source: trunk/kernel/mm/vmm.h @ 571

Last change on this file since 571 was 567, checked in by alain, 4 years ago

Complete restructuration of kernel locks.

File size: 26.1 KB
Line 
1/*
2 * vmm.h - virtual memory management related operations
3 *
4 * Authors   Ghassan Almaless (2008,2009,2010,2011, 2012)
5 *           Mohamed Lamine Karaoui (2015)
6 *           Alain Greiner (2016,2017,2018)
7 *
8 * Copyright (c) UPMC Sorbonne Universites
9 *
10 * This file is part of ALMOS-MKH.
11 *
12 * ALMOS-MKH is free software; you can redistribute it and/or modify it
13 * under the terms of the GNU General Public License as published by
14 * the Free Software Foundation; version 2.0 of the License.
15 *
16 * ALMOS-MKH is distributed in the hope that it will be useful, but
17 * WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
19 * General Public License for more details.
20 *
21 * You should have received a copy of the GNU General Public License
22 * along with ALMOS-MKH; if not, write to the Free Software Foundation,
23 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
24 */
25
26#ifndef _VMM_H_
27#define _VMM_H_
28
29#include <hal_kernel_types.h>
30#include <bits.h>
31#include <list.h>
32#include <queuelock.h>
33#include <hal_gpt.h>
34#include <vseg.h>
35#include <page.h>
36
37/****  Forward declarations  ****/
38
39struct process_s;
40
41/*********************************************************************************************
42 * This structure defines the STACK allocator used by the VMM to dynamically handle
43 * a STACK vseg requested or released by an user process.
44 * This allocator handles a fixed size array of fixed size slots in the STACK zone.
45 * The stack size and the number of slots are defined by the CONFIG_VMM_STACK_SIZE, and
46 * CONFIG_VMM_STACK_BASE parameters.
47 * Each slot can contain one user stack vseg. The first page in the slot is not allocated
48 * to detect stack overflow.
49 * The slot index can be computed form the slot base address, and reversely.
50 * All allocation / release operations are registered in the stack_bitmap, that completely
51 * define the STACK zone state.
52 ********************************************************************************************/
53
54typedef struct stack_mgr_s
55{
56    busylock_t     lock;               /*! lock protecting STACK allocator                  */
57    vpn_t          vpn_base;           /*! first page of STACK zone                         */
58    bitmap_t       bitmap;             /*! bit bector of allocated stacks                   */
59}
60stack_mgr_t;
61
62/*********************************************************************************************
63 * This structure defines the MMAP allocator used by the VMM to dynamically handle 
64 * MMAP vsegs requested or released by an user process.
65 * This allocator should be only used in the reference cluster.
66 * - allocation policy : all allocated vsegs occupy an integer number of pages that is
67 *   power of 2, and are aligned on a page boundary. The requested number of pages is
68 *   rounded if required. The first_free_vpn variable defines completely the MMAP zone state.
69 *   It is never decremented, as the released vsegs are simply registered in a zombi_list.
70 *   The relevant zombi_list is checked first for each allocation request.
71 * - release policy : a released MMAP vseg is registered in an array of zombi_lists.
72 *   This array is indexed by ln(number of pages), and each entry contains the root of
73 *   a local list of zombi vsegs that have the same size. The physical memory allocated
74 *   for a zombi vseg descriptor is not released, to use the "list" field.
75 *   This physical memory allocated for MMAP vseg descriptors is actually released
76 *   when the VMM is destroyed.
77 ********************************************************************************************/
78
79typedef struct mmap_mgr_s
80{
81    busylock_t     lock;               /*! lock protecting MMAP allocator                   */
82    vpn_t          vpn_base;           /*! first page of MMAP zone                          */
83    vpn_t          vpn_size;           /*! number of pages in MMAP zone                     */
84    vpn_t          first_free_vpn;     /*! first free page in MMAP zone                     */
85    list_entry_t   zombi_list[32];     /*! array of roots of released vsegs lists           */
86}
87mmap_mgr_t;
88
89/*********************************************************************************************
90 * This structure defines the Virtual Memory Manager for a given process in a given cluster.
91 * This local VMM provides four main services:
92 * 1) It contains the local copy of vseg list (VSL), only complete in referrence.
93 * 2) It contains the local copy of the generic page table (GPT), only complete in reference.
94 * 3) The stack manager dynamically allocates virtual memory space for the STACK vsegs.
95 * 4) The mmap manager dynamically allocates virtual memory for the (FILE/ANON/REMOTE) vsegs.
