Rev 67 | Rev 69 | Go to most recent revision | Only display areas with differences | Ignore whitespace | Details | Blame | Last modification | View Log | RSS feed
Rev 67 | Rev 68 | ||
---|---|---|---|
1 | <?xml version="1.0" encoding="UTF-8"?> |
1 | <?xml version="1.0" encoding="UTF-8"?> |
2 | <chapter id="mm"> |
2 | <chapter id="mm"> |
3 | <?dbhtml filename="mm.html"?> |
3 | <?dbhtml filename="mm.html"?> |
4 | 4 | ||
5 | <title>Memory management</title> |
5 | <title>Memory management</title> |
6 | 6 | ||
7 | <para>In previous chapters, this book described the scheduling subsystem as |
7 | <para>In previous chapters, this book described the scheduling subsystem as |
8 | the creator of the impression that threads execute in parallel. The memory |
8 | the creator of the impression that threads execute in parallel. The memory |
9 | management subsystem, on the other hand, creates the impression that there |
9 | management subsystem, on the other hand, creates the impression that there |
10 | is enough physical memory for the kernel and that userspace tasks have the |
10 | is enough physical memory for the kernel and that userspace tasks have the |
11 | entire address space only for themselves.</para> |
11 | entire address space only for themselves.</para> |
12 | 12 | ||
13 | <section> |
13 | <section> |
14 | <title>Physical memory management</title> |
14 | <title>Physical memory management</title> |
15 | 15 | ||
16 | <section id="zones_and_frames"> |
16 | <section id="zones_and_frames"> |
17 | <title>Zones and frames</title> |
17 | <title>Zones and frames</title> |
18 | 18 | ||
19 | <para>HelenOS represents continuous areas of physical memory in |
19 | <para>HelenOS represents continuous areas of physical memory in |
20 | structures called frame zones (abbreviated as zones). Each zone contains |
20 | structures called frame zones (abbreviated as zones). Each zone contains |
21 | information about the number of allocated and unallocated physical |
21 | information about the number of allocated and unallocated physical |
22 | memory frames as well as the physical base address of the zone and |
22 | memory frames as well as the physical base address of the zone and |
23 | number of frames contained in it. A zone also contains an array of frame |
23 | number of frames contained in it. A zone also contains an array of frame |
24 | structures describing each frame of the zone and, in the last, but not |
24 | structures describing each frame of the zone and, in the last, but not |
25 | the least important, front, each zone is equipped with a buddy system |
25 | the least important, front, each zone is equipped with a buddy system |
26 | that faciliates effective allocation of power-of-two sized block of |
26 | that faciliates effective allocation of power-of-two sized block of |
27 | frames.</para> |
27 | frames.</para> |
28 | 28 | ||
29 | <para>This organization of physical memory provides good preconditions |
29 | <para>This organization of physical memory provides good preconditions |
30 | for hot-plugging of more zones. There is also one currently unused zone |
30 | for hot-plugging of more zones. There is also one currently unused zone |
31 | attribute: <code>flags</code>. The attribute could be used to give a |
31 | attribute: <code>flags</code>. The attribute could be used to give a |
32 | special meaning to some zones in the future.</para> |
32 | special meaning to some zones in the future.</para> |
33 | 33 | ||
34 | <para>The zones are linked in a doubly-linked list. This might seem a |
34 | <para>The zones are linked in a doubly-linked list. This might seem a |
35 | bit ineffective because the zone list is walked everytime a frame is |
35 | bit ineffective because the zone list is walked everytime a frame is |
36 | allocated or deallocated. However, this does not represent a significant |
36 | allocated or deallocated. However, this does not represent a significant |
37 | performance problem as it is expected that the number of zones will be |
37 | performance problem as it is expected that the number of zones will be |
38 | rather low. Moreover, most architectures merge all zones into |
38 | rather low. Moreover, most architectures merge all zones into |
39 | one.</para> |
39 | one.</para> |
40 | 40 | ||
41 | <para>For each physical memory frame found in a zone, there is a frame |
41 | <para>For each physical memory frame found in a zone, there is a frame |
42 | structure that contains number of references and data used by buddy |
42 | structure that contains number of references and data used by buddy |
43 | system.</para> |
43 | system.</para> |
44 | </section> |
44 | </section> |
45 | 45 | ||
46 | <section id="frame_allocator"> |
46 | <section id="frame_allocator"> |
47 | <title>Frame allocator</title> |
47 | <title>Frame allocator</title> |
48 | 48 | ||
49 | <para>The frame allocator satisfies kernel requests to allocate |
49 | <para>The frame allocator satisfies kernel requests to allocate |
50 | power-of-two sized blocks of physical memory. Because of zonal |
50 | power-of-two sized blocks of physical memory. Because of zonal |
51 | organization of physical memory, the frame allocator is always working |
51 | organization of physical memory, the frame allocator is always working |
52 | within a context of some frame zone. In order to carry out the |
52 | within a context of some frame zone. In order to carry out the |
53 | allocation requests, the frame allocator is tightly integrated with the |
53 | allocation requests, the frame allocator is tightly integrated with the |
54 | buddy system belonging to the zone. The frame allocator is also |
54 | buddy system belonging to the zone. The frame allocator is also |
55 | responsible for updating information about the number of free and busy |
55 | responsible for updating information about the number of free and busy |
56 | frames in the zone. <figure> |
56 | frames in the zone. <figure> |
57 | <mediaobject id="frame_alloc"> |
57 | <mediaobject id="frame_alloc"> |
58 | <imageobject role="html"> |
58 | <imageobject role="html"> |
59 | <imagedata fileref="images/frame_alloc.png" format="PNG" /> |
59 | <imagedata fileref="images/frame_alloc.png" format="PNG" /> |
60 | </imageobject> |
60 | </imageobject> |
61 | 61 | ||
62 | <imageobject role="fop"> |
62 | <imageobject role="fop"> |
63 | <imagedata fileref="images.vector/frame_alloc.svg" format="SVG" /> |
63 | <imagedata fileref="images.vector/frame_alloc.svg" format="SVG" /> |
64 | </imageobject> |
64 | </imageobject> |
65 | </mediaobject> |
65 | </mediaobject> |
66 | 66 | ||
67 | <title>Frame allocator scheme.</title> |
67 | <title>Frame allocator scheme.</title> |
68 | </figure></para> |
68 | </figure></para> |
69 | 69 | ||
70 | <formalpara> |
70 | <formalpara> |
71 | <title>Allocation / deallocation</title> |
71 | <title>Allocation / deallocation</title> |
72 | 72 | ||
73 | <para>Upon allocation request via function <code>frame_alloc</code>, |
73 | <para>Upon allocation request via function <code>frame_alloc</code>, |
74 | the frame allocator first tries to find a zone that can satisfy the |
74 | the frame allocator first tries to find a zone that can satisfy the |
75 | request (i.e. has the required amount of free frames). Once a suitable |
75 | request (i.e. has the required amount of free frames). Once a suitable |
