 4.79 9.00 66.1 106. 239. 8.61 30.4 2858 10.K 24.K ^^^^ ^^^^ ^^^ File & VM system latencies in microseconds - smaller is better ------------------------------------------------------------------------------- Host OS 0K File 10K File Mmap Prot Page 100fd Create Delete Create Delete Latency Fault Fault selct --------- ------------- ------ ------ ------ ------ ------- ----- ------- ----- 3.9-rc6-a Linux 3.9.0-r 317.8 204.2 1122.3 375.1 3522.0 4.288 20.7 126.8 3.9-rc6-b Linux 3.9.0-r 298.7 223.0 1141.6 367.8 3531.0 4.866 20.9 126.4 3.9-rc7-c Linux 3.9.0-r 278.4 179.2 862.1 339.3 3705.0 3.223 20.3 126.6 ^^^^^ ^^^^^ ^^^^^ ^^^^ (2) Customer Silicon @ 500 MHz (166 MHz mem) ------------------------------------------------------------------------------ Host OS Mhz null null open slct sig sig fork exec sh call I/O stat clos TCP inst hndl proc proc proc --------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- abilis-ba Linux 3.9.0-r 497 0.71 1.38 4.58 12.0 35.5 1.40 3.89 2070 5525 13.K abilis-ca Linux 3.9.0-r 497 0.71 1.40 4.61 11.8 35.6 1.37 3.92 1411 4317 10.K ^^^^ ^^^^ ^^^ Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
2013-04-16 16:40:48 +08:00
int dirty = !test_and_set_bit(PG_dc_clean, &page->flags);
ARC: [mm] Lazy D-cache flush (non aliasing VIPT) flush_dcache_page( ) is MM hook to ensure that a page has consistent views between kernel and userspace. Thus it is called when * kernel writes to a page which at some later point could get mapped to userspace (so kernel mapping needs to be flushed-n-inv) * kernel is about to read from a page with possible userspace mappings (so userspace mappings needs to be made coherent with kernel ones) However for Non aliasing VIPT dcache, any userspace mapping will always be congruent to kernel mapping. Thus d-cache need need not be flushed at all (or delayed indefinitely). The only reason it does need to be flushed is when mapping code pages. Since icache doesn't snoop dcache, those dirty dcache lines need to be written back to memory and icache line invalidated so that icache lines fetch will get the right data. Decent gains on LMBench fork/exec/sh and File I/O micro-benchmarks. (1) FPGA @ 80 MHZ Processor, Processes - times in microseconds - smaller is better ------------------------------------------------------------------------------ Host OS Mhz null null open slct sig sig fork exec sh call I/O stat clos TCP inst hndl proc proc proc --------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- 3.9-rc6-a Linux 3.9.0-r 80 4.79 8.72 66.7 116. 239. 8.39 30.4 4798 14.K 34.K 3.9-rc6-b Linux 3.9.0-r 80 4.79 8.62 65.4 111. 239. 8.35 29.0 3995 12.K 30.K 3.9-rc7-c Linux 3.9.0-r 80 4.79 9.00 66.1 106. 239. 8.61 30.4 2858 10.K 24.K ^^^^ ^^^^ ^^^ File & VM system latencies in microseconds - smaller is better ------------------------------------------------------------------------------- Host OS 0K File 10K File Mmap Prot Page 100fd Create Delete Create Delete Latency Fault Fault selct --------- ------------- ------ ------ ------ ------ ------- ----- ------- ----- 3.9-rc6-a Linux 3.9.0-r 317.8 204.2 1122.3 375.1 3522.0 4.288 20.7 126.8 3.9-rc6-b Linux 3.9.0-r 298.7 223.0 1141.6 367.8 3531.0 4.866 20.9 126.4 3.9-rc7-c Linux 3.9.0-r 278.4 179.2 862.1 339.3 3705.0 3.223 20.3 126.6 ^^^^^ ^^^^^ ^^^^^ ^^^^ (2) Customer Silicon @ 500 MHz (166 MHz mem) ------------------------------------------------------------------------------ Host OS Mhz null null open slct sig sig fork exec sh call I/O stat clos TCP inst hndl proc proc proc --------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- abilis-ba Linux 3.9.0-r 497 0.71 1.38 4.58 12.0 35.5 1.40 3.89 2070 5525 13.K abilis-ca Linux 3.9.0-r 497 0.71 1.40 4.61 11.8 35.6 1.37 3.92 1411 4317 10.K ^^^^ ^^^^ ^^^ Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