96 ******************************************************a**************************************
97 * Implementation notes:
98 * 1. The VSL contains only local vsegs, but it is implemented as an xlist, and protected by
99 *    a remote_rwlock, because it can be accessed by a thread running in a remote cluster.
100 *    An exemple is the vmm_fork_copy() function.
101 * 2. In most clusters, the VSL and GPT are only partial copies of the reference VSL and GPT
102 *    structures, stored in the reference cluster.
103 ********************************************************************************************/
104
105typedef struct vmm_s
106{
107        remote_rwlock_t  vsegs_lock;         /*! lock protecting the local VSL                  */
108        xlist_entry_t    vsegs_root;         /*! VSL root (VSL only complete in reference)      */
109        uint32_t         vsegs_nr;           /*! total number of local vsegs                    */
110
111    gpt_t            gpt;                /*! Generic Page Table (complete in reference)     */
112
113    stack_mgr_t      stack_mgr;          /*! embedded STACK vsegs allocator                 */
114    mmap_mgr_t       mmap_mgr;           /*! embedded MMAP vsegs allocator                  */
115
116        uint32_t         pgfault_nr;         /*! page fault counter (instrumentation)           */
117
118    vpn_t            kent_vpn_base;      /*! kentry vseg first page                         */
119    vpn_t            args_vpn_base;      /*! args vseg first page                           */
120    vpn_t            envs_vpn_base;      /*! envs zone first page                           */
121    vpn_t            heap_vpn_base;      /*! envs zone first page                           */
122        vpn_t            code_vpn_base;      /*! code zone first page                           */
123        vpn_t            data_vpn_base;      /*! data zone first page                           */
124
125        intptr_t         entry_point;        /*! main thread entry point                        */
126}
127vmm_t;
128
129/*********************************************************************************************
130 * This function initialises the virtual memory manager attached to an user process.
131 * - It initializes the STACK and MMAP allocators.
132 * - It registers the "kentry", "args", "envs" vsegs in the VSL.
133 * - It initializes the generic page table, calling the HAL specific hal_gpt_init() function.
134 * - For TSAR it map all pages for the "kentry" vseg, that must be identity mapping.
135 * Note:
136 * - The "code" and "data" vsegs are registered by the elf_load_process() function.
137 * - The "stack" vsegs are dynamically created by the thread_user_create() function.
138 * - The "file", "anon", "remote" vsegs are dynamically created by the mmap() syscall.
139 *********************************************************************************************
140 * @ process   : pointer on process descriptor
141 * @ return 0 if success / return -1 if failure.
142 ********************************************************************************************/
143error_t vmm_init( struct process_s * process );
144
145/*********************************************************************************************
146 * This function displays on TXY0 the list or registered vsegs for a given <process>.
147 * It must be executed by a thread running in reference cluster.
148 * If the <mapping> argument is true, it displays for each vseg all mapped PTEs in GPT.
149 *********************************************************************************************
150 * @ process   : pointer on process descriptor.
151 * @ mapping   : detailed mapping if true.
152 ********************************************************************************************/
153void vmm_display( struct process_s * process,
154                  bool_t             mapping );
155
156/*******************************************************************************************
157 * This function adds a vseg descriptor in the VSL of a given VMM,
158 * and updates the vmm field in the vseg descriptor.
159 * It takes the lock protecting VSL.
160 *******************************************************************************************
161 * @ vmm       : pointer on the VMM
162 * @ vseg      : pointer on the vseg descriptor
163 ******************************************************************************************/
164void vmm_vseg_attach( struct vmm_s  * vmm,
165                      vseg_t        * vseg );
166
167/*******************************************************************************************
168 * This function removes a vseg descriptor from the set of vsegs controlled by a given VMM,
169 * and updates the vmm field in the vseg descriptor. No memory is released.
170 * It takes the lock protecting VSL.
171 *******************************************************************************************
172 * @ vmm       : pointer on the VMM
173 * @ vseg      : pointer on the vseg descriptor
174 ******************************************************************************************/
175void vmm_vseg_detach( struct vmm_s  * vmm,
176                      vseg_t        * vseg );
177
178/*********************************************************************************************
179 * This function is called by the process_make_fork() function. It partially copies
180 * the content of a remote parent process VMM to the local child process VMM:
181 * - all DATA, MMAP, REMOTE vsegs registered in the parent VSL are registered in the child
182 *   VSL, and all valid GPT entries in parent GPT are copied to the child GPT.
