2 * SN Platform GRU Driver
4 * GRU DRIVER TABLES, MACROS, externs, etc
6 * Copyright (c) 2008 Silicon Graphics, Inc. All Rights Reserved.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 #ifndef __GRUTABLES_H__
24 #define __GRUTABLES_H__
28 * The GRU is a user addressible memory accelerator. It provides
29 * several forms of load, store, memset, bcopy instructions. In addition, it
30 * contains special instructions for AMOs, sending messages to message
33 * The GRU is an integral part of the node controller. It connects
34 * directly to the cpu socket. In its current implementation, there are 2
35 * GRU chiplets in the node controller on each blade (~node).
37 * The entire GRU memory space is fully coherent and cacheable by the cpus.
39 * Each GRU chiplet has a physical memory map that looks like the following:
52 * +-----------------+ _______ +-------------+
53 * |/////////////////| / | |
54 * |/////////////////| / | |
55 * |/////////////////| / | instructions|
56 * |/////////////////| / | |
57 * |/////////////////| / | |
58 * |/////////////////| / |-------------|
59 * |/////////////////| / | |
60 * +-----------------+ | |
61 * | context 15 | | data |
62 * +-----------------+ | |
64 * +-----------------+ \____________ +-------------+
70 * Each of the "contexts" is a chunk of memory that can be mmaped into user
71 * space. The context consists of 2 parts:
73 * - an instruction space that can be directly accessed by the user
74 * to issue GRU instructions and to check instruction status.
76 * - a data area that acts as normal RAM.
78 * User instructions contain virtual addresses of data to be accessed by the
79 * GRU. The GRU contains a TLB that is used to convert these user virtual
80 * addresses to physical addresses.
82 * The "system control" area of the GRU chiplet is used by the kernel driver
83 * to manage user contexts and to perform functions such as TLB dropin and
86 * One context may be reserved for the kernel and used for cross-partition
87 * communication. The GRU will also be used to asynchronously zero out
88 * large blocks of memory (not currently implemented).
93 * VDATA-VMA Data - Holds a few parameters. Head of linked list of
94 * GTS tables for threads using the GSEG
95 * GTS - Gru Thread State - contains info for managing a GSEG context. A
96 * GTS is allocated for each thread accessing a
98 * GTD - GRU Thread Data - contains shadow copy of GRU data when GSEG is
99 * not loaded into a GRU
100 * GMS - GRU Memory Struct - Used to manage TLB shootdowns. Tracks GRUs
101 * where a GSEG has been loaded. Similar to
102 * an mm_struct but for GRU.
104 * GS - GRU State - Used to manage the state of a GRU chiplet
105 * BS - Blade State - Used to manage state of all GRU chiplets
109 * Normal task tables for task using GRU.
110 * - 2 threads in process
111 * - 2 GSEGs open in process
112 * - GSEG1 is being used by both threads
113 * - GSEG2 is used only by thread 2
116 * task ---+---> mm ->------ (notifier) -------+-> gms
118 * |--> vma -> vdata ---> gts--->| GSEG1 (thread1)
120 * | +-> gts--->| GSEG1 (thread2)
122 * |--> vma -> vdata ---> gts--->| GSEG2 (thread2)
126 * GSEGs are marked DONTCOPY on fork
129 * file.private_data -> NULL
134 * After gseg reference
139 * vma -> vdata -> gts
141 * (vma is not copied)
145 #include <linux/rmap.h>
146 #include <linux/interrupt.h>
147 #include <linux/mutex.h>
148 #include <linux/wait.h>
149 #include <linux/mmu_notifier.h>
151 #include "gruhandles.h"
153 extern struct gru_stats_s gru_stats;
154 extern struct gru_blade_state *gru_base[];
155 extern unsigned long gru_start_paddr, gru_end_paddr;
157 #define GRU_MAX_BLADES MAX_NUMNODES
158 #define GRU_MAX_GRUS (GRU_MAX_BLADES * GRU_CHIPLETS_PER_BLADE)
160 #define GRU_DRIVER_ID_STR "SGI GRU Device Driver"
161 #define GRU_DRIVER_VERSION_STR "0.80"
