1 // SPDX-License-Identifier: GPL-2.0
3 * Implementation of Gasket page table support.
5 * Copyright (C) 2018 Google, Inc.
9 * Implementation of Gasket page table support.
11 * This file assumes 4kB pages throughout; can be factored out when necessary.
13 * There is a configurable number of page table entries, as well as a
14 * configurable bit index for the extended address flag. Both of these are
15 * specified in gasket_page_table_init through the page_table_config parameter.
17 * The following example assumes:
18 * page_table_config->total_entries = 8192
19 * page_table_config->extended_bit = 63
22 * Simple addresses - those whose containing pages are directly placed in the
23 * device's address translation registers - are laid out as:
24 * [ 63 - 25: 0 | 24 - 12: page index | 11 - 0: page offset ]
25 * page index: The index of the containing page in the device's address
26 * translation registers.
27 * page offset: The index of the address into the containing page.
29 * Extended address - those whose containing pages are contained in a second-
30 * level page table whose address is present in the device's address translation
31 * registers - are laid out as:
32 * [ 63: flag | 62 - 34: 0 | 33 - 21: dev/level 0 index |
33 * 20 - 12: host/level 1 index | 11 - 0: page offset ]
34 * flag: Marker indicating that this is an extended address. Always 1.
35 * dev index: The index of the first-level page in the device's extended
36 * address translation registers.
37 * host index: The index of the containing page in the [host-resident] second-
39 * page offset: The index of the address into the containing [second-level]
42 #include "gasket_page_table.h"
44 #include <linux/device.h>
45 #include <linux/file.h>
46 #include <linux/init.h>
47 #include <linux/kernel.h>
48 #include <linux/module.h>
49 #include <linux/moduleparam.h>
50 #include <linux/pagemap.h>
51 #include <linux/vmalloc.h>
53 #include "gasket_constants.h"
54 #include "gasket_core.h"
56 /* Constants & utility macros */
57 /* The number of pages that can be mapped into each second-level page table. */
58 #define GASKET_PAGES_PER_SUBTABLE 512
60 /* The starting position of the page index in a simple virtual address. */
61 #define GASKET_SIMPLE_PAGE_SHIFT 12
63 /* Flag indicating that a [device] slot is valid for use. */
64 #define GASKET_VALID_SLOT_FLAG 1
67 * The starting position of the level 0 page index (i.e., the entry in the
68 * device's extended address registers) in an extended address.
69 * Also can be thought of as (log2(PAGE_SIZE) + log2(PAGES_PER_SUBTABLE)),
72 #define GASKET_EXTENDED_LVL0_SHIFT 21
75 * Number of first level pages that Gasket chips support. Equivalent to
76 * log2(NUM_LVL0_PAGE_TABLES)
78 * At a maximum, allowing for a 34 bits address space (or 16GB)
79 * = GASKET_EXTENDED_LVL0_WIDTH + (log2(PAGE_SIZE) + log2(PAGES_PER_SUBTABLE)
82 #define GASKET_EXTENDED_LVL0_WIDTH 13
85 * The starting position of the level 1 page index (i.e., the entry in the
86 * host second-level/sub- table) in an extended address.
88 #define GASKET_EXTENDED_LVL1_SHIFT 12
90 /* Type declarations */
91 /* Valid states for a struct gasket_page_table_entry. */
98 * Mapping metadata for a single page.
100 * In this file, host-side page table entries are referred to as that (or PTEs).
101 * Where device vs. host entries are differentiated, device-side or -visible
102 * entries are called "slots". A slot may be either an entry in the device's
103 * address translation table registers or an entry in a second-level page
104 * table ("subtable").
106 * The full data in this structure is visible on the host [of course]. Only
107 * the address contained in dma_addr is communicated to the device; that points
108 * to the actual page mapped and described by this structure.
110 struct gasket_page_table_entry {
111 /* The status of this entry/slot: free or in use. */
112 enum pte_status status;
115 * Index for alignment into host vaddrs.
116 * When a user specifies a host address for a mapping, that address may
117 * not be page-aligned. Offset is the index into the containing page of
118 * the host address (i.e., host_vaddr & (PAGE_SIZE - 1)).
119 * This is necessary for translating between user-specified addresses
120 * and page-aligned addresses.
124 /* Address of the page in DMA space. */
127 /* Linux page descriptor for the page described by this structure. */
131 * If this is an extended and first-level entry, sublevel points
132 * to the second-level entries underneath this entry.
134 struct gasket_page_table_entry *sublevel;
138 * Maintains virtual to physical address mapping for a coherent page that is
139 * allocated by this module for a given device.
140 * Note that coherent pages mappings virt mapping cannot be tracked by the
141 * Linux kernel, and coherent pages don't have a struct page associated,
142 * hence Linux kernel cannot perform a get_user_page_xx() on a phys address
143 * that was allocated coherent.
144 * This structure trivially implements this mechanism.
146 struct gasket_coherent_page_entry {
147 /* Phys address, dma'able by the owner device */
150 /* Kernel virtual address */
153 /* User virtual address that was mapped by the mmap kernel subsystem */
157 * Whether this page has been mapped into a user land process virtual
164 * [Host-side] page table descriptor.
166 * This structure tracks the metadata necessary to manage both simple and
167 * extended page tables.
