// SPDX-License-Identifier: GPL-2.0-only /* * Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved. * Copyright 2004-2011 Red Hat, Inc.
*/
/** * gfs2_update_stats - Update time based stats * @s: The stats to update (local or global) * @index: The index inside @s * @sample: New data to include
*/ staticinlinevoid gfs2_update_stats(struct gfs2_lkstats *s, unsigned index,
s64 sample)
{ /* * @delta is the difference between the current rtt sample and the * running average srtt. We add 1/8 of that to the srtt in order to * update the current srtt estimate. The variance estimate is a bit * more complicated. We subtract the current variance estimate from * the abs value of the @delta and add 1/4 of that to the running * total. That's equivalent to 3/4 of the current variance * estimate plus 1/4 of the abs of @delta. * * Note that the index points at the array entry containing the * smoothed mean value, and the variance is always in the following * entry * * Reference: TCP/IP Illustrated, vol 2, p. 831,832 * All times are in units of integer nanoseconds. Unlike the TCP/IP * case, they are not scaled fixed point.
*/
/** * gfs2_update_reply_times - Update locking statistics * @gl: The glock to update * @blocking: The operation may have been blocking * * This assumes that gl->gl_dstamp has been set earlier. * * The rtt (lock round trip time) is an estimate of the time * taken to perform a dlm lock request. We update it on each * reply from the dlm. * * The blocking flag is set on the glock for all dlm requests * which may potentially block due to lock requests from other nodes. * DLM requests where the current lock state is exclusive, the * requested state is null (or unlocked) or where the TRY or * TRY_1CB flags are set are classified as non-blocking. All * other DLM requests are counted as (potentially) blocking.
*/ staticinlinevoid gfs2_update_reply_times(struct gfs2_glock *gl, bool blocking)
{ struct gfs2_pcpu_lkstats *lks; constunsigned gltype = gl->gl_name.ln_type; unsigned index = blocking ? GFS2_LKS_SRTTB : GFS2_LKS_SRTT;
s64 rtt;
/** * gfs2_update_request_times - Update locking statistics * @gl: The glock to update * * The irt (lock inter-request times) measures the average time * between requests to the dlm. It is updated immediately before * each dlm call.
*/
/* If the glock is dead, we only react to a dlm_unlock() reply. */ if (__lockref_is_dead(&gl->gl_lockref) &&
gl->gl_lksb.sb_status != -DLM_EUNLOCK) return;
BUG_ON(gl->gl_lksb.sb_flags & DLM_SBF_DEMOTED);
if ((gl->gl_lksb.sb_flags & DLM_SBF_VALNOTVALID) && gl->gl_lksb.sb_lvbptr)
memset(gl->gl_lksb.sb_lvbptr, 0, GDLM_LVB_SIZE);
switch (gl->gl_lksb.sb_status) { case -DLM_EUNLOCK: /* Unlocked, so glock can be freed */ if (gl->gl_ops->go_unlocked)
gl->gl_ops->go_unlocked(gl);
gfs2_glock_free(gl); return; case -DLM_ECANCEL: /* Cancel while getting lock */
ret = LM_OUT_CANCELED; goto out; case -EAGAIN: /* Try lock fails */
ret = LM_OUT_TRY_AGAIN; goto out; case -EDEADLK: /* Deadlock detected */
ret = LM_OUT_DEADLOCK; goto out; case -ETIMEDOUT: /* Canceled due to timeout */
ret = LM_OUT_ERROR; goto out; case 0: /* Success */ break; default: /* Something unexpected */
BUG();
}
ret = gl->gl_req; if (gl->gl_lksb.sb_flags & DLM_SBF_ALTMODE) { if (gl->gl_req == LM_ST_SHARED)
ret = LM_ST_DEFERRED; elseif (gl->gl_req == LM_ST_DEFERRED)
ret = LM_ST_SHARED; else
BUG();
}
/* * The GLF_INITIAL flag is initially set for new glocks. Upon the * first successful new (non-conversion) request, we clear this flag to * indicate that a DLM lock exists and that gl->gl_lksb.sb_lkid is the * identifier to use for identifying it. * * Any failed initial requests do not create a DLM lock, so we ignore * the gl->gl_lksb.sb_lkid values that come with such requests.
