[IA64-SGI] SGI Altix cross partition functionality [2/3]

This patch contains the communication module (XPC) for cross partition
communication on a partitioned SGI Altix.

Signed-off-by: Dean Nelson <dcn@sgi.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
This commit is contained in:
Dean Nelson 2005-03-23 19:50:00 -07:00 committed by Tony Luck
parent 21223a9e78
commit 89eb8eb927
5 changed files with 5325 additions and 0 deletions

View File

@ -13,3 +13,5 @@ obj-$(CONFIG_IA64_GENERIC) += machvec.o
obj-$(CONFIG_SGI_TIOCX) += tiocx.o
obj-$(CONFIG_IA64_SGI_SN_XP) += xp.o
xp-y := xp_main.o xp_nofault.o
obj-$(CONFIG_IA64_SGI_SN_XP) += xpc.o
xpc-y := xpc_main.o xpc_channel.o xpc_partition.o

991
arch/ia64/sn/kernel/xpc.h Normal file
View File

@ -0,0 +1,991 @@
/*
* This file is subject to the terms and conditions of the GNU General Public
* License. See the file "COPYING" in the main directory of this archive
* for more details.
*
* Copyright (c) 2004-2005 Silicon Graphics, Inc. All Rights Reserved.
*/
/*
* Cross Partition Communication (XPC) structures and macros.
*/
#ifndef _IA64_SN_KERNEL_XPC_H
#define _IA64_SN_KERNEL_XPC_H
#include <linux/config.h>
#include <linux/interrupt.h>
#include <linux/sysctl.h>
#include <linux/device.h>
#include <asm/pgtable.h>
#include <asm/processor.h>
#include <asm/sn/bte.h>
#include <asm/sn/clksupport.h>
#include <asm/sn/addrs.h>
#include <asm/sn/mspec.h>
#include <asm/sn/shub_mmr.h>
#include <asm/sn/xp.h>
/*
* XPC Version numbers consist of a major and minor number. XPC can always
* talk to versions with same major #, and never talk to versions with a
* different major #.
*/
#define _XPC_VERSION(_maj, _min) (((_maj) << 4) | ((_min) & 0xf))
#define XPC_VERSION_MAJOR(_v) ((_v) >> 4)
#define XPC_VERSION_MINOR(_v) ((_v) & 0xf)
/*
* The next macros define word or bit representations for given
* C-brick nasid in either the SAL provided bit array representing
* nasids in the partition/machine or the AMO_t array used for
* inter-partition initiation communications.
*
* For SN2 machines, C-Bricks are alway even numbered NASIDs. As
* such, some space will be saved by insisting that nasid information
* passed from SAL always be packed for C-Bricks and the
* cross-partition interrupts use the same packing scheme.
*/
#define XPC_NASID_W_INDEX(_n) (((_n) / 64) / 2)
#define XPC_NASID_B_INDEX(_n) (((_n) / 2) & (64 - 1))
#define XPC_NASID_IN_ARRAY(_n, _p) ((_p)[XPC_NASID_W_INDEX(_n)] & \
(1UL << XPC_NASID_B_INDEX(_n)))
#define XPC_NASID_FROM_W_B(_w, _b) (((_w) * 64 + (_b)) * 2)
#define XPC_HB_DEFAULT_INTERVAL 5 /* incr HB every x secs */
#define XPC_HB_CHECK_DEFAULT_TIMEOUT 20 /* check HB every x secs */
/* define the process name of HB checker and the CPU it is pinned to */
#define XPC_HB_CHECK_THREAD_NAME "xpc_hb"
#define XPC_HB_CHECK_CPU 0
/* define the process name of the discovery thread */
#define XPC_DISCOVERY_THREAD_NAME "xpc_discovery"
#define XPC_HB_ALLOWED(_p, _v) ((_v)->heartbeating_to_mask & (1UL << (_p)))
#define XPC_ALLOW_HB(_p, _v) (_v)->heartbeating_to_mask |= (1UL << (_p))
#define XPC_DISALLOW_HB(_p, _v) (_v)->heartbeating_to_mask &= (~(1UL << (_p)))
/*
* Reserved Page provided by SAL.
*
* SAL provides one page per partition of reserved memory. When SAL
* initialization is complete, SAL_signature, SAL_version, partid,
* part_nasids, and mach_nasids are set.
*
* Note: Until vars_pa is set, the partition XPC code has not been initialized.
*/
struct xpc_rsvd_page {
u64 SAL_signature; /* SAL unique signature */
u64 SAL_version; /* SAL specified version */
u8 partid; /* partition ID from SAL */
u8 version;
u8 pad[6]; /* pad to u64 align */
u64 vars_pa;
u64 part_nasids[XP_NASID_MASK_WORDS] ____cacheline_aligned;
u64 mach_nasids[XP_NASID_MASK_WORDS] ____cacheline_aligned;
};
#define XPC_RP_VERSION _XPC_VERSION(1,0) /* version 1.0 of the reserved page */
#define XPC_RSVD_PAGE_ALIGNED_SIZE \
(L1_CACHE_ALIGN(sizeof(struct xpc_rsvd_page)))
/*
* Define the structures by which XPC variables can be exported to other
* partitions. (There are two: struct xpc_vars and struct xpc_vars_part)
*/
/*
* The following structure describes the partition generic variables
* needed by other partitions in order to properly initialize.
*
* struct xpc_vars version number also applies to struct xpc_vars_part.
* Changes to either structure and/or related functionality should be
* reflected by incrementing either the major or minor version numbers
* of struct xpc_vars.
*/
struct xpc_vars {
u8 version;
u64 heartbeat;
u64 heartbeating_to_mask;
u64 kdb_status; /* 0 = machine running */
int act_nasid;
int act_phys_cpuid;
u64 vars_part_pa;
u64 amos_page_pa; /* paddr of page of AMOs from MSPEC driver */
AMO_t *amos_page; /* vaddr of page of AMOs from MSPEC driver */
AMO_t *act_amos; /* pointer to the first activation AMO */
};
#define XPC_V_VERSION _XPC_VERSION(3,0) /* version 3.0 of the cross vars */
#define XPC_VARS_ALIGNED_SIZE (L1_CACHE_ALIGN(sizeof(struct xpc_vars)))
/*
* The following structure describes the per partition specific variables.
*
* An array of these structures, one per partition, will be defined. As a
* partition becomes active XPC will copy the array entry corresponding to
* itself from that partition. It is desirable that the size of this
* structure evenly divide into a cacheline, such that none of the entries
* in this array crosses a cacheline boundary. As it is now, each entry
* occupies half a cacheline.
*/
struct xpc_vars_part {
u64 magic;
u64 openclose_args_pa; /* physical address of open and close args */
u64 GPs_pa; /* physical address of Get/Put values */
u64 IPI_amo_pa; /* physical address of IPI AMO_t structure */
int IPI_nasid; /* nasid of where to send IPIs */
int IPI_phys_cpuid; /* physical CPU ID of where to send IPIs */
u8 nchannels; /* #of defined channels supported */
u8 reserved[23]; /* pad to a full 64 bytes */
};
/*
* The vars_part MAGIC numbers play a part in the first contact protocol.
*
* MAGIC1 indicates that the per partition specific variables for a remote
* partition have been initialized by this partition.
*
* MAGIC2 indicates that this partition has pulled the remote partititions
* per partition variables that pertain to this partition.
*/
#define XPC_VP_MAGIC1 0x0053524156435058L /* 'XPCVARS\0'L (little endian) */
#define XPC_VP_MAGIC2 0x0073726176435058L /* 'XPCvars\0'L (little endian) */
/*
* Functions registered by add_timer() or called by kernel_thread() only
* allow for a single 64-bit argument. The following macros can be used to
* pack and unpack two (32-bit, 16-bit or 8-bit) arguments into or out from
* the passed argument.
*/
#define XPC_PACK_ARGS(_arg1, _arg2) \
((((u64) _arg1) & 0xffffffff) | \
((((u64) _arg2) & 0xffffffff) << 32))
#define XPC_UNPACK_ARG1(_args) (((u64) _args) & 0xffffffff)
#define XPC_UNPACK_ARG2(_args) ((((u64) _args) >> 32) & 0xffffffff)
/*
* Define a Get/Put value pair (pointers) used with a message queue.
