/* * Utility Routine to erase a J-TLB entry * Caller needs to setup Index Reg (manually or via getIndex)
*/ staticinlinevoid __tlb_entry_erase(void)
{
write_aux_reg(ARC_REG_TLBPD1, 0);
if (is_pae40_enabled())
write_aux_reg(ARC_REG_TLBPD1HI, 0);
/* 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);
}
}
/* * 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 doesn't 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);
}
/* * Flush the entire 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 explicitly
*/ void local_flush_tlb_range(struct vm_area_struct *vma, unsignedlong start, unsignedlong end)
{ constunsignedint cpu = smp_processor_id(); unsignedlong flags;
/* If range @start to @end is more than 32 TLB entries deep, * it's 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;
}
}
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(unsignedlong start, unsignedlong end)
{ unsignedlong flags;
/* exactly same as above, except for TLB entry not taking ASID */
/* * 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. * -More importantly it makes this handler inconsistent with fast-path * TLB Refill handler which always deals with "current" * * Let's 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;
/* * 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 */
/* * 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_range(struct vm_fault *vmf, struct vm_area_struct *vma, unsignedlong vaddr_unaligned, pte_t *ptep, unsignedint nr)
{ unsignedlong vaddr = vaddr_unaligned & PAGE_MASK;
phys_addr_t paddr = pte_val(*ptep) & PAGE_MASK_PHYS; struct page *page = pfn_to_page(pte_pfn(*ptep));
create_tlb(vma, vaddr, ptep);
if (page == ZERO_PAGE(0)) return;
/* * For executable pages, since icache doesn't snoop dcache, any * dirty K-mapping of a code page needs to be wback+inv so that * icache fetch by userspace sees code correctly.
*/ if (vma->vm_flags & VM_EXEC) { struct folio *folio = page_folio(page); int dirty = !test_and_set_bit(PG_dc_clean, &folio->flags); if (dirty) { unsignedlong offset = offset_in_folio(folio, paddr);
nr = folio_nr_pages(folio);
paddr -= offset;
vaddr -= offset; /* wback + inv dcache lines (K-mapping) */
__flush_dcache_pages(paddr, paddr, nr);
/* invalidate any existing icache lines (U-mapping) */ if (vma->vm_flags & VM_EXEC)
__inv_icache_pages(paddr, vaddr, nr);
}
}
}
#ifdef CONFIG_TRANSPARENT_HUGEPAGE
/* * MMUv4 in HS38x cores supports Super Pages which are basis for Linux THP * support. * * Normal and Super pages can co-exist (ofcourse not overlap) in TLB with a * new bit "SZ" in TLB page descriptor to distinguish between them. * Super Page size is configurable in hardware (4K to 16M), but fixed once * RTL builds. * * The exact THP size a Linux configuration will support is a function of: * - MMU page size (typical 8K, RTL fixed) * - software page walker address split between PGD:PTE:PFN (typical * 11:8:13, but can be changed with 1 line) * So for above default, THP size supported is 8K * (2^8) = 2M * * Default Page Walker is 2 levels, PGD:PTE:PFN, which in THP regime * reduces to 1 level (as PTE is folded into PGD and canonically referred * to as PMD). * Thus THP PMD accessors are implemented in terms of PTE (just like sparc)
*/
/* No need to loop here: this will always be for 1 Huge Page */
tlb_entry_erase(start | _PAGE_HW_SZ | asid);
}
local_irq_restore(flags);
}
#endif
/* Read the Cache Build Configuration Registers, Decode them and save into * the cpuinfo structure for later use. * No Validation is done here, simply read/convert the BCRs
*/ int arc_mmu_mumbojumbo(int c, char *buf, int len)
{ struct cpuinfo_arc_mmu *mmu = &mmuinfo; unsignedint bcr, u_dtlb, u_itlb, sasid; struct bcr_mmu_3 *mmu3; struct bcr_mmu_4 *mmu4; char super_pg[64] = ""; int n = 0;
/* * Can't be done in processor.h due to header include dependencies
*/
BUILD_BUG_ON(!IS_ALIGNED((CONFIG_ARC_KVADDR_SIZE << 20), PMD_SIZE));
/* * stack top size sanity check, * Can't be done in processor.h due to header include dependencies
*/
BUILD_BUG_ON(!IS_ALIGNED(STACK_TOP, PMD_SIZE));
/* * Ensure that MMU features assumed by kernel exist in hardware. * - For older ARC700 cpus, only v3 supported * - For HS cpus, v4 was baseline and v5 is backwards compatible * (will run older software).
*/ if (is_isa_arcompact() && mmu->ver == 3)
compat = 1; elseif (is_isa_arcv2() && mmu->ver >= 4)
compat = 1;
if (!compat)
panic("MMU ver %d doesn't match kernel built for\n", mmu->ver);
if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
mmu->s_pg_sz_m != TO_MB(HPAGE_PMD_SIZE))
panic("MMU Super pg size != Linux HPAGE_PMD_SIZE (%luM)\n",
(unsignedlong)TO_MB(HPAGE_PMD_SIZE));
if (IS_ENABLED(CONFIG_ARC_HAS_PAE40) && !mmu->pae)
panic("Hardware doesn't support PAE40\n");
/* Enable the MMU with ASID 0 */
mmu_setup_asid(NULL, 0);
/* cache the pgd pointer in MMU SCRATCH reg (ARCv2 only) */
mmu_setup_pgd(NULL, swapper_pg_dir);
if (pae40_exist_but_not_enab())
write_aux_reg(ARC_REG_TLBPD1HI, 0);
}
/* * 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)
*/ volatileint dup_pd_silent; /* Be silent abt it or complain (default) */
/* loop thru all sets of TLB */ for (set = 0; set < mmu->sets; set++) {
int is_valid, way; unsignedint pd0[4];
/* read out all the ways of current set */ for (way = 0, is_valid = 0; way < n_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);
is_valid |= pd0[way] & _PAGE_PRESENT;
pd0[way] &= PAGE_MASK;
}
/* 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 < n_ways - 1; way++) {
int n;
if (!pd0[way]) continue;
for (n = way + 1; n < n_ways; n++) { if (pd0[way] != pd0[n]) continue;
if (!dup_pd_silent)
pr_info("Dup TLB PD0 %08x @ set %d ways %d,%d\n",
pd0[way], set, way, n);
/* * clear entry @way and not @n. * This is critical to our optimised loop
*/
pd0[way] = 0;
write_aux_reg(ARC_REG_TLBINDEX,
SET_WAY_TO_IDX(mmu, set, way));
__tlb_entry_erase();
}
}
}
local_irq_restore(flags);
}
Messung V0.5
¤ Dauer der Verarbeitung: 0.13 Sekunden
(vorverarbeitet)
¤
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.