1 .. SPDX-License-Identifier: GPL-2.0
3 ====================================================
4 pin_user_pages() and related calls
5 ====================================================
12 This document describes the following functions::
16 pin_user_pages_remote()
18 Basic description of FOLL_PIN
19 =============================
21 FOLL_PIN and FOLL_LONGTERM are flags that can be passed to the get_user_pages*()
22 ("gup") family of functions. FOLL_PIN has significant interactions and
23 interdependencies with FOLL_LONGTERM, so both are covered here.
25 FOLL_PIN is internal to gup, meaning that it should not appear at the gup call
26 sites. This allows the associated wrapper functions (pin_user_pages*() and
27 others) to set the correct combination of these flags, and to check for problems
30 FOLL_LONGTERM, on the other hand, *is* allowed to be set at the gup call sites.
31 This is in order to avoid creating a large number of wrapper functions to cover
32 all combinations of get*(), pin*(), FOLL_LONGTERM, and more. Also, the
33 pin_user_pages*() APIs are clearly distinct from the get_user_pages*() APIs, so
34 that's a natural dividing line, and a good point to make separate wrapper calls.
35 In other words, use pin_user_pages*() for DMA-pinned pages, and
36 get_user_pages*() for other cases. There are five cases described later on in
37 this document, to further clarify that concept.
39 FOLL_PIN and FOLL_GET are mutually exclusive for a given gup call. However,
40 multiple threads and call sites are free to pin the same struct pages, via both
41 FOLL_PIN and FOLL_GET. It's just the call site that needs to choose one or the
42 other, not the struct page(s).
44 The FOLL_PIN implementation is nearly the same as FOLL_GET, except that FOLL_PIN
45 uses a different reference counting technique.
47 FOLL_PIN is a prerequisite to FOLL_LONGTERM. Another way of saying that is,
48 FOLL_LONGTERM is a specific case, more restrictive case of FOLL_PIN.
50 Which flags are set by each wrapper
51 ===================================
53 For these pin_user_pages*() functions, FOLL_PIN is OR'd in with whatever gup
54 flags the caller provides. The caller is required to pass in a non-null struct
55 pages* array, and the function then pins pages by incrementing each by a special
56 value: GUP_PIN_COUNTING_BIAS.
58 For huge pages (and in fact, any compound page of more than 2 pages), the
59 GUP_PIN_COUNTING_BIAS scheme is not used. Instead, an exact form of pin counting
60 is achieved, by using the 3rd struct page in the compound page. A new struct
61 page field, hpage_pinned_refcount, has been added in order to support this.
63 This approach for compound pages avoids the counting upper limit problems that
64 are discussed below. Those limitations would have been aggravated severely by
65 huge pages, because each tail page adds a refcount to the head page. And in
66 fact, testing revealed that, without a separate hpage_pinned_refcount field,
67 page overflows were seen in some huge page stress tests.
69 This also means that huge pages and compound pages (of order > 1) do not suffer
70 from the false positives problem that is mentioned below.::
74 pin_user_pages FOLL_PIN is always set internally by this function.
75 pin_user_pages_fast FOLL_PIN is always set internally by this function.
76 pin_user_pages_remote FOLL_PIN is always set internally by this function.
78 For these get_user_pages*() functions, FOLL_GET might not even be specified.
79 Behavior is a little more complex than above. If FOLL_GET was *not* specified,
80 but the caller passed in a non-null struct pages* array, then the function
81 sets FOLL_GET for you, and proceeds to pin pages by incrementing the refcount
86 get_user_pages FOLL_GET is sometimes set internally by this function.
87 get_user_pages_fast FOLL_GET is sometimes set internally by this function.
88 get_user_pages_remote FOLL_GET is sometimes set internally by this function.
90 Tracking dma-pinned pages
91 =========================
93 Some of the key design constraints, and solutions, for tracking dma-pinned
96 * An actual reference count, per struct page, is required. This is because
97 multiple processes may pin and unpin a page.