76 | zone is found, the frame allocator uses the buddy allocator on the |
76 | zone is found, the frame allocator uses the buddy allocator on the |
77 | zone's buddy system to perform the allocation. During deallocation, |
77 | zone's buddy system to perform the allocation. During deallocation, |
78 | which is triggered by a call to <code>frame_free</code>, the frame |
78 | which is triggered by a call to <code>frame_free</code>, the frame |
79 | allocator looks up the respective zone that contains the frame being |
79 | allocator looks up the respective zone that contains the frame being |
80 | deallocated. Afterwards, it calls the buddy allocator again, this time |
80 | deallocated. Afterwards, it calls the buddy allocator again, this time |
81 | to take care of deallocation within the zone's buddy system.</para> |
81 | to take care of deallocation within the zone's buddy system.</para> |
82 | </formalpara> |
82 | </formalpara> |
83 | </section> |
83 | </section> |
84 | 84 | ||
85 | <section id="buddy_allocator"> |
85 | <section id="buddy_allocator"> |
86 | <title>Buddy allocator</title> |
86 | <title>Buddy allocator</title> |
87 | 87 | ||
88 | <para>In the buddy system, the memory is broken down into power-of-two |
88 | <para>In the buddy system, the memory is broken down into power-of-two |
89 | sized naturally aligned blocks. These blocks are organized in an array |
89 | sized naturally aligned blocks. These blocks are organized in an array |
90 | of lists, in which the list with index i contains all unallocated blocks |
90 | of lists, in which the list with index i contains all unallocated blocks |
91 | of size <mathphrase>2<superscript>i</superscript></mathphrase>. The |
91 | of size <mathphrase>2<superscript>i</superscript></mathphrase>. The |
92 | index i is called the order of block. Should there be two adjacent |
92 | index i is called the order of block. Should there be two adjacent |
93 | equally sized blocks in the list i<mathphrase />(i.e. buddies), the |
93 | equally sized blocks in the list i<mathphrase />(i.e. buddies), the |
94 | buddy allocator would coalesce them and put the resulting block in list |
94 | buddy allocator would coalesce them and put the resulting block in list |
95 | <mathphrase>i + 1</mathphrase>, provided that the resulting block would |
95 | <mathphrase>i + 1</mathphrase>, provided that the resulting block would |
96 | be naturally aligned. Similarily, when the allocator is asked to |
96 | be naturally aligned. Similarily, when the allocator is asked to |
97 | allocate a block of size |
97 | allocate a block of size |
98 | <mathphrase>2<superscript>i</superscript></mathphrase>, it first tries |
98 | <mathphrase>2<superscript>i</superscript></mathphrase>, it first tries |
99 | to satisfy the request from the list with index i. If the request cannot |
99 | to satisfy the request from the list with index i. If the request cannot |
100 | be satisfied (i.e. the list i is empty), the buddy allocator will try to |
100 | be satisfied (i.e. the list i is empty), the buddy allocator will try to |
101 | allocate and split a larger block from the list with index i + 1. Both |
101 | allocate and split a larger block from the list with index i + 1. Both |
102 | of these algorithms are recursive. The recursion ends either when there |
102 | of these algorithms are recursive. The recursion ends either when there |
103 | are no blocks to coalesce in the former case or when there are no blocks |
103 | are no blocks to coalesce in the former case or when there are no blocks |
104 | that can be split in the latter case.</para> |
104 | that can be split in the latter case.</para> |
105 | 105 | ||
106 | <para>This approach greatly reduces external fragmentation of memory and |
106 | <para>This approach greatly reduces external fragmentation of memory and |
107 | helps in allocating bigger continuous blocks of memory aligned to their |
107 | helps in allocating bigger continuous blocks of memory aligned to their |
108 | size. On the other hand, the buddy allocator suffers increased internal |
108 | size. On the other hand, the buddy allocator suffers increased internal |
109 | fragmentation of memory and is not suitable for general kernel |
109 | fragmentation of memory and is not suitable for general kernel |
110 | allocations. This purpose is better addressed by the <link |
110 | allocations. This purpose is better addressed by the <link |
111 | linkend="slab">slab allocator</link>.<figure> |
111 | linkend="slab">slab allocator</link>.<figure> |
112 | <mediaobject id="buddy_alloc"> |
112 | <mediaobject id="buddy_alloc"> |
113 | <imageobject role="html"> |
113 | <imageobject role="html"> |
114 | <imagedata fileref="images/buddy_alloc.png" format="PNG" /> |
114 | <imagedata fileref="images/buddy_alloc.png" format="PNG" /> |
115 | </imageobject> |
115 | </imageobject> |
116 | 116 | ||
117 | <imageobject role="fop"> |
117 | <imageobject role="fop"> |
118 | <imagedata fileref="images.vector/buddy_alloc.svg" format="SVG" /> |
118 | <imagedata fileref="images.vector/buddy_alloc.svg" format="SVG" /> |
119 | </imageobject> |
119 | </imageobject> |
120 | </mediaobject> |
120 | </mediaobject> |
121 | 121 | ||
122 | <title>Buddy system scheme.</title> |
122 | <title>Buddy system scheme.</title> |
123 | </figure></para> |
123 | </figure></para> |
124 | 124 | ||
125 | <section> |
125 | <section> |
126 | <title>Implementation</title> |
126 | <title>Implementation</title> |
127 | 127 | ||
128 | <para>The buddy allocator is, in fact, an abstract framework wich can |
128 | <para>The buddy allocator is, in fact, an abstract framework wich can |
129 | be easily specialized to serve one particular task. It knows nothing |
129 | be easily specialized to serve one particular task. It knows nothing |
130 | about the nature of memory it helps to allocate. In order to beat the |
130 | about the nature of memory it helps to allocate. In order to beat the |
131 | lack of this knowledge, the buddy allocator exports an interface that |
131 | lack of this knowledge, the buddy allocator exports an interface that |
132 | each of its clients is required to implement. When supplied with an |
132 | each of its clients is required to implement. When supplied with an |
133 | implementation of this interface, the buddy allocator can use |
133 | implementation of this interface, the buddy allocator can use |
134 | specialized external functions to find a buddy for a block, split and |
134 | specialized external functions to find a buddy for a block, split and |
135 | coalesce blocks, manipulate block order and mark blocks busy or |
135 | coalesce blocks, manipulate block order and mark blocks busy or |
136 | available.</para> |
136 | available.</para> |
137 | 137 | ||
138 | <formalpara> |
138 | <formalpara> |
139 | <title>Data organization</title> |
139 | <title>Data organization</title> |
140 | 140 | ||
141 | <para>Each entity allocable by the buddy allocator is required to |
141 | <para>Each entity allocable by the buddy allocator is required to |
142 | contain space for storing block order number and a link variable |
142 | contain space for storing block order number and a link variable |
143 | used to interconnect blocks within the same order.</para> |
143 | used to interconnect blocks within the same order.</para> |
144 | 144 | ||