*/
if (!test_bit(GLF_INITIAL, &gl->gl_flags)) {
lkf |= DLM_LKF_CONVERT;
/* * The DLM_LKF_QUECVT flag needs to be set for "first come, * first served" semantics, but it must only be set for * "upward" lock conversions or else DLM will reject the * request as invalid.
*/ if (blocking)
lkf |= DLM_LKF_QUECVT;
}
/* * When the lockspace is released, all remaining glocks will be * unlocked automatically. This is more efficient than unlocking them * individually, but when the lock is held in DLM_LOCK_EX or * DLM_LOCK_PW mode, the lock value block (LVB) would be lost.
*/
/* * dlm/gfs2 recovery coordination using dlm_recover callbacks * * 0. gfs2 checks for another cluster node withdraw, needing journal replay * 1. dlm_controld sees lockspace members change * 2. dlm_controld blocks dlm-kernel locking activity * 3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep) * 4. dlm_controld starts and finishes its own user level recovery * 5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery * 6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot) * 7. dlm_recoverd does its own lock recovery * 8. dlm_recoverd unblocks dlm-kernel locking activity * 9. dlm_recoverd notifies gfs2 when done (recover_done with new generation) * 10. gfs2_control updates control_lock lvb with new generation and jid bits * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none) * 12. gfs2_recover dequeues and recovers journals of failed nodes * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result) * 14. gfs2_control updates control_lock lvb jid bits for recovered journals * 15. gfs2_control unblocks normal locking when all journals are recovered * * - failures during recovery * * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control * clears BLOCK_LOCKS (step 15), e.g. another node fails while still * recovering for a prior failure. gfs2_control needs a way to detect * this so it can leave BLOCK_LOCKS set in step 15. This is managed using * the recover_block and recover_start values. * * recover_done() provides a new lockspace generation number each time it * is called (step 9). This generation number is saved as recover_start. * When recover_prep() is called, it sets BLOCK_LOCKS and sets * recover_block = recover_start. So, while recover_block is equal to * recover_start, BLOCK_LOCKS should remain set. (recover_spin must * be held around the BLOCK_LOCKS/recover_block/recover_start logic.) * * - more specific gfs2 steps in sequence above * * 3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start * 6. recover_slot records any failed jids (maybe none) * 9. recover_done sets recover_start = new generation number * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids * 12. gfs2_recover does journal recoveries for failed jids identified above * 14. gfs2_control clears control_lock lvb bits for recovered jids * 15. gfs2_control checks if recover_block == recover_start (step 3 occured * again) then do nothing, otherwise if recover_start > recover_block * then clear BLOCK_LOCKS. * * - parallel recovery steps across all nodes * * All nodes attempt to update the control_lock lvb with the new generation * number and jid bits, but only the first to get the control_lock EX will * do so; others will see that it's already done (lvb already contains new * generation number.) * * . All nodes get the same recover_prep/recover_slot/recover_done callbacks * . All nodes attempt to set control_lock lvb gen + bits for the new gen * . One node gets control_lock first and writes the lvb, others see it's done * . All nodes attempt to recover jids for which they see control_lock bits set * . One node succeeds for a jid, and that one clears the jid bit in the lvb * . All nodes will eventually see all lvb bits clear and unblock locks * * - is there a problem with clearing an lvb bit that should be set * and missing a journal recovery? * * 1. jid fails * 2. lvb bit set for step 1 * 3. jid recovered for step 1 * 4. jid taken again (new mount) * 5. jid fails (for step 4) * 6. lvb bit set for step 5 (will already be set) * 7. lvb bit cleared for step 3 * * This is not a problem because the failure in step 5 does not * require recovery, because the mount in step 4 could not have * progressed far enough to unblock locks and access the fs. The * control_mount() function waits for all recoveries to be complete * for the latest lockspace generation before ever unblocking locks * and returning. The mount