99 * False positives (reporting that a page is dma-pinned, when in fact it is not)
100 are acceptable, but false negatives are not.
102 * struct page may not be increased in size for this, and all fields are already
105 * Given the above, we can overload the page->_refcount field by using, sort of,
106 the upper bits in that field for a dma-pinned count. "Sort of", means that,
107 rather than dividing page->_refcount into bit fields, we simple add a medium-
108 large value (GUP_PIN_COUNTING_BIAS, initially chosen to be 1024: 10 bits) to
109 page->_refcount. This provides fuzzy behavior: if a page has get_page() called
110 on it 1024 times, then it will appear to have a single dma-pinned count.
111 And again, that's acceptable.
113 This also leads to limitations: there are only 31-10==21 bits available for a
114 counter that increments 10 bits at a time.
116 * Callers must specifically request "dma-pinned tracking of pages". In other
117 words, just calling get_user_pages() will not suffice; a new set of functions,
118 pin_user_page() and related, must be used.
120 FOLL_PIN, FOLL_GET, FOLL_LONGTERM: when to use which flags
121 ==========================================================
123 Thanks to Jan Kara, Vlastimil Babka and several other -mm people, for describing
126 CASE 1: Direct IO (DIO)
127 -----------------------
128 There are GUP references to pages that are serving
129 as DIO buffers. These buffers are needed for a relatively short time (so they
130 are not "long term"). No special synchronization with page_mkclean() or
131 munmap() is provided. Therefore, flags to set at the call site are: ::
135 ...but rather than setting FOLL_PIN directly, call sites should use one of
136 the pin_user_pages*() routines that set FOLL_PIN.
140 There are GUP references to pages that are serving as DMA
141 buffers. These buffers are needed for a long time ("long term"). No special
142 synchronization with page_mkclean() or munmap() is provided. Therefore, flags
143 to set at the call site are: ::
145 FOLL_PIN | FOLL_LONGTERM
147 NOTE: Some pages, such as DAX pages, cannot be pinned with longterm pins. That's
148 because DAX pages do not have a separate page cache, and so "pinning" implies
149 locking down file system blocks, which is not (yet) supported in that way.
151 CASE 3: MMU notifier registration, with or without page faulting hardware
152 -------------------------------------------------------------------------
153 Device drivers can pin pages via get_user_pages*(), and register for mmu
154 notifier callbacks for the memory range. Then, upon receiving a notifier
155 "invalidate range" callback , stop the device from using the range, and unpin
156 the pages. There may be other possible schemes, such as for example explicitly
157 synchronizing against pending IO, that accomplish approximately the same thing.
159 Or, if the hardware supports replayable page faults, then the device driver can
160 avoid pinning entirely (this is ideal), as follows: register for mmu notifier
161 callbacks as above, but instead of stopping the device and unpinning in the
162 callback, simply remove the range from the device's page tables.
164 Either way, as long as the driver unpins the pages upon mmu notifier callback,
165 then there is proper synchronization with both filesystem and mm
166 (page_mkclean(), munmap(), etc). Therefore, neither flag needs to be set.
168 CASE 4: Pinning for struct page manipulation only
169 -------------------------------------------------
170 If only struct page data (as opposed to the actual memory contents that a page
171 is tracking) is affected, then normal GUP calls are sufficient, and neither flag
174 CASE 5: Pinning in order to write to the data within the page
175 -------------------------------------------------------------
176 Even though neither DMA nor Direct IO is involved, just a simple case of "pin,
177 write to a page's data, unpin" can cause a problem. Case 5 may be considered a
178 superset of Case 1, plus Case 2, plus anything that invokes that pattern. In
179 other words, if the code is neither Case 1 nor Case 2, it may still require
180 FOLL_PIN, for patterns like this:
182 Correct (uses FOLL_PIN calls):
184 write to the data within the pages
187 INCORRECT (uses FOLL_GET calls):
189 write to the data within the pages
192 page_maybe_dma_pinned(): the whole point of pinning
193 ===================================================
195 The whole point of marking pages as "DMA-pinned" or "gup-pinned" is to be able
196 to query, "is this page DMA-pinned?" That allows code such as page_mkclean()
197 (and file system writeback code in general) to make informed decisions about
198 what to do when a page cannot be unmapped due to such pins.
