Date: Tue, 19 Jul 2022 09:57:50 GMT From: Andrew Turner <andrew@FreeBSD.org> To: src-committers@FreeBSD.org, dev-commits-src-all@FreeBSD.org, dev-commits-src-main@FreeBSD.org Subject: git: 36f1526a598c - main - Add experimental 16k page support on arm64 Message-ID: <202207190957.26J9voj4012964@gitrepo.freebsd.org>
next in thread | raw e-mail | index | archive | help
The branch main has been updated by andrew: URL: https://cgit.FreeBSD.org/src/commit/?id=36f1526a598c373ca660910c9772d28a61383c3b commit 36f1526a598c373ca660910c9772d28a61383c3b Author: Andrew Turner <andrew@FreeBSD.org> AuthorDate: 2022-03-23 17:39:58 +0000 Commit: Andrew Turner <andrew@FreeBSD.org> CommitDate: 2022-07-19 09:57:03 +0000 Add experimental 16k page support on arm64 Add initial 16k page support on arm64. It is considered experimental, with no guarantee of compatibility with a userspace or kernel modules built with the current a 4k page size as code will likely try to pass in a too small size when working with APIs that take a multiple of a page, e.g. mmap. As this is experimental, and because userspace and the kernel need to have the PAGE_SIZE macro kept in sync there is no kernel option to enable this. To test a new image should be built with the PAGE_{SIZE,SHIFT,MASK} macros changed to the 16k versions. There are currently known issues with loading modules from an old loader as it can misalign them to load on a non-16k boundary. Testing has shown good results in kernel workloads that allocate and free large amounts of memory as only a quarter of the number of calls into the VM subsystem are needed in the best case. Reviewed by: markj Tested by: gallatin Sponsored by: The FreeBSD Foundation Differential Revision: https://reviews.freebsd.org/D34793 --- sys/arm64/arm64/locore.S | 136 ++++++++++++++++++++++++++++++++-- sys/arm64/arm64/minidump_machdep.c | 6 ++ sys/arm64/arm64/pmap.c | 146 +++++++++++++++++++++++++------------ sys/arm64/include/pmap.h | 11 +++ sys/arm64/include/pte.h | 42 ++++++++--- 5 files changed, 277 insertions(+), 64 deletions(-) diff --git a/sys/arm64/arm64/locore.S b/sys/arm64/arm64/locore.S index ba85bb4e46b2..518c6c812aa9 100644 --- a/sys/arm64/arm64/locore.S +++ b/sys/arm64/arm64/locore.S @@ -39,6 +39,14 @@ #define VIRT_BITS 48 +#if PAGE_SIZE == PAGE_SIZE_16K +/* + * The number of level 3 tables to create. 32 will allow for 1G of address + * space, the same as a single level 2 page with 4k pages. + */ +#define L3_PAGE_COUNT 32 +#endif + .globl kernbase .set kernbase, KERNBASE @@ -431,8 +439,13 @@ LENTRY(create_pagetables) /* Booted with modules pointer */ /* Find modulep - begin */ sub x8, x0, x6 - /* Add two 2MiB pages for the module data and round up */ - ldr x7, =(3 * L2_SIZE - 1) + /* + * Add space for the module data. When PAGE_SIZE is 4k this will + * add at least 2 level 2 blocks (2 * 2MiB). When PAGE_SIZE is + * larger it will be at least as large as we use smaller level 3 + * pages. + */ + ldr x7, =((6 * 1024 * 1024) - 1) add x8, x8, x7 b common @@ -457,6 +470,34 @@ booti_no_fdt: #endif common: +#if PAGE_SIZE != PAGE_SIZE_4K + /* + * Create L3 pages. The kernel will be loaded at a 2M aligned + * address, however L2 blocks are too large when the page size is + * not 4k to map