*/
struct xpc_gp {
s64 get; /* Get value */
s64 put; /* Put value */
};
#define XPC_GP_SIZE \
L1_CACHE_ALIGN(sizeof(struct xpc_gp) * XPC_NCHANNELS)
/*
* Define a structure that contains arguments associated with opening and
* closing a channel.
*/
struct xpc_openclose_args {
u16 reason; /* reason why channel is closing */
u16 msg_size; /* sizeof each message entry */
u16 remote_nentries; /* #of message entries in remote msg queue */
u16 local_nentries; /* #of message entries in local msg queue */
u64 local_msgqueue_pa; /* physical address of local message queue */
};
#define XPC_OPENCLOSE_ARGS_SIZE \
L1_CACHE_ALIGN(sizeof(struct xpc_openclose_args) * XPC_NCHANNELS)
/* struct xpc_msg flags */
#define XPC_M_DONE 0x01 /* msg has been received/consumed */
#define XPC_M_READY 0x02 /* msg is ready to be sent */
#define XPC_M_INTERRUPT 0x04 /* send interrupt when msg consumed */
#define XPC_MSG_ADDRESS(_payload) \
((struct xpc_msg *)((u8 *)(_payload) - XPC_MSG_PAYLOAD_OFFSET))
/*
* Defines notify entry.
*
* This is used to notify a message's sender that their message was received
* and consumed by the intended recipient.
*/
struct xpc_notify {
struct semaphore sema; /* notify semaphore */
u8 type; /* type of notification */
/* the following two fields are only used if type == XPC_N_CALL */
xpc_notify_func func; /* user's notify function */
void *key; /* pointer to user's key */
};
/* struct xpc_notify type of notification */
#define XPC_N_CALL 0x01 /* notify function provided by user */
/*
* Define the structure that manages all the stuff required by a channel. In
* particular, they are used to manage the messages sent across the channel.
*
* This structure is private to a partition, and is NOT shared across the
* partition boundary.
*
* There is an array of these structures for each remote partition. It is
* allocated at the time a partition becomes active. The array contains one
* of these structures for each potential channel connection to that partition.
*
* Each of these structures manages two message queues (circular buffers).
* They are allocated at the time a channel connection is made. One of
* these message queues (local_msgqueue) holds the locally created messages
* that are destined for the remote partition. The other of these message
* queues (remote_msgqueue) is a locally cached copy of the remote partition's
* own local_msgqueue.
*
* The following is a description of the Get/Put pointers used to manage these
* two message queues. Consider the local_msgqueue to be on one partition
* and the remote_msgqueue to be its cached copy on another partition. A
* description of what each of the lettered areas contains is included.
*
*
* local_msgqueue remote_msgqueue
*
* |/////////| |/////////|
* w_remote_GP.get --> +---------+ |/////////|
* | F | |/////////|
* remote_GP.get --> +---------+ +---------+ <-- local_GP->get
* | | | |
* | | | E |
* | | | |
* | | +---------+ <-- w_local_GP.get
* | B | |/////////|
* | | |////D////|
* | | |/////////|
* | | +---------+ <-- w_remote_GP.put
* | | |////C////|
* local_GP->put --> +---------+ +---------+ <-- remote_GP.put
* | | |/////////|
* | A | |/////////|
* | | |/////////|
* w_local_GP.put --> +---------+ |/////////|
* |/////////| |/////////|
*
*
* ( remote_GP.[get|put] are cached copies of the remote
* partition's local_GP->[get|put], and thus their values can
* lag behind their counterparts on the remote partition. )
*
*
* A - Messages that have been allocated, but have not yet been sent to the
* remote partition.
*
* B - Messages that have been sent, but have not yet been acknowledged by the
* remote partition as having been received.
*
* C - Area that needs to be prepared for the copying of sent messages, by
* the clearing of the message flags of any previously received messages.
*
* D - Area into which sent messages are to be copied from the remote
* partition's local_msgqueue and then delivered to their intended
* recipients. [ To allow for a multi-message copy, another pointer
* (next_msg_to_pull) has been added to keep track of the next message
* number needing to be copied (pulled). It chases after w_remote_GP.put.
* Any messages lying between w_local_GP.get and next_msg_to_pull have
* been copied and are ready to be delivered. ]
*
* E - Messages that have been copied and delivered, but have not yet been
* acknowledged by the recipient as having been received.
*
* F - Messages that have been acknowledged, but XPC has not yet notified the
* sender that the message was received by its intended recipient.
* This is also an area that needs to be prepared for the allocating of
* new messages, by the clearing of the message flags of the acknowledged
* messages.
*/
struct xpc_channel {
partid_t partid; /* ID of remote partition connected */
spinlock_t lock; /* lock for updating this structure */
u32 flags; /* general flags */
enum xpc_retval reason; /* reason why channel is disconnect'g */
int reason_line; /* line# disconnect initiated from */
u16 number; /* channel # */
u16 msg_size; /* sizeof each msg entry */
u16 local_nentries; /* #of msg entries in local msg queue */
u16 remote_nentries; /* #of msg entries in remote msg queue*/
void *local_msgqueue_base; /* base address of kmalloc'd space */
struct xpc_msg *local_msgqueue; /* local message queue */
void *remote_msgqueue_base; /* base address of kmalloc'd space */
struct xpc_msg *remote_msgqueue;/* cached copy of remote partition's */
/* local message queue */
u64 remote_msgqueue_pa; /* phys addr of remote partition's */
/* local message queue */
atomic_t references; /* #of external references to queues */
atomic_t n_on_msg_allocate_wq; /* #on msg allocation wait queue */
wait_queue_head_t msg_allocate_wq; /* msg allocation wait queue */
/* queue of msg senders who want to be notified when msg received */
atomic_t n_to_notify; /* #of msg senders to notify */
struct xpc_notify *notify_queue;/* notify queue for messages sent */
xpc_channel_func func; /* user's channel function */
void *key; /* pointer to user's key */
struct semaphore msg_to_pull_sema; /* next msg to pull serialization */
struct semaphore teardown_sema; /* wait for teardown completion */
struct xpc_openclose_args *local_openclose_args; /* args passed on */
/* opening or closing of channel */
/* various flavors of local and remote Get/Put values */
struct xpc_gp *local_GP; /* local Get/Put values */
struct xpc_gp remote_GP; /* remote Get/Put values */
struct xpc_gp w_local_GP; /* working local Get/Put values */
struct xpc_gp w_remote_GP; /* working remote Get/Put values */
s64 next_msg_to_pull; /* Put value of next msg to pull */
/* kthread management related fields */
// >>> rethink having kthreads_assigned_limit and kthreads_idle_limit; perhaps
// >>> allow the assigned limit be unbounded and let the idle limit be dynamic
// >>> dependent on activity over the last interval of time
atomic_t kthreads_assigned; /* #of kthreads assigned to channel */
u32 kthreads_assigned_limit; /* limit on #of kthreads assigned */
atomic_t kthreads_idle; /* #of kthreads idle waiting for work */
u32 kthreads_idle_limit; /* limit on #of kthreads idle */
atomic_t kthreads_active; /* #of kthreads actively working */
// >>> following field is temporary
u32 kthreads_created; /* total #of kthreads created */
wait_queue_head_t idle_wq; /* idle kthread wait queue */
} ____cacheline_aligned;
/* struct xpc_channel flags */
#define XPC_C_WASCONNECTED 0x00000001 /* channel was connected */
#define XPC_C_ROPENREPLY 0x00000002 /* remote open channel reply */
#define XPC_C_OPENREPLY 0x00000004 /* local open channel reply */
#define XPC_C_ROPENREQUEST 0x00000008 /* remote open channel request */
#define XPC_C_OPENREQUEST 0x00000010 /* local open channel request */
#define XPC_C_SETUP 0x00000020 /* channel's msgqueues are alloc'd */
#define XPC_C_CONNECTCALLOUT 0x00000040 /* channel connected callout made */
#define XPC_C_CONNECTED 0x00000080 /* local channel is connected */
#define XPC_C_CONNECTING 0x00000100 /* channel is being connected */
#define XPC_C_RCLOSEREPLY 0x00000200 /* remote close channel reply */
#define XPC_C_CLOSEREPLY 0x00000400 /* local close channel reply */
#define XPC_C_RCLOSEREQUEST 0x00000800 /* remote close channel request */
#define XPC_C_CLOSEREQUEST 0x00001000 /* local close channel request */
#define XPC_C_DISCONNECTED 0x00002000 /* channel is disconnected */
#define XPC_C_DISCONNECTING 0x00004000 /* channel is being disconnected */
/*
* Manages channels on a partition basis. There is one of these structures
* for each partition (a partition will never utilize the structure that
* represents itself).