183 *   The WRITABLE flag is reset and the COW flag is set in child GPT.
184 * - all CODE vsegs registered in the parent VSL are registered in the child VSL, but the
185 *   GPT entries are not copied in the chilf GPT, that will be dynamically updated from
186 *   the .elf file when a page fault is reported.
187 * - all FILE vsegs registered in the parent VSL are registered in the child VSL, and all
188 *   valid GPT entries in parent GPT are copied to the child GPT. The COW flag is not set.
189 * - no STACK vseg is copied from  parent VMM to child VMM, because the child STACK vseg
190 *   must be copied later from the cluster containing the user thread requesting the fork().
191 *********************************************************************************************
192 * @ child_process     : local pointer on local child process descriptor.
193 * @ parent_process_xp : extended pointer on remote parent process descriptor.
194 * @ return 0 if success / return -1 if failure.
195 ********************************************************************************************/
196error_t vmm_fork_copy( struct process_s * child_process,
197                       xptr_t             parent_process_xp );
198
199/*********************************************************************************************
200 * This function is called by the process_make_fork() function executing the fork syscall.
201 * It set the COW flag, and reset the WRITABLE flag of all GPT entries of the DATA, MMAP,
202 * and REMOTE vsegs of a process identified by the <process> argument.
203 * It must be called by a thread running in the reference cluster, that contains the complete
204 * VSL and GPT (use the rpc_vmm_set_cow_client() when the calling thread client is remote).
205 * It updates all copies of the process in all clusters, to maintain coherence in GPT copies,
206 * using the list of copies stored in the owner process, and using remote_write accesses to
207 * update the remote GPTs. It atomically increment the pending_fork counter, in all involved
208 * physical page descriptors. It cannot fail, as only mapped entries in GPTs are updated.
209 *********************************************************************************************
210 * @ process   : local pointer on local reference process descriptor.
211 ********************************************************************************************/
212void vmm_set_cow( struct process_s * process );
213
214/*********************************************************************************************
215 * This global function modifies a GPT entry identified  by the <process> and <vpn>
216 * arguments in all clusters containing a process copy.
217 * It must be called by a thread running in the reference cluster.
218 * It updates all copies of the process in all clusters, to maintain coherence in GPT copies,
219 * using the list of copies stored in the owner process, and using remote_write accesses to
220 * update the remote GPTs. It cannot fail, as only mapped entries in GPT copies are updated.
221 *********************************************************************************************
222 * @ process   : local pointer on local process descriptor.
223 * @ vpn       : PTE index.
224 * @ attr      : PTE / attributes.
225 * @ ppn       : PTE / physical page index.
226 ********************************************************************************************/
227void vmm_global_update_pte( struct process_s * process,
228                            vpn_t              vpn,
229                            uint32_t           attr,
230                            ppn_t              ppn );
231
232/*********************************************************************************************
233 * This function unmaps from the local GPT all mapped PTEs of a vseg identified by the
234 * <process> and <vseg> arguments. It can be used for any type of vseg.
235 * If this function is executed in the reference cluster, it handles for each referenced
236 * physical pages the pending forks counter :
237 * - if counter is non-zero, it decrements it.
238 * - if counter is zero, it releases the physical page to local kmem allocator.
239 *********************************************************************************************
240 * @ process  : pointer on process descriptor.
241 * @ vseg     : pointer on the vseg to be unmapped.
242 ********************************************************************************************/
243void vmm_unmap_vseg( struct process_s * process,
244                     vseg_t           * vseg );
245
246/*********************************************************************************************
247 * This function deletes, in the local cluster, all vsegs registered in the VSL
248 * of the process identified by the <process> argument. For each vseg:
249 * - it unmaps all vseg PTEs from the GPT (release the physical pages when required).
250 * - it removes the vseg from the local VSL.
251 * - it releases the memory allocated to the local vseg descriptors.
252 * Finally, it releases the memory allocated to the GPT itself.
253 *********************************************************************************************
254 * @ process   : pointer on process descriptor.
255 ********************************************************************************************/
256void vmm_destroy( struct process_s * process );
257
258/*********************************************************************************************
259 * This function scans the list of vsegs registered in the VMM of a given process descriptor
260 * to check if a given virtual region (defined by a base and size) overlap an existing vseg.
261 *********************************************************************************************
262 * @ process  : pointer on process descriptor.
263 * @ base     : region virtual base address.
264 * @ size     : region size (bytes).
265 * @ returns NULL if no conflict / return conflicting vseg pointer if conflict.