169 struct gasket_page_table {
170 /* The config used to create this page table. */
171 struct gasket_page_table_config config;
173 /* The number of simple (single-level) entries in the page table. */
174 uint num_simple_entries;
176 /* The number of extended (two-level) entries in the page table. */
177 uint num_extended_entries;
179 /* Array of [host-side] page table entries. */
180 struct gasket_page_table_entry *entries;
182 /* Number of actively mapped kernel pages in this table. */
183 uint num_active_pages;
185 /* Device register: base of/first slot in the page table. */
186 u64 __iomem *base_slot;
188 /* Device register: holds the offset indicating the start of the
189 * extended address region of the device's address translation table.
191 u64 __iomem *extended_offset_reg;
193 /* Device structure for the underlying device. Only used for logging. */
194 struct device *device;
196 /* PCI system descriptor for the underlying device. */
197 struct pci_dev *pci_dev;
199 /* Location of the extended address bit for this Gasket device. */
202 /* Mutex to protect page table internals. */
205 /* Number of coherent pages accessible thru by this page table */
206 int num_coherent_pages;
209 * List of coherent memory (physical) allocated for a device.
211 * This structure also remembers the user virtual mapping, this is
212 * hacky, but we need to do this because the kernel doesn't keep track
213 * of the user coherent pages (pfn pages), and virt to coherent page
215 * TODO: use find_vma() APIs to convert host address to vm_area, to
216 * dma_addr_t instead of storing user virtu address in
217 * gasket_coherent_page_entry
219 * Note that the user virtual mapping is created by the driver, in
220 * gasket_mmap function, so user_virt belongs in the driver anyhow.
222 struct gasket_coherent_page_entry *coherent_pages;
225 /* See gasket_page_table.h for description. */
226 int gasket_page_table_init(struct gasket_page_table **ppg_tbl,
227 const struct gasket_bar_data *bar_data,
228 const struct gasket_page_table_config *page_table_config,
229 struct device *device, struct pci_dev *pci_dev)
232 struct gasket_page_table *pg_tbl;
233 ulong total_entries = page_table_config->total_entries;
236 * TODO: Verify config->total_entries against value read from the
237 * hardware register that contains the page table size.
239 if (total_entries == ULONG_MAX) {
241 "Error reading page table size. Initializing page table with size 0\n");
246 "Attempting to initialize page table of size 0x%lx\n",
250 "Table has base reg 0x%x, extended offset reg 0x%x\n",
251 page_table_config->base_reg,
252 page_table_config->extended_reg);
254 *ppg_tbl = kzalloc(sizeof(**ppg_tbl), GFP_KERNEL);
256 dev_dbg(device, "No memory for page table\n");
261 bytes = total_entries * sizeof(struct gasket_page_table_entry);
263 pg_tbl->entries = vzalloc(bytes);
264 if (!pg_tbl->entries) {
266 "No memory for address translation metadata\n");
273 mutex_init(&pg_tbl->mutex);
274 memcpy(&pg_tbl->config, page_table_config, sizeof(*page_table_config));
275 if (pg_tbl->config.mode == GASKET_PAGE_TABLE_MODE_NORMAL ||
276 pg_tbl->config.mode == GASKET_PAGE_TABLE_MODE_SIMPLE) {
277 pg_tbl->num_simple_entries = total_entries;
278 pg_tbl->num_extended_entries = 0;
279 pg_tbl->extended_flag = 1ull << page_table_config->extended_bit;
281 pg_tbl->num_simple_entries = 0;
282 pg_tbl->num_extended_entries = total_entries;
283 pg_tbl->extended_flag = 0;
285 pg_tbl->num_active_pages = 0;
287 (u64 __iomem *)&bar_data->virt_base[page_table_config->base_reg];
288 pg_tbl->extended_offset_reg =
289 (u64 __iomem *)&bar_data->virt_base[page_table_config->extended_reg];
290 pg_tbl->device = get_device(device);
291 pg_tbl->pci_dev = pci_dev;
293 dev_dbg(device, "Page table initialized successfully\n");
299 * Check if a range of PTEs is free.
300 * The page table mutex must be held by the caller.
302 static bool gasket_is_pte_range_free(struct gasket_page_table_entry *ptes,
307 for (i = 0; i < num_entries; i++) {
308 if (ptes[i].status != PTE_FREE)
316 * Free a second level page [sub]table.
317 * The page table mutex must be held before this call.
319 static void gasket_free_extended_subtable(struct gasket_page_table *pg_tbl,
320 struct gasket_page_table_entry *pte,
323 /* Release the page table from the driver */
324 pte->status = PTE_FREE;
326 /* Release the page table from the device */
330 dma_unmap_page(pg_tbl->device, pte->dma_addr, PAGE_SIZE,
333 vfree(pte->sublevel);
336 free_page((ulong)page_address(pte->page));
338 memset(pte, 0, sizeof(struct gasket_page_table_entry));
342 * Actually perform collection.
343 * The page table mutex must be held by the caller.