200 What to do in those cases is the subject of a years-long series of discussions
201 and debates (see the References at the end of this document). It's a TODO item
202 here: fill in the details once that's worked out. Meanwhile, it's safe to say
203 that having this available: ::
205 static inline bool page_maybe_dma_pinned(struct page *page)
207 ...is a prerequisite to solving the long-running gup+DMA problem.
209 Another way of thinking about FOLL_GET, FOLL_PIN, and FOLL_LONGTERM
210 ===================================================================
212 Another way of thinking about these flags is as a progression of restrictions:
213 FOLL_GET is for struct page manipulation, without affecting the data that the
214 struct page refers to. FOLL_PIN is a *replacement* for FOLL_GET, and is for
215 short term pins on pages whose data *will* get accessed. As such, FOLL_PIN is
216 a "more severe" form of pinning. And finally, FOLL_LONGTERM is an even more
217 restrictive case that has FOLL_PIN as a prerequisite: this is for pages that
218 will be pinned longterm, and whose data will be accessed.
224 tools/testing/selftests/vm/gup_benchmark.c
226 has the following new calls to exercise the new pin*() wrapper functions:
228 * PIN_FAST_BENCHMARK (./gup_benchmark -a)
229 * PIN_BENCHMARK (./gup_benchmark -b)
231 You can monitor how many total dma-pinned pages have been acquired and released
232 since the system was booted, via two new /proc/vmstat entries: ::
234 /proc/vmstat/nr_foll_pin_acquired
235 /proc/vmstat/nr_foll_pin_released
237 Under normal conditions, these two values will be equal unless there are any
238 long-term [R]DMA pins in place, or during pin/unpin transitions.
240 * nr_foll_pin_acquired: This is the number of logical pins that have been
241 acquired since the system was powered on. For huge pages, the head page is
242 pinned once for each page (head page and each tail page) within the huge page.
243 This follows the same sort of behavior that get_user_pages() uses for huge
244 pages: the head page is refcounted once for each tail or head page in the huge
245 page, when get_user_pages() is applied to a huge page.
247 * nr_foll_pin_released: The number of logical pins that have been released since
248 the system was powered on. Note that pages are released (unpinned) on a
249 PAGE_SIZE granularity, even if the original pin was applied to a huge page.
250 Becaused of the pin count behavior described above in "nr_foll_pin_acquired",
251 the accounting balances out, so that after doing this::
253 pin_user_pages(huge_page);
254 for (each page in huge_page)
255 unpin_user_page(page);
257 ...the following is expected::
259 nr_foll_pin_released == nr_foll_pin_acquired
261 (...unless it was already out of balance due to a long-term RDMA pin being in
267 dump_page() has been enhanced slightly, to handle these new counting fields, and
268 to better report on compound pages in general. Specifically, for compound pages
269 with order > 1, the exact (hpage_pinned_refcount) pincount is reported.
274 * `Some slow progress on get_user_pages() (Apr 2, 2019) <https://lwn.net/Articles/784574/>`_
275 * `DMA and get_user_pages() (LPC: Dec 12, 2018) <https://lwn.net/Articles/774411/>`_
276 * `The trouble with get_user_pages() (Apr 30, 2018) <https://lwn.net/Articles/753027/>`_
277 * `LWN kernel index: get_user_pages() <https://lwn.net/Kernel/Index/#Memory_management-get_user_pages>`_
279 John Hubbard, October, 2019