145 | <para>Whatever entities are allocated by the buddy allocator, the |
145 | <para>Whatever entities are allocated by the buddy allocator, the |
146 | first entity within a block is used to represent the entire block. |
146 | first entity within a block is used to represent the entire block. |
147 | The first entity keeps the order of the whole block. Other entities |
147 | The first entity keeps the order of the whole block. Other entities |
148 | within the block are assigned the magic value |
148 | within the block are assigned the magic value |
149 | <constant>BUDDY_INNER_BLOCK</constant>. This is especially important |
149 | <constant>BUDDY_INNER_BLOCK</constant>. This is especially important |
150 | for effective identification of buddies in a one-dimensional array |
150 | for effective identification of buddies in a one-dimensional array |
151 | because the entity that represents a potential buddy cannot be |
151 | because the entity that represents a potential buddy cannot be |
152 | associated with <constant>BUDDY_INNER_BLOCK</constant> (i.e. if it |
152 | associated with <constant>BUDDY_INNER_BLOCK</constant> (i.e. if it |
153 | is associated with <constant>BUDDY_INNER_BLOCK</constant> then it is |
153 | is associated with <constant>BUDDY_INNER_BLOCK</constant> then it is |
154 | not a buddy).</para> |
154 | not a buddy).</para> |
155 | </formalpara> |
155 | </formalpara> |
156 | </section> |
156 | </section> |
157 | </section> |
157 | </section> |
158 | 158 | ||
159 | <section id="slab"> |
159 | <section id="slab"> |
160 | <title>Slab allocator</title> |
160 | <title>Slab allocator</title> |
161 | 161 | ||
162 | <para>The majority of memory allocation requests in the kernel is for |
162 | <para>The majority of memory allocation requests in the kernel is for |
163 | small, frequently used data structures. The basic idea behind the slab |
163 | small, frequently used data structures. The basic idea behind the slab |
164 | allocator is that commonly used objects are preallocated in continuous |
164 | allocator is that commonly used objects are preallocated in continuous |
165 | areas of physical memory called slabs<footnote> |
165 | areas of physical memory called slabs<footnote> |
166 | <para>Slabs are in fact blocks of physical memory frames allocated |
166 | <para>Slabs are in fact blocks of physical memory frames allocated |
167 | from the frame allocator.</para> |
167 | from the frame allocator.</para> |
168 | </footnote>. Whenever an object is to be allocated, the slab allocator |
168 | </footnote>. Whenever an object is to be allocated, the slab allocator |
169 | returns the first available item from a suitable slab corresponding to |
169 | returns the first available item from a suitable slab corresponding to |
170 | the object type<footnote> |
170 | the object type<footnote> |
171 | <para>The mechanism is rather more complicated, see the next |
171 | <para>The mechanism is rather more complicated, see the next |
172 | paragraph.</para> |
172 | paragraph.</para> |
173 | </footnote>. Due to the fact that the sizes of the requested and |
173 | </footnote>. Due to the fact that the sizes of the requested and |
174 | allocated object match, the slab allocator significantly reduces |
174 | allocated object match, the slab allocator significantly reduces |
175 | internal fragmentation.</para> |
175 | internal fragmentation.</para> |
176 | 176 | ||
177 | <para>Slabs of one object type are organized in a structure called slab |
177 | <para>Slabs of one object type are organized in a structure called slab |
178 | cache. There are ususally more slabs in the slab cache, depending on |
178 | cache. There are ususally more slabs in the slab cache, depending on |
179 | previous allocations. If the the slab cache runs out of available slabs, |
179 | previous allocations. If the the slab cache runs out of available slabs, |
180 | new slabs are allocated. In order to exploit parallelism and to avoid |
180 | new slabs are allocated. In order to exploit parallelism and to avoid |
181 | locking of shared spinlocks, slab caches can have variants of |
181 | locking of shared spinlocks, slab caches can have variants of |
182 | processor-private slabs called magazines. On each processor, there is a |
182 | processor-private slabs called magazines. On each processor, there is a |
183 | two-magazine cache. Full magazines that are not part of any |
183 | two-magazine cache. Full magazines that are not part of any |
184 | per-processor magazine cache are stored in a global list of full |
184 | per-processor magazine cache are stored in a global list of full |
185 | magazines.</para> |
185 | magazines.</para> |
186 | 186 | ||
187 | <para>Each object begins its life in a slab. When it is allocated from |
187 | <para>Each object begins its life in a slab. When it is allocated from |
188 | there, the slab allocator calls a constructor that is registered in the |
188 | there, the slab allocator calls a constructor that is registered in the |
189 | respective slab cache. The constructor initializes and brings the object |
189 | respective slab cache. The constructor initializes and brings the object |
190 | into a known state. The object is then used by the user. When the user |
190 | into a known state. The object is then used by the user. When the user |
191 | later frees the object, the slab allocator puts it into a processor |
191 | later frees the object, the slab allocator puts it into a processor |
192 | private magazine cache, from where it can be precedently allocated |
192 | private magazine cache, from where it can be precedently allocated |
193 | again. Note that allocations satisfied from a magazine are already |
193 | again. Note that allocations satisfied from a magazine are already |
194 | initialized by the constructor. When both of the processor cached |
194 | initialized by the constructor. When both of the processor cached |
195 | magazines get full, the allocator will move one of the magazines to the |
195 | magazines get full, the allocator will move one of the magazines to the |
196 | list of full magazines. Similarily, when allocating from an empty |
196 | list of full magazines. Similarily, when allocating from an empty |
197 | processor magazine cache, the kernel will reload only one magazine from |
197 | processor magazine cache, the kernel will reload only one magazine from |
198 | the list of full magazines. In other words, the slab allocator tries to |
198 | the list of full magazines. In other words, the slab allocator tries to |
199 | keep the processor magazine cache only half-full in order to prevent |
199 | keep the processor magazine cache only half-full in order to prevent |
200 | thrashing when allocations and deallocations interleave on magazine |
200 | thrashing when allocations and deallocations interleave on magazine |
201 | boundaries.</para> |
201 | boundaries.</para> |
202 | 202 | ||
203 | <para>Should HelenOS run short of memory, it would start deallocating |
203 | <para>Should HelenOS run short of memory, it would start deallocating |
204 | objects from magazines, calling slab cache destructor on them and |
204 | objects from magazines, calling slab cache destructor on them and |
205 | putting them back into slabs. When a slab contanins no allocated object, |
205 | putting them back into slabs. When a slab contanins no allocated object, |