2013-04-16 16:40:48 +08:00
if (dirty) {
/* wback + inv dcache lines */
__flush_dcache_page(paddr, paddr);
/* invalidate any existing icache lines */
if (vma->vm_flags & VM_EXEC)
__inv_icache_page(paddr, vaddr);
ARC: [mm] Lazy D-cache flush (non aliasing VIPT) flush_dcache_page( ) is MM hook to ensure that a page has consistent views between kernel and userspace. Thus it is called when * kernel writes to a page which at some later point could get mapped to userspace (so kernel mapping needs to be flushed-n-inv) * kernel is about to read from a page with possible userspace mappings (so userspace mappings needs to be made coherent with kernel ones) However for Non aliasing VIPT dcache, any userspace mapping will always be congruent to kernel mapping. Thus d-cache need need not be flushed at all (or delayed indefinitely). The only reason it does need to be flushed is when mapping code pages. Since icache doesn't snoop dcache, those dirty dcache lines need to be written back to memory and icache line invalidated so that icache lines fetch will get the right data. Decent gains on LMBench fork/exec/sh and File I/O micro-benchmarks. (1) FPGA @ 80 MHZ Processor, Processes - times in microseconds - smaller is better ------------------------------------------------------------------------------ Host OS Mhz null null open slct sig sig fork exec sh call I/O stat clos TCP inst hndl proc proc proc --------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- 3.9-rc6-a Linux 3.9.0-r 80 4.79 8.72 66.7 116. 239. 8.39 30.4 4798 14.K 34.K 3.9-rc6-b Linux 3.9.0-r 80 4.79 8.62 65.4 111. 239. 8.35 29.0 3995 12.K 30.K 3.9-rc7-c Linux 3.9.0-r 80 4.79 9.00 66.1 106. 239. 8.61 30.4 2858 10.K 24.K ^^^^ ^^^^ ^^^ File & VM system latencies in microseconds - smaller is better ------------------------------------------------------------------------------- Host OS 0K File 10K File Mmap Prot Page 100fd Create Delete Create Delete Latency Fault Fault selct --------- ------------- ------ ------ ------ ------ ------- ----- ------- ----- 3.9-rc6-a Linux 3.9.0-r 317.8 204.2 1122.3 375.1 3522.0 4.288 20.7 126.8 3.9-rc6-b Linux 3.9.0-r 298.7 223.0 1141.6 367.8 3531.0 4.866 20.9 126.4 3.9-rc7-c Linux 3.9.0-r 278.4 179.2 862.1 339.3 3705.0 3.223 20.3 126.6 ^^^^^ ^^^^^ ^^^^^ ^^^^ (2) Customer Silicon @ 500 MHz (166 MHz mem) ------------------------------------------------------------------------------ Host OS Mhz null null open slct sig sig fork exec sh call I/O stat clos TCP inst hndl proc proc proc --------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- abilis-ba Linux 3.9.0-r 497 0.71 1.38 4.58 12.0 35.5 1.40 3.89 2070 5525 13.K abilis-ca Linux 3.9.0-r 497 0.71 1.40 4.61 11.8 35.6 1.37 3.92 1411 4317 10.K ^^^^ ^^^^ ^^^ Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
2013-04-16 16:40:48 +08:00
}
}
}
/* Read the Cache Build Confuration Registers, Decode them and save into
* the cpuinfo structure for later use.