346 gasket_page_table_garbage_collect_nolock(struct gasket_page_table *pg_tbl)
348 struct gasket_page_table_entry *pte;
351 /* XXX FIX ME XXX -- more efficient to keep a usage count */
352 /* rather than scanning the second level page tables */
354 for (pte = pg_tbl->entries + pg_tbl->num_simple_entries,
355 slot = pg_tbl->base_slot + pg_tbl->num_simple_entries;
356 pte < pg_tbl->entries + pg_tbl->config.total_entries;
358 if (pte->status == PTE_INUSE) {
359 if (gasket_is_pte_range_free(pte->sublevel,
360 GASKET_PAGES_PER_SUBTABLE))
361 gasket_free_extended_subtable(pg_tbl, pte,
367 /* See gasket_page_table.h for description. */
368 void gasket_page_table_garbage_collect(struct gasket_page_table *pg_tbl)
370 mutex_lock(&pg_tbl->mutex);
371 gasket_page_table_garbage_collect_nolock(pg_tbl);
372 mutex_unlock(&pg_tbl->mutex);
375 /* See gasket_page_table.h for description. */
376 void gasket_page_table_cleanup(struct gasket_page_table *pg_tbl)
378 /* Deallocate free second-level tables. */
379 gasket_page_table_garbage_collect(pg_tbl);
381 /* TODO: Check that all PTEs have been freed? */
383 vfree(pg_tbl->entries);
384 pg_tbl->entries = NULL;
386 put_device(pg_tbl->device);
390 /* See gasket_page_table.h for description. */
391 int gasket_page_table_partition(struct gasket_page_table *pg_tbl,
392 uint num_simple_entries)
396 mutex_lock(&pg_tbl->mutex);
397 if (num_simple_entries > pg_tbl->config.total_entries) {
398 mutex_unlock(&pg_tbl->mutex);
402 gasket_page_table_garbage_collect_nolock(pg_tbl);
404 start = min(pg_tbl->num_simple_entries, num_simple_entries);
406 for (i = start; i < pg_tbl->config.total_entries; i++) {
407 if (pg_tbl->entries[i].status != PTE_FREE) {
408 dev_err(pg_tbl->device, "entry %d is not free\n", i);
409 mutex_unlock(&pg_tbl->mutex);
414 pg_tbl->num_simple_entries = num_simple_entries;
415 pg_tbl->num_extended_entries =
416 pg_tbl->config.total_entries - num_simple_entries;
417 writeq(num_simple_entries, pg_tbl->extended_offset_reg);
419 mutex_unlock(&pg_tbl->mutex);
422 EXPORT_SYMBOL(gasket_page_table_partition);
425 * Return whether a host buffer was mapped as coherent memory.
427 * A Gasket page_table currently support one contiguous dma range, mapped to one
428 * contiguous virtual memory range. Check if the host_addr is within that range.
430 static int is_coherent(struct gasket_page_table *pg_tbl, ulong host_addr)
434 /* whether the host address is within user virt range */
435 if (!pg_tbl->coherent_pages)
438 min = (u64)pg_tbl->coherent_pages[0].user_virt;
439 max = min + PAGE_SIZE * pg_tbl->num_coherent_pages;
441 return min <= host_addr && host_addr < max;
444 /* Safely return a page to the OS. */
445 static bool gasket_release_page(struct page *page)
450 if (!PageReserved(page))
452 unpin_user_page(page);
458 * Get and map last level page table buffers.
460 * slots is the location(s) to write device-mapped page address. If this is a
461 * simple mapping, these will be address translation registers. If this is
462 * an extended mapping, these will be within a second-level page table
463 * allocated by the host and so must have their __iomem attribute casted away.
465 static int gasket_perform_mapping(struct gasket_page_table *pg_tbl,
466 struct gasket_page_table_entry *ptes,
467 u64 __iomem *slots, ulong host_addr,
468 uint num_pages, int is_simple_mapping)
477 for (i = 0; i < num_pages; i++) {
478 page_addr = host_addr + i * PAGE_SIZE;
479 offset = page_addr & (PAGE_SIZE - 1);
480 if (is_coherent(pg_tbl, host_addr)) {
483 (u64)pg_tbl->coherent_pages[0].user_virt;
485 ptes[i].offset = offset;
486 ptes[i].dma_addr = pg_tbl->coherent_pages[0].paddr +
489 ret = pin_user_pages_fast(page_addr - offset, 1,
493 dev_err(pg_tbl->device,
494 "pin user pages failed for addr=0x%lx, offset=0x%lx [ret=%d]\n",
495 page_addr, offset, ret);
496 return ret ? ret : -ENOMEM;
498 ++pg_tbl->num_active_pages;
501 ptes[i].offset = offset;
503 /* Map the page into DMA space. */
505 dma_map_page(pg_tbl->device, page, 0, PAGE_SIZE,
508 if (dma_mapping_error(pg_tbl->device,
510 if (gasket_release_page(ptes[i].page))
511 --pg_tbl->num_active_pages;
514 sizeof(struct gasket_page_table_entry));
519 /* Make the DMA-space address available to the device. */
520 dma_addr = (ptes[i].dma_addr + offset) | GASKET_VALID_SLOT_FLAG;
522 if (is_simple_mapping) {
523 writeq(dma_addr, &slots[i]);
525 ((u64 __force *)slots)[i] = dma_addr;
526 /* Extended page table vectors are in DRAM,
527 * and so need to be synced each time they are updated.
529 dma_map_single(pg_tbl->device,
530 (void *)&((u64 __force *)slots)[i],
531 sizeof(u64), DMA_TO_DEVICE);
533 ptes[i].status = PTE_INUSE;
539 * Return the index of the page for the address in the simple table.
540 * Does not perform validity checking.
542 static int gasket_simple_page_idx(struct gasket_page_table *pg_tbl,
545 return (dev_addr >> GASKET_SIMPLE_PAGE_SHIFT) &
546 (pg_tbl->config.total_entries - 1);
550 * Return the level 0 page index for the given address.