206 | it is immediately freed.</para> |
206 | it is immediately freed.</para> |
207 | 207 | ||
208 | <para><figure> |
208 | <para><figure> |
209 | <mediaobject id="slab_alloc"> |
209 | <mediaobject id="slab_alloc"> |
210 | <imageobject role="html"> |
210 | <imageobject role="html"> |
211 | <imagedata fileref="images/slab_alloc.png" format="PNG" /> |
211 | <imagedata fileref="images/slab_alloc.png" format="PNG" /> |
212 | </imageobject> |
212 | </imageobject> |
213 | </mediaobject> |
213 | </mediaobject> |
214 | 214 | ||
215 | <title>Slab allocator scheme.</title> |
215 | <title>Slab allocator scheme.</title> |
216 | </figure></para> |
216 | </figure></para> |
217 | 217 | ||
218 | <section> |
218 | <section> |
219 | <title>Implementation</title> |
219 | <title>Implementation</title> |
220 | 220 | ||
221 | <para>The slab allocator is closely modelled after OpenSolaris slab |
221 | <para>The slab allocator is closely modelled after OpenSolaris slab |
222 | allocator by Jeff Bonwick and Jonathan Adams with the following |
222 | allocator by Jeff Bonwick and Jonathan Adams with the following |
223 | exceptions:<itemizedlist> |
223 | exceptions:<itemizedlist> |
224 | <listitem> |
224 | <listitem> |
225 | empty slabs are immediately deallocated |
225 | empty slabs are immediately deallocated |
226 | </listitem> |
226 | </listitem> |
227 | 227 | ||
228 | <listitem> |
228 | <listitem> |
229 | <para>empty magazines are deallocated when not needed</para> |
229 | <para>empty magazines are deallocated when not needed</para> |
230 | </listitem> |
230 | </listitem> |
231 | </itemizedlist> Following features are not currently supported but |
231 | </itemizedlist> Following features are not currently supported but |
232 | would be easy to do: <itemizedlist> |
232 | would be easy to do: <itemizedlist> |
233 | <listitem> |
233 | <listitem> |
234 | cache coloring |
234 | cache coloring |
235 | </listitem> |
235 | </listitem> |
236 | 236 | ||
237 | <listitem> |
237 | <listitem> |
238 | dynamic magazine grow (different magazine sizes are already supported, but the allocation strategy would need to be adjusted) |
238 | dynamic magazine grow (different magazine sizes are already supported, but the allocation strategy would need to be adjusted) |
239 | </listitem> |
239 | </listitem> |
240 | </itemizedlist></para> |
240 | </itemizedlist></para> |
241 | 241 | ||
242 | <section> |
242 | <section> |
243 | <title>Magazine layer</title> |
243 | <title>Magazine layer</title> |
244 | 244 | ||
245 | <para>Due to the extensive bottleneck on SMP architures, caused by |
245 | <para>Due to the extensive bottleneck on SMP architures, caused by |
246 | global slab locking mechanism, making processing of all slab |
246 | global slab locking mechanism, making processing of all slab |
247 | allocation requests serialized, a new layer was introduced to the |
247 | allocation requests serialized, a new layer was introduced to the |
248 | classic slab allocator design. Slab allocator was extended to |
248 | classic slab allocator design. Slab allocator was extended to |
249 | support per-CPU caches 'magazines' to achieve good SMP scaling. |
249 | support per-CPU caches 'magazines' to achieve good SMP scaling. |
250 | <termdef>Slab SMP perfromance bottleneck was resolved by introducing |
250 | <termdef>Slab SMP perfromance bottleneck was resolved by introducing |
251 | a per-CPU caching scheme called as <glossterm>magazine |
251 | a per-CPU caching scheme called as <glossterm>magazine |
252 | layer</glossterm></termdef>.</para> |
252 | layer</glossterm></termdef>.</para> |
253 | 253 | ||
254 | <para>Magazine is a N-element cache of objects, so each magazine can |
254 | <para>Magazine is a N-element cache of objects, so each magazine can |
255 | satisfy N allocations. Magazine behaves like a automatic weapon |
255 | satisfy N allocations. Magazine behaves like a automatic weapon |
256 | magazine (LIFO, stack), so the allocation/deallocation become simple |
256 | magazine (LIFO, stack), so the allocation/deallocation become simple |
257 | push/pop pointer operation. Trick is that CPU does not access global |
257 | push/pop pointer operation. Trick is that CPU does not access global |
258 | slab allocator data during the allocation from its magazine, thus |
258 | slab allocator data during the allocation from its magazine, thus |
259 | making possible parallel allocations between CPUs.</para> |
259 | making possible parallel allocations between CPUs.</para> |
260 | 260 | ||
261 | <para>Implementation also requires adding another feature as the |
261 | <para>Implementation also requires adding another feature as the |
262 | CPU-bound magazine is actually a pair of magazines to avoid |
262 | CPU-bound magazine is actually a pair of magazines to avoid |
263 | thrashing when during allocation/deallocatiion of 1 item at the |
263 | thrashing when during allocation/deallocatiion of 1 item at the |
264 | magazine size boundary. LIFO order is enforced, which should avoid |
264 | magazine size boundary. LIFO order is enforced, which should avoid |
265 | fragmentation as much as possible.</para> |
265 | fragmentation as much as possible.</para> |
266 | 266 | ||
267 | <para>Another important entity of magazine layer is the common full |
267 | <para>Another important entity of magazine layer is the common full |
268 | magazine list (also called a depot), that stores full magazines that |
268 | magazine list (also called a depot), that stores full magazines that |
269 | may be used by any of the CPU magazine caches to reload active CPU |
269 | may be used by any of the CPU magazine caches to reload active CPU |
270 | magazine. This list of magazines can be pre-filled with full |
270 | magazine. This list of magazines can be pre-filled with full |
271 | magazines during initialization, but in current implementation it is |
271 | magazines during initialization, but in current implementation it is |
272 | filled during object deallocation, when CPU magazine becomes |
272 | filled during object deallocation, when CPU magazine becomes |
273 | full.</para> |
273 | full.</para> |
274 | 274 | ||
275 | <para>Slab allocator control structures are allocated from special |
275 | <para>Slab allocator control structures are allocated from special |
276 | slabs, that are marked by special flag, indicating that it should |
276 | slabs, that are marked by special flag, indicating that it should |
277 | not be used for slab magazine layer. This is done to avoid possible |
277 | not be used for slab magazine layer. This is done to avoid possible |
278 | infinite recursions and deadlock during conventional slab allocaiton |
278 | infinite recursions and deadlock during conventional slab allocaiton |
279 | requests.</para> |
279 | requests.</para> |
280 | </section> |
280 | </section> |
281 | 281 | ||
282 | <section> |
282 | <section> |
283 | <title>Allocation/deallocation</title> |
283 | <title>Allocation/deallocation</title> |
284 | 284 | ||
285 | <para>Every cache contains list of full slabs and list of partialy |
285 | <para>Every cache contains list of full slabs and list of partialy |
286 | full slabs. Empty slabs are immediately freed (thrashing will be |
286 | full slabs. Empty slabs are immediately freed (thrashing will be |
287 | avoided because of magazines).</para> |
