| blob | a501792e3b57d1a358b9f8e2bbc1a43d0e250ca0 |
1 Frontswap provides a "transcendent memory" interface for swap pages.
2 In some environments, dramatic performance savings may be obtained because
3 swapped pages are saved in RAM (or a RAM-like device) instead of a swap disk.
5 Frontswap is so named because it can be thought of as the opposite of
6 a "backing" store for a swap device. The storage is assumed to be
7 a synchronous concurrency-safe page-oriented "pseudo-RAM device" conforming
8 to the requirements of transcendent memory (such as Xen's "tmem", or
9 in-kernel compressed memory, aka "zcache", or future RAM-like devices);
10 this pseudo-RAM device is not directly accessible or addressable by the
11 kernel and is of unknown and possibly time-varying size. The driver
12 links itself to frontswap by calling frontswap_register_ops to set the
13 frontswap_ops funcs appropriately and the functions it provides must
14 conform to certain policies as follows:
16 An "init" prepares the device to receive frontswap pages associated
17 with the specified swap device number (aka "type"). A "put_page" will
18 copy the page to transcendent memory and associate it with the type and
19 offset associated with the page. A "get_page" will copy the page, if found,
20 from transcendent memory into kernel memory, but will NOT remove the page
21 from from transcendent memory. A "flush_page" will remove the page from
22 transcendent memory and a "flush_area" will remove ALL pages associated
23 with the swap type (e.g., like swapoff) and notify the "device" to refuse
24 further puts with that swap type.
26 Once a page is successfully put, a matching get on the page will normally
27 succeed. So when the kernel finds itself in a situation where it needs
28 to swap out a page, it first attempts to use frontswap. If the put returns
29 success, the data has been successfully saved to transcendent memory and
30 a disk write and, if the data is later read back, a disk read are avoided.
31 If a put returns failure, transcendent memory has rejected the data, and the
32 page can be written to swap as usual.
34 Note that if a page is put and the page already exists in transcendent memory
35 (a "duplicate" put), either the put succeeds and the data is overwritten,
36 or the put fails AND the page is flushed. This ensures stale data may
37 never be obtained from frontswap.
39 Monitoring and control of frontswap is done by sysfs files in the
40 /sys/kernel/mm/frontswap directory. The effectiveness of frontswap can
41 be measured (across all swap devices) with:
43 curr_pages - number of pages currently contained in frontswap
44 failed_puts - how many put attempts have failed
45 gets - how many gets were attempted (all should succeed)
46 succ_puts - how many put attempts have succeeded
47 flushes - how many flushes were attempted
49 The number of pages currently contained in frontswap can be reduced by root by
50 writing an integer target to curr_pages, which results in a "partial swapoff",
51 thus reducing the number of frontswap pages to that target if memory
52 constraints permit.
54 FAQ
56 1) Where's the value?
58 When a workload starts swapping, performance falls through the floor.
59 Frontswap significantly increases performance in many such workloads by
60 providing a clean, dynamic interface to read and write swap pages to
61 "transcendent memory" that is otherwise not directly addressable to the kernel.
62 This interface is ideal when data is transformed to a different form
63 and size (such as with compression) or secretly moved (as might be
64 useful for write-balancing for some RAM-like devices). Swap pages (and
65 evicted page-cache pages) are a great use for this kind of slower-than-RAM-
66 but-much-faster-than-disk "pseudo-RAM device" and the frontswap (and
67 cleancache) interface to transcendent memory provides a nice way to read
68 and write -- and indirectly "name" -- the pages.
70 In the virtual case, the whole point of virtualization is to statistically
71 multiplex physical resources acrosst the varying demands of multiple
72 virtual machines. This is really hard to do with RAM and efforts to do
73 it well with no kernel changes have essentially failed (except in some
74 well-publicized special-case workloads). Frontswap -- and cleancache --
75 with a fairly small impact on the kernel, provides a huge amount
76 of flexibility for more dynamic, flexible RAM multiplexing.
77 Specifically, the Xen Transcendent Memory backend allows otherwise
78 "fallow" hypervisor-owned RAM to not only be "time-shared" between multiple
79 virtual machines, but the pages can be compressed and deduplicated to
80 optimize RAM utilization. And when guest OS's are induced to surrender
81 underutilized RAM (e.g. with "self-ballooning"), sudden unexpected
82 memory pressure may result in swapping; frontswap allows those pages
83 to be swapped to and from hypervisor RAM if overall host system memory
84 conditions allow.
86 2) Sure there may be performance advantages in some situations, but
87 what's the space/time overhead of frontswap?
89 If CONFIG_FRONTSWAP is disabled, every frontswap hook compiles into
90 nothingness and the only overhead is a few extra bytes per swapon'ed
91 swap device. If CONFIG_FRONTSWAP is enabled but no frontswap "backend"
92 registers, there is one extra global variable compared to zero for
93 every swap page read or written. If CONFIG_FRONTSWAP is enabled
94 AND a frontswap backend registers AND the backend fails every "put"
95 request (i.e. provides no memory despite claiming it might),
96 CPU overhead is still negligible -- and since every frontswap fail
97 precedes a swap page write-to-disk, the system is highly likely
98 to be I/O bound and using a small fraction of a percent of a CPU
99 will be irrelevant anyway.