551 * Does not perform validity checking.
553 static ulong gasket_extended_lvl0_page_idx(struct gasket_page_table *pg_tbl,
556 return (dev_addr >> GASKET_EXTENDED_LVL0_SHIFT) &
557 (pg_tbl->config.total_entries - 1);
561 * Return the level 1 page index for the given address.
562 * Does not perform validity checking.
564 static ulong gasket_extended_lvl1_page_idx(struct gasket_page_table *pg_tbl,
567 return (dev_addr >> GASKET_EXTENDED_LVL1_SHIFT) &
568 (GASKET_PAGES_PER_SUBTABLE - 1);
572 * Allocate page table entries in a simple table.
573 * The page table mutex must be held by the caller.
575 static int gasket_alloc_simple_entries(struct gasket_page_table *pg_tbl,
576 ulong dev_addr, uint num_pages)
578 if (!gasket_is_pte_range_free(pg_tbl->entries +
579 gasket_simple_page_idx(pg_tbl, dev_addr),
587 * Unmap and release mapped pages.
588 * The page table mutex must be held by the caller.
590 static void gasket_perform_unmapping(struct gasket_page_table *pg_tbl,
591 struct gasket_page_table_entry *ptes,
592 u64 __iomem *slots, uint num_pages,
593 int is_simple_mapping)
597 * For each page table entry and corresponding entry in the device's
598 * address translation table:
600 for (i = 0; i < num_pages; i++) {
601 /* release the address from the device, */
602 if (is_simple_mapping || ptes[i].status == PTE_INUSE) {
603 writeq(0, &slots[i]);
605 ((u64 __force *)slots)[i] = 0;
606 /* sync above PTE update before updating mappings */
610 /* release the address from the driver, */
611 if (ptes[i].status == PTE_INUSE) {
612 if (ptes[i].page && ptes[i].dma_addr) {
613 dma_unmap_page(pg_tbl->device, ptes[i].dma_addr,
614 PAGE_SIZE, DMA_BIDIRECTIONAL);
616 if (gasket_release_page(ptes[i].page))
617 --pg_tbl->num_active_pages;
620 /* and clear the PTE. */
621 memset(&ptes[i], 0, sizeof(struct gasket_page_table_entry));
626 * Unmap and release pages mapped to simple addresses.
627 * The page table mutex must be held by the caller.
629 static void gasket_unmap_simple_pages(struct gasket_page_table *pg_tbl,
630 ulong dev_addr, uint num_pages)
632 uint slot = gasket_simple_page_idx(pg_tbl, dev_addr);
634 gasket_perform_unmapping(pg_tbl, pg_tbl->entries + slot,
635 pg_tbl->base_slot + slot, num_pages, 1);
639 * Unmap and release buffers to extended addresses.
640 * The page table mutex must be held by the caller.
642 static void gasket_unmap_extended_pages(struct gasket_page_table *pg_tbl,
643 ulong dev_addr, uint num_pages)
645 uint slot_idx, remain, len;
646 struct gasket_page_table_entry *pte;
647 u64 __iomem *slot_base;
650 slot_idx = gasket_extended_lvl1_page_idx(pg_tbl, dev_addr);
651 pte = pg_tbl->entries + pg_tbl->num_simple_entries +
652 gasket_extended_lvl0_page_idx(pg_tbl, dev_addr);
655 /* TODO: Add check to ensure pte remains valid? */
656 len = min(remain, GASKET_PAGES_PER_SUBTABLE - slot_idx);
658 if (pte->status == PTE_INUSE) {
659 slot_base = (u64 __iomem *)(page_address(pte->page) +
661 gasket_perform_unmapping(pg_tbl,
662 pte->sublevel + slot_idx,
663 slot_base + slot_idx, len, 0);
672 /* Evaluates to nonzero if the specified virtual address is simple. */
673 static inline bool gasket_addr_is_simple(struct gasket_page_table *pg_tbl,
676 return !((addr) & (pg_tbl)->extended_flag);
680 * Convert (simple, page, offset) into a device address.
682 * Simple page 0, offset 32:
683 * Input (1, 0, 32), Output 0x20
684 * Simple page 1000, offset 511:
685 * Input (1, 1000, 511), Output 0x3E81FF
686 * Extended page 0, offset 32:
687 * Input (0, 0, 32), Output 0x8000000020
688 * Extended page 1000, offset 511:
689 * Input (0, 1000, 511), Output 0x8003E81FF
691 static ulong gasket_components_to_dev_address(struct gasket_page_table *pg_tbl,
692 int is_simple, uint page_index,
695 ulong dev_addr = (page_index << GASKET_SIMPLE_PAGE_SHIFT) | offset;
697 return is_simple ? dev_addr : (pg_tbl->extended_flag | dev_addr);
701 * Validity checking for simple addresses.
703 * Verify that address translation commutes (from address to/from page + offset)
704 * and that the requested page range starts and ends within the set of
705 * currently-partitioned simple pages.
