diff options
Diffstat (limited to 'xlators/cluster/ec/src/ec-common.c')
| -rw-r--r-- | xlators/cluster/ec/src/ec-common.c | 2148 |
1 files changed, 1544 insertions, 604 deletions
diff --git a/xlators/cluster/ec/src/ec-common.c b/xlators/cluster/ec/src/ec-common.c index d0c9f97ab28..b955efd8c2d 100644 --- a/xlators/cluster/ec/src/ec-common.c +++ b/xlators/cluster/ec/src/ec-common.c @@ -8,10 +8,11 @@ cases as published by the Free Software Foundation. */ -#include "byte-order.h" +#include <glusterfs/byte-order.h> +#include <glusterfs/hashfn.h> #include "ec-mem-types.h" -#include "ec-data.h" +#include "ec-types.h" #include "ec-helpers.h" #include "ec-combine.h" #include "ec-common.h" @@ -20,102 +21,332 @@ #include "ec.h" #include "ec-messages.h" -int32_t ec_child_valid(ec_t * ec, ec_fop_data_t * fop, int32_t idx) +#define EC_INVALID_INDEX UINT32_MAX + +void +ec_update_fd_status(fd_t *fd, xlator_t *xl, int idx, int32_t ret_status) { - return (idx < ec->nodes) && (((fop->remaining >> idx) & 1) == 1); + ec_fd_t *fd_ctx; + + if (fd == NULL) + return; + + LOCK(&fd->lock); + { + fd_ctx = __ec_fd_get(fd, xl); + if (fd_ctx) { + if (ret_status >= 0) + fd_ctx->fd_status[idx] = EC_FD_OPENED; + else + fd_ctx->fd_status[idx] = EC_FD_NOT_OPENED; + } + } + UNLOCK(&fd->lock); } -int32_t ec_child_next(ec_t * ec, ec_fop_data_t * fop, int32_t idx) +static uintptr_t +ec_fd_ctx_need_open(fd_t *fd, xlator_t *this, uintptr_t mask) { - while (!ec_child_valid(ec, fop, idx)) + int i = 0; + int count = 0; + ec_t *ec = NULL; + ec_fd_t *fd_ctx = NULL; + uintptr_t need_open = 0; + + ec = this->private; + + fd_ctx = ec_fd_get(fd, this); + if (!fd_ctx) + return count; + + LOCK(&fd->lock); { - if (++idx >= ec->nodes) - { + for (i = 0; i < ec->nodes; i++) { + if ((fd_ctx->fd_status[i] == EC_FD_NOT_OPENED) && + ((ec->xl_up & (1 << i)) != 0) && ((mask & (1 << i)) != 0)) { + fd_ctx->fd_status[i] = EC_FD_OPENING; + need_open |= (1 << i); + count++; + } + } + } + UNLOCK(&fd->lock); + + /* If fd needs to open on minimum number of nodes + * then ignore fixing the fd as it has been + * requested from heal operation. + */ + if (count >= ec->fragments) { + need_open = 0; + } + + return need_open; +} + +static gf_boolean_t +ec_is_fd_fixable(fd_t *fd) +{ + if (!fd || !fd->inode) + return _gf_false; + else if (fd_is_anonymous(fd)) + return _gf_false; + else if (gf_uuid_is_null(fd->inode->gfid)) + return _gf_false; + + return _gf_true; +} + +static void +ec_fix_open(ec_fop_data_t *fop, uintptr_t mask) +{ + uintptr_t need_open = 0; + int ret = 0; + int32_t flags = 0; + loc_t loc = { + 0, + }; + + if (!ec_is_fd_fixable(fop->fd)) + goto out; + + /* Evaluate how many remote fd's to be opened */ + need_open = ec_fd_ctx_need_open(fop->fd, fop->xl, mask); + if (need_open == 0) { + goto out; + } + + loc.inode = inode_ref(fop->fd->inode); + gf_uuid_copy(loc.gfid, fop->fd->inode->gfid); + ret = loc_path(&loc, NULL); + if (ret < 0) { + goto out; + } + + flags = fop->fd->flags & (~(O_TRUNC | O_APPEND | O_CREAT | O_EXCL)); + if (IA_IFDIR == fop->fd->inode->ia_type) { + ec_opendir(fop->frame, fop->xl, need_open, + EC_MINIMUM_ONE | EC_FOP_NO_PROPAGATE_ERROR, NULL, NULL, + &fop->loc[0], fop->fd, NULL); + } else { + ec_open(fop->frame, fop->xl, need_open, + EC_MINIMUM_ONE | EC_FOP_NO_PROPAGATE_ERROR, NULL, NULL, &loc, + flags, fop->fd, NULL); + } + +out: + loc_wipe(&loc); +} + +static off_t +ec_range_end_get(off_t fl_start, uint64_t fl_size) +{ + if (fl_size > 0) { + if (fl_size >= EC_RANGE_FULL) { + /* Infinity */ + fl_start = LLONG_MAX; + } else { + fl_start += fl_size - 1; + if (fl_start < 0) { + /* Overflow */ + fl_start = LLONG_MAX; + } + } + } + + return fl_start; +} + +static gf_boolean_t +ec_is_range_conflict(ec_lock_link_t *l1, ec_lock_link_t *l2) +{ + return ((l1->fl_end >= l2->fl_start) && (l2->fl_end >= l1->fl_start)); +} + +static gf_boolean_t +ec_lock_conflict(ec_lock_link_t *l1, ec_lock_link_t *l2) +{ + ec_t *ec = l1->fop->xl->private; + + /* Fops like access/stat won't have to worry what the other fops are + * modifying as the fop is wound only to one brick. So it can be + * executed in parallel*/ + if (l1->fop->minimum == EC_MINIMUM_ONE || + l2->fop->minimum == EC_MINIMUM_ONE) + return _gf_false; + + if ((l1->fop->flags & EC_FLAG_LOCK_SHARED) && + (l2->fop->flags & EC_FLAG_LOCK_SHARED)) + return _gf_false; + + if (!ec->parallel_writes) { + return _gf_true; + } + + return ec_is_range_conflict(l1, l2); +} + +uint32_t +ec_select_first_by_read_policy(ec_t *ec, ec_fop_data_t *fop) +{ + if (ec->read_policy == EC_ROUND_ROBIN) { + return ec->idx; + } else if (ec->read_policy == EC_GFID_HASH) { + if (fop->use_fd) { + return SuperFastHash((char *)fop->fd->inode->gfid, + sizeof(fop->fd->inode->gfid)) % + ec->nodes; + } else { + if (gf_uuid_is_null(fop->loc[0].gfid)) + loc_gfid(&fop->loc[0], fop->loc[0].gfid); + return SuperFastHash((char *)fop->loc[0].gfid, + sizeof(fop->loc[0].gfid)) % + ec->nodes; + } + } + return 0; +} + +static gf_boolean_t +ec_child_valid(ec_t *ec, ec_fop_data_t *fop, uint32_t idx) +{ + return (idx < ec->nodes) && (((fop->remaining >> idx) & 1) == 1); +} + +static uint32_t +ec_child_next(ec_t *ec, ec_fop_data_t *fop, uint32_t idx) +{ + while (!ec_child_valid(ec, fop, idx)) { + if (++idx >= ec->nodes) { idx = 0; } - if (idx == fop->first) - { - return -1; + if (idx == fop->first) { + return EC_INVALID_INDEX; } } return idx; } -int32_t ec_heal_report(call_frame_t * frame, void * cookie, xlator_t * this, - int32_t op_ret, int32_t op_errno, uintptr_t mask, - uintptr_t good, uintptr_t bad, dict_t * xdata) +int32_t +ec_heal_report(call_frame_t *frame, void *cookie, xlator_t *this, + int32_t op_ret, int32_t op_errno, uintptr_t mask, uintptr_t good, + uintptr_t bad, uint32_t pending, dict_t *xdata) { if (op_ret < 0) { - gf_msg (this->name, GF_LOG_WARNING, op_errno, - EC_MSG_HEAL_FAIL, "Heal failed"); + gf_msg(this->name, GF_LOG_DEBUG, op_errno, EC_MSG_HEAL_FAIL, + "Heal failed"); } else { if ((mask & ~good) != 0) { - gf_msg (this->name, GF_LOG_INFO, 0, - EC_MSG_HEAL_SUCCESS, "Heal succeeded on %d/%d " - "subvolumes", - ec_bits_count(mask & ~(good | bad)), - ec_bits_count(mask & ~good)); + gf_msg(this->name, GF_LOG_DEBUG, 0, EC_MSG_HEAL_SUCCESS, + "Heal succeeded on %d/%d " + "subvolumes", + gf_bits_count(mask & ~(good | bad)), + gf_bits_count(mask & ~good)); } } return 0; } -int32_t ec_fop_needs_heal(ec_fop_data_t *fop) +static uintptr_t +ec_fop_needs_name_heal(ec_fop_data_t *fop) +{ + ec_t *ec = NULL; + ec_cbk_data_t *cbk = NULL; + ec_cbk_data_t *enoent_cbk = NULL; + + ec = fop->xl->private; + if (fop->id != GF_FOP_LOOKUP) + return 0; + + if (!fop->loc[0].name || strlen(fop->loc[0].name) == 0) + return 0; + + list_for_each_entry(cbk, &fop->cbk_list, list) + { + if (cbk->op_ret < 0 && cbk->op_errno == ENOENT) { + enoent_cbk = cbk; + break; + } + } + + if (!enoent_cbk) + return 0; + + return ec->xl_up & ~enoent_cbk->mask; +} + +int32_t +ec_fop_needs_heal(ec_fop_data_t *fop) { ec_t *ec = fop->xl->private; + if (fop->lock_count == 0) { + /* + * if fop->lock_count is zero that means it saw version mismatch + * without any locks so it can't be trusted. If we launch a heal + * based on this it will lead to INODELKs which will affect I/O + * performance. Considering self-heal-daemon and operations on + * the inode from client which take locks can still trigger the + * heal we can choose to not attempt a heal when fop->lock_count + * is zero. + */ + return 0; + } return (ec->xl_up & ~(fop->remaining | fop->good)) != 0; } -void ec_check_status(ec_fop_data_t * fop) +void +ec_check_status(ec_fop_data_t *fop) { - ec_t * ec = fop->xl->private; + ec_t *ec = fop->xl->private; int32_t partial = 0; + char str1[32], str2[32], str3[32], str4[32], str5[32]; - if (!ec_fop_needs_heal(fop)) { + if (!ec_fop_needs_name_heal(fop) && !ec_fop_needs_heal(fop)) { return; } - if (fop->answer->op_ret >= 0) { - if ((fop->id == GF_FOP_LOOKUP) || - (fop->id == GF_FOP_STAT) || (fop->id == GF_FOP_FSTAT)) { + if (fop->answer && fop->answer->op_ret >= 0) { + if ((fop->id == GF_FOP_LOOKUP) || (fop->id == GF_FOP_STAT) || + (fop->id == GF_FOP_FSTAT)) { partial = fop->answer->iatt[0].ia_type == IA_IFDIR; } else if (fop->id == GF_FOP_OPENDIR) { partial = 1; } } - gf_msg (fop->xl->name, GF_LOG_WARNING, 0, - EC_MSG_OP_FAIL_ON_SUBVOLS, - "Operation failed on some " - "subvolumes (up=%lX, mask=%lX, " - "remaining=%lX, good=%lX, bad=%lX)", - ec->xl_up, fop->mask, fop->remaining, fop->good, - ec->xl_up & ~(fop->remaining | fop->good)); - - if (fop->use_fd) - { + gf_msg( + fop->xl->name, GF_LOG_WARNING, 0, EC_MSG_OP_FAIL_ON_SUBVOLS, + "Operation failed on %d of %d subvolumes.(up=%s, mask=%s, " + "remaining=%s, good=%s, bad=%s," + "(Least significant bit represents first client/brick of subvol), %s)", + gf_bits_count(ec->xl_up & ~(fop->remaining | fop->good)), ec->nodes, + ec_bin(str1, sizeof(str1), ec->xl_up, ec->nodes), + ec_bin(str2, sizeof(str2), fop->mask, ec->nodes), + ec_bin(str3, sizeof(str3), fop->remaining, ec->nodes), + ec_bin(str4, sizeof(str4), fop->good, ec->nodes), + ec_bin(str5, sizeof(str5), ec->xl_up & ~(fop->remaining | fop->good), + ec->nodes), + ec_msg_str(fop)); + if (fop->use_fd) { if (fop->fd != NULL) { ec_fheal(NULL, fop->xl, -1, EC_MINIMUM_ONE, ec_heal_report, NULL, fop->fd, partial, NULL); } - } - else - { + } else { ec_heal(NULL, fop->xl, -1, EC_MINIMUM_ONE, ec_heal_report, NULL, &fop->loc[0], partial, NULL); - if (fop->loc[1].inode != NULL) - { + if (fop->loc[1].inode != NULL) { ec_heal(NULL, fop->xl, -1, EC_MINIMUM_ONE, ec_heal_report, NULL, &fop->loc[1], partial, NULL); } } } -void ec_update_good(ec_fop_data_t *fop, uintptr_t good) +void +ec_update_good(ec_fop_data_t *fop, uintptr_t good) { fop->good = good; @@ -126,7 +357,8 @@ void ec_update_good(ec_fop_data_t *fop, uintptr_t good) } } -void ec_lock_update_good(ec_lock_t *lock, ec_fop_data_t *fop) +void +ec_lock_update_good(ec_lock_t *lock, ec_fop_data_t *fop) { /* Fops that are executed only on one brick do not have enough information * to update the global mask of good bricks. */ @@ -134,22 +366,24 @@ void ec_lock_update_good(ec_lock_t *lock, ec_fop_data_t *fop) return; } - /* When updating the good mask of the lock, we only take into - * consideration those bits corresponding to the bricks where - * the fop has been executed. */ - lock->good_mask &= ~fop->mask | fop->remaining; - lock->good_mask |= fop->good; + /* When updating the good mask of the lock, we only take into consideration + * those bits corresponding to the bricks where the fop has been executed. + * Bad bricks are removed from good_mask, but once marked as bad it's never + * set to good until the lock is released and reacquired */ + + lock->good_mask &= fop->good | fop->remaining; } -void __ec_fop_set_error(ec_fop_data_t * fop, int32_t error) +void +__ec_fop_set_error(ec_fop_data_t *fop, int32_t error) { - if ((error != 0) && (fop->error == 0)) - { + if ((error != 0) && (fop->error == 0)) { fop->error = error; } } -void ec_fop_set_error(ec_fop_data_t * fop, int32_t error) +void +ec_fop_set_error(ec_fop_data_t *fop, int32_t error) { LOCK(&fop->lock); @@ -201,18 +435,20 @@ ec_fop_prepare_answer(ec_fop_data_t *fop, gf_boolean_t ro) return cbk; } -void ec_sleep(ec_fop_data_t *fop) +void +ec_sleep(ec_fop_data_t *fop) { LOCK(&fop->lock); - GF_ASSERT (fop->refs > 0); + GF_ASSERT(fop->refs > 0); fop->refs++; fop->jobs++; UNLOCK(&fop->lock); } -int32_t ec_check_complete(ec_fop_data_t * fop, ec_resume_f resume) +int32_t +ec_check_complete(ec_fop_data_t *fop, ec_resume_f resume) { int32_t error = -1; @@ -220,14 +456,11 @@ int32_t ec_check_complete(ec_fop_data_t * fop, ec_resume_f resume) GF_ASSERT(fop->resume == NULL); - if (--fop->jobs != 0) - { + if (--fop->jobs != 0) { ec_trace("WAIT", fop, "resume=%p", resume); fop->resume = resume; - } - else - { + } else { error = fop->error; fop->error = 0; } @@ -237,7 +470,8 @@ int32_t ec_check_complete(ec_fop_data_t * fop, ec_resume_f resume) return error; } -void ec_resume(ec_fop_data_t * fop, int32_t error) +void +ec_resume(ec_fop_data_t *fop, int32_t error) { ec_resume_f resume = NULL; @@ -245,16 +479,13 @@ void ec_resume(ec_fop_data_t * fop, int32_t error) __ec_fop_set_error(fop, error); - if (--fop->jobs == 0) - { + if (--fop->jobs == 0) { resume = fop->resume; fop->resume = NULL; - if (resume != NULL) - { + if (resume != NULL) { ec_trace("RESUME", fop, "error=%d", error); - if (fop->error != 0) - { + if (fop->error != 0) { error = fop->error; } fop->error = 0; @@ -263,21 +494,24 @@ void ec_resume(ec_fop_data_t * fop, int32_t error) UNLOCK(&fop->lock); - if (resume != NULL) - { + if (resume != NULL) { resume(fop, error); } ec_fop_data_release(fop); } -void ec_resume_parent(ec_fop_data_t * fop, int32_t error) +void +ec_resume_parent(ec_fop_data_t *fop) { - ec_fop_data_t * parent; + ec_fop_data_t *parent; + int32_t error = 0; parent = fop->parent; - if (parent != NULL) - { + if (parent != NULL) { + if ((fop->fop_flags & EC_FOP_NO_PROPAGATE_ERROR) == 0) { + error = fop->error; + } ec_trace("RESUME_PARENT", fop, "error=%u", error); fop->parent = NULL; ec_resume(parent, error); @@ -285,22 +519,23 @@ void ec_resume_parent(ec_fop_data_t * fop, int32_t error) } gf_boolean_t -ec_is_recoverable_error (int32_t op_errno) +ec_is_recoverable_error(int32_t op_errno) { - switch (op_errno) { + switch (op_errno) { case ENOTCONN: case ESTALE: case ENOENT: - case EBADFD:/*Opened fd but brick is disconnected*/ - case EIO:/*Backend-fs crash like XFS/ext4 etc*/ - return _gf_true; - } - return _gf_false; + case EBADFD: /*Opened fd but brick is disconnected*/ + case EIO: /*Backend-fs crash like XFS/ext4 etc*/ + return _gf_true; + } + return _gf_false; } -void ec_complete(ec_fop_data_t * fop) +void +ec_complete(ec_fop_data_t *fop) { - ec_cbk_data_t * cbk = NULL; + ec_cbk_data_t *cbk = NULL; int32_t resume = 0, update = 0; int healing_count = 0; @@ -312,9 +547,9 @@ void ec_complete(ec_fop_data_t * fop) if (fop->answer == NULL) { if (!list_empty(&fop->cbk_list)) { cbk = list_entry(fop->cbk_list.next, ec_cbk_data_t, list); - healing_count = ec_bits_count (cbk->mask & fop->healing); - /* fop shouldn't be treated as success if it is not - * successful on at least fop->minimum good copies*/ + healing_count = gf_bits_count(cbk->mask & fop->healing); + /* fop shouldn't be treated as success if it is not + * successful on at least fop->minimum good copies*/ if ((cbk->count - healing_count) >= fop->minimum) { fop->answer = cbk; @@ -336,8 +571,7 @@ void ec_complete(ec_fop_data_t * fop) ec_update_good(fop, cbk->mask); } - if (resume) - { + if (resume) { ec_resume(fop, 0); } @@ -347,36 +581,95 @@ void ec_complete(ec_fop_data_t * fop) /* There could be already granted locks sitting on the bricks, unlock for which * must be wound at all costs*/ static gf_boolean_t -ec_must_wind (ec_fop_data_t *fop) +ec_must_wind(ec_fop_data_t *fop) { - if ((fop->id == GF_FOP_INODELK) || (fop->id == GF_FOP_FINODELK) || - (fop->id == GF_FOP_LK)) { - if (fop->flock.l_type == F_UNLCK) - return _gf_true; - } else if ((fop->id == GF_FOP_ENTRYLK) || - (fop->id == GF_FOP_FENTRYLK)) { - if (fop->entrylk_cmd == ENTRYLK_UNLOCK) - return _gf_true; - } + if ((fop->id == GF_FOP_INODELK) || (fop->id == GF_FOP_FINODELK) || + (fop->id == GF_FOP_LK)) { + if (fop->flock.l_type == F_UNLCK) + return _gf_true; + } else if ((fop->id == GF_FOP_ENTRYLK) || (fop->id == GF_FOP_FENTRYLK)) { + if (fop->entrylk_cmd == ENTRYLK_UNLOCK) + return _gf_true; + } - return _gf_false; + return _gf_false; } static gf_boolean_t -ec_internal_op (ec_fop_data_t *fop) -{ - if (ec_must_wind (fop)) - return _gf_true; - if (fop->id == GF_FOP_XATTROP) - return _gf_true; - if (fop->id == GF_FOP_FXATTROP) - return _gf_true; - return _gf_false; +ec_internal_op(ec_fop_data_t *fop) +{ + if (ec_must_wind(fop)) + return _gf_true; + if (fop->id == GF_FOP_XATTROP) + return _gf_true; + if (fop->id == GF_FOP_FXATTROP) + return _gf_true; + if (fop->id == GF_FOP_OPEN) + return _gf_true; + return _gf_false; +} + +char * +ec_msg_str(ec_fop_data_t *fop) +{ + loc_t *loc1 = NULL; + loc_t *loc2 = NULL; + char gfid1[64] = {0}; + char gfid2[64] = {0}; + ec_fop_data_t *parent = fop->parent; + + if (fop->errstr) + return fop->errstr; + if (!fop->use_fd) { + loc1 = &fop->loc[0]; + loc2 = &fop->loc[1]; + + if (fop->id == GF_FOP_RENAME) { + gf_asprintf(&fop->errstr, + "FOP : '%s' failed on '%s' and '%s' with gfids " + "%s and %s respectively. Parent FOP: %s", + ec_fop_name(fop->id), loc1->path, loc2->path, + uuid_utoa_r(loc1->gfid, gfid1), + uuid_utoa_r(loc2->gfid, gfid2), + parent ? ec_fop_name(parent->id) : "No Parent"); + } else { + gf_asprintf( + &fop->errstr, + "FOP : '%s' failed on '%s' with gfid %s. Parent FOP: %s", + ec_fop_name(fop->id), loc1->path, + uuid_utoa_r(loc1->gfid, gfid1), + parent ? ec_fop_name(parent->id) : "No Parent"); + } + } else { + gf_asprintf( + &fop->errstr, "FOP : '%s' failed on gfid %s. Parent FOP: %s", + ec_fop_name(fop->id), uuid_utoa_r(fop->fd->inode->gfid, gfid1), + parent ? ec_fop_name(parent->id) : "No Parent"); + } + return fop->errstr; } -int32_t ec_child_select(ec_fop_data_t * fop) +static void +ec_log_insufficient_vol(ec_fop_data_t *fop, int32_t have, uint32_t need, + int32_t loglevel) { - ec_t * ec = fop->xl->private; + ec_t *ec = fop->xl->private; + char str1[32], str2[32], str3[32]; + + gf_msg(ec->xl->name, loglevel, 0, EC_MSG_CHILDS_INSUFFICIENT, + "Insufficient available children for this request: " + "Have : %d, Need : %u : Child UP : %s " + "Mask: %s, Healing : %s : %s ", + have, need, ec_bin(str1, sizeof(str1), ec->xl_up, ec->nodes), + ec_bin(str2, sizeof(str2), fop->mask, ec->nodes), + ec_bin(str3, sizeof(str3), fop->healing, ec->nodes), + ec_msg_str(fop)); +} + +static int32_t +ec_child_select(ec_fop_data_t *fop) +{ + ec_t *ec = fop->xl->private; int32_t first = 0, num = 0; ec_fop_cleanup(fop); @@ -385,27 +678,25 @@ int32_t ec_child_select(ec_fop_data_t * fop) /* Wind the fop on same subvols as parent for any internal extra fops like * head/tail read in case of writev fop. Unlocks shouldn't do this because * unlock should go on all subvols where lock is performed*/ - if (fop->parent && !ec_internal_op (fop)) { - fop->mask &= (fop->parent->mask & ~fop->parent->healing); + if (fop->parent && !ec_internal_op(fop)) { + fop->mask &= (fop->parent->mask & ~fop->parent->healing); + if (ec_is_data_fop(fop->id)) { + fop->healing |= fop->parent->healing; + } } - if ((fop->mask & ~ec->xl_up) != 0) - { - gf_msg (fop->xl->name, GF_LOG_WARNING, 0, - EC_MSG_OP_EXEC_UNAVAIL, - "Executing operation with " - "some subvolumes unavailable " - "(%lX)", fop->mask & ~ec->xl_up); - + if ((fop->mask & ~ec->xl_up) != 0) { + gf_msg(fop->xl->name, GF_LOG_WARNING, 0, EC_MSG_OP_EXEC_UNAVAIL, + "Executing operation with " + "some subvolumes unavailable. (%" PRIXPTR "). %s ", + fop->mask & ~ec->xl_up, ec_msg_str(fop)); fop->mask &= ec->xl_up; } - switch (fop->minimum) - { + switch (fop->minimum) { case EC_MINIMUM_ALL: - fop->minimum = ec_bits_count(fop->mask); - if (fop->minimum >= ec->fragments) - { + fop->minimum = gf_bits_count(fop->mask); + if (fop->minimum >= ec->fragments) { break; } case EC_MINIMUM_MIN: @@ -415,13 +706,15 @@ int32_t ec_child_select(ec_fop_data_t * fop) fop->minimum = 1; } - first = ec->idx; - if (++first >= ec->nodes) - { - first = 0; + if (ec->read_policy == EC_ROUND_ROBIN) { + first = ec->idx; + if (++first >= ec->nodes) { + first = 0; + } + ec->idx = first; } - ec->idx = first; + num = gf_bits_count(fop->mask); /*Unconditionally wind on healing subvolumes*/ fop->mask |= fop->healing; fop->remaining = fop->mask; @@ -429,32 +722,35 @@ int32_t ec_child_select(ec_fop_data_t * fop) ec_trace("SELECT", fop, ""); - num = ec_bits_count(fop->mask); - if ((num < fop->minimum) && (num < ec->fragments)) - { - gf_msg (ec->xl->name, GF_LOG_ERROR, 0, - EC_MSG_CHILDS_INSUFFICIENT, - "Insufficient available childs " - "for this request (have %d, need " - "%d)", num, fop->minimum); - + if ((num < fop->minimum) && (num < ec->fragments)) { + ec_log_insufficient_vol(fop, num, fop->minimum, GF_LOG_ERROR); return 0; } - ec_sleep(fop); + if (!fop->parent && fop->lock_count && + (fop->locks[0].update[EC_DATA_TXN] || + fop->locks[0].update[EC_METADATA_TXN])) { + if (ec->quorum_count && (num < ec->quorum_count)) { + ec_log_insufficient_vol(fop, num, ec->quorum_count, GF_LOG_ERROR); + return 0; + } + } return 1; } -int32_t ec_dispatch_next(ec_fop_data_t * fop, int32_t idx) +void +ec_dispatch_next(ec_fop_data_t *fop, uint32_t idx) { - ec_t * ec = fop->xl->private; + uint32_t i = EC_INVALID_INDEX; + ec_t *ec = fop->xl->private; LOCK(&fop->lock); - idx = ec_child_next(ec, fop, idx); - if (idx >= 0) - { + i = ec_child_next(ec, fop, idx); + if (i < EC_MAX_NODES) { + idx = i; + fop->remaining ^= 1ULL << idx; ec_trace("EXECUTE", fop, "idx=%d", idx); @@ -465,20 +761,18 @@ int32_t ec_dispatch_next(ec_fop_data_t * fop, int32_t idx) UNLOCK(&fop->lock); - if (idx >= 0) - { + if (i < EC_MAX_NODES) { fop->wind(ec, fop, idx); } - - return idx; } -void ec_dispatch_mask(ec_fop_data_t * fop, uintptr_t mask) +void +ec_dispatch_mask(ec_fop_data_t *fop, uintptr_t mask) { - ec_t * ec = fop->xl->private; + ec_t *ec = fop->xl->private; int32_t count, idx; - count = ec_bits_count(mask); + count = gf_bits_count(mask); LOCK(&fop->lock); @@ -492,10 +786,8 @@ void ec_dispatch_mask(ec_fop_data_t * fop, uintptr_t mask) UNLOCK(&fop->lock); idx = 0; - while (mask != 0) - { - if ((mask & 1) != 0) - { + while (mask != 0) { + if ((mask & 1) != 0) { fop->wind(ec, fop, idx); } idx++; @@ -503,29 +795,29 @@ void ec_dispatch_mask(ec_fop_data_t * fop, uintptr_t mask) } } -void ec_dispatch_start(ec_fop_data_t * fop) +void +ec_dispatch_start(ec_fop_data_t *fop) { fop->answer = NULL; fop->good = 0; INIT_LIST_HEAD(&fop->cbk_list); - if (fop->lock_count > 0) - { + if (fop->lock_count > 0) { ec_owner_copy(fop->frame, &fop->req_frame->root->lk_owner); } } -void