* No Validation is done here, simply read/convert the BCRs
*/
void read_decode_mmu_bcr(void)
{
struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
unsigned int tmp;
struct bcr_mmu_1_2 {
#ifdef CONFIG_CPU_BIG_ENDIAN
unsigned int ver:8, ways:4, sets:4, u_itlb:8, u_dtlb:8;
#else
unsigned int u_dtlb:8, u_itlb:8, sets:4, ways:4, ver:8;
#endif
} *mmu2;
struct bcr_mmu_3 {
#ifdef CONFIG_CPU_BIG_ENDIAN
unsigned int ver:8, ways:4, sets:4, osm:1, reserv:3, pg_sz:4,
u_itlb:4, u_dtlb:4;
#else
unsigned int u_dtlb:4, u_itlb:4, pg_sz:4, reserv:3, osm:1, sets:4,
ways:4, ver:8;
#endif
} *mmu3;
tmp = read_aux_reg(ARC_REG_MMU_BCR);
mmu->ver = (tmp >> 24);
if (mmu->ver <= 2) {
mmu2 = (struct bcr_mmu_1_2 *)&tmp;
mmu->pg_sz = PAGE_SIZE;
mmu->sets = 1 << mmu2->sets;
mmu->ways = 1 << mmu2->ways;
mmu->u_dtlb = mmu2->u_dtlb;
mmu->u_itlb = mmu2->u_itlb;
} else {
mmu3 = (struct bcr_mmu_3 *)&tmp;
mmu->pg_sz = 512 << mmu3->pg_sz;
mmu->sets = 1 << mmu3->sets;
mmu->ways = 1 << mmu3->ways;
mmu->u_dtlb = mmu3->u_dtlb;
mmu->u_itlb = mmu3->u_itlb;
}
mmu->num_tlb = mmu->sets * mmu->ways;
}
char *arc_mmu_mumbojumbo(int cpu_id, char *buf, int len)
{
int n = 0;
struct cpuinfo_arc_mmu *p_mmu = &cpuinfo_arc700[cpu_id].mmu;
n += scnprintf(buf + n, len - n,
"MMU [v%x]\t: %dk PAGE, JTLB %d (%dx%d), uDTLB %d, uITLB %d %s\n",
p_mmu->ver, TO_KB(p_mmu->pg_sz),
p_mmu->num_tlb, p_mmu->sets, p_mmu->ways,
p_mmu->u_dtlb, p_mmu->u_itlb,
IS_ENABLED(CONFIG_ARC_MMU_SASID) ? ",SASID" : "");
return buf;
}
void arc_mmu_init(void)
{
char str[256];
struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
printk(arc_mmu_mumbojumbo(0, str, sizeof(str)));
/* For efficiency sake, kernel is compile time built for a MMU ver
* This must match the hardware it is running on.
* Linux built for MMU V2, if run on MMU V1 will break down because V1
* hardware doesn't understand cmds such as WriteNI, or IVUTLB
* On the other hand, Linux built for V1 if run on MMU V2 will do
* un-needed workarounds to prevent memcpy thrashing.
* Similarly MMU V3 has new features which won't work on older MMU
*/
if (mmu->ver != CONFIG_ARC_MMU_VER) {
panic("MMU ver %d doesn't match kernel built for %d...\n",
mmu->ver, CONFIG_ARC_MMU_VER);
}
if (mmu->pg_sz != PAGE_SIZE)
panic("MMU pg size != PAGE_SIZE (%luk)\n", TO_KB(PAGE_SIZE));
/* Enable the MMU */
write_aux_reg(ARC_REG_PID, MMU_ENABLE);
/* In smp we use this reg for interrupt 1 scratch */
#ifndef CONFIG_SMP
/* swapper_pg_dir is the pgd for the kernel, used by vmalloc */
write_aux_reg(ARC_REG_SCRATCH_DATA0, swapper_pg_dir);
#endif
}
/*
* TLB Programmer's Model uses Linear Indexes: 0 to {255, 511} for 128 x {2,4}
* The mapping is Column-first.
* --------------------- -----------
* |way0|way1|way2|way3| |way0|way1|
* --------------------- -----------
* [set0] | 0 | 1 | 2 | 3 | | 0 | 1 |
* [set1] | 4 | 5 | 6 | 7 | | 2 | 3 |
* ~ ~ ~ ~
* [set127] | 508| 509| 510| 511| | 254| 255|
* --------------------- -----------
* For normal operations we don't(must not) care how above works since
* MMU cmd getIndex(vaddr) abstracts that out.
