1 /*        $NetBSD: rf_raid.h,v 1.53 2023/09/25 21:59:38 oster Exp $   */
2 /*
3  * Copyright (c) 1995 Carnegie-Mellon University.
4  * All rights reserved.
5  *
6  * Author: Mark Holland
7  *
8  * Permission to use, copy, modify and distribute this software and
9  * its documentation is hereby granted, provided that both the copyright
10  * notice and this permission notice appear in all copies of the
11  * software, derivative works or modified versions, and any portions
12  * thereof, and that both notices appear in supporting documentation.
13  *
14  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
15  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
16  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
17  *
18  * Carnegie Mellon requests users of this software to return to
19  *
20  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
21  *  School of Computer Science
22  *  Carnegie Mellon University
23  *  Pittsburgh PA 15213-3890
24  *
25  * any improvements or extensions that they make and grant Carnegie the
26  * rights to redistribute these changes.
27  */
28 
29 /**********************************************
30  * rf_raid.h -- main header file for RAID driver
31  **********************************************/
32 
33 
34 #ifndef _RF__RF_RAID_H_
35 #define _RF__RF_RAID_H_
36 
37 #include <dev/raidframe/raidframevar.h>
38 #include "rf_archs.h"
39 #include "rf_threadstuff.h"
40 
41 #include "rf_netbsd.h"
42 
43 #include <sys/disklabel.h>
44 #include <sys/types.h>
45 #include <sys/queue.h>
46 
47 #include "rf_alloclist.h"
48 #include "rf_stripelocks.h"
49 #include "rf_layout.h"
50 #include "rf_disks.h"
51 #include "rf_debugMem.h"
52 #include "rf_diskqueue.h"
53 #include "rf_reconstruct.h"
54 #include "rf_acctrace.h"
55 #include "rf_fifo.h"
56 
57 #if RF_INCLUDE_PARITYLOGGING > 0
58 #include "rf_paritylog.h"
59 #endif                                  /* RF_INCLUDE_PARITYLOGGING > 0 */
60 
61 #define RF_COMPONENT_LABEL_VERSION_1 1
62 #define RF_COMPONENT_LABEL_VERSION 2
63 #define RF_RAID_DIRTY 0
64 #define RF_RAID_CLEAN 1
65 
66 
67 /*
68  * Each row in the array is a distinct parity group, so
69  * each has its own status, which is one of the following.
70  */
71 typedef enum RF_RowStatus_e {
72           rf_rs_optimal,
73           rf_rs_degraded,
74           rf_rs_reconstructing,
75           rf_rs_reconfigured
76 }       RF_RowStatus_t;
77 
78 struct RF_CumulativeStats_s {
79           struct timeval start;         /* the time when the stats were last started */
80           struct timeval stop;          /* the time when the stats were last stopped */
81           long    sum_io_us;  /* sum of all user response times (us) */
82           long    num_ios;    /* total number of I/Os serviced */
83           long    num_sect_moved;       /* total number of sectors read or written */
84 };
85 
86 struct RF_ThroughputStats_s {
87           rf_declare_mutex2(mutex);/* a mutex used to lock the configuration
88                                          * stuff */
89           struct timeval start;         /* timer started when numOutstandingRequests
90                                          * moves from 0 to 1 */
91           struct timeval stop;          /* timer stopped when numOutstandingRequests
92                                          * moves from 1 to 0 */
93           RF_uint64 sum_io_us;          /* total time timer is enabled */
94           RF_uint64 num_ios;  /* total number of ios processed by RAIDframe */
95           long    num_out_ios;          /* number of outstanding ios */
96 };
97 
98 struct RF_Raid_s {
99           /* This portion never changes, and can be accessed without locking */
100           /* an exception is Disks[][].status, which requires locking when it is
101            * changed.  XXX this is no longer true.  numSpare and friends can
102            * change now.