707 static bool gasket_is_simple_dev_addr_bad(struct gasket_page_table *pg_tbl,
708 ulong dev_addr, uint num_pages)
710 ulong page_offset = dev_addr & (PAGE_SIZE - 1);
712 (dev_addr / PAGE_SIZE) & (pg_tbl->config.total_entries - 1);
714 if (gasket_components_to_dev_address(pg_tbl, 1, page_index,
715 page_offset) != dev_addr) {
716 dev_err(pg_tbl->device, "address is invalid, 0x%lX\n",
721 if (page_index >= pg_tbl->num_simple_entries) {
722 dev_err(pg_tbl->device,
723 "starting slot at %lu is too large, max is < %u\n",
724 page_index, pg_tbl->num_simple_entries);
728 if (page_index + num_pages > pg_tbl->num_simple_entries) {
729 dev_err(pg_tbl->device,
730 "ending slot at %lu is too large, max is <= %u\n",
731 page_index + num_pages, pg_tbl->num_simple_entries);
739 * Validity checking for extended addresses.
741 * Verify that address translation commutes (from address to/from page +
742 * offset) and that the requested page range starts and ends within the set of
743 * currently-partitioned extended pages.
745 static bool gasket_is_extended_dev_addr_bad(struct gasket_page_table *pg_tbl,
746 ulong dev_addr, uint num_pages)
748 /* Starting byte index of dev_addr into the first mapped page */
749 ulong page_offset = dev_addr & (PAGE_SIZE - 1);
750 ulong page_global_idx, page_lvl0_idx;
751 ulong num_lvl0_pages;
754 /* check if the device address is out of bound */
755 addr = dev_addr & ~((pg_tbl)->extended_flag);
756 if (addr >> (GASKET_EXTENDED_LVL0_WIDTH + GASKET_EXTENDED_LVL0_SHIFT)) {
757 dev_err(pg_tbl->device, "device address out of bounds: 0x%lx\n",
762 /* Find the starting sub-page index in the space of all sub-pages. */
763 page_global_idx = (dev_addr / PAGE_SIZE) &
764 (pg_tbl->config.total_entries * GASKET_PAGES_PER_SUBTABLE - 1);
766 /* Find the starting level 0 index. */
767 page_lvl0_idx = gasket_extended_lvl0_page_idx(pg_tbl, dev_addr);
769 /* Get the count of affected level 0 pages. */
770 num_lvl0_pages = DIV_ROUND_UP(num_pages, GASKET_PAGES_PER_SUBTABLE);
772 if (gasket_components_to_dev_address(pg_tbl, 0, page_global_idx,
773 page_offset) != dev_addr) {
774 dev_err(pg_tbl->device, "address is invalid: 0x%lx\n",
779 if (page_lvl0_idx >= pg_tbl->num_extended_entries) {
780 dev_err(pg_tbl->device,
781 "starting level 0 slot at %lu is too large, max is < %u\n",
782 page_lvl0_idx, pg_tbl->num_extended_entries);
786 if (page_lvl0_idx + num_lvl0_pages > pg_tbl->num_extended_entries) {
787 dev_err(pg_tbl->device,
788 "ending level 0 slot at %lu is too large, max is <= %u\n",
789 page_lvl0_idx + num_lvl0_pages,
790 pg_tbl->num_extended_entries);
798 * Non-locking entry to unmapping routines.
799 * The page table mutex must be held by the caller.
801 static void gasket_page_table_unmap_nolock(struct gasket_page_table *pg_tbl,
802 ulong dev_addr, uint num_pages)
807 if (gasket_addr_is_simple(pg_tbl, dev_addr))
808 gasket_unmap_simple_pages(pg_tbl, dev_addr, num_pages);
810 gasket_unmap_extended_pages(pg_tbl, dev_addr, num_pages);
814 * Allocate and map pages to simple addresses.
815 * If there is an error, no pages are mapped.
817 static int gasket_map_simple_pages(struct gasket_page_table *pg_tbl,
818 ulong host_addr, ulong dev_addr,
822 uint slot_idx = gasket_simple_page_idx(pg_tbl, dev_addr);
824 ret = gasket_alloc_simple_entries(pg_tbl, dev_addr, num_pages);
826 dev_err(pg_tbl->device,
827 "page table slots %u (@ 0x%lx) to %u are not available\n",
828 slot_idx, dev_addr, slot_idx + num_pages - 1);
832 ret = gasket_perform_mapping(pg_tbl, pg_tbl->entries + slot_idx,
833 pg_tbl->base_slot + slot_idx, host_addr,
837 gasket_page_table_unmap_nolock(pg_tbl, dev_addr, num_pages);
838 dev_err(pg_tbl->device, "gasket_perform_mapping %d\n", ret);
844 * Allocate a second level page table.
845 * The page table mutex must be held by the caller.
847 static int gasket_alloc_extended_subtable(struct gasket_page_table *pg_tbl,
848 struct gasket_page_table_entry *pte,
851 ulong page_addr, subtable_bytes;
854 /* XXX FIX ME XXX this is inefficient for non-4K page sizes */
856 /* GFP_DMA flag must be passed to architectures for which
857 * part of the memory range is not considered DMA'able.