the kernel with such an aligned address. However, + * when the page size is larger than 4k, L2 blocks are too large to + * map the kernel with such an alignment. + */ + + /* Get the number of l3 pages to allocate, rounded down */ + lsr x10, x8, #(L3_SHIFT) + + /* Create the kernel space L2 table */ + mov x6, x26 + mov x7, #(ATTR_S1_IDX(VM_MEMATTR_WRITE_BACK)) + mov x8, #(KERNBASE) + mov x9, x28 + bl build_l3_page_pagetable + + /* Move to the l2 table */ + ldr x9, =(PAGE_SIZE * L3_PAGE_COUNT) + add x26, x26, x9 + + /* Link the l2 -> l3 table */ + mov x9, x6 + mov x6, x26 + bl link_l2_pagetable +#else /* Get the number of l2 pages to allocate, rounded down */ lsr x10, x8, #(L2_SHIFT) @@ -466,6 +507,7 @@ common: mov x8, #(KERNBASE) mov x9, x28 bl build_l2_block_pagetable +#endif /* Move to the l1 table */ add x26, x26, #PAGE_SIZE @@ -504,7 +546,8 @@ common: #if defined(SOCDEV_PA) /* Create a table for the UART */ mov x7, #(ATTR_S1_nG | ATTR_S1_IDX(VM_MEMATTR_DEVICE)) - add x16, x16, #(L2_SIZE) /* VA start */ + ldr x9, =(L2_SIZE) + add x16, x16, x9 /* VA start */ mov x8, x16 /* Store the socdev virtual address */ @@ -523,7 +566,8 @@ common: /* Create the mapping for FDT data (2 MiB max) */ mov x7, #(ATTR_S1_nG | ATTR_S1_IDX(VM_MEMATTR_WRITE_BACK)) - add x16, x16, #(L2_SIZE) /* VA start */ + ldr x9, =(L2_SIZE) + add x16, x16, x9 /* VA start */ mov x8, x16 mov x9, x0 /* PA start */ /* Update the module pointer to point at the allocated memory */ @@ -662,6 +706,76 @@ LENTRY(build_l2_block_pagetable) ret LEND(build_l2_block_pagetable) +#if PAGE_SIZE != PAGE_SIZE_4K +/* + * Builds an L2 -> L3 table descriptor + * + * x6 = L2 table + * x8 = Virtual Address + * x9 = L3 PA (trashed) + * x11, x12 and x13 are trashed + */ +LENTRY(link_l2_pagetable) + /* + * Link an L2 -> L3 table entry. + */ + /* Find the table index */ + lsr x11, x8, #L2_SHIFT + and x11, x11, #Ln_ADDR_MASK + + /* Build the L1 block entry */ + mov x12, #L2_TABLE + + /* Only use the output address bits */ + lsr x9, x9, #PAGE_SHIFT + orr x13, x12, x9, lsl #PAGE_SHIFT + + /* Store the entry */ + str x13, [x6, x11, lsl #3] + + ret +LEND(link_l2_pagetable) + +/* + * Builds count level 3 page table entries + * x6 = L3 table + * x7 = Block attributes + * x8 = VA start + * x9 = PA start (trashed) + * x10 = Entry count (trashed) + * x11, x12 and x13 are trashed + */ +LENTRY(build_l3_page_pagetable) + /* + * Build the L3 table entry. + */ + /* Find the table index */ + lsr x11, x8, #L3_SHIFT + and x11, x11, #Ln_ADDR_MASK + + /* Build the L3 page entry */ + orr x12, x7, #L3_PAGE + orr x12, x12, #(ATTR_DEFAULT) + orr x12, x12, #(ATTR_S1_UXN) + + /* Only use the output address bits */ + lsr x9, x9, #L3_SHIFT + + /* Set the physical address for this virtual address */ +1: orr x13, x12, x9, lsl #L3_SHIFT + + /* Store the entry */ + str x13, [x6, x11, lsl #3] + + sub x10, x10, #1 + add x11, x11, #1 + add x9, x9, #1 + cbnz x10, 1b + + ret +LEND(build_l3_page_pagetable) +#endif + LENTRY(start_mmu) dsb sy @@ -743,7 +857,15 @@ mair: MAIR_ATTR(MAIR_NORMAL_WT, VM_MEMATTR_WRITE_THROUGH) | \ MAIR_ATTR(MAIR_DEVICE_nGnRE, VM_MEMATTR_DEVICE_nGnRE) tcr: - .quad (TCR_TxSZ(64 - VIRT_BITS) | TCR_TG1_4K | TCR_TG0_4K | \ +#if PAGE_SIZE == PAGE_SIZE_4K +#define