103          */
104           u_int   numCol;               /* number of columns of disks, typically == #
105                                          * of disks/rank */
106           u_int   numSpare;   /* number of spare disks */
107           int     maxQueueDepth;        /* max disk queue depth */
108           RF_SectorCount_t totalSectors;          /* total number of sectors in the
109                                                    * array */
110           RF_SectorCount_t sectorsPerDisk;        /* number of sectors on each
111                                                              * disk */
112           u_int   logBytesPerSector;    /* base-2 log of the number of bytes
113                                                    * in a sector */
114           u_int   bytesPerSector;       /* bytes in a sector */
115           RF_int32 sectorMask;          /* mask of bytes-per-sector */
116 
117           RF_RaidLayout_t Layout;       /* all information related to layout */
118           RF_RaidDisk_t *Disks;         /* all information related to physical disks */
119           RF_DiskQueue_t *Queues; /* all information related to disk queues */
120           u_int   maxQueue;   /* initialized queues in Queues array */
121           const RF_DiskQueueSW_t *qType;/* pointer to the DiskQueueSW used for the
122                                                    component queues. */
123           /* NOTE:  This is an anchor point via which the queues can be
124            * accessed, but the enqueue/dequeue routines in diskqueue.c use a
125            * local copy of this pointer for the actual accesses. */
126           /* The remainder of the structure can change, and therefore requires
127            * locking on reads and updates */
128           rf_declare_mutex2(mutex);/* mutex used to serialize access to
129                                          * the fields below */
130           RF_RowStatus_t status;        /* the status of each row in the array */
131           int     valid;                /* indicates successful configuration */
132           RF_LockTableEntry_t *lockTable;         /* stripe-lock table */
133           RF_LockTableEntry_t *quiesceLock;       /* quiesnce table */
134           int     numFailures;          /* total number of failures in the array */
135           int     numNewFailures; /* number of *new* failures (that havn't
136                                            caused a mod_counter update */
137 
138           int     parity_good;    /* !0 if parity is known to be correct */
139           int     serial_number;  /* a "serial number" for this set */
140           int     mod_counter;    /* modification counter for component labels */
141           int     clean;          /* completely unused and should be removed */
142 
143           int     openings;       /* Number of IO's which can be scheduled
144                                            simultaneously (high-level - not a
145                                            per-component limit)*/
146 
147           int maxOutstanding;   /* maxOutstanding requests (per-component) */
148           int autoconfigure;    /* automatically configure this RAID set.
149                                          0 == no, 1 == yes */
150           int root_partition;   /* Use this set as /
151                                          0 == no, 1 == yes*/
152           int last_unit;        /* last unit number (e.g. 0 for /dev/raid0)
153                                          of this component.  Used for autoconfigure
154                                          only. */
155           int config_order;     /* 0 .. n.  The order in which the component
156                                          should be auto-configured.  E.g. 0 is will
157                                          done first, (and would become raid0).
158                                          This may be in conflict with last_unit!!?! */
159                                 /* Not currently used. */
160 
161           /* queue to gather up requests from KernelWakeupFunc() and let
162              a kernel thread deal with calling rf_DiskIOComplete and any
163              callback functions. */
164           TAILQ_HEAD(iodone_q,RF_DiskQueueData_s) iodone;
165           /* and a lock/cv to protect it */
166           rf_declare_mutex2(iodone_lock);
167           rf_declare_cond2(iodone_cv);
168 
169 
170           RF_VoidPointerListElem_t *iobuf;       /* I/O buffer free list */
171           int iobuf_count;             /* count of I/O buffers on the freelist */
172           int numEmergencyBuffers;     /* number of these buffers to pre-allocate */
173 
174           RF_VoidPointerListElem_t *stripebuf;   /* Full-stripe buffer free list */
175           int stripebuf_count;           /* count of full-stripe buffers on the freelist */
176           int numEmergencyStripeBuffers; /* number of these buffers to pre-allocate */
177 
178           /*
179          * Cleanup stuff
180          */
181           RF_ShutdownList_t *shutdownList;        /* shutdown activities */
182           RF_AllocListElem_t *cleanupList;        /* memory to be freed at
183                                                              * shutdown time */
184 
185           /*
186          * Recon stuff
187          */
188           RF_HeadSepLimit_t headSepLimit;
189           int     numFloatingReconBufs;
190           int     reconInProgress;
191           int     forceRecon;
192           rf_declare_cond2(waitForReconCond);     /* goes with raidPtr->mutex */
193           RF_RaidReconDesc_t *reconDesc;          /* reconstruction descriptor */
194           RF_ReconCtrl_t *reconControl; /* reconstruction control structure
195                                                    * pointers for each row in the array */
196 
197           /*
198          * Array-quiescence stuff
199          */
200           rf_declare_mutex2(access_suspend_mutex);
201           rf_declare_cond2(access_suspend_cv);
202           RF_IoCount_t accesses_suspended;
203           RF_IoCount_t accs_in_flight;
204           int     access_suspend_release;
205           int     waiting_for_quiescence;
206           RF_CallbackFuncDesc_t *quiesce_wait_list;
207 
208           /*
209          * Statistics
210          */
211           RF_StripeCount_t     parity_rewrite_stripes_done;
212 
213           int     recon_in_progress;
214           int     parity_rewrite_in_progress;
215           int     changing_components;
216 
217           rf_declare_cond2(parity_rewrite_cv);
218           rf_declare_cond2(changing_components_cv);
219 
220           /*
221          * Engine thread control
222          */
223           rf_declare_mutex2(node_queue_mutex);
224           rf_declare_cond2(node_queue_cv);
225           RF_DagNode_t *node_queue;
226           RF_Thread_t parity_rewrite_thread;
227           RF_Thread_t engine_thread;
228           RF_Thread_t engine_helper_thread;
229           RF_Thread_t recon_thread;
230           int     shutdown_engine;
231           int     shutdown_raidio;
232           int     dags_in_flight;       /* debug */
233 
234           /*
235          * PSS (Parity Stripe Status) stuff
236          */
237           long    pssTableSize;
238 
239           /*
240          * Reconstruction stuff
241          */
242           int     procsInBufWait;
243           int     numFullReconBuffers;
244 #if RF_ACC_TRACE > 0
245           RF_AccTraceEntry_t *recon_tracerecs;
246 #endif
247           unsigned long accumXorTimeUs;
248 
249           /*
250          * nAccOutstanding, waitShutdown protected by desc freelist lock
251          * (This may seem strange, since that's a central serialization point
252          * for a per-array piece of data, but otherwise, it'd be an extra
253          * per-array lock, and that'd only be less efficient...)