167 atomic_long_t vdata_alloc;
168 atomic_long_t vdata_free;
169 atomic_long_t gts_alloc;
170 atomic_long_t gts_free;
171 atomic_long_t vdata_double_alloc;
172 atomic_long_t gts_double_allocate;
173 atomic_long_t assign_context;
174 atomic_long_t assign_context_failed;
175 atomic_long_t free_context;
176 atomic_long_t load_context;
177 atomic_long_t unload_context;
178 atomic_long_t steal_context;
179 atomic_long_t steal_context_failed;
181 atomic_long_t break_cow;
182 atomic_long_t asid_new;
183 atomic_long_t asid_next;
184 atomic_long_t asid_wrap;
185 atomic_long_t asid_reuse;
187 atomic_long_t call_os;
188 atomic_long_t call_os_check_for_bug;
189 atomic_long_t call_os_wait_queue;
190 atomic_long_t user_flush_tlb;
191 atomic_long_t user_unload_context;
192 atomic_long_t user_exception;
193 atomic_long_t set_task_slice;
194 atomic_long_t migrate_check;
195 atomic_long_t migrated_retarget;
196 atomic_long_t migrated_unload;
197 atomic_long_t migrated_unload_delay;
198 atomic_long_t migrated_nopfn_retarget;
199 atomic_long_t migrated_nopfn_unload;
200 atomic_long_t tlb_dropin;
201 atomic_long_t tlb_dropin_fail_no_asid;
202 atomic_long_t tlb_dropin_fail_upm;
203 atomic_long_t tlb_dropin_fail_invalid;
204 atomic_long_t tlb_dropin_fail_range_active;
205 atomic_long_t tlb_dropin_fail_idle;
206 atomic_long_t tlb_dropin_fail_fmm;
207 atomic_long_t mmu_invalidate_range;
208 atomic_long_t mmu_invalidate_page;
209 atomic_long_t mmu_clear_flush_young;
210 atomic_long_t flush_tlb;
211 atomic_long_t flush_tlb_gru;
212 atomic_long_t flush_tlb_gru_tgh;
213 atomic_long_t flush_tlb_gru_zero_asid;
215 atomic_long_t copy_gpa;
217 atomic_long_t mesq_receive;
218 atomic_long_t mesq_receive_none;
219 atomic_long_t mesq_send;
220 atomic_long_t mesq_send_failed;
221 atomic_long_t mesq_noop;
222 atomic_long_t mesq_send_unexpected_error;
223 atomic_long_t mesq_send_lb_overflow;
224 atomic_long_t mesq_send_qlimit_reached;
225 atomic_long_t mesq_send_amo_nacked;
226 atomic_long_t mesq_send_put_nacked;
227 atomic_long_t mesq_qf_not_full;
228 atomic_long_t mesq_qf_locked;
229 atomic_long_t mesq_qf_noop_not_full;
230 atomic_long_t mesq_qf_switch_head_failed;
231 atomic_long_t mesq_qf_unexpected_error;
232 atomic_long_t mesq_noop_unexpected_error;
233 atomic_long_t mesq_noop_lb_overflow;
234 atomic_long_t mesq_noop_qlimit_reached;
235 atomic_long_t mesq_noop_amo_nacked;
236 atomic_long_t mesq_noop_put_nacked;
242 #define GRU_QUICKLOOK 4
245 #define IRQ_GRU 110 /* Starting IRQ number for interrupts */
247 /* Delay in jiffies between attempts to assign a GRU context */
248 #define GRU_ASSIGN_DELAY ((HZ * 20) / 1000)
251 * If a process has it's context stolen, min delay in jiffies before trying to
252 * steal a context from another process.
254 #define GRU_STEAL_DELAY ((HZ * 200) / 1000)
256 #define STAT(id) do { \
257 if (gru_options & OPT_STATS) \
258 atomic_long_inc(&gru_stats.id); \
261 #ifdef CONFIG_SGI_GRU_DEBUG
262 #define gru_dbg(dev, fmt, x...) \
264 if (gru_options & OPT_DPRINT) \
265 dev_dbg(dev, "%s: " fmt, __func__, x); \
268 #define gru_dbg(x...)
271 /*-----------------------------------------------------------------------------
274 #define MAX_ASID 0xfffff0
276 #define ASID_INC 8 /* number of regions */
278 /* Generate a GRU asid value from a GRU base asid & a virtual address. */
279 #if defined CONFIG_IA64
280 #define VADDR_HI_BIT 64
281 #define GRUREGION(addr) ((addr) >> (VADDR_HI_BIT - 3) & 3)
282 #elif defined CONFIG_X86_64
283 #define VADDR_HI_BIT 48
284 #define GRUREGION(addr) (0) /* ZZZ could do better */
286 #error "Unsupported architecture"
288 #define GRUASID(asid, addr) ((asid) + GRUREGION(addr))
290 /*------------------------------------------------------------------------------
297 * This structure is pointed to from the mmstruct via the notifier pointer.