858 * This seems to be the case for Juno board with 4.5.0 Linaro kernel
860 page_addr = get_zeroed_page(GFP_KERNEL | GFP_DMA);
863 pte->page = virt_to_page((void *)page_addr);
866 subtable_bytes = sizeof(struct gasket_page_table_entry) *
867 GASKET_PAGES_PER_SUBTABLE;
868 pte->sublevel = vzalloc(subtable_bytes);
869 if (!pte->sublevel) {
870 free_page(page_addr);
871 memset(pte, 0, sizeof(struct gasket_page_table_entry));
875 /* Map the page into DMA space. */
876 pte->dma_addr = dma_map_page(pg_tbl->device, pte->page, 0, PAGE_SIZE,
878 if (dma_mapping_error(pg_tbl->device, pte->dma_addr)) {
879 free_page(page_addr);
880 vfree(pte->sublevel);
881 memset(pte, 0, sizeof(struct gasket_page_table_entry));
885 /* make the addresses available to the device */
886 dma_addr = (pte->dma_addr + pte->offset) | GASKET_VALID_SLOT_FLAG;
887 writeq(dma_addr, slot);
889 pte->status = PTE_INUSE;
895 * Allocate slots in an extended page table. Check to see if a range of page
896 * table slots are available. If necessary, memory is allocated for second level
899 * Note that memory for second level page tables is allocated as needed, but
900 * that memory is only freed on the final close of the device file, when the
901 * page tables are repartitioned, or the device is removed. If there is an
902 * error or if the full range of slots is not available, any memory
903 * allocated for second level page tables remains allocated until final close,
904 * repartition, or device removal.
906 * The page table mutex must be held by the caller.
908 static int gasket_alloc_extended_entries(struct gasket_page_table *pg_tbl,
909 ulong dev_addr, uint num_entries)
912 uint remain, subtable_slot_idx, len;
913 struct gasket_page_table_entry *pte;
916 remain = num_entries;
917 subtable_slot_idx = gasket_extended_lvl1_page_idx(pg_tbl, dev_addr);
918 pte = pg_tbl->entries + pg_tbl->num_simple_entries +
919 gasket_extended_lvl0_page_idx(pg_tbl, dev_addr);
920 slot = pg_tbl->base_slot + pg_tbl->num_simple_entries +
921 gasket_extended_lvl0_page_idx(pg_tbl, dev_addr);
925 GASKET_PAGES_PER_SUBTABLE - subtable_slot_idx);
927 if (pte->status == PTE_FREE) {
928 ret = gasket_alloc_extended_subtable(pg_tbl, pte, slot);
930 dev_err(pg_tbl->device,
931 "no memory for extended addr subtable\n");
935 if (!gasket_is_pte_range_free(pte->sublevel +
936 subtable_slot_idx, len))
941 subtable_slot_idx = 0;
950 * gasket_map_extended_pages - Get and map buffers to extended addresses.
951 * If there is an error, no pages are mapped.
953 static int gasket_map_extended_pages(struct gasket_page_table *pg_tbl,
954 ulong host_addr, ulong dev_addr,
959 uint slot_idx, remain, len;
960 struct gasket_page_table_entry *pte;
961 u64 __iomem *slot_base;
963 ret = gasket_alloc_extended_entries(pg_tbl, dev_addr, num_pages);
965 dev_addr_end = dev_addr + (num_pages / PAGE_SIZE) - 1;
966 dev_err(pg_tbl->device,
967 "page table slots (%lu,%lu) (@ 0x%lx) to (%lu,%lu) are not available\n",
968 gasket_extended_lvl0_page_idx(pg_tbl, dev_addr),
970 gasket_extended_lvl1_page_idx(pg_tbl, dev_addr),
971 gasket_extended_lvl0_page_idx(pg_tbl, dev_addr_end),
972 gasket_extended_lvl1_page_idx(pg_tbl, dev_addr_end));
977 slot_idx = gasket_extended_lvl1_page_idx(pg_tbl, dev_addr);
978 pte = pg_tbl->entries + pg_tbl->num_simple_entries +
979 gasket_extended_lvl0_page_idx(pg_tbl, dev_addr);
982 len = min(remain, GASKET_PAGES_PER_SUBTABLE - slot_idx);
985 (u64 __iomem *)(page_address(pte->page) + pte->offset);
986 ret = gasket_perform_mapping(pg_tbl, pte->sublevel + slot_idx,
987 slot_base + slot_idx, host_addr,
990 gasket_page_table_unmap_nolock(pg_tbl, dev_addr,
998 host_addr += len * PAGE_SIZE;
1005 * See gasket_page_table.h for general description.
1007 * gasket_page_table_map calls either gasket_map_simple_pages() or
1008 * gasket_map_extended_pages() to actually perform the mapping.
1010 * The page table mutex is held for the entire operation.
1012 int gasket_page_table_map(struct gasket_page_table *pg_tbl, ulong host_addr,
1013 ulong dev_addr, uint num_pages)
1020 mutex_lock(&pg_tbl->mutex);
1022 if (gasket_addr_is_simple(pg_tbl, dev_addr)) {
1023 ret = gasket_map_simple_pages(pg_tbl, host_addr, dev_addr,
1026 ret = gasket_map_extended_pages(pg_tbl, host_addr, dev_addr,
1030 mutex_unlock(&pg_tbl->mutex);
1033 EXPORT_SYMBOL(gasket_page_table_map);
1036 * See gasket_page_table.h for general description.
1038 * gasket_page_table_unmap takes the page table lock and calls either
1039 * gasket_unmap_simple_pages() or gasket_unmap_extended_pages() to
1040 * actually unmap the pages from device space.
1042 * The page table mutex is held for the entire operation.