287 | avoided because of magazines).</para> |
288 | 288 | ||
289 | <para>The SLAB allocator allocates lots of space and does not free |
289 | <para>The SLAB allocator allocates lots of space and does not free |
290 | it. When frame allocator fails to allocate the frame, it calls |
290 | it. When frame allocator fails to allocate the frame, it calls |
291 | slab_reclaim(). It tries 'light reclaim' first, then brutal reclaim. |
291 | slab_reclaim(). It tries 'light reclaim' first, then brutal reclaim. |
292 | The light reclaim releases slabs from cpu-shared magazine-list, |
292 | The light reclaim releases slabs from cpu-shared magazine-list, |
293 | until at least 1 slab is deallocated in each cache (this algorithm |
293 | until at least 1 slab is deallocated in each cache (this algorithm |
294 | should probably change). The brutal reclaim removes all cached |
294 | should probably change). The brutal reclaim removes all cached |
295 | objects, even from CPU-bound magazines.</para> |
295 | objects, even from CPU-bound magazines.</para> |
296 | 296 | ||
297 | <formalpara> |
297 | <formalpara> |
298 | <title>Allocation</title> |
298 | <title>Allocation</title> |
299 | 299 | ||
300 | <para><emphasis>Step 1.</emphasis> When it comes to the allocation |
300 | <para><emphasis>Step 1.</emphasis> When it comes to the allocation |
301 | request, slab allocator first of all checks availability of memory |
301 | request, slab allocator first of all checks availability of memory |
302 | in local CPU-bound magazine. If it is there, we would just "pop" |
302 | in local CPU-bound magazine. If it is there, we would just "pop" |
303 | the CPU magazine and return the pointer to object.</para> |
303 | the CPU magazine and return the pointer to object.</para> |
304 | 304 | ||
305 | <para><emphasis>Step 2.</emphasis> If the CPU-bound magazine is |
305 | <para><emphasis>Step 2.</emphasis> If the CPU-bound magazine is |
306 | empty, allocator will attempt to reload magazin, swapping it with |
306 | empty, allocator will attempt to reload magazin, swapping it with |
307 | second CPU magazine and returns to the first step.</para> |
307 | second CPU magazine and returns to the first step.</para> |
308 | 308 | ||
309 | <para><emphasis>Step 3.</emphasis> Now we are in the situation |
309 | <para><emphasis>Step 3.</emphasis> Now we are in the situation |
310 | when both CPU-bound magazines are empty, which makes allocator to |
310 | when both CPU-bound magazines are empty, which makes allocator to |
311 | access shared full-magazines depot to reload CPU-bound magazines. |
311 | access shared full-magazines depot to reload CPU-bound magazines. |
312 | If reload is succesful (meaning there are full magazines in depot) |
312 | If reload is succesful (meaning there are full magazines in depot) |
313 | algoritm continues at Step 1.</para> |
313 | algoritm continues at Step 1.</para> |
314 | 314 | ||
315 | <para><emphasis>Step 4.</emphasis> Final step of the allocation. |
315 | <para><emphasis>Step 4.</emphasis> Final step of the allocation. |
316 | In this step object is allocated from the conventional slab layer |
316 | In this step object is allocated from the conventional slab layer |
317 | and pointer is returned.</para> |
317 | and pointer is returned.</para> |
318 | </formalpara> |
318 | </formalpara> |
319 | 319 | ||
320 | <formalpara> |
320 | <formalpara> |
321 | <title>Deallocation</title> |
321 | <title>Deallocation</title> |
322 | 322 | ||
323 | <para><emphasis>Step 1.</emphasis> During deallocation request, |
323 | <para><emphasis>Step 1.</emphasis> During deallocation request, |
324 | slab allocator will check if the local CPU-bound magazine is not |
324 | slab allocator will check if the local CPU-bound magazine is not |
325 | full. In this case we will just push the pointer to this |
325 | full. In this case we will just push the pointer to this |
326 | magazine.</para> |
326 | magazine.</para> |
327 | 327 | ||
328 | <para><emphasis>Step 2.</emphasis> If the CPU-bound magazine is |
328 | <para><emphasis>Step 2.</emphasis> If the CPU-bound magazine is |
329 | full, allocator will attempt to reload magazin, swapping it with |
329 | full, allocator will attempt to reload magazin, swapping it with |
330 | second CPU magazine and returns to the first step.</para> |
330 | second CPU magazine and returns to the first step.</para> |
331 | 331 | ||
332 | <para><emphasis>Step 3.</emphasis> Now we are in the situation |
332 | <para><emphasis>Step 3.</emphasis> Now we are in the situation |
333 | when both CPU-bound magazines are full, which makes allocator to |
333 | when both CPU-bound magazines are full, which makes allocator to |
334 | access shared full-magazines depot to put one of the magazines to |
334 | access shared full-magazines depot to put one of the magazines to |
335 | the depot and creating new empty magazine. Algoritm continues at |
335 | the depot and creating new empty magazine. Algoritm continues at |
336 | Step 1.</para> |
336 | Step 1.</para> |
337 | </formalpara> |
337 | </formalpara> |
338 | </section> |
338 | </section> |
339 | </section> |
339 | </section> |
340 | </section> |
340 | </section> |
341 | 341 | ||
342 | <!-- End of Physmem --> |
342 | <!-- End of Physmem --> |
343 | </section> |
343 | </section> |
344 | 344 | ||
345 | <section> |
345 | <section> |
346 | <title>Virtual memory management</title> |
346 | <title>Virtual memory management</title> |
347 | 347 | ||
348 | <section> |
348 | <section> |
349 | <title>Introduction</title> |
349 | <title>Introduction</title> |
350 | 350 | ||
351 | <para>Virtual memory is a special memory management technique, used by |
351 | <para>Virtual memory is a special memory management technique, used by |
352 | kernel to achieve a bunch of mission critical goals. <itemizedlist> |
352 | kernel to achieve a bunch of mission critical goals. <itemizedlist> |
353 | <listitem> |
353 | <listitem> |
354 | Isolate each task from other tasks that are running on the system at the same time. |
354 | Isolate each task from other tasks that are running on the system at the same time. |
355 | </listitem> |
355 | </listitem> |
356 | 356 | ||
357 | <listitem> |
357 | <listitem> |
358 | Allow to allocate more memory, than is actual physical memory size of the machine. |
358 | Allow to allocate more memory, than is actual physical memory size of the machine. |
359 | </listitem> |
359 | </listitem> |
360 | 360 | ||
361 | <listitem> |
361 | <listitem> |
362 | Allowing, in general, to load and execute two programs that are linked on the same address without complicated relocations. |
362 | Allowing, in general, to load and execute two programs that are linked on the same address without complicated relocations. |
363 | </listitem> |
363 | </listitem> |
364 | </itemizedlist></para> |
364 | </itemizedlist></para> |
365 | 365 | ||
366 | <para><!-- |
366 | <para><!-- |
367 | 367 | ||
368 | TLB shootdown ASID/ASID:PAGE/ALL. |
- | |
369 | TLB shootdown requests can come in asynchroniously |
- | |
370 | so there is a cache of TLB shootdown requests. Upon cache overflow TLB shootdown ALL is executed |
- | |
371 | - | ||
372 | - | ||
373 | <para> |
368 | <para> |
374 | Address spaces. Address space area (B+ tree). Only for uspace. Set of syscalls (shrink/extend etc). |