298 * There is one of these per address space.
300 struct gru_mm_tracker {
301 unsigned int mt_asid_gen; /* ASID wrap count */
302 int mt_asid; /* current base ASID for gru */
303 unsigned short mt_ctxbitmap; /* bitmap of contexts using
307 struct gru_mm_struct {
308 struct mmu_notifier ms_notifier;
310 spinlock_t ms_asid_lock; /* protects ASID assignment */
311 atomic_t ms_range_active;/* num range_invals active */
313 wait_queue_head_t ms_wait_queue;
314 DECLARE_BITMAP(ms_asidmap, GRU_MAX_GRUS);
315 struct gru_mm_tracker ms_asids[GRU_MAX_GRUS];
319 * One of these structures is allocated when a GSEG is mmaped. The
320 * structure is pointed to by the vma->vm_private_data field in the vma struct.
322 struct gru_vma_data {
323 spinlock_t vd_lock; /* Serialize access to vma */
324 struct list_head vd_head; /* head of linked list of gts */
325 long vd_user_options;/* misc user option flags */
331 * One of these is allocated for each thread accessing a mmaped GRU. A linked
332 * list of these structure is hung off the struct gru_vma_data in the mm_struct.
334 struct gru_thread_state {
335 struct list_head ts_next; /* list - head at vma-private */
336 struct mutex ts_ctxlock; /* load/unload CTX lock */
337 struct mm_struct *ts_mm; /* mm currently mapped to
339 struct vm_area_struct *ts_vma; /* vma of GRU context */
340 struct gru_state *ts_gru; /* GRU where the context is
342 struct gru_mm_struct *ts_gms; /* asid & ioproc struct */
343 unsigned long ts_cbr_map; /* map of allocated CBRs */
344 unsigned long ts_dsr_map; /* map of allocated DATA
346 unsigned long ts_steal_jiffies;/* jiffies when context last
348 long ts_user_options;/* misc user option flags */
349 pid_t ts_tgid_owner; /* task that is using the
350 context - for migration */
351 int ts_tsid; /* thread that owns the
353 int ts_tlb_int_select;/* target cpu if interrupts
355 int ts_ctxnum; /* context number where the
357 atomic_t ts_refcnt; /* reference count GTS */
358 unsigned char ts_dsr_au_count;/* Number of DSR resources
359 required for contest */
360 unsigned char ts_cbr_au_count;/* Number of CBR resources
361 required for contest */
362 char ts_force_unload;/* force context to be unloaded
364 char ts_cbr_idx[GRU_CBR_AU];/* CBR numbers of each
366 unsigned long ts_gdata[0]; /* save area for GRU data (CB,
371 * Threaded programs actually allocate an array of GSEGs when a context is
372 * created. Each thread uses a separate GSEG. TSID is the index into the GSEG
375 #define TSID(a, v) (((a) - (v)->vm_start) / GRU_GSEG_PAGESIZE)
376 #define UGRUADDR(gts) ((gts)->ts_vma->vm_start + \
377 (gts)->ts_tsid * GRU_GSEG_PAGESIZE)
379 #define NULLCTX (-1) /* if context not loaded into GRU */
381 /*-----------------------------------------------------------------------------
386 * One of these exists for each GRU chiplet.
389 struct gru_blade_state *gs_blade; /* GRU state for entire
391 unsigned long gs_gru_base_paddr; /* Physical address of
393 void *gs_gru_base_vaddr; /* Virtual address of
395 unsigned char gs_gid; /* unique GRU number */
396 unsigned char gs_tgh_local_shift; /* used to pick TGH for
398 unsigned char gs_tgh_first_remote; /* starting TGH# for
400 unsigned short gs_blade_id; /* blade of GRU */
401 spinlock_t gs_asid_lock; /* lock used for
403 spinlock_t gs_lock; /* lock used for
404 assigning contexts */
406 /* -- the following are protected by the gs_asid_lock spinlock ---- */
407 unsigned int gs_asid; /* Next availe ASID */
408 unsigned int gs_asid_limit; /* Limit of available
410 unsigned int gs_asid_gen; /* asid generation.