266 ********************************************************************************************/
267vseg_t * vmm_check_conflict( struct process_s * process,
268                             vpn_t              base,
269                             vpn_t              size );
270
271/*********************************************************************************************
272 * This function allocates memory for a vseg descriptor, initialises it, and register it
273 * in the VMM of the local process descriptor, that should be the reference process.
274 * For the 'stack", "file", "anon", & "remote" types, it does not use the <base> argument,
275 * but uses the STACK and MMAP virtual memory allocators.
276 * It checks collision with all pre-existing vsegs.
277 * To comply with the "on-demand" paging policy, this function does NOT modify the page table,
278 * and does not allocate physical memory for vseg data.
279 * It should be called by a local thread (could be a RPC thread if the client thread is not
280 * running in the regerence cluster).
281 *********************************************************************************************
282 * @ process     : pointer on local processor descriptor.
283 * @ type        : vseg type.
284 * @ base        : vseg base address (not used for dynamically allocated vsegs).
285 * @ size        : vseg size (bytes).
286 * @ file_offset : offset in file for CODE, DATA, FILE types.
287 * @ file_size   : can be smaller than "size" for DATA type.
288 * @ mapper_xp   : extended pointer on mapper for CODE, DATA, FILE types.
289 * @ cxy         : physical mapping cluster (for non distributed vsegs).
290 * @ returns pointer on vseg if success / returns NULL if no memory, or conflict.
291 ********************************************************************************************/
292vseg_t * vmm_create_vseg( struct process_s * process,
293                          vseg_type_t        type,
294                          intptr_t           base,
295                              uint32_t           size,
296                          uint32_t           file_offset,
297                          uint32_t           file_size,
298                          xptr_t             mapper_xp,
299                          cxy_t              cxy );
300
301/*********************************************************************************************
302 * This function removes a vseg identified by it's pointer from the VMM of the calling process.
303 * - If the vseg has not the STACK or MMAP type, it is removed from the vsegs list,
304 *   and the physical memory allocated to vseg descriptor is released to KMEM.
305 * - If the vseg has the STACK type, it is removed from the vsegs list, the physical memory
306 *   allocated to vseg descriptor is released to KMEM, and the stack slot is returned to the
307 *   VMM STACK allocator.
308 * - If the vseg has the MMAP type, it is removed from the vsegs list and is registered
309 *   in the zombi_list of the VMM MMAP allocator for future reuse. The physical memory
310 *   allocated to vseg descriptor is NOT released to KMEM.
311 *********************************************************************************************
312 * @ vseg      : pointer on vseg to be removed.
313 ********************************************************************************************/
314void vmm_remove_vseg( vseg_t * vseg );
315
316/*********************************************************************************************
317 * This function removes a given region (defined by a base address and a size) from
318 * the VMM of a given process descriptor. This can modify the number of vsegs:
319 * (a) if the region is not entirely mapped in an existing vseg, it's an error.
320 * (b) if the region has same base and size as an existing vseg, the vseg is removed.
321 * (c) if the removed region cut the vseg in two parts, it is modified.
322 * (d) if the removed region cut the vseg in three parts, it is modified, and a new
323 *     vseg is created with same type.
324 * FIXME [AG] this function must be called by a thread running in the reference cluster,
325 * and the VMM must be updated in all process descriptors copies.
326 *********************************************************************************************
327 * @ process   : pointer on process descriptor
328 * @ base      : vseg base address
329 * @ size      : vseg size (bytes)
330 ********************************************************************************************/
331error_t vmm_resize_vseg( struct process_s * process,
332                         intptr_t           base,
333                         intptr_t           size );
334
335/*********************************************************************************************
336 * This function checks that a given virtual address is contained in a registered vseg.
337 * It can be called by any thread running in any cluster:
338 * - if the vseg is registered in the local process VMM, it returns the local vseg pointer.
339 * - if the vseg is missing in local VMM, it uses a RPC to get it from the reference cluster,
340 *   register it in local VMM and returns the local vseg pointer, if success.
341 * - it returns an user error if the vseg is missing in the reference VMM, or if there is
342 *   not enough memory for a new vseg descriptor in cluster containing the calling thread.
343 *********************************************************************************************
344 * @ process   : [in] pointer on process descriptor
345 * @ vaddr     : [in] virtual address
346 * @ vseg      : [out] local pointer on local vseg
347 * @ returns 0 if success / returns -1 if user error (out of segment).