369 | Address spaces. Address space area (B+ tree). Only for uspace. Set of syscalls (shrink/extend etc). |
375 | Special address space area type - device - prohibits shrink/extend syscalls to call on it. |
370 | Special address space area type - device - prohibits shrink/extend syscalls to call on it. |
376 | Address space has link to mapping tables (hierarchical - per Address space, hash - global tables). |
371 | Address space has link to mapping tables (hierarchical - per Address space, hash - global tables). |
377 | </para> |
372 | </para> |
378 | 373 | ||
379 | --></para> |
374 | --></para> |
380 | </section> |
375 | </section> |
381 | 376 | ||
382 | <section> |
377 | <section> |
383 | <title>Paging</title> |
- | |
384 | - | ||
385 | <para>Virtual memory is usually using paged memory model, where virtual |
- | |
386 | memory address space is divided into the <emphasis>pages</emphasis> |
- | |
387 | (usually having size 4096 bytes) and physical memory is divided into the |
- | |
388 | frames (same sized as a page, of course). Each page may be mapped to |
- | |
389 | some frame and then, upon memory access to the virtual address, CPU |
- | |
390 | performs <emphasis>address translation</emphasis> during the instruction |
- | |
391 | execution. Non-existing mapping generates page fault exception, calling |
- | |
392 | kernel exception handler, thus allowing kernel to manipulate rules of |
- | |
393 | memory access. Information for pages mapping is stored by kernel in the |
- | |
394 | <link linkend="page_tables">page tables</link></para> |
- | |
395 | - | ||
396 | <para>The majority of the architectures use multi-level page tables, |
- | |
397 | which means need to access physical memory several times before getting |
- | |
398 | physical address. This fact would make serios performance overhead in |
- | |
399 | virtual memory management. To avoid this <link linkend="tlb">Traslation |
- | |
400 | Lookaside Buffer (TLB)</link> is used.</para> |
- | |
401 | </section> |
- | |
402 | - | ||
403 | <section> |
- | |
404 | <title>Address spaces</title> |
378 | <title>Address spaces</title> |
405 | 379 | ||
406 | <section> |
380 | <section> |
407 | <title>Address space areas</title> |
381 | <title>Address space areas</title> |
408 | 382 | ||
409 | <para>Each address space consists of mutually disjunctive continuous |
383 | <para>Each address space consists of mutually disjunctive continuous |
410 | address space areas. Address space area is precisely defined by its |
384 | address space areas. Address space area is precisely defined by its |
411 | base address and the number of frames/pages is contains.</para> |
385 | base address and the number of frames/pages is contains.</para> |
412 | 386 | ||
413 | <para>Address space area , that define behaviour and permissions on |
387 | <para>Address space area , that define behaviour and permissions on |
414 | the particular area. <itemizedlist> |
388 | the particular area. <itemizedlist> |
415 | <listitem> |
389 | <listitem> |
416 | 390 | ||
417 | 391 | ||
418 | <emphasis>AS_AREA_READ</emphasis> |
392 | <emphasis>AS_AREA_READ</emphasis> |
419 | 393 | ||
420 | flag indicates reading permission. |
394 | flag indicates reading permission. |
421 | </listitem> |
395 | </listitem> |
422 | 396 | ||
423 | <listitem> |
397 | <listitem> |
424 | 398 | ||
425 | 399 | ||
426 | <emphasis>AS_AREA_WRITE</emphasis> |
400 | <emphasis>AS_AREA_WRITE</emphasis> |
427 | 401 | ||
428 | flag indicates writing permission. |
402 | flag indicates writing permission. |
429 | </listitem> |
403 | </listitem> |
430 | 404 | ||
431 | <listitem> |
405 | <listitem> |
432 | 406 | ||
433 | 407 | ||
434 | <emphasis>AS_AREA_EXEC</emphasis> |
408 | <emphasis>AS_AREA_EXEC</emphasis> |
435 | 409 | ||
436 | flag indicates code execution permission. Some architectures do not support execution persmission restriction. In this case this flag has no effect. |
410 | flag indicates code execution permission. Some architectures do not support execution persmission restriction. In this case this flag has no effect. |
437 | </listitem> |
411 | </listitem> |
438 | 412 | ||
439 | <listitem> |
413 | <listitem> |
440 | 414 | ||
441 | 415 | ||
442 | <emphasis>AS_AREA_DEVICE</emphasis> |
416 | <emphasis>AS_AREA_DEVICE</emphasis> |
443 | 417 | ||
444 | marks area as mapped to the device memory. |
418 | marks area as mapped to the device memory. |
445 | </listitem> |
419 | </listitem> |
446 | </itemizedlist></para> |
420 | </itemizedlist></para> |
447 | 421 | ||
448 | <para>Kernel provides possibility tasks create/expand/shrink/share its |
422 | <para>Kernel provides possibility tasks create/expand/shrink/share its |
449 | address space via the set of syscalls.</para> |
423 | address space via the set of syscalls.</para> |
450 | </section> |
424 | </section> |
451 | 425 | ||
452 | <section> |
426 | <section> |
453 | <title>Address Space ID (ASID)</title> |
427 | <title>Address Space ID (ASID)</title> |
454 | 428 | ||
455 | <para>When switching to the different task, kernel also require to |
429 | <para>When switching to the different task, kernel also require to |
456 | switch mappings to the different address space. In case TLB cannot |
430 | switch mappings to the different address space. In case TLB cannot |
457 | distinguish address space mappings, all mapping information in TLB |
431 | distinguish address space mappings, all mapping information in TLB |
458 | from the old address space must be flushed, which can create certain |
432 | from the old address space must be flushed, which can create certain |
459 | uncessary overhead during the task switching. To avoid this, some |
433 | uncessary overhead during the task switching. To avoid this, some |
460 | architectures have capability to segregate different address spaces on |
434 | architectures have capability to segregate different address spaces on |
461 | hardware level introducing the address space identifier as a part of |
435 | hardware level introducing the address space identifier as a part of |
462 | TLB record, telling the virtual address space translation unit to |
436 | TLB record, telling the virtual address space translation unit to |
463 | which address space this record is applicable.</para> |
437 | which address space this record is applicable.</para> |
464 | 438 | ||
465 | <para>HelenOS kernel can take advantage of this hardware supported |
439 | <para>HelenOS kernel can take advantage of this hardware supported |
466 | identifier by having an ASID abstraction which is somehow related to |
440 | identifier by having an ASID abstraction which is somehow related to |
467 | the corresponding architecture identifier. I.e. on ia64 kernel ASID is |
441 | the corresponding architecture identifier. I.e. on ia64 kernel ASID is |
468 | derived from RID (region identifier) and on the mips32 kernel ASID is |
442 | derived from RID (region identifier) and on the mips32 kernel ASID is |
469 | actually the hardware identifier. As expected, this ASID information |
443 | actually the hardware identifier. As expected, this ASID information |
470 | record is the part of <emphasis>as_t</emphasis> structure.</para> |
444 | record is the part of <emphasis>as_t</emphasis> structure.</para> |
471 | 445 | ||
472 | <para>Due to the hardware limitations, hardware ASID has limited |
446 | <para>Due to the hardware limitations, hardware ASID has limited |