*/
struct xpc_partition {
/* XPC HB infrastructure */
u64 remote_rp_pa; /* phys addr of partition's rsvd pg */
u64 remote_vars_pa; /* phys addr of partition's vars */
u64 remote_vars_part_pa; /* phys addr of partition's vars part */
u64 last_heartbeat; /* HB at last read */
u64 remote_amos_page_pa; /* phys addr of partition's amos page */
int remote_act_nasid; /* active part's act/deact nasid */
int remote_act_phys_cpuid; /* active part's act/deact phys cpuid */
u32 act_IRQ_rcvd; /* IRQs since activation */
spinlock_t act_lock; /* protect updating of act_state */
u8 act_state; /* from XPC HB viewpoint */
enum xpc_retval reason; /* reason partition is deactivating */
int reason_line; /* line# deactivation initiated from */
int reactivate_nasid; /* nasid in partition to reactivate */
/* XPC infrastructure referencing and teardown control */
u8 setup_state; /* infrastructure setup state */
wait_queue_head_t teardown_wq; /* kthread waiting to teardown infra */
atomic_t references; /* #of references to infrastructure */
/*
* NONE OF THE PRECEDING FIELDS OF THIS STRUCTURE WILL BE CLEARED WHEN
* XPC SETS UP THE NECESSARY INFRASTRUCTURE TO SUPPORT CROSS PARTITION
* COMMUNICATION. ALL OF THE FOLLOWING FIELDS WILL BE CLEARED. (THE
* 'nchannels' FIELD MUST BE THE FIRST OF THE FIELDS TO BE CLEARED.)
*/
u8 nchannels; /* #of defined channels supported */
atomic_t nchannels_active; /* #of channels that are not DISCONNECTED */
struct xpc_channel *channels;/* array of channel structures */
void *local_GPs_base; /* base address of kmalloc'd space */
struct xpc_gp *local_GPs; /* local Get/Put values */
void *remote_GPs_base; /* base address of kmalloc'd space */
struct xpc_gp *remote_GPs;/* copy of remote partition's local Get/Put */
/* values */
u64 remote_GPs_pa; /* phys address of remote partition's local */
/* Get/Put values */
/* fields used to pass args when opening or closing a channel */
void *local_openclose_args_base; /* base address of kmalloc'd space */
struct xpc_openclose_args *local_openclose_args; /* local's args */
void *remote_openclose_args_base; /* base address of kmalloc'd space */
struct xpc_openclose_args *remote_openclose_args; /* copy of remote's */
/* args */
u64 remote_openclose_args_pa; /* phys addr of remote's args */
/* IPI sending, receiving and handling related fields */
int remote_IPI_nasid; /* nasid of where to send IPIs */
int remote_IPI_phys_cpuid; /* phys CPU ID of where to send IPIs */
AMO_t *remote_IPI_amo_va; /* address of remote IPI AMO_t structure */
AMO_t *local_IPI_amo_va; /* address of IPI AMO_t structure */
u64 local_IPI_amo; /* IPI amo flags yet to be handled */
char IPI_owner[8]; /* IPI owner's name */
struct timer_list dropped_IPI_timer; /* dropped IPI timer */
spinlock_t IPI_lock; /* IPI handler lock */
/* channel manager related fields */
atomic_t channel_mgr_requests; /* #of requests to activate chan mgr */
wait_queue_head_t channel_mgr_wq; /* channel mgr's wait queue */
} ____cacheline_aligned;
/* struct xpc_partition act_state values (for XPC HB) */
#define XPC_P_INACTIVE 0x00 /* partition is not active */
#define XPC_P_ACTIVATION_REQ 0x01 /* created thread to activate */
#define XPC_P_ACTIVATING 0x02 /* activation thread started */
#define XPC_P_ACTIVE 0x03 /* xpc_partition_up() was called */
#define XPC_P_DEACTIVATING 0x04 /* partition deactivation initiated */
#define XPC_DEACTIVATE_PARTITION(_p, _reason) \
xpc_deactivate_partition(__LINE__, (_p), (_reason))
/* struct xpc_partition setup_state values */
#define XPC_P_UNSET 0x00 /* infrastructure was never setup */
#define XPC_P_SETUP 0x01 /* infrastructure is setup */
#define XPC_P_WTEARDOWN 0x02 /* waiting to teardown infrastructure */
#define XPC_P_TORNDOWN 0x03 /* infrastructure is torndown */
/*
* struct xpc_partition IPI_timer #of seconds to wait before checking for
* dropped IPIs. These occur whenever an IPI amo write doesn't complete until
* after the IPI was received.
*/
#define XPC_P_DROPPED_IPI_WAIT (0.25 * HZ)
#define XPC_PARTID(_p) ((partid_t) ((_p) - &xpc_partitions[0]))
/* found in xp_main.c */
extern struct xpc_registration xpc_registrations[];
/* >>> found in xpc_main.c only */
extern struct device *xpc_part;
extern struct device *xpc_chan;
extern irqreturn_t xpc_notify_IRQ_handler(int, void *, struct pt_regs *);
extern void xpc_dropped_IPI_check(struct xpc_partition *);
extern void xpc_activate_kthreads(struct xpc_channel *, int);
extern void xpc_create_kthreads(struct xpc_channel *, int);
extern void xpc_disconnect_wait(int);
/* found in xpc_main.c and efi-xpc.c */
extern void xpc_activate_partition(struct xpc_partition *);
/* found in xpc_partition.c */
extern int xpc_exiting;
extern int xpc_hb_interval;
extern int xpc_hb_check_interval;
extern struct xpc_vars *xpc_vars;
extern struct xpc_rsvd_page *xpc_rsvd_page;
extern struct xpc_vars_part *xpc_vars_part;
extern struct xpc_partition xpc_partitions[XP_MAX_PARTITIONS + 1];
extern char xpc_remote_copy_buffer[];
extern struct xpc_rsvd_page *xpc_rsvd_page_init(void);
extern void xpc_allow_IPI_ops(void);
extern void xpc_restrict_IPI_ops(void);
extern int xpc_identify_act_IRQ_sender(void);
extern enum xpc_retval xpc_mark_partition_active(struct xpc_partition *);
extern void xpc_mark_partition_inactive(struct xpc_partition *);
extern void xpc_discovery(void);
extern void xpc_check_remote_hb(void);
extern void xpc_deactivate_partition(const int, struct xpc_partition *,
enum xpc_retval);
extern enum xpc_retval xpc_initiate_partid_to_nasids(partid_t, void *);
/* found in xpc_channel.c */
extern void xpc_initiate_connect(int);
extern void xpc_initiate_disconnect(int);
extern enum xpc_retval xpc_initiate_allocate(partid_t, int, u32, void **);
extern enum xpc_retval xpc_initiate_send(partid_t, int, void *);
extern enum xpc_retval xpc_initiate_send_notify(partid_t, int, void *,
xpc_notify_func, void *);
extern void xpc_initiate_received(partid_t, int, void *);
extern enum xpc_retval xpc_setup_infrastructure(struct xpc_partition *);
extern enum xpc_retval xpc_pull_remote_vars_part(struct xpc_partition *);
extern void xpc_process_channel_activity(struct xpc_partition *);
extern void xpc_connected_callout(struct xpc_channel *);
extern void xpc_deliver_msg(struct xpc_channel *);
extern void xpc_disconnect_channel(const int, struct xpc_channel *,
enum xpc_retval, unsigned long *);
extern void xpc_disconnected_callout(struct xpc_channel *);
extern void xpc_partition_down(struct xpc_partition *, enum xpc_retval);
extern void xpc_teardown_infrastructure(struct xpc_partition *);
static inline void
xpc_wakeup_channel_mgr(struct xpc_partition *part)
{
if (atomic_inc_return(&part->channel_mgr_requests) == 1) {
wake_up(&part->channel_mgr_wq);
}
}
/*
* These next two inlines are used to keep us from tearing down a channel's
* msg queues while a thread may be referencing them.