* However for walking WAYS of a SET, we need to know this
*/
#define SET_WAY_TO_IDX(mmu, set, way) ((set) * mmu->ways + (way))
/* Handling of Duplicate PD (TLB entry) in MMU.
* -Could be due to buggy customer tapeouts or obscure kernel bugs
* -MMU complaints not at the time of duplicate PD installation, but at the
* time of lookup matching multiple ways.
* -Ideally these should never happen - but if they do - workaround by deleting
* the duplicate one.
* -Knob to be verbose abt it.(TODO: hook them up to debugfs)
*/
volatile int dup_pd_verbose = 1;/* Be slient abt it or complain (default) */
void do_tlb_overlap_fault(unsigned long cause, unsigned long address,
struct pt_regs *regs)
{
int set, way, n;
unsigned long flags, is_valid;
struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
unsigned int pd0[mmu->ways], pd1[mmu->ways];
local_irq_save(flags);
/* re-enable the MMU */
write_aux_reg(ARC_REG_PID, MMU_ENABLE | read_aux_reg(ARC_REG_PID));
/* loop thru all sets of TLB */
for (set = 0; set < mmu->sets; set++) {
/* read out all the ways of current set */
for (way = 0, is_valid = 0; way < mmu->ways; way++) {
write_aux_reg(ARC_REG_TLBINDEX,
SET_WAY_TO_IDX(mmu, set, way));
write_aux_reg(ARC_REG_TLBCOMMAND, TLBRead);
pd0[way] = read_aux_reg(ARC_REG_TLBPD0);
pd1[way] = read_aux_reg(ARC_REG_TLBPD1);
is_valid |= pd0[way] & _PAGE_PRESENT;
}
/* If all the WAYS in SET are empty, skip to next SET */
if (!is_valid)
continue;
/* Scan the set for duplicate ways: needs a nested loop */
for (way = 0; way < mmu->ways - 1; way++) {
if (!pd0[way])
continue;
for (n = way + 1; n < mmu->ways; n++) {
if ((pd0[way] & PAGE_MASK) ==
(pd0[n] & PAGE_MASK)) {
if (dup_pd_verbose) {
pr_info("Duplicate PD's @"
"[%d:%d]/[%d:%d]\n",
set, way, set, n);
pr_info("TLBPD0[%u]: %08x\n",
way, pd0[way]);
}
/*
* clear entry @way and not @n. This is
* critical to our optimised loop
*/
pd0[way] = pd1[way] = 0;
write_aux_reg(ARC_REG_TLBINDEX,
SET_WAY_TO_IDX(mmu, set, way));
__tlb_entry_erase();
}
}
}
}
local_irq_restore(flags);
}
/***********************************************************************
* Diagnostic Routines
* -Called from Low Level TLB Hanlders if things don;t look good
**********************************************************************/
#ifdef CONFIG_ARC_DBG_TLB_PARANOIA
/*
* Low Level ASM TLB handler calls this if it finds that HW and SW ASIDS
* don't match
*/
void print_asid_mismatch(int mm_asid, int mmu_asid, int is_fast_path)
{
pr_emerg("ASID Mismatch in %s Path Handler: sw-pid=0x%x hw-pid=0x%x\n",
is_fast_path ? "Fast" : "Slow", mm_asid, mmu_asid);
__asm__ __volatile__("flag 1");
}
void tlb_paranoid_check(unsigned int mm_asid, unsigned long addr)
{
unsigned int mmu_asid;
mmu_asid = read_aux_reg(ARC_REG_PID) & 0xff;
/*
* At the time of a TLB miss/installation
* - HW version needs to match SW version
* - SW needs to have a valid ASID
*/
if (addr < 0x70000000 &&
((mm_asid == MM_CTXT_NO_ASID) ||
(mmu_asid != (mm_asid & MM_CTXT_ASID_MASK))))
print_asid_mismatch(mm_asid, mmu_asid, 0);
}
#endif