TCR_TG (TCR_TG1_4K | TCR_TG0_4K) +#elif PAGE_SIZE == PAGE_SIZE_16K +#define TCR_TG (TCR_TG1_16K | TCR_TG0_16K) +#else +#error Unsupported page size +#endif + + .quad (TCR_TxSZ(64 - VIRT_BITS) | TCR_TG | \ TCR_CACHE_ATTRS | TCR_SMP_ATTRS) sctlr_set: /* Bits to set */ @@ -774,6 +896,10 @@ END(abort) */ .globl pagetable_l0_ttbr1 pagetable: +#if PAGE_SIZE != PAGE_SIZE_4K + .space (PAGE_SIZE * L3_PAGE_COUNT) +pagetable_l2_ttbr1: +#endif .space PAGE_SIZE pagetable_l1_ttbr1: .space PAGE_SIZE diff --git a/sys/arm64/arm64/minidump_machdep.c b/sys/arm64/arm64/minidump_machdep.c index 3dfeb3dfef1e..ee2b1be9b0b1 100644 --- a/sys/arm64/arm64/minidump_machdep.c +++ b/sys/arm64/arm64/minidump_machdep.c @@ -239,7 +239,13 @@ cpu_minidumpsys(struct dumperinfo *di, const struct minidumpstate *state) mdhdr.dmapbase = DMAP_MIN_ADDRESS; mdhdr.dmapend = DMAP_MAX_ADDRESS; mdhdr.dumpavailsize = round_page(sizeof(dump_avail)); +#if PAGE_SIZE == PAGE_SIZE_4K mdhdr.flags = MINIDUMP_FLAG_PS_4K; +#elif PAGE_SIZE == PAGE_SIZE_16K + mdhdr.flags = MINIDUMP_FLAG_PS_16K; +#else +#error Unsupported page size +#endif dump_init_header(di, &kdh, KERNELDUMPMAGIC, KERNELDUMP_AARCH64_VERSION, dumpsize); diff --git a/sys/arm64/arm64/pmap.c b/sys/arm64/arm64/pmap.c index 3b37ce214664..d95eccb445a5 100644 --- a/sys/arm64/arm64/pmap.c +++ b/sys/arm64/arm64/pmap.c @@ -286,10 +286,6 @@ vm_paddr_t dmap_phys_base; /* The start of the dmap region */ vm_paddr_t dmap_phys_max; /* The limit of the dmap region */ vm_offset_t dmap_max_addr; /* The virtual address limit of the dmap */ -/* This code assumes all L1 DMAP entries will be used */ -CTASSERT((DMAP_MIN_ADDRESS & ~L0_OFFSET) == DMAP_MIN_ADDRESS); -CTASSERT((DMAP_MAX_ADDRESS & ~L0_OFFSET) == DMAP_MAX_ADDRESS); - extern pt_entry_t pagetable_l0_ttbr1[]; #define PHYSMAP_SIZE (2 * (VM_PHYSSEG_MAX - 1)) @@ -299,6 +295,15 @@ static u_int physmap_idx; static SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "VM/pmap parameters"); +#if PAGE_SIZE == PAGE_SIZE_4K +#define L1_BLOCKS_SUPPORTED 1 +#else +/* TODO: Make this dynamic when we support FEAT_LPA2 (TCR_EL1.DS == 1) */ +#define L1_BLOCKS_SUPPORTED 0 +#endif + +#define PMAP_ASSERT_L1_BLOCKS_SUPPORTED MPASS(L1_BLOCKS_SUPPORTED) + /* * This ASID allocator uses a bit vector ("asid_set") to remember which ASIDs * that it has currently allocated to a pmap, a cursor ("asid_next") to @@ -571,6 +576,7 @@ pmap_pte(pmap_t pmap, vm_offset_t va, int *level) } desc = pmap_load(l1) & ATTR_DESCR_MASK; if (desc == L1_BLOCK) { + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; *level = 1; return (l1); } @@ -621,9 +627,11 @@ pmap_pte_exists(pmap_t pmap, vm_offset_t va, int level, const char *diag) if (desc == L0_TABLE && level > 0) { l1p = pmap_l0_to_l1(l0p, va); desc = pmap_load(l1p) & ATTR_DESCR_MASK; - if (desc == L1_BLOCK && level == 1) + if (desc == L1_BLOCK && level == 1) { + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; return (l1p); - else if (desc == L1_TABLE && level > 1) { + } + if (desc == L1_TABLE && level > 1) { l2p = pmap_l1_to_l2(l1p, va); desc = pmap_load(l2p) & ATTR_DESCR_MASK; if (desc == L2_BLOCK && level == 2) @@ -673,6 +681,7 @@ pmap_get_tables(pmap_t pmap, vm_offset_t va, pd_entry_t **l0, pd_entry_t **l1, *l1 = l1p; if ((pmap_load(l1p) & ATTR_DESCR_MASK) == L1_BLOCK) { + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; *l2 = NULL; *l3 = NULL; return (true); @@ -1013,29 +1022,36 @@ pmap_bootstrap_dmap(vm_offset_t kern_l1, vm_paddr_t min_pa, pmap_bootstrap_dmap_l3_page(&state, i); MPASS(state.pa <= physmap[i + 1]); - /* Create L2 mappings at the start of the region */ - if ((state.pa & L1_OFFSET) != 0) - pmap_bootstrap_dmap_l2_block(&state, i); - MPASS(state.pa <= physmap[i + 1]); + if (L1_BLOCKS_SUPPORTED) { + /* Create L2 mappings at the start of the region */ + if ((state.pa & L1_OFFSET) != 0) + pmap_bootstrap_dmap_l2_block(&state, i); + MPASS(state.pa <= physmap[i + 1]); + + /* Create the main L1 block mappings */ + for (; state.va < DMAP_MAX_ADDRESS && + (physmap[i + 1] - state.pa) >= L1_SIZE; + state.va += L1_SIZE, state.pa += L1_SIZE) { + /* Make sure there is a valid L1 table */ + pmap_bootstrap_dmap_l0_table(&state); + MPASS((state.pa & L1_OFFSET) == 0); + pmap_store(&state.l1[pmap_l1_index(state.va)], + state.pa | ATTR_DEFAULT | ATTR_S1_XN | + ATTR_S1_IDX(VM_MEMATTR_WRITE_BACK) | + L1_BLOCK); + } + MPASS(state.pa <= physmap[i + 1]); - /* Create the main L1 block mappings */ - for (; state.va < DMAP_MAX_ADDRESS && - (physmap[i + 1] - state.pa) >= L1_SIZE; - state.va += L1_SIZE, state.pa += L1_SIZE) { - /* Make sure there is a valid L1 table */ - pmap_bootstrap_dmap_l0_table(&state); - MPASS((state.pa & L1_OFFSET) == 0); - pmap_store(&state.l1[pmap_l1_index(state.va)], - state.pa | ATTR_DEFAULT | ATTR_S1_XN | - ATTR_S1_IDX(VM_MEMATTR_WRITE_BACK) | - L1_BLOCK); + /* Create L2 mappings at the end of the region */ + pmap_bootstrap_dmap_l2_block(&state, i); + } else { + while (state.va < DMAP_MAX_ADDRESS && + (physmap[i + 1] - state.pa) >= L2_SIZE) { + pmap_bootstrap_dmap_l2_block(&state, i); + } } MPASS(state.pa <= physmap[i + 1]); - /* Create L2 mappings at the end of the region */ - pmap_bootstrap_dmap_l2_block(&state, i); - MPASS(state.pa <= physmap[i + 1]); - /* Create L3 mappings at the end of the region */ pmap_bootstrap_dmap_l3_page(&state, i); MPASS(state.pa == physmap[i + 1]); @@ -1261,9 +1277,11 @@ pmap_init(void) KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0, ("pmap_init: can't assign to pagesizes[1]")); pagesizes[1] = L2_SIZE; - KASSERT(MAXPAGESIZES > 2 && pagesizes[2] == 0, - ("pmap_init: can't assign to pagesizes[2]")); - pagesizes[2] = L1_SIZE; + if (L1_BLOCKS_SUPPORTED) { + KASSERT(MAXPAGESIZES > 2 && pagesizes[2] == 0, + ("pmap_init: can't assign to pagesizes[2]")); + pagesizes[2] = L1_SIZE; + } } /* @@ -1483,6 +1501,7 @@ pmap_extract(pmap_t pmap, vm_offset_t va) pa = tpte & ~ATTR_MASK; switch(lvl) { case 1: + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT((tpte & ATTR_DESCR_MASK) == L1_BLOCK, ("pmap_extract: Invalid L1 pte found: %lx", tpte & ATTR_DESCR_MASK)); @@ -1530,6 +1549,10 @@ pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) KASSERT(lvl > 0 && lvl <= 3, ("pmap_extract_and_hold: Invalid level %d", lvl)); + /* + * Check that the pte is either a L3 page, or a L1 or L2 block + * entry. We can assume L1_BLOCK == L2_BLOCK. + */ KASSERT((lvl == 3 && (tpte & ATTR_DESCR_MASK) == L3_PAGE) || (lvl < 3 && (tpte & ATTR_DESCR_MASK) == L1_BLOCK), ("pmap_extract_and_hold: Invalid pte at L%d: %lx", lvl, @@ -2426,8 +2449,13 @@ pmap_growkernel(vm_offset_t addr) ***************************************************/ CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE); +#if PAGE_SIZE == PAGE_SIZE_4K CTASSERT(_NPCM == 3); CTASSERT(_NPCPV == 168); +#else +CTASSERT(_NPCM == 11); +CTASSERT(_NPCPV == 677); +#endif static __inline struct pv_chunk * pv_to_chunk(pv_entry_t pv) @@ -2438,11 +2466,30 @@ pv_to_chunk(pv_entry_t pv) #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap) -#define PC_FREE0 0xfffffffffffffffful -#define PC_FREE1 0xfffffffffffffffful -#define PC_FREE2 0x000000fffffffffful +#define PC_FREEN 0xfffffffffffffffful +#if _NPCM == 3 +#define PC_FREEL 0x000000fffffffffful +#elif _NPCM == 11 +#define PC_FREEL 0x0000001ffffffffful +#endif + +#if _NPCM == 3 +#define PC_IS_FREE(pc) ((pc)->pc_map[0] == PC_FREEN && \ + (pc)->pc_map[1] == PC_FREEN && (pc)->pc_map[2] == PC_FREEL) +#else +#define PC_IS_FREE(pc) \ + (memcmp((pc)->pc_map, pc_freemask, sizeof(pc_freemask)) == 0) +#endif -static const uint64_t pc_freemask[_NPCM] = { PC_FREE0, PC_FREE1, PC_FREE2 }; +static const uint64_t pc_freemask[] = { PC_FREEN, PC_FREEN, +#if _NPCM > 3 + PC_FREEN, PC_FREEN, PC_FREEN, PC_FREEN, PC_FREEN, PC_FREEN, PC_FREEN, + PC_FREEN, +#endif + PC_FREEL +}; + +CTASSERT(nitems(pc_freemask) == _NPCM); #ifdef PV_STATS static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail; @@ -2608,8 +2655,7 @@ reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp) PV_STAT(atomic_add_int(&pv_entry_spare, freed)); PV_STAT(atomic_subtract_long(&pv_entry_count, freed)); TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); - if (pc->pc_map[0] == PC_FREE0 && pc->pc_map[1] == PC_FREE1 && - pc->pc_map[2] == PC_FREE2) { + if (PC_IS_FREE(pc)) { PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV)); PV_STAT(atomic_subtract_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_frees, 1)); @@ -2678,8 +2724,7 @@ free_pv_entry(pmap_t pmap, pv_entry_t pv) field = idx / 64; bit = idx % 64; pc->pc_map[field] |= 1ul << bit; - if (pc->pc_map[0] != PC_FREE0 || pc->pc_map[1] != PC_FREE1 || - pc->pc_map[2] != PC_FREE2) { + if (!PC_IS_FREE(pc)) { /* 98% of the time, pc is already at the head of the list. */ if (__predict_false(pc != TAILQ_FIRST(&pmap->pm_pvchunk))) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); @@ -2767,9 +2812,8 @@ retry: dump_add_page(m->phys_addr); pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; - pc->pc_map[0] = PC_FREE0 & ~1ul; /* preallocated bit 0 */ - pc->pc_map[1] = PC_FREE1; - pc->pc_map[2] = PC_FREE2; + memcpy(pc->pc_map, pc_freemask, sizeof(pc_freemask)); + pc->pc_map[0] &= ~1ul; /* preallocated bit 0 */ mtx_lock(&pv_chunks_mutex); TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru); mtx_unlock(&pv_chunks_mutex); @@ -2829,9 +2873,7 @@ retry: dump_add_page(m->phys_addr); pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; - pc->pc_map[0] = PC_FREE0; - pc->pc_map[1] = PC_FREE1; - pc->pc_map[2] = PC_FREE2; + memcpy(pc->pc_map, pc_freemask, sizeof(pc_freemask)); TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&new_tail, pc, pc_lru); PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV)); @@ -3265,6 +3307,7 @@ pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) if (pmap_load(l1) == 0) continue; if ((pmap_load(l1) & ATTR_DESCR_MASK) == L1_BLOCK) { + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT(va_next <= eva, ("partial update of non-transparent 1G page " "l1 %#lx sva %#lx eva %#lx va_next %#lx", @@ -3518,6 +3561,7 @@ pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) if (pmap_load(l1) == 0) continue; if ((pmap_load(l1) & ATTR_DESCR_MASK) == L1_BLOCK) { + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT(va_next <= eva, ("partial update of non-transparent 1G page " "l1 %#lx sva %#lx eva %#lx va_next %#lx", @@ -3848,9 +3892,10 @@ restart: mp->ref_count++; } } - KASSERT((origpte & ATTR_DESCR_VALID) == 0 || - ((origpte & ATTR_DESCR_MASK) == L1_BLOCK && - (origpte & ~ATTR_MASK) == (newpte & ~ATTR_MASK)), + KASSERT((origpte & ~ATTR_MASK) == (newpte & ~ATTR_MASK) || + (L1_BLOCKS_SUPPORTED && + (origpte & ATTR_DESCR_MASK) == L1_BLOCK && + (origpte & ATTR_DESCR_VALID) == 0), ("va %#lx changing 1G phys page l1 %#lx newpte %#lx", va, origpte, newpte)); pmap_store(l1p, newpte); @@ -3980,9 +4025,10 @@ pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed largepage va %#lx flags %#x", va, flags)); new_l3 &= ~L3_PAGE; - if (psind == 2) + if (psind == 2) { + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; new_l3 |= L1_BLOCK; - else /* (psind == 1) */ + } else /* (psind == 1) */ new_l3 |= L2_BLOCK; rv = pmap_enter_largepage(pmap, va, new_l3, flags, psind); goto out; @@ -4660,6 +4706,7 @@ pmap_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) continue; if ((pmap_load(l1) & ATTR_DESCR_MASK) == L1_BLOCK) { + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT(va_next <= eva, ("partial update of non-transparent 1G page " "l1 %#lx sva %#lx eva %#lx va_next %#lx", @@ -4772,6 +4819,7 @@ pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, if (pmap_load(l1) == 0) continue; if ((pmap_load(l1) & ATTR_DESCR_MASK) == L1_BLOCK) { + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT(va_next <= end_addr, ("partial update of non-transparent 1G page " "l1 %#lx addr %#lx end_addr %#lx va_next %#lx", @@ -5730,6 +5778,7 @@ pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice) if (pmap_load(l1) == 0) continue; if ((pmap_load(l1) & ATTR_DESCR_MASK) == L1_BLOCK) { + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT(va_next <= eva, ("partial update of non-transparent 1G page " "l1 %#lx sva %#lx eva %#lx va_next %#lx", @@ -6243,6 +6292,7 @@ pmap_change_props_locked(vm_offset_t va, vm_size_t size, vm_prot_t prot, default: panic("Invalid DMAP table level: %d\n", lvl); case 1: + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; if ((tmpva & L1_OFFSET) == 0 && (base + size - tmpva) >= L1_SIZE) { pte_size = L1_SIZE; @@ -6318,6 +6368,7 @@ pmap_demote_l1(pmap_t pmap, pt_entry_t *l1, vm_offset_t va) PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldl1 = pmap_load(l1); + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; KASSERT((oldl1 & ATTR_DESCR_MASK) == L1_BLOCK, ("pmap_demote_l1: Demoting a non-block entry")); KASSERT((va & L1_OFFSET) == 0, @@ -7400,6 +7451,7 @@ sysctl_kmaps(SYSCTL_HANDLER_ARGS) continue; } if ((l1e & ATTR_DESCR_MASK) == L1_BLOCK) { + PMAP_ASSERT_L1_BLOCKS_SUPPORTED; sysctl_kmaps_check(sb, &range, sva, l0e, l1e, 