*/
static inline void
xpc_msgqueue_ref(struct xpc_channel *ch)
{
atomic_inc(&ch->references);
}
static inline void
xpc_msgqueue_deref(struct xpc_channel *ch)
{
s32 refs = atomic_dec_return(&ch->references);
DBUG_ON(refs < 0);
if (refs == 0) {
xpc_wakeup_channel_mgr(&xpc_partitions[ch->partid]);
}
}
#define XPC_DISCONNECT_CHANNEL(_ch, _reason, _irqflgs) \
xpc_disconnect_channel(__LINE__, _ch, _reason, _irqflgs)
/*
* These two inlines are used to keep us from tearing down a partition's
* setup infrastructure while a thread may be referencing it.
*/
static inline void
xpc_part_deref(struct xpc_partition *part)
{
s32 refs = atomic_dec_return(&part->references);
DBUG_ON(refs < 0);
if (refs == 0 && part->setup_state == XPC_P_WTEARDOWN) {
wake_up(&part->teardown_wq);
}
}
static inline int
xpc_part_ref(struct xpc_partition *part)
{
int setup;
atomic_inc(&part->references);
setup = (part->setup_state == XPC_P_SETUP);
if (!setup) {
xpc_part_deref(part);
}
return setup;
}
/*
* The following macro is to be used for the setting of the reason and
* reason_line fields in both the struct xpc_channel and struct xpc_partition
* structures.
*/
#define XPC_SET_REASON(_p, _reason, _line) \
{ \
(_p)->reason = _reason; \
(_p)->reason_line = _line; \
}
/*
* The following set of macros and inlines are used for the sending and
* receiving of IPIs (also known as IRQs). There are two flavors of IPIs,
* one that is associated with partition activity (SGI_XPC_ACTIVATE) and
* the other that is associated with channel activity (SGI_XPC_NOTIFY).
*/
static inline u64
xpc_IPI_receive(AMO_t *amo)
{
return FETCHOP_LOAD_OP(TO_AMO((u64) &amo->variable), FETCHOP_CLEAR);
}
static inline enum xpc_retval
xpc_IPI_send(AMO_t *amo, u64 flag, int nasid, int phys_cpuid, int vector)
{
int ret = 0;
unsigned long irq_flags;
local_irq_save(irq_flags);
FETCHOP_STORE_OP(TO_AMO((u64) &amo->variable), FETCHOP_OR, flag);
sn_send_IPI_phys(nasid, phys_cpuid, vector, 0);
/*
* We must always use the nofault function regardless of whether we
* are on a Shub 1.1 system or a Shub 1.2 slice 0xc processor. If we
* didn't, we'd never know that the other partition is down and would
* keep sending IPIs and AMOs to it until the heartbeat times out.
*/
ret = xp_nofault_PIOR((u64 *) GLOBAL_MMR_ADDR(NASID_GET(&amo->variable),
xp_nofault_PIOR_target));
local_irq_restore(irq_flags);
return ((ret == 0) ? xpcSuccess : xpcPioReadError);
}
/*
* IPIs associated with SGI_XPC_ACTIVATE IRQ.
*/
/*
* Flag the appropriate AMO variable and send an IPI to the specified node.
*/
static inline void
xpc_activate_IRQ_send(u64 amos_page, int from_nasid, int to_nasid,
int to_phys_cpuid)
{
int w_index = XPC_NASID_W_INDEX(from_nasid);
int b_index = XPC_NASID_B_INDEX(from_nasid);
AMO_t *amos = (AMO_t *) __va(amos_page +
(XP_MAX_PARTITIONS * sizeof(AMO_t)));
(void) xpc_IPI_send(&amos[w_index], (1UL << b_index), to_nasid,
to_phys_cpuid, SGI_XPC_ACTIVATE);
}
static inline void
xpc_IPI_send_activate(struct xpc_vars *vars)
{
xpc_activate_IRQ_send(vars->amos_page_pa, cnodeid_to_nasid(0),
vars->act_nasid, vars->act_phys_cpuid);
}
static inline void
xpc_IPI_send_activated(struct xpc_partition *part)
{
xpc_activate_IRQ_send(part->remote_amos_page_pa, cnodeid_to_nasid(0),
part->remote_act_nasid, part->remote_act_phys_cpuid);
}
static inline void
xpc_IPI_send_reactivate(struct xpc_partition *part)
{
xpc_activate_IRQ_send(xpc_vars->amos_page_pa, part->reactivate_nasid,
xpc_vars->act_nasid, xpc_vars->act_phys_cpuid);
}
/*
* IPIs associated with SGI_XPC_NOTIFY IRQ.
*/
/*
* Send an IPI to the remote partition that is associated with the
* specified channel.
*/
#define XPC_NOTIFY_IRQ_SEND(_ch, _ipi_f, _irq_f) \
xpc_notify_IRQ_send(_ch, _ipi_f, #_ipi_f, _irq_f)
static inline void
xpc_notify_IRQ_send(struct xpc_channel *ch, u8 ipi_flag, char *ipi_flag_string,
unsigned long *irq_flags)
{
struct xpc_partition *part = &xpc_partitions[ch->partid];
enum xpc_retval ret;
if (likely(part->act_state != XPC_P_DEACTIVATING)) {
ret = xpc_IPI_send(part->remote_IPI_amo_va,
(u64) ipi_flag << (ch->number * 8),
part->remote_IPI_nasid,
part->remote_IPI_phys_cpuid,
SGI_XPC_NOTIFY);
dev_dbg(xpc_chan, "%s sent to partid=%d, channel=%d, ret=%d\n",
ipi_flag_string, ch->partid, ch->number, ret);
if (unlikely(ret != xpcSuccess)) {
if (irq_flags != NULL) {
spin_unlock_irqrestore(&ch->lock, *irq_flags);
}
XPC_DEACTIVATE_PARTITION(part, ret);
if (irq_flags != NULL) {
spin_lock_irqsave(&ch->lock, *irq_flags);
}
}
}
}
/*
* Make it look like the remote partition, which is associated with the
* specified channel, sent us an IPI. This faked IPI will be handled
* by xpc_dropped_IPI_check().
*/
#define XPC_NOTIFY_IRQ_SEND_LOCAL(_ch, _ipi_f) \
xpc_notify_IRQ_send_local(_ch, _ipi_f, #_ipi_f)
static inline void
xpc_notify_IRQ_send_local(struct xpc_channel *ch, u8 ipi_flag,
char *ipi_flag_string)
{
struct xpc_partition *part = &xpc_partitions[ch->partid];
FETCHOP_STORE_OP(TO_AMO((u64) &part->local_IPI_amo_va->variable),
FETCHOP_OR, ((u64) ipi_flag << (ch->number * 8)));
dev_dbg(xpc_chan, "%s sent local from partid=%d, channel=%d\n",
ipi_flag_string, ch->partid, ch->number);
}
/*
* The sending and receiving of IPIs includes the setting of an AMO variable
* to indicate the reason the IPI was sent. The 64-bit variable is divided
* up into eight bytes, ordered from right to left. Byte zero pertains to
* channel 0, byte one to channel 1, and so on. Each byte is described by
* the following IPI flags.