in step 4 waits until the recovery in * step 1 is done. * * - special case of first mounter: first node to mount the fs * * The first node to mount a gfs2 fs needs to check all the journals * and recover any that need recovery before other nodes are allowed * to mount the fs. (Others may begin mounting, but they must wait * for the first mounter to be done before taking locks on the fs * or accessing the fs.) This has two parts: * * 1. The mounted_lock tells a node it's the first to mount the fs. * Each node holds the mounted_lock in PR while it's mounted. * Each node tries to acquire the mounted_lock in EX when it mounts. * If a node is granted the mounted_lock EX it means there are no * other mounted nodes (no PR locks exist), and it is the first mounter. * The mounted_lock is demoted to PR when first recovery is done, so * others will fail to get an EX lock, but will get a PR lock. * * 2. The control_lock blocks others in control_mount() while the first * mounter is doing first mount recovery of all journals. * A mounting node needs to acquire control_lock in EX mode before * it can proceed. The first mounter holds control_lock in EX while doing * the first mount recovery, blocking mounts from other nodes, then demotes * control_lock to NL when it's done (others_may_mount/first_done), * allowing other nodes to continue mounting. * * first mounter: * control_lock EX/NOQUEUE success * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters) * set first=1 * do first mounter recovery * mounted_lock EX->PR * control_lock EX->NL, write lvb generation * * other mounter: * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry) * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR) * mounted_lock PR/NOQUEUE success * read lvb generation * control_lock EX->NL * set first=0 * * - mount during recovery * * If a node mounts while others are doing recovery (not first mounter), * the mounting node will get its initial recover_done() callback without * having seen any previous failures/callbacks. * * It must wait for all recoveries preceding its mount to be finished * before it unblocks locks. It does this by repeating the "other mounter" * steps above until the lvb generation number is >= its mount generation * number (from initial recover_done) and all lvb bits are clear. * * - control_lock lvb format * * 4 bytes generation number: the latest dlm lockspace generation number * from recover_done callback. Indicates the jid bitmap has been updated * to reflect all slot failures through that generation. * 4 bytes unused. * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates * that jid N needs recovery.
*/
/** * remote_withdraw - react to a node withdrawing from the file system * @sdp: The superblock
*/ staticvoid remote_withdraw(struct gfs2_sbd *sdp)
{ struct gfs2_jdesc *jd; int ret = 0, count = 0;
list_for_each_entry(jd, &sdp->sd_jindex_list, jd_list) { if (jd->jd_jid == sdp->sd_lockstruct.ls_jid) continue;
ret = gfs2_recover_journal(jd, true); if (ret) break;
count++;
}
/* Now drop the additional reference we acquired */
fs_err(sdp, "Journals checked: %d, ret = %d.\n", count, ret);
}
staticvoid gfs2_control_func(struct work_struct *work)
{ struct gfs2_sbd *sdp = container_of(work, struct gfs2_sbd, sd_control_work.work); struct lm_lockstruct *ls = &sdp->sd_lockstruct;
uint32_t block_gen, start_gen, lvb_gen, flags; int recover_set = 0; int write_lvb = 0; int recover_size; int i, error;
/* First check for other nodes that may have done a withdraw. */ if (test_bit(SDF_REMOTE_WITHDRAW, &sdp->sd_flags)) {
remote_withdraw(sdp);
clear_bit(SDF_REMOTE_WITHDRAW, &sdp->sd_flags); return;
}
spin_lock(&ls->ls_recover_spin); /* * No MOUNT_DONE means we're still mounting; control_mount() * will set this flag, after which this thread will take over * all further clearing of BLOCK_LOCKS. * * FIRST_MOUNT means this node is doing first mounter recovery, * for which recovery control is handled by * control_mount()/control_first_done(), not this thread.
*/ if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
spin_unlock(&ls->ls_recover_spin); return;
}
block_gen = ls->ls_recover_block;
start_gen = ls->ls_recover_start;
spin_unlock(&ls->ls_recover_spin);
/* * Equal block_gen and start_gen implies we are between * recover_prep and recover_done callbacks, which means * dlm recovery is in progress and dlm locking is blocked. * There's no point trying to do any work until recover_done.