473 | length from 8 bits on ia64 to 24 bits on mips32, which makes it |
447 | length from 8 bits on ia64 to 24 bits on mips32, which makes it |
474 | impossible to use it as unique address space identifier for all tasks |
448 | impossible to use it as unique address space identifier for all tasks |
475 | running in the system. In such situations special ASID stealing |
449 | running in the system. In such situations special ASID stealing |
476 | algoritm is used, which takes ASID from inactive task and assigns it |
450 | algoritm is used, which takes ASID from inactive task and assigns it |
477 | to the active task.</para> |
451 | to the active task.</para> |
478 | 452 | ||
479 | <para><classname>ASID stealing algoritm here.</classname></para> |
453 | <para><classname>ASID stealing algoritm here.</classname></para> |
480 | </section> |
454 | </section> |
481 | </section> |
455 | </section> |
482 | 456 | ||
483 | <section> |
457 | <section> |
484 | <title>Virtual address translation</title> |
458 | <title>Virtual address translation</title> |
485 | 459 | ||
486 | <section id="page_tables"> |
460 | <section id="paging"> |
487 | <title>Page tables</title> |
461 | <title>Paging</title> |
488 | 462 | ||
489 | <para>HelenOS kernel has two different approaches to the paging |
- | |
490 | implementation: <emphasis>4 level page tables</emphasis> and |
- | |
491 | <emphasis>global hash tables</emphasis>, which are accessible via |
- | |
492 | generic paging abstraction layer. Such different functionality was |
- | |
493 | caused by the major architectural differences between supported |
- | |
494 | platforms. This abstraction is implemented with help of the global |
- | |
495 | structure of pointers to basic mapping functions |
- | |
496 | <emphasis>page_mapping_operations</emphasis>. To achieve different |
- | |
497 | functionality of page tables, corresponding layer must implement |
- | |
498 | functions, declared in |
463 | <section> |
499 | <emphasis>page_mapping_operations</emphasis></para> |
464 | <title>Introduction</title> |
500 | 465 | ||
- | 466 | <para>Virtual memory is usually using paged memory model, where |
|
- | 467 | virtual memory address space is divided into the |
|
- | 468 | <emphasis>pages</emphasis> (usually having size 4096 bytes) and |
|
- | 469 | physical memory is divided into the frames (same sized as a page, of |
|
- | 470 | course). Each page may be mapped to some frame and then, upon memory |
|
- | 471 | access to the virtual address, CPU performs <emphasis>address |
|
- | 472 | translation</emphasis> during the instruction execution. |
|
- | 473 | Non-existing mapping generates page fault exception, calling kernel |
|
- | 474 | exception handler, thus allowing kernel to manipulate rules of |
|
- | 475 | memory access. Information for pages mapping is stored by kernel in |
|
- | 476 | the <link linkend="page_tables">page tables</link></para> |
|
- | 477 | ||
- | 478 | <para>The majority of the architectures use multi-level page tables, |
|
- | 479 | which means need to access physical memory several times before |
|
- | 480 | getting physical address. This fact would make serios performance |
|
- | 481 | overhead in virtual memory management. To avoid this <link |
|
- | 482 | linkend="tlb">Traslation Lookaside Buffer (TLB)</link> is |
|
501 | <formalpara> |
483 | used.</para> |
- | 484 | ||
- | 485 | <para>HelenOS kernel has two different approaches to the paging |
|
- | 486 | implementation: <emphasis>4 level page tables</emphasis> and |
|
- | 487 | <emphasis>global hash table</emphasis>, which are accessible via |
|
- | 488 | generic paging abstraction layer. Such different functionality was |
|
- | 489 | caused by the major architectural differences between supported |
|
- | 490 | platforms. This abstraction is implemented with help of the global |
|
- | 491 | structure of pointers to basic mapping functions |
|
- | 492 | <emphasis>page_mapping_operations</emphasis>. To achieve different |
|
- | 493 | functionality of page tables, corresponding layer must implement |
|
- | 494 | functions, declared in |
|
- | 495 | <emphasis>page_mapping_operations</emphasis></para> |
|
- | 496 | ||
- | 497 | <para>Thanks to the abstract paging interface, there was a place |
|
- | 498 | left for more paging implementations (besides already implemented |
|
- | 499 | hieararchical page tables and hash table), for example B-Tree based |
|
- | 500 | page tables.</para> |
|
- | 501 | </section> |
|
- | 502 | ||
- | 503 | <section> |
|
502 | <title>4-level page tables</title> |
504 | <title>Hierarchical 4-level page tables</title> |
503 | 505 | ||
504 | <para>4-level page tables are the generalization of the hardware |
506 | <para>Hierarchical 4-level page tables are the generalization of the |
- | 507 | hardware capabilities of most architectures. Each address space has |
|
505 | capabilities of several architectures.<itemizedlist> |
508 | its own page tables.<itemizedlist> |
506 | <listitem> |
509 | <listitem> |
507 | ia32 uses 2-level page tables, with full hardware support. |
510 | ia32 uses 2-level page tables, with full hardware support. |
508 | </listitem> |
511 | </listitem> |
509 | 512 | ||
510 | <listitem> |
513 | <listitem> |
511 | amd64 uses 4-level page tables, also coming with full hardware support. |
514 | amd64 uses 4-level page tables, also coming with full hardware support. |
512 | </listitem> |
515 | </listitem> |
513 | 516 | ||
514 | <listitem> |
517 | <listitem> |
515 | mips and ppc32 have 2-level tables, software simulated support. |
518 | mips and ppc32 have 2-level tables, software simulated support. |
516 | </listitem> |
519 | </listitem> |
517 | </itemizedlist></para> |
520 | </itemizedlist></para> |
518 | </formalpara> |
521 | </section> |
- | 522 | ||
- | 523 | <section> |
|
- | 524 | <title>Global hash table</title> |
|
519 | 525 | ||
- | 526 | <para>Implementation of the global hash table was encouraged by the |
|
- | 527 | ia64 architecture support. One of the major differences between |
|
- | 528 | global hash table and hierarchical tables is that global hash table |
|
- | 529 | exists only once in the system and the hierarchical tables are |
|
520 | <formalpara> |
530 | maintained per address space.</para> |
- | 531 | ||
- | 532 | <para>Thus, hash table contains information about all address spaces |
|
- | 533 | mappings in the system, so, the hash of an entry must contain |
|
- | 534 | information of both address space pointer or id and the virtual |
|
- | 535 | address of the page. Generic hash table implementation assumes that |
|
- | 536 | the addresses of the pointers to the address spaces are likely to be |
|
- | 537 | on the close addresses, so it uses least significant bits for hash; |
|
- | 538 | also it assumes that the virtual page addresses have roughly the |
|
- | 539 | same probability of occurring, so the least significant bits of VPN |
|
521 | <title>Global hash tables</title> |
540 | compose the hash index.</para> |
522 | 541 | ||
523 | <para>- global page hash table: existuje jen jedna v celem systemu |
- | |
524 | (vyuziva ji ia64), pozn. ia64 ma zatim vypnuty VHPT. Pouziva se |
- | |
525 | genericke hash table s oddelenymi collision chains. ASID support is |
- | |
526 | required to use global hash tables.</para> |
542 | <para>Collision chains ...</para> |