ec_dispatch_one(ec_fop_data_t * fop) +void +ec_dispatch_one(ec_fop_data_t *fop) { - ec_t * ec = fop->xl->private; - ec_dispatch_start(fop); - if (ec_child_select(fop)) - { + if (ec_child_select(fop)) { + ec_sleep(fop); + fop->expected = 1; - fop->first = ec->idx; + fop->first = ec_select_first_by_read_policy(fop->xl->private, fop); ec_dispatch_next(fop, fop->first); } @@ -541,8 +833,8 @@ ec_dispatch_one_retry(ec_fop_data_t *fop, ec_cbk_data_t **cbk) *cbk = tmp; } if ((tmp != NULL) && (tmp->op_ret < 0) && - ec_is_recoverable_error (tmp->op_errno)) { - GF_ASSERT (fop->mask & (1ULL << tmp->idx)); + ec_is_recoverable_error(tmp->op_errno)) { + GF_ASSERT(fop->mask & (1ULL << tmp->idx)); fop->mask ^= (1ULL << tmp->idx); if (fop->mask) { return _gf_true; @@ -552,13 +844,15 @@ ec_dispatch_one_retry(ec_fop_data_t *fop, ec_cbk_data_t **cbk) return _gf_false; } -void ec_dispatch_inc(ec_fop_data_t * fop) +void +ec_dispatch_inc(ec_fop_data_t *fop) { ec_dispatch_start(fop); - if (ec_child_select(fop)) - { - fop->expected = ec_bits_count(fop->remaining); + if (ec_child_select(fop)) { + ec_sleep(fop); + + fop->expected = gf_bits_count(fop->remaining); fop->first = 0; ec_dispatch_next(fop, 0); @@ -566,55 +860,76 @@ void ec_dispatch_inc(ec_fop_data_t * fop) } void -ec_dispatch_all (ec_fop_data_t *fop) +ec_dispatch_all(ec_fop_data_t *fop) { - ec_dispatch_start(fop); + ec_dispatch_start(fop); - if (ec_child_select(fop)) { - fop->expected = ec_bits_count(fop->remaining); - fop->first = 0; + if (ec_child_select(fop)) { + ec_sleep(fop); - ec_dispatch_mask(fop, fop->remaining); - } + fop->expected = gf_bits_count(fop->remaining); + fop->first = 0; + + ec_dispatch_mask(fop, fop->remaining); + } } -void ec_dispatch_min(ec_fop_data_t * fop) +void +ec_dispatch_min(ec_fop_data_t *fop) { - ec_t * ec = fop->xl->private; + ec_t *ec = fop->xl->private; uintptr_t mask; - int32_t idx, count; + uint32_t idx; + int32_t count; ec_dispatch_start(fop); - if (ec_child_select(fop)) - { + if (ec_child_select(fop)) { + ec_sleep(fop); + fop->expected = count = ec->fragments; - fop->first = ec->idx; + fop->first = ec_select_first_by_read_policy(fop->xl->private, fop); idx = fop->first - 1; mask = 0; - while (count-- > 0) - { + while (count-- > 0) { idx = ec_child_next(ec, fop, idx + 1); - mask |= 1ULL << idx; + if (idx < EC_MAX_NODES) + mask |= 1ULL << idx; } ec_dispatch_mask(fop, mask); } } -ec_lock_t *ec_lock_allocate(ec_fop_data_t *fop, loc_t *loc) +void +ec_succeed_all(ec_fop_data_t *fop) +{ + ec_dispatch_start(fop); + + if (ec_child_select(fop)) { + fop->expected = gf_bits_count(fop->remaining); + fop->first = 0; + + /* Simulate a successful execution on all bricks */ + ec_trace("SUCCEED", fop, ""); + + fop->good = fop->remaining; + fop->remaining = 0; + } +} + +ec_lock_t * +ec_lock_allocate(ec_fop_data_t *fop, loc_t *loc) { ec_t *ec = fop->xl->private; - ec_lock_t * lock; + ec_lock_t *lock; int32_t err; if ((loc->inode == NULL) || - (gf_uuid_is_null(loc->gfid) && gf_uuid_is_null(loc->inode->gfid))) - { - gf_msg (fop->xl->name, GF_LOG_ERROR, EINVAL, - EC_MSG_INVALID_INODE, - "Trying to lock based on an invalid " - "inode"); + (gf_uuid_is_null(loc->gfid) && gf_uuid_is_null(loc->inode->gfid))) { + gf_msg(fop->xl->name, GF_LOG_ERROR, EINVAL, EC_MSG_INVALID_INODE, + "Trying to lock based on an invalid " + "inode"); __ec_fop_set_error(fop, EINVAL); @@ -622,9 +937,8 @@ ec_lock_t *ec_lock_allocate(ec_fop_data_t *fop, loc_t *loc) } lock = mem_get0(ec->lock_pool); - if (lock != NULL) - { - lock->good_mask = -1ULL; + if (lock != NULL) { + lock->good_mask = UINTPTR_MAX; INIT_LIST_HEAD(&lock->owners); INIT_LIST_HEAD(&lock->waiting); INIT_LIST_HEAD(&lock->frozen); @@ -640,7 +954,8 @@ ec_lock_t *ec_lock_allocate(ec_fop_data_t *fop, loc_t *loc) return lock; } -void ec_lock_destroy(ec_lock_t * lock) +void +ec_lock_destroy(ec_lock_t *lock) { loc_wipe(&lock->loc); if (lock->fd != NULL) { @@ -650,13 +965,15 @@ void ec_lock_destroy(ec_lock_t * lock) mem_put(lock); } -int32_t ec_lock_compare(ec_lock_t * lock1, ec_lock_t * lock2) +int32_t +ec_lock_compare(ec_lock_t *lock1, ec_lock_t *lock2) { return gf_uuid_compare(lock1->loc.gfid, lock2->loc.gfid); } -void ec_lock_insert(ec_fop_data_t *fop, ec_lock_t *lock, uint32_t flags, - loc_t *base) +static void +ec_lock_insert(ec_fop_data_t *fop, ec_lock_t *lock, uint32_t flags, loc_t *base, + off_t fl_start, uint64_t fl_size) { ec_lock_link_t *link; @@ -690,13 +1007,15 @@ void ec_lock_insert(ec_fop_data_t *fop, ec_lock_t *lock, uint32_t flags, link->update[EC_DATA_TXN] = (flags & EC_UPDATE_DATA) != 0; link->update[EC_METADATA_TXN] = (flags & EC_UPDATE_META) != 0; link->base = base; + link->fl_start = fl_start; + link->fl_end = ec_range_end_get(fl_start, fl_size); - lock->refs++; - lock->inserted++; + lock->refs_pending++; } -void ec_lock_prepare_inode_internal(ec_fop_data_t *fop, loc_t *loc, - uint32_t flags, loc_t *base) +static void +ec_lock_prepare_inode_internal(ec_fop_data_t *fop, loc_t *loc, uint32_t flags, + loc_t *base, off_t fl_start, uint64_t fl_size) { ec_lock_t *lock = NULL; ec_inode_t *ctx; @@ -725,8 +1044,8 @@ void ec_lock_prepare_inode_internal(ec_fop_data_t *fop, loc_t *loc, if ((fop->lock_count > 0) && (fop->locks[0].lock == lock)) { /* Combine data/meta updates */ fop->locks[0].update[EC_DATA_TXN] |= (flags & EC_UPDATE_DATA) != 0; - fop->locks[0].update[EC_METADATA_TXN] |= - (flags & EC_UPDATE_META) != 0; + fop->locks[0].update[EC_METADATA_TXN] |= (flags & EC_UPDATE_META) != + 0; /* Only one base inode is allowed per fop, so there shouldn't be * overwrites here. */ @@ -737,8 +1056,10 @@ void ec_lock_prepare_inode_internal(ec_fop_data_t *fop, loc_t *loc, goto update_query; } - ec_trace("LOCK_INODELK", fop, "lock=%p, inode=%p. Lock already " - "acquired", lock, loc->inode); + ec_trace("LOCK_INODELK", fop, + "lock=%p, inode=%p. Lock already " + "acquired", + lock, loc->inode); goto insert; } @@ -757,22 +1078,25 @@ void ec_lock_prepare_inode_internal(ec_fop_data_t *fop, loc_t *loc, ctx->inode_lock = lock; insert: - ec_lock_insert(fop, lock, flags, base); + ec_lock_insert(fop, lock, flags, base, fl_start, fl_size); update_query: lock->query |= (flags & EC_QUERY_INFO) != 0; unlock: UNLOCK(&loc->inode->lock); } -void ec_lock_prepare_inode(ec_fop_data_t *fop, loc_t *loc, uint32_t flags) +void +ec_lock_prepare_inode(ec_fop_data_t *fop, loc_t *loc, uint32_t flags, + off_t fl_start, uint64_t fl_size) { - ec_lock_prepare_inode_internal(fop, loc, flags, NULL); + ec_lock_prepare_inode_internal(fop, loc, flags, NULL, fl_start, fl_size); } -void ec_lock_prepare_parent_inode(ec_fop_data_t *fop, loc_t *loc, - uint32_t flags) +void +ec_lock_prepare_parent_inode(ec_fop_data_t *fop, loc_t *loc, loc_t *base, + uint32_t flags) { - loc_t tmp, *base = NULL; + loc_t tmp; int32_t err; if (fop->error != 0) { @@ -787,16 +1111,19 @@ void ec_lock_prepare_parent_inode(ec_fop_data_t *fop, loc_t *loc, } if ((flags & EC_INODE_SIZE) != 0) { - base = loc; flags ^= EC_INODE_SIZE; + } else { + base = NULL; } - ec_lock_prepare_inode_internal(fop, &tmp, flags, base); + ec_lock_prepare_inode_internal(fop, &tmp, flags, base, 0, EC_RANGE_FULL); loc_wipe(&tmp); } -void ec_lock_prepare_fd(ec_fop_data_t *fop, fd_t *fd, uint32_t flags) +void +ec_lock_prepare_fd(ec_fop_data_t *fop, fd_t *fd, uint32_t flags, off_t fl_start, + uint64_t fl_size) { loc_t loc; int32_t err; @@ -812,21 +1139,20 @@ void ec_lock_prepare_fd(ec_fop_data_t *fop, fd_t *fd, uint32_t flags) return; } - ec_lock_prepare_inode_internal(fop, &loc, flags, NULL); + ec_lock_prepare_inode_internal(fop, &loc, flags, NULL, fl_start, fl_size); loc_wipe(&loc); } gf_boolean_t -ec_config_check (ec_fop_data_t *fop, ec_config_t *config) +ec_config_check(xlator_t *xl, ec_config_t *config) { ec_t *ec; - ec = fop->xl->private; + ec = xl->private; if ((config->version != EC_CONFIG_VERSION) || (config->algorithm != EC_CONFIG_ALGORITHM) || - (config->gf_word_size != EC_GF_BITS) || - (config->bricks != ec->nodes) || + (config->gf_word_size != EC_GF_BITS) || (config->bricks != ec->nodes) || (config->redundancy != ec->redundancy) || (config->chunk_size != EC_METHOD_CHUNK_SIZE)) { uint32_t data_bricks; @@ -845,20 +1171,17 @@ ec_config_check (ec_fop_data_t *fop, ec_config_t *config) if ((config->redundancy < 1) || (config->redundancy * 2 >= config->bricks) || !ec_is_power_of_2(config->gf_word_size) || - ((config->chunk_size * 8) % (config->gf_word_size * data_bricks) - != 0)) { - gf_msg (fop->xl->name, GF_LOG_ERROR, EINVAL, - EC_MSG_INVALID_CONFIG, - "Invalid or corrupted config"); + ((config->chunk_size * 8) % (config->gf_word_size * data_bricks) != + 0)) { + gf_msg(xl->name, GF_LOG_ERROR, EINVAL, EC_MSG_INVALID_CONFIG, + "Invalid or corrupted config"); } else { - gf_msg (fop->xl->name, GF_LOG_ERROR, EINVAL, - EC_MSG_INVALID_CONFIG, - "Unsupported config " - "(V=%u, A=%u, W=%u, " - "N=%u, R=%u, S=%u)", - config->version, config->algorithm, - config->gf_word_size, config->bricks, - config->redundancy, config->chunk_size); + gf_msg(xl->name, GF_LOG_ERROR, EINVAL, EC_MSG_INVALID_CONFIG, + "Unsupported config " + "(V=%u, A=%u, W=%u, " + "N=%u, R=%u, S=%u)", + config->version, config->algorithm, config->gf_word_size, + config->bricks, config->redundancy, config->chunk_size); } return _gf_false; @@ -867,111 +1190,164 @@ ec_config_check (ec_fop_data_t *fop, ec_config_t *config) return _gf_true; } +gf_boolean_t +ec_set_dirty_flag(ec_lock_link_t *link, ec_inode_t *ctx, uint64_t *dirty) +{ + gf_boolean_t set_dirty = _gf_false; + + if (link->update[EC_DATA_TXN] && !ctx->dirty[EC_DATA_TXN]) { + if (!link->optimistic_changelog) + dirty[EC_DATA_TXN] = 1; + } + + if (link->update[EC_METADATA_TXN] && !ctx->dirty[EC_METADATA_TXN]) { + if (!link->optimistic_changelog) + dirty[EC_METADATA_TXN] = 1; + } + + if (dirty[EC_METADATA_TXN] || dirty[EC_DATA_TXN]) { + set_dirty = _gf_true; + } + + return set_dirty; +} + int32_t -ec_prepare_update_cbk (call_frame_t *frame, void *cookie, - xlator_t *this, int32_t op_ret, int32_t op_errno, - dict_t *dict, dict_t *xdata) +ec_prepare_update_cbk(call_frame_t *frame, void *cookie, xlator_t *this, + int32_t op_ret, int32_t op_errno, dict_t *dict, + dict_t *xdata) { struct list_head list; ec_fop_data_t *fop = cookie, *parent, *tmp; - ec_lock_link_t *link = fop->data; + ec_lock_link_t *parent_link = fop->data; + ec_lock_link_t *link = NULL; ec_lock_t *lock = NULL; ec_inode_t *ctx; - - lock = link->lock; - parent = link->fop; + gf_boolean_t release = _gf_false; + uint64_t provided_flags = 0; + uint64_t dirty[EC_VERSION_SIZE] = {0, 0}; + lock = parent_link->lock; + parent = parent_link->fop; ctx = lock->ctx; INIT_LIST_HEAD(&list); + provided_flags = EC_PROVIDED_FLAGS(parent_link->waiting_flags); LOCK(&lock->loc.inode->lock); - list_for_each_entry(tmp, &lock->owners, owner_list) { - if ((tmp->flags & EC_FLAG_WAITING_SIZE) != 0) { - tmp->flags ^= EC_FLAG_WAITING_SIZE; - - list_add_tail(&tmp->cbk_list, &list); + list_for_each_entry(link, &lock->owners, owner_list) + { + if ((link->waiting_flags & provided_flags) != 0) { + link->waiting_flags ^= (link->waiting_flags & provided_flags); + if (EC_NEEDED_FLAGS(link->waiting_flags) == 0) + list_add_tail(&link->fop->cbk_list, &list); } } - if (op_ret < 0) { - gf_msg (this->name, GF_LOG_WARNING, op_errno, - EC_MSG_SIZE_VERS_GET_FAIL, - "Failed to get size and version"); + gf_msg(this->name, GF_LOG_WARNING, op_errno, EC_MSG_SIZE_VERS_GET_FAIL, + "Failed to get size and version : %s", ec_msg_str(fop)); goto unlock; } - op_errno = -ec_dict_del_array(dict, EC_XATTR_VERSION, ctx->pre_version, - EC_VERSION_SIZE); - if (op_errno != 0) { - gf_msg (this->name, GF_LOG_ERROR, op_errno, - EC_MSG_VER_XATTR_GET_FAIL, - "Unable to get version xattr"); - - goto unlock; - } - ctx->post_version[0] += ctx->pre_version[0]; - ctx->post_version[1] += ctx->pre_version[1]; - - ctx->have_version = _gf_true; - - if (lock->loc.inode->ia_type == IA_IFREG) { - op_errno = -ec_dict_del_number(dict, EC_XATTR_SIZE, &ctx->pre_size); + if (EC_FLAGS_HAVE(provided_flags, EC_FLAG_XATTROP)) { + op_errno = -ec_dict_del_array(dict, EC_XATTR_VERSION, ctx->pre_version, + EC_VERSION_SIZE); if (op_errno != 0) { - gf_msg (this->name, GF_LOG_ERROR, op_errno, - EC_MSG_SIZE_XATTR_GET_FAIL, "Unable to get size xattr"); - + gf_msg(this->name, GF_LOG_ERROR, op_errno, + EC_MSG_VER_XATTR_GET_FAIL, "Unable to get version xattr. %s", + ec_msg_str(fop)); goto unlock; } - ctx->post_size = ctx->pre_size; + ctx->post_version[0] += ctx->pre_version[0]; + ctx->post_version[1] += ctx->pre_version[1]; + + ctx->have_version = _gf_true; + + if (lock->loc.inode->ia_type == IA_IFREG || + lock->loc.inode->ia_type == IA_INVAL) { + op_errno = -ec_dict_del_number(dict, EC_XATTR_SIZE, &ctx->pre_size); + if (op_errno != 0) { + if (lock->loc.inode->ia_type == IA_IFREG) { + gf_msg(this->name, GF_LOG_ERROR, op_errno, + EC_MSG_SIZE_XATTR_GET_FAIL, + "Unable to get size xattr. %s", ec_msg_str(fop)); + goto unlock; + } + } else { + ctx->post_size = ctx->pre_size; - ctx->have_size = _gf_true; + ctx->have_size = _gf_true; + } - op_errno = -ec_dict_del_config(dict, EC_XATTR_CONFIG, &ctx->config); - if (op_errno != 0) { - gf_msg (this->name, GF_LOG_ERROR, op_errno, - EC_MSG_CONFIG_XATTR_GET_FAIL, - "Unable to get config xattr"); + op_errno = -ec_dict_del_config(dict, EC_XATTR_CONFIG, &ctx->config); + if (op_errno != 0) { + if ((lock->loc.inode->ia_type == IA_IFREG) || + (op_errno != ENODATA)) { + gf_msg(this->name, GF_LOG_ERROR, op_errno, + EC_MSG_CONFIG_XATTR_GET_FAIL, + "Unable to get config xattr. %s", ec_msg_str(fop)); - goto unlock; - } - if (!ec_config_check(parent, &ctx->config)) { - gf_msg (this->name, GF_LOG_ERROR, EINVAL, - EC_MSG_CONFIG_XATTR_INVALID, - "Invalid config xattr"); + goto unlock; + } + } else { + if (!ec_config_check(parent->xl, &ctx->config)) { + gf_msg(this->name, GF_LOG_ERROR, EINVAL, + EC_MSG_CONFIG_XATTR_INVALID, "Invalid config xattr"); - op_errno = EINVAL; + op_errno = EINVAL; - goto unlock; + goto unlock; + } + ctx->have_config = _gf_true; + } } - - ctx->have_config = _gf_true; + ctx->have_info = _gf_true; } - ctx->have_info = _gf_true; + ec_set_dirty_flag(fop->data, ctx, dirty); + if (dirty[EC_METADATA_TXN] && + (EC_FLAGS_HAVE(provided_flags, EC_FLAG_METADATA_DIRTY))) { + GF_ASSERT(!ctx->dirty[EC_METADATA_TXN]); + ctx->dirty[EC_METADATA_TXN] = 1; + } + if (dirty[EC_DATA_TXN] && + (EC_FLAGS_HAVE(provided_flags, EC_FLAG_DATA_DIRTY))) { + GF_ASSERT(!ctx->dirty[EC_DATA_TXN]); + ctx->dirty[EC_DATA_TXN] = 1; + } op_errno = 0; - unlock: - lock->getting_size = _gf_false; - UNLOCK(&lock->loc.inode->lock); + lock->waiting_flags ^= provided_flags; if (op_errno == 0) { + /* If the fop fails on any of the good bricks, it is important to mark + * it dirty and update versions right away if dirty was not set before. + */ + if (lock->good_mask & ~(fop->good | fop->remaining)) { + release = _gf_true; + } + + if (parent_link->update[0] && !parent_link->dirty[0]) { + lock->release |= release; + } + + if (parent_link->update[1] && !parent_link->dirty[1]) { + lock->release |= release; + } + /* We don't allow the main fop to be executed on bricks that have not * succeeded the initial xattrop. */ - parent->mask &= fop->good; + ec_lock_update_good(lock, fop); /*As of now only data healing marks bricks as healing*/ lock->healing |= fop->healing; - if (ec_is_data_fop (parent->id)) { - parent->healing |= fop->healing; - } - } else { - ec_fop_set_error(parent, op_errno); } + UNLOCK(&lock->loc.inode->lock); + while (!list_empty(&list)) { tmp = list_entry(list.next, ec_fop_data_t, cbk_list); list_del_init(&tmp->cbk_list); @@ -980,95 +1356,163 @@ unlock: tmp->mask &= fop->good; /*As of now only data healing marks bricks as healing*/ - if (ec_is_data_fop (tmp->id)) { + if (ec_is_data_fop(tmp->id)) { tmp->healing |= fop->healing; } - } else { - ec_fop_set_error(tmp, op_errno); } - ec_resume(tmp, 0); + ec_resume(tmp, op_errno); } return 0; } -void ec_get_size_version(ec_lock_link_t *link) +static gf_boolean_t +ec_set_needed_flag(ec_lock_t *lock, ec_lock_link_t *link, uint64_t flag) +{ + uint64_t current; + + link->waiting_flags |= EC_FLAG_NEEDS(flag); + + current = EC_NEEDED_FLAGS(lock->waiting_flags); + if (!EC_FLAGS_HAVE(current, flag)) { + lock->waiting_flags |= EC_FLAG_NEEDS(flag); + link->waiting_flags |= EC_FLAG_PROVIDES(flag); + + return _gf_true; + } + + return _gf_false; +} + +static uint64_t +ec_set_xattrop_flags_and_params(ec_lock_t *lock, ec_lock_link_t *link, + uint64_t *dirty) +{ + uint64_t oldflags = 0; + uint64_t newflags = 0; + ec_inode_t *ctx = lock->ctx; + + oldflags = EC_NEEDED_FLAGS(lock->waiting_flags); + + if (lock->query && !ctx->have_info) { + ec_set_needed_flag(lock, link, EC_FLAG_XATTROP); + } + + if (dirty[EC_DATA_TXN]) { + if (!ec_set_needed_flag(lock, link, EC_FLAG_DATA_DIRTY)) { + dirty[EC_DATA_TXN] = 0; + } + } + + if (dirty[EC_METADATA_TXN]) { + if (!ec_set_needed_flag(lock, link, EC_FLAG_METADATA_DIRTY)) { + dirty[EC_METADATA_TXN] = 0; + } + } + newflags = EC_NEEDED_FLAGS(lock->waiting_flags); + + return oldflags ^ newflags; +} + +void +ec_get_size_version(ec_lock_link_t *link) { loc_t loc; ec_lock_t *lock; ec_inode_t *ctx; ec_fop_data_t *fop; dict_t *dict = NULL; - uid_t uid; - gid_t gid; - int32_t error = -ENOMEM; - gf_boolean_t getting_size; + dict_t *xdata = NULL; + ec_t *ec = NULL; + int32_t error = 0; + gf_boolean_t set_dirty = _gf_false; uint64_t allzero[EC_VERSION_SIZE] = {0, 0}; - + uint64_t dirty[EC_VERSION_SIZE] = {0, 0}; lock = link->lock; ctx = lock->ctx; fop = link->fop; + ec = fop->xl->private; + uint64_t changed_flags = 0; + + if (ec->optimistic_changelog && !(ec->node_mask & ~link->lock->good_mask) && + !ec_is_data_fop(fop->id)) + link->optimistic_changelog = _gf_true; + + memset(&loc, 0, sizeof(loc)); + + LOCK(&lock->loc.inode->lock); + + set_dirty = ec_set_dirty_flag(link, ctx, dirty); /* If ec metadata has already been retrieved, do not try again. */ if (ctx->have_info) { - if (ec_is_data_fop (fop->id)) { + if (ec_is_data_fop(fop->id)) { fop->healing |= lock->healing; } - return; + if (!set_dirty) + goto unlock; } /* Determine if there's something we need to retrieve for the current * operation. */ - if (!lock->query && (lock->loc.inode->ia_type != IA_IFREG)) { - return; + if (!set_dirty && !lock->query && (lock->loc.inode->ia_type != IA_IFREG) && + (lock->loc.inode->ia_type != IA_INVAL)) { + goto unlock; } - uid = fop->frame->root->uid; - gid = fop->frame->root->gid; - - memset(&loc, 0, sizeof(loc)); - - LOCK(&lock->loc.inode->lock); - - getting_size = lock->getting_size; - lock->getting_size = _gf_true; - if (getting_size) { - fop->flags |= EC_FLAG_WAITING_SIZE; - + changed_flags = ec_set_xattrop_flags_and_params(lock, link, dirty); + if (link->waiting_flags) { + /* This fop needs to wait until all its flags are cleared which + * potentially can be cleared by other xattrops that are already + * wound*/ ec_sleep(fop); + } else { + GF_ASSERT(!changed_flags); } +unlock: UNLOCK(&lock->loc.inode->lock); - if (getting_size) { - error = 0; - + if (!changed_flags) goto out; - } dict = dict_new(); if (dict == NULL) { + error = -ENOMEM; goto out; } - /* Once we know that an xattrop will be needed, we try to get all available - * information in a single call. */ - error = ec_dict_set_array(dict, EC_XATTR_VERSION, allzero, - EC_VERSION_SIZE); - if (error == 0) { - error = ec_dict_set_array(dict, EC_XATTR_DIRTY, allzero, + if (EC_FLAGS_HAVE(changed_flags, EC_FLAG_XATTROP)) { + /* Once we know that an xattrop will be needed, + * we try to get all available information in a + * single call. */ + error = ec_dict_set_array(dict, EC_XATTR_VERSION, allzero, EC_VERSION_SIZE); - } - if (error != 0) { - goto out; - } + if (error != 0) { + goto out; + } - if (lock->loc.inode->ia_type == IA_IFREG) { - error = ec_dict_set_number(dict, EC_XATTR_SIZE, 0); - if (error == 0) { - error = ec_dict_set_number(dict, EC_XATTR_CONFIG, 0); + if (lock->loc.inode->ia_type == IA_IFREG || + lock->loc.inode->ia_type == IA_INVAL) { + error = ec_dict_set_number(dict, EC_XATTR_SIZE, 0); + if (error == 0) { + error = ec_dict_set_number(dict, EC_XATTR_CONFIG, 0); + } + if (error != 0) { + goto out; + } + + xdata = dict_new(); + if (xdata == NULL || dict_set_int32(xdata, GF_GET_SIZE, 1)) { + error = -ENOMEM; + goto out; + } } + } + + if (memcmp(allzero, dirty, sizeof(allzero))) { + error = ec_dict_set_array(dict, EC_XATTR_DIRTY, dirty, EC_VERSION_SIZE); if (error != 0) { goto out; } @@ -1099,20 +1543,20 @@ void ec_get_size_version(ec_lock_link_t *link) loc.name = NULL; } - ec_xattrop (fop->frame, fop->xl, fop->mask, fop->minimum, - ec_prepare_update_cbk, link, &loc, - GF_XATTROP_ADD_ARRAY64, dict, NULL); + ec_xattrop(fop->frame, fop->xl, fop->mask, fop->minimum, + ec_prepare_update_cbk, link, &loc, GF_XATTROP_ADD_ARRAY64, + dict, xdata); } else { ec_fxattrop(fop->frame, fop->xl, fop->mask, fop->minimum, - ec_prepare_update_cbk, link, lock->fd, - GF_XATTROP_ADD_ARRAY64, dict, NULL); + ec_prepare_update_cbk, link, lock->fd, + GF_XATTROP_ADD_ARRAY64, dict, xdata); } error = 0; out: - fop->frame->root->uid = uid; - fop->frame->root->gid = gid; + fop->frame->root->uid = fop->uid; + fop->frame->root->gid = fop->gid; loc_wipe(&loc); @@ -1120,22 +1564,24 @@ out: dict_unref(dict); } + if (xdata != NULL) { + dict_unref(xdata); + } + if (error != 0) { ec_fop_set_error(fop, -error); } } -gf_boolean_t ec_get_inode_size(ec_fop_data_t *fop, inode_t *inode, - uint64_t *size) +gf_boolean_t +__ec_get_inode_size(ec_fop_data_t *fop, inode_t *inode, uint64_t *size) { ec_inode_t *ctx; gf_boolean_t found = _gf_false; - LOCK(&inode->lock); - ctx = __ec_inode_get(inode, fop->xl); if (ctx == NULL) { - goto unlock; + goto out; } if (ctx->have_size) { @@ -1143,23 +1589,33 @@ gf_boolean_t ec_get_inode_size(ec_fop_data_t *fop, inode_t *inode, found = _gf_true; } -unlock: +out: + return found; +} + +gf_boolean_t +ec_get_inode_size(ec_fop_data_t *fop, inode_t *inode, uint64_t *size) +{ + gf_boolean_t found = _gf_false; + + LOCK(&inode->lock); + { + found = __ec_get_inode_size(fop, inode, size); + } UNLOCK(&inode->lock); return found; } -gf_boolean_t ec_set_inode_size(ec_fop_data_t *fop, inode_t *inode, - uint64_t size) +gf_boolean_t +__ec_set_inode_size(ec_fop_data_t *fop, inode_t *inode, uint64_t size) { ec_inode_t *ctx; gf_boolean_t found = _gf_false; - LOCK(&inode->lock); - ctx = __ec_inode_get(inode, fop->xl); if (ctx == NULL) { - goto unlock; + goto out; } /* Normal fops always have ctx->have_size set. However self-heal calls this @@ -1174,13 +1630,42 @@ gf_boolean_t ec_set_inode_size(ec_fop_data_t *fop, inode_t *inode, found = _gf_true; -unlock: +out: + return found; +} + +gf_boolean_t +ec_set_inode_size(ec_fop_data_t *fop, inode_t *inode, uint64_t size) +{ + gf_boolean_t found = _gf_false; + + LOCK(&inode->lock); + { + found = __ec_set_inode_size(fop, inode, size); + } UNLOCK(&inode->lock); return found; } -void ec_clear_inode_info(ec_fop_data_t *fop, inode_t *inode) +static void +ec_release_stripe_cache(ec_inode_t *ctx) +{ + ec_stripe_list_t *stripe_cache = NULL; + ec_stripe_t *stripe = NULL; + + stripe_cache = &ctx->stripe_cache; + while (!list_empty(&stripe_cache->lru)) { + stripe = list_first_entry(&stripe_cache->lru, ec_stripe_t, lru); + list_del(&stripe->lru); + GF_FREE(stripe); + } + stripe_cache->count = 0; + stripe_cache->max = 0; +} + +void +ec_clear_inode_info(ec_fop_data_t *fop, inode_t *inode) { ec_inode_t *ctx; @@ -1191,6 +1676,7 @@ void ec_clear_inode_info(ec_fop_data_t *fop, inode_t *inode) goto unlock; } + ec_release_stripe_cache(ctx); ctx->have_info = _gf_false; ctx->have_config = _gf_false; ctx->have_version = _gf_false; @@ -1206,10 +1692,10 @@ unlock: UNLOCK(&inode->lock); } -int32_t ec_get_real_size_cbk(call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno, inode_t *inode, - struct iatt *buf, dict_t *xdata, - struct iatt *postparent) +int32_t +ec_get_real_size_cbk(call_frame_t *frame, void *cookie, xlator_t *this, + int32_t op_ret, int32_t op_errno, inode_t *inode, + struct iatt *buf, dict_t *xdata, struct iatt *postparent) { ec_fop_data_t *fop = cookie; ec_lock_link_t *link; @@ -1226,14 +1712,15 @@ int32_t ec_get_real_size_cbk(call_frame_t *frame, void *cookie, xlator_t *this, } /* This function is used to get the trusted.ec.size xattr from a file when - * no lock is needed on the inode. This is only required to maintan iatt + * no lock is needed on the inode. This is only required to maintain iatt * structs on fops that manipulate directory entries but do not operate * directly on the inode, like link, rename, ... * * Any error processing this request is ignored. In the worst case, an invalid * or not up to date value in the iatt could cause some cache invalidation. */ -void ec_get_real_size(ec_lock_link_t *link) +void +ec_get_real_size(ec_lock_link_t *link) { ec_fop_data_t *fop; dict_t *xdata; @@ -1277,48 +1764,66 @@ ec_lock_update_fd(ec_lock_t *lock, ec_fop_data_t *fop) { /* If the fop has an fd available, attach it to the lock structure to be * able to do fxattrop calls instead of xattrop. */ - if (fop->use_fd) { - if (lock->fd != NULL) { - __fd_unref(lock->fd); - } + if (fop->use_fd && (lock->fd == NULL)) { lock->fd = __fd_ref(fop->fd); } } +static gf_boolean_t +ec_link_has_lock_conflict(ec_lock_link_t *link, gf_boolean_t waitlist_check) +{ + ec_lock_link_t *trav_link = NULL; + + list_for_each_entry(trav_link, &link->lock->owners, owner_list) + { + if (ec_lock_conflict(trav_link, link)) + return _gf_true; + } + + if (!waitlist_check) + return _gf_false; + + list_for_each_entry(trav_link, &link->lock->waiting, wait_list) + { + if (ec_lock_conflict(trav_link, link)) + return _gf_true; + } + + return _gf_false; +} + static void ec_lock_wake_shared(ec_lock_t *lock, struct list_head *list) { ec_fop_data_t *fop; ec_lock_link_t *link; - gf_boolean_t exclusive = _gf_false; + gf_boolean_t conflict = _gf_false; - while (!exclusive && !list_empty(&lock->waiting)) { + while (!conflict && !list_empty(&lock->waiting)) { link = list_entry(lock->waiting.next, ec_lock_link_t, wait_list); fop = link->fop; /* If lock is not acquired, at most one fop can be assigned as owner. * The following fops will need to wait in the lock->waiting queue * until the lock has been fully acquired. */ - exclusive = !lock->acquired; + conflict = !lock->acquired; /* If the fop is not shareable, only this fop can be assigned as owner. * Other fops will need to wait until this one finishes. */ - if ((fop->flags & EC_FLAG_LOCK_SHARED) == 0) { - exclusive = _gf_true; - - /* Avoid other requests to be assigned as owners. */ - lock->exclusive = 1; + if (ec_link_has_lock_conflict(link, _gf_false)) { + conflict = _gf_true; } /* If only one fop is allowed, it can be assigned as the owner of the * lock only if there weren't any other owner. */ - if (exclusive && !list_empty(&lock->owners)) { + if (conflict && !list_empty(&lock->owners)) { break; } list_move_tail(&link->wait_list, list); - list_add_tail(&fop->owner_list, &lock->owners); + list_add_tail(&link->owner_list, &lock->owners); + lock->refs_owners++; ec_lock_update_fd(lock, fop); } @@ -1336,7 +1841,8 @@ ec_lock_apply(ec_lock_link_t *link) ec_get_real_size(link); } -gf_boolean_t ec_lock_acquire(ec_lock_link_t *link); +gf_boolean_t +ec_lock_acquire(ec_lock_link_t *link); static void ec_lock_resume_shared(struct list_head *list) @@ -1360,7 +1866,8 @@ ec_lock_resume_shared(struct list_head *list) } } -void ec_lock_acquired(ec_lock_link_t *link) +void +ec_lock_acquired(ec_lock_link_t *link) { struct list_head list; ec_lock_t *lock; @@ -1376,61 +1883,76 @@ void ec_lock_acquired(ec_lock_link_t *link) LOCK(&lock->loc.inode->lock); lock->acquired = _gf_true; + if (lock->contention) { + lock->release = _gf_true; + lock->contention = _gf_false; + } ec_lock_update_fd(lock, fop); - if ((fop->flags & EC_FLAG_LOCK_SHARED) != 0) { - ec_lock_wake_shared(lock, &list); - } + ec_lock_wake_shared(lock, &list); UNLOCK(&lock->loc.inode->lock); ec_lock_apply(link); + if (fop->use_fd && + (link->update[EC_DATA_TXN] || link->update[EC_METADATA_TXN])) { + /* Try to reopen closed fd's only if lock has succeeded. */ + ec_fix_open(fop, lock->mask); + } + ec_lock_resume_shared(&list); } -int32_t ec_locked(call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno, dict_t *xdata) +int32_t +ec_locked(call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, + int32_t op_errno, dict_t *xdata) { ec_fop_data_t *fop = cookie; ec_lock_link_t *link = NULL; ec_lock_t *lock = NULL; + link = fop->data; + lock = link->lock; if (op_ret >= 0) { - link = fop->data; - lock = link->lock; lock->mask = lock->good_mask = fop->good; lock->healing = 0; ec_lock_acquired(link); ec_lock(fop->parent); } else { - gf_msg (this->name, GF_LOG_WARNING, op_errno, - EC_MSG_PREOP_LOCK_FAILED, - "Failed to complete preop lock"); + LOCK(&lock->loc.inode->lock); + { + lock->contention = _gf_false; + } + UNLOCK(&lock->loc.inode->lock); + gf_msg(this->name, GF_LOG_WARNING, op_errno, EC_MSG_PREOP_LOCK_FAILED, + "Failed to complete preop lock"); } return 0; } -gf_boolean_t ec_lock_acquire(ec_lock_link_t *link) +gf_boolean_t +ec_lock_acquire(ec_lock_link_t *link) { ec_lock_t *lock; ec_fop_data_t *fop; + gf_lkowner_t lk_owner; lock = link->lock; fop = link->fop; if (!lock->acquired) { - ec_owner_set(fop->frame, lock); + set_lk_owner_from_ptr(&lk_owner, lock); ec_trace("LOCK_ACQUIRE", fop, "lock=%p, inode=%p", lock, lock->loc.inode); lock->flock.l_type = F_WRLCK; - ec_inodelk(fop->frame, fop->xl, -1, EC_MINIMUM_ALL, ec_locked, - link, fop->xl->name, &lock->loc, F_SETLKW, &lock->flock, - NULL); + ec_inodelk(fop->frame, fop->xl, &lk_owner, -1, EC_MINIMUM_ALL, + ec_locked, link, fop->xl->name, &lock->loc, F_SETLKW, + &lock->flock, NULL); return _gf_false; } @@ -1442,6 +1964,67 @@ gf_boolean_t ec_lock_acquire(ec_lock_link_t *link) return _gf_true; } +static ec_lock_link_t * +ec_lock_timer_cancel(xlator_t *xl, ec_lock_t *lock) +{ + ec_lock_link_t *timer_link; + + /* If we don't have any timer, there's nothing to cancel. */ + if (lock->timer == NULL) { + return NULL; + } + + /* We are trying to access a lock that has an unlock timer active. + * This means that the lock must be idle, i.e. no fop can be in the + * owner, waiting or frozen lists. It also means that the lock cannot + * have been marked as being released (this is done without timers). + * There should only be one owner reference, but it's possible that + * some fops are being prepared to use this lock. */ + GF_ASSERT((lock->refs_owners == 1) && list_empty(&lock->owners) && + list_empty(&lock->waiting)); + + /* We take the timer_link before cancelling the timer, since a + * successful cancellation will destroy it. It must not be NULL + * because it references the fop responsible for the delayed unlock + * that we are currently trying to cancel. */ + timer_link = lock->timer->data; + GF_ASSERT(timer_link != NULL); + + if (gf_timer_call_cancel(xl->ctx, lock->timer) < 0) { + /* It's too late to avoid the execution of the timer callback. + * Since we need to be sure that the callback has access to all + * needed resources, we cannot resume the execution of the + * timer fop now. This will be done in the callback. */ + timer_link = NULL; + } else { + /* The timer has been cancelled. The fop referenced by + * timer_link holds the last reference. The caller is + * responsible to release it when not needed anymore. */ + ec_trace("UNLOCK_CANCELLED", timer_link->fop, "lock=%p", lock); + } + + /* We have two options here: + * + * 1. The timer has been successfully cancelled. + * + * This is the easiest case and we can continue with the currently + * acquired lock. + * + * 2. The timer callback has already been fired. + * + * In this case we have not been able to cancel the timer before + * the timer callback has been fired, but we also know that + * lock->timer != NULL. This means that the timer callback is still + * trying to acquire the inode mutex that we currently own. We are + * safe until we release it. In this case we can safely clear + * lock->timer. This will cause that the timer callback does nothing + * once it acquires the mutex. + */ + lock->timer = NULL; + + return timer_link; +} + static gf_boolean_t ec_lock_assign_owner(ec_lock_link_t *link) { @@ -1450,6 +2033,8 @@ ec_lock_assign_owner(ec_lock_link_t *link) ec_lock_link_t *timer_link = NULL; gf_boolean_t assigned = _gf_false; + /* The link cannot be in any list because we have just finished preparing + * it. */ GF_ASSERT(list_empty(&link->wait_list)); fop = link->fop; @@ -1457,64 +2042,72 @@ ec_lock_assign_owner(ec_lock_link_t *link) LOCK(&lock->loc.inode->lock); - GF_ASSERT (lock->inserted > 0); - lock->inserted--; + /* Since the link has just been prepared but it's not active yet, the + * refs_pending must be one at least (the ref owned by