1044 void gasket_page_table_unmap(struct gasket_page_table *pg_tbl, ulong dev_addr,
1050 mutex_lock(&pg_tbl->mutex);
1051 gasket_page_table_unmap_nolock(pg_tbl, dev_addr, num_pages);
1052 mutex_unlock(&pg_tbl->mutex);
1054 EXPORT_SYMBOL(gasket_page_table_unmap);
1056 static void gasket_page_table_unmap_all_nolock(struct gasket_page_table *pg_tbl)
1058 gasket_unmap_simple_pages(pg_tbl,
1059 gasket_components_to_dev_address(pg_tbl, 1, 0,
1061 pg_tbl->num_simple_entries);
1062 gasket_unmap_extended_pages(pg_tbl,
1063 gasket_components_to_dev_address(pg_tbl, 0,
1065 pg_tbl->num_extended_entries *
1066 GASKET_PAGES_PER_SUBTABLE);
1069 /* See gasket_page_table.h for description. */
1070 void gasket_page_table_unmap_all(struct gasket_page_table *pg_tbl)
1072 mutex_lock(&pg_tbl->mutex);
1073 gasket_page_table_unmap_all_nolock(pg_tbl);
1074 mutex_unlock(&pg_tbl->mutex);
1076 EXPORT_SYMBOL(gasket_page_table_unmap_all);
1078 /* See gasket_page_table.h for description. */
1079 void gasket_page_table_reset(struct gasket_page_table *pg_tbl)
1081 mutex_lock(&pg_tbl->mutex);
1082 gasket_page_table_unmap_all_nolock(pg_tbl);
1083 writeq(pg_tbl->config.total_entries, pg_tbl->extended_offset_reg);
1084 mutex_unlock(&pg_tbl->mutex);
1087 /* See gasket_page_table.h for description. */
1088 int gasket_page_table_lookup_page(struct gasket_page_table *pg_tbl,
1089 ulong dev_addr, struct page **ppage,
1093 struct gasket_page_table_entry *pte;
1095 mutex_lock(&pg_tbl->mutex);
1096 if (gasket_addr_is_simple(pg_tbl, dev_addr)) {
1097 page_num = gasket_simple_page_idx(pg_tbl, dev_addr);
1098 if (page_num >= pg_tbl->num_simple_entries)
1101 pte = pg_tbl->entries + page_num;
1102 if (pte->status != PTE_INUSE)
1105 /* Find the level 0 entry, */
1106 page_num = gasket_extended_lvl0_page_idx(pg_tbl, dev_addr);
1107 if (page_num >= pg_tbl->num_extended_entries)
1110 pte = pg_tbl->entries + pg_tbl->num_simple_entries + page_num;
1111 if (pte->status != PTE_INUSE)
1114 /* and its contained level 1 entry. */
1115 page_num = gasket_extended_lvl1_page_idx(pg_tbl, dev_addr);
1116 pte = pte->sublevel + page_num;
1117 if (pte->status != PTE_INUSE)
1122 *poffset = pte->offset;
1123 mutex_unlock(&pg_tbl->mutex);
1129 mutex_unlock(&pg_tbl->mutex);
1133 /* See gasket_page_table.h for description. */
1134 bool gasket_page_table_are_addrs_bad(struct gasket_page_table *pg_tbl,
1135 ulong host_addr, ulong dev_addr,
1138 if (host_addr & (PAGE_SIZE - 1)) {
1139 dev_err(pg_tbl->device,
1140 "host mapping address 0x%lx must be page aligned\n",
1145 return gasket_page_table_is_dev_addr_bad(pg_tbl, dev_addr, bytes);
1147 EXPORT_SYMBOL(gasket_page_table_are_addrs_bad);
1149 /* See gasket_page_table.h for description. */
1150 bool gasket_page_table_is_dev_addr_bad(struct gasket_page_table *pg_tbl,
1151 ulong dev_addr, ulong bytes)
1153 uint num_pages = bytes / PAGE_SIZE;
1155 if (bytes & (PAGE_SIZE - 1)) {
1156 dev_err(pg_tbl->device,
1157 "mapping size 0x%lX must be page aligned\n", bytes);
1161 if (num_pages == 0) {
1162 dev_err(pg_tbl->device,
1163 "requested mapping is less than one page: %lu / %lu\n",
1168 if (gasket_addr_is_simple(pg_tbl, dev_addr))
1169 return gasket_is_simple_dev_addr_bad(pg_tbl, dev_addr,
1171 return gasket_is_extended_dev_addr_bad(pg_tbl, dev_addr, num_pages);
1173 EXPORT_SYMBOL(gasket_page_table_is_dev_addr_bad);
1175 /* See gasket_page_table.h for description. */
1176 uint gasket_page_table_max_size(struct gasket_page_table *page_table)
1180 return page_table->config.total_entries;
1182 EXPORT_SYMBOL(gasket_page_table_max_size);
1184 /* See gasket_page_table.h for description. */
1185 uint gasket_page_table_num_entries(struct gasket_page_table *pg_tbl)
1189 return pg_tbl->num_simple_entries + pg_tbl->num_extended_entries;
1191 EXPORT_SYMBOL(gasket_page_table_num_entries);
1193 /* See gasket_page_table.h for description. */
1194 uint gasket_page_table_num_simple_entries(struct gasket_page_table *pg_tbl)
1198 return pg_tbl->num_simple_entries;
1200 EXPORT_SYMBOL(gasket_page_table_num_simple_entries);
1202 /* See gasket_page_table.h for description. */
1203 uint gasket_page_table_num_active_pages(struct gasket_page_table *pg_tbl)
1207 return pg_tbl->num_active_pages;