348 *********************************************************************************************/
349error_t vmm_get_vseg( struct process_s  * process,
350                      intptr_t            vaddr,
351                      vseg_t           ** vseg );           
352
353/*********************************************************************************************
354 * This function is called by the generic exception handler in case of page-fault,
355 * or copy-on-write event locally detected for a given <vpn> in a given <process>
356 * as defined by the <is_cow> argument.
357 * 1) For a Page-Fault:
358 * - If the local cluster is the reference, or for the STACK and CODE segment types,
359 *   it call directly the vmm_get_pte() function to access the local VMM.
360 * - Otherwise, it send a RPC_VMM_GET_PTE to the reference cluster to get the missing
361 *   PTE attributes and PPN.
362 * This function check that the missing VPN belongs to a registered vseg, allocates
363 * a new physical page if required, and updates the local page table.
364 * 2) For a Copy-On-Write:
365 * - If no pending fork, it reset the COW flag and set the WRITE flag in the reference
366 *   GPT entry, and in all the GPT copies.
367 * - If there is a pending fork, it allocates a new physical page from the cluster defined
368 *   by the vseg type, copies the old physical page content to the new physical page,
369 *   and decrements the pending_fork counter in old physical page descriptor.
370 *********************************************************************************************
371 * @ process   : pointer on local process descriptor copy.
372 * @ vpn       : VPN of the missing or faulting PTE.
373 * @ is_cow    : Copy-On-Write event if true / Page-fault if false.
374 * @ returns 0 if success / returns ENOMEM if no memory or illegal VPN.
375 ********************************************************************************************/
376error_t vmm_handle_page_fault( struct process_s * process,
377                               vpn_t              vpn,
378                               bool_t             is_cow );
379
380/*********************************************************************************************
381 * This function is called by the vmm_handle_page_fault() to handle both the "page-fault",
382 * and the "copy-on_write" events for a given <vpn> in a given <process>, as defined
383 * by the <is_cow> argument.
384 * The vseg containing the searched VPN must be registered in the reference VMM.
385 * - for an page-fault, it allocates the missing physical page from the target cluster
386 *   defined by the vseg type, initializes it, and updates the reference GPT, but not
387 *   the copies GPT, that will be updated on demand.
388 * - for a copy-on-write, it allocates a new physical page from the target cluster,
389 *   initialise it from the old physical page, and updates the reference GPT and all
390 *   the GPT copies, for coherence.
391 * It calls the RPC_PMEM_GET_PAGES to get the new physical page when the target cluster
392 * is not the local cluster,
393 * It returns in the <attr> and <ppn> arguments the accessed or modified PTE.
394 *********************************************************************************************
395 * @ process   : [in] pointer on process descriptor.
396 * @ vpn       : [in] VPN defining the missing PTE.
397 * @ is_cow    : [in] "copy_on_write" if true / "page_fault" if false.
398 * @ attr      : [out] PTE attributes.
399 * @ ppn       : [out] PTE ppn.
400 * @ returns 0 if success / returns ENOMEM if error.
401 ********************************************************************************************/
402error_t vmm_get_pte( struct process_s * process,
403                     vpn_t              vpn,
404                     bool_t             is_cow,
405                     uint32_t         * attr,
406                     ppn_t            * ppn );
407
408/*********************************************************************************************
409 * This function is called by the vmm_get_pte() function when a page is unmapped.
410 * Depending on the vseg type, defined by the <vseg> argument, it returns the PPN
411 * (Physical Page Number) associated to a missing page defined by the <vpn> argument.
412 * - For the FILE type, it returns directly the physical page from the file mapper.
413 * - For the CODE and DATA types, it allocates a new physical page from the cluster defined
414 *   by the <vseg->cxy> field, or by the <vpn> MSB bits for a distributed vseg,
415 *   and initialize this page from the .elf file mapper.
416 * - For all other types, it allocates a new physical page from the cluster defined
417 *   by the <vseg->cxy> field, or by the <vpn> MSB bits for a distributed vseg,
418 *   but the new page is not initialized.
419 *********************************************************************************************
420 * @ vseg   : local pointer on vseg containing the mising page.
421 * @ vpn    : Virtual Page Number identifying the missing page.
422 * @ ppn    : [out] returned Physical Page Number.
423 * return 0 if success / return EINVAL or ENOMEM if error.
424 ********************************************************************************************/
425error_t vmm_get_one_ppn( vseg_t * vseg,
426                         vpn_t    vpn,
427                         ppn_t  * ppn );
428
429
430#endif /* _VMM_H_ */
Note: See TracBrowser for help on using the repository browser.