527 | </formalpara> |
543 | </section> |
528 | - | ||
529 | <para>Thanks to the abstract paging interface, there is possibility |
- | |
530 | left have more paging implementations, for example B-Tree page |
- | |
531 | tables.</para> |
- | |
532 | </section> |
544 | </section> |
533 | 545 | ||
534 | <section id="tlb"> |
546 | <section id="tlb"> |
535 | <title>Translation Lookaside buffer</title> |
547 | <title>Translation Lookaside buffer</title> |
536 | 548 | ||
537 | <para>Due to the extensive overhead during the page mapping lookup in |
549 | <para>Due to the extensive overhead during the page mapping lookup in |
538 | the page tables, all architectures has fast assotiative cache memory |
550 | the page tables, all architectures has fast assotiative cache memory |
539 | built-in CPU. This memory called TLB stores recently used page table |
551 | built-in CPU. This memory called TLB stores recently used page table |
540 | entries.</para> |
552 | entries.</para> |
541 | 553 | ||
542 | <section id="tlb_shootdown"> |
554 | <section id="tlb_shootdown"> |
543 | <title>TLB consistency. TLB shootdown algorithm.</title> |
555 | <title>TLB consistency. TLB shootdown algorithm.</title> |
544 | 556 | ||
545 | <para>Operating system is responsible for keeping TLB consistent by |
557 | <para>Operating system is responsible for keeping TLB consistent by |
546 | invalidating the contents of TLB, whenever there is some change in |
558 | invalidating the contents of TLB, whenever there is some change in |
547 | page tables. Those changes may occur when page or group of pages |
559 | page tables. Those changes may occur when page or group of pages |
548 | were unmapped, mapping is changed or system switching active address |
560 | were unmapped, mapping is changed or system switching active address |
549 | space to schedule a new system task (which is a batch unmap of all |
561 | space to schedule a new system task. Moreover, this invalidation |
550 | address space mappings). Moreover, this invalidation operation must |
562 | operation must be done an all system CPUs because each CPU has its |
551 | be done an all system CPUs because each CPU has its own independent |
563 | own independent TLB cache. Thus maintaining TLB consistency on SMP |
552 | TLB cache. Thus maintaining TLB consistency on SMP configuration as |
564 | configuration as not as trivial task as it looks on the first |
553 | not as trivial task as it looks at the first glance. Naive solution |
565 | glance. Naive solution would assume that is the CPU which wants to |
554 | would assume remote TLB invalidatation, which is not possible on the |
566 | invalidate TLB will invalidate TLB caches on other CPUs. It is not |
555 | most of the architectures, because of the simple fact - flushing TLB |
567 | possible on the most of the architectures, because of the simple |
556 | is allowed only on the local CPU and there is no possibility to |
568 | fact - flushing TLB is allowed only on the local CPU and there is no |
557 | access other CPUs' TLB caches.</para> |
569 | possibility to access other CPUs' TLB caches, thus invalidate TLB |
- | 570 | remotely.</para> |
|
558 | 571 | ||
559 | <para>Technique of remote invalidation of TLB entries is called "TLB |
572 | <para>Technique of remote invalidation of TLB entries is called "TLB |
560 | shootdown". HelenOS uses a variation of the algorithm described by |
573 | shootdown". HelenOS uses a variation of the algorithm described by |
561 | D. Black et al., "Translation Lookaside Buffer Consistency: A |
574 | D. Black et al., "Translation Lookaside Buffer Consistency: A |
562 | Software Approach," Proc. Third Int'l Conf. Architectural Support |
575 | Software Approach," Proc. Third Int'l Conf. Architectural Support |
563 | for Programming Languages and Operating Systems, 1989, pp. |
576 | for Programming Languages and Operating Systems, 1989, pp. |
564 | 113-122.</para> |
577 | 113-122.</para> |
565 | 578 | ||
566 | <para>As the situation demands, you will want partitial invalidation |
579 | <para>As the situation demands, you will want partitial invalidation |
567 | of TLB caches. In case of simple memory mapping change it is |
580 | of TLB caches. In case of simple memory mapping change it is |
568 | necessary to invalidate only one or more adjacent pages. In case if |
581 | necessary to invalidate only one or more adjacent pages. In case if |
569 | the architecture is aware of ASIDs, during the address space |
582 | the architecture is aware of ASIDs, when kernel needs to dump some |
570 | switching, kernel invalidates only entries from this particular |
583 | ASID to use by another task, it invalidates only entries from this |
571 | address space. Final option of the TLB invalidation is the complete |
584 | particular address space. Final option of the TLB invalidation is |
572 | TLB cache invalidation, which is the operation that flushes all |
585 | the complete TLB cache invalidation, which is the operation that |
573 | entries in TLB.</para> |
586 | flushes all entries in TLB.</para> |
574 | - | ||
575 | <para>TLB shootdown is performed in two phases. First, the initiator |
- | |
576 | process sends an IPI message indicating the TLB shootdown request to |
- | |
577 | the rest of the CPUs. Then, it waits until all CPUs confirm TLB |
- | |
578 | invalidating action execution.</para> |
- | |
579 | </section> |
- | |
580 | </section> |
- | |
581 | </section> |
- | |
582 | 587 | ||
583 | <section> |
- | |
584 | <title>---</title> |
588 | <para>TLB shootdown is performed in two phases.</para> |
585 | 589 | ||
- | 590 | <formalpara> |
|
586 | <para>At the moment HelenOS does not support swapping.</para> |
591 | <title>Phase 1.</title> |
587 | 592 | ||
588 | <para>- pouzivame vypadky stranky k alokaci ramcu on-demand v ramci |
593 | <para>First, initiator locks a global TLB spinlock, then request |
- | 594 | is being put to the local request cache of every other CPU in the |
|
- | 595 | system protected by its spinlock. In case the cache is full, all |
|
589 | as_area - na architekturach, ktere to podporuji, podporujeme non-exec |
596 | requests in the cache are replaced by one request, indicating |
- | 597 | global TLB flush. Then the initiator thread sends an IPI message |
|
- | 598 | indicating the TLB shootdown request to the rest of the CPUs and |
|
- | 599 | waits actively until all CPUs confirm TLB invalidating action |
|
- | 600 | execution by setting up a special flag. After setting this flag |
|
- | 601 | this thread is blocked on the TLB spinlock, held by the |
|
- | 602 | initiator.</para> |
|
- | 603 | </formalpara> |
|
- | 604 | ||
- | 605 | <formalpara> |
|
- | 606 | <title>Phase 2.</title> |
|
- | 607 | ||
- | 608 | <para>All CPUs are waiting on the TLB spinlock to execute TLB |
|
- | 609 | invalidation action and have indicated their intention to the |
|
- | 610 | initiator. Initiator continues, cleaning up its TLB and releasing |
|
- | 611 | the global TLB spinlock. After this all other CPUs gain and |
|
- | 612 | immidiately release TLB spinlock and perform TLB invalidation |
|
590 | stranky</para> |
613 | actions.</para> |
- | 614 | </formalpara> |
|
- | 615 | </section> |
|
- | 616 | </section> |
|
591 | </section> |
617 | </section> |
592 | </section> |
618 | </section> |
593 | </chapter> |
619 | </chapter> |