*/
#define XPC_IPI_CLOSEREQUEST 0x01
#define XPC_IPI_CLOSEREPLY 0x02
#define XPC_IPI_OPENREQUEST 0x04
#define XPC_IPI_OPENREPLY 0x08
#define XPC_IPI_MSGREQUEST 0x10
/* given an AMO variable and a channel#, get its associated IPI flags */
#define XPC_GET_IPI_FLAGS(_amo, _c) ((u8) (((_amo) >> ((_c) * 8)) & 0xff))
#define XPC_ANY_OPENCLOSE_IPI_FLAGS_SET(_amo) ((_amo) & 0x0f0f0f0f0f0f0f0f)
#define XPC_ANY_MSG_IPI_FLAGS_SET(_amo) ((_amo) & 0x1010101010101010)
static inline void
xpc_IPI_send_closerequest(struct xpc_channel *ch, unsigned long *irq_flags)
{
struct xpc_openclose_args *args = ch->local_openclose_args;
args->reason = ch->reason;
XPC_NOTIFY_IRQ_SEND(ch, XPC_IPI_CLOSEREQUEST, irq_flags);
}
static inline void
xpc_IPI_send_closereply(struct xpc_channel *ch, unsigned long *irq_flags)
{
XPC_NOTIFY_IRQ_SEND(ch, XPC_IPI_CLOSEREPLY, irq_flags);
}
static inline void
xpc_IPI_send_openrequest(struct xpc_channel *ch, unsigned long *irq_flags)
{
struct xpc_openclose_args *args = ch->local_openclose_args;
args->msg_size = ch->msg_size;
args->local_nentries = ch->local_nentries;
XPC_NOTIFY_IRQ_SEND(ch, XPC_IPI_OPENREQUEST, irq_flags);
}
static inline void
xpc_IPI_send_openreply(struct xpc_channel *ch, unsigned long *irq_flags)
{
struct xpc_openclose_args *args = ch->local_openclose_args;
args->remote_nentries = ch->remote_nentries;
args->local_nentries = ch->local_nentries;
args->local_msgqueue_pa = __pa(ch->local_msgqueue);
XPC_NOTIFY_IRQ_SEND(ch, XPC_IPI_OPENREPLY, irq_flags);
}
static inline void
xpc_IPI_send_msgrequest(struct xpc_channel *ch)
{
XPC_NOTIFY_IRQ_SEND(ch, XPC_IPI_MSGREQUEST, NULL);
}
static inline void
xpc_IPI_send_local_msgrequest(struct xpc_channel *ch)
{
XPC_NOTIFY_IRQ_SEND_LOCAL(ch, XPC_IPI_MSGREQUEST);
}
/*
* Memory for XPC's AMO variables is allocated by the MSPEC driver. These
* pages are located in the lowest granule. The lowest granule uses 4k pages
* for cached references and an alternate TLB handler to never provide a
* cacheable mapping for the entire region. This will prevent speculative
* reading of cached copies of our lines from being issued which will cause
* a PI FSB Protocol error to be generated by the SHUB. For XPC, we need 64
* (XP_MAX_PARTITIONS) AMO variables for message notification (xpc_main.c)
* and an additional 16 AMO variables for partition activation (xpc_hb.c).
*/
static inline AMO_t *
xpc_IPI_init(partid_t partid)
{
AMO_t *part_amo = xpc_vars->amos_page + partid;
xpc_IPI_receive(part_amo);
return part_amo;
}
static inline enum xpc_retval
xpc_map_bte_errors(bte_result_t error)
{
switch (error) {
case BTE_SUCCESS: return xpcSuccess;
case BTEFAIL_DIR: return xpcBteDirectoryError;
case BTEFAIL_POISON: return xpcBtePoisonError;
case BTEFAIL_WERR: return xpcBteWriteError;
case BTEFAIL_ACCESS: return xpcBteAccessError;
case BTEFAIL_PWERR: return xpcBtePWriteError;
case BTEFAIL_PRERR: return xpcBtePReadError;
case BTEFAIL_TOUT: return xpcBteTimeOutError;
case BTEFAIL_XTERR: return xpcBteXtalkError;
case BTEFAIL_NOTAVAIL: return xpcBteNotAvailable;
default: return xpcBteUnmappedError;
}
}
static inline void *
xpc_kmalloc_cacheline_aligned(size_t size, int flags, void **base)
{
/* see if kmalloc will give us cachline aligned memory by default */
*base = kmalloc(size, flags);
if (*base == NULL) {
return NULL;
}
if ((u64) *base == L1_CACHE_ALIGN((u64) *base)) {
return *base;
}
kfree(*base);
/* nope, we'll have to do it ourselves */
*base = kmalloc(size + L1_CACHE_BYTES, flags);
if (*base == NULL) {
return NULL;
}
return (void *) L1_CACHE_ALIGN((u64) *base);
}
/*
* Check to see if there is any channel activity to/from the specified
* partition.
*/
static inline void
xpc_check_for_channel_activity(struct xpc_partition *part)
{
u64 IPI_amo;
unsigned long irq_flags;
IPI_amo = xpc_IPI_receive(part->local_IPI_amo_va);
if (IPI_amo == 0) {
return;
}
spin_lock_irqsave(&part->IPI_lock, irq_flags);
part->local_IPI_amo |= IPI_amo;
spin_unlock_irqrestore(&part->IPI_lock, irq_flags);
dev_dbg(xpc_chan, "received IPI from partid=%d, IPI_amo=0x%lx\n",
XPC_PARTID(part), IPI_amo);
xpc_wakeup_channel_mgr(part);
}
#endif /* _IA64_SN_KERNEL_XPC_H */

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

View File

@ -0,0 +1,971 @@
/*
* This file is subject to the terms and conditions of the GNU General Public
* License. See the file "COPYING" in the main directory of this archive
* for more details.
*
* Copyright (c) 2004-2005 Silicon Graphics, Inc. All Rights Reserved.
*/
/*
* Cross Partition Communication (XPC) partition support.
*
* This is the part of XPC that detects the presence/absence of
* other partitions. It provides a heartbeat and monitors the
* heartbeats of other partitions.
*
*/
#include <linux/kernel.h>
#include <linux/sysctl.h>
#include <linux/cache.h>
#include <linux/mmzone.h>
#include <linux/nodemask.h>
#include <asm/sn/bte.h>
#include <asm/sn/intr.h>
#include <asm/sn/sn_sal.h>
#include <asm/sn/nodepda.h>
#include <asm/sn/addrs.h>
#include "xpc.h"
/* XPC is exiting flag */
int xpc_exiting;
/* SH_IPI_ACCESS shub register value on startup */
static u64 xpc_sh1_IPI_access;
static u64 xpc_sh2_IPI_access0;
static u64 xpc_sh2_IPI_access1;
static u64 xpc_sh2_IPI_access2;
static u64 xpc_sh2_IPI_access3;
/* original protection values for each node */
u64 xpc_prot_vec[MAX_COMPACT_NODES];
/* this partition's reserved page */
struct xpc_rsvd_page *xpc_rsvd_page;
/* this partition's XPC variables (within the reserved page) */
struct xpc_vars *xpc_vars;
struct xpc_vars_part *xpc_vars_part;
/*
* For performance reasons, each entry of xpc_partitions[] is cacheline
* aligned. And xpc_partitions[] is padded with an additional entry at the
* end so that the last legitimate entry doesn't share its cacheline with
* another variable.
*/
struct xpc_partition xpc_partitions[XP_MAX_PARTITIONS + 1];
/*
* Generic buffer used to store a local copy of the remote partitions
* reserved page or XPC variables.
*
* xpc_discovery runs only once and is a seperate thread that is
* very likely going to be processing in parallel with receiving
* interrupts.
*/
char ____cacheline_aligned
xpc_remote_copy_buffer[XPC_RSVD_PAGE_ALIGNED_SIZE];
/* systune related variables */
int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL;
int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_TIMEOUT;
/*
* Given a nasid, get the physical address of the partition's reserved page
* for that nasid. This function returns 0 on any error.
*/
static u64
xpc_get_rsvd_page_pa(int nasid, u64 buf, u64 buf_size)
{
bte_result_t bte_res;
s64 status;
u64 cookie = 0;
u64 rp_pa = nasid; /* seed with nasid */
u64 len = 0;
while (1) {
status = sn_partition_reserved_page_pa(buf, &cookie, &rp_pa,
&len);
dev_dbg(xpc_part, "SAL returned with status=%li, cookie="
"0x%016lx, address=0x%016lx, len=0x%016lx\n",
status, cookie, rp_pa, len);
if (status != SALRET_MORE_PASSES) {
break;
}
if (len > buf_size) {
dev_err(xpc_part, "len (=0x%016lx) > buf_size\n", len);
status = SALRET_ERROR;
break;
}
bte_res = xp_bte_copy(rp_pa, ia64_tpa(buf), buf_size,
(BTE_NOTIFY | BTE_WACQUIRE), NULL);
if (bte_res != BTE_SUCCESS) {
dev_dbg(xpc_part, "xp_bte_copy failed %i\n", bte_res);
status = SALRET_ERROR;
break;
}
}
if (status != SALRET_OK) {
rp_pa = 0;
}
dev_dbg(xpc_part, "reserved page at phys address 0x%016lx\n", rp_pa);
return rp_pa;
}
/*
* Fill the partition reserved page with the information needed by
* other partitions to discover we are alive and establish initial
* communications.