*/
if (block_gen == start_gen) return;
/* * Propagate recover_submit[] and recover_result[] to lvb: * dlm_recoverd adds to recover_submit[] jids needing recovery * gfs2_recover adds to recover_result[] journal recovery results * * set lvb bit for jids in recover_submit[] if the lvb has not * yet been updated for the generation of the failure * * clear lvb bit for jids in recover_result[] if the result of * the journal recovery is SUCCESS
*/
error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_VALBLK); if (error) {
fs_err(sdp, "control lock EX error %d\n", error); return;
}
if (lvb_gen <= start_gen) { /* * Clear lvb bits for jids we've successfully recovered. * Because all nodes attempt to recover failed journals, * a journal can be recovered multiple times successfully * in succession. Only the first will really do recovery, * the others find it clean, but still report a successful * recovery. So, another node may have already recovered * the jid and cleared the lvb bit for it.
*/ for (i = 0; i < recover_size; i++) { if (ls->ls_recover_result[i] != LM_RD_SUCCESS) continue;
ls->ls_recover_result[i] = 0;
if (!test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET)) continue;
if (lvb_gen == start_gen) { /* * Failed slots before start_gen are already set in lvb.
*/ for (i = 0; i < recover_size; i++) { if (!ls->ls_recover_submit[i]) continue; if (ls->ls_recover_submit[i] < lvb_gen)
ls->ls_recover_submit[i] = 0;
}
} elseif (lvb_gen < start_gen) { /* * Failed slots before start_gen are not yet set in lvb.
*/ for (i = 0; i < recover_size; i++) { if (!ls->ls_recover_submit[i]) continue; if (ls->ls_recover_submit[i] < start_gen) {
ls->ls_recover_submit[i] = 0;
__set_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
}
} /* even if there are no bits to set, we need to write the
latest generation to the lvb */
write_lvb = 1;
} else { /* * we should be getting a recover_done() for lvb_gen soon
*/
}
spin_unlock(&ls->ls_recover_spin);
/* * Everyone will see jid bits set in the lvb, run gfs2_recover_set(), * and clear a jid bit in the lvb if the recovery is a success. * Eventually all journals will be recovered, all jid bits will * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
*/
for (i = 0; i < recover_size; i++) { if (test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET)) {
fs_info(sdp, "recover generation %u jid %d\n",
start_gen, i);
gfs2_recover_set(sdp, i);
recover_set++;
}
} if (recover_set) return;
/* * No more jid bits set in lvb, all recovery is done, unblock locks * (unless a new recover_prep callback has occured blocking locks * again while working above)
*/
/* * We always start with both locks in NL. control_lock is * demoted to NL below so we don't need to do it here.
*/
if (mounted_mode != DLM_LOCK_NL) {
error = mounted_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT); if (error) goto fail;
mounted_mode = DLM_LOCK_NL;
}
/* * Other nodes need to do some work in dlm recovery and gfs2_control * before the recover_done and control_lock will be ready for us below. * A delay here is not required but often avoids having to retry.
*/
msleep_interruptible(500);
/* * Acquire control_lock in EX and mounted_lock in either EX or PR. * control_lock lvb keeps track of any pending journal recoveries. * mounted_lock indicates if any other nodes have the fs mounted.
*/
/** * If we're a spectator, we don't want to take the lock in EX because * we cannot do the first-mount responsibility it implies: recovery.
*/ if (sdp->sd_args.ar_spectator) goto locks_done;
error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE); if (!error) {
mounted_mode = DLM_LOCK_PR; goto locks_done;
} else { /* not even -EAGAIN should happen here */
fs_err(sdp, "control_mount mounted_lock PR error %d\n", error); goto fail;
}
locks_done: /* * If we got both locks above in EX, then we're the first mounter. * If not, then we need to wait for the control_lock lvb to be * updated by other mounted nodes to reflect our mount generation. * * In simple first mounter cases, first mounter will see zero lvb_gen, * but in cases where all existing nodes leave/fail before mounting * nodes finish control_mount, then all nodes will be mounting and * lvb_gen will be non-zero.