413 /* --- the following fields are protected by the gs_lock spinlock --- */
414 unsigned long gs_context_map; /* bitmap to manage
416 unsigned long gs_cbr_map; /* bitmap to manage CB
418 unsigned long gs_dsr_map; /* bitmap used to manage
420 unsigned int gs_reserved_cbrs; /* Number of kernel-
422 unsigned int gs_reserved_dsr_bytes; /* Bytes of kernel-
424 unsigned short gs_active_contexts; /* number of contexts
426 struct gru_thread_state *gs_gts[GRU_NUM_CCH]; /* GTS currently using
431 * This structure contains the GRU state for all the GRUs on a blade.
433 struct gru_blade_state {
434 void *kernel_cb; /* First kernel
436 void *kernel_dsr; /* First kernel
438 /* ---- the following are protected by the bs_lock spinlock ---- */
439 spinlock_t bs_lock; /* lock used for
441 int bs_lru_ctxnum; /* STEAL - last context
443 struct gru_state *bs_lru_gru; /* STEAL - last gru
446 struct gru_state bs_grus[GRU_CHIPLETS_PER_BLADE];
449 /*-----------------------------------------------------------------------------
452 #define get_tfm_for_cpu(g, c) \
453 ((struct gru_tlb_fault_map *)get_tfm((g)->gs_gru_base_vaddr, (c)))
454 #define get_tfh_by_index(g, i) \
455 ((struct gru_tlb_fault_handle *)get_tfh((g)->gs_gru_base_vaddr, (i)))
456 #define get_tgh_by_index(g, i) \
457 ((struct gru_tlb_global_handle *)get_tgh((g)->gs_gru_base_vaddr, (i)))
458 #define get_cbe_by_index(g, i) \
459 ((struct gru_control_block_extended *)get_cbe((g)->gs_gru_base_vaddr,\
462 /*-----------------------------------------------------------------------------
466 /* Given a blade# & chiplet#, get a pointer to the GRU */
467 #define get_gru(b, c) (&gru_base[b]->bs_grus[c])
469 /* Number of bytes to save/restore when unloading/loading GRU contexts */
470 #define DSR_BYTES(dsr) ((dsr) * GRU_DSR_AU_BYTES)
471 #define CBR_BYTES(cbr) ((cbr) * GRU_HANDLE_BYTES * GRU_CBR_AU_SIZE * 2)
473 /* Convert a user CB number to the actual CBRNUM */
474 #define thread_cbr_number(gts, n) ((gts)->ts_cbr_idx[(n) / GRU_CBR_AU_SIZE] \
475 * GRU_CBR_AU_SIZE + (n) % GRU_CBR_AU_SIZE)
477 /* Convert a gid to a pointer to the GRU */
478 #define GID_TO_GRU(gid) \
479 (gru_base[(gid) / GRU_CHIPLETS_PER_BLADE] ? \
480 (&gru_base[(gid) / GRU_CHIPLETS_PER_BLADE]-> \
481 bs_grus[(gid) % GRU_CHIPLETS_PER_BLADE]) : \
484 /* Scan all active GRUs in a GRU bitmap */
485 #define for_each_gru_in_bitmap(gid, map) \
486 for ((gid) = find_first_bit((map), GRU_MAX_GRUS); (gid) < GRU_MAX_GRUS;\
487 (gid)++, (gid) = find_next_bit((map), GRU_MAX_GRUS, (gid)))
489 /* Scan all active GRUs on a specific blade */
490 #define for_each_gru_on_blade(gru, nid, i) \
491 for ((gru) = gru_base[nid]->bs_grus, (i) = 0; \
492 (i) < GRU_CHIPLETS_PER_BLADE; \
495 /* Scan all active GTSs on a gru. Note: must hold ss_lock to use this macro. */
496 #define for_each_gts_on_gru(gts, gru, ctxnum) \
497 for ((ctxnum) = 0; (ctxnum) < GRU_NUM_CCH; (ctxnum)++) \
498 if (((gts) = (gru)->gs_gts[ctxnum]))
500 /* Scan each CBR whose bit is set in a TFM (or copy of) */
501 #define for_each_cbr_in_tfm(i, map) \
502 for ((i) = find_first_bit(map, GRU_NUM_CBE); \
504 (i)++, (i) = find_next_bit(map, GRU_NUM_CBE, i))
506 /* Scan each CBR in a CBR bitmap. Note: multiple CBRs in an allocation unit */
507 #define for_each_cbr_in_allocation_map(i, map, k) \
508 for ((k) = find_first_bit(map, GRU_CBR_AU); (k) < GRU_CBR_AU; \