*/
struct xpc_rsvd_page *
xpc_rsvd_page_init(void)
{
struct xpc_rsvd_page *rp;
AMO_t *amos_page;
u64 rp_pa, next_cl, nasid_array = 0;
int i, ret;
/* get the local reserved page's address */
rp_pa = xpc_get_rsvd_page_pa(cnodeid_to_nasid(0),
(u64) xpc_remote_copy_buffer,
XPC_RSVD_PAGE_ALIGNED_SIZE);
if (rp_pa == 0) {
dev_err(xpc_part, "SAL failed to locate the reserved page\n");
return NULL;
}
rp = (struct xpc_rsvd_page *) __va(rp_pa);
if (rp->partid != sn_partition_id) {
dev_err(xpc_part, "the reserved page's partid of %d should be "
"%d\n", rp->partid, sn_partition_id);
return NULL;
}
rp->version = XPC_RP_VERSION;
/*
* Place the XPC variables on the cache line following the
* reserved page structure.
*/
next_cl = (u64) rp + XPC_RSVD_PAGE_ALIGNED_SIZE;
xpc_vars = (struct xpc_vars *) next_cl;
/*
* Before clearing xpc_vars, see if a page of AMOs had been previously
* allocated. If not we'll need to allocate one and set permissions
* so that cross-partition AMOs are allowed.
*
* The allocated AMO page needs MCA reporting to remain disabled after
* XPC has unloaded. To make this work, we keep a copy of the pointer
* to this page (i.e., amos_page) in the struct xpc_vars structure,
* which is pointed to by the reserved page, and re-use that saved copy
* on subsequent loads of XPC. This AMO page is never freed, and its
* memory protections are never restricted.
*/
if ((amos_page = xpc_vars->amos_page) == NULL) {
amos_page = (AMO_t *) mspec_kalloc_page(0);
if (amos_page == NULL) {
dev_err(xpc_part, "can't allocate page of AMOs\n");
return NULL;
}
/*
* Open up AMO-R/W to cpu. This is done for Shub 1.1 systems
* when xpc_allow_IPI_ops() is called via xpc_hb_init().
*/
if (!enable_shub_wars_1_1()) {
ret = sn_change_memprotect(ia64_tpa((u64) amos_page),
PAGE_SIZE, SN_MEMPROT_ACCESS_CLASS_1,
&nasid_array);
if (ret != 0) {
dev_err(xpc_part, "can't change memory "
"protections\n");
mspec_kfree_page((unsigned long) amos_page);
return NULL;
}
}
}
memset(xpc_vars, 0, sizeof(struct xpc_vars));
/*
* Place the XPC per partition specific variables on the cache line
* following the XPC variables structure.
*/
next_cl += XPC_VARS_ALIGNED_SIZE;
memset((u64 *) next_cl, 0, sizeof(struct xpc_vars_part) *
XP_MAX_PARTITIONS);
xpc_vars_part = (struct xpc_vars_part *) next_cl;
xpc_vars->vars_part_pa = __pa(next_cl);
xpc_vars->version = XPC_V_VERSION;
xpc_vars->act_nasid = cpuid_to_nasid(0);
xpc_vars->act_phys_cpuid = cpu_physical_id(0);
xpc_vars->amos_page = amos_page; /* save for next load of XPC */
/*
* Initialize the activation related AMO variables.
*/
xpc_vars->act_amos = xpc_IPI_init(XP_MAX_PARTITIONS);
for (i = 1; i < XP_NASID_MASK_WORDS; i++) {
xpc_IPI_init(i + XP_MAX_PARTITIONS);
}
/* export AMO page's physical address to other partitions */
xpc_vars->amos_page_pa = ia64_tpa((u64) xpc_vars->amos_page);
/*
* This signifies to the remote partition that our reserved
* page is initialized.
*/
(volatile u64) rp->vars_pa = __pa(xpc_vars);
return rp;
}
/*
* Change protections to allow IPI operations (and AMO operations on
* Shub 1.1 systems).
*/
void
xpc_allow_IPI_ops(void)
{
int node;
int nasid;
// >>> Change SH_IPI_ACCESS code to use SAL call once it is available.
if (is_shub2()) {
xpc_sh2_IPI_access0 =
(u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH2_IPI_ACCESS0));
xpc_sh2_IPI_access1 =
(u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH2_IPI_ACCESS1));
xpc_sh2_IPI_access2 =
(u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH2_IPI_ACCESS2));
xpc_sh2_IPI_access3 =
(u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH2_IPI_ACCESS3));
for_each_online_node(node) {
nasid = cnodeid_to_nasid(node);
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS0),
-1UL);
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS1),
-1UL);
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS2),
-1UL);
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS3),
-1UL);
}
} else {
xpc_sh1_IPI_access =
(u64) HUB_L((u64 *) LOCAL_MMR_ADDR(SH1_IPI_ACCESS));
for_each_online_node(node) {
nasid = cnodeid_to_nasid(node);
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH1_IPI_ACCESS),
-1UL);
/*
* Since the BIST collides with memory operations on
* SHUB 1.1 sn_change_memprotect() cannot be used.
*/
if (enable_shub_wars_1_1()) {
/* open up everything */
xpc_prot_vec[node] = (u64) HUB_L((u64 *)
GLOBAL_MMR_ADDR(nasid,
SH1_MD_DQLP_MMR_DIR_PRIVEC0));
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid,
SH1_MD_DQLP_MMR_DIR_PRIVEC0),
-1UL);
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid,
SH1_MD_DQRP_MMR_DIR_PRIVEC0),
-1UL);
}
}
}
}
/*
* Restrict protections to disallow IPI operations (and AMO operations on
* Shub 1.1 systems).
*/
void
xpc_restrict_IPI_ops(void)
{
int node;
int nasid;
// >>> Change SH_IPI_ACCESS code to use SAL call once it is available.
if (is_shub2()) {
for_each_online_node(node) {
nasid = cnodeid_to_nasid(node);
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS0),
xpc_sh2_IPI_access0);
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS1),
xpc_sh2_IPI_access1);
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS2),
xpc_sh2_IPI_access2);
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH2_IPI_ACCESS3),
xpc_sh2_IPI_access3);
}
} else {
for_each_online_node(node) {
nasid = cnodeid_to_nasid(node);
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid, SH1_IPI_ACCESS),
xpc_sh1_IPI_access);
if (enable_shub_wars_1_1()) {
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid,
SH1_MD_DQLP_MMR_DIR_PRIVEC0),
xpc_prot_vec[node]);
HUB_S((u64 *) GLOBAL_MMR_ADDR(nasid,
SH1_MD_DQRP_MMR_DIR_PRIVEC0),
xpc_prot_vec[node]);
}
}
}
}
/*
* At periodic intervals, scan through all active partitions and ensure
* their heartbeat is still active. If not, the partition is deactivated.
*/
void
xpc_check_remote_hb(void)
{
struct xpc_vars *remote_vars;
struct xpc_partition *part;
partid_t partid;
bte_result_t bres;
remote_vars = (struct xpc_vars *) xpc_remote_copy_buffer;
for (partid = 1; partid < XP_MAX_PARTITIONS; partid++) {
if (partid == sn_partition_id) {
continue;
}
part = &xpc_partitions[partid];
if (part->act_state == XPC_P_INACTIVE ||
part->act_state == XPC_P_DEACTIVATING) {
continue;
}
/* pull the remote_hb cache line */
bres = xp_bte_copy(part->remote_vars_pa,
ia64_tpa((u64) remote_vars),
XPC_VARS_ALIGNED_SIZE,
(BTE_NOTIFY | BTE_WACQUIRE), NULL);
if (bres != BTE_SUCCESS) {
XPC_DEACTIVATE_PARTITION(part,
xpc_map_bte_errors(bres));
continue;
}
dev_dbg(xpc_part, "partid = %d, heartbeat = %ld, last_heartbeat"
" = %ld, kdb_status = %ld, HB_mask = 0x%lx\n", partid,
remote_vars->heartbeat, part->last_heartbeat,
remote_vars->kdb_status,
remote_vars->heartbeating_to_mask);
if (((remote_vars->heartbeat == part->last_heartbeat) &&
(remote_vars->kdb_status == 0)) ||
!XPC_HB_ALLOWED(sn_partition_id, remote_vars)) {
XPC_DEACTIVATE_PARTITION(part, xpcNoHeartbeat);
continue;
}
part->last_heartbeat = remote_vars->heartbeat;
}
}
/*
* Get a copy of the remote partition's rsvd page.