*/
control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
if (lvb_gen == 0xFFFFFFFF) { /* special value to force mount attempts to fail */
fs_err(sdp, "control_mount control_lock disabled\n");
error = -EINVAL; goto fail;
}
if (mounted_mode == DLM_LOCK_EX) { /* first mounter, keep both EX while doing first recovery */
spin_lock(&ls->ls_recover_spin);
clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
set_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
spin_unlock(&ls->ls_recover_spin);
fs_info(sdp, "first mounter control generation %u\n", lvb_gen); return 0;
}
error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT); if (error) goto fail;
/* * We are not first mounter, now we need to wait for the control_lock * lvb generation to be >= the generation from our first recover_done * and all lvb bits to be clear (no pending journal recoveries.)
*/
if (!all_jid_bits_clear(ls->ls_lvb_bits)) { /* journals need recovery, wait until all are clear */
fs_info(sdp, "control_mount wait for journal recovery\n"); goto restart;
}
if (start_gen == block_gen) { /* * Wait for the end of a dlm recovery cycle to switch from * first mounter recovery. We can ignore any recover_slot * callbacks between the recover_prep and next recover_done * because we are still the first mounter and any failed nodes * have not fully mounted, so they don't need recovery.
*/
spin_unlock(&ls->ls_recover_spin);
fs_info(sdp, "control_first_done wait gen %u\n", start_gen);
error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK); if (error)
fs_err(sdp, "control_first_done control NL error %d\n", error);
return error;
}
/* * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC) * to accommodate the largest slot number. (NB dlm slot numbers start at 1, * gfs2 jids start at 0, so jid = slot - 1)
*/
#define RECOVER_SIZE_INC 16
staticint set_recover_size(struct gfs2_sbd *sdp, struct dlm_slot *slots, int num_slots)
{ struct lm_lockstruct *ls = &sdp->sd_lockstruct;
uint32_t *submit = NULL;
uint32_t *result = NULL;
uint32_t old_size, new_size; int i, max_jid;
if (!ls->ls_lvb_bits) {
ls->ls_lvb_bits = kzalloc(GDLM_LVB_SIZE, GFP_NOFS); if (!ls->ls_lvb_bits) return -ENOMEM;
}
max_jid = 0; for (i = 0; i < num_slots; i++) { if (max_jid < slots[i].slot - 1)
max_jid = slots[i].slot - 1;
}
if (gfs2_withdrawing_or_withdrawn(sdp)) {
fs_err(sdp, "recover_done ignored due to withdraw.\n"); return;
} /* ensure the ls jid arrays are large enough */
set_recover_size(sdp, slots, num_slots);
if (gfs2_withdrawing_or_withdrawn(sdp)) {
fs_err(sdp, "recovery_result jid %d ignored due to withdraw.\n",
jid); return;
} if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags)) return;
/* don't care about the recovery of own journal during mount */ if (jid == ls->ls_jid) return;
spin_lock(&ls->ls_recover_spin); if (test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
spin_unlock(&ls->ls_recover_spin); return;
} if (ls->ls_recover_size < jid + 1) {
fs_err(sdp, "recovery_result jid %d short size %d\n",
jid, ls->ls_recover_size);
spin_unlock(&ls->ls_recover_spin); return;
}
fs_info(sdp, "recover jid %d result %s\n", jid,
result == LM_RD_GAVEUP ? "busy" : "success");
ls->ls_recover_result[jid] = result;
/* GAVEUP means another node is recovering the journal; delay our next attempt to recover it, to give the other node a chance to
finish before trying again */
if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work,
result == LM_RD_GAVEUP ? HZ : 0);
spin_unlock(&ls->ls_recover_spin);
}
if (ops_result < 0) { /* * dlm does not support ops callbacks, * old dlm_controld/gfs_controld are used, try without ops.
*/
fs_info(sdp, "dlm lockspace ops not used\n");
free_recover_size(ls);
set_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags); return 0;
}
/* mounted_lock and control_lock will be purged in dlm recovery */
release:
down_write(&ls->ls_sem); if (ls->ls_dlm) {
dlm_release_lockspace(ls->ls_dlm, 2);
ls->ls_dlm = NULL;
}
up_write(&ls->ls_sem);
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.
