1// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2/*
3 * Copyright(c) 2015-2017 Intel Corporation.
4 */
5
6#include <linux/mm.h>
7#include <linux/sched/signal.h>
8#include <linux/device.h>
9#include <linux/module.h>
10
11#include "hfi.h"
12
13static unsigned long cache_size = 256;
14module_param(cache_size, ulong, S_IRUGO | S_IWUSR);
15MODULE_PARM_DESC(cache_size, "Send and receive side cache size limit (in MB)");
16
17/*
18 * Determine whether the caller can pin pages.
19 *
20 * This function should be used in the implementation of buffer caches.
21 * The cache implementation should call this function prior to attempting
22 * to pin buffer pages in order to determine whether they should do so.
23 * The function computes cache limits based on the configured ulimit and
24 * cache size. Use of this function is especially important for caches
25 * which are not limited in any other way (e.g. by HW resources) and, thus,
26 * could keeping caching buffers.
27 *
28 */
29bool hfi1_can_pin_pages(struct hfi1_devdata *dd, struct mm_struct *mm,
30			u32 nlocked, u32 npages)
31{
32	unsigned long ulimit_pages;
33	unsigned long cache_limit_pages;
34	unsigned int usr_ctxts;
35
36	/*
37	 * Perform RLIMIT_MEMLOCK based checks unless CAP_IPC_LOCK is present.
38	 */
39	if (!capable(CAP_IPC_LOCK)) {
40		ulimit_pages =
41			DIV_ROUND_DOWN_ULL(rlimit(RLIMIT_MEMLOCK), PAGE_SIZE);
42
43		/*
44		 * Pinning these pages would exceed this process's locked memory
45		 * limit.
46		 */
47		if (atomic64_read(&mm->pinned_vm) + npages > ulimit_pages)
48			return false;
49
50		/*
51		 * Only allow 1/4 of the user's RLIMIT_MEMLOCK to be used for HFI
52		 * caches.  This fraction is then equally distributed among all
53		 * existing user contexts.  Note that if RLIMIT_MEMLOCK is
54		 * 'unlimited' (-1), the value of this limit will be > 2^42 pages
55		 * (2^64 / 2^12 / 2^8 / 2^2).
56		 *
57		 * The effectiveness of this check may be reduced if I/O occurs on
58		 * some user contexts before all user contexts are created.  This
59		 * check assumes that this process is the only one using this
60		 * context (e.g., the corresponding fd was not passed to another
61		 * process for concurrent access) as there is no per-context,
62		 * per-process tracking of pinned pages.  It also assumes that each
63		 * user context has only one cache to limit.
64		 */
65		usr_ctxts = dd->num_rcv_contexts - dd->first_dyn_alloc_ctxt;
66		if (nlocked + npages > (ulimit_pages / usr_ctxts / 4))
67			return false;
68	}
69
70	/*
71	 * Pinning these pages would exceed the size limit for this cache.
72	 */
73	cache_limit_pages = cache_size * (1024 * 1024) / PAGE_SIZE;
74	if (nlocked + npages > cache_limit_pages)
75		return false;
76
77	return true;
78}
79
80int hfi1_acquire_user_pages(struct mm_struct *mm, unsigned long vaddr, size_t npages,
81			    bool writable, struct page **pages)
82{
83	int ret;
84	unsigned int gup_flags = FOLL_LONGTERM | (writable ? FOLL_WRITE : 0);
85
86	ret = pin_user_pages_fast(vaddr, npages, gup_flags, pages);
87	if (ret < 0)
88		return ret;
89
90	atomic64_add(ret, &mm->pinned_vm);
91
92	return ret;
93}
94
95void hfi1_release_user_pages(struct mm_struct *mm, struct page **p,
96			     size_t npages, bool dirty)
97{
98	unpin_user_pages_dirty_lock(p, npages, dirty);
99
100	if (mm) { /* during close after signal, mm can be NULL */
101		atomic64_sub(npages, &mm->pinned_vm);
102	}
103}
104