*
* remote_rp points to a buffer that is cacheline aligned for BTE copies and
* assumed to be of size XPC_RSVD_PAGE_ALIGNED_SIZE.
*/
static enum xpc_retval
xpc_get_remote_rp(int nasid, u64 *discovered_nasids,
struct xpc_rsvd_page *remote_rp, u64 *remote_rsvd_page_pa)
{
int bres, i;
/* get the reserved page's physical address */
*remote_rsvd_page_pa = xpc_get_rsvd_page_pa(nasid, (u64) remote_rp,
XPC_RSVD_PAGE_ALIGNED_SIZE);
if (*remote_rsvd_page_pa == 0) {
return xpcNoRsvdPageAddr;
}
/* pull over the reserved page structure */
bres = xp_bte_copy(*remote_rsvd_page_pa, ia64_tpa((u64) remote_rp),
XPC_RSVD_PAGE_ALIGNED_SIZE,
(BTE_NOTIFY | BTE_WACQUIRE), NULL);
if (bres != BTE_SUCCESS) {
return xpc_map_bte_errors(bres);
}
if (discovered_nasids != NULL) {
for (i = 0; i < XP_NASID_MASK_WORDS; i++) {
discovered_nasids[i] |= remote_rp->part_nasids[i];
}
}
/* check that the partid is for another partition */
if (remote_rp->partid < 1 ||
remote_rp->partid > (XP_MAX_PARTITIONS - 1)) {
return xpcInvalidPartid;
}
if (remote_rp->partid == sn_partition_id) {
return xpcLocalPartid;
}
if (XPC_VERSION_MAJOR(remote_rp->version) !=
XPC_VERSION_MAJOR(XPC_RP_VERSION)) {
return xpcBadVersion;
}
return xpcSuccess;
}
/*
* Get a copy of the remote partition's XPC variables.
*
* remote_vars points to a buffer that is cacheline aligned for BTE copies and
* assumed to be of size XPC_VARS_ALIGNED_SIZE.
*/
static enum xpc_retval
xpc_get_remote_vars(u64 remote_vars_pa, struct xpc_vars *remote_vars)
{
int bres;
if (remote_vars_pa == 0) {
return xpcVarsNotSet;
}
/* pull over the cross partition variables */
bres = xp_bte_copy(remote_vars_pa, ia64_tpa((u64) remote_vars),
XPC_VARS_ALIGNED_SIZE,
(BTE_NOTIFY | BTE_WACQUIRE), NULL);
if (bres != BTE_SUCCESS) {
return xpc_map_bte_errors(bres);
}
if (XPC_VERSION_MAJOR(remote_vars->version) !=
XPC_VERSION_MAJOR(XPC_V_VERSION)) {
return xpcBadVersion;
}
return xpcSuccess;
}
/*
* Prior code has determine the nasid which generated an IPI. Inspect
* that nasid to determine if its partition needs to be activated or
* deactivated.
*
* A partition is consider "awaiting activation" if our partition
* flags indicate it is not active and it has a heartbeat. A
* partition is considered "awaiting deactivation" if our partition
* flags indicate it is active but it has no heartbeat or it is not
* sending its heartbeat to us.
*
* To determine the heartbeat, the remote nasid must have a properly
* initialized reserved page.
*/
static void
xpc_identify_act_IRQ_req(int nasid)
{
struct xpc_rsvd_page *remote_rp;
struct xpc_vars *remote_vars;
u64 remote_rsvd_page_pa;
u64 remote_vars_pa;
partid_t partid;
struct xpc_partition *part;
enum xpc_retval ret;
/* pull over the reserved page structure */
remote_rp = (struct xpc_rsvd_page *) xpc_remote_copy_buffer;
ret = xpc_get_remote_rp(nasid, NULL, remote_rp, &remote_rsvd_page_pa);
if (ret != xpcSuccess) {
dev_warn(xpc_part, "unable to get reserved page from nasid %d, "
"which sent interrupt, reason=%d\n", nasid, ret);
return;
}
remote_vars_pa = remote_rp->vars_pa;
partid = remote_rp->partid;
part = &xpc_partitions[partid];
/* pull over the cross partition variables */
remote_vars = (struct xpc_vars *) xpc_remote_copy_buffer;
ret = xpc_get_remote_vars(remote_vars_pa, remote_vars);
if (ret != xpcSuccess) {
dev_warn(xpc_part, "unable to get XPC variables from nasid %d, "
"which sent interrupt, reason=%d\n", nasid, ret);
XPC_DEACTIVATE_PARTITION(part, ret);
return;
}
part->act_IRQ_rcvd++;
dev_dbg(xpc_part, "partid for nasid %d is %d; IRQs = %d; HB = "
"%ld:0x%lx\n", (int) nasid, (int) partid, part->act_IRQ_rcvd,
remote_vars->heartbeat, remote_vars->heartbeating_to_mask);
if (part->act_state == XPC_P_INACTIVE) {
part->remote_rp_pa = remote_rsvd_page_pa;
dev_dbg(xpc_part, " remote_rp_pa = 0x%016lx\n",
part->remote_rp_pa);
part->remote_vars_pa = remote_vars_pa;
dev_dbg(xpc_part, " remote_vars_pa = 0x%016lx\n",
part->remote_vars_pa);
part->last_heartbeat = remote_vars->heartbeat;
dev_dbg(xpc_part, " last_heartbeat = 0x%016lx\n",
part->last_heartbeat);
part->remote_vars_part_pa = remote_vars->vars_part_pa;
dev_dbg(xpc_part, " remote_vars_part_pa = 0x%016lx\n",
part->remote_vars_part_pa);
part->remote_act_nasid = remote_vars->act_nasid;
dev_dbg(xpc_part, " remote_act_nasid = 0x%x\n",
part->remote_act_nasid);
part->remote_act_phys_cpuid = remote_vars->act_phys_cpuid;
dev_dbg(xpc_part, " remote_act_phys_cpuid = 0x%x\n",
part->remote_act_phys_cpuid);
part->remote_amos_page_pa = remote_vars->amos_page_pa;
dev_dbg(xpc_part, " remote_amos_page_pa = 0x%lx\n",
part->remote_amos_page_pa);
xpc_activate_partition(part);
} else if (part->remote_amos_page_pa != remote_vars->amos_page_pa ||
!XPC_HB_ALLOWED(sn_partition_id, remote_vars)) {
part->reactivate_nasid = nasid;
XPC_DEACTIVATE_PARTITION(part, xpcReactivating);
}
}
/*
* Loop through the activation AMO variables and process any bits
* which are set. Each bit indicates a nasid sending a partition
* activation or deactivation request.
*
* Return #of IRQs detected.
*/
int
xpc_identify_act_IRQ_sender(void)
{
int word, bit;
u64 nasid_mask;
u64 nasid; /* remote nasid */
int n_IRQs_detected = 0;
AMO_t *act_amos;
struct xpc_rsvd_page *rp = (struct xpc_rsvd_page *) xpc_rsvd_page;
act_amos = xpc_vars->act_amos;
/* scan through act AMO variable looking for non-zero entries */
for (word = 0; word < XP_NASID_MASK_WORDS; word++) {
nasid_mask = xpc_IPI_receive(&act_amos[word]);
if (nasid_mask == 0) {
/* no IRQs from nasids in this variable */
continue;
}
dev_dbg(xpc_part, "AMO[%d] gave back 0x%lx\n", word,
nasid_mask);
/*
* If this nasid has been added to the machine since
* our partition was reset, this will retain the
* remote nasid in our reserved pages machine mask.
* This is used in the event of module reload.
*/
rp->mach_nasids[word] |= nasid_mask;
/* locate the nasid(s) which sent interrupts */
for (bit = 0; bit < (8 * sizeof(u64)); bit++) {
if (nasid_mask & (1UL << bit)) {
n_IRQs_detected++;
nasid = XPC_NASID_FROM_W_B(word, bit);
dev_dbg(xpc_part, "interrupt from nasid %ld\n",
nasid);
xpc_identify_act_IRQ_req(nasid);
}
}
}
return n_IRQs_detected;
}
/*
* Mark specified partition as active.