101 As for space, if CONFIG_FRONTSWAP is enabled AND a frontswap backend
102 registers, one bit is allocated for every swap page for every swap
103 device that is swapon'd. This is added to the EIGHT bits (which
104 was sixteen until about 2.6.34) that the kernel already allocates
105 for every swap page for every swap device that is swapon'd. (Hugh
106 Dickins has observed that frontswap could probably steal one of
107 the existing eight bits, but let's worry about that minor optimization
108 later.) For very large swap disks (which are rare) on a standard
109 4K pagesize, this is 1MB per 32GB swap.
111 3) OK, how about a quick overview of what this frontswap patch does
112 in terms that a kernel hacker can grok?
114 Let's assume that a frontswap "backend" has registered during
115 kernel initialization; this registration indicates that this
116 frontswap backend has access to some "memory" that is not directly
117 accessible by the kernel. Exactly how much memory it provides is
118 entirely dynamic and random.
120 Whenever a swap-device is swapon'd frontswap_init() is called,
121 passing the swap device number (aka "type") as a parameter.
122 This notifies frontswap to expect attempts to "put" swap pages
123 associated with that number.
125 Whenever the swap subsystem is readying a page to write to a swap
126 device (c.f swap_writepage()), frontswap_put_page is called. Frontswap
127 consults with the frontswap backend and if the backend says it does NOT
128 have room, frontswap_put_page returns -1 and the kernel swaps the page
129 to the swap device as normal. Note that the response from the frontswap
130 backend is unpredictable to the kernel; it may choose to never accept a
131 page, it could accept every ninth page, or it might accept every
132 page. But if the backend does accept a page, the data from the page
133 has already been copied and associated with the type and offset,
134 and the backend guarantees the persistence of the data. In this case,
135 frontswap sets a bit in the "frontswap_map" for the swap device
136 corresponding to the page offset on the swap device to which it would
137 otherwise have written the data.
139 When the swap subsystem needs to swap-in a page (swap_readpage()),
140 it first calls frontswap_get_page() which checks the frontswap_map to
141 see if the page was earlier accepted by the frontswap backend. If
142 it was, the page of data is filled from the frontswap backend and
143 the swap-in is complete. If not, the normal swap-in code is
144 executed to obtain the page of data from the real swap device.
146 So every time the frontswap backend accepts a page, a swap device read
147 and (potentially) a swap device write are replaced by a "frontswap backend
148 put" and (possibly) a "frontswap backend get", which are presumably much
149 faster.
151 4) Can't frontswap be configured as a "special" swap device that is
152 just higher priority than any real swap device (e.g. like zswap)?
154 No. Recall that acceptance of any swap page by the frontswap
155 backend is entirely unpredictable. This is critical to the definition
156 of frontswap because it grants completely dynamic discretion to the
157 backend. But since any "put" might fail, there must always be a real
158 slot on a real swap device to swap the page. Thus frontswap must be
159 implemented as a "shadow" to every swapon'd device with the potential
160 capability of holding every page that the swap device might have held
161 and the possibility that it might hold no pages at all.
162 On the downside, this also means that frontswap cannot contain more
163 pages than the total of swapon'd swap devices. For example, if NO
164 swap device is configured on some installation, frontswap is useless.
166 Further, frontswap is entirely synchronous whereas a real swap
167 device is, by definition, asynchronous and uses block I/O. The
168 block I/O layer is not only unnecessary, but may perform "optimizations"
169 that are inappropriate for a RAM-oriented device including delaying
170 the write of some pages for a significant amount of time. Synchrony is
171 required to ensure the dynamicity of the backend and to avoid thorny race
172 conditions that would unnecessarily and greatly complicate frontswap
173 and/or the block I/O subsystem.
175 In a virtualized environment, the dynamicity allows the hypervisor
176 (or host OS) to do "intelligent overcommit". For example, it can
177 choose to accept pages only until host-swapping might be imminent,
178 then force guests to do their own swapping. In zcache, "poorly"
179 compressible pages can be rejected, where "poorly" can itself be defined
180 dynamically depending on current memory constraints.
182 5) Why this weird definition about "duplicate puts"? If a page
183 has been previously successfully put, can't it always be
184 successfully overwritten?
186 Nearly always it can, but no, sometimes it cannot. Consider an example
187 where data is compressed and the original 4K page has been compressed
188 to 1K. Now an attempt is made to overwrite the page with data that
189 is non-compressible and so would take the entire 4K. But the backend
190 has no more space. In this case, the put must be rejected. Whenever
191 frontswap rejects a put that would overwrite, it also must flush
192 the old data and ensure that it is no longer accessible. Since the
193 swap subsystem then writes the new data to the read swap device,
194 this is the correct course of action to ensure coherency.
196 6) What is frontswap_shrink for?
198 When the (non-frontswap) swap subsystem swaps out a page to a real
199 swap device, that page is only taking up low-value pre-allocated disk
200 space. But if frontswap has placed a page in transcendent memory, that
201 page may be taking up valuable real estate. The frontswap_shrink
202 routine allows a process outside of the swap subsystem (such as
203 a userland service via the sysfs interface, or a kernel thread)
204 to force pages out of the memory managed by frontswap and back into
205 kernel-addressable memory.
207 7) Why does the frontswap patch create the new include file swapfile.h?
209 The frontswap code depends on some swap-subsystem-internal data
210 structures that have, over the years, moved back and forth between
211 static and global. This seemed a reasonable compromise: Define
212 them as global but declare them in a new include file that isn't
213 included by the large number of source files that include swap.h.
215 Dan Magenheimer, last updated August 8, 2011