this link). */ + GF_ASSERT(lock->refs_pending > 0); + /* The link is not pending any more. It will be assigned to the owner, + * waiting or frozen list. */ + lock->refs_pending--; if (lock->release) { ec_trace("LOCK_QUEUE_FREEZE", fop, "lock=%p", lock); - list_add_tail(&link->wait_list, &lock->frozen); + /* When lock->release is set, we'll unlock the lock as soon as + * possible, meaning that we won't use a timer. */ + GF_ASSERT(lock->timer == NULL); - /* The lock is frozen, so we move the current reference to refs_frozen. - * After that, there should remain at least one ref belonging to the - * lock that is processing the release. */ - lock->refs--; - GF_ASSERT(lock->refs > 0); - lock->refs_frozen++; + /* The lock is marked to be released. We can still have owners and fops + * in the waiting ilist f they have been added before the lock has been + * marked to be released. However new fops are put into the frozen list + * to wait for the next unlock/lock cycle. */ + list_add_tail(&link->wait_list, &lock->frozen); goto unlock; } - lock->exclusive |= (fop->flags & EC_FLAG_LOCK_SHARED) == 0; + /* The lock is not marked to be released, so the frozen list should be + * empty. */ + GF_ASSERT(list_empty(&lock->frozen)); + + timer_link = ec_lock_timer_cancel(fop->xl, lock); if (!list_empty(&lock->owners)) { - if (!lock->acquired || (lock->exclusive != 0)) { + /* There are other owners of this lock. We can only take ownership if + * the lock is already acquired and doesn't have conflict with existing + * owners, or waiters(to prevent starvation). + * Otherwise we need to wait. + */ + if (!lock->acquired || ec_link_has_lock_conflict(link, _gf_true)) { ec_trace("LOCK_QUEUE_WAIT", fop, "lock=%p", lock); list_add_tail(&link->wait_list, &lock->waiting); goto unlock; } - } else if (lock->timer != NULL) { - GF_ASSERT (lock->release == _gf_false); - - timer_link = lock->timer->data; - if (gf_timer_call_cancel(fop->xl->ctx, lock->timer) == 0) { - ec_trace("UNLOCK_CANCELLED", timer_link->fop, "lock=%p", lock); - lock->timer = NULL; - lock->refs--; - /* There should remain at least 1 ref, the current one. */ - GF_ASSERT(lock->refs > 0); - } else { - /* Timer expired and on the way to unlock. - * Set lock->release to _gf_true, so that this - * lock will be put in frozen list*/ - timer_link = NULL; - lock->release = _gf_true; - } } - list_add_tail(&fop->owner_list, &lock->owners); + list_add_tail(&link->owner_list, &lock->owners); + + /* If timer_link is not NULL, it means that we have inherited the owner + * reference assigned to the timer fop. In this case we simply reuse it. + * Otherwise we need to increase the number of owners. */ + if (timer_link == NULL) { + lock->refs_owners++; + } assigned = _gf_true; unlock: if (!assigned) { + /* We have not been able to take ownership of this lock. The fop must + * be put to sleep. */ ec_sleep(fop); } UNLOCK(&lock->loc.inode->lock); + /* If we have cancelled the timer, we need to resume the fop that was + * waiting for it. */ if (timer_link != NULL) { ec_resume(timer_link->fop, 0); } @@ -1530,7 +2123,6 @@ ec_lock_next_owner(ec_lock_link_t *link, ec_cbk_data_t *cbk, ec_lock_t *lock = link->lock; ec_fop_data_t *fop = link->fop; ec_inode_t *ctx = lock->ctx; - ec_t *ec = fop->xl->private; INIT_LIST_HEAD(&list); @@ -1538,38 +2130,46 @@ ec_lock_next_owner(ec_lock_link_t *link, ec_cbk_data_t *cbk, ec_trace("LOCK_DONE", fop, "lock=%p", lock); - GF_ASSERT(!list_empty(&fop->owner_list)); - list_del_init(&fop->owner_list); + /* Current link must belong to the owner list of the lock. We don't + * decrement lock->refs_owners here because the inode mutex is released + * before ec_unlock() is called and we need to know when the last owner + * unlocks the lock to do proper cleanup. lock->refs_owners is used for + * this task. */ + GF_ASSERT((lock->refs_owners > 0) && !list_empty(&link->owner_list)); + list_del_init(&link->owner_list); + lock->release |= release; if ((fop->error == 0) && (cbk != NULL) && (cbk->op_ret >= 0)) { if (link->update[0]) { ctx->post_version[0]++; - if (ec->node_mask & ~fop->good) { - ctx->dirty[0]++; - } } if (link->update[1]) { ctx->post_version[1]++; - if (ec->node_mask & ~fop->good) { - ctx->dirty[1]++; + } + /* If the fop fails on any of the good bricks, it is important to mark + * it dirty and update versions right away. */ + if (link->update[0] || link->update[1]) { + if (lock->good_mask & ~(fop->good | fop->remaining)) { + lock->release = _gf_true; } } } + if (fop->healing) { + lock->healing = fop->healing & (fop->good | fop->remaining); + } ec_lock_update_good(lock, fop); - lock->exclusive -= (fop->flags & EC_FLAG_LOCK_SHARED) == 0; - if (list_empty(&lock->owners)) { - ec_lock_wake_shared(lock, &list); - } + ec_lock_wake_shared(lock, &list); UNLOCK(&lock->loc.inode->lock); ec_lock_resume_shared(&list); } -void ec_lock(ec_fop_data_t *fop) +void +ec_lock(ec_fop_data_t *fop) { ec_lock_link_t *link; @@ -1577,7 +2177,7 @@ void ec_lock(ec_fop_data_t *fop) * Which can result in refs == 0 for fop leading to use after free in this * function when it calls ec_sleep so do ec_sleep at start and ec_resume at * the end of this function.*/ - ec_sleep (fop); + ec_sleep(fop); while (fop->locked < fop->lock_count) { /* Since there are only up to 2 locks per fop, this xor will change @@ -1597,6 +2197,7 @@ ec_lock_unfreeze(ec_lock_link_t *link) { struct list_head list; ec_lock_t *lock; + gf_boolean_t destroy = _gf_false; lock = link->lock; @@ -1604,18 +2205,30 @@ ec_lock_unfreeze(ec_lock_link_t *link) LOCK(&lock->loc.inode->lock); - lock->acquired = _gf_false; + /* The lock must be marked to be released here, since we have just released + * it and any attempt to assign it to more fops must have added them to the + * frozen list. We can only have one active reference here: the one that + * is processing this unfreeze. */ + GF_ASSERT(lock->release && (lock->refs_owners == 1)); lock->release = _gf_false; - lock->refs--; + lock->refs_owners = 0; - GF_ASSERT (lock->refs == lock->inserted); - GF_ASSERT(lock->exclusive == 0); - GF_ASSERT(list_empty(&lock->waiting) && list_empty(&lock->owners)); + lock->acquired = _gf_false; + + /* We are unfreezing a lock. This means that the lock has already been + * released. In this state it shouldn't have a pending timer nor have any + * owner, and the waiting list should be empty. Only the frozen list can + * contain some fop. */ + GF_ASSERT((lock->timer == NULL) && list_empty(&lock->waiting) && + list_empty(&lock->owners)); + /* We move all frozen fops to the waiting list. */ list_splice_init(&lock->frozen, &lock->waiting); - lock->refs += lock->refs_frozen; - lock->refs_frozen = 0; - if (lock->refs == 0) { + + /* If we don't have any fop waiting nor there are any prepared fops using + * this lock, we can finally dispose it. */ + destroy = list_empty(&lock->waiting) && (lock->refs_pending == 0); + if (destroy) { ec_trace("LOCK_DESTROY", link->fop, "lock=%p", lock); lock->ctx->inode_lock = NULL; @@ -1629,22 +2242,21 @@ ec_lock_unfreeze(ec_lock_link_t *link) ec_lock_resume_shared(&list); - if (lock->refs == 0) { + if (destroy) { ec_lock_destroy(lock); } } -int32_t ec_unlocked(call_frame_t *frame, void *cookie, xlator_t *this, - int32_t op_ret, int32_t op_errno, dict_t *xdata) +int32_t +ec_unlocked(call_frame_t *frame, void *cookie, xlator_t *this, int32_t op_ret, + int32_t op_errno, dict_t *xdata) { ec_fop_data_t *fop = cookie; ec_lock_link_t *link = fop->data; if (op_ret < 0) { - gf_msg (this->name, GF_LOG_WARNING, op_errno, - EC_MSG_UNLOCK_FAILED, - "entry/inode unlocking failed (%s)", - ec_fop_name(link->fop->id)); + gf_msg(this->name, GF_LOG_WARNING, op_errno, EC_MSG_UNLOCK_FAILED, + "entry/inode unlocking failed :(%s)", ec_msg_str(link->fop)); } else { ec_trace("UNLOCKED", link->fop, "lock=%p", link->lock); } @@ -1654,24 +2266,26 @@ int32_t ec_unlocked(call_frame_t *frame, void *cookie, xlator_t *this, return 0; } -void ec_unlock_lock(ec_lock_link_t *link) +void +ec_unlock_lock(ec_lock_link_t *link) { ec_lock_t *lock; ec_fop_data_t *fop; + gf_lkowner_t lk_owner; lock = link->lock; fop = link->fop; + lock->unlock_now = _gf_false; ec_clear_inode_info(fop, lock->loc.inode); if ((lock->mask != 0) && lock->acquired) { - ec_owner_set(fop->frame, lock); - + set_lk_owner_from_ptr(&lk_owner, lock); lock->flock.l_type = F_UNLCK; ec_trace("UNLOCK_INODELK", fop, "lock=%p, inode=%p", lock, lock->loc.inode); - ec_inodelk(fop->frame, fop->xl, lock->mask, EC_MINIMUM_ONE, + ec_inodelk(fop->frame, fop->xl, &lk_owner, lock->mask, EC_MINIMUM_ONE, ec_unlocked, link, fop->xl->name, &lock->loc, F_SETLK, &lock->flock, NULL); } else { @@ -1679,25 +2293,49 @@ void ec_unlock_lock(ec_lock_link_t *link) } } -int32_t ec_update_size_version_done(call_frame_t * frame, void * cookie, - xlator_t * this, int32_t op_ret, - int32_t op_errno, dict_t * xattr, - dict_t * xdata) +void +ec_inode_bad_inc(inode_t *inode, xlator_t *xl) +{ + ec_inode_t *ctx = NULL; + + LOCK(&inode->lock); + { + ctx = __ec_inode_get(inode, xl); + if (ctx == NULL) { + goto unlock; + } + ctx->bad_version++; + } +unlock: + UNLOCK(&inode->lock); +} + +int32_t +ec_update_size_version_done(call_frame_t *frame, void *cookie, xlator_t *this, + int32_t op_ret, int32_t op_errno, dict_t *xattr, + dict_t *xdata) { ec_fop_data_t *fop = cookie; ec_lock_link_t *link; ec_lock_t *lock; ec_inode_t *ctx; + link = fop->data; + lock = link->lock; + ctx = lock->ctx; + if (op_ret < 0) { - gf_msg(fop->xl->name, fop_log_level (fop->id, op_errno), op_errno, + if (link->lock->fd == NULL) { + ec_inode_bad_inc(link->lock->loc.inode, this); + } else { + ec_inode_bad_inc(link->lock->fd->inode, this); + } + + gf_msg(fop->xl->name, fop_log_level(fop->id, op_errno), op_errno, EC_MSG_SIZE_VERS_UPDATE_FAIL, - "Failed to update version and size"); + "Failed to update version and size. %s", ec_msg_str(fop)); } else { fop->parent->good &= fop->good; - link = fop->data; - lock = link->lock; - ctx = lock->ctx; ec_lock_update_good(lock, fop); @@ -1714,16 +2352,17 @@ int32_t ec_update_size_version_done(call_frame_t * frame, void * cookie, ctx->have_size = _gf_true; } if ((ec_dict_del_config(xdata, EC_XATTR_CONFIG, &ctx->config) == 0) && - ec_config_check(fop->parent, &ctx->config)) { + ec_config_check(fop->xl, &ctx->config)) { ctx->have_config = _gf_true; } ctx->have_info = _gf_true; } - - if ((fop->parent->id != GF_FOP_FLUSH) && - (fop->parent->id != GF_FOP_FSYNC) && - (fop->parent->id != GF_FOP_FSYNCDIR)) { + /* If we are here because of fop's and other than unlock request, + * that means we are still holding a lock. That make sure + * lock->unlock_now can not be modified. + */ + if (lock->unlock_now) { ec_unlock_lock(fop->data); } @@ -1731,21 +2370,19 @@ int32_t ec_update_size_version_done(call_frame_t * frame, void * cookie, } void -ec_update_size_version(ec_lock_link_t *link, uint64_t *version, - uint64_t size, uint64_t *dirty) +ec_update_size_version(ec_lock_link_t *link, uint64_t *version, uint64_t size, + uint64_t *dirty) { ec_fop_data_t *fop; ec_lock_t *lock; ec_inode_t *ctx; - dict_t * dict; - uid_t uid; - gid_t gid; + dict_t *dict = NULL; + uintptr_t update_on = 0; int32_t err = -ENOMEM; fop = link->fop; - - GF_ASSERT(version[0] < 0x100000000); - GF_ASSERT(version[1] < 0x100000000); + lock = link->lock; + ctx = lock->ctx; ec_trace("UPDATE", fop, "version=%ld/%ld, size=%ld, dirty=%ld/%ld", version[0], version[1], size, dirty[0], dirty[1]); @@ -1755,9 +2392,6 @@ ec_update_size_version(ec_lock_link_t *link, uint64_t *version, goto out; } - lock = link->lock; - ctx = lock->ctx; - /* If we don't have version information or it has been modified, we * update it. */ if (!ctx->have_version || (version[0] != 0) || (version[1] != 0)) { @@ -1769,8 +2403,8 @@ ec_update_size_version(ec_lock_link_t *link, uint64_t *version, } if (size != 0) { - /* If size has been changed, we should already know the previous size - * of the file. */ + /* If size has been changed, we should already + * know the previous size of the file. */ GF_ASSERT(ctx->have_size); err = ec_dict_set_number(dict, EC_XATTR_SIZE, size); @@ -1779,40 +2413,37 @@ ec_update_size_version(ec_lock_link_t *link, uint64_t *version, } } - /* If we don't have dirty information or it has been modified, we update - * it. */ - if ((dirty[0] != 0) || (dirty[1] != 0)) { + if (dirty[0] || dirty[1]) { err = ec_dict_set_array(dict, EC_XATTR_DIRTY, dirty, EC_VERSION_SIZE); if (err != 0) { goto out; } } - /* If config information is not know, we request it now. */ + /* If config information is not known, we request it now. */ if ((lock->loc.inode->ia_type == IA_IFREG) && !ctx->have_config) { /* A failure requesting this xattr is ignored because it's not * absolutely required right now. */ - ec_dict_set_number(dict, EC_XATTR_CONFIG, 0); + (void)ec_dict_set_number(dict, EC_XATTR_CONFIG, 0); } - uid = fop->frame->root->uid; - gid = fop->frame->root->gid; - fop->frame->root->uid = 0; fop->frame->root->gid = 0; + update_on = lock->good_mask | lock->healing; + if (link->lock->fd == NULL) { - ec_xattrop(fop->frame, fop->xl, fop->good, EC_MINIMUM_MIN, - ec_update_size_version_done, link, &link->lock->loc, - GF_XATTROP_ADD_ARRAY64, dict, NULL); + ec_xattrop(fop->frame, fop->xl, update_on, EC_MINIMUM_MIN, + ec_update_size_version_done, link, &link->lock->loc, + GF_XATTROP_ADD_ARRAY64, dict, NULL); } else { - ec_fxattrop(fop->frame, fop->xl, fop->good, EC_MINIMUM_MIN, - ec_update_size_version_done, link, link->lock->fd, - GF_XATTROP_ADD_ARRAY64, dict, NULL); + ec_fxattrop(fop->frame, fop->xl, update_on, EC_MINIMUM_MIN, + ec_update_size_version_done, link, link->lock->fd, + GF_XATTROP_ADD_ARRAY64, dict, NULL); } - fop->frame->root->uid = uid; - fop->frame->root->gid = gid; + fop->frame->root->uid = fop->uid; + fop->frame->root->gid = fop->gid; dict_unref(dict); @@ -1825,8 +2456,12 @@ out: ec_fop_set_error(fop, -err); - gf_msg (fop->xl->name, GF_LOG_ERROR, -err, EC_MSG_SIZE_VERS_UPDATE_FAIL, - "Unable to update version and size"); + gf_msg(fop->xl->name, GF_LOG_ERROR, -err, EC_MSG_SIZE_VERS_UPDATE_FAIL, + "Unable to update version and size. %s", ec_msg_str(fop)); + + if (lock->unlock_now) { + ec_unlock_lock(fop->data); + } } gf_boolean_t @@ -1834,28 +2469,62 @@ ec_update_info(ec_lock_link_t *link) { ec_lock_t *lock; ec_inode_t *ctx; - uint64_t version[2]; - uint64_t dirty[2]; + uint64_t version[2] = {0, 0}; + uint64_t dirty[2] = {0, 0}; uint64_t size; + ec_t *ec = NULL; + uintptr_t mask; lock = link->lock; ctx = lock->ctx; + ec = link->fop->xl->private; /* pre_version[*] will be 0 if have_version is false */ - version[0] = ctx->post_version[0] - ctx->pre_version[0]; - version[1] = ctx->post_version[1] - ctx->pre_version[1]; + version[EC_DATA_TXN] = ctx->post_version[EC_DATA_TXN] - + ctx->pre_version[EC_DATA_TXN]; + version[EC_METADATA_TXN] = ctx->post_version[EC_METADATA_TXN] - + ctx->pre_version[EC_METADATA_TXN]; size = ctx->post_size - ctx->pre_size; + /* If we set the dirty flag for update fop, we have to unset it. + * If fop has failed on some bricks, leave the dirty as marked. */ + + if (lock->unlock_now) { + if (version[EC_DATA_TXN]) { + /*A data fop will have difference in post and pre version + *and for data fop we send writes on healing bricks also */ + mask = lock->good_mask | lock->healing; + } else { + mask = lock->good_mask; + } + /* Ensure that nodes are up while doing final + * metadata update.*/ + if (!(ec->node_mask & ~(mask)) && !(ec->node_mask & ~ec->xl_up)) { + if (ctx->dirty[EC_DATA_TXN] != 0) { + dirty[EC_DATA_TXN] = -1; + } + if (ctx->dirty[EC_METADATA_TXN] != 0) { + dirty[EC_METADATA_TXN] = -1; + } + /*If everything is fine and we already + *have version xattr set on entry, there + *is no need to update version again*/ + if (ctx->pre_version[EC_DATA_TXN]) { + version[EC_DATA_TXN] = 0; + } + if (ctx->pre_version[EC_METADATA_TXN]) { + version[EC_METADATA_TXN] = 0; + } + } else { + link->optimistic_changelog = _gf_false; + ec_set_dirty_flag(link, ctx, dirty); + } + memset(ctx->dirty, 0, sizeof(ctx->dirty)); + } - dirty[0] = ctx->dirty[0]; - dirty[1] = ctx->dirty[1]; - /*Dirty is not combined so just reset it right here*/ - memset(ctx->dirty, 0, sizeof(ctx->dirty)); - - if ((version[0] != 0) || (version[1] != 0) || - (dirty[0] != 0) || (dirty[1] != 0)) { + if ((version[EC_DATA_TXN] != 0) || (version[EC_METADATA_TXN] != 0) || + (dirty[EC_DATA_TXN] != 0) || (dirty[EC_METADATA_TXN] != 0)) { ec_update_size_version(link, version, size, dirty); - return _gf_true; } @@ -1865,7 +2534,15 @@ ec_update_info(ec_lock_link_t *link) void ec_unlock_now(ec_lock_link_t *link) { + ec_lock_t *lock; + lock = link->lock; + ec_trace("UNLOCK_NOW", link->fop, "lock=%p", link->lock); + /*At this point, lock is not being used by any fop and + *can not be reused by any fop as it is going to be released. + *lock->unlock_now can not be modified at any other place. + */ + lock->unlock_now = _gf_true; if (!ec_update_info(link)) { ec_unlock_lock(link); @@ -1875,102 +2552,254 @@ ec_unlock_now(ec_lock_link_t *link) } void +ec_lock_release(ec_t *ec, inode_t *inode) +{ + ec_lock_t *lock; + ec_inode_t *ctx; + ec_lock_link_t *timer_link = NULL; + + LOCK(&inode->lock); + + ctx = __ec_inode_get(inode, ec->xl); + if (ctx == NULL) { + goto done; + } + lock = ctx->inode_lock; + if ((lock == NULL) || lock->release) { + goto done; + } + + gf_msg_debug(ec->xl->name, 0, "Releasing inode %p due to lock contention", + inode); + + if (!lock->acquired) { + /* This happens if some bricks already got the lock while inodelk is in + * progress. Set release to true after lock is acquired*/ + lock->contention = _gf_true; + goto done; + } + + /* The lock is not marked to be released, so the frozen list should be + * empty. */ + GF_ASSERT(list_empty(&lock->frozen)); + + timer_link = ec_lock_timer_cancel(ec->xl, lock); + + /* We mark the lock to be released as soon as possible. */ + lock->release = _gf_true; + +done: + UNLOCK(&inode->lock); + + /* If we have cancelled the timer, we need to start the unlock of the + * inode. If there was a timer but we have been unable to cancel it + * because it was just triggered, the timer callback will take care + * of releasing the inode. */ + if (timer_link != NULL) { + ec_unlock_now(timer_link); + } +} + +void +ec_unlock_timer_add(ec_lock_link_t *link); + +void ec_unlock_timer_del(ec_lock_link_t *link) { - int32_t before = 0; - ec_lock_t *lock; - inode_t *inode; - gf_boolean_t now = _gf_false; + ec_lock_t *lock; + inode_t *inode; + gf_boolean_t now = _gf_false; - lock = link->lock; + /* If we are here, it means that the timer has expired before having + * been cancelled. This guarantees that 'link' is still valid because + * the fop that contains it must be pending (if timer cancellation in + * ec_lock_assign_owner() fails, the fop is left sleeping). + * + * At the same time, the fop still has a reference to the lock, so + * it must also be valid. + */ + lock = link->lock; - /* A race condition can happen if timer expires, calls this function - * and the lock is released (lock->loc is wiped) but the fop is not - * fully completed yet (it's still on the list of pending fops). In - * this case, this function can also be called if ec_unlock_force() is - * called. */ - inode = lock->loc.inode; - if (inode == NULL) { - return; - } + /* 'lock' must have a valid inode since it can only be destroyed + * when the lock itself is destroyed, but we have a reference to the + * lock to avoid this. + */ + inode = lock->loc.inode; - LOCK(&inode->lock); + LOCK(&inode->lock); - if (lock->timer != NULL) { - ec_trace("UNLOCK_DELAYED", link->fop, "lock=%p", lock); + if (lock->timer != NULL) { + ec_trace("UNLOCK_DELAYED", link->fop, "lock=%p", lock); - gf_timer_call_cancel(link->fop->xl->ctx, lock->timer); - lock->timer = NULL; + /* The unlock timer has expired without anyone cancelling it. + * This means that it shouldn't have any owner, and the waiting + * and frozen lists should be empty. It must have only one + * owner reference, but there can be fops being prepared + * though. + * */ + GF_ASSERT(!lock->release && (lock->refs_owners == 1) && + list_empty(&lock->owners) && list_empty(&lock->waiting) && + list_empty(&lock->frozen)); - lock->release = now = _gf_true; + gf_timer_call_cancel(link->fop->xl->ctx, lock->timer); + lock->timer = NULL; - /* TODO: If the assertion is really true, following code is - * not needed. */ - GF_ASSERT(list_empty(&lock->waiting)); + /* Any fop being processed from now on, will need to wait + * until the next unlock/lock cycle. */ + lock->release = now = _gf_true; + } - before = lock->refs + lock->refs_frozen; - list_splice_init(&lock->waiting, &lock->frozen); - lock->refs_frozen += lock->refs - lock->inserted - 1; - lock->refs = 1 + lock->inserted; - /* We moved around the locks, so total number of locks shouldn't - * change by this operation*/ - GF_ASSERT (before == (lock->refs + lock->refs_frozen)); - } + UNLOCK(&inode->lock); - UNLOCK(&inode->lock); + if (now) { + ec_unlock_now(link); + } else { + /* The timer has been cancelled just after firing it but before + * getting here. This means that another fop has used the lock + * and everything should be handled as if this callback were + * have not been executed. However we still have an owner + * reference. + * + * We need to release our reference. If this is not the last + * reference (the most common case because another fop has + * taken another ref) we only need to decrement the counter. + * Otherwise we have been delayed enough so that the other fop + * has had time to acquire the reference, do its operation and + * release it. At the time of releasing it, the fop did found + * that the ref counter was > 1 (our reference), so the delayed + * unlock timer wasn't started. We need to start it again if we + * are the last reference. + * + * ec_unlock_timer_add() handles both cases. + */ + ec_unlock_timer_add(link); - if (now) { - ec_unlock_now(link); - } + /* We need to resume the fop that was waiting for the delayed + * unlock. + */ + ec_resume(link->fop, 0); + } +} + +void +ec_unlock_timer_cbk(void *data) +{ + ec_unlock_timer_del(data); } -void ec_unlock_timer_cbk(void *data) +static gf_boolean_t +ec_eager_lock_used(ec_t *ec, ec_fop_data_t *fop) { - ec_unlock_timer_del(data); + /* Fops with no locks at this point mean that they are sent as sub-fops + * of other higher level fops. In this case we simply assume that the + * parent fop will take correct care of the eager lock. */ + if (fop->lock_count == 0) { + return _gf_true; + } + + /* We may have more than one lock, but this only happens in the rename + * fop, and both locks will reference an inode of the same type (a + * directory in this case), so we only need to check the first lock. */ + if (fop->locks[0].lock->loc.inode->ia_type == IA_IFREG) { + return ec->eager_lock; + } + + return ec->other_eager_lock; } -void ec_unlock_timer_add(ec_lock_link_t *link) +static uint32_t +ec_eager_lock_timeout(ec_t *ec, ec_lock_t *lock) +{ + if (lock->loc.inode->ia_type == IA_IFREG) { + return ec->eager_lock_timeout; + } + + return ec->other_eager_lock_timeout; +} + +static gf_boolean_t +ec_lock_delay_create(ec_lock_link_t *link) { struct timespec delay; ec_fop_data_t *fop = link->fop; ec_lock_t *lock = link->lock; + + delay.tv_sec = ec_eager_lock_timeout(fop->xl->private, lock); + delay.tv_nsec = 0; + lock->timer = gf_timer_call_after(fop->xl->ctx, delay, ec_unlock_timer_cbk, + link); + if (lock->timer == NULL) { + gf_msg(fop->xl->name, GF_LOG_WARNING, ENOMEM, + EC_MSG_UNLOCK_DELAY_FAILED, "Unable to delay an unlock"); + + return _gf_false; + } + + return _gf_true; +} + +void +ec_unlock_timer_add(ec_lock_link_t *link) +{ + ec_fop_data_t *fop = link->fop; + ec_lock_t *lock = link->lock; gf_boolean_t now = _gf_false; LOCK(&lock->loc.inode->lock); - GF_ASSERT(lock->timer == NULL); + /* We are trying to unlock the lock. We can have multiple scenarios here, + * but all of them need to have lock->timer == NULL: + * + * 1. There are other owners currently running that can call ec_unlock(). + * + * None of them can have started the timer until the last one. But this + * call should be the consequence of this lastest one. + * + * 2. There are fops in the waiting or frozen lists. + * + * These fops cannot call ec_unlock(). So we should be here. + * + * We must reach here with at least one owner reference. + */ + GF_ASSERT((lock->timer == NULL) && (lock->refs_owners > 0)); - if ((lock->refs - lock->inserted) > 1) { + /* If the fop detects that a heal is needed, we mark the lock to be + * released as soon as possible. */ + lock->release |= ec_fop_needs_heal(fop); + + if (lock->refs_owners > 1) { ec_trace("UNLOCK_SKIP", fop, "lock=%p", lock); - lock->refs--; + /* If there are other owners we cannot do anything else with the lock. + * Note that the current fop has already been removed from the owners + * list in ec_lock_reuse(). */ + lock->refs_owners--; UNLOCK(&lock->loc.inode->lock); } else if (lock->acquired) { - ec_t *ec = fop->xl->private; + /* There are no other owners and the lock is acquired. If there were + * fops waiting, at least one of them should have been promoted to an + * owner, so the waiting list should be empty. */ + GF_ASSERT(list_empty(&lock->owners) && list_empty(&lock->waiting)); - GF_ASSERT(list_empty(&lock->owners)); + ec_t *ec = fop->xl->private; + /* If everything goes as expected this fop will be put to sleep until + * the timer callback is executed. */ ec_sleep(fop); - /* If healing is needed, the lock needs to be released due to - * contention, or ec is shutting down, do not delay lock release. */ - if (!lock->release && !ec_fop_needs_heal(fop) && !ec->shutdown) { + /* If the lock needs to be released, or ec is shutting down, do not + * delay lock release. */ + if (!lock->release && !ec->shutdown) { ec_trace("UNLOCK_DELAY", fop, "lock=%p, release=%d", lock, lock->release); - delay.tv_sec = 1; - delay.tv_nsec = 0; - lock->timer = gf_timer_call_after(fop->xl->ctx, delay, - ec_unlock_timer_cbk, link); - if (lock->timer == NULL) { - gf_msg(fop->xl->name, GF_LOG_WARNING, ENOMEM, - EC_MSG_UNLOCK_DELAY_FAILED, - "Unable to delay an " - "unlock"); - + if (!ec_lock_delay_create(link)) { + /* We are unable to create a new timer. We immediately release + * the lock. */ lock->release = now = _gf_true; } + } else { ec_trace("UNLOCK_FORCE", fop, "lock=%p, release=%d", lock, lock->release); @@ -1983,17 +2812,25 @@ void ec_unlock_timer_add(ec_lock_link_t *link) ec_unlock_now(link); } } else { + /* There are no owners and the lock is not acquired. This can only + * happen if a lock attempt has failed and we get to the unlock step + * of the fop. As in the previous case, the waiting list must be + * empty. */ + GF_ASSERT(list_empty(&lock->owners) && list_empty(&lock->waiting)); + + /* We need to mark the lock to be released to correctly handle fops + * that may get in after we release the inode mutex but before + * ec_lock_unfreeze() is processed. */ lock->release = _gf_true; - GF_ASSERT(list_empty(&lock->owners)); - UNLOCK(&lock->loc.inode->lock); ec_lock_unfreeze(link); } } -void ec_unlock(ec_fop_data_t *fop) +void +ec_unlock(ec_fop_data_t *fop) { int32_t i; @@ -2002,60 +2839,153 @@ void ec_unlock(ec_fop_data_t *fop) } } -void ec_flush_size_version(ec_fop_data_t * fop) +void +ec_flush_size_version(ec_fop_data_t *fop) { GF_ASSERT(fop->lock_count == 1); + ec_update_info(&fop->locks[0]); +} - /* In normal circumstances, ec_update_info() is called after having - * executed a normal fop, and it uses fop->good to update only those bricks - * that succeeded. In this case we haven't executed any fop, so fop->good - * is 0. We use the current good mask of the lock itself to send the - * updates.*/ - fop->good = fop->locks[0].lock->good_mask; +static void +ec_update_stripe(ec_t *ec, ec_stripe_list_t *stripe_cache, ec_stripe_t *stripe, + ec_fop_data_t *fop) +{ + off_t base; + + /* On write fops, we only update existing fragments if the write has + * succeeded. Otherwise, we remove them from the cache. */ + if ((fop->id == GF_FOP_WRITE) && (fop->answer != NULL) && + (fop->answer->op_ret >= 0)) { + base = stripe->frag_offset - fop->frag_range.first; + base *= ec->fragments; + + /* We check if the stripe offset falls inside the real region + * modified by the write fop (a write request is allowed, + * though uncommon, to write less bytes than requested). The + * current write fop implementation doesn't allow partial + * writes of fragments, so if there's no error, we are sure + * that a full stripe has been completely modified or not + * touched at all. The value of op_ret may not be a multiple + * of the stripe size because it depends on the requested + * size by the user, so we update the stripe if the write has + * modified at least one byte (meaning ec has written the full + * stripe). */ + if (base < fop->answer->op_ret + fop->head) { + memcpy(stripe->data, fop->vector[0].iov_base + base, + ec->stripe_size); + list_move_tail(&stripe->lru, &stripe_cache->lru); + + GF_ATOMIC_INC(ec->stats.stripe_cache.updates); + } + } else { + stripe->frag_offset = -1; + list_move(&stripe->lru, &stripe_cache->lru); - ec_update_info(&fop->locks[0]); + GF_ATOMIC_INC(ec->stats.stripe_cache.invals); + } +} + +static void +ec_update_cached_stripes(ec_fop_data_t *fop) +{ + uint64_t first; + uint64_t last; + ec_stripe_t *stripe = NULL; + ec_inode_t *ctx = NULL; + ec_stripe_list_t *stripe_cache = NULL; + inode_t *inode = NULL; + struct list_head *temp; + struct list_head sentinel; + + first = fop->frag_range.first; + /* 'last' represents the first stripe not touched by the operation */ + last = fop->frag_range.last; + + /* If there are no modified stripes, we don't need to do anything + * else. */ + if (last <= first) { + return; + } + + if (!fop->use_fd) { + inode = fop->loc[0].inode; + } else { + inode = fop->fd->inode; + } + + LOCK(&inode->lock); + + ctx = __ec_inode_get(inode, fop->xl); + if (ctx == NULL) { + goto out; + } + stripe_cache = &ctx->stripe_cache; + + /* Since we'll be moving elements of the list to the tail, we might + * end in an infinite loop. To avoid it, we insert a sentinel element + * into the list, so that it will be used to detect when we have + * traversed all existing elements once. */ + list_add_tail(&sentinel, &stripe_cache->lru); + temp = stripe_cache->lru.next; + while (temp != &sentinel) { + stripe = list_entry(temp, ec_stripe_t, lru); + temp = temp->next; + if ((first <= stripe->frag_offset) && (stripe->frag_offset < last)) { + ec_update_stripe(fop->xl->private, stripe_cache, stripe, fop); + } + } + list_del(&sentinel); + +out: + UNLOCK(&inode->lock); } -void ec_lock_reuse(ec_fop_data_t *fop) +void +ec_lock_reuse(ec_fop_data_t *fop) { ec_cbk_data_t *cbk; + ec_t *ec = NULL; int32_t i, count; gf_boolean_t release = _gf_false; - + ec = fop->xl->private; cbk = fop->answer; - if (cbk != NULL) { + + if (ec_eager_lock_used(ec, fop) && cbk != NULL) { if (cbk->xdata != NULL) { - if ((dict_get_int32(cbk->xdata, GLUSTERFS_INODELK_COUNT, - &count) == 0) && (count > 1)) { + if ((dict_get_int32(cbk->xdata, GLUSTERFS_INODELK_COUNT, &count) == + 0) && + (count > 1)) { release = _gf_true; } if (release) { - gf_msg_debug (fop->xl->name, 0, - "Lock contention detected"); + gf_msg_debug(fop->xl->name, 0, "Lock contention detected"); } } } else { - /* If we haven't get an answer with enough quorum, we always release + /* If eager lock is disabled or if we haven't get + * an answer with enough quorum, we always release * the lock. */ release = _gf_true; } + ec_update_cached_stripes(fop); for (i = 0; i < fop->lock_count; i++) { ec_lock_next_owner(&fop->locks[i], cbk, release); } } -void __ec_manager(ec_fop_data_t * fop, int32_t error) +void +__ec_manager(ec_fop_data_t *fop, int32_t error) { ec_t *ec = fop->xl->private; do { ec_trace("MANAGER", fop, "error=%d", error); - if (!ec_must_wind (fop)) { - if (ec->xl_up_count < ec->fragments) { - error = ENOTCONN; - } + if (!ec_must_wind(fop)) { + if (ec->xl_up_count < ec->fragments) { + error = ENOTCONN; + } } if (error != 0) { @@ -2081,22 +3011,32 @@ void __ec_manager(ec_fop_data_t * fop, int32_t error) fop->jobs = 1; fop->state = fop->handler(fop, fop->state); - GF_ASSERT (fop->state >= 0); + GF_ASSERT(fop->state >= 0); error = ec_check_complete(fop, __ec_manager); } while (error >= 0); } -void ec_manager(ec_fop_data_t * fop, int32_t error) +void +ec_manager(ec_fop_data_t *fop, int32_t error) { GF_ASSERT(fop->jobs == 0); GF_ASSERT(fop->winds == 0); GF_ASSERT(fop->error == 0); - if (fop->state == EC_STATE_START) - { + if (fop->state == EC_STATE_START) { fop->state = EC_STATE_INIT; } __ec_manager(fop, error); } + +gf_boolean_t +__ec_is_last_fop(ec_t *ec) +{ + if ((list_empty(&ec->pending_fops)) && + (GF_ATOMIC_GET(ec->async_fop_count) == 0)) { + return _gf_true; + } + return _gf_false; +} |