1209 EXPORT_SYMBOL(gasket_page_table_num_active_pages);
1211 /* See gasket_page_table.h */
1212 int gasket_page_table_system_status(struct gasket_page_table *page_table)
1215 return GASKET_STATUS_LAMED;
1217 if (gasket_page_table_num_entries(page_table) == 0) {
1218 dev_dbg(page_table->device, "Page table size is 0\n");
1219 return GASKET_STATUS_LAMED;
1222 return GASKET_STATUS_ALIVE;
1225 /* Record the host_addr to coherent dma memory mapping. */
1226 int gasket_set_user_virt(struct gasket_dev *gasket_dev, u64 size,
1227 dma_addr_t dma_address, ulong vma)
1230 struct gasket_page_table *pg_tbl;
1232 unsigned int num_pages = size / PAGE_SIZE;
1235 * TODO: for future chipset, better handling of the case where multiple
1236 * page tables are supported on a given device
1238 pg_tbl = gasket_dev->page_table[0];
1240 dev_dbg(gasket_dev->dev, "%s: invalid page table index\n",
1244 for (j = 0; j < num_pages; j++) {
1245 pg_tbl->coherent_pages[j].user_virt =
1246 (u64)vma + j * PAGE_SIZE;
1251 /* Allocate a block of coherent memory. */
1252 int gasket_alloc_coherent_memory(struct gasket_dev *gasket_dev, u64 size,
1253 dma_addr_t *dma_address, u64 index)
1258 unsigned int num_pages = DIV_ROUND_UP(size, PAGE_SIZE);
1259 const struct gasket_driver_desc *driver_desc =
1260 gasket_get_driver_desc(gasket_dev);
1262 if (!gasket_dev->page_table[index])
1268 mem = dma_alloc_coherent(gasket_get_device(gasket_dev),
1269 num_pages * PAGE_SIZE, &handle, GFP_KERNEL);
1273 gasket_dev->page_table[index]->num_coherent_pages = num_pages;
1275 /* allocate the physical memory block */
1276 gasket_dev->page_table[index]->coherent_pages =
1278 sizeof(*gasket_dev->page_table[index]->coherent_pages),
1280 if (!gasket_dev->page_table[index]->coherent_pages)
1283 gasket_dev->coherent_buffer.length_bytes =
1284 PAGE_SIZE * (num_pages);
1285 gasket_dev->coherent_buffer.phys_base = handle;
1286 gasket_dev->coherent_buffer.virt_base = mem;
1288 *dma_address = driver_desc->coherent_buffer_description.base;
1289 for (j = 0; j < num_pages; j++) {
1290 gasket_dev->page_table[index]->coherent_pages[j].paddr =
1291 handle + j * PAGE_SIZE;
1292 gasket_dev->page_table[index]->coherent_pages[j].kernel_virt =
1293 (u64)mem + j * PAGE_SIZE;
1300 dma_free_coherent(gasket_get_device(gasket_dev),
1301 num_pages * PAGE_SIZE, mem, handle);
1302 gasket_dev->coherent_buffer.length_bytes = 0;
1303 gasket_dev->coherent_buffer.virt_base = NULL;
1304 gasket_dev->coherent_buffer.phys_base = 0;
1307 kfree(gasket_dev->page_table[index]->coherent_pages);
1308 gasket_dev->page_table[index]->coherent_pages = NULL;
1309 gasket_dev->page_table[index]->num_coherent_pages = 0;
1313 /* Free a block of coherent memory. */
1314 int gasket_free_coherent_memory(struct gasket_dev *gasket_dev, u64 size,
1315 dma_addr_t dma_address, u64 index)
1317 const struct gasket_driver_desc *driver_desc;
1319 if (!gasket_dev->page_table[index])
1322 driver_desc = gasket_get_driver_desc(gasket_dev);
1324 if (driver_desc->coherent_buffer_description.base != dma_address)
1325 return -EADDRNOTAVAIL;
1327 if (gasket_dev->coherent_buffer.length_bytes) {
1328 dma_free_coherent(gasket_get_device(gasket_dev),
1329 gasket_dev->coherent_buffer.length_bytes,
1330 gasket_dev->coherent_buffer.virt_base,
1331 gasket_dev->coherent_buffer.phys_base);
1332 gasket_dev->coherent_buffer.length_bytes = 0;
1333 gasket_dev->coherent_buffer.virt_base = NULL;
1334 gasket_dev->coherent_buffer.phys_base = 0;
1337 kfree(gasket_dev->page_table[index]->coherent_pages);
1338 gasket_dev->page_table[index]->coherent_pages = NULL;
1339 gasket_dev->page_table[index]->num_coherent_pages = 0;
1344 /* Release all coherent memory. */
1345 void gasket_free_coherent_memory_all(struct gasket_dev *gasket_dev, u64 index)
1347 if (!gasket_dev->page_table[index])
1350 if (gasket_dev->coherent_buffer.length_bytes) {
1351 dma_free_coherent(gasket_get_device(gasket_dev),
1352 gasket_dev->coherent_buffer.length_bytes,
1353 gasket_dev->coherent_buffer.virt_base,
1354 gasket_dev->coherent_buffer.phys_base);
1355 gasket_dev->coherent_buffer.length_bytes = 0;
1356 gasket_dev->coherent_buffer.virt_base = NULL;
1357 gasket_dev->coherent_buffer.phys_base = 0;