0, 0); range.l1blocks++; diff --git a/sys/arm64/include/pmap.h b/sys/arm64/include/pmap.h index 8c7c26fce8a0..87527c390f57 100644 --- a/sys/arm64/include/pmap.h +++ b/sys/arm64/include/pmap.h @@ -106,8 +106,18 @@ typedef struct pv_entry { * pv_entries are allocated in chunks per-process. This avoids the * need to track per-pmap assignments. */ +#if PAGE_SIZE == PAGE_SIZE_4K #define _NPCM 3 #define _NPCPV 168 +#define _NPAD 0 +#elif PAGE_SIZE == PAGE_SIZE_16K +#define _NPCM 11 +#define _NPCPV 677 +#define _NPAD 1 +#else +#error Unsupported page size +#endif + #define PV_CHUNK_HEADER \ pmap_t pc_pmap; \ TAILQ_ENTRY(pv_chunk) pc_list; \ @@ -121,6 +131,7 @@ struct pv_chunk_header { struct pv_chunk { PV_CHUNK_HEADER struct pv_entry pc_pventry[_NPCPV]; + uint64_t pc_pad[_NPAD]; }; struct thread; diff --git a/sys/arm64/include/pte.h b/sys/arm64/include/pte.h index 3ce11133e2ef..eaf6745f9679 100644 --- a/sys/arm64/include/pte.h +++ b/sys/arm64/include/pte.h @@ -109,33 +109,43 @@ typedef uint64_t pt_entry_t; /* page table entry */ #define ATTR_DESCR_TYPE_PAGE 2 #define ATTR_DESCR_TYPE_BLOCK 0 -/* Level 0 table, 512GiB per entry */ +#if PAGE_SIZE == PAGE_SIZE_4K #define L0_SHIFT 39 -#define L0_SIZE (1ul << L0_SHIFT) +#define L1_SHIFT 30 +#define L2_SHIFT 21 +#define L3_SHIFT 12 +#elif PAGE_SIZE == PAGE_SIZE_16K +#define L0_SHIFT 47 +#define L1_SHIFT 36 +#define L2_SHIFT 25 +#define L3_SHIFT 14 +#else +#error Unsupported page size +#endif + +/* Level 0 table, 512GiB/128TiB per entry */ +#define L0_SIZE (UINT64_C(1) << L0_SHIFT) #define L0_OFFSET (L0_SIZE - 1ul) #define L0_INVAL 0x0 /* An invalid address */ /* 0x1 Level 0 doesn't support block translation */ /* 0x2 also marks an invalid address */ #define L0_TABLE 0x3 /* A next-level table */ -/* Level 1 table, 1GiB per entry */ -#define L1_SHIFT 30 -#define L1_SIZE (1 << L1_SHIFT) +/* Level 1 table, 1GiB/64GiB per entry */ +#define L1_SIZE (UINT64_C(1) << L1_SHIFT) #define L1_OFFSET (L1_SIZE - 1) #define L1_INVAL L0_INVAL #define L1_BLOCK 0x1 #define L1_TABLE L0_TABLE -/* Level 2 table, 2MiB per entry */ -#define L2_SHIFT 21 -#define L2_SIZE (1 << L2_SHIFT) +/* Level 2 table, 2MiB/32MiB per entry */ +#define L2_SIZE (UINT64_C(1) << L2_SHIFT) #define L2_OFFSET (L2_SIZE - 1) #define L2_INVAL L1_INVAL -#define L2_BLOCK L1_BLOCK +#define L2_BLOCK 0x1 #define L2_TABLE L1_TABLE -/* Level 3 table, 4KiB per entry */ -#define L3_SHIFT 12 +/* Level 3 table, 4KiB/16KiB per entry */ #define L3_SIZE (1 << L3_SHIFT) #define L3_OFFSET (L3_SIZE - 1) #define L3_INVAL 0x0 @@ -145,11 +155,19 @@ typedef uint64_t pt_entry_t; /* page table entry */ #define PMAP_MAPDEV_EARLY_SIZE (L2_SIZE * 8) +#if PAGE_SIZE == PAGE_SIZE_4K #define L0_ENTRIES_SHIFT 9 +#define Ln_ENTRIES_SHIFT 9 +#elif PAGE_SIZE == PAGE_SIZE_16K +#define L0_ENTRIES_SHIFT 1 +#define Ln_ENTRIES_SHIFT 11 +#else +#error Unsupported page size +#endif + #define L0_ENTRIES (1 << L0_ENTRIES_SHIFT) #define L0_ADDR_MASK (L0_ENTRIES - 1) -#define Ln_ENTRIES_SHIFT 9 #define Ln_ENTRIES (1 << Ln_ENTRIES_SHIFT) #define Ln_ADDR_MASK (Ln_ENTRIES - 1) #define Ln_TABLE_MASK ((1 << 12) - 1)
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?202207190957.26J9voj4012964>