*/
enum xpc_retval
xpc_mark_partition_active(struct xpc_partition *part)
{
unsigned long irq_flags;
enum xpc_retval ret;
dev_dbg(xpc_part, "setting partition %d to ACTIVE\n", XPC_PARTID(part));
spin_lock_irqsave(&part->act_lock, irq_flags);
if (part->act_state == XPC_P_ACTIVATING) {
part->act_state = XPC_P_ACTIVE;
ret = xpcSuccess;
} else {
DBUG_ON(part->reason == xpcSuccess);
ret = part->reason;
}
spin_unlock_irqrestore(&part->act_lock, irq_flags);
return ret;
}
/*
* Notify XPC that the partition is down.
*/
void
xpc_deactivate_partition(const int line, struct xpc_partition *part,
enum xpc_retval reason)
{
unsigned long irq_flags;
partid_t partid = XPC_PARTID(part);
spin_lock_irqsave(&part->act_lock, irq_flags);
if (part->act_state == XPC_P_INACTIVE) {
XPC_SET_REASON(part, reason, line);
spin_unlock_irqrestore(&part->act_lock, irq_flags);
if (reason == xpcReactivating) {
/* we interrupt ourselves to reactivate partition */
xpc_IPI_send_reactivate(part);
}
return;
}
if (part->act_state == XPC_P_DEACTIVATING) {
if ((part->reason == xpcUnloading && reason != xpcUnloading) ||
reason == xpcReactivating) {
XPC_SET_REASON(part, reason, line);
}
spin_unlock_irqrestore(&part->act_lock, irq_flags);
return;
}
part->act_state = XPC_P_DEACTIVATING;
XPC_SET_REASON(part, reason, line);
spin_unlock_irqrestore(&part->act_lock, irq_flags);
XPC_DISALLOW_HB(partid, xpc_vars);
dev_dbg(xpc_part, "bringing partition %d down, reason = %d\n", partid,
reason);
xpc_partition_down(part, reason);
}
/*
* Mark specified partition as active.
*/
void
xpc_mark_partition_inactive(struct xpc_partition *part)
{
unsigned long irq_flags;
dev_dbg(xpc_part, "setting partition %d to INACTIVE\n",
XPC_PARTID(part));
spin_lock_irqsave(&part->act_lock, irq_flags);
part->act_state = XPC_P_INACTIVE;
spin_unlock_irqrestore(&part->act_lock, irq_flags);
part->remote_rp_pa = 0;
}
/*
* SAL has provided a partition and machine mask. The partition mask
* contains a bit for each even nasid in our partition. The machine
* mask contains a bit for each even nasid in the entire machine.
*
* Using those two bit arrays, we can determine which nasids are
* known in the machine. Each should also have a reserved page
* initialized if they are available for partitioning.
*/
void
xpc_discovery(void)
{
void *remote_rp_base;
struct xpc_rsvd_page *remote_rp;
struct xpc_vars *remote_vars;
u64 remote_rsvd_page_pa;
u64 remote_vars_pa;
int region;
int max_regions;
int nasid;
struct xpc_rsvd_page *rp;
partid_t partid;
struct xpc_partition *part;
u64 *discovered_nasids;
enum xpc_retval ret;
remote_rp = xpc_kmalloc_cacheline_aligned(XPC_RSVD_PAGE_ALIGNED_SIZE,
GFP_KERNEL, &remote_rp_base);
if (remote_rp == NULL) {
return;
}
remote_vars = (struct xpc_vars *) remote_rp;
discovered_nasids = kmalloc(sizeof(u64) * XP_NASID_MASK_WORDS,
GFP_KERNEL);
if (discovered_nasids == NULL) {
kfree(remote_rp_base);
return;
}
memset(discovered_nasids, 0, sizeof(u64) * XP_NASID_MASK_WORDS);
rp = (struct xpc_rsvd_page *) xpc_rsvd_page;
/*
* The term 'region' in this context refers to the minimum number of
* nodes that can comprise an access protection grouping. The access
* protection is in regards to memory, IOI and IPI.
*/
//>>> move the next two #defines into either include/asm-ia64/sn/arch.h or
//>>> include/asm-ia64/sn/addrs.h
#define SH1_MAX_REGIONS 64
#define SH2_MAX_REGIONS 256
max_regions = is_shub2() ? SH2_MAX_REGIONS : SH1_MAX_REGIONS;
for (region = 0; region < max_regions; region++) {
if ((volatile int) xpc_exiting) {
break;
}
dev_dbg(xpc_part, "searching region %d\n", region);
for (nasid = (region * sn_region_size * 2);
nasid < ((region + 1) * sn_region_size * 2);
nasid += 2) {
if ((volatile int) xpc_exiting) {
break;
}
dev_dbg(xpc_part, "checking nasid %d\n", nasid);
if (XPC_NASID_IN_ARRAY(nasid, rp->part_nasids)) {
dev_dbg(xpc_part, "PROM indicates Nasid %d is "
"part of the local partition; skipping "
"region\n", nasid);
break;
}
if (!(XPC_NASID_IN_ARRAY(nasid, rp->mach_nasids))) {
dev_dbg(xpc_part, "PROM indicates Nasid %d was "
"not on Numa-Link network at reset\n",
nasid);
continue;
}
if (XPC_NASID_IN_ARRAY(nasid, discovered_nasids)) {
dev_dbg(xpc_part, "Nasid %d is part of a "
"partition which was previously "
"discovered\n", nasid);
continue;
}
/* pull over the reserved page structure */
ret = xpc_get_remote_rp(nasid, discovered_nasids,
remote_rp, &remote_rsvd_page_pa);
if (ret != xpcSuccess) {
dev_dbg(xpc_part, "unable to get reserved page "
"from nasid %d, reason=%d\n", nasid,
ret);
if (ret == xpcLocalPartid) {
break;
}
continue;
}
remote_vars_pa = remote_rp->vars_pa;
partid = remote_rp->partid;
part = &xpc_partitions[partid];
/* pull over the cross partition variables */
ret = xpc_get_remote_vars(remote_vars_pa, remote_vars);
if (ret != xpcSuccess) {
dev_dbg(xpc_part, "unable to get XPC variables "
"from nasid %d, reason=%d\n", nasid,
ret);
XPC_DEACTIVATE_PARTITION(part, ret);
continue;
}
if (part->act_state != XPC_P_INACTIVE) {
dev_dbg(xpc_part, "partition %d on nasid %d is "
"already activating\n", partid, nasid);
break;
}
/*
* Register the remote partition's AMOs with SAL so it
* can handle and cleanup errors within that address
* range should the remote partition go down. We don't
* unregister this range because it is difficult to
* tell when outstanding writes to the remote partition
* are finished and thus when it is thus safe to
* unregister. This should not result in wasted space
* in the SAL xp_addr_region table because we should
* get the same page for remote_act_amos_pa after
* module reloads and system reboots.
*/
if (sn_register_xp_addr_region(
remote_vars->amos_page_pa,
PAGE_SIZE, 1) < 0) {
dev_dbg(xpc_part, "partition %d failed to "
"register xp_addr region 0x%016lx\n",
partid, remote_vars->amos_page_pa);
XPC_SET_REASON(part, xpcPhysAddrRegFailed,
__LINE__);
break;
}
/*
* The remote nasid is valid and available.
* Send an interrupt to that nasid to notify
* it that we are ready to begin activation.
*/
dev_dbg(xpc_part, "sending an interrupt to AMO 0x%lx, "
"nasid %d, phys_cpuid 0x%x\n",
remote_vars->amos_page_pa,
remote_vars->act_nasid,
remote_vars->act_phys_cpuid);
xpc_IPI_send_activate(remote_vars);
}
}
kfree(discovered_nasids);
kfree(remote_rp_base);
}
/*
* Given a partid, get the nasids owned by that partition from the
* remote partitions reserved page.
*/
enum xpc_retval
xpc_initiate_partid_to_nasids(partid_t partid, void *nasid_mask)
{
struct xpc_partition *part;
u64 part_nasid_pa;
int bte_res;
part = &xpc_partitions[partid];
if (part->remote_rp_pa == 0) {
return xpcPartitionDown;
}
part_nasid_pa = part->remote_rp_pa +
(u64) &((struct xpc_rsvd_page *) 0)->part_nasids;
bte_res = xp_bte_copy(part_nasid_pa, ia64_tpa((u64) nasid_mask),
L1_CACHE_ALIGN(XP_NASID_MASK_BYTES),
(BTE_NOTIFY | BTE_WACQUIRE), NULL);
return xpc_map_bte_errors(bte_res);
}