254          */
255           rf_declare_mutex2(rad_lock);
256           rf_declare_cond2(outstandingCond);
257           int     waitShutdown;
258           int     nAccOutstanding;
259 
260           int     *abortRecon;          /* Abort background operations requested */
261 
262           RF_DiskId_t **diskids;
263 
264           int     raidid;
265           void   *softc;
266           RF_AccTotals_t acc_totals;
267           int     keep_acc_totals;
268 
269           struct raidcinfo *raid_cinfo; /* array of component info */
270 
271           int     terminate_disk_queues;
272 
273           /*
274          * XXX
275          *
276          * config-specific information should be moved
277          * somewhere else, or at least hung off this
278          * in some generic way
279          */
280 #if RF_INCLUDE_CHAINDECLUSTER > 0
281 
282           /* used by rf_compute_workload_shift */
283           RF_RowCol_t hist_diskreq[RF_MAXCOL];
284 #endif
285           /* used by declustering */
286           int     noRotate;
287 
288 #if RF_INCLUDE_PARITYLOGGING > 0
289           /* used by parity logging */
290           RF_SectorCount_t regionLogCapacity;
291           RF_ParityLogQueue_t parityLogPool;      /* pool of unused parity logs */
292           RF_RegionInfo_t *regionInfo;  /* array of region state */
293           int     numParityLogs;
294           int     numSectorsPerLog;
295           int     regionParityRange;
296           int     logsInUse;  /* debugging */
297           RF_ParityLogDiskQueue_t parityLogDiskQueue;       /* state of parity
298                                                                        * logging disk work */
299           RF_RegionBufferQueue_t regionBufferPool;          /* buffers for holding
300                                                                        * region log */
301           RF_RegionBufferQueue_t parityBufferPool;          /* buffers for holding
302                                                                        * parity */
303           void *parityLogBufferHeap;    /* pool of unused parity logs */
304           RF_Thread_t pLogDiskThreadHandle;
305 
306 #endif                                  /* RF_INCLUDE_PARITYLOGGING > 0 */
307           struct rf_paritymap *parity_map;
308           struct RF_Pools_s pools;
309           struct RF_PoolNames_s poolNames;
310 };
311 
312 struct raid_softc {
313           struct dk_softc sc_dksc;
314           int       sc_unit;
315           int     sc_flags;   /* flags */
316           int     sc_cflags;  /* configuration flags */
317           kmutex_t sc_mutex;  /* interlock mutex */
318           kcondvar_t sc_cv;   /* and the condvar */
319           uint64_t sc_size;   /* size of the raid device */
320           char    sc_xname[20];         /* XXX external name */
321           RF_Raid_t sc_r;
322           LIST_ENTRY(raid_softc) sc_link;
323 };
324 /* sc_flags */
325 #define RAIDF_INITED                    0x01      /* unit has been initialized */
326 #define RAIDF_SHUTDOWN                  0x02      /* unit is being shutdown */
327 #define RAIDF_DETACH                    0x04      /* detach after final close */
328 #define RAIDF_WANTED                    0x08      /* someone waiting to obtain a lock */
329 #define RAIDF_LOCKED                    0x10      /* unit is locked */
330 #define RAIDF_UNIT_CHANGED    0x20      /* unit is being changed */
331 
332 
333 int rf_fail_disk(RF_Raid_t *, struct rf_recon_req *);
334 
335 int rf_inited(const struct raid_softc *);
336 int rf_get_unit(const struct raid_softc *);
337 RF_Raid_t *rf_get_raid(struct raid_softc *);
338 int rf_construct(struct raid_softc *, RF_Config_t *);
339 
340 #endif                                  /* !_RF__RF_RAID_H_ */
341