509 (k) = find_next_bit(map, GRU_CBR_AU, (k) + 1)) \
510 for ((i) = (k)*GRU_CBR_AU_SIZE; \
511 (i) < ((k) + 1) * GRU_CBR_AU_SIZE; (i)++)
513 /* Scan each DSR in a DSR bitmap. Note: multiple DSRs in an allocation unit */
514 #define for_each_dsr_in_allocation_map(i, map, k) \
515 for ((k) = find_first_bit((const unsigned long *)map, GRU_DSR_AU);\
517 (k) = find_next_bit((const unsigned long *)map, \
518 GRU_DSR_AU, (k) + 1)) \
519 for ((i) = (k) * GRU_DSR_AU_CL; \
520 (i) < ((k) + 1) * GRU_DSR_AU_CL; (i)++)
522 #define gseg_physical_address(gru, ctxnum) \
523 ((gru)->gs_gru_base_paddr + ctxnum * GRU_GSEG_STRIDE)
524 #define gseg_virtual_address(gru, ctxnum) \
525 ((gru)->gs_gru_base_vaddr + ctxnum * GRU_GSEG_STRIDE)
527 /*-----------------------------------------------------------------------------
528 * Lock / Unlock GRU handles
529 * Use the "delresp" bit in the handle as a "lock" bit.
532 /* Lock hierarchy checking enabled only in emulator */
534 static inline void __lock_handle(void *h)
536 while (test_and_set_bit(1, h))
540 static inline void __unlock_handle(void *h)
545 static inline void lock_cch_handle(struct gru_context_configuration_handle *cch)
550 static inline void unlock_cch_handle(struct gru_context_configuration_handle
553 __unlock_handle(cch);
556 static inline void lock_tgh_handle(struct gru_tlb_global_handle *tgh)
561 static inline void unlock_tgh_handle(struct gru_tlb_global_handle *tgh)
563 __unlock_handle(tgh);
566 /*-----------------------------------------------------------------------------
567 * Function prototypes & externs
569 struct gru_unload_context_req;
571 extern struct vm_operations_struct gru_vm_ops;
572 extern struct device *grudev;
574 extern struct gru_vma_data *gru_alloc_vma_data(struct vm_area_struct *vma,
576 extern struct gru_thread_state *gru_find_thread_state(struct vm_area_struct
578 extern struct gru_thread_state *gru_alloc_thread_state(struct vm_area_struct
580 extern void gru_unload_context(struct gru_thread_state *gts, int savestate);
581 extern void gts_drop(struct gru_thread_state *gts);
582 extern void gru_tgh_flush_init(struct gru_state *gru);
583 extern int gru_kservices_init(struct gru_state *gru);
584 extern irqreturn_t gru_intr(int irq, void *dev_id);
585 extern int gru_handle_user_call_os(unsigned long address);
586 extern int gru_user_flush_tlb(unsigned long arg);
587 extern int gru_user_unload_context(unsigned long arg);
588 extern int gru_get_exception_detail(unsigned long arg);
589 extern int gru_set_task_slice(long address);
590 extern int gru_cpu_fault_map_id(void);
591 extern struct vm_area_struct *gru_find_vma(unsigned long vaddr);
592 extern void gru_flush_all_tlb(struct gru_state *gru);
593 extern int gru_proc_init(void);
594 extern void gru_proc_exit(void);
596 extern unsigned long gru_reserve_cb_resources(struct gru_state *gru,
597 int cbr_au_count, char *cbmap);
598 extern unsigned long gru_reserve_ds_resources(struct gru_state *gru,
599 int dsr_au_count, char *dsmap);
600 extern int gru_fault(struct vm_area_struct *, struct vm_fault *vmf);
601 extern struct gru_mm_struct *gru_register_mmu_notifier(void);
602 extern void gru_drop_mmu_notifier(struct gru_mm_struct *gms);
604 extern void gru_flush_tlb_range(struct gru_mm_struct *gms, unsigned long start,
607 extern unsigned long gru_options;
609 #